ADVANCES IN IMAGING AND ELECTRON PHYSICS VOLUME 126
EDITOR-IN-CHIEF
PETER W. HAWKES CEMES-CNRS Toulouse, France
ASSOCIATE EDITORS
BENJAMIN KAZAN Xerox Corporation Palo Alto Research Center Palo Alto, California
TOM MULVEY Department of Electronic Engineering and Applied Physics Aston University Birmingham, United Kingdom
Advances in
Imaging and Electron Physics Edited by
PETER W. HAWKES CEMES-CNRS Toulouse, France
VOLUME 126
Amsterdam Boston London New York Oxford Paris San Diego San Francisco Singapore Sydney Tokyo
This book is printed on acid-free paper. Copyright ß 2003, Elsevier Science (USA).
All Rights Reserved. No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopy, recording, or any information storage and retrieval system, without permission in writing from the Publisher. The appearance of the code at the bottom of the first page of a chapter in this book indicates the Publisher’s consent that copies of the chapter may be made for personal or internal use of specific clients. This consent is given on the condition, however, that the copier pay the stated per copy fee through the Copyright Clearance Center, Inc. (222 Rosewood Drive, Danvers, Massachusetts 01923), for copying beyond that permitted by Sections 107 or 108 of the U.S. Copyright Law. This consent does not extend to other kinds of copying, such as copying for general distribution, for advertising or promotional purposes, for creating new collective works, or for resale. Copy fees for pre-2003 chapters are as shown on the title pages. If no fee code appears on the title page, the copy fee is the same as for current chapters. 1076-5670/2003 $35.00 Permissionions may be sought directly from Elsevier’s Science & Technology Rights Department in Oxford, UK: phone: (+44) 1865 843830, fax: (+44) 1865 853333, e-mail:
[email protected]. You may also complete your request on-line via the Elsevier Science homepage (http://elsevier.com), by selecting ‘‘Customer Support’’ and then ‘‘Obtaining Permissions.’’
Academic Press An Elsevier Science Imprint. 525 B Street, Suite 1900, San Diego, California 92101-4495, USA http://www.academicpress.com
Academic Press 84 Theobald’s Road, London WC1X 8RR, UK http://www.academicpress.com International Standard Book Number: 0-12-014768-8 PRINTED IN THE UNITED STATES OF AMERICA 03 04 05 06 07 08 9 8 7 6 5 4 3
2
1
CONTENTS
Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Future Contributions . . . . . . . . . . . . . . . . . . . . . . . .
ix xi xiii
A Wavelet-Based Method for Multifractal Image Analysis: From Theoretical Concepts to Experimental Applications A. ArnØodo, N. Decoster, P. Kestener, and S. G. Roux I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . II. Image Processing with the 2D Continuous Wavelet Transform III. Test Applications of the WTMM Method to Monofractal and Multifractal Rough Surfaces . . . . . . . . . . . . . . . . IV. Multifractal Analysis of High-Resolution Satellite Images of Cloud Structure . . . . . . . . . . . . . . . . . . . . . . . . V. Multifractal Analysis of 3D Turbulence Simulation Data . . . VI. Multifractal Analysis of Digitized Mammograms . . . . . . . VII. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2 7 23 41 53 73 80 82
An Analysis of the Geometric Distortions Produced by Median and Related Image Processing Filters E. R. Davies I. II. III. IV. V. VI. VII. VIII. IX. X. XI. XII.
Introduction . . . . . . . . . . . . . . . . . . . . . . . . Image Filters . . . . . . . . . . . . . . . . . . . . . . . . Shifts Produced by Median Filters in Continuous Images Shifts Produced by Median Filters in Digital Images . . Shifts Produced by Mean Filters . . . . . . . . . . . . . Shifts Produced by Mode Filters . . . . . . . . . . . . . Shifts Produced by Rank-Order Filters . . . . . . . . . . Rank-Order Filters—a Didactic Example . . . . . . . . . A Problem with Closing . . . . . . . . . . . . . . . . . . A Median-Based Corner Detector . . . . . . . . . . . . Boundary Length Measurement Problem . . . . . . . . . Concluding Remarks . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . .
v
. . . . . . . . . . . . .
. . . . . . . . . . . . .
. . . . . . . . . . . . .
94 96 105 122 146 150 156 170 174 178 183 188 191
vi
CONTENTS
Two-Photon Excitation Microscopy Alberto Diaspro and Giuseppe Chirico I. Introduction . . . . . . . . . . . . . . . . . . . . . . II. Historical Notes . . . . . . . . . . . . . . . . . . . . III. Basic Principles of Two-Photon Excitation of Fluorescent Molecules . . . . . . . . . . . . . . . . . IV. Behavior of Fluorescent Molecules under TPE Regime V. Optical Consequences and Resolution Aspects . . . . VI. Architecture of Two-Photon Microscopy . . . . . . . VII. Application Gallery . . . . . . . . . . . . . . . . . . VIII. Conclusions . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . .
195 198
. . . . . . .
. . . . . . .
202 212 219 225 257 273 276
. . . . . . . . . .
288 293
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
296 299 299 305 307 309 313 315 317 319 320 321 321 323 327
. . . .
. . . .
. . . .
. . . .
. . . .
329 330 337 340
. . . . . . .
. . . . . . .
Phase Closure Imaging AndrE´ Lannes I. Introduction . . . . . . . . . . . . . . . . . . . . . II. Phase Space and Integer Lattices . . . . . . . . . . III. Phase Closure Operator, Phase Closure Projection, and Related Properties . . . . . . . . . . . . . . . . IV. Variance–Covariance Matrix of the Closure Phases . V. Spectral Phase Closure Projection . . . . . . . . . . VI. Reference Algebraic Framework . . . . . . . . . . . VII. Statement of the Phase Calibration Problem . . . . VIII. Phase Calibration Discrepancy and Related Results IX. Optimal Model Phase Shift and Related Results . . X. Special Cases . . . . . . . . . . . . . . . . . . . . . XI. Simulated Example . . . . . . . . . . . . . . . . . . XII. Concluding Comments . . . . . . . . . . . . . . . . Appendix 1 . . . . . . . . . . . . . . . . . . . . . . Appendix 2 . . . . . . . . . . . . . . . . . . . . . . Appendix 3 . . . . . . . . . . . . . . . . . . . . . . Appendix 4 . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . .
Three-Dimensional Image Processing and Optical Scanning Holography Ting-Chung Poon I. II. III. IV.
Introduction . . . . . . . . . . . . . . . Two-Pupil Optical Heterodyne Scanning Three-Dimensional Imaging Properties . Optical Scanning Holography . . . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
vii
CONTENTS
V. Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . 347 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
Nonlinear Image Processing using Artificial Neural Networks Dick de Ridder, Robert P. W. Duin, Michael EgmontPetersen, Lucas J. Van Vliet, and Piet W. Verbeek I. II. III. IV. V. VI. VII.
Introduction . . . . . . . . . . . . . . . . . . . . . . Applications of ANNs in Image Processing . . . . . Shared Weight Networks for Object Recognition . . Feature Extraction in Shared Weight Networks . . . Regression Networks for Image Restoration . . . . . Inspection and Improvement of Regression Networks Conclusions . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
352 356 366 377 399 418 442 447
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
This Page Intentionally Left Blank
CONTRIBUTORS
Numbers in parentheses indicate the pages on which the authors’ contributions begins.
A. ArnØodo (1), Centre de Recherche Paul Pascal, 33600 Pessac, France Giuseppe Chirico (195), LAMBS-INFM and Department of Physics, University of Milano Biocca, 20126 Milano, Italy E. R. Davies (93), Machine Vision Group, Department of Physics, Royal Holloway College, University of London, Egham, Surrey TW20 0EX, United Kingdom N. Decoster (1), Noveltis, Parc Technologique du Canal, 31520 Ramonville Saint Agne, France Alberto Diaspro (195), LAMBS-INFM and Department of Physics, University of Genoa, 16146 Genova, Italy Robert P. W. Duin (351), Pattern Recognition Group, Department of Applied Physics, Delft University of Technology, 2628 CJ Delft, The Netherlands Michael Egmont-Petersen (351), Decision Support Systems Group, Institute of Information and Computing Sciences, Utrecht University, 3508 TB Utrecht, The Netherlands P. Kestener (1), Centre de Recherche Paul Pascal, 33600 Pessac, France AndrE´ Lannes (287), Sciences de l’Univers du Centre Europe´en de Recherche et de Formation Avance´e en Calcul Scientifique (Suc-Cerfacs), F-31057 Toulouse cedex, France Ting-Chung Poon (329), Optical Image Processing Laboratory, Bradley Department of Electrical and Computer Engineering, Virginia Polytechnic Institute and State University, Blacksburg, Virginia 24061 Dick de Ridder (351), Pattern Recognition Group, Department of Applied Physics, Delft University of Technology, 2628 CJ Delft, The Netherlands ix
x
CONTRIBUTORS
S. G. Roux (1), Laboratoire de Physique, Ecole Normale Supe´rieure de Lyon, 69364 Lyon cedex 07, France Lucas J. Van Vliet (351), Pattern Recognition Group, Department of Applied Physics, Delft University of Technology, 2628 CJ Delft, The Netherlands Piet W. Verbeek (351), Pattern Recognition Group, Department of Applied Physics, Delft University of Technology, 2628 CJ Delft, The Netherlands
PREFACE
This latest volume of these Advances is dominated by image processing and by a major contribution on microscopy, which is also the object of much of the image processing. The collection opens with a contribution by A. Arne´odo, N. Decoster, R. Kestener, and S. G. Roux on multifractal analysis, inspired by the need to find a consistent way of characterizing surface roughness. One of the authors of this chapter has shown that use of the continuous wavelet transform has many advantages here and that the method can be extended to the two-dimensional situation, which is of course of most practical interest. This chapter provides a full account of the method and of many realistic applications. The second contribution is by E. R. Davies, whose work on median and rank-order filters is well known in the field of image processing. Despite their attractive features, these nonlinear filters also create distortions and these are analysed very thoroughly in this chapter. The author examines median and rank-order filters, mode filters and morphological filters and discusses the shifts that they are liable to create. A knowledge of the artefacts that can be generated is essential for anyone using these filters. This brings us to a chapter by A. Diaspro and G. Chirico, who present a form of microscopy that is attracting great interest, namely, two-photon excitation microscopy. Alberto Diaspro is one of the leaders of this discipline and here, the principles of the technique and a range of applications are set out very clearly. Although two-photon microscopy is a fairly recent development, the basic physics has a long history, which is summarized at the beginning of the chapter with a wealth of historic illustrations. The problem of phase calibration is particularly acute in multi-antenna radio imaging. A radically new approach based on graph-theoretic reasoning has been pioneered by A. Lannes, who explains the procedure in the fifth chapter, entitled ‘Phase closure imaging’. The mathematical fundamentals are first presented and related to realistic practical situations, after which a number of special cases are explored. This very full statement of Lannes’ solution to an important and difficult problem is thus very welcome here. Optical scanning holography was introduced by T.-C. Poon, author of the fourth chapter. This technique is explained in terms of the two-pupil optical heterodyne scanning image processor, which leads on to the notion xi
xii
PREFACE
of three-dimensional point-spread functions. The sine- and cosine-Fresnel zone plate hologram and the complex hologram are then introduced and finally, the use of these for three-dimensional reconstruction is explained. This is an authoritative account of a very exciting development in imaging. We close with a long contribution on nonlinear image processing in which the operations are performed on artificial neural networks. Here, D. de Ridder and colleagues from the Pattern Recognition Group in Delft University and the Institute of Information and Computing Sciences in Utrecht University first explain how image processing is performed by neural networks. They describe the various kinds of network and then discuss in detail object recognition, feature extraction, image restoration and finally, the inspection and improvement of regression networks. This chapter has the status of a monograph on the subject and will, I am sure, be heavily used. In conclusion, I thank most sincerely all the contributors for taking so much trouble to make their specialized knowledge available to a wider audience and list the contributions planned for future volumes. Peter W. Hawkes
FUTURE CONTRIBUTIONS
T. Aach (vol. 128) Lapped transforms G. Abbate New developments in liquid-crystal-based photonic devices S. Ando Gradient operators and edge and corner detection C. Beeli Structure and microscopy of quasicrystals I. Bloch (vol. 128) Fuzzy distance measures in image processing G. Borgefors Distance transforms B. L. Breton, D. McMullan and K. C. A. Smith (Eds) Sir Charles Oatley and the scanning electron microscope A. Bretto Hypergraphs and their use in image modelling Y. Cho (vol. 127) Scanning nonlinear dielectric microscopy H. Delingette Surface reconstruction based on simplex meshes R. G. Forbes Liquid metal ion sources E. Fo¨rster and F. N. Chukhovsky X-ray optics A. Fox The critical-voltage effect L. Frank and I. Mu¨llerova´ (vol.128) Scanning low-energy electron microscopy L. Godo and V. Torra Aggregation operators xiii
xiv
FUTURE CONTRIBUTIONS
A. Go¨lzha¨user Recent advances in electron holography with point sources A. M. Grigoryan and S. S. Agaian Transform-based image enhancement algorithms with performance measure. A. Hanbury (vol. 128) Morphology on a circle H. F. Harmuth and B. Meffert Calculus of finite differences in quantum electrodynamics P. W. Hawkes (vol. 127) Electron optics and electron microscopy: conference proceedings and abstracts as source material M. I. Herrera The development of electron microscopy in Spain J. S. Hesthaven (vol. 127) Higher-order accuracy computational methods for time-domain electromagnetics D. Hitz Recent progress on HF ECR ion sources K. Ishizuka Contrast transfer and crystal images G. Ko¨gel Positron microscopy W. Krakow Sideband imaging N. Krueger The application of statistical and deterministic regularities in biological and artificial vision systems B. Lahme Karhunen-Loeve decomposition B. Lencova´ Modern developments in electron optical calculations M. A. O’Keefe Electron image simulation
FUTURE CONTRIBUTIONS
xv
N. Papamarkos and A. Kesidis The inverse Hough transform M. G. A. Paris and G. d’Ariano (vol. 128) Quantum tomography K. S. Pedersen, A. Lee and M. Nielsen The scale-space properties of natural images E. Petajan HDTV M. Petrou Image registration M. Rainforth Recent developments in the microscopy of ceramics, ferroelectric materials and glass E. Rau Energy analysers for electron microscopes H. Rauch The wave-particle dualism J. J. W. M. Rosink and N. van der Vaart HEC sources for the CRT O. Scherzer (vol. 128) Regularization techniques G. Schmahl X-ray microscopy S. Shirai CRT gun design methods T. Soma Focus-deflection systems and their applications J.-L. Starck The curvelet transform I. Talmon Study of complex fluids by transmission electron microscopy M. Tonouchi Terahertz radiation imaging
xvi
FUTURE CONTRIBUTIONS
N. M. Towghi Ip norm optimal filters Y. Uchikawa Electron gun optics D. van Dyck Very high resolution electron microscopy K. Vaeth and G. Rajeswaran Organic light-emitting arrays C. D. Wright and E. W. Hill Magnetic force microscopy F. Yang and M. Paindavoine (vol. 127) Pre-filtering for pattern recognition using wavelet transforms and neural networks M. Yeadon Instrumentation for surface studies
ADVANCES IN IMAGING AND ELECTRON PHYSICS, VOL. 126
A Wavelet-Based Method for Multifractal Image Analysis: From Theoretical Concepts to Experimental Applications A. ARNE´ODO,1 N. DECOSTER,2 P. KESTENER,1 AND S. G. ROUX3 1
Centre de Recherche Paul Pascal, Avenue Schweitzer, 33600, Pessac, France 2 Noveltis, Parc Technologique du Canal, 2 avenue de l’Europe, 31520, Ramonville Saint Agne, France 3 Laboratoire de Physique, Ecole Normale Supe´rieure de Lyon, 46 alle´e d’Italie, 69364, Lyon cedex 07, France
I. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . II. Image Processing with the 2D Continuous Wavelet Transform . . . . . . A. Analyzing Wavelets for Multiscale Edge Detection . . . . . . . . . B. Characterizing the Local Regularity Properties of Rough Surfaces with the Wavelet Transform Modulus Maxima . . . . . . . . . . 1. Isotropic Dilations . . . . . . . . . . . . . . . . . . . . 2. Anisotropic Dilations . . . . . . . . . . . . . . . . . . . C. The 2D Wavelet Transform Modulus Maxima (WTMM) Method . . . 1. Definition . . . . . . . . . . . . . . . . . . . . . . . 2. Methodology . . . . . . . . . . . . . . . . . . . . . . 3. Remark . . . . . . . . . . . . . . . . . . . . . . . . 4. Numerical Implementation . . . . . . . . . . . . . . . . . III. Test Applications of the WTMM Method to Monofractal and Multifractal Rough Surfaces . . . . . . . . . . . . . . . . . . . A. Fractional Brownian Surfaces . . . . . . . . . . . . . . . . . B. Multifractal Rough Surfaces Generated by Random Cascades on Separable Wavelet Orthogonal Basis . . . . . . . . . . . . . . 1. Remark . . . . . . . . . . . . . . . . . . . . . . . . C. Distinguishing ‘‘Multiplicative from Additive’’ Processes Underlying the Scale Invariance Properties of Rough Surfaces from Space-Scale Correlation Analysis . . . . . . . . . . . . . . . . . . . . D. Using the 2D WTMM Method to Perform Image Processing Tasks . . IV. Multifractal Analysis of High-Resolution Satellite Images of Cloud Structure A. Landsat Data of Marine Stratocumulus Cloud Scenes . . . . . . . . B. Application of the 2D WTMM Method to Landsat Images of Stratocumulus Clouds . . . . . . . . . . . . . . . . . . . . 1. Numerical Computation of the Multifractal (q) and D(h) Spectra . . 2. WTMMM Probability Density Functions . . . . . . . . . . . C. Space-Scale Correlation Function Analysis of Radiance Landsat Images . D. Comparative WTMM Multifractal Analysis of Landsat Radiance Field and Velocity and Temperature Fields in Fully Developed Turbulence . . V. Multifractal Analysis of 3D Turbulence Simulation Data . . . . . . . . A. Multifractal Description of Intermittency . . . . . . . . . . . . . 1. Intermittency Based on the Velocity Field . . . . . . . . . . .
. . . . . . . . . . . . . .
2 7 7
. . . . . . . .
9 10 10 17 17 18 19 21
. . . .
23 23
. . . .
31 35
. . . .
. . . .
36 38 41 43
. . . .
. . . .
44 45 48 50
. . . .
. . . .
51 53 53 53
1 Copyright 2003 Elsevier Science (USA). All rights reserved. ISSN 1076-5670/03
2
´ ODO ET AL. ARNE
2. Intermittency Based on the Energy Dissipation Field . . . . . . 3. Intermittency Based on the Enstrophy Field . . . . . . . . . . B. Application of the 2D WTMM Method to 2D Cuts of a Turbulent 3D Dissipation Field . . . . . . . . . . . . . . . . . . . . . 1. Remark . . . . . . . . . . . . . . . . . . . . . . . . 2. Numerical Computation of the ðqÞ and f ðqÞ Multifractal Spectra . 3. WTMMM Probability Density Functions . . . . . . . . . . . 4. Space-Scale Correlation Function Analysis . . . . . . . . . . C. Application of the 2D WTMM Method to 2D Cuts of a Turbulent 3D Enstrophy Field . . . . . . . . . . . . . . . . . . . . . . 1. Numerical Computation of the Multifractal ðqÞ and f ðqÞ Spectra . 2. WTMMM Probability Density Functions . . . . . . . . . . . 3. Space-Scale Correlation Function Analysis . . . . . . . . . . D. Discussion . . . . . . . . . . . . . . . . . . . . . . . . VI. Multifractal Analysis of Digitized Mammograms . . . . . . . . . . A. Application of the 2D WTMM Method to Mammographic Tissue Classification: Dense and Fatty Tissues . . . . . . . . . . . . . B. Detecting Microcalcifications through WT Skeleton Segmentation . . VII. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . .
. . . .
56 59
. . . . .
. . . . .
61 62 63 66 67
. . . . . .
. . . . . .
68 68 70 71 71 73
. . . .
. . . .
74 77 80 82
I. Introduction Ever since the explosive propagation of fractal ideas [1,2] throughout the scientific community in the late 1970s and early 1980s, there have been numerous applications to surface science [3–13]. Both real space imaging techniques (including scanning tunneling microscopy, atomic force microscopy, transmission electron microscopy, secondary electron microscopy, and optical imaging techniques) and diffraction techniques (including electron, atom, light, and X-ray scattering) have been extensively used to study rough surfaces [12]. The characterization of surface roughness is an important problem from a fundamental point of view as well as for the wealth of potential applications in applied sciences. Indeed, a wide variety of natural and technological processes lead to the formation of complex interfaces [1–18]. Assigning a fractal dimension to those irregular surfaces has now become routine in various fields including topography, defect and fracture studies, growth phenomena, erosion and corrosion processes, catalysis, and many other areas in physics, chemistry, biology, geology, meteorology, and material sciences [1–18]. For isotropic and self-similar interfaces when magnified equally in all directions, algorithms (e.g., box-counting algorithms, fixed-size and fixed-mass
MULTIFRACTAL IMAGE ANALYSIS
3
correlation algorithms) were designed and shown to provide a good estimate of the fractal dimension DF [19–27]. For rough surfaces that are well described by self-affine fractals displaying anisotropic scale invariance [1,2,4,5,7,28–31], various methods (e.g., divider, box, triangle, slit-island, power spectral, variogram, and distribution methods) of computing DF were shown to give different results [32–36]. Limited resolution as well as finite-size effects are well known for introducing biases in the estimate of DF , which are indeed method dependent [32,36,37]. For a documented discussion of the possible reasons for these differences in fractal dimension measurements, we refer the reader to the review article of Lea-Cox and Wang [38]. An alternative strategy consists in computing the so-called roughness exponent H [1,2,4,7] that describes the scaling of the width (or thickness) of the rough interface with respect to measurement scale. Different methods (e.g., height–height correlation function, variance and power spectral methods, detrented fluctuation analysis, first return and multireturn probability distributions) [33–36,39–42] are available to estimate this exponent that is supposed to be related to the fractal dimension DF ¼ d H of self-affine surfaces embedded in a d-dimensional space. Again a number of artifacts may pollute the estimate of the roughness exponent [36]. Since sensitivity and accuracy are method dependent, the usually recommendation is to simultaneously use different tools in order to appreciate, in a quantitative way, the level of confidence in the measured exponent. But beyond some practical algorithmic limitations, there exists a more fundamental intrinsic insufficiency of fractal dimension measurement in the sense that the fractal dimension DF as well as the roughness exponent H are global quantities that do not account for possible fluctuations (from point to point) of the local regularity properties of a fractal surface. Box-counting and correlation algorithms were successfully adapted to resolve multifractal scaling for isotropic self-similar fractals by computation of the generalized fractal dimensions Dq [20–26]. As to self-affine fractals, Parisi and Frisch [43] proposed, in the context of the analysis of fully developed turbulence data, an alternative multifractal description based on the investigation of the scaling behavior of the so-called structure functions [18,44]: Sp ðlÞ ¼ < ðfl Þp > l p ( p integer > 0), where fl ðxÞ ¼ f ðx þ lÞ f ðxÞ is an increment of the recorded signal over a distance l. Then, after reinterpreting the roughness exponent as a local quantity [43,45–49]: fl ðxÞ l hðxÞ , the D(h) singularity spectrum is defined as the Hausdorff dimension of the set of points x where the local roughness (or Ho¨lder) exponent h(x) of f is h. In principle, D(h) can be attained by Legendre transforming the structure function scaling exponents p [43,48,49]. Unfortunately, as noticed by Muzy et al. [50], there are some fundamental drawbacks to the structure
4
´ ODO ET AL. ARNE
function method. Indeed, it generally fails to fully characterize the D(h) singularity spectrum since only the strongest singularities of the function f itself (and not the singularities present in the derivatives of f ) are a priori amenable to this analysis. Even though one can extend this study from integer to real positive p values by considering the increment absolute value, the structure functions generally do not exist for p < 0. Moreover, singularities corresponding to h > 1, as well as regular behavior, bias the estimate of p [48–50]. In previous work [47–50], one of the authors (A.A.), in collaboration with Bacry and Muzy, has shown that there exists a natural way of performing a multifractal analysis of self (multi)affine functions, which consists in using the continuous wavelet transform [51–66]. By using wavelets instead of boxes, as in classic multifractal formalism [24,67–71], one can take advantage of freedom in the choice of these ‘‘Generalized oscillating boxes’’ to get rid of possible smooth behavior that might either mask singularities or perturb the estimation of their strength h [47–50]. The other fundamental advantage of using wavelets is that the skeleton defined by the wavelet transform modulus maxima (WTMM) [72,73] provides an adaptative space-scale partitioning from which one can extract the D(h) singularity spectrum via the scaling exponents ðqÞ of some partition functions defined on the skeleton. The so-called WTMM method [47–50] therefore provides access to the entire D(h) spectrum via the usual Legendre transform DðhÞ ¼ minq ½qh ðqÞ. We refer the reader to Refs. [74,75] for rigorous mathematical results. Since the WTMM method is mainly devoted to practical applications to stochastic systems, let us point out that the theoretical treatment of random multifractal functions requires special attention. A priori, there is no reason that all the realizations of the same stochastic multifractal process correspond to a unique D(h) curve. Each realization has its own unique distribution of singularities and one crucial issue is to relate these distributions to some averaged versions computed experimentally. As emphasized by Hentschel [76], one can take advantage of the analogy that links the multifractal description to statistical thermodynamics [24,49,67,68,77], by using methods created specifically to study disorder in spin-glass theory [78]. When carrying out replica averages of the random partition function associated with a stochastic function, one gets multifractal spectra ðq; nÞ that generally depend on the number of members n in the replica average (let us note that n ¼ 0 and n ¼ 1, respectively, correspond to commonly used quenched and annealed averaging [76]). Then, by Legendre transforming ðq; nÞ, some type of average D(h) spectra is found [76]. Some care is thus required when interpreting these average spectra in order to avoid some misunderstanding of the underlying physics.
MULTIFRACTAL IMAGE ANALYSIS
5
Applications of the WTMM method to one-dimensional (ID) signals have already provided insight into a wide variety of outstanding problems [62], e.g., the validation of the cascade phenomenology of fully developed turbulence [47–49,79–87], the discovery of a Fibonacci structural ordering in 1D cuts of diffusion-limited aggregates (DLA) [88–92], the characterization and the understanding of long-range correlations in DNA sequences [93– 98], and the demonstration of the existence of a causal cascade of information from large to small scales in financial time series [99,100]. Let us also note that from a fundamental point of view, the WTMM multifractal formalism [47–50,74] has been recently revisited [101–104] in order to incorporate in this statistical ‘‘canonical’’ description (which applies for cusp-like singularities only), the possible existence of oscillating singularities [73,101,105]. This new ‘‘grand canonical’’ description [102–104] allows us to compute the singularity spectrum Dðh; Þ, which accounts for the statistical contribution of singularities of Ho¨lder exponent h and oscillation exponent (where characterizes the local power-law divergence of the instantaneous frequency). In a recent work [106–110], we have generalized the canonical WTMM method from 1D to two-dimensional (2D), with the specific goal of achieving multifractal analysis of rough surfaces with fractal dimension DF anywhere between 2 and 3. During the past few years, increasing interest has been paid to the application of the wavelet transform (WT) to image processing [26,61,62,65,111–113]. In this context, Mallat and collaborators [72,73] have extended the WTMM representation in 2D in a manner inspired from Canny’s multiscale edge detectors commonly used in computer vision [114]. Our strategy [107,108] consists of using this representation to define a three-dimensional (3D) WT skeleton from which one can compute partition functions and ultimately extract multifractal spectra. This article is mainly devoted to a detailed description of the 2D WTMM methodology with some test applications to random monofractal and multifractal self-affine surfaces displaying isotropic as well as anisotropic (with respect to space variables) scale similarity properties. As an illustration of the efficiency and reliability of this method, we will report the main results of its application to experimental 2D data in various domains, namely geophysics, hydrodynamics, and medecine. The article is organized as follows. In Section II, we describe the 2D WTMM representation introduced by Mallat et al. [72,73] as the equivalent of multiscale Canny edge detection. We present the continuous WT as a mathematical microscope that is well suited for characterizing the local regularity of rough surfaces. For practical purposes, the WTMM representation is emphasized as a very efficient and accurate numerical tool for scanning the singularities of fractal landscapes. We then describe the
6
´ ODO ET AL. ARNE
2D WTMM method as a natural generalization of box-counting algorithms and structure function techniques previously used for multifractal analysis of isotropic self-similar interfaces and multiaffine surfaces [107,108]. Section III is devoted to the application of the 2D WTMM method to fractional Brownian surfaces [1,2,4,28] that display isotropic (with respect to space variables) scaling properties. For this class of isotropic homogeneous random rough surfaces, we address the issues of statistical convergence and finite-size effects [108]. We illustrate the ability of the 2D WTMM method to reveal and to master anisotropic scale invariance hidden in the roughness fluctuations of a random surface. We also report the results of test applications to synthetic random multifractal rough surfaces generated with a random W-cascade process on a separable wavelet orthogonal basis [109]. On a more general ground, we show that the 2D WTMM method can be used for many purposes in image processing including edge detection, pattern recognition, and image denoising. The next sections are devoted to the description of the most significant results obtained when applying the 2D WTMM method to three different experimental situations. In Section IV, we review the outcomes of the statistical analysis of high-resolution LANDSAT satellite images of cloudy scenes. This study brings into light the underlying multiplicative structure of marine stratocumulus clouds [107,110]. The multifractal properties of the stratocumulus radiance fields are further compared to previous experimental estimates performed on velocity and temperature fluctuations in high Reynolds number turbulence. In Section V, we report the preliminary results of the application of the 2D WTMM method to 2D cuts of the dissipation and enstrophy fields computed from direct high-resolution numerical simulations of statistically stationary 3D homogeneous and isotropic fully developed turbulent flows at a Reynolds number around 1000 ðR ’ 1150Þ. This study reveals that both fields display log-normal multifractal properties but that the enstrophy field turns out to be much more intermittent than the dissipation field. From a comparison with previous experimental investigations of 1D data, we comment about the reliability of the results obtained when using 1D surrogate dissipation data. In Section VI, we apply the 2D WTMM method to perform a multifractal analysis of digitized mammograms [115]. We show that this method can be used to classify fatty and dense areas of breast tissue. We further demonstrate that this method provides a very efficient way to detect tumors as well as microcalcifications, which correspond to much stronger singularities than those involved in the background tissue roughness fluctuations. These preliminary results indicate that the texture discriminatory power of the 2D WTMM method may lead to significant improvement in computer-assisted diagnosis in digitized mammograms. We conclude in Section VII.
7
MULTIFRACTAL IMAGE ANALYSIS
II. Image Processing with the 2D Continuous Wavelet Transform A. Analyzing Wavelets for Multiscale Edge Detection The edges of the different structures that appear in an image are often the most important features for pattern recognition. Hence, in computer vision [116,117], a large class of edge detectors looks for points where the gradient of the image intensity has a modulus that is locally maximum in its direction. As originally noticed by Mallat and collaborators [72,73], with an appropriate choice of the analyzing wavelet, one can reformalize the Canny’s multiscale edge detector [114] in terms of a 2D wavelet transform. The general idea is to start by smoothing the discrete image data by convolving it with a filter and then to compute the gradient on the smoothed signal. Let us consider two wavelets that are, respectively, the partial derivative with respect to x and y of a 2D -smoothing function ðx; yÞ: 1 ðx; yÞ
¼
@ðx; yÞ @x
and
2 ðx; yÞ
¼
@ðx; yÞ @y
ð1Þ
We will assume that is a well-localized (around x ¼ y ¼ 0) isotropic function that depends on jxj only. In this work, we will mainly use the Gaussian function: ðx; yÞ ¼ eðx
2 þy2 Þ=2
¼ ejxj
2
=2
ð2Þ
as well as the isotropic Mexican hat: ðxÞ ¼ ð2 x2 Þejxj
2
=2
ð3Þ
The corresponding analyzing wavelets 1 and 2 are illustrated in Figure 1. They have one and three vanishing moments when using, respectively, the Gaussian function [Eq. (2)] and the Mexican hat [Eq. (3)] as smoothing function. For any function f ðx; yÞ 2 L2 ðRÞ, the wavelet transform with respect to 1 and 2 has two components and therefore can be expressed in a vectorial form: Z 9 8 2 > ½ f ¼ a d 2 x 1 ½a1 ðx bÞ f ðxÞ > T = < 1 Z ð4Þ T ½ f ðb; aÞ ¼ > ; : T 2 ½ f ¼ a2 d 2 x 2 ½a1 ðx bÞ f ðxÞ >
´ ODO ET AL. ARNE
8
Figure 1. The analyzing wavelets 1 and 2 defined in Eq. (1). First-order analyzing wavelets obtained from a Gaussian smoothing function [Eq. (2)]: (a) 1 ; (b) 2 . Third-order analyzing wavelets obtained from the isotropic Mexican hat smoothing function [Eq. (3)]: (c) 1 ; (d) 2 .
Then, after a straightforward integration by parts, one gets: Z 1 2 2 T ½ f ðb; aÞ ¼ a r d x a ðx bÞ f ðxÞ ¼ rfT ½ f ðb; aÞg ¼ rfb;a f g
ð5Þ
If ðxÞ is simply a smoothing filter like the Gaussian function [Eq. (2)], then Eq. (5) amounts to define the 2D wavelet transform as the gradient vector of f (x) smoothed by dilated versions ða1 xÞ of this filter. If ðxÞ has some vanishing moments, then T ½ f ðb; aÞ in Eq. (5) is nothing but the continuous 2D wavelet transform of f (x) as originally defined by Murenzi [118,119], provided ðxÞ be an isotropic analyzing wavelet so that the integration over the angle becomes trivial. As far as notations are concerned, we will mainly use the representation involving the modulus and the argument of the wavelet transform: T ½ f ðb; aÞ ¼ M ½ f ðb; aÞ; A ½ f ðb; aÞ ð6Þ with
M ½ f ðb; aÞ ¼
n 2 2 o1=2 T 1 ½ f ðb; aÞ þ T 2 ½ f ðb; aÞ
ð7Þ
MULTIFRACTAL IMAGE ANALYSIS
9
and A ½ f ðb; aÞ ¼ Arg T 1 ½ f ðb; aÞ þ iT 2 ½ f ðb; aÞ
ð8Þ
B. Characterizing the Local Regularity Properties of Rough Surfaces with the Wavelet Transform Modulus Maxima In the present work, we will use the term rough surface for an irregular surface on which there are no overhanging regions. This means that the surface can be correctly described by a single-valued self-affine function satisfying 8x0 ¼ ðx0 ; y0 Þ 2 R2 ; 8x ¼ ðx; yÞ 2 R2 in the neighborhood of x0 ; 9H 2 R such that, for any > 0, one has [1,2,4,5,7,28–30]: f ðx0 þ x; y0 þ yÞ f ðx0 ; y0 Þ ’ H ½ f ðx0 þ x; y0 þ yÞ f ðx0 ; y0 Þ
ð9Þ
If f is a stochastic process, this identity holds in law for fixed and x0 . According to the value of the exponent , this self-affine function will display either isotropic scale invariance with respect to the space variables ( ¼ 1) or anisotropic scale invariance ( 6¼ 1) [36,120–123]. The Hurst exponent H characterizes the global regularity of the function f. Let us note that if H < 1, then f is nowhere differentiable and that the smaller the exponent H, the more singular f. For H ¼ 1 and ¼ 1, the rough surface defined by f in R3 , is a self-similar fractal in the sense that it is invariant under some isotropic dilations [1,2,36,121–123]. In various contexts [1–18], several methods have been used to estimate the Hurst exponent of self-affine functions. In most studies, isotropic scale invariance was used as a prerequisite for the application of commonly used methods to the analysis of 1D fractal landscapes, e.g., the height–height correlation function, the variance and power spectral methods, the detrented fluctuation analysis, and the first return and multireturn probability distributions [33–36,39–42]. The strategy followed in these studies reduces the analysis of rough surfaces to the investigation of self-affine (1D) profiles obtained through 2D cuts in a three-dimensional representation. As long as the estimate of the Hurst exponent H is independent of the intersection plane, there is no inconsistency in the methodology. When H is found to be sensitive to the orientation of the intersecting plane, this means that the isotropic scale invariance hypothesis does not apply and that one needs to have recourse to methods fully adapted to the characterization of rough surfaces. Unfortunately, to our knowledge, most of the methods listed above have been extended to self-affine functions from R2 to R under the implicit assumption of isotropic scaling.
´ ODO ET AL. ARNE
10
But fractal functions generally display multiaffine properties in the sense that their roughness (or regularity) fluctuates from point to point [43,45–49]. To describe these multifractal functions, one thus needs to change slightly the definition of the Hurst regularity of f so that it becomes a local quantity hðx0 Þ. A rigorous definition of the Ho¨lder exponent (as the strength of a singularity of a function f at the point x0 ) is given by the largest exponent hðx0 Þ such that there exists a polynomial of degree n < hðx0 Þ and a constant C > 0, so that for any point x in the neighborhood of x0 one has [72,73,106–108] j f ðxÞ Pn ðx x0 Þj Cjx x0 jhðx0 Þ
ð10Þ
If f is n times continuously differentiable at the point x0 , then one can use for the polynomial Pn ðx x0 Þ the order-n Taylor series of f at x0 and thus prove that hðx0 Þ > n. Thus hðx0 Þ measures how irregular the function f is at the point x0 . The higher the exponent hðx0 Þ, the more regular the function f. In this work, we will mainly consider fractal functions of two variables that possess only cusp-like singularities. (We refer the reader to Ref. [124], for rigorous mathematical results concerning 2D oscillating singularities or chirps.) But the situation is a little more tricky than in 1D. Indeed one has to distinguish two main cases depending on whether scale invariance is under isotropic or anisotropic dilations [1,2,36,108,121–123,125]. 1. Isotropic Dilations Local scale invariance under isotropic dilations means that locally, around the point x0 , the function f behaves as f ðx0 þ uÞ f ðx0 Þ ’ 1hðx0 Þ ½ f ðx0 þ uÞ f ðx0 Þ
ð11Þ
where > 0 and u is a unit vector. If the scaling exponent hðx0 Þ does not depend upon the direction of u, then f displays isotropic local scale invariance around x0 and the corresponding singularity is of Ho¨lder exponent hðx0 Þ. If, on the contrary, the scaling exponent depends upon the direction of u, then the Ho¨lder exponent is the minimum value of h over all the possible orientations of u. Thus f displays anisotropic scale invariance around x0 with one, several, or a continuum of privileged directions along which the variation of f defines the Ho¨lder exponent of the singularity located at x0 . 2. Anisotropic Dilations Local scale invariance under anisotropic dilations means that locally around the point x0 , the function f behaves as [120–123,125] f ½x0 þ ðÞr u f ðx0 Þ ’ 1hðx0 Þ ½ f ðx0 þ uÞ f ðx0 Þ
ð12Þ
MULTIFRACTAL IMAGE ANALYSIS
11
where > 0 and u is a unit vector. r is a rotation matrix and ðÞ is a positive diagonal 2 2 matrix that accounts for anisotropic self-affine scale transformation in the -rotated referential with origin x0 : 0 ðÞ ¼ ð13Þ 0
The function f thus displays anisotropic scale invariance around x0 and the Ho¨lder exponent is given by the behavior of f in the direction ð < 1Þ or þ =2ð > 1Þ. Very much like the wavelet transform analysis of cusp singularities in 1D [47–49,74], in order to recover the Ho¨lder exponent hðx0 Þ of a function f from R2 to R, one needs to study the behavior of the wavelet transform modulus inside a cone jx x0 j < Ca in the (space-scale) half space [106,108,126]. As originally proposed by Mallat and collaborators [72,73], a very efficient way to perform point-wise regularity analysis is to use the wavelet transform modulus maxima. In the spirit of Canny edge detection [114], at a given scale a, the WTMM are defined as the points b where the wavelet transform modulus M ½ f ðb; aÞ [Eq. (7)] is locally maximum along the gradient direction given by the wavelet transform argument A ½ f ðb; aÞ [Eq. (8)]. These modulus maxima are inflection points of f a ðxÞ. As illustrated in the examples just below, these WTMM lie on connected chains hereafter called maxima chains [106–108]. In theory, one only needs to record the position of the local maxima of M along the maxima chains together with the value of M ½ f and A ½ f at the corresponding locations. At each scale a, our wavelet analysis thus reduces to store those WTMM maxima (WTMMM) only. They indicate locally the direction where the signal has the sharpest variation. This orientation component is the main difference between 1D and 2D wavelet transform analysis. These WTMMM are disposed along connected curves across scales called maxima lines [107,108]. We will define the WT skeleton as the set of maxima lines that converges to the (x, y)-plane in the limit a ! 0þ . This WT skeleton is likely to contain all the information concerning the local Ho¨lder regularity properties of the function f under consideration [108]. Example 1. Isotropic singularity interacting with a localized smooth structure. Let us first illustrate the above definitions on the function f1 shown in Figure 2: 2
2
f1 ðxÞ ¼ Aeðxx1 Þ =2 þ Bjx x0 j0:3
ð14Þ
This function is C 1 everywhere except at x ¼ x0 where f1 is isotropically singular with a Ho¨lder exponent hðx0 Þ ¼ 0:3. Its 2D wavelet transform [Eq. (4)] with a first-order analyzing wavelet [the smoothing function ðxÞ is
12
´ ODO ET AL. ARNE
Figure 2. Three-dimensional representation of the function f1 ðxÞ ¼ Aeðxx1 Þ2=2 2 þBjx x0 j0:3 . The isotropic singularity S is located at x0 ¼ ð256; 256Þ. The Gaussian localized structure G of width ¼ 128 is located at x1 ¼ ð256; 256Þ. The parameter values are A ¼ 1 and B ¼ 1.
Figure 3. Wavelet transform [Eq. (4)] of the function f1 shown in Figure 2, with a first-order analyzing wavelet ( is the isotropic Gaussian function). (a) T 1 ½ f1 ; (b) T 2 ½ f1 coded using 32 gray levels from white (min T ) to black (max T ). (c) M ½ f1 coded from white (M ¼ 0) to black (max M ). (d) jA ½ f1 j coded from white (jA j ¼ 0) to black (jA j ¼ ). The considered scale is a ¼ 23 W where W ¼ 13 (pixels) is the characteristic size of at the smallest resolved scale.
the isotropic Gaussian function] is shown in Figure 3 for a given scale a ¼ 23 w , where w ¼ 13 is the width (in pixel units) of the analyzing wavelet at the smallest scale where it is still well enough resolved. Indeed w is the smallest scale (or the highest resolution) accessible to our wavelet transform microscope. T 1 ½ f1 and T 2 ½ f1 [Eq. (4)] are shown in Figure 3a and b,
MULTIFRACTAL IMAGE ANALYSIS
13
respectively. The corresponding modulus M ½ f1 and argument A ½ f1 are represented in Figure 3c and d. From a simple visual inspection of Figure 3c, one can convince oneself that the modulus is radially symmetric around x0 where is located the singularity S. This is confirmed in Figure 3d where A ½ f1 rotates uniformly from 0 to 2 around x0 . The WTMM as well as the WTMMM are shown in Figure 4 for various values of the scale parameter a ranging from a ¼ 23:5 W (Fig. 4a) to 27:5 (Fig. 4f ). At small scale, there exist mainly two maxima chains. One is a closed curve around x0 at which the
Figure 4. Maxima chains (solid line) defined by the WTMM of the function f1 (Fig. 3). The local maxima (respectively minima) along these chains are indicated by () [respectively ( )] from which originates an arrow whose length is proportional to M ½ f1 and its direction (with respect to the x-axis) is given by the WTMM argument A ½ f1 . The scale parameter is a ¼ 23:5 (a), 24:7 (b), 25:5 (c), 26:3 (d), 26:8 (e), and 27:5 (f ) in w units. Same first-order analyzing wavelet as in Figure 3.
14
´ ODO ET AL. ARNE
Figure 5. Three-dimensional representation of the topological evolution of the WTMM chains of f1 in the space-scale half-hyperplane. The WTMMM () are disposed on connected curves called maxima lines. These maxima lines are obtained by linking each WTMMM computed at a given scale to the nearest WTMMM computed at the scale just above. There exist two maxima lines, Lx0 ðaÞ and Lx1 ðaÞ, pointing, respectively,to the singularity S and to the smooth localized structure G in the limit a ! 0þ .
singularity S is located. The other one is an open curve that partially surrounds G. On each of these maxima chains, one finds only one WTMMM () whose corresponding arguments are such that the gradient vector points to S and G, respectively. As far as the singularity S is concerned, this means that the direction of largest variation of f1 around S is given by x0 ¼ A ½ f1 þ , where A ½ f1 is the argument of the corresponding WTMMM. When increasing the scale parameter, the maxima chains evolve; in particular the closed maxima chain around S swells (its characteristic size behaves like a) until it connects with the maxima chain associated with G (Fig. 4d) to form a single closed curve surrounding both S and G (Fig. 4f ). The topological evolution of the maxima chains in the space-scale half-hyperplane is illustrated in Figure 5. This three-dimensional representation enlightens the existence of two maxima lines obtained by linking the WTMMM step by step (i.e., as continuously as possible) from small to large scales. One of these maxima
MULTIFRACTAL IMAGE ANALYSIS
15
Figure 6. Evolution of M ½ f1 and A ½ f1 when following, from large scale to small scale, the maxima lines Lx0 ðaÞ and Lx1 ðaÞ pointing, respectively, to the singularity S [(a) and (c), respectively] and to the localized smooth structure G [(b) and (d), respectively]. The symbols () and ( ) have the same meaning as in Figure 4. Same first-order analyzing wavelet as in Figure 3.
lines points to the singularity S in the limit a ! 0þ . As shown in Figure 6a, along this maxima line [Lx0 ðaÞ], the wavelet transform modulus behaves as [72,73] M ½ f1 ½Lx0 ðaÞ a hðx0 Þ ;
a ! 0þ
ð15Þ
where hðx0 Þ ¼ 0:3 is the Ho¨lder exponent of S. Moreover, along this maxima line, the wavelet transform argument evolves toward the value (Fig. 6c): A ½ f1 ðLx0 ðaÞÞ ¼ þ x0
ð16Þ
in the limit a ! 0þ , where x0 is nothing but the direction of the largest variation of f1 around x0 , i.e., the direction to follow from x0 to cross the maxima line at a given (small) scale. From the maxima line Lx0 ðaÞ, one thus gets the required amplitude as well as directional informations to characterize the local Ho¨lder regularity of f1 at x0 . Note that along the other maxima line Lx1 ðaÞ that points to x1 where the smooth localized structure G is located, the wavelet transform modulus behaves as (Fig. 6b) M ½ f1 ðLx1 ðaÞÞ an ;
a ! 0þ
where n ¼ 1 is the order of the analyzing wavelet.
ð17Þ
16
´ ODO ET AL. ARNE
Figure 7. WTMM analysis of the function f2 ðxÞ defined in Eq. (18). (a) f2 ðxÞ as coded using 32 gray levels from white (min f2 ) to black (max f2 ). The maxima chains (solid line) and the WTMMM () are shown for the following values of the scale parameter a ¼ 2 (b), 28 (c), and 211 (d) in w units. Same first-order analyzing wavelet as in Figure 3.
Example 2. Anisotropic singularity. Let us illustrate with a specific example, the possibility for a function f2 ðxÞ to display anisotropic local scale invariance with respect to isotropic dilations. In Figure 7a the following function is represented: f2 ðxÞ ¼ f2 ð ; Þ ¼ hð Þ
ð18Þ
hð Þ ¼ 0:3 sin ð 2=3Þ þ 0:5
ð19Þ
with
The exponent hð Þ is nothing but the Ho¨lder exponent at ¼ 0 of the 1D profile obtained when intersecting the image in Figure 7a along the direction . As far as the whole 2D problem is concerned, the Ho¨lder exponent of the singularity S is hðx0 Þ ¼ min hð Þ ¼ 0:2. It quantifies the sharpest variation of f2 ðxÞ, which occurs in the direction x0 ¼ =6. As shown in Figure 7b–d for different zooms, there exists at each scale only one WTMMM, which belongs to a unique maxima line Lx0 ðaÞ pointing to the singularity S. Note that this WTMMM is located in the direction x0 ¼ =6 from the origin. When following Lx0 ðaÞ from large to small scales, M ½ f2 ½Lx0 ðaÞ behaves as a power law with an exponent hðx0 Þ ¼ 0:2 (Fig. 8a), in remarkable agreement with the theoretical prediction for the Ho¨lder exponent of S. Moreover, when investigating A ½ f2 ½Lx0 ðaÞ, one further gets directional
MULTIFRACTAL IMAGE ANALYSIS
17
Figure 8. Evolution of (a) M ½ f2 and (b) A ½ f2 when following, from large to small scales, the maxima line Lx0 ðaÞ ðÞ, which points to the singularity S. Same first-order analyzing wavelet as in Figure 7.
information: A ½ f2 ¼ 5=6 ¼ x0 , from which one learns about the possible existence of some preferential direction as far as the Ho¨lder regularity properties are concerned. We will not treat here the case of local scale invariance with respect to anisotropic self-affine dilations. We send the reader to Arneodo et al. [108] where the 2D WTMM method has been be applied to random self-affine rough surfaces. C. The 2D Wavelet Transform Modulus Maxima (WTMM) Method Before describing the methodology to be used to perform a multifractal analysis of rough surfaces, we need to define the notion of singularity spectrum of a fractal function from R2 into R [108]. 1. Definition Let f be a function from R2 into R and Sh the set of all the points x0 so that the Ho¨lder exponent [Eq. (10)] of f at x0 is h. The singularity spectrum D(h) of f is the function that associates with any h, the Hausdorff dimension of Sh : DðhÞ ¼ dH x 2 R2 ; hðxÞ ¼ h ð20Þ
In the previous section, we have seen that the maxima lines defined from the WTMMM computed at different scales can be used as a scanner of singularities. They allow us to detect the positions where the singularities are located as well as to estimate their strength h. A rather naive way to compute the D(h) singularity spectrum would thus consist in identifying the
18
´ ODO ET AL. ARNE
set of maxima lines along which the wavelet transform modulus behaves with a power-law exponent h [Eq. (15)] and then to use classical boxcounting techniques [19–27] to compute the fractal dimension D(h) of the set of points fxn g R2 to which these maxima lines converge. Unfortunately, when investigating deterministic as well as random fractal functions, the situation is somewhat more intricate than when dealing with isolated singularities. The characteristic feature of these singular functions is the existence of a hierarchical distribution of singularities [47–50,62–65]. Locally, the Ho¨lder exponent hðx0 Þ is then governed by the singularities that accumulate at x0 . This results in unavoidable oscillations around the expected power-law behavior of the wavelet transform modulus [47–50,79]. The exact determination of h from log–log plots on a finite range of scales is therefore somewhat uncertain [127,128]. Note that there have been many attempts to circumvent these difficulties in 1D [79,129]. But in 2D (rough surfaces) as well as in 1D (multiaffine profiles), there exist fundamental limitations (which are not intrinsic to the wavelet technique) to the local measurement of the Ho¨lder exponents of a fractal function. Therefore, the determination of statistical quantities like the D(h) singularity spectrum requires a method that is more feasible and more appropriate than a systematic investigation of the wavelet transform local scaling behavior as experienced [127,128].
2. Methodology Our strategy will consist in mapping the methodology developed [47–50] for multifractal analysis of irregular 1D landscapes to the statistical characterization of roughness fluctuations of 2D surfaces [107,108]. The 2D WTMM method relies upon the space-scale partitioning given by the wavelet transform skeleton. As discussed in Section II.B, this skeleton (see Fig. 12) is defined by the set of maxima lines that points to the singularities of the considered function and therefore is likely to contain all the information concerning the fluctuations of point-wise Ho¨lder regularity. Let us define LðaÞ as the set of all maxima lines that exist at the scale a and that contain maxima at any scale a0 a. The important feature is that each time the analyzed image has a Ho¨lder exponent hðx0 Þ < n , there is at least one maxima line pointing toward x0 along which Eq. (15) is expected to hold. In the case of fractal functions, we thus expect that the number of maxima lines will diverge in the limit a ! 0þ , as the signature of the hierarchical organization of the singularities. The WTMM method consists in defining the following partitions functions directly from the WTMMM that belong to the wavelet transform skeleton:
19
MULTIFRACTAL IMAGE ANALYSIS
Zðq; aÞ ¼
X
L 2 LðaÞ
"
sup ðx; a0 Þ 2 L; a0 a
M ½ f ðx; a0 Þ
#q
ð21Þ
where q 2 R. As compared to classic box-counting techniques [19–27], the analyzing wavelet plays the role of a generalized ‘‘oscillating box’’, the scale a defines its size, while the WTMM skeleton indicates how to position our oscillating boxes to obtain a partition (of S ¼ [Sh ) at the considered scale. Without the ‘‘sup’’ in Eq. (21), one would have implicitely considered a uniform covering with wavelets of the same size a. As emphasized [47– 50,74], the ‘‘sup’’ can be regarded as a way of defining a ‘‘Hausdorff-like’’ scale-adaptative partition that will prevent divergencies to show up in the calculation of Zðq; aÞ for q < 0. Now, from the analogy that links the multifractal formalism to thermodynamics [48,49,67–69,76,77], one can define the exponent ðqÞ from the power-law behavior of the partition function: Zðq; aÞ aðqÞ ;
a ! 0þ
ð22Þ
where q and ðqÞ play, respectively, the role of the inverse temperature and the free energy. The main result of the wavelet-based multifractal formalism is that in place of the energy and the entropy (i.e., the variables conjugated to q and ), one has the Ho¨lder exponent h [Eq. (10)] and the singularity spectrum D(h) [Eq. (20)]. This means that the D(h) singularity spectrum of f can be determined from the Legendre transform of the partition function scaling exponent ðqÞ: DðhÞ ¼ min ½qh ðqÞ q
ð23Þ
From the properties of the Legendre transform, it is easy to convince oneself that homogeneous (monofractal) fractal functions that involve singularities of unique Ho¨lder exponent h ¼ @ =@q are characterized by a ðqÞ spectrum that is a linear function of q. On the contrary, a nonlinear ðqÞ curve is the signature of nonhomogeneous functions that display multifractal properties, in the sense that the Ho¨lder exponent h(x) is a fluctuating quantity that depends upon the spatial position x (in other words the local roughness exponent is fluctuating from point to point). 3. Remark The exponents ðqÞ are much more than simply some intermediate quantities of a rather easy experimental access. For some specific values of q, they have well known meaning [48].
´ ODO ET AL. ARNE
20
q ¼ 0: From Eqs. (21) and (22), one deduces that the exponent ð0Þ accounts for the divergence of the number of maxima lines in the limit a ! 0þ . This number basically corresponds to the number of wavelets of size a required to cover the set S of singularities of f. In full analogy with standard box-counting arguments [19–27], ð0Þ can be identified to the fractal dimension (capacity) of this set: ð0Þ ¼ dF fx; hðxÞ < þ1g
ð24Þ
q ¼ 1: As pointed out [48], the value of the exponent ð1Þ is related to the fractal dimension (capacity) of the rough surface S defined by the function f. More precisely [130]: dF ðSÞ ¼ max½2; 1 ð1Þ
ð25Þ
q ¼ 2: It is easy to show that the exponent ð2Þ is intimately related to the scaling exponent of the spectral density: Z 1 d j ^ f ðk; Þj2 k SðkÞ ¼ ð26Þ 2 where ¼ 4 þ ð2Þ
ð27Þ
From a practical point of view, the computation of the D(h) singularity spectrum, via the Legendre transform defined in Eq. (23), first requires a smoothing of the ðqÞ curve. This procedure has a main disadvantage. This smoothing operation prevents the observation of any nonanalycity in the curves ðqÞ and D(h) and the interesting physics of phase transitions [49,71,131,132] in the scaling properties of fractal functions can be completely missed. As suggested [49,131,133–137], one can avoid directly performing the Legendre transform by considering the quantities h and D(h) as mean quantities defined in a canonical ensemble, i.e., with respect to their Boltzmann weights computed from the WTMMM [49,79]: W ½ f ðq; L; aÞ ¼
jsupðx; a0 Þ 2 L; a0 a M ½ f ðx; a0 Þjq Zðq; aÞ
ð28Þ
where Zðq; aÞ is the partition function defined in Eq. (21). Then one computes the expectation values: hðq; aÞ ¼
X
L 2 LðaÞ
ln
sup ðx; a0 Þ 2 L; a0 a
M ½ f ðx; a Þ
W ½ f ðq; L; aÞ 0
ð29Þ
MULTIFRACTAL IMAGE ANALYSIS
21
and Dðq; aÞ ¼
X
L 2 LðaÞ
W ½ f ðq; L; aÞ ln ½W ½ f ðq; L; aÞ
ð30Þ
from which one extracts hðqÞ ¼ lim hðq; aÞ=lna
ð31Þ
DðqÞ ¼ lim Dðq; aÞ=lna
ð32Þ
a!0þ
a!0þ
and therefore the D(h) singularity spectrum. 4. Numerical Implementation In this section, we briefly review the main steps of the numerical implementation of the 2D WTMM method. Let us consider an n n digitized image of a rough surface. Step 1: Computation of the 2D wavelet transform. We compute the two components T 1 and T 2 of the wavelet transform [Eq. (4)] in the Fourier domain, using 2D Fast Fourier Transform (FFT) [138] and inverse FFT. We start our analysis by choosing the analyzing wavelet among the class of radially isotropic wavelets defined in Section II.A (Fig. 1). To master edge effects we focus only on the n=2 n=2 central part of the image where our wavelet coefficients can be shown to be not affected by the boundary of the original image. This means that we will be careful not to increase the scale parameter a above a critical value amax so that the n=2 n=2 central wavelet coefficients remain safe of finite-size effects. In the opposite limit, we will define a lower bound amin to the accessible range of scales so that the analyzing wavelet is still well resolved at that scale. (We refer the reader to Section 1.3.3 of Decoster’s Ph.D. thesis [139] for a detailed practical definition of the accessible [amin ; amax ] range of scales.) Under those precautions, one can be confident of our wavelet transform microscope as far as the investigation of the scale invariance properties in the range a 2 ½amin ; amax is concerned. Step 2: Computation of the wavelet transform skeleton. As explained in Section II.B, at a given scale a, we identify the wavelet transform modulus maxima as the points where M ½ f ðb; aÞ [Eq. (7)] is locally maximum along the gradient direction given by A ½ f ðb; aÞ [Eq. (8)]. Then we chain the points that are nearest neighbors (which actually have compatible arguments). Along each of these maxima chains, we locate the local maxima previously called WTMMM. Note that the two ends of an open maxima chain are not allowed positions for the WTMMM. Once
22
´ ODO ET AL. ARNE
computed the set of WTMMM for a finite number of scales ranging from amin to amax , one proceeds to the connection of these WTMMM from scale to scale. One starts at the smallest scale amin and we link each WTMMM to their nearest neighbor found at the next scale just above. We proceed iteratively from scale to scale up to amax . All the WTMMM that then remain isolated are suppressed. All the WTMMM that are connected on a curve across scales that does not originate from the smallest scale amin are also suppressed. We then store the modulus M and the argument A of the WTMMM that belong to the so-called maxima lines. Those lines are supposed to converge, in the limit a ! 0þ , to the points where the singularities of the image under study are located. As explained in Section II.B, to define the wavelet transform skeleton, one has to select these maxima lines that satisfy Eq. (15) from those that satisfy Eq. (17) and that are wavelet dependent. This is done by increasing the order of the analyzing wavelet; for n large enough, the spurious maxima lines are suppressed by a simple thresholding on M at the smallest scale amin . Their roots are definitely rejected as misleading singularity locations. Step 3: Computation of the multifractal spectrum. According to Eq. (21), one uses the wavelet transform skeleton to compute the partition function Zðq; aÞ on the discrete set of considered scales amin a amax . Then, for a given value of q 2 ½qmin ; qmax , one extracts the exponent ðqÞ [Eq. (22)] from a linear regression fit of lnZðq; aÞ vs. lna. As a test of the robustness of our measurement, we examine the stability of our estimate of ðqÞ with respect to the range of scales ½amin ; amax ½amin ; amax over which the linear regression fit is performed. After estimating the exponent ðqÞ for a discrete set of q-values, we smooth the ðqÞ curve using standard procedure. Then, one determines the D(h) singularity spectrum by Legendre transforming the ðqÞ curve according to Eq. (23). As a check of the reliability of our results, we use the alternative strategy defined in Eq. (28) to (32) to estimate the D(h) singularity spectrum without performing explicitly the Legendre transform. When dealing specifically with stochastic process, we generally have several images at our disposal somehow corresponding to different realizations of this process. In this case, we will mainly proceed to two different averagings corresponding to the following: Quenched averaging: We extract the ðqÞ curve from averaging < lnZðq; aÞ > over the number of images: e < lnZðq;aÞ> aðqÞ ;
a ! 0þ
ð33Þ
In other words, the ðqÞ spectrum is obtained by averaging over the ðqÞ curves extracted from each individual image.
MULTIFRACTAL IMAGE ANALYSIS
23
Annealed averaging: One can alternatively compute the ðqÞ spectrum after averaging the partition functions obtained for each image: < Zðq; aÞ > aðqÞ ;
a ! 0þ
ð34Þ
Note that in most of the examples discussed in this work, we have not observed any significant discrepency between the ðqÞ spectra obtained using either one of these averagings. Consequently, in the following we will mainly show the results obtained when estimating the ðqÞ and D(h) multifractal spectra using annealed averaging. Step 4: Computation of the WTMMM probability density functions. From the computation of the joint probability density function Pa ðM; AÞ, we first proceed to a test of the possible independence of M and A. If it is so, we then investigate separately the scale dependence of Pa ðMÞ and Pa ðAÞ. From the investigation of the shape of Pa ðAÞ, and of its possible evolution when varying a, one can then quantify some possible departure from isotropic scaling as well as the existence of possible privileged directions. When Pa ðM; AÞ does not factorize, then M and A are intimately related. In this case, one can try to compute the A ðqÞ and DA ðhÞ multifractal spectra by conditioning the statistics of the modulus fluctuations to a given value of the argument. The A-dependence of these spectra quantifies what one could call anisotropic multifractal scaling properties.
III. Test Applications of the WTMM Method to Monofractal and Multifractal Rough Surfaces A. Fractional Brownian Surfaces Since its introduction by Mandelbrot and Van Ness [140], the fractional Brownian motion (f Bm) has become a very popular model in signal and image processing [1–18,28–30]. In one dimension, f Bm has proved useful for modeling various physical phenomena with long-range dependence, e.g., ‘‘1/f ’’ noises. The f Bm exhibits a power spectral density Sð!Þ 1=! , where the spectral exponent ¼ 2H þ 1 is related to the Hurst exponent H. 1D f Bm has been extensively used as test stochastic signals for Hurst exponent measurements. The performances of classic methods [33–36,39–42,141–143] (e.g., height–height correlation function, variance and power spectral methods, first return and multireturn probability distributions, maximum likelihood techniques) have been recently competed by wavelet-based
24
´ ODO ET AL. ARNE
techniques [144–157]. Comparative analysis of different wavelet-based estimators for the self-similarity parameter H of f Bm can be found [152–154]. FBm’s are homogeneous random self-affine functions that have been specifically used to calibrate the 1D WTMM methodology [47–49,79]. This method was shown to be a very efficient tool to diagnose the monofractal scaling properties of f Bm. Moreover, it provides very accurate new estimators of the Hurst exponent with remarkable performances [158]. The purpose of this section is to carry out a test application of the 2D WTMM methodology described in Section II, on several realizations of 2D f Bm [108]. The generalization of Brownian motion to more than one dimension was first considered by Levy [159]. The generalization of f Bm follows along similar lines. A 2D f Bm BH ðxÞ indexed by H 2 ½0; 1 is a process with stationary zero-mean Gaussian increments and whose correlation function is given by [1,2,28,159,160] < BH ðxÞBH ðyÞ > ¼
2 ðjxj2H þ jyj2H jx yj2H Þ 2
ð35Þ
where < > represents the ensemble mean value. The variance of such a process is varðBH ðxÞÞ ¼ 2 jxj2H
ð36Þ
from which one recovers the classic behavior var½B1=2 ðxÞ ¼ 2 jxj for uncorrelated Brownian motion with H ¼ 1=2. 2D fBms are self-affine processes that are statistically invariant under isotropic dilations [Eq. (11)]: BH ðx0 þ uÞ BH ðx0 Þ ’ H ½BH ðx0 þ uÞ BH ðx0 Þ
ð37Þ
where u is a unitary vector and ’ stands for the equality in law. The index H corresponds to the Hurst exponent; the higher the exponent H, the more regular the f Bm surface. But since Eq. (37) holds for any x0 and any direction u, this means that almost all realizations of the f Bm process are continuous, everywhere nondifferentiable, isotropically scale-invariant as characterized by a unique Ho¨lder exponent hðxÞ ¼ H; 8x [1,2,28,158]. Thus f Bm surfaces are the representation of homogeneous stochastic fractal functions characterized by a singularity spectrum that reduces to a single point DðhÞ ¼ 2 ¼ 1
if h ¼ H if h ¼ 6 H
ð38Þ
By Legendre transforming D(h) according to Eq. (23), one gets the following expression for the partition function exponent [Eq. (22)]:
MULTIFRACTAL IMAGE ANALYSIS
25
Figure 9. FBm surfaces (128 128) generated with the Fourier transform filtering synthesis method. (a) H ¼ 1=3; (b) H ¼ 1=2; (c) H ¼ 2=3. In the top panels, BH ðxÞ is coded using 32 gray levels from white (min BH ) to black (max BH ).
ðqÞ ¼ qH 2
ð39Þ
ðqÞ is a linear function of q, the signature of monofractal scaling, with a slope given by the index H of the f Bm. We have tested the 2D WTMM method described in Section II on f Bm surfaces generated by the so-called Fourier transform filtering method [28,29,160]. We have used this particular synthesis method because of its implementation simplicity. Indeed it amounts to a fractional integration of a 2D ‘‘white noise’’ and therefore it is expected to reproduce quite faithfully the expected isotropic scaling invariance properties [Eqs. (37)–(39)]. From a visual inspection of Figure 9a (H ¼ 1=3), 9b (H ¼ 1=2), and 9c (H ¼ 2=3), one can convince oneself that the f Bm surfaces become less and less irregular when increasing the index H. This is nothing but the traduction that the fractal dimension of f Bm surfaces increases from 2 to 3 when H covers [0,1] [Eq. (25)]: dF ðf BmSÞ ¼ 1 ð1Þ ¼ 3 H
ð40Þ
When increasing H, a f Bm surface becomes more and more similar to a smooth Euclidean 2D surface. In Figure 10 are reported the results of a power-spectral analysis of a (1024 1024) image of a f Bm rough surface with Hurst exponent H ¼ 1=3.
26
´ ODO ET AL. ARNE
Figure 10. Power spectrum analysis of a (1024 1024) image of a fBm surface B1=3 ðxÞ. (a) in jB^1=3 ðkÞj as coded using 32 gray levels from white (min lnjB^1=3 j) to black (max lnjB^1=3 j). (b) The spectral density SðjkjÞ vs. jkj in a logarithmic representation. The solid line corresponds to the theoretical power-law prediction with exponent ¼ 2H þ 2 ¼ 8=3 [Eq. (41)].
In Figure 10a, the Fourier transform of B1=3 ðxÞ does not display any significant departure from radial symmetry. Isotropic scaling is actually ^ 1=3 ðkÞ over several of such images. In Figure confirmed when averaging B 10b, the power spectral density is shown to behave as a power law as a function of the wavevector modulus jkj, with an exponent that is in perfect agreement with the theoretical prediction for the spectral exponent [Eq. (27)]: ¼ 4 þ ð2Þ ¼ 2 þ 2H
ð41Þ
Along the lines of the numerical implementation procedure described in Section II.C, we have wavelet transformed 32 (1024 1024) images of BH¼1=3 with an isotropic first-order analyzing wavelet. To master edge effects, we then restrain our analysis to the 512 512 central part of the wavelet transform of each image. In Figure 11 the computation of the maxima chains and the WTMMM for an individual image at three different scales is illustrated. In Figure 11b the convolution of the original image (Fig. 11a) with the isotropic Gaussian smoothing filter [Eq. (5)] is shown. According to the definition of the wavelet transform modulus maxima, the maxima chains correspond to well-defined edge curves of the smoothed image. The local maxima of M along these curves are located at the points where the sharpest intensity variation is observed. The corresponding arrows clearly indicate that locally, the gradient vector points in the direction (as given by A ) of maximum change of the intensity surface. When going from large scale (Fig. 11d) to small scale (Fig. 11c), the characteristic average distance between two nearest neighbor WTMMM decreases like a. This means that the number of WTMMM and, in turn, the number of maxima lines proliferate across scales like a2 . The corresponding wavelet transform skeleton is shown in Figure 12. As confirmed just
MULTIFRACTAL IMAGE ANALYSIS
27
Figure 11. 2D wavelet transform analysis of BH¼1=3 ðxÞ. is a first-order radially symmetric analyzing function (see Fig. 1). (a) Thirty-two gray-scale coding of the central 512 512 portion of the original image. In (b) a ¼ 2 W , (c) a ¼ 20:1 W , and (d) a ¼ 21:9 W are shown the maxima chains; the local maxima of M along these chains are indicated by () from which originate an arrow whose length is proportional to M and its direction (with respect to the x-axis) is given by A . In (b), the smoothed image b;a B1=3 [Eq. (5)] is shown as a grayscale coded background from white (min) to black (max).
below, when extrapolating the arborescent structure of this skeleton to the limit a ! 0þ , one recovers the theoretical result that the support of the singularities of a 2D f Bm has a dimension dF ¼ 2, i.e., BH¼1=3 ðxÞ is nowhere differentiable [1,2,28,29,159]. The local scale invariance properties of a f Bm rough surface are investigated in Figure 13. When looking at the behavior of M along some maxima lines belonging to the wavelet transform skeleton, despite some superimposed fluctuations, one observes a rather convincing power-law decrease with an exponent hðx0 Þ that does not seem to depend upon the spatial location x0 . Moreover, the theoretical value for the Ho¨lder exponent hðx0 Þ ¼ H ¼ 1=3 provides a rather good fit of the slopes obtained at small scale in a logarithmic representation of M vs. a [Eq. (15)]. When looking at the simultaneous evolution of A along the same maxima lines, one observes random fluctuations. Unfortunately, because of the rather limited range of scales accessible to our mathematical microscope, a 2 ½ W ; 24 W , there is no hope of demostrating numerically that A actually performs a random walk over [0, 2].
28
´ ODO ET AL. ARNE
Figure 12. Wavelet transform skeleton of the 2D f Bm image shown in Figure 11a. This skeleton is defined by the set of maxima lines obtained after linking the WTMMM detected at different scales. Same analyzing wavelet as in Figure 11.
Figure 13. Characterizing the local Ho¨lder regularity of BH¼1=3 ðxÞ from the behavior of the WTMMM along the maxima lines. Three maxima lines are investigated. (a) log2 M vs. log2 a; (b) A vs. log2 a. Same analyzing wavelet as in Figure 11. The solid line in (a) corresponds to the theoretical slope h ¼ H ¼ 1=3. a is expressed in W units.
In Figure 14 the results of the computation of the ðqÞ and D(h) spectra using the 2D WTMM method described in Section II are reported. As shown in Figure 14a, the annealed average partition function Zðq; aÞ [over 32 images of B1=3 ðxÞ] displays a remarkable scaling behavior over more than three octaves when plotted versus a in a logarithmic representation
MULTIFRACTAL IMAGE ANALYSIS
29
Figure 14. Determination of the ðqÞ and D(h) spectra of 2D fBm with the 2D WTMM method. (a) log2 Zðq; aÞ vs. log2 a; the solid lines correspond to the theoretical predictions ðqÞ ¼ qH 2 [Eq. (39)] with H ¼ 1=3. (b) h(q, a) vs. log2 a; the solid lines correspond to the theoretical slope H ¼ 1=3. (c) ðqÞ vs. q for H ¼ 1=3ðÞ, 1/2 (&), and 2/3 (~); the solid lines correspond to linear regression fit estimates of H. (d) D(h) vs. h as obtained from the scaling behavior of D(q, a) vs. log2 a [Eq. (30)]; the symbols have the same meaning as in (c). Same analyzing wavelet as in Figure 11. These results correspond to annealed averaging over 32 (1024 1024) fBm images. a is expressed in W units.
[Eqs. (21) and (22)]. Moreover, for a wide range of values of q 2 ½4; 6, the data are in good agreement with the theoretical ðqÞ spectrum [Eq. (39)]. When proceeding to a linear regression fit of the data over the first two octaves, one gets the ðqÞ spectra shown in Figure 14c for three values of the f Bm index H ¼ 1=3, 1/2, and 2/3. Whatever H, the data systematically fall on a straight line, the signature of homogeneous (monofractal) scaling properties. However, the slope of this straight line provides a slight
30
´ ODO ET AL. ARNE
Figure 15. Pdfs of the WTMMM coefficients of B1=3 ðxÞ as computed at different scales a ¼ 1; 2; 4, and 8 (in W units). (a) Pa ðMÞ vs. M. (b) Pa ðAÞ vs. A. is the first-order analyzing wavelet shown in Figure 1. These results correspond to averaging over 32 (1024 1024) fBm images.
underestimate of the corresponding Hurst exponent H. Let us point out that a few percent underestimate has also been reported when performing similar analysis of 1D f Bm [47–49,98]. Theoretical investigation of finite-size effects and statistical convergence has been recently performed to explain this experimental observation [98]. R R In Figure 15 are shown the pdfs Pa ðMÞ ¼ dAPa ðM; AÞ and Pa ðAÞ ¼ dM Pa ðM; AÞ, computed for four different values of the scale parameter with B1=3 ðxÞ. As seen in Figure 15a, Pa ðMÞ is not a Gaussian [in contrast to the pdf of the continuous 2D wavelet coefficients when using Eq. (7)], but decreases fast to zero at zero. This explains that when concentrating on the wavelet transform skeleton, the discrete sum on the r.h.s. of Eq. (21) no longer diverges when considering negative q values. This remark is at the heart of the 2D WTMM method; by allowing us to compute the ðqÞ spectrum for negative as well as positive q values, the 2D WTMM method is a definite step beyond the 2D structure function method that is intrinsically restricted to positive q values. The corresponding pdfs Pa ðAÞ are represented in Figure 15b. Pa ðAÞ clearly does not evolve across scales. Moreover, except some small amplitude fluctuations observed at the largest scale, Pa ðAÞ ¼ 1=2 is a flat distribution as expected for statistically isotropic scale-invariant rough surfaces. The results reported in Figure 16 not only corroborate statistical isotropy but they bring unambiguous evidence for the independence of M and A. For two different scales, the pdf of M, when conditioned by the argument A, is shown to be shape invariant. We refer the reader to Arneodo et al. [108] for a similar detailed discussion of the results of the application of the 2D WTMM method to anisotropic monofractal self-affine rough surfaces.
MULTIFRACTAL IMAGE ANALYSIS
31
Figure 16. Pdf of M as conditioned by A. The different curves correspond to fixing A (mod ) to 0 =8; =4 =8; =2 =8 and 3=4 =8. (a) a ¼ 1; (b) a ¼ 2 (in W units). Same 2D WTMM computations for B1=3 as in Figure 15.
B. Multifractal Rough Surfaces Generated by Random Cascades on Separable Wavelet Orthogonal Basis This section is devoted to the application of the 2D WTMM method to multifractal functions synthetized from W-cascades on separable wavelet orthogonal basis as defined in Decoster et al. [109]. A 2D random Wcascade is built recursively on the two-dimensional square grid of separable wavelet orthogonal basis, involving only scales that range between a given large scale L and the scale 0 (excluded). Thus the corresponding fractal function f (x) will not involve scales greater than L. For that purpose, we will use compactly supported wavelets defined by Daubechies [58,109]. Moreover we will mainly concentrate here on multifractal rough surfaces that display isotropic scaling and that are generated with a 2D log-normal W-cascade. If m and 2 are, respectively, the mean and the variance of ln W, where W is a multiplicative random variable with log-normal probability distribution, then, as shown in Decoster et al. [109], a straightforward computation leads to the following ðqÞ spectrum: ðqÞ ¼ log2 < W q > 2; 8q 2 R ¼
2 2 m q q2 2 ln 2 ln 2
ð42Þ
where < > means ensemble average. The corresponding D(h) singularity spectrum is obtained by Legendre transforming ðqÞ [Eq. (23)]: DðhÞ ¼
ðh þ m=ln 2Þ2 þ2 2 2 =ln 2
ð43Þ
´ ODO ET AL. ARNE
32
According to the convergence criteria established in 1D [161], we will consider only parameter values that satisfy the conditions: m<0
and
pffiffiffiffiffiffiffiffi jmj > 2 ln 2 2
ð44Þ
Moreover, by solving DðhÞ ¼ 0, one gets the extremal values hmin and hmax : m 2 pffiffiffiffiffiffiffiffi ln 2 ln 2 m 2 þ pffiffiffiffiffiffiffiffi ¼ ln 2 ln 2
hmin ¼ hmax
ð45Þ
Figure 17 illustrates the computation of the maxima chains and the WTMMM for an individual image of a multifractal rough surface generated with the log-normal W-cascade model with parameter values: m ¼ 0:38 ln 2 and 2 ¼ 0:03 ln 2. Again Figure 17b illustrates perfectly the fact that the maxima chains correspond to edge curves of the original
Figure 17. 2D wavelet transform analysis of a multifractal rough surface generated with the log-normal W-cascade model with parameter values m ¼ 0:38 ln 2 and 2 ¼ 0:03 ln 2. is the first-order radially symmetric analyzing wavelet shown in Figure 1. (a) Thirty-two grayscale coding of the original (1024 1024) image. In (b) a ¼ 22:9 W , (c) a ¼ 21:9 W , and (d) a ¼ 23:9 W are shown the maxima chains and the WTMMM for the central (512 512) part of the original image [dashed square in (a)]. In (b), the smoothed image b;a f is shown as a grayscale-coded background from white (min) to black (max).
MULTIFRACTAL IMAGE ANALYSIS
33
image after smoothing by a Gaussian filter . From the WTMMM defined on these maxima chains, one constructs the WT skeleton according to the procedure described in Section II.C. From the WT skeleton of 32 (1024 1024) images like the one in Figure 17a, one computes the annealed average of the partition functions Zðq; aÞ. As shown in Figure 18a, when plotted versus the scale parameter a in a logarithmic representation, these annealed average partition functions display a rather impressive scaling
Figure 18. Determination of the ðqÞ and D(h) spectra of multifractal rough surfaces generated with the log-normal ( ) random W-cascade models, using the 2D WTMM method. is the first-order radially symmetric analyzing wavelet shown in Figure 1. (a) log2 Zðq; aÞ vs. log2 a; the solid lines correspond to linear regression fit of the data over the first four octaves. (b) h(q, a) vs. log2 a; the solid lines correspond to linear regression fit estimates of h(q). (c) ðqÞ vs. q as obtained from linear regression fit of the data in (a) over the first four octaves. (d) D(h) vs. h, after Legendre transforming the ðqÞ curve in (c). In (c) and (d), the solid lines represent the theoretical log-normal spectra given by Eqs. (42) and (43), respectively.
34
´ ODO ET AL. ARNE
behavior over a range of scales of about four octaves (i.e., W < a< 16 W , where W ¼ 13 pixels). Let us point out that scaling of quite good quality is found for a rather wide range of values of q: 6< q< 8. When processing to a linear regression fit of the data over the first four octaves, one gets the ðqÞ spectrum ( ) shown in Figure 18c. For the range of q values where scaling is operating, the numerical data are in remarkable agreement with the theoretical nonlinear ðqÞ spectrum given by Eq. (42). Similar quantitative agreement is observed on the D(h) singularity spectrum in Figure 18d. Let us note that consistant parabolic shapes are obtained when using either the Legendre transform of the ðqÞ data [Eq. (23)] or the formula (31) and (32) to compute h(q) and D(q). In Figure 18b are reported the results for the expectation values h(q,a) [Eq. (29)] vs. log2 a; it is clear on this figure that the slope h(q) depends upon q, the hallmark of multifractal scaling. Note that again, the theoretical predictions hðqÞ ¼ @ =@q ¼ 2 q=ln 2 m=ln 2 provide very satisfactory fits of the numerical data. From Eq. (45), the multifractal rough surfaces under study display intermittent fluctuations corresponding to Ho¨lder exponent values ranging from hmin ¼ 0:034 to hmax ¼ 0:726. Unfortunately, to capture the strongest and weakest singularities, one needs to compute the ðqÞ spectrum for very large values of jqj. This requires the processing of many more images of much larger size, which is not within current computer capabilities. Note that with the statistical sample studied here, one has Dðhðq ¼ 0Þ ¼ 0:38Þ ¼ 2:00 0:02, which allows us to conclude that the rough surfaces under consideration are singular everywhere. From the construction rule of these synthetic log-normal rough surfaces [109], the multifractal nature of these random functions is expected to be contained in the way the shape of the WT modulus pdf Pa ðMÞ evolves when varying the scale parameter a, as shown in Figure 19a. Indeed the joint probability distribution Pa ðM; AÞ is expected to factorize as the signature of the implicit decoupling of M and A in the construction process. This decoupling is numerically retrieved in Figure 20 where, for two different scales, the pdf of M, when conditioned by the argument A, is shown to be shape invariant. When varying the scale parameter a, no significant angulardependent evolution is observed in the distribution of the WTMMM. As seen in Figure 19b, Pa ðAÞ does not exhibit any significant change when increasing a, except some loss in regularity at large scales due to the rarefaction of the maxima lines. Let us point out that even though Pa ðAÞ looks globally rather flat, one can notice some small amplitude almost periodic oscillations at the smallest scales that reflects the existence of privileged directions in the wavelet cascading process. These oscillations are maximum for A ¼ 0; =2; and 3=2, as the witness to the square lattice anisotropy underlying the 2D wavelet tree decomposition.
MULTIFRACTAL IMAGE ANALYSIS
35
Figure 19. Pdfs of the WTMMM coefficients of synthetic multifractal rough surfaces generated with the log-normal W-cascade model (m ¼ 0:38 ln 2 and 2 ¼ 0:03 ln 2). (a) Pa ðMÞ vs. M. (b) Pa (A) vs. A. is a first-order radially symmetric analyzing wavelet. Four different scales a ¼ 1; 2; 4; 8 (in W units) are shown. These results correspond to averaging over 32 (1024 1024) images.
Figure 20. Pdfs of M when conditioned by A. The different curves correspond to fixing A (mod ) to 0 =8; =4 =8; =2 =8, and 3=4 =8. (a) a ¼ 20:1 ; (b) a ¼ 21:1 (in W units). Same 2D WTMM computations as in Figure 19.
1. Remark We have reported results obtained with the first-order radially symmetric analyzing wavelets shown in Figure 1. Possibly because of the range of Ho¨lder exponent values that is restricted to h 2 ½0; 1, but more probably because of the underlying multiplicative structure of the multifractal surface itself, a first-order analyzing wavelet leads to numerical multifractal spectra that are in remarkable agreement with the theoretical predictions. Let us point out that quite robust results are obtained with the third-order analyzing wavelet used in the previous subsection.
´ ODO ET AL. ARNE
36
C. Distinguishing ‘‘Multiplicative from Additive’’ Processes Underlying the Scale Invariance Properties of Rough Surfaces from Space-Scale Correlation Analysis Correlations in multifractals have already been experienced in the literature [162–164]. However, all these studies rely upon the computation of the scaling behavior of some partition functions involving different points; they thus mainly concentrate on spatial correlations of the local singularity exponents. The approach recently developed [85,100,165] is different since it does not focus on (or suppose) any scaling property but rather consists in studying the correlations of the logarithms of the amplitute of a space-scale decomposition of the signal. More specifically, if ðxÞ is a bump function such that kk1 ¼ 1, then by taking Z ð46Þ "2 ðx; aÞ ¼ a4 ½ðx yÞ=ajT ½ f ðy; aÞj2 d 2 y one has k f k22 ¼
ZZ
"2 ðx; aÞd 2 xda
ð47Þ
Thus, "2 ðx; aÞ can be interpreted as the local space-scale energy density of the considered multifractal function f(x). Since "2 ðx; aÞ is a positive quantity, we can define the magnitude of the function f at the point x and scale a as !ðx; aÞ ¼
1 ln "2 ðx; aÞ 2
ð48Þ
We have shown [109] that a multiplicative process can be revealed and characterized through the correlations of its space-scale magnitudes: ~ðx1 ; a1 Þ! ~ðx2 ; a2 Þ > Cðx1 ; x2 ; a1 ; a2 Þ ¼ < !
ð49Þ
~ for the centered process where < > stands for ensemble average and ! ! < ! >. When using W-cascade process, one can compute analytically the ‘‘two-scale’’ correlation function Cðx; a1 ; a2 Þ, between the magnitude at scale a1 and the magnitude at scale a2 . The function displays a logarithmic behavior as long as x is greater than the supremum of a1 and a2 , namely [109,161,165]:
L x 2 Cðx; a1 ; a2 Þ ¼ log2 2þ2 ð50Þ x L when supða1 ; a2 Þ x < L
MULTIFRACTAL IMAGE ANALYSIS
37
Figure 21. Magnitude correlation function Cðx; a1 ; a2 Þ vs. log2 ðxÞ, as computed from the continuous wavelet transform of 32 (1024 1024) images. The analyzing wavelet is the radially symmetric first-order wavelet shown in Figure 1. (a) Log-normal W-cascades for parameter values m ¼ 0:38 ln 2 and 2 ¼ 0:03 ln 2. (b) Fractional Brownian surfaces BH ¼ 1=3 (x). The symbols have the following meaning: a1 ¼ a2 ¼ 2ð Þ; a1 ¼ 1; a2 ¼ 2ð4Þ; a1 ¼ 1; a2 ¼ 22 ðhÞ and a1 ¼ 2; a2 ¼ 22 ( ) in W units. In (a) the solid line represents the theoretical prediction given by Eq. (50). We have not shown any data points for x W ð 13pixelsÞ.
Thus, the ultrametric structure of the wavelet representation of multifractal rough surfaces generated with the random W-cascade model implies that the cross-correlation functions (across scales) decrease very slowly, independently of a1 and a2 , as a logarithm function of the spatial distance x. In Figure 21a the results of the computation of Cðx; a1 ; a2 Þ when averaging over 32ð1024 1024Þ images of multifractal rough surfaces generated with the log-normal W-cascade model for the same parameter values as in Figure 17a are shown. One can see that for x > supða1 ; a2 Þ, all the data points fall onto a unique curve when plotted versus log2 ðxÞ, independently of the considered pair of scales (a1 ; a2 ). Moreover, although the analyzing wavelet is different from the one used in the construction process of the W-cascade, these numerical data are in striking good agreement with the theoretical prediction given by Eq. (50) for 2 ¼ 0:03 ln 2 and L ¼ 1024. The observed slow (logarithmic) decay of the space-scale correlation functions is thus a clear indication that magnitudes in random cascades are correlated over very long distances [100,109,161,165–167]. Note that both the scale independence and the logarithmic decay are features that are not observed in ‘‘additive’’ models like fractional Brownian motions whose long-range correlations originate from the sign of their variations rather than from the amplitudes. In Figure 21b are plotted the correlation functions Cðx; a1 ; a2 Þ computed from 32ð1024 1024Þ images of isotropic fractional Brownian surfaces with index H ¼ 1=3 (see Fig. 9a). When compared with Figure 21a, the difference is
38
´ ODO ET AL. ARNE
impressive: for x > supða1 ; a2 Þ, the magnitudes of BH ¼ 1=3 (x) are found uncorrelated.
D. Using the 2D WTMM Method to Perform Image Processing Tasks We now want to discuss the ability to use the WTMM method for specific purposes in image processing. We refer the reader to previous work [108] for edge detection and image denoising applications, and also to the work of Levy-Vehel [168,169] for previous attempts to use multifractal concepts for image analysis. In this subsection, we want to address a specific image processing segmentation problem that will be helpful when dealing with medical applications in Section VI. Indeed, in the past 20 years, many signal and image processing works have been devoted to medical research, especially mammography [170,171]. A major point is the detection and the characterization of clusters of microcalcifications, which are early signs of breast cancer (Section VI). Our goal is to demonstrate the ability of the 2D WTMM method to do, in a very attractive way, such a task with both synthetic surfaces and genuine mammographic scenes. Here by cluster or aggregate we mean a set of small objects in which the distances between them are small as compared to the size of the aggregate itself; otherwise there is no way of speaking of aggregate—all we have are isolated objects. Indeed the WTMM method allows us to discriminate two classes of singularities from the space-scale information embedded in the WT skeleton, and then to characterize separately the two resulting subsets by computing the corresponding partition functions and multifractal spectra [115]. In Figure 22, we show synthetic images of clusters of small spots of various heights over a background 2D f Bm BH (x) rough surface of Hurst exponent H ¼ 0:6. The trivial case of a single isolated spot is shown in Figure 22a. Figure 22b–d displays small spots located on a straight line, on the border of a filled-in Julia set and on a dense area, respectively. Let us recall that Julia sets are beautiful objects that arise in the study of iteration of rational functions on the complex plane [1,2,28]. Here we use the wellknown example of a quadratic polynom as iterating function fc : z ! z2 þ c, with c ¼ 0:85 þ 0:20 i. The Julia set is just the set of initial seed z0 such that the iterated sequence ðzn Þn2N ; znþ1 ¼ z2n þ c does not go to infinity. In Figure 23 is shown the filled-in Julia set that we have used to compute Figure 22c. Each of these clusters has a known fractal dimension, respectively 0, 1, 1.68, and 2 for the point, line, Julia, and dense clusters. Although these spots are not singularities but localized structures with Gaussian shape of width ¼ 3 pixels and random heights, the 2D WTMM
MULTIFRACTAL IMAGE ANALYSIS
39
Figure 22. Synthetic rough surfaces (512 512) with a 2D fBm background of Hurst exponent H ¼ 0:6 and containing a cluster of localized spots. (a) The cluster contains only one spot in the middle of the image; (b) the spots are located on a straight line; (c) the spots are located on a Julia set; (d) the spots are randomly distributed in a square. The spots are modeled by a gaussian of width ¼ 3 pixels and height randomly chosen in the range [1.2, 1.8] in BH ¼ 0:6 unit. Same gray coding as in Figure 9.
Figure 23. Filled-in Julia set with parameter c ¼ 0:85 þ 0:20 i.
method can be used in a very efficient way to identify them and characterize the geometric properties of the aggregate to which they belong. As shown in Section II.B, in the WT skeleton, one expects the maxima lines pointing to the background texture to display local scaling properties corresponding to a 2D f Bm surface, i.e. hðxÞ ¼ H ¼ 0:6 [Eq. (15)], whereas maxima lines pointing to clustered spots are expected to display different local scaling
40
´ ODO ET AL. ARNE
Figure 24. Scaling behavior of the WT modulus along some maxima lines of the WT skeletons computed from the images shown in Figure 22a, 22b, 22c and 22d respectively using the first-order radially symmetric analyzing wavelet shown in Figure 1. Symbols ( ) are for maxima lines pointing to background BH ¼ 0:6 texture singularities and (4) for those pointing to clustered small spots. The solid (resp. dashed) line corresponding to scaling exponent h ¼ 0:6 (resp h ¼ 1) is drawn the guide the eyes.
properties with exponent h ¼ 1 since they are seen (at scales a > W > by our W -resolved WT microscope) as Dirac singularities. Notice that because these spots are quite smooth localized structures, one expects the WTMM on these maxima lines to display a crossover at small scales (a> W ) toward the behavior M ½ f ðx0 ; aÞ an ; a ! 0þ [Eq. (17)] dictated by the number of zero moment of the analyzing wavelet. In Figure 24 are shown, in a logarithmic representation, the WT modulus versus scale parameter a for various maxima lines belonging to the WT skeletons computed from the four images in Figure 22. For each of the analyzed images, maxima lines pointing to small spots clearly display a crossover from some increase of M at small scales to a clear power-law decrease at large scales with a local scaling exponent h> 1 that is negative and thus can be easily distinguished from the monotonous power-law increase M a0:6 observed along the maxima line pointing to a h ¼ H ¼ 0:6 background singularity. Now if one proceeds to the computation of the partition functions Z(q, a) on the subskeleton corresponding to identified small spots, one gets the results reported in Figure 25. Even though there are quite few maxima lines in this
MULTIFRACTAL IMAGE ANALYSIS
41
Figure 25. Determination of the fractal dimension DF ¼ ð0Þ of the cluster of localized spots in Figures 22a ( : isolated spot), 22b (4: linear cluster), 22c (h: Julia cluster) and 22d (*: dense cluster). log2 Zðq ¼ 0; aÞvs: log2 ðaÞ as computed with the 2D WTMM method after discriminating the WT subskeleton corresponding to clustered spots as illustrated in Figure 24. The solid lines correspond to the theoretical fractal dimensions DF ¼ 0; 1; 1:68 and 2 respectively.
subskeleton, one gets a rather nice scaling behavior for small values of q. In particular the estimate of the exponent ð0Þ for q ¼ 0 achieves our aim to classify geometrically these clusters of localized spots. Within numerical uncertainties, one obtains the following estimates of the fractal dimensions DF ¼ ð0Þ ¼ 0 0:02; 1 0:02; 1:7 0:04 and 2 0:02 for single spot, linear cluster, Julia cluster, and dense cluster. These results are in quite good agreement with the theoretical DF values. They illustrate the ability of WTMM methodology to extract clustered objects from a nontrivial background and to retrieve a geometric characterization of the cluster, via the estimate of its fractal dimension DF .
IV. Multifractal Analysis of High-resolution Satellite Images of Cloud Structure The problem of nonlinear variability over a wide range of scales has been considered for a long time with respect to the highly intermittent nature of turbulent flows in fluid dynamics [18,44]. Special attention has been paid to their asymptotic and possibly universal behavior when the dissipation length goes to zero, i.e., when the Reynolds number goes to infinity. Besides windtunnel and laboratory (grid, jet, etc.) experiments, the atmosphere is a huge natural laboratory in which high Reynolds number (fully developed) turbulent dynamics can be studied. Clouds, which are at the source of the
42
´ ODO ET AL. ARNE
hydrological cycle, are the most obvious manifestation of the earth’s turbulent atmospheric dynamics [10,172,173]. By modulating the input of solar radiation, they play a critical role in the maintenance of the earth’s climate [174]. They are also one of the main sources of uncertainty in current climate modeling [175], where clouds are assumed to be homogeneous media lying parallel to the earth’s surface; at best, a linear combination of cloudy and clear portions according to cloud fraction is used to account for horizontal inhomogeneity when predicting radiative properties. For many years, the lack of data hindered our understanding of cloud microphysics and cloud–radiation interactions. It is now well-recognized that clouds are variable in all directions and that fractal [172,173,176–181] and multifractal [10,182–184] concepts are likely to be relevant to the description of the complex 3D geometry of clouds. Until quite recently, the internal structure of clouds was probed by balloons or aircrafts that penetrated the cloud layer, revealing an extreme variability of 1D cuts of some cloud fields [184– 192]. In particular, in situ measurements of cloud liquid water content (LWC) were performed during many intensive field programs (FIRE [193], ASTEX [194], SOCEX [195], etc.). Indeed, during the past 15 years, vast amounts of data on the distribution of atmospheric liquid water from a variety of sources were collected and analyzed in many different ways. All these data contain information on spatial and/or temporal correlations in cloudiness, enabling the investigation of scale invariance over a range from a few centimeters to hundred of kilometers. An attractive alternative to in situ probing is to use high-resolution satellite imagery that now provides direct information about the fluctuations in liquid water concentration in the depths of clouds [177,179–181,196–202]. These rather sophisticated remote sensing systems called ‘‘millimeter radars’’ are actually sensitive not only to precipating raindrops but also to suspended cloud droplets. Spectral analysis of the recorded 2D radiance field [196–202] confirms previous 1D findings that make it likely that cloud scenes display scaling over a wide range of scales. One has to give credit to Lovejoy and co-workers [120–123,182,183,203– 206] for applying the multifractal description to atmospheric phenomena. Using trace moment and double trace moment techniques [120–123, 204–206], they have brought experimental evidence for multiple scaling (or in other words, the existence of a continuum of scaling exponent values) in various geophysical fields. More recently, Davis and co-workers [184,192,202] have used the structure function method to study LWC data recorded during ASTEX and FIRE programs. Both these analyses converge to the conclusion that the internal marine stratocumulus (Sc) structure is multifractal over at least three decades in scales. Similar multifractal behavior has been reported by Wiscombe et al. [201] when analyzing liquid
MULTIFRACTAL IMAGE ANALYSIS
43
water path (LWP) data (i.e., column integrated LWC), from the Atmosperic Radiation Measurement (ARM) archives. Even though all these studies seem to agree, at least as far as their common diagnostic of multifractal scaling of the cloud structure, they all concern 1D data. To our knowledge, the structure function method has also been applied to 1D cuts of highresolution satellite images [197,207], but we are not aware of any results coming from a specific 2D analysis. Our goal here is to take advantage of the 2D WTMM method to carry out a multifractal analysis of highresolution satellite images of Sc cloudy scenes [106,107,110]. Beyond the issue of improving statistical characterization of in situ and remotly sensed data, there is a most challenging aspect, which consists in extracting structural information to constraint stochastic cloud models, which in turn will be used for radiative transfer simulations [180,182,202,208–215]. Then by comparing the multifractal properties of the numerically generated artificial radiation fields with those of actual measurements, one can hope to achieve some degree of closure.
A. Landsat Data of Marine Stratocumulus Cloud Scenes Over the past 15 years, Landsat imagery has provided the remote sensing community at large with a very attractive and reliable tool for studying the Earth’s environment [177,179–181,196–202,216,217]. One of the main advantages of high-resolution satellite imagery is its rather low effective cost as compared to outfitting and flying research aircraft. Moreover this instrument is well calibrated and it offers the possibility of reaching unusual high spatial, spectral, and radiometric resolutions [197,216]. Mainly two types of statistical analysis have been applied so far to Landsat imagery: spectral analysis of the 2D radiance field [196–200,216] and joint area and perimeter distributions for ensembles of individual clouds [177,179–181] defined by some threshold in radiance. One of the most remarkable properties of Landsat cloud scenes is their statistical scale-invariance over a rather large range of scales, which explains why fractal and multifractal concepts have progressively gained more acceptance in the atmospheric scientific community [10]. Of all cloud types, marine stratocumulus (Sc) are without doubt the ones that have attracted the most attention, mainly because of their first–order effect on the Earth’s energy balance [10,173,197,216,218]. Being at once very persistent and horizontally extented, marine Sc layers carry considerable weight in the overall reflectance (albedo) of the planet and, from there, command a strong effect on its global climate [174]. Furthermore, with respect to climate modeling [175] and the major problem of cloud–radiation
44
´ ODO ET AL. ARNE
interaction [182,196,197,208–211], they are presumably at their simplest in marine Sc that are relatively thin ( 300 500 m), with well-defined (quasiplanar) top and bottom, thus approximating the plane-parallel geometry in which radiative transfer theory is well developed [173,182, 197,209,210,213]. However, because of its internal homogeneity assumption, plane-parallel theory shows systematic biases in large-scale average reflectance [210,219] relevant to Global Circulation Model (GCM) energetics and large random errors in small-scale values [213,220] relevant to remote-sensing applications. Indeed, marine Sc have huge internal variability [184,192], not necessarily apparent to the remote observer. In this section we challenge previous analysis [177,179–181,196–202, 216,217] of Landsat imagery using the 2D WTMM methodology [106–110] with the specific goal of improving statistical characterization of the highly intermittent radiance fluctuations of marine Sc, a prerequisite for developing better models of cloud structure and, in turn, furthering our understanding of cloud–radiation interaction. For that purpose, we analyze [110] a (’ 196 168 km2 ) original cloudy Landsat 5 scene captured with the TM camera (1 pixel ¼ 30 m) in the 0.6–0.7 m channel (i.e., reflected solar photons as opposed to their counterparts emitted in the thermal infrared) during the first ISCCP (International Satellite Cloud Climatology Project) Research Experiment (FIRE) field program [193], which took place over the Pacific Ocean off San Diego in the summer of 1987. For computational convenience, we actually select 32 overlapping 1024 1024 pixels2 subscenes in this cloudy region. The overall extent of the explored area is about 7840 km2 . Figure 26a shows a typical (1024 1024) portion of the original image where the eight-bit gray scale coding of the quasinadir viewing radiance clearly reveals the presence of some anisotropic texture induced by convective structures that are generally aligned to the wind direction.
B. Application of the 2D WTMM Method to Landsat Images of Stratocumulus Clouds We systematically follow the numerical implementation procedure described in Section II.C. We first wavelet transform the 32 overlapping (1024 1024) images, cut out of the original image, with the first-order (n ¼ 1) radially symmetric analyzing wavelet defined in Figure 1. From the wavelet transform skeleton defined by the WTMMM, we compute the partition functions from which we extract the ðqÞ and D(h) multifractal spectra. We systematically test the robutness of our estimates with respect to some change in the shape of the analyzing wavelet, in particular when increasing the number of zero moments.
MULTIFRACTAL IMAGE ANALYSIS
45
Figure 26. 2D wavelet transform analysis of a Landsat image of marine Sc clouds [110]. (x) is the first-order radially symmetric analyzing wavelet shown in Figure 1. (a) A 256 grayscale coding of a (1024 1024) portion of the original radiance image. In (b) a ¼ 22:9 W , (c) a ¼ 21:9 W , and (d) a ¼ 23:9 W (where W ¼ 13 pixels ’ 390 m), are shown the maxima chains; the local maxima of M along these chains are indicated by () from which originates an arrow whose length is proportional to M and its direction (with respect to the x-axis) is given by A ; only the central (512 512) part delimited by a dashed square in (a) is taken into account to define the WT skeleton. In (b), the smoothed image b;a I is shown as a gray-scale coded background from white (min) to black (max).
1. Numerical Computation of the Multifractal ðqÞ and D(h) Spectra Figure 26 illustrates the computation of the maxima chains and the WTMMM for the marine Sc subscene. After linking these WTMMM across scales, one constructs the WT skeleton from which one computes the partition functions Z(q, a) [Eq. (21)]. As reported in Figure 27a, the annealed average partition functions () display some well-defined scaling behavior over the first three octaves, i.e., over the range of scales 390 m < a < 3120 m, when plotted versus a in a logarithmic representation. Indeed the scaling deteriorates progressively from the large scale side when one goes to large values of jqj> 3. As discussed [110], besides the fact that we are suffering from insufficient sampling, the presence of localized Dirac-like structures likely explains the fact that the observed crossover to a steeper power-law decay occurs at a smaller and a smaller scale when one increases q > 0. Actually for q> 3, the crossover scale a* < 1200 m becomes significantly smaller than the so-called integral scale, which is approximately
46
´ ODO ET AL. ARNE
Figure 27. Determination of the ðqÞ and D(h) spectra of radiance Landsat images of marine Sc. The 2D WTMM method is used with either a first-order () or a third-order ( ) radially symmetric analyzing wavelet (see Fig. 1). (a) log2 Zðq; aÞ vs. log2 a; the solid lines correspond to linear regression fits of the data over the first octave and a half. (b) ðqÞ vs. q as obtained from a linear regression fit of the data in (a). (c) D(h) vs. h, after Legendre transforming the ðqÞ curve in (b). In (b) and (c), the solid lines correspond to the theoretical multifractal spectra for log-normal W-cascades with parameter values m ¼ 0:38 ln 2 and 2 ¼ 0:07 ln 2 [Eqs. (42) and (43)]. The D(h) singularity spectrum of velocity (dotted line) and temperature (dashed line) fluctuations in fully developed turbulence are shown for comparison in (c).
given by the characteristic width ’ 1 5–6 km of the convective rolls (Fig. 26a). When processing to a linear regression fit of the data in Figure 27a over the first octave and a half (in order to avoid any bias induced by the presence of the observed crossover at large scales), one gets the ðqÞ spectrum () shown in Figure 27b. In contrast to the fractional Brownian rough surfaces studied in Section III.A [108], this ðqÞ spectrum unambiguously deviates from a straight line. When Legendre transforming this nonlinear ðqÞ curve, one gets the D(h) singularity spectrum reported in Figure 27c. Its characteristic single humped shape over a finite range of Ho¨lder exponents is a clear indication of the multifractal nature of the marine Sc radiance fluctuations. We have checked [110] that the estimate of the D(h) singularity spectrum from the scaling behavior of the partition functions h(q, a) [Eq. (29)] and D(q, a) [Eq. (30)] yields similar quantitative results.
MULTIFRACTAL IMAGE ANALYSIS
47
In Figure 27 are also shown for comparison the results ( ) obtained when applying the 2D WTMM method with a third-order (n ¼ 3) radially symmetric analyzing wavelet (the smoothing function being the isotropic 2D Mexican hat). As seen in Figure 27a, the use of a wavelet that has more zero moments seems to somehow improve scaling. For the range of q-values investigated, the crossover scale turns out to be rejected at a larger scale, enlarging by some amount the range of scales over which scaling properties can be measured, especially for the largest values of jqj. The fact that one improves scaling when increasing the order of the analyzing wavelet suggests that perhaps some smooth behavior unfortunately deteriorates our statistical estimate of the multifractal spectra of the original Landsat radiance image. Let us recall that, as explained in Section II.B, smooth C 1 behavior may give rise to maxima lines along which M an (Fig. 6b); hence the larger n , the smaller is the overall contribution of those ‘‘spurious’’ maxima lines in the partition function summation over the WT skeleton. As seen in Figures 26a, the anisotropic texture induced by the convective streets or rolls might well be at the origin of the relative lack of well-defined scale invariance. When looking at the corresponding ðqÞ spectrum ( ) extracted from the data in Figure 27b, one gets quantitatively the same estimates for q> 1. For more negative values of q, the data obtained with the third-order analyzing wavelet clearly depart from the previous estimates with the first-order wavelet. The slope of the new ðqÞ spectrum is somehow weakened, which implies, from the Legendre transform properties, that the corresponding values of hðqÞ ¼ @=@q are reduced. The computation of the D(h) singularity spectrum ( ) in Figure 27c enlightens this phenomenon: while the increasing left-hand branch (which corresponds to the strongest singularities) of the D(h) curve appears to be quite robust with respect to the choise of , the decreasing right-hand branch (associated to the weakest singularities) is modified when increasing the number of zero moments of . As shown in Figure 27b and c, the D(h) spectrum as well as the ðqÞ spectrum data are very well fitted by the theoretical quadratic spectra predicted for log-normal random W-cascades [Eqs. (42) and (43)]. However, with the first-order analyzing wavelet, the best fit is obtained with the parameter values m ¼ 0:38 ln 2 ¼ 0:263 and 2 ¼ 0:07 ln 2 ¼ 0:049, while for the third-order wavelet these parameters take slightly different values, namely m ¼ 0:366 ln 2 ¼ 0:254 and 2 ¼ 0:06 ln 2 ¼ 0:042. The variance parameter 2 that characterizes the inter mittent nature of marine Sc radiance fluctuations is therefore somehow reduced when going from n ¼ 1 to n ¼ 3. Actually the lack of statistical convergence because of insufficient sampling is actually the main reason for this uncertainty in the estimate of 2 [110]. As previously experienced [109] for synthetic multifractal rough surfaces, an accurate
48
´ ODO ET AL. ARNE
Figure 28. Pdfs of the WTMMM coefficients of the 32 (1024 1024) radiance Landsat images as computed with the first-order radially symmetric analyzing wavelet. (a) Pa ðMÞ vs. M; (b) Pa ðAÞ vs. A; the symbols correspond to the following scales: a ¼ 20:3 W ¼ 480; mðÞ; 21:3 W ¼ 960 mð Þ, and 22:3 W ¼ 1920 m ðÞ. The solid lines in (a) correspond to lognormal distributions.
estimate of the exponents ðqÞ for q< 3 requires more than 32 (1024 1024) images. With the statistical sample of Landsat images we have at our disposal, one gets Dðhðq ¼ 0Þ ¼ 0:37 0:02Þ ¼ 2:00 0:01, which is a strong indication that the radiance field is singular everywhere. From the estimate of ðq ¼ 2Þ ¼ 1:38 0:02, one gets the following estimate of the spectral exponent: ¼ ð2Þ þ 4 ¼ 2:62 0:02, i.e., a value in good agreement with previous estimates [185–189,191,196–200,216]. 2. WTMMM Probability Density Functions This subsection is mainly devoted to the analysis of the joint probability distribution function Pa ðM; AÞ [108–110] as computed from the wavelet transform skeletons of the 32 (1024 1024) radiance images with the first-order radially symmetric analyzing wavelet R (n ¼ 1). In Figure 28a and b are respectively shown the pdfs P ðMÞ ¼ dAPa ðM; AÞ and Pa ðAÞ ¼ a R dMPa ðM; AÞ for three different values of the scale parameter a ¼ 20:3 W (480 m), 21:3 W (960 m), and 22:3 W (1920 m). First let us focus on the results shown in Figure 28b for Pa ðAÞ. This distribution is clearly scale dependent with some evidence of anisotropy enhancement when going from small to large scales, in particular when one reaches scales that become comparable to the characteristic width of the convective structures (i.e., a few kilometers wide). Two peaks around the values A ’ 1 =6 and 5=6 become more and more pronounced as the signature of a privileged direction in the analyzed images. As one can check from a visual inspection of Figure 26a, this direction is nothing but the perpendicular to the mean direction of the convective rolls that are generally aligned to the wind direction. This is another clear indication that at large scales, the wavelet
MULTIFRACTAL IMAGE ANALYSIS
49
Figure 29. Pdfs of the WTMMM coefficients of the 32 (1024 1024) radiance Landsat images as computed with a first-order radially symmetric analyzing wavelet. Pdfs of M when conditioned by A. The different symbols correspond to fixing A (mod ) to 0 =8ð Þ; =4 =8 (h), =2 =8 ð4Þ, and 3=4 =8 (&). (a) a ¼ 20:3 W ¼ 480; m; (b) a ¼ 21:3 W ¼ 960; m.
transform microscope is sensitive to the convective roll texture, a rather regular modulation superimposed to the background radiance fluctuations [107–110]. Another important message that comes out from our analysis is illustrated in Figure 29. When conditioning the pdf of M by the argument A, the shape of this pdf is shown to be independent of the considered value of A, as long as the value of the scale parameter a remains small as compared to the characteristic width of the convective structures. The observation that the joint probability distribution actually factorizes, i.e., Pa ðM; AÞ ¼ Pa ðMÞPa ðAÞ, indicates that M and A are likely to be independent [107,110]. This implies that all the multifractal properties of the marine Sc radiance fluctuations are contained in the way the shape of the pdf of M evolves when one decreases the scale parameter a. This evolution is illustrated in Figure 28a when using a first-order radially symmetric analyzing wavelet. Since by definition the WTMMM are different from zero, Pa ðMÞ decreases exponentially fast to zero at zero. As previously emphasized [108], this observation is at the heart of the 2D WTMM method, which, for this reason, does not suffer any problem with divergency when estimating the ðqÞ spectrum for q < 0. As shown in Figure 28a for any scale significantly smaller than the integral scale ( 5–6 km, as given by the characteristic width of the convective structures), all the data points fall, within a good approximation, on a log-normal curve [106,110]. As shown [110], this experimental feature is not specific to some particular shape of the analyzing wavelet since log-normal pdfs are also found when using a third-order radially symmetric analyzing wavelet.
50
´ ODO ET AL. ARNE
Figure 30. Magnitude correlation function Cðx; a1 ; a2 Þ vs. log2 ðxÞ, as computed from the 32 (1024 1024) radiance Landsat images using a first-order radially symmetric analyzing wavelet. (a) WTMMM magnitude: !ðx; aÞ ¼ ln½M ½ f ½Lx ðaÞ. (b) Continuous WT magnitude: !ðx; aÞ ¼ 12 ln "2 ðx; aÞ [Eq. (48)]. The symbols have the following meaning: a1 ¼ a2 ¼ 2 W ¼ 780; mð Þ; a1 ¼ W ¼ 390; m; a2 ¼ 2 aW ¼ 780 mð4Þ; a1 ¼ W ¼ 390 m; a2 ¼ 4 W ¼ 1560 m ðhÞ; a1 ¼ 2 W ¼ 780; m; a2 ¼ W ¼ 1560; m ( ). The solid (dashed) lines correspond to the theoretical prediction [Eq. (50)] for multifractal rough surfaces generated with the random W-cascade model with parameters 2 ¼ 0:08 ln 2 (0.16 ln 2) and L ¼ 220 pixels ¼ 6.6 km.
C. Space-Scale Correlation Function Analysis of Radiance Landsat Images As pointed out in Section III.C, the real demonstration of the existence of an underlying multiplicative structure consists in taking advantage of the space-scale unfolding provided by the continuous wavelet transform to compute the cross-scale correlation functions. In Figure 30 the results of the computation of Cðx; a1 ; a2 Þ when averaging over the 32 (1024 1024) radiance Landsat images, using either the WTMMM (Fig. 30a) or the continuous WT (Fig. 30b) definition of the magnitude of f (Section III.C [110]) are reported. One can see that for x > supða1 ; a2 Þ, all the data points fall, in good approximation, onto a unique curve when plotted versus log2 ðxÞ, independently of the considered pair of scales (a1 ; a2 ). Moreover, a straight line of slope 2 ¼ 0:012 provides a rather reasonable fit of the data up to a separation distance x ’ 27 pixels ’ 3.8 km, where decorrelation seems to be attained. Note that using the WTMMM instead of the continuous WT does not make any difference; this is a strong indication of the existence of some ultrametric properties underlying the branching structure of the space-scale wavelet representation of the radiance fluctuations. On top of the data in both Figure 30a and b, we have shown, for comparison, the theoretical prediction [Eq. (50)] for the ‘‘two-scale’’ correlation function of multifractal rough surfaces generated by the random W-cascade model. This formula provides a reasonable fit of the data when adjusting the model parameters to 2 ¼ 0:16 ln 2 and L ¼ 220
MULTIFRACTAL IMAGE ANALYSIS
51
pixels ¼ 6.6 km. Although the estimate of the integral scale seems to be of the right order of magnitude as regard to the characteristic width ( 5–6 km) of the convective rolls, the value obtained for the intermittency parameter 2 is about twice as large as previous estimates derived from the WTMM computation of the ðqÞ and D(h) multifractal spectra in Figure 27. At this point, let us emphasize that a similar discrepancy has been previously noticed in the WTMM analysis of wind tunnel turbulent velocity fields [85,152]. It may suggest that simple scale-invariant self-similar cascades as pictured by the random W-cascade model are not sophisticated enough to account for the space-scale structure of the radiance fluctuations in marine Sc clouds. The interpretation of this feature in terms of correlations between weights at a given cascade step or in terms of a more complex geometry of the tree underlying the multiplicative structure of the radiance field is underway. The possible importance of the intermittently distributed localized downward spike structures is also under consideration. Before drawing definite conclusions, there is clearly a need to repeat the ‘‘two-point’’ correlation function analysis on the background radiance fluctuations, once all the maxima lines corresponding to those Dirac-like singularities are removed from the WT skeleton.
D. Comparative WTMM Multifractal Analysis of Landsat Radiance Field and Velocity and Temperature Fields in Fully Developed Turbulence Let us point out that a similar 1D WTMM analysis of the velocity fluctuations in high Reynolds number turbulence has come to conclusions very close to those of the present study [81–87,221]. Besides the presence of rather localized Dirac-like structures that witness the probing of vorticity filaments [62,84,127,221], the multifractal nature of turbulent velocity is likely to be understood in terms of a log-normal cascading process that is expected to be scale-invariant in the limit of very high Reynolds numbers [81–87]. In Figure 27c are shown for comparison the results obtained for the D(h) singularity spectrum of the radiance Landsat images together with the D(h) data extracted from the 1D analysis of a turbulent velocity signal recorded at the Modane wind tunnel (R ’ 2000) [82,85] [indeed DðhÞ þ 1 is represented for the latter in order to compare 1D to 2D data]. The turbulent velocity D(h) spectrum significantly differs from the results obtained for the marine Sc cloud. They have a common feature, i.e., the Ho¨lder exponent most frequently encountered in the radiance field h ¼ m= ln 2 ¼ hðq ¼ 0Þ ¼ @ =@qjq¼0 ¼ 0:38 0:01 is indistinguishable from the corresponding exponent h ¼ hðq ¼ 0Þ ¼ 0:39 0:01 found for the turbulent velocity field. Note that these values are significantly larger than the theoretical value
52
´ ODO ET AL. ARNE
h ¼ 1=3 predicted by Kolmogorov in 1941 [222] to account for the observed k5=3 power-spectrum behavior. The main difference comes from the intermittency parameter, which is much stronger for the cloud, 2 = ln 2 ¼ 0:07 0:01ðn ¼ 1Þ or 2 = ln 2 ¼ 0:06 0:01ðn ¼ 3Þ, than for the turbulent velocity, 2 =ln 2 ¼ 0:036 0:004. This indicates that the radiance field is much more intermittent than the velocity field: the D(h) singularity spectrum for the former is unambiguously wider than the corresponding spectrum for the later. For the sake of comparison, in Figure 27c we have also reported the multifractal D(h) spectrum of the temperature fluctuations recorded in a R ¼ 400 turbulent flow [223]. The corresponding single humped curve is definitely much wider than the velocity D(h) spectrum and it is rather close to the data corresponding to the marine Sc radiance field. It is well recognized, however, that liquid water is not really passive and that its identification with a passive component in atmospheric dynamics offers limited insight into cloud structure since, by definition, near-saturation conditions prevail and latent heat production affects buoyancy [202]. So cloud microphysical processes are expected to interact with the circulation at some, if not all, scales [224]. Nevertheless, our results in Figure 27c indicate that from a multifractal point of view, the intermittency captured by the Landsat satellite looks statistically equivalent to the intermittency of a passive scalar in fully developed 3D turbulence. The fact that the internal structure of Sc cloud somehow reflects some statistical properties of atmospheric turbulence is not such a surprise in this highly turbulent environment. The investigation of different sets of Landsat data is urgently required in order to test the degree of generality of the results reported in this first WTMM analysis of high-resolution satellite images. In particular, one may wonder to what extent the marine Sc Landsat data collected off the coast of San Diego on July 7, 1987 under specific observation conditions actually reflect the specific internal structure of Sc clouds. Work in this direction is currently in progress. Finally, with respect to the issue of cloud modeling, the WTMM analysis of marine Sc Landsat data indicates that the 2D random W-cascade models introduced [109] are much more realistic hierarchical models than commonly used multifractal models such as the fractionally integrated singular cascade [120,123,200,205,216] or the bounded cascade models [218,225]. We are quite optimistic in view of using the log-normal Wcascade models with realistic parameter values for radiation transfer simulations. To our opinion, random W-cascade models are a real breakthrough, not only for the general purpose of image synthesis, but more specifically for cloud modeling. It is likely that better cloud modeling will enable further progress in our understanding of cloud–radiation interactions possible.
53
MULTIFRACTAL IMAGE ANALYSIS
V. Multifractal Analysis of 3D Turbulence Simulation Data A. Multifractal Description of Intermittency 1. Intermittency Based on the Velocity Field Since Kolmogorov’s founding work in 1941 (K41) [222], fully developed turbulence has been intensively studied theoretically, numerically, and experimentally [18,44,226–229]. A standard way of analyzing a turbulent flow is to look for some universal statistical properties of the fluctuations of the velocity increments over a distance l: vðr; leÞ ¼ vðr þ leÞ vðrÞ
ð51Þ
where e is an arbitrary unit vector. For instance, investigating the scaling properties of the longitudinal structure functions: Sp ðlÞ ¼ < ðe:vðr; leÞÞp > l p ;
p>0
ð52Þ
where < > stands for ensemble average, leads to a spectrum of scaling exponents p that has been widely used as a statistical characterization of turbulent fields [18,44,226,228,229]. Based upon assumptions of statistical homogeneity, isotropy, and constant mean energy dissipation per unit mass , K41 asymptotic theory predicts the existence of an inertial range l L for which the structure functions behave as Sp ðlÞ p=3 l p=3
ð53Þ
where is the Kolmogorov dissipative scale and L the so-called integral scale. Although these assumptions are usually considered to be correct, there has been increasing numerical [18,226,230,231] and experimental [18,44,226,228,229,232–240] evidence that p deviates substantially from the K41 prediction p ¼ 13 p, at large p. The observed nonlinear behavior of the p spectrum actually characterizes some evolution of the longitudinal velocity increment probability density function (pdf) in the inertial range, from a Gaussian shape at large scales to stretched exponential tails toward smaller scales [228,234,235,241–245]. This evolution of the longitudinal velocity increment statistics across scales is at the heart of the multifractal description of the intermittency of small scales, pioneered by Parisi and Frisch in 1985 [43]. K41 theory is actually based on the assumption that at each point r of the fluid, the velocity field has the same scaling behavior e:vðr; leÞ l 1=3 , which yields the well-known EðkÞ ¼ k5=3 energy spectrum [18]. By interpreting the nonlinear behavior of p as a direct consequence of the existence of spatial fluctuations in the local regularity of the velocity field, namely
54
´ ODO ET AL. ARNE
e:vðr; leÞ l hðrÞ
ð54Þ
where the exponent h depends upon r, Parisi and Frisch [43] propose to capture intermittency in a geometric framework. For each h, if one calls D(h) the fractal dimension of the set of spatial points r for which hðrÞ ¼ h, then by suitably inserting this local scaling behavior [Eq. (54)] into Eq. (52), one can bridge the so-called singularity spectrum D(h) and the set of scaling exponent p by a Legendre transform: DðhÞ ¼ minð ph p þ dÞ p
ð55Þ
where d ¼ 3 is the dimension of the velocity field. From the properties of the Legendre transform, a nonlinear p spectrum is equivalent to the assumption that there is more than a single exponent h. But as already mentioned in the introduction (Section I), Eq. (55) is valid for positive (integer) p values only, which precludes the computation of the entire D(h) spectrum (in particular its right decreasing part corresponding to the weakest singularities is inaccessible to the structure function method) [50]. In the early 1990s, the 1D WTMM method [47–50] was introduced to overcome the insufficiencies of the numerical techniques commonly used to perform multifractal analysis (e.g., the structure function method, and the box-counting techniques). The use of wavelets (instead of increments or boxes) actually allows us to compute partition functions that scale like Zðq; aÞ aðqÞ , where the exponents ðqÞ are nothing but a generalization of the exponents p in the sense that q is now a real number going from 1 to þ1. Then, as demonstrated [74,75], one can prove that by Legendre transforming the ðqÞ spectrum, one gets both the increasing ðq > 0Þ and the decreasing ðq < 0Þ parts of the D(h) singularity spectrum. Preliminary results obtained for high Reynolds wind tunnel experimental data with the 1D WTMM method have confirmed the nonlinearity of the ðqÞ spectrum and consequently the multifractal nature of the longitudinal velocity fluctuations [47–49]. Let us note that from low to moderate Reynolds number turbulence, the inertial range revealed in numerical simulations as well as in experiments is rather small, which makes the estimate of the scaling exponents p and ðqÞ not very accurate. Actually, the existence of scaling laws such as Eq. (52) for the structure functions [240,246,247], as well as for the WTMM partition functions [81–83,85], is not clear experimentally, even at the highest accessible Reynolds numbers. Indeed, there is a persistent curvature when one plots ln ½Sp ðlÞ vs. ln (l), which means that, rigorously speaking, there is no scale invariance. This observation somehow questions the validity of the multifractal description. Benzi et al. [248–250] proposed some remedy to the observed departure from scale invariance by looking at the scaling behavior of one structure function against another. More precisely, p can be
MULTIFRACTAL IMAGE ANALYSIS
55
estimated from the behavior Sp ðlÞ S3 ðlÞp , if one assumes that 3 ¼ 1 ½18. The relevance of the so-called extended self-similarity (ESS) hypothesis improves and further extends the scaling behavior toward the dissipative range [230,248–250]. From the application of ESS, some broad consensus among European researchers was reached in 1996 [240], at least as far as isotropic homogeneous turbulence is concerned. In this context, the ESS hypothesis has received strong support from the ‘‘propagator (across scales)’’ approach originally developed by Castaing and co-workers [235,246,251–257] and recently revisited with the wavelet transform methodology [81–83,85,221,258]. Let us notice that Castaing’s approach can be linked to the recently proposed Fokker–Planck/Langevin description of intermittency [259–261]. According to this description, the velocity field is a Markov process across scales which suggests that the velocity increment pdf at different scales obeys a Fokker–Planck differential equation characterized by a drift and a diffusion coefficient. Even though this description remains, to a large extent, formal from a mathematical point of view and very phenomenological, it can be interesting because of its great versatility as far as scaling behavior is concerned [262]. Let us note that some theoretical works have tried to build some bridge between the Fokker–Planck approach and the Navier–Stokes dynamics [263,264]. Very recently, a systematic computation of the cumulants of the magnitude ln je:vðr; leÞj of 1D longitudinal velocity profiles stemming from three different experimental setups and covering a broad range of Taylor-scaled Reynolds numbers from R ¼ 89 to 2500 has clearly revealed some inconsistency with the ESS hypothesis [87]. Indeed this study shows that the breaking of scale invariance is mainly contained in the first-order cumulant, which is found to strongly depend on Reynolds number and experimental conditions, whereas, surprisingly, the second-order cumulant displays universal scale invariance behavior from R values as low as R ’ 100. Furthermore, when extrapolating these results to the limit of infinite Reynolds number, this study confirms the asymptotic validity of the log-normal multifractal description of the intermittency phenomenon; the p spectrum is quadratic: p ¼ ð pÞ þ d ¼ C1 p C2
p2 2
ð56Þ
with a well-defined intermittency parameter C2 ¼ 0:025 0:003 [87]. Note that a plausible explanation to the scale invariance symmetry breaking observed in the magnitude first-order cumulant at finite Reynolds number [and which turns out to pollute the scaling behavior of Sp ðlÞ for every p] is the presence of anisotropic velocity fluctuations in the inertial range that are likely to originate from large-scale boundary and forcing effects. We refer
56
´ ODO ET AL. ARNE
the reader to Refs. [265–268], which show how to master these anisotropic effects using the irreducible representations of the rotation group. 2. Intermittency Based on the Energy Dissipation Field A central quantity in the K41 theory [222] is the mean energy dissipation , which is supposed to be constant [Eq. (53)]. The observed nonlinear behavior of the p spectrum [Eq. (56)] is generally interpreted as a direct consequence of the intermittency phenomenon displayed by , which is not spatially homogeneous but undergoes local intermittent fluctuations [18,226,227,235]. Under the so-called Kolmogorov refined similarity hypothesis (RSH) [269,270], the velocity structure functions can be rewritten as Sp ðlÞ < l ðrÞp=3 > l p=3 l ð p=3Þþp=3
ð57Þ
where l ðrÞ is the spatial average of the energy dissipation over a ball of radius l centered at the point r and of volume Vl l d : Z 1 ðr0 Þdd r0 ð58Þ l ðrÞ ¼ Vl Vl Note that the dissipation rate is related to the symmetric part of the strain tensor ði; j ¼ 1; 2; 3Þ: X ð@ þ @i j Þ2 ¼ i; j j i 2 ð59Þ X S S ; ¼ 2 ij ji i; j where
1 Sij ¼ ð@j i þ @i j Þ 2
ð60Þ
According to Eq. (57), the scaling exponents of Sp are thus related to those of l ðrÞ [18]: p ¼ ð p=3Þ þ p=3
ð61Þ
By Legendre transforming both sides of this equation, one gets the following relationship between the singularity spectra of and v:
¼ 3h;
f ð Þ ¼ DðhÞ
ð62Þ
where f ð Þ is the Hausdorff dimension of the set of spatial points such that l ðrÞ behaves like
MULTIFRACTAL IMAGE ANALYSIS
l ðrÞ l 1
as
l!0
57 ð63Þ
Considered as a measure, the dissipation has singularities of exponent 1 on sets of dimension: f ð Þ ¼ minðqð 1Þ ðqÞ þ dÞ q
ð64Þ
Several experimental and numerical works have tested various facets of the RSH hypothesis [164,221,227,250,251,271–282]. The support for the RSH is strong but not unequivocal. In the experiments, besides some possible artifact that may result from the use of the Taylor’s hypothesis (which consists in substituting time derivatives for space derivatives) [18], the socalled surrogacy issue concerns the shortcoming of replacing by its surrogate: 2 @u 0 ¼ 15 ð65Þ @x where u is the recorded longitudinal velocity component. Indeed the necessity of working with the surrogate dissipation amounts to assuming that the local dissipation is well approximated by an isotropic form, which is strictly valid in an ensemble-averaged sense in high Reynolds number flows and not obviously satisfied in real experimental conditions [283,284]. In the direct Navier–Stokes simulations (DNS), there are strong indications that the detailed structures of the pdfs of the energy dissipation and its 1D surrogate are different and that the velocity increments conditioned on 0l do not follow the RSH to the same degree as those conditioned on l [282]. Since Richardson’s cascade pioneering picture [285], multiplicative cascade models have enjoyed a lot of interest as the paradigm of methods for obtaining multifractal dissipation measures [1,2,18,67–71,76,122, 123,164,206,227]. The notion of cascade actually refers to a self-similar process whose properties are defined multiplicatively from coarse to fine scales. In that respect, it occupies a central place in the statistical theory of turbulence [18,69,226–228]. Over the past 40 years, since the log-normal model proposed by Kolmogorov [269] and Obukhov [270] (KO62) to account for the correction to K41 theory, refined cascade models have flourished in the literature such as the random -model, the -model, the p-model (for reviews see [18,69,227]), the log-stable models [120–123,286], and more recently the log-infinitely divisible cascade models [254,287–291] including the rather popular log-Poisson model advocated by She and Leveque [292]. Very generally, a self-similar cascade is defined by the way the scales are refined and the statistics of the multiplicative factor at each steps of the process [76,123,167,227,293,294]. One can thus distinguish
58
´ ODO ET AL. ARNE
discrete cascades that involve discrete scale ratios leading to log-periodic corrections to scaling (discrete scale invariance [295,296]) from continuous cascades without preferable scale factors (continuous scale invariance). As far as the fragmentation process is concerned, one can specify whether some conservation laws are operating or not [76]; in particular one can discriminate between conservative (the measure is conserved at each cascade step) and nonconservative (only some fraction of the measure is transferred at each step) cascades. More fundamentally, there are two main classes of self-similar cascade processes: deterministic cascades that generally correspond to solvable models [69,227] and random cascades that are likely to provide more realistic models but for which some theoretical care is required as far as their multifractal limit and some basic multifractal properties (including multifractal phase transitions) are concerned [76]. As a notable member of the later class, the independent random cascades introduced by Mandelbrot [297,298] as a general model of random curdling in fully developed turbulence have a special status since they are the main random cascade model for which deep mathematical results have been obtained [299,300]. Recently, these multiplicative random cascade models have been recast in a Fokker–Planck/Langevin description of the pdf of lnðl Þ across scales [301,302]. There has been early experimental attemps to measure the f ð Þ singularity spectrum [Eq. (64)] of the dissipation rate with the specific goal to discriminate between the most popular multiplicative cascade models [69,227,303–305]. Surprisingly, the simplest version of the weighted curdling models proposed by Mandelbrot [297,298], namely the binomial model, turns out to account reasonably well (at least at a certain level of description) for the observed multifractal ðqÞ and f ð Þ spectra (see Ref. [296] for a recent analysis). Indeed, all the existing cascade models appeal to adjustable parameters that are difficult to determine by plausible physical arguments and that generally provide enough freedom to account for the experimental data. Moreover, a quantitative validation of any model seems rather illusive since various technical difficulties may have disturbed the measurement of the dissipation multifractal spectra. We refer the reader to Ref. [306] for a review of the possible problems involved in the experimental process. We will mention only two main experimental limitations. The first one results from the fact that the multifractal model of turbulence implies a dependence of the viscous cutoff on the singularity exponent ð Þ=L ¼ Re3=ð3þ Þ [18,306–310]. It is thus a crucial question if the current hot-wire probes can resolve the scales implied by exponents significantly less than 1, i.e., those that correspond to the strongest singularities of the dissipation measure. The second one is the fact that single probe measurement of the longitudinal velocity requires the use of the 1D surrogate dissipation 0
MULTIFRACTAL IMAGE ANALYSIS
59
approximation [Eq. (65)], which may introduce severe bias in the estimate of the multifractal spectra mainly because of the presence of global and local anisotropic effects. A genuine 3D multifractal processing of turbulence dissipation data is at the moment feasible only for numerically simulated flows. But there is a price to pay for the additional gain of not using Taylor’s frozen flow hypothesis; these simulations are still somehow limited in Reynolds number to regimes where scaling just begins to manifest itself, thus making reliable measurements of multifractal properties difficult [247,274,282,311]. Nevertheless several numerical studies [274,312] agree that, at least at low and moderate Reynolds numbers, the 1D-surrogate energy dissipation is in general more intermittent than the full field, which is found nearly log-normal in the inertial range [274,313]. Note that some departure from log-normality can be observed for high-order moments (large q > 0) [274] and is likely to define local anisotropic effects induced by strongly localized events [314,315]. Besides the experimental difficulties of measuring the energy dissipation field, there is some additional intrinsic limitation to the multifractal analysis of turbulent fields that comes from the numerical techniques commonly used in the literature to process the experimental as well as the numerical data. For instance, the multifractal spectra p of the longitudinal velocity and ðqÞ of the energy dissipation are commonly computed using, respectively, the structure functions [18,43,44] and the box-counting [24,67,227] methods. The fact that the former method allows us to compute the longitudinal velocity exponents p for positive p values only explains why for many years the validity of the RSH relationships (61) and (62) has been partially tested [227,274]. More recent checks using the 1D WTMM method [221] and an alternative two-scale method [316,317] have clearly revealed the failure of Eq. (61) for negative p values when identifying with its surrogate 0 . This means that the decreasing part of the singularity spectra f0 ð Þ and D(h) (corresponding to the weakest singularities of both fields) significantly differs with respect to numerical uncertainty. Moreover, there is an implicit normalization constraint inherent to the box-counting technique, namely ð1Þ ¼ 0 ð1Þ ¼ 0, which makes this method quite inappropriate for studying nonconservative multiplicative cascade processes. Indeed a blind use of boxcounting algorithms will always yield multifractal spectra that can be misleading compared to the theoretical spectra of the conservative cascading process. 3. Intermittency Based on the Enstrophy Field An important step in the understanding of small-scale turbulence driven by expectations of universality is to proceed to a comparative statistical
60
´ ODO ET AL. ARNE
analysis of dissipation and enstrophy in isotropic turbulence. Note that the enstrophy is related to the antisymmetric part of the strain tensor: 1 i; j ð@j i @i j Þ2 2 ¼ 2i; j !ij !ji
ð66Þ
1 !ij ¼ ð@j i @i j Þ 2
ð67Þ
¼ where
The relationship between ¼ 2S2 and ¼ 2!2 ð! ¼ r ^ v is the usual vorticity pseudovector) is X ¼ þ2 ð@j i @i j Þ ð68Þ i;j From the incompressibility condition, one can show that the global averages of dissipation and enstrophy are related: <>¼< >
ð69Þ
But this does not imply that their local averages l ðrÞ [Eq. (58)] and l ðrÞ scale identically, where Z 1
l ðrÞ ¼
ðr0 Þdd r0 : ð70Þ Vl Vl Nevertheless, if they do, this will imply that the power-law scaling of < ql > and < ql > in the inertial range must be the same ðqÞ ¼ ðqÞ and in turn the corresponding singularity spectra f ð Þ ¼ f ð Þ. There has been interesting recent controversy concerning the relative scaling properties of enstrophy and dissipation densities. Different theoretical studies have converged to the conclusion that the asymptotic scaling exponents must be equal in the limit of infinite Reynolds number [265,266,318–320]. Pioneering numerical DNS studies [321,322] have shown that the field is more intermittent than the field. The 1D measurements of the streamwise components of and !, obtained at both high and low Reynolds numbers [323,324], conclude that the degrees of intermittency in the dissipation and the enstrophy fields are not the same. This observation is corroborated by the analysis of circulation data [325]. More recent DNS studies at moderate Reynolds number (R ¼ 216) [326] confirm that there are differences between the two scalings. As suggested by Chen et al. [327], this difference is likely to result from the difference observed in the scaling exponents pL and pT of longitudinal and transverse structure functions, respectively
MULTIFRACTAL IMAGE ANALYSIS
61
[256,328–333]. More precisely, Chen et al. [327] reported numerical results that demonstrate the possible validity of a different RSH for the transverse direction (RSHT) that connects the statistics of the transverse velocity increments with the locally averaged enstrophy in the inertial range. The important implication of RSHT is the possible existence of two independent sets of scaling exponents related, respectively, to the symmetric (dissipation physics) and antisymmetric (vortex dynamics) parts of the strain rate. But some caution should be taken when extrapolating these results to high Reynolds numbers. The statistical analysis of the dissipation and enstrophy fields induced by a set of Burger vortices in He et al. [319] is very eloquent in that respect. For this model system, finite-range scaling exponents for and
are different but the asymptotic scaling exponents can be shown to be equal in the limit of infinite Reynolds number.
B. Application of the 2D WTMM Method to 2D Cuts of a Turbulent 3D Dissipation Field In this section, we want to revisit previous multifractal analysis of the dissipation field ðrÞ in isotropic turbulence using the 2D WTMM methodology described in Section II.C. Given the uncertain nature of the existing knowledge, it is important to study the scaling properties of both the dissipation and enstrophy fields without resorting to the artifacts mentioned in Section V.A. We thus employ the numerical data from DNS of isotropic turbulence carried out by Meneguzzi [334] with the same numerical code as previously developed by Vincent and Meneguzzi [231] but at a higher resolution. The DNS were performed using 5123 mesh points in a 3D periodic box and a viscosity of 5 104 . A statistically steady state was obtained by forcing low Fourier modes in a deterministic way. The Taylor microscale Reynolds number R ¼ 216 is close to the value attained in the DNS reported [274,282,326,327]. Here we will examine only one snapshot of both dissipation and enstrophy 3D spatial fields. Indeed we will mainly proceed to a comparative multifractal analysis of 2D cuts of both fields using classical box-counting techniques and the 2D WTMM method. The corresponding ðqÞ, ðqÞ, and f ð Þ, f ð Þ spectra will result from an annealed averaging over 512 (512 512) 2D cuts in a 5123 cube. Figure 31a is a picture of the fluctuations of the local dissipation as seen on an arbitrary 2D cut when using a 256 gray-level coding. The highly intermittent nature of these fluctuations is striking and further illustrated in Figure 31c on an arbitrary 1D spatial profile. We systematically follow the numerical implementation procedure described in Section II.C. We first wavelet transform the 512 (512 512) images of with the first-order
62
´ ODO ET AL. ARNE
Figure 31. 5123 DNS of the dissipation and enstrophy fields at R ¼ 216 [334]. Dissipation field: (a) 2D cut of ðrÞ; (b) 2D cut of ln ðrÞ; (c) 1D cut of ðrÞ. Enstrophy field: (d) 2D cut of ðrÞ; (e) 2D cut of ln ðrÞ; (f) 1D cut of ðrÞ. In (a), (b), (d), and (e), and are represented using a 256 gray scale coding from black (min) to white (max).
(n ¼ 1) and the third-order (n ¼ 3) radially symmetric analyzing wavelets defined in Figure 1. From the wavelet transform skeleton defined by the WTMMM, we compute the partition functions Z(q, a) from which we extract the ðqÞ and f ð Þ multifractal spectra. 1. Remark Let us point out that the WTMM definition of the WT ðqÞ spectrum [Eq. (22)] is slightly different from the one defined in Eq. (57) from the moments of the l pdf (< ql > l ðqÞ ) and from the ‘‘standard’’ boxcounting definition BC ðqÞ found in the literature, for example in Meneveau and Sreenivasan [227]:
MULTIFRACTAL IMAGE ANALYSIS
63
Figure 32. 2D wavelet transform analysis of the 2D cuts of the dissipation and enstrophy fields shown in Figure 31a and d, respectively. ðxÞ is the first-order radially symmetric analyzing wavelet shown in Figure 1. Dissipation field: (a) a ¼ 22 W ; (b) a ¼ 24 W . Enstrophy field: (c) a ¼ 22 W ; (d) a ¼ 24 W . The local maxima of M along the maxima chains are indicated by () from which originates an arrow whose length is proportional to M and whose direction (with respect to the x-axis) is given by A .
WT ðqÞ ¼ ðqÞ d ¼ BC ðqÞ dq ¼ ðq 1ÞDq dq
ð71Þ
where d ¼ 2 when investigating 2D cuts of the 3D dissipation field and Dq are the generalized fractal dimensions defined in Refs. [20–26]. Note that the Legendre transforms used in the three different cases lead to the same estimate of the f ð Þ singularity spectrum. 2. Numerical Computation of the ðqÞ and f ðqÞ Multifractal Spectra Figure 32a and b illustrates the computation of both the maxima chains and the WTMMM of the 2D cut of shown in Figure 31a when using the firstorder analyzing wavelet at two different scales. After linking these WTMMM across scales, one constructs the WT skeleton from which one computes the partition functions Z(q, a) [Eq. (21)]. As shown in Figure 33a, the annealed average of a2 Zðq; aÞðÞ displays some well-defined scaling 4 behavior over the range of scales 2 W < a< 2 W (where W is the characteristic size of at the smallest scale), when plotted versus a in a logarithmic representation and this for values of q in the interval [2; 4] for
64
´ ODO ET AL. ARNE
Figure 33. Determination of the ðqÞ and f ð Þ spectra of 512 2D cuts of the dissipation field. The 2D WTMM method is used with either a first-order () or a third-order ( ) radially symmetric analyzing wavelet (see Fig. 1). Results obtained with box-counting techniques () are shown for comparison. (a) log2 ½a2 Zðq; aÞ vs. log2 a; (b) ðqÞ vs. q; (c) f ð Þ vs. , after Legendre transforming the ðqÞ curve in (b). In (a) the different data curves have been arbitrarily vertically shifted for the sake of clarity. In (b) and (c), the solid lines correspond, respectively, to the theoretical log-normal multifractal spectra (72) and (74) for the parameter values C1 ¼ 0:11 and C2 ¼ 0:18 [Eq. (73)]. In (c), the dashed line corresponds to the average f ð Þ spectrum obtained by Meneveau and Sreenivasan from the analysis of surrogate dissipation data using a box-counting algorithm [227].
which statistical convergence turns out to be achieved. Indeed some curvature can be observed in this logarithmic representation as the indication of some scale symmetry breaking as previously observed for the longitudinal velocity [81–83,85,221,246–258,267]. The extension of this statistical analysis to time averaging over a few turnover times is currently under progress. When processing to a linear regression fit of the data in Figure 33 over the range 21:1 W a 23:6 W , one gets the ðqÞ spectrum () shown in Figure 33b. In this accessible range of q values, the ðqÞ spectrum obtained unambiguously deviates from a monofractal linear spectrum. Actually the data are remarkably well fitted by a parabola, the hallmark of log-normal multifractal spectra:
MULTIFRACTAL IMAGE ANALYSIS
ðqÞ ¼ C1q C2
q2 2
65 ð72Þ
with C1 ¼ 0:11 0:01;
C2 ¼ 0:18 0:01
ð73Þ
By Legendre transforming this quadratic ðqÞ spectrum, one gets data for the f ð Þ singularity spectrum that are well parameterized by the corresponding parabolic log-normal singularity spectrum: f ð Þ ¼ 2
ð 1 þ C1 Þ2 2C2
ð74Þ
We have checked that the estimate of the f ð Þ singularity spectrum from the scaling behavior of the partition functions ðq; aÞ ¼ hðq; aÞ þ 1 [Eq. (29)] and f ðq; aÞ ¼ Dðq; aÞ [Eq. (30)] yields similar quantitative results. Figure 33 also shows for comparison the results ( ) obtained when applying the 2D WTMM method with a third-order (n ¼ 3) radially symmetric analyzing wavelet (the smoothing function being the isotropic 2D Mexican hat). An overall comparison with the previous results shows a remarkable robustness of the estimates of the ðqÞ and f ðqÞ spectra with respect to the order of the analyzing wavelet. Both spectra are still strikingly well fitted by the lognormal multifractal model predictions [Eqs. (72) and (74), respectively] with the parameter values C1 ¼ 0:07 0:01;
C2 ¼ 0:19 0:01
ð75Þ
which, up to the numerical uncertainty, are quite consistent with the previous values in Eq. (73). Figure 33 also shows for comparison the results (~) obtained when using classical box-counting techniques (indeed we use boxes with Gaussian shape in order to take advantage of part of our 2D WT software). It is clear in Figure 33a that the data obtained for a2 Zðq; aÞ with the box-counting method significantly differ from those obtained with the 2D WTMM methodology. Actually, as reported in Figure 33b, the ðqÞ data are still reasonably well accounted by the theoretical log-normal spectrum [Eq. (72)], but with significantly different parameter values: C1 ¼ 0:09 0:01;
C2 ¼ 0:20 0:01
ð76Þ
Note that the difference is not so much in the intermittency coefficient C2 , which is found to be robust to the method used to estimate it and in good agreement with the results of previous DNS studies [231,274,282,313, 326,327]. Let us emphasize that the C2 values in Eqs. (73), (75), and (76) are at the lower bound of the range of values (0.20 to 0.28) found in
66
´ ODO ET AL. ARNE
experimental measurements based on surrogate dissipation data [227,277, 303–305,323,324]. The main difference between the 2D WTMM and the box-counting results concerns the estimate of the coefficient C1 of the linear term in ðqÞ. This is a direct consequence of the normalization constraint ð1Þ ¼ 0 intrinsic to the box-counting method, which implies the relationship C1 ¼ C2 =2 between the two parameters of the log-normal ðqÞ spectrum [Eq. (72)]. The results reported in Figure 33 dramatically reveal the failure of commonly used box-counting algorithms when the considered measure results from a nonconservative log-normal multifractal process as characterized by a negative cancellation exponent [335–337] ð1Þ ¼ 0:20 0:01, the signature of a signed measure (i.e., a distribution that varies in sign on small scales). As a consequence, the f ð Þ spectrum is misleadingly shifted to the right when using box-counting techniques as illustrated in Figure 33c (this shift is also present when studying 1D cuts of the dissipation field as reported in Roux’s thesis [221]). This observation seriously questions the validity of most of the experimental and numerical box-counting estimates of the f ð Þ singularity spectrum reported so far in the literature. Figure 33c shows for comparison some average f ð Þ spectrum obtained by Meneveau and Sreenivasan [227] from the analysis of surrogate dissipation data; the agreement with our box-counting estimate is very good for the left increasing (q > 0) branch while the right decreasing (q < 0) branch departs somehow to a larger value of as an indication of a slightly larger intermittency coefficient C2 ¼ 0:25 as compared to the value C2 ¼ 0:20 in Eq. (76). This is an additional indication that surrogate dissipation is likely to be more intermittent than real dissipation [282]. 3. WTMMM Probability Density Functions This subsection is mainly devoted to the analysis of the joint probability density function Pa ðM; AÞ as computed from the WT skeletons of 512 2D cuts of the dissipation field with the first-order radially symmetric analyzing R wavelet ðn ¼ 1Þ. Figure 34a and b show the pdfs P ðMÞ ¼ dAP a a ðM; AÞ R and Pa ðAÞ ¼ dMPa ðM; AÞ for three different values of the scale parameter a in the scaling range. First let us focus on the results shown in Figure 34b for Pa ðAÞ. Pa ðAÞ does not evolve accross scales and is almost flat. Actually some oscillations are observed with maxima for A ¼ 0; =2; and 3=2 as an indication of some anisotropy induced by the cubic lattice discretization in the DNS. All the multifractal properties of 2D cuts are thus contained in the way the shape of Pa ðMÞ evolves when one decreases the scale parameter a as shown in Figure 34a. Actually, for the three selected scales, all the data points fall, within a good approximation, on a log-normal curve (see for comparison the pdfs in Fig. 28a), which is a strong indication
MULTIFRACTAL IMAGE ANALYSIS
67
Figure 34. Pdfs of the WTMMM coefficients of 512 2D cuts of the dissipation and enstrophy fields as computed with the first-order radially symmetric analyzing wavelet. Dissipation field: (a) Pa ðMÞ vs. M; (b) Pa ðAÞ vs. A. Enstrophy field: (c) Pa ðMÞ vs. M; (d) Pa ðAÞ vs. A. The symbols correspond to the following scales a ¼ 21 W ðÞ; 22 W ð Þ, and 23 W ðÞ.
that the WTMMM have a log-normal distribution in the inertial range. This observation of log-normal statistics strengthens the previous estimates of log-normal quadratic ðqÞ and f ð Þ spectra. (We refer the reader to Wang et al. [282] for similar conclusions on the entire 3D dissipation field when using box-counting techniques.) 4. Space-Scale Correlation Function Analysis As pointed out in Section III.C, to go from log-normal diagnosis to the demonstration of the existence of an underlying multiplicative structure in the 2D fluctuations of the dissipation field, one can take advantage of the space-scale unfolding provided by the WT skeleton to compute the crossscale correlation functions. Figure 35a shows the results of the computation of C ðx; a1 ; a2 Þ when averaging over the 512 2D cuts of . One can see that for x > supða1 ; a2 Þ, all the data points fall onto a unique curve when plotted versus log2 ðxÞ, independently of the considered pair of scales
68
´ ODO ET AL. ARNE
Figure 35. Magnitude correlation function Cðx; a1 ; a2 Þ [Eq. (49)] vs. log2 ðxÞ, as computed from the WT skeleton of 512 images. The analyzing wavelet is the radially symmetric first-order wavelet shown in Figure 1. The symbols have the following meaning: a1 ¼ 1; a2 ¼ 2ð Þ; a1 ¼ 2; a2 ¼ 3ðhÞ; and a1 ¼ 1; a2 ¼ 3ð4Þ in W units. (a) Dissipation field ; the solid line represents the theoretical prediction given by Eq. (50) with 2 ¼ C2 ln 2 ¼ 0:12 [C2 ¼ 0:18 as given by Eq. (73)]. (b) Enstrophy field ; the solid line has the same meaning as in (a); the dashed line represents the theoretical curve given by Eq. (50) when fixing 2 ¼ C2 ln 2 ¼ 0:20 [C2 ¼ 0:29 as given by Eq. (78)].
ða1 ; a2 Þ. Moreover this curve is in remarkable agreement with the theoretical prediction [Eq. (50)] for the random cascading process when plugging into this equation the value 2 ¼ C2 ln 2 ¼ 0:12 as previously estimated in Eq. (73). These consistent observations strongly suggest that a 2D nonconservative log-normal cascading process provides a reasonable model for the intermittent fluctuations observed along 2D cuts of the dissipation field. C. Application of the 2D WTMM Method to 2D Cuts of a Turbulent 3D Enstrophy Field Let us now proceed to a comparative statistical analysis of the corresponding numerical enstrophy field ðrÞ. Figure 31d and illustrates a 2D cut of
ðrÞ in linear and in semilogarithmic representations, respectively. The intermittent aspect of ðrÞ is enlightened on the 1D cut shown in Figure 31f. We proceed, as for the dissipation field in Section V.B, by applying the 2D WTMM method described in Section II.C to 512 (512 512) images of
with analyzing wavelets of different orders and we compare the ðqÞ and f ð Þ multifractal spectra obtained with the corresponding estimates from box-counting computations. 1. Numerical Computation of the Multifractal ðqÞ and f ðqÞ Spectra Figure 32c and d illustrates the maxima chains and the WTMMM of the 2D cut of shown in Figure 31d as computed with the first-order (n ¼ 1)
MULTIFRACTAL IMAGE ANALYSIS
69
Figure 36. Determination of the ðqÞ and f ð Þ spectra of 512 2D cuts of the enstrophy field. The 2D WTMM method is used with either a first-order () or a third-order ( ) radially symmetric analyzing wavelet (see Fig. 1). Results obtained with box-counting techniques (4) are shown for comparison. (a) log2 ½a2 Zðq; aÞ vs. log2 a; (b) ðqÞ vs. q; (c) f ð Þ vs. , after Legendre transforming the ðqÞ curve in (b). In (a) the different data curves have been arbitrarily vertically shifted for the sake of clarity. In (b) and (c), the solid lines correspond to the theoretical log-normal multifractal spectra (77) and (79) for the parameter values C1 ¼ 0:19 and C2 ¼ 0:29 [Eq. (78)].
analyzing wavelet (Fig. 1) at two different scales. After linking these WTMMM across scales, one constructs the WT skeleton from which one computes the partition functions Z(q, a) [Eq. (21)]. As shown in Figure 36a, the annealed average of a2 Zðq; aÞðÞ displays some well-defined scaling behavior 4 over the range of scales 2 W < a< q< 2 W for 2 < 4. Indeed some slight but systematic curvature can be noticed in the log–log plots very much like what has been observed for the dissipation in Figure 33a. If we proceed as in Section V.B to a linear regression fit of the data over the range 21:0 W a 23:4 W , one gets the ðqÞ spectrum () shown in Figure 36b, which is again in quite good agreement with a parabolic log-normal spectrum: ðqÞ ¼ C1 q C2
q2 2
ð77Þ
70
´ ODO ET AL. ARNE
with C1 ¼ 0:19 0:01;
C2 ¼ 0:29 0:01
ð78Þ
Consistently, we find in Figure 36c that the corresponding f ð Þ singularity spectrum is remarkably well fitted by the parabolic log-normal curve: f ð Þ ¼ 2
ð 1 þ C1 Þ2 2C2
ð79Þ
Figure 36b and c also show for comparison the results ( ) obtained when using the third-order ðn ¼ 3Þ analyzing wavelet. The estimates of the ðqÞ and f ð Þ spectra are in very good agreement with the results obtained previously with the first-order ðn ¼ 1Þ analyzing wavelet. These spectra are still remarkably approximated by a parabola [Eqs. (77) and (79)] with the following parameter values: C1 ¼ 0:18 0:01;
C2 ¼ 0:28 0:01
ð80Þ
which are within the error bars of the values reported in Eq. (78). The robustness of these multifractal spectra estimates with respect to some change in the shape of the analyzing wavelet is even more striking when one compares these estimates to those extracted from a box-counting algorithm (4). Very much like what we have observed for , this standard technique also yields parabolic spectra but with significantly different parameter values (mainly for C1 ): C1 ¼ 0:13 0:02;
C2 ¼ 0:29 0:01
ð81Þ
because of the normalization requirement ð1Þ ¼ 0, i.e., C1 ¼ C2 =2, inherent to this method. Let us point out that whatever the technique, the estimate of the intermittency parameter C2 of the enstrophy [Eqs. (78), (80), (81)] is much larger than the corresponding value found for the dissipation [Eqs. (73), (75), (76)]. This confirms that the enstrophy field is likely to be more intermittent than the dissipation field as previously suggested [321–327]. However the WTMM method reveals that the f ð Þ spectrum is noticeably shifted toward smaller values (corresponding to stronger singularities) as compared to the box-counting estimate (Fig. 36c). We will come back to this point as well as to the possible nonconservative nature ½ ð1Þ ’ 0:34 < 0 of the underlying log-normal multiplicative structure. 2. WTMMM Probability Density Functions The pdfs Pa ðMÞ and Pa ðAÞ of the WTMMM modulus and argument of the enstrophy field ðrÞ are shown in Figure 34c and d, respectively.
MULTIFRACTAL IMAGE ANALYSIS
71
Quantitatively one recovers similar results as previously observed for (r) (Fig. 34a and b). In Figure 34d, Pa ðAÞ is rather flat (with some small amplitude oscillations induced by the cubic lattice discretization in the DNS) and does not evolve across scales. When looking at Pa ðMÞ in Figure 34c, one sees that at each scale, the data points fall on a curve, which is well approximated by a log-normal pdf and which evolves across scales as governed by the log-normal ðqÞ spectrum computed just above [Eqs. (77) and (78)]. 3. Space-Scale Correlation Function Analysis In Figure 35b the results of the computation of the magnitude correlation function C ðx; a1 ; a2 Þ when averaging over the 512 2D cuts of (r) are reported. One can see that consistently with a multiplicative cascade structure, all the data points fall onto a unique curve when plotted versus log2 ðxÞ, for x > supða1 ; a2 Þ, and this independently of the considered pair of scales ða1 ; a2 Þ. As far as the pertinence of Eq. (50) for modeling the numerical data, it seems that when plugging this theoretical prediction into the parameter value 2 ¼ C2 ln 2 ¼ 0:20, according to the previous estimate of the intermittency parameter C2 in Eq. (78), one gets a poorer agreement than when comparing it with the theoretical curve predicted for the dissipation field ð 2 ¼ 0:12; C2 ¼ 0:18Þ. Actually, if one focuses on spatial distances x that are not too large (i.e., smaller than the integral scale L), for which the linear term 2 log2 ðL=xÞ becomes dominant in Eq. (50), then the observed slope of C ðx; a1 ; a2 Þ is quite in agreement with the expected value 2 ¼ C2 ln 2 ¼ 0:20. The results in Figure 35b are thus an additional indication that a 2D nonconservative log-normal multiplicative process can be used to model the intermittent fluctuations observed in 2D cuts of the enstrophy field. D. Discussion We have used the 2D WTMM method to characterize statistically the multifractal properties of 2D cuts of both the dissipation and the enstrophy fields issued from ð512Þ3 DNS at R ¼ 216 [334]. As a general result, we find that the intermittent nature of the corresponding spatial landscape can be well modeled by a 2D nonconservative log-normal multiplicative process. To some extent this result is not so surprising since it is most likely that dissipation and enstrophy are not conserved along 2D cuts. We hope that the generalization in 3D of the WTMM method will allow us to decide whether this nonconservativity is a 2D cut effect that is likely to disappear when increasing the Reynolds number or if it is an intrinsic property of the
72
´ ODO ET AL. ARNE
underlying 3D multiplicative spatial structures of both fields. Moreover the averaging over several turnover times will allow us to investigate larger values of jqj (i.e., higher order moments) and possibly to evidence some departure from the theoretical log-normal multifractal spectra as suggested in Refs. [274,314,315] as an indication of some local anisotropy induced by strongly localized events (e.g., vorticity filaments). Work in this direction is in progress. One of the main disturbing results reported in this section is the numerical demonstration that most of the numerical and experimental estimates of the multifractal spectra of and fields previously reported in the literature are strongly biased by the normalization constraint ðq ¼ 1Þ ¼ 0 inherent to the commonly used box-counting techniques, which turn out to be quite inappropriate to study nonconservative multiplicative cascading processes. These techniques yield f; ð Þ spectra, which have almost the right width as given by the intermittency exponent C2 but which are significantly shifted to the right (i.e., to larger values corresponding to weakest singularities) with an estimate of the most frequent singularity hðq ¼ 0Þ ¼ ðq ¼ 0Þ 1 ¼ C1 , which is misleadingly found positive instead of negative as revealed by our 2D WTMM analysis (Figs. 33c and 36c). Finally, our comparative 2D WTMM multifractal analysis of the dissipation and enstrophy fields shows an unambiguous quantitative difference between the f ð Þ and f ð Þ singularity spectra. The width of the later is significantly larger than the one of the former as given by the respective values of the intermittency parameter: C2 ¼ 0:29 0:01 > C2 ¼ 0:19 0:01. Moreover f ð Þ is maximum for
ðq ¼ 0Þ ¼ h ðq ¼ 0Þ þ 1 ¼ 1 C1 ’ 0:80, a value that is smaller than ðq ¼ 0Þ ¼ h ðq ¼ 0Þ þ 1 ¼ 1 C1 ’ 0:90 for which f ð Þ is maximum. These results demonstrate that the enstrophy spatial landscape is more intermittent than the dissipation spatial landscape in the sense that the support of its singularity exponent is wider and that it reaches smaller values of h ¼ 1 corresponding to stronger singularities. Note that for both fields the maximum of the f; ð Þ curves is equal to 2 ½ ðq ¼ 0Þ ¼ ðq ¼ 0Þ ¼ 0, which means that the corresponding 2D spatial landscapes are singular everywhere. These results confirm the conclusions of preliminary comparative box-counting studies of the dissipation and enstrophy fields [321–327]. We hope to extend this 2D WTMM analysis to the current highest accessible Reynolds number DNS with the specific goal of investigating the validity of several theoretical studies [265,266,318–320] that predict the asymptotic ðR ! þ1Þ equality of the multifractal spectra of both fields, namely ðqÞ ¼ ðqÞ and f ð Þ ¼ f ð Þ.
MULTIFRACTAL IMAGE ANALYSIS
73
VI. Multifractal Analysis of Digitized Mammograms Breast cancer, the most common cancer among women in western countries, has become a major problem of public health. Statistics indicate that in the United States approximately 1 in 10 women will develop breast cancer during her lifetime [338]. Each year, breast cancer kills about 10,000 women in France (120,000 in the world); it is still the leading cause of cancer-related death in women. It is a slowly evolving dicease; the average duration of tumor growth to obtain a palpable mass is about 10 to 15 years. Mammography (X-ray examination) is widely regarded as the most effective method for early detection of breast cancer. In the past 20 years, several national mass screening mammography programs [e.g., Health Insurance Plan of Greatest New-York (1982) and the Swedish 2-county Program of Mammography Screening for Breast Cancer (1992)] have shown that early diagnosis can significantly decrease breast cancer mortality about 23 to 31% in women aged 49 to 69 (see Dilhuydy and Barreau [339] for a complete discussion of the pros and cons of mass mammography). Because no way to prevent breast cancer (as opposed to lung cancer, for example) has been found so far, mammography actually plays a vital role in diagnosis of the decease as well as pretherapeutic management and control during and after treatment, whereas MRI (magnetic resonance imaging) and echography are helpful only when the mammogram is questionable. However, the radiological interpretation of mammograms is a rather difficult task since the mammographic appearance of normal tissue is highly variable. In the context of breast cancer screening, abnormalities have to be detected at an early stage in a large number of asymptomatic women. For this reason, independent reading of screening mammograms by two expert radiologists is required to reduce the number of interpretation errors. In spite of this, about 10 % to 30 % of cancers that could have been detected are missed and a high percentage of patients called back at screening turn out not to have cancer. Recently, much research has been devoted to developing reliable computer-aided diagnosis (CAD) methods (see Doi et al. [340] for a general review). Many of these methods are based on multiresolution analysis [341–344], difference image technique and global and local thresholding [345–349], statistical approaches [350–354], neural networks [355–360], fuzzy logic [361–363], and the wavelet transform (WT) and related techniques [342–344,360,362,364–371]. Currently most of these methods are often combined to detect and classify clusters of microcalcifications (MC), which are an important mammographic sign of early (in situ) breast cancer despite the fact that several benign diseases show MC as well [347,353,354,359,
74
´ ODO ET AL. ARNE
360,362,366,367,369–373]. In the mid-1990s, fractal methods were applied to the analysis of radiographic images with some success in improving the performances of previous CAD schemes [352,374–379]. But most of these methods have been intrinsically elaborated based on the prerequesite that the background roughness fluctuations of normal breast texture are statistically homogeneous (i.e., monofractal) and uncorrelated. Regions that contain statistical aberrations that deviate from this monofractal picture are considered as abnormal regions in which tumors or MC are likely to be found. Our goal here is to propose the 2D WTMM method as an alternative method to perform multifractal analysis of digitized mammograms [115]. As we want to study scaling properties of digitized mammograms, we chose to use full-breast images from the Digital Database for Screening Mammography (DDSM) project [380], which provides online more than 2600 studies1 sorted into three categories: normal, cancer, and benign. Mammograms were digitized using a 12-bit scanner with both a good spatial resolution of 43:5 m. Full-breast images enable us to select about 50 overlapping 512 512 pixel squares; indeed, to master edge effects, only cores of the images were used for the computation of the WT skeleton and partition functions.
A. Application of the 2D WTMM Method to Mammographic Tissue Classification: Dense and Fatty Tissues Several studies in the mid-1970s showed that an association existed between mammographic parenchymal patterns and the risk of developing breast cancer [381–383]. However, it appears that very few image processing works [363] have been devoted to automatic breast tissue density measurement, since Boyd et al. [383] studied the relation between mammographic densities and breast cancer risk using both radiological classification and semiautomatic user-assisted computer measurement based on gray-level histogram thresholding. Here we analyze normal mammary parenchyma with our multifractal 2D WTMM method with the specific goal of proposing a computerized method to calculate a breast density fluctuations index. We have selected a set of 10 images in the DDSM database according to ACR breast density rating with some index ranking from 1 to 4, as assigned by an experienced mammographer: five fatty (rated 1 on ACR density scale) and five dense (rated 4) breasts. The main steps of the 2D WTMM computations are illustrated in Figure 37 on two full-breast images selected, 1
http://marathon.csee.usf.edu/Mammography/Database.html.
MULTIFRACTAL IMAGE ANALYSIS
75
respectively, to be representative of dense-glandular and fatty breasts. Figure 37a and e shows the original images, respectively, with the (superimposed) grid used to cut out 49 (512 512 pixels) subscenes. Figure 37b and f represents a zoom in the respective central subscenes. The corresponding WT modulus landscape and WT maxima chains computed
Figure 37. 2D wavelet transform analysis of two mammograms: (a–d) dense breast tissue and (e–h) fatty breast tissue. The analyzing wavelet is the first-order isotropic wavelet ( is the isotropic gaussian function) shown in Figure 1. (a) and (e) are the two full breast images. (b) and (f) represent some zooming in the central part of the two original images. (c) and (g) show the WT modulus at the scale a ¼ 3 W with the same gray-level coding as in Figure 3c; the maxima chains are shown for comparison. In (d) and (h) only the maxima chains and the local maxima of M along these chains are represented () at the scale a ¼ 2:5 W .
76
´ ODO ET AL. ARNE
at the scale a ¼ 39 pixels are shown in Figure 37c and g, respectively. Figure 37d and h represents, at a smaller scale, the location of the WTMMM () from which originate an arrow that represents the WT vector T [f] (b, a). Figure 38 shows the results of the computation of the partition functions Z (q, a) [Eq. (21)], h (q, a) [Eq. (29)], and D (q, a) [Eq. (30)] obtained when averaging over 49 nonoverlapping (512 512) images cut out of the original dense and fatty mammograms. As shown in Figure 38a and b, both dense and fatty tissues display rather good scaling properties over two and a half octaves. The scaling actually deteriorates progressively when considering large scales, due to finite size effects. When proceeding to a linear regression fit of log2 [Z (q, a)] vs. log2 ðaÞ over the range of scales extending from amin ¼ 1:6 W to amax ¼ 4 W , one obtains the ðqÞ spectra reported in Figure 38c. From a simple visual inspection, one realizes that dense and fatty breast tissues display quite different scaling properties. The latter presents a ðqÞ spectrum, which is remarkably linear in the range q 2 ½3; 3 with a slope H ¼ 0:25 0:05, while the former presents a larger slope H ¼ 0:65 0:05 with some possible nonlinear departure, which might indicate multifractality. This monofractal vs. multifractal discrimination between fatty and dense breast tissues is also evidenced by the computation of the corresponding D(h) singularity spectra in Figure 38d. However, the multifractal diagnosis for dense tissues requires further numerical analysis to ensure statistical convergence of the ðqÞ exponents for large values of jqj. Nevertheless, what seems to be robust, considering the whole set of processed images, is the fact that fatty tissues display monofractal scaling behavior with a Hurst exponent H taking a value in the range [0.20,0.35] as an indication of antipersistent roughness fluctuations while dense tissues display (possibly multifractal) scaling with H 2 [0.55, 0.75] as an indication of persistent long-range correlations. Furthermore, in the most general case, we have shown that in any full-breast mammogram, those two kinds of tissue are present and only those two. In particular, one can assign a color (e.g., blue or red) to each square of the working grid according to its dense or fatty area identification. Work is in progress to make this segmentation independent of the square grid used to cut out subscenes. Finally, let us note that in previous work, Heine et al. [379,384] already used self-similarity (fractal) analysis to study mammographic density, using the Fourier power spectrum method to extract the scaling exponent ¼ 2H þ 2 [Eq. (41)]. They obtained a histogram of values with an average H of 0.469 and a rather small standard deviation of 0.045. This finding may be interpreted in light of our results. Indeed we may think that most of the images analyzed by Heine et al. Clearly contain both fatty (H 2 [0.20, 0.35]) and dense (H 2 [0.55, 0.75]) areas, so that the power spectrum exponent is an average of two dictinct behaviors.
MULTIFRACTAL IMAGE ANALYSIS
77
Figure 38. Determination of the ðqÞ and D(h) spectra of dense () and fatty ( ) breasts with the 2D WTMM method. (a) log2 z(q, a) vs. log2a. (b) h(q, a) vs. log2 a. (c) ðqÞ vs. q. (d) D(h) vs. h obtained from Eqs. (31) and (32). Same analyzing wavelet as in Figure 37. These results correspond to annealed averaging over 49 (512 512) squares cut out of full-breast images. a is expressed in W units. In (a) and (b), q goes from 1 to 3 from bottom to top.
B. Detecting Microcalcifications through WT Skeleton Segmentation The presence of clustered MC is one of the most important and sometimes the only sign of cancer in a mammogram. As a potential computer-aided diagnostic tool, let us show how our WT methodology can identify MC that are small calcium deposits in tissue, appearing as clusters of bright spots. Figure 39 illustrates how one can actually detect MC by inspecting the WT maxima chains. Indeed, at the smallest scale resolved by our WT microscope ( W ¼ 13 pixels), MC, which can be considered as strong singularities, are
78
´ ODO ET AL. ARNE
Figure 39. Detection and characterization of microcalcifications. (a) Original 726 726 image of dense breast tissue containing MC. (b) Scaling behavior of the WT modulus M along some maxima lines pointing toward dense tissue background ( ) and microcalcifications (h). The solid (respectively dashed) straight line corresponds to the slope h ¼ 0:65 (respectively 1) characteristic of background tissue roughness fluctuations (respectively MC). (c) and (d) show the maxima chains obtained after eliminating background tissue maxima chains at scales a ¼ W (c) and 2:5 W (d), when using the WT skeleton space-scale information.
contour shaped by some maxima chains. Because the average size of MC is about 200 m (5 pixels), these singularities are seen by our mathematical microscope as Dirac singularities; thus the corresponding maxima lines pointing to the MC are likely to display scaling properties with a local Ho¨lder exponent h ¼ 1ðM ½ f a1 Þ down to scales of the order of the MC size where one should observe a crossover to the value h ¼ 0ðM ½ f cst Þ as an indication of the discontinuity induced by the MC boundary. The behavior of the WT modulus along several maxima lines pointing to background points and to MC is illustrated in Figure 39b. One can thus classify these lines according to the behavior of M ½ f along these lines, and then separate MC (h 1) from dense background tissue (h 0:65 0:05) as experienced on synthetic images in Section III.D. Figure 39c and d shows the maxima chains that are found to correspond to MC at two different scales. We see that these maxima chains can be used not only to detect MC at the smallest resolved scale (Fig. 39c), but also to perform MC clustering when investigating largest scales (Fig. 39d).
MULTIFRACTAL IMAGE ANALYSIS
79
As pointed out in Section III.D, the MC WT subskeleton can be used to compute the corresponding partition functions, and thus to fully characterize the fractal geometry of the MC cluster. Figure 40 shows the results of the computation of the partition functions from the subskeleton of WT maxima lines pointing toward MC (h 1). Let us recall that in this case h(q, a) (Fig. 40a) is simply the average scaling behavior (see Fig. 39b) along all the maxima lines of this subskeleton. As expected, one retrieves a crossover between small scale scaling properties (h 0 induced by the MC boundaries) and larger scale (2 W a 3:7 W ) scaling properties
Figure 40. Determination of the ðqÞ and D(h) spectra of the MC cluster shown in Figure 39a. The partition functions are computed from the MC WT skeleton obtained after eliminating the background tissue maxima lines (see Figure 39). (a) h(q, a) vs. log2 a. (b) D(q, a) vs. log2 a. (c) ðqÞ vs. q. (d) D(h) vs. h obtained from Eqs. (31) and (32). Same analyzing wavelet as in Figure 37. a is expressed in W units. The solid line in (a) corresponds to the slope h ¼ 0:45 and in (b) to DF ¼ 1:2.
80
´ ODO ET AL. ARNE
(h 0:4) since maxima lines, pointing toward MC, have not all reached the asymptotic (h ¼ 1) Dirac singularities behavior because of finite size effects. In the same finite range of scales, D(q, a) (Fig. 40b) displays good scaling properties for q values between 1 and 3, which results, to a good approximation, in a D(h) singularity spectrum that reduces to a single point h 0:4 (Fig. 40d). This result is consistent with the slope of the corresponding ðqÞ spectrum (Fig. 40c) that is found to be linear up to numerical uncertainty. Moreover, from the slope of D(q ¼ 0; a) vs. log a, as well as from the estimate of ð0Þ ¼ DF , one can assign with no ambiguity the fractal dimension DF ¼ 1:2 0:05 to the MC cluster, which is definitely larger than 1 and smaller than 2, the hallmark of fractal geometry. We have also applied our methodology to a small number of benign and malign clusters; work is in progress to determine to what extent the fractal dimension of a MC cluster can be used as a discriminating index between a benign state and malignancy. We have presented a new space-scale methodology for studying, within the same algorithmic framework, background tissue properties and abnormal singularities associated with breast cancer. For its ability to reveal and distinguish persistent and nonpersistent long-range correlations, the 2D WTMM method looks very promising in classifying tissues by quantifying breast density fluctuations in a very accurate way. Furthermore, we plan to improve detection and segmentation of MC by mixing and combining the 2D WTMM method with neural network techniques to assist in diagnosis of digitized mammograms.
VII. Conclusion To summarize, we have presented a first step toward a statistical theory of multifractal images based on the wavelet theory. The 2D WTMM method [106–110] relies on the computation of partition functions from the WT skeleton defined by the wavelet transform modulus maxima. This skeleton provides an adaptative space-scale partition of the fractal distribution under study from which one can extract the ðqÞ and D(h) [or f ( )] multifractal spectra as the equivalent of thermodynamic functions. With some appropriate choice of the analyzing wavelet, we have shown that the WTMM method provides a natural but necessary generalization of the classical box-counting and structure function techniques that both have intrinsic and fundamental limitations. Indeed we believe that the 2D WTMM method for characterizing the roughness fluctuations of a fractal landscape, a rough surface, a turbulent flow, or the image of a fractal object
MULTIFRACTAL IMAGE ANALYSIS
81
is likely to become as useful as the well-known phase portrait reconstruction, Poincare´ section, and first return map techniques for the analysis of chaotic time series [385–388]. Besides the new concepts involved in this methodology and its potential theoretical interest, there is a more concrete and technical contribution [139] that is likely to have a strong impact on future research. For both image analysis [107,108,139] and image synthesis [109,139] purposes, we have implemented new algorithms and developed new software that can be routinely used to analyze as well as to model experimental data. In particular, some of these numerical tools take advantage of the space-scale information contained in the WT skeleton to go beyond the classical (one-point) multifractal description via the estimate of (two-point) space-scale correlation functions. Prior to experimental applications, all these numerical tools were calibrated via systematic test applications on random self-affine surfaces (e.g., isotropic fractional Brownian surfaces and anisotropic monofractal rough surfaces [108]) as well as on synthetic multifractal rough surfaces [109]. To illustrate the wide range of potential applications of this wavelet-based image processing method, we have reported the most significant results obtained when applying the 2D WTMM methodology to three rather different experimental situations, namely the statistical analysis of high-resolution satellite images of the cloud structure, of 2D cuts of the dissipation and enstrophy fields in 3D direct numerical simulations of homogeneous and isotropic trubulence, and of digitized mammograms. We are convinced that this methodology will lead to significant progress in the understanding of the multiscale mechanisms that underly the formation of rough surfaces and the spatiotemporal evolution of intermittent fields in various domains of fundamental as well as applied sciences such as erosion and corrosion processes, deposition and growth phenomena, catalysis, fracture propagation, turbulence, medical imaging, and many other areas in physics, astrophysics, chemistry, biology, geology, meteorology, and material sciences. Acknowledgments We are very grateful to E. Bacry, R. F. Cahalan, A. Davis, M. H. Dilhuydy, S. Jaffard, L. Lalonde, J. M. Lina, A. Marshak, J. F. Muzy, and P. SaintJean for very interesting and helpful discussions. We are very indebted to Y. Gagne, Y. Male´cot, and S. Ciliberto for permission to use their experimental turbulent signals and to M. Meneguzzi for allowing us access to his DNS numerical data. We want to acknowledge M. H. Dilhuydy and L. Lalonde for many helpful and illuminating conversations relevant to
82
´ ODO ET AL. ARNE
mammography and for providing additional training mammograms. The work concerning the analysis of Landsat satellite images of cloud structure was supported by NATO (Grant CRG 960176) and was performed while S. G. Roux held a National Research Council–NASA/GSFC Research Associateship. The work concerning the analysis of DNS turbulent dissipation and enstrophy fields is currently supported by the Centre National de la Recherche Scientifique under GDR ‘‘Turbulence.’’
References 1. B. B. Mandelbrot, Fractals: Form, Chance and Dimensions (Freeman, San Francisco, 1977). 2. B. B. Mandelbrot, The Fractal Geometry of Nature (Freeman, San Francisco, 1982). 3. Random Fluctuations and Pattern Growth, edited by H. E. Stanley and N. Ostrowski (Kluwer Academic, Dordrecht, 1988). 4. J. Feder, Fractals (Pergamon, New York, 1988). 5. T. Vicsek, Fractal Growth Phenomena (World Scientific, Singapore, 1989). 6. The Fractal Approach to Heterogeneous Chemistry: Surfaces, Colloids, Polymers, edited by D. Avnir (John Wiley & Sons, New York, 1989). 7. F. Family and T. Vicsek, Dynamics of Fractal Surfaces (World Scientific, Singapore, 1991). 8. Fractals and Disordered Systems, edited by A. Bunde and S. Havlin (Springer-Verlag, Berlin, 1991). 9. Fractals in Natural Science, edited by T. Vicsek, M. Schlesinger, and M. Matsushita (World Scientific, Singapore, 1994). 10. Fractals in Geoscience and Remote Sensing, Image Understanding Research Series, Vol. 1, ECSC-EC-EAEC, edited by G. G. Wilkinson, J. Kanellopoulos, and J. Megier (Brussels, Luxemburg, 1995). 11. A. L. Baraba´si and H. E. Stanley, Fractal Concepts in Surface Growth (Cambridge University Press, Cambridge, 1995). 12. Fractal Aspects of Materials, Material Research Society Symposium Proceeding, Vol. 367, edited by F. Family, P. Meakin, B. Sapoval, and R. Wool (Pittsburg, 1995). 13. B. Sapoval, Les Fractales (Aditech, Paris, 1988). 14. On Growth and Form: Fractal and Non-Fractal Patterns in Physics, edited by H. E. Stanley and N. Ostrowski (Martinus Nijhof, Dordrecht, 1986). 15. Fractals in Physics, edited by L. Pietronero and E. Tosatti (North-Holland, Amsterdam, 1986). 16. Fractals in Physics, Essays in honour of B. B. Mandelbrot, Physica D, Vol. 38, Fractals in Physics, Essays in honour of B. B. Mandelbrot, Physica D, Vol. 38, edited by A. Aharony and J. Feder (North-Holland, Amsterdam, 1989). 17. B. J. West, Fractal Physiology and Chaos in Medecine (World Scientific, Singapore, 1990). 18. U. Frisch, Turbulence (Cambridge University Press, Cambridge, 1995). 19. J. D. Farmer, E. Ott, and J. A. Yorke, Physica D 7, 153 (1983). 20. P. Grassberger and I. Procaccia, Phys. Rev. Lett. 50, 346 (1983). 21. P. Grassberger and I. Procaccia, Physica D 9, 189 (1983). 22. R. Badii and A. Politi, Phys. Rev. Lett. 52, 1661 (1984). 23. R. Badii and A. Politi, J. Stat. Phys. 40, 725 (1985). 24. P. Grassberger, R. Badii, and A. Politi, J. Stat. Phys. 51, 135 (1988).
MULTIFRACTAL IMAGE ANALYSIS
83
25. G. Grasseau, Ph.D. thesis, University of Bordeaux I, 1989. 26. F. Argoul, A. Arneodo, J. Elezgaray, G. Grasseau, and R. Murenzi, Phys. Rev. A 41, 5537 (1990). 27. L. V. Meisel, M. Johnson, and P. J. Cote, Phys. Rev. A 45, 6989 (1992). 28. The Science of Fractal Images, The Science of Fractal Images, edited by H. O. Peitgen and D. Saupe (Springer-Verlag, New York, 1987). 29. R. F. Voss, Physica D 38, 362 (1989). 30. G. A. Edgard, Measures, Topology and Fractal Geometry (Springer-Verlag, Berlin, 1990). 31. S. Davies and P. Hall, Technical Report No. SRR 96-008, School of Mathematical Sciences, National Australian University (1996). 32. B. Dubuc, J. F. Quiniou, C. Roques-Carmes, C. Tricot, and S. W. Zucker, Phys. Rev. A 39, 1500 (1989). 33. T. Higuchi, Physica D 46, 254 (1990). 34. N. P. Greis and H. P. Greenside, Phys. Rev. A 44, 2324 (1991). 35. W. Li, Int. J. Bifurcation Chaos. 1, 583 (1991). 36. J. Schmittbuhl, J. P. Violette, and S. Roux, Phys. Rev. E 51, 131 (1995). 37. A. Scotti, C. Meneveau, and S. G. Saddoughi, Phys. Rev. E 51, 5594 (1995). 38. B. Lea-Cox and J. S. Y. Wang, Fractals 1, 87 (1993). 39. C. K. Peng, S. V. Buldyrev, M. Simons, H. E. Stanley, and A. L. Goldberger, Phys. Rev. E 49, 1685 (1994). 40. M. S. Taqqu, V. Teverovsky, and W. Willinger, Fractals 3, 785 (1995). 41. A. R. Mehrabi, H. Rassamdana, and M. Sahimi, Phys. Rev. E 56, 712 (1997). 42. B. Pilgram and D. T. Kaplan, Physica D 114, 108 (1998). 43. G. Parisi and U. Frisch, in Turbulence and Predictability in Geophysical Fluid Dynamics and Climate Dynamics, Proc. of Int. School, edited by M. Ghil, R. Benzi, and G. Parisi (NorthHolland, Amsterdam, 1985), p. 84. 44. A. S. Monin and A. M. Yaglom, Statistical Fluid Mechanics (MIT Press, Cambridge, MA, 1975), Vol. 2. 45. A. L. Baraba´si and T. Vicsek, Phys. Rev. A 44, 2730 (1991). 46. A. L. Baraba´si, P. Sze´falusy, and T. Vicsek, Physica A 178, 17 (1991). 47. J. F. Muzy, E. Bacry, and A. Arneodo, Phys. Rev. Lett. 67, 3515 (1991). 48. J. F. Muzy, E. Bacry, and A. Arneodo, Int. J. Bifurcation Chaos 4, 245 (1994). 49. A. Arneodo, E. Bacry, and J. F. Muzy, Physica A 213, 232 (1995). 50. J. F. Muzy, E. Bacry, and A. Arneodo, Phys. Rev. E 47, 875 (1993). 51. A. Grossmann and J. Morlet, S.I.A.M.J. Math. Anal. 15, 723 (1984). 52. A. Grossmann and J. Morlet, in Mathematics and Physics, Lectures on Recent Results, edited by L. Streit (World Scientific, Singapore, 1985), p. 135. 53. P. Goupillaud, A. Grossmann, and J. Morlet, Geoexploration 23, 85 (1984). 54. Wavelets, Wavelets, edited by J. M. Combes, A. Grossmann, and P. Tchamitchian (Springer-Verlag, Berlin, 1989). 55. Y. Meyer, Ondelettes (Herman, Paris, 1990). 56. Les Ondelettes en 1989, Les Ondelettes en 1989, edited by P. G. Lemarie´ (Springer-Verlag, Berlin, 1990). 57. Wavelets and Applications, Wavelets and Applications, edited by Y. Meyer (Springer, Berlin, 1992). 58. I. Daubechies, Ten Lectures on Wavelets (S.I.A.M, Philadelphia, 1992). 59. Wavelets and Their Applications, Wavelets and Their Applications, edited by M. B. Ruskai, G. Beylkin, R. Coifman, I. Daubechies, S. Mallat, Y. Meyer, and L. Raphael (Jones and Barlett, Boston, 1992). 60. C. K. Chui, An Introduction to Wavelets (Academic Press, Boston, 1992).
84
´ ODO ET AL. ARNE
61. Progress in Wavelets Analysis and Applications, Progress in Wavelets Analysis and Applications, edited by Y. Meyer and S. Roques (Editions frontie`res, Gif-sur-Yvette, 1993). 62. A. Arneodo, F. Argoul, E. Bacry, J. Elezgaray, and J. F. Muzy, Ondelettes, Multifractales et Turbulences: de l’ADN aux croissances cristallines (Diderot Editeur, Art et Sciences, Paris, 1995). 63. Wavelets: Theory and Applications, Wavelets: Theory and Applications, edited by G. Erlebacher, M. Y. Hussaini, and L. M. Jameson (Oxford University Press, Oxford, 1996). 64. M. Holschneider, Wavelets: An Analysis Tool (Oxford University Press, Oxford, 1996). 65. S. Mallat, A Wavelet Tour in Signal Processing (Academic Press, New York, 1998). 66. B. Torresani, Analyse Continue par Ondelettes (Editions de Physique, Les Ulis, 1998). 67. T. C. Halsey, M. H. Jensen, L. P. Kadanoff, I. Procaccia, and B. I. Shraiman, Phys. Rev. A 33, 1141 (1986). 68. P. Collet, J. Lebowitz, and A. Porzio, J. Stat. Phys. 47, 609 (1987). 69. G. Paladin A. Vulpiani, Phys. Rep. 156, 148 (1987). 70. B. B. Mandelbrot, Fractals and Multifractals: Noise, Turbulence and Galaxies, Vol. 1 of Selecta (Springer-Verlag, Berlin, 1989). 71. D. Rand, Ergod. Th. Dyn. Sys. 9, 527 (1989). 72. S. Zhong and S. Mallat, IEEE Trans. Pattern Anal. Machine Intelligence 14, 710 (1992). 73. S. Hwang and W. L. Mallat, IEEE Trans. Inform. Theory 38, 617 (1992). 74. E. Bacry, J. F. Muzy, and A. Arneodo, J. Stat. Phys. 70, 635 (1993). 75. S. Jaffard, SIAM J. Math. Anal. 28, 944 (1997). 76. H. G. E. Hentschel, Phys. Rev. E 50, 243 (1994). 77. T. Bohr and T. Te`l, in Direction in Chaos, Vol. 2, edited by B. L. Hao (World Scientific, Singapore, 1988), p. 194. 78. S. F. Edwards and P. W. Anderson, J. Phys. F 5, 965 (1975). 79. A. Arneodo, in Ref. [63], p. 349. 80. J. F. Muzy, E. Bacry, and A. Arneodo, in Ref. [61], p. 323. 81. A. Arneodo, J. F. Muzy, and S. G. Roux, J. Phys. II France 7, 363 (1997). 82. A. Arneodo, S. Manneville, and J. F. Muzy, Eur. Phys. J. B 1, 129 (1998). 83. A. Arneodo, B. Audit, E. Bacry, S. Manneville, J. F. Muzy, and S. G. Roux, Physica A 254, 24 (1998). 84. S. G. Roux, J. F. Muzy, and A. Arneodo, Eur. Phys. J. B 8, 301 (1999). 85. A. Arneodo, S. Manneville, J. F. Muzy, and S. G. Roux, Phil. Trans. R. Soc. London A 357, 2415 (1999). 86. A. Arneodo, J. Delour, and J. F. Muzy, in Wavelet Applications in Signal and Image Processing VIII, edited by A. Aldroubi, A. F. Laine, and M. A. Unser p. 58 (2000). 87. J. Delour, J. F. Muzy, and A. Arneodo, Eur. Phys. J. B 23, 243 (2001). 88. A. Arneodo, F. Argoul, E. Bacry, J. F. Muzy, and M. Tabard, Phys. Rev. Lett. 68, 3456 (1992). 89. A. Arneodo, F. Argoul, J. F. Muzy, M. Tabard, and E. Bacry, Fractals. 1, 629 (1993). 90. A. Arneodo, F. Argoul, J. F. Muzy, and M Tabard, Phys. Lett. A 171, 31 (1992). 91. A. Arneodo, F. Argoul, J. F. Muzy, and M. Tabard, Physica A 188, 217 (1992). 92. A. Kuhn, F. Argoul, J. F. Muzy, and A. Arneodo, Phys. Rev. Lett. 73, 2998 (1994). 93. A. Arneodo, E. Bacry, P. V. Graves, and J. F. Muzy, Phys. Rev. Lett. 74, 3293 (1995). 94. A. Arneodo, Y. Daubenton-Carafa, E. Bacry, P. V. Graves, J. F. Muzy, and C. Thermes, Physica D 96, 291 (1996). 95. A. Arneodo, Y. Daubenton-Carafa, B. Audit, E. Bacry, J. F. Muzy, and C. Thermes, Eur. Phys. J. B 1, 259 (1998). 96. A. Arneodo, Y. Daubenton-Carafa, B. Audit, E. Bacry, J. F. Muzy, and C. Thermes, Physica A 249, 439 (1998).
MULTIFRACTAL IMAGE ANALYSIS
85
97. B. Audit, C. Thermes, C. Vaillant, Y. Daubenton-Carafa, J. F. Muzy, and A. Arneodo, Phys. Rev. Lett. 86, 2471 (2001). 98. B. Audit, C. Vaillant, A. Arneodo, Y. Daubenton-Carafa, and C. Thermes, J. Mol. Biol. 316, 903 (2002). 99. A. Arneodo, J. P. Bouchaud, R. Cont, J. F. Muzy, M. Potters, and D. Sornette, preprint cond-mat/9607120 at http://xxx.lanl.gov. 100. A. Arneodo, J. F. Muzy, and D. Sornette, Eur. Phys. J. B 2, 277 (1998). 101. A. Arneodo, E. Bacry, and J. F. Muzy, Phys. Rev. Lett. 74, 4823 (1995). 102. A. Arneodo, E. Bacry, S. Jaffard, and J. F. Muzy, J. Stat. Phys. 87, 179 (1997). 103. A. Arneodo, E. Bacry, S. Jaffard, and J. F. Muzy, J. Fourier Anal. Appl. 4, 159 (1998). 104. A. Arneodo, E. Bacry, S. Jaffard, and J. F. Muzy, CRM Proc. Lecture Notes. 18, 315 (1999). 105. J. C. Vassilicos and J. C. Hunt, Proc. R. Soc. London. 435, 505 (1991). 106. J. Arrault, A. Arneodo, A. Davis, and A. Marshak, Phys. Rev. Lett. 79, 75 (1997). 107. A. Arneodo, N. Decoster, and S. G. Roux, Phys. Rev. Lett. 83, 1255 (1999). 108. A. Arneodo, N. Decoster, and S. G. Roux, Eur. Phys. J. B 15, 567 (2000). 109. N. Decoster, S. G. Roux, and A. Arneodo, Eur. Phys. J. B 15, 739 (2000). 110. S. G. Roux, A. Arneodo, and N. Decoster, Eur. Phys. J. B 15, 765 (2000). 111. J. P. Antoine, P. Carette, R. Murenzi, and B. Piette, Signal Process. 31, 241 (1993). 112. E. Freysz, B. Pouligny, F. Argoul, and A. Arneodo, Phys. Rev. Lett. 64, 745 (1990). 113. A. Arneodo, F. Argoul, J. F. Muzy, B. Pouligny, and E. Freysz, in Ref. [59], p. 241. 114. J. Canny, IEEE Trans. Pattern Anal. Machine Intelligence. 8, 679 (1986). 115. P. Kestener, J. Lina, P. Saint-Jean, and A. Arneodo, Image Anal. Stereol. 20, 169 (2001). 116. D. Marr, Vision (W. H. Freemann and Co, San Francisco, 1982). 117. A. Rosenfeld M. Thurston, IEEE Trans. Comput. C 20, 562 (1971). 118. R. Murenzi, Ph.D. thesis, University of Louvain la Neuve, 1990. 119. R. Murenzi, in Ref. [54], p. 239. 120. D. Schertzer and S. Lovejoy, J. Geophys. Res. 92, 9693 (1987). 121. D. Schertzer and S. Lovejoy, Phys. Chem. Hyd. J. 6, 623 (1985). 122. S. Lovejoy and D. Schertzer, in Ref. [10], p. 102. 123. D. Schertzer, S. Lovejoy, F. Schmitt, Y. Ghigisinskaya, and D. Marsan, Fractals. 5, 427 (1997). 124. S. Jaffard and Y. Meyer, Memoirs A.M.S. 123, n.587 (1996). 125. M. Ben Slimane, Ph.D. thesis, E.N.P.C., France, 1996. 126. S. Jaffard, Pub. Math. 35, 155 (1991). 127. E. Bacry, A. Arneodo, U. Frisch, Y. Gagne, and E. Hopfinger, in Turbulence and Coherent Structures, edited by M. Lesieur and O. Metais (Kluwer, Dordrecht, 1991), p. 203. 128. M. Vergassola, R. Benzi, L. Biferale, and D. Pisarenko, J. Phys. A 26, 6093 (1993). 129. M. Vergassola and U. Frisch, Physica D 54, 58 (1991). 130. S. Jaffard, C. R. Acad. Sci. Paris, Serie I: Math. 326, 555 (1998). 131. R. Badii, Ph.D. thesis, University of Zurich, 1987. 132. P. Cvitanovic, in Proceedings Group Theoretical Methods in Physics, edited by R. Gilmore (World Scientific, Singapore, 1987). 133. M. J. Feigenbaum, M. H. Jensen, and I. Procaccia, Phys. Rev. Lett. 57, 1503 (1986). 134. M. H. Jensen, L. P. Kadanoff, and I. Procaccia, Phys. Rev. A 36, 1409 (1987). 135. A. B. Chhabra, R. V. Jensen, and K. R. Sreenivasan, Phys. Rev. A 40, 4593 (1989). 136. A. B. Chhabra and R. V. Jensen, Phys. Rev. Lett. 62, 1327 (1989). 137. A. B. Chhabra, C. Meneveau, R. V. Jensen, and K. R. Sreenivasan, Phys. Rev. A 40, 5284 (1989). 138. W. H. Press, B. P. Flannery, S. A. Teukolsky, and W. T. Vetterling, Numerical Recipes (Cambridge University Press, Cambridge, 1992).
86 139. 140. 141. 142. 143. 144. 145. 146. 147. 148. 149.
150. 151. 152. 153. 154. 155. 156. 157. 158. 159. 160. 161. 162. 163. 164. 165. 166. 167. 168. 169. 170. 171.
172. 173. 174. 175.
´ ODO ET AL. ARNE N. Decoster, Ph.D. thesis, University of Bordeaux I, 1999. B. B. Mandelbrot and J. W. Van Ness, S.I.A.M. Rev. 10, 422 (1968). J. Beran, Statistics for Long-Memory Process (Chapman & Hall, New York, 1994). G. Wornell and A. V. Oppenheim, IEEE Trans. Signal Proc. 40, 611 (1992). R. F. Peltier and J. Levy Ve´hel, INRIA Report No. 2396 (1994). P. Flandrin, IEEE Trans. Inform. Theory. 35, 197 (1989). P. Flandrin, IEEE Trans. Inform. Theory. 38, 910 (1992). P. Flandrin, Temps-Fre´quence (Herme`s, Paris, 1993). E. Masry, IEEE Trans. Inform. Theory. 39, 260 (1993). P. Abry, P. Goncalve`s, and P. Flandrin, Lectures Note Statistics. 105, 15 (1995). P. Abry, Ondelettes et Turbulence—Multire´solution, Algorithmes de De´composition, Invariance d’Echelles et Signaux de Pression (Diderot Editeur, Arts et Sciences, Paris, 1997). P. Abry and D. Veitch, IEEE Trans. Inform. Theory. 44, 2 (1998). P. Abry, D. Veitch, and P. Flandrin, J. Time Ser. Anal. 19, 253 (1998). L. Abry and P. Delbeke, Stoch. Proc. Applic. 86, 177 (2000). D. Veitch and P. Abry, IEEE Trans. Inform. Theory. 45, 878 (1999). P. Abry, P. Flandrin, M. S. Taqqu, and D. Veitch, in Self-Similarity in Network Traffic, edited by K. Parks and W. Willinger (John Wiley & Sons, New York, 1998). A. H. Tewfik and M. Kim, IEEE Trans. Inform. Theory. 38, 904 (1992). J. Pando and L. Z. Fang, Phys. Rev. E 57, 3593 (1998). J. Simonsen, A. Hansen, and O. M. Nes, Phys. Rev. E 58, 2779 (1998). B. Audit, E. Bacry, J. F. Muzy, and A. Arneodo, IEEE Trans. Inform. Theory 48, 2938 (2002). P. Le´vy, Processus Stochastiques et Mouvement Brownien (Gauthier-Villars, Paris, 1965). R. F. Voss, in Fundamental Algorithms for Computer Graphics, edited by R. A. Earnshaw (Springer-Verlag, Heidelberg, 1985), p. 805. A. Arneodo, E. Bacry, and J. F. Muzy, J. Math. Phys. 39, 4142 (1998). M. E. Cates and J. M. Deutsch, Phys. Rev. A 35, 4907 (1987). A. P. Siebesma, in Universality in Condensed Matter, edited by R. Julien, L. Peliti, R. Rammal, and N. Boccara (Springer-Verlag, Heidelberg, 1988), p. 188. J. O’Neil and C. Meneveau, Phys. Fluids A 5, 158 (1993). A. Arneodo, E. Bacry, S. Manneville, and J. F. Muzy, Phys. Rev. Lett. 80, 708 (1998). M. Greiner, J. Geisemann, P. Lipa, and P. Carruthers, Z. Phys. C 69, 305 (1996). M. Greiner, J. Geisemann, and P. Lipa, Phys. Rev. E 56, 4263 (1997). J. Le´vy-Ve´hel, Fractals. 3, 755 (1995). J. Le´vy-Ve´hel, in Ref. [10], p. 85. M. Unser and A. Aldroubi, Proc. IEEE 84, 626 (1996). Wavelet Applications in Signal and Image Processing VIII., Wavelet Applications in Signal and Image Processing VIII., Vol. 4119 of SPIE Conference Proceedings edited by A. Aldroubi, A. F. Laine, and M. A. Unser (2000). S. Lovejoy, Science. 216, 185 (1982). R. F. Cahalan, in Advances in Remote Sensing and Retrieval Methods, edited by A. Deepak, H. Fleming, and J. Theon (Deepak Pub, Hampton, 1989), p. 371. V. Ramanatahn, R. D. Cess, E. F. Harrison, P. Minnis, B. R. Barkston, E. Ahmad, and D. Hartmann, Science 243, 57 (1989). R. D. Cess, G. L. Potter, J. P. Blanchet, G. J. Boer, S. J. Ghan, J. T. Kiehl, M. Le Treut, Z.-X. Li, X.-Z. Lang, J. F. B. Mitchell, J.-J. Morcrette, D. A. Randall, M. R. Riches, E. Roeckner, U. Schlese, A. Slingo, K. E. Taylor, W. M. Washington, R. T. Wetherald, and I. Yagai, Science 245, 513 (1989).
MULTIFRACTAL IMAGE ANALYSIS
87
176. F. S. Rys and A. Waldvogel, in Fractal in Physics, edited by L. Pietronero and E. Tosatti (North-Holland, Amsterdam, 1986), p. 461. 177. R. M. Welch and B. A. Wielicki, Clim. Appl. Meteorol. 25, 261 (1986). 178. J. I. Yano and Y. Takeuchi, J. Meteorol. Soc. Jpn. 65, 661 (1987). 179. R. M. Welch, K. S. Kuo, B. A. Wielicki, S. K. Sengupta, and L. Parker, J. Appl. Meteorol. 27, 341 (1988). 180. R. F. Cahalan and J. H. Joseph, Mon. Weather Rev. 117, 261 (1989). 181. G. Se`ze and L. Smith, in Proceedings of the Seventh Conference on Atmospheric Radiation, American Meteorological Society, San Francisco, CA (1990), p. 47. 182. A. Davis, S. Lovejoy, and D. Schertzer, in Scaling, Fractals and Nonlinear Variability in Geophysics, edited by S. Lovejoy and D. Schertzer (Kluwer, Dordrecht, 1991), p. 303. 183. Y. Tessier, S. Lovejoy, and D. Schertzer, J. Appl. Meteorol. 32, 223 (1993). 184. A. Davis, A. Marshak, W. J. Wiscombe, and R. F. Cahalan, J. Geophys. Res. 99, 8055 (1994). 185. W. D. King, C. T. Maher, and G. A. Hepburn, J. Appl. Meteorol. 20, 195 (1981). 186. C. Duroure and B. Guillemet, Atmos. Res. 25, 331 (1990). 187. B. Baker, J. Atmos. Sci. 49, 387 (1992). 188. S. P. Malinowski and I. Zawadski, J. Atmos. Sci. 50, 5 (1993). 189. A. V. Korolev and I. P. Mazin, J. Appl. Meteorol. 32, 760 (1993). 190. S. P. Malinowski, M. Y. Leclerc, and D. G. Baumgardner, J. Atmos. Sci. 51, 397 (1994). 191. A. Davis, A. Marshak, W. J. Wiscombe, and R. F. Cahalan, J. Atmos. Sci. 53, 1538 (1996). 192. A. Marshak, A. Davis, W. J. Wiscombe, and R. F. Cahalan, J. Atmos. Sci. 54, 1423 (1997). 193. S. Cox, D. McDougal, D. Randall, and R. Schiffer, Bull. Am. Meteorol. Soc. 68, 114 (1987). 194. B. A. Albrecht, C. S. Bretherton, D. Jonhson, W. H. Schubert, and A. S. Frisch, Bull. Am. Meteorol. Soc. 76, 889 (1995). 195. R. Boers, J. B. Jensen, P. B. Krummel, and H. Gerber, Quart. J. R. Meteorol. Soc. 122, 1307 (1996). 196. H. W. Baker and J. A. Davies, Remote Sens. Environ. 42, 51 (1992). 197. A. Davis, A. Marshak, R. F. Cahalan, and W. J. Wiscombe, J. Atmos. Sci. 54, 241 (1997). 198. R. F. Cahalan and J. B. Snider, Remote Sens. Environ. 28, 95 (1989). 199. S. Lovejoy, D. Schertzer, P. Silas, Y. Tessier, and D. Lavalle´e, Ann. Geophys. 11, 119 (1993). 200. S. M. Gollmer, M. Harshvardan, R. F. Cahalan, and J. S. Snider, J. Atmos. Sci. 52, 3013 (1995). 201. W. J. Wiscombe, A. Davis, A. Marshak, and R. F. Cahalan, Proceedings of the Fourth Atmospheric Radiation Measurement (ARM) Science Team Meeting, Charleston, U.S. Department of Energy 11 (1995). 202. A. Davis, A. Marshak, H. Gerber, and W. J. Wiscombe, J. Geophys. Res (1998) Unpublished. 203. D. Lovejoy and S. Schertzer, in Turbulence and Chaotic Phenomena in Fluids, edited by T. Tatsumi (North-Holland, Amsterdam, 1984), p. 505. 204. D. Lovejoy and S. Schertzer, in Fractals: Their Physical Origin and Properties, edited by L. Pietronero (Plenum, New York, 1989), p. 49. 205. J. Wilson, D. Schertzer, and S. Lovejoy, in Scaling, Fractals and Nonlinear Variability in Geophysics, edited by D. Schertzer and S. Lovejoy (Kluwer, Dordrecht, 1991), p. 185. 206. D. Schertzer and S. Lovejoy, in Ref. [10], p. 11.
88
´ ODO ET AL. ARNE
207. A. Davis, A. Marshak, W. J. Wiscombe, and R. F. Cahalan, Proceedings of the 2nd Workshop on Nonstationary Random Processes and Their Applications (1995), preprint. 208. L. M. Romanova, Izv. Acad. Sci. USSR Atmos. Oceanic Phys. 11, 509 (1975). 209. A. Davis, Ph.D. thesis, McGill University, Montreal, 1992. 210. R. F. Cahalan, W. Ridgway, W. J. Wiscombe, T. L. Bell, and J. B. Snider, J. Atmos. Sci. 51, 2434 (1994). 211. R. D. Cess, M. H. Zhang, Y. Zhou, X. Jing, and V. Dvortsov, J. Geophys. Res. 101, 23299 (1996). 212. K. Stamnes, S.-C. Tsay, W. J. Wiscombe, and K. Jayaweera, Appl. Opt. 27, 2502 (1988). 213. R. F. Cahalan, W. Ridgway, W. J. Wiscombe, S. Gollmer, and M. Harshvardan, J. Atmos. Sci. 51, 3776 (1994). 214. A. Marshak,A.Davis,W.J. Wiscombe, andR.F. Cahalan,J.Geophys.Res. 100,26247(1995). 215. M. Tiedke, Mon. Weather Res. 124, 745 (1996). 216. A. Davis, A. Marshak, W. J. Wiscombe, and R. F. Cahalan, in Current Topics in Nonstationary Analysis, edited by G. Trevin˜o et al. (World Scientific, Singapore, 1996), p. 97. 217. M. Harshvardan, B. A. Wielicki, and K. M. Ginger, J. Climate. 7, 1987 (1994). 218. R. F. Cahalan, M. Nestler, W. Ridgway, W. J. Wiscombe, and T. L. Bell, in Proceedings the 4th International Meeting on Statistical Climatology, edited by J. Sansom (New Zealand Meteorological Service, Wellington, 1990), p. 28. 219. A. Davis, S. Lovejoy, and D. Schertzer, SPIE Proc. 1558, 37 (1991). 220. A. Marshak, A. Davis, W. J. Wiscombe, and G. Titov, Remote Sens. Environ. 52, 72 (1995). 221. S. G. Roux, Ph.D. thesis, University of Aix-Marseille II, 1996. 222. A. N. Kolmogorov, C. R. Acad. Sci. USSR. 30, 301 (1941). 223. G. Ruiz-Chavarria, C. Baudet, and S. Ciliberto, Physica D 99, 369 (1996). 224. C. H. Meong, W. R. Cotton, C. Bretherton, A. Chlond, M. Khairoutdinov, S. Krueger, W. S. Lewellen, M. K. McVean, J. R. M. Pasquier, H. A. Rand, A. P. Siebesma, B. Stevens, and R. I. Sykes, Bull. Am. Meteorol. Soc. 77, 261 (1996). 225. A. Marshak, A. Davis, R. F. Cahalan, and W. J. Wiscombe, Phys. Rev. E 49, 55 (1994). 226. U. Frisch and S. A. Orszag, Phys. Today 24, (1990). 227. C. Meneveau and K. R. Sreenivasan, J. Fluid Mech. 224, 429 (1991). 228. Turbulence: A Tentative Dictionary, Turbulence: A Tentative Dictionary, edited by P. Tabeling and O. Cardoso (Plenum, New York, 1995). 229. K. R. Sreenivasan and R. A. Antonia, Annu. Rev. Fluid Mech. 29, 435 (1997). 230. M. Briscolini, P. Santangelo, S. Succi, and R. Benzi, Phys. Rev. E 50, R1745 (1994). 231. A. Vincent and M. Meneguzzi, J. Fluid Mech. 225, 1 (1995). 232. C. W. Van Atta and W. Y. Chen, J. Fluid Mech. 44, 145 (1970). 233. F. Anselmet, Y. Gagne, E. J. Hopfinger, and R. A. Antonia, J. Fluid Mech. 140, 63 (1984). 234. Y. Gagne, Ph.D. thesis, University of Grenoble, 1987. 235. B. Castaing, Y. Gagne, and E. J. Hopfinger, Physica D 46, 177 (1990). 236. C. Baudet, S. Ciliberto, and Phan Nhan Tien, J. Phys. II France 3, 293 (1993). 237. G. Stolovitzky and K. R. Sreenivasan, Phys. Rev. E 48, R33 (1993). 238. J. Maurer, P. Tabeling, and G. Zocchi, Europhys. Lett. 26, 31 (1994). 239. J. Herweijer and W. Van de Water, Phys. Rev. Lett. 74, 4651 (1995). 240. A. Arneodo et al., Europhys. Lett. 34, 411 (1996). 241. R. Benzi, L. Biferale, G. Paladin, A. Vulpiani, and M. Vergassola, Phys. Rev. Lett. 67, 2299 (1991). 242. P. Kailasnath, K. R. Sreenivasan, and G. Stolovitzky, Phys. Rev. Lett. 68, 2766 (1992). 243. A. Praskovsky and S. Oncley, Phys. Rev. Lett. 7, 3999 (1994).
MULTIFRACTAL IMAGE ANALYSIS
89
244. P. Tabeling, G. Zocchi, F. Belin, J. Maurer, and H. Willaime, Phys. Rev. E 53, 1613 (1996). 245. F. Belin, P. Tabeling, and H. Willaime, Physica D 93, 52 (1996). 246. B. Castaing, Y. Gagne, and M. Marchand, Physica D 68, 387 (1993). 247. G. Pedrizetti, E. Novikov, and A. Praskovsky, Phys. Rev. E 53, 475 (1996). 248. R. Benzi, S. Ciliberto, R. Tripiccione, C. Baudet, F. Massaioli, and S. Succi, Phys. Rev. E 48, R29 (1993). 249. R. Benzi, S. Ciliberto, C. Baudet, G. R. Chavarria, and R. Tripiccione, Europhys. Lett. 24, 275 (1993). 250. R. Benzi, S. Ciliberto, C. Baudet, and G. R. Chavarria, Physica D 80, 385 (1995). 251. Y. Gagne, M. Marchand, and B. Castaing, J. Phys. II France 4, 1 (1994). 252. A. Naert, L. Puech, B. Chabaud, J. Peinke, and B. Castaing B. Hebral, J. Phys. II France 4, 215 (1994). 253. B. Chabaud, A. Naert, J. Peinke, F. Chilla`, B. Castaing, and B. Hebral, Phys. Rev. Lett. 73, 3227 (1994). 254. B. Dubrulle and B. Castaing, J. Phys. II France 5, 895 (1995). 255. F. Chilla`, J. Peinke, and B. Castaing, J. Phys. II France 6, 455 (1996). 256. Y. Male´cot, C. Auriault, H. Kahalerras, Y. Gagne, O. Chanal, B. Chabaud, and B. Castaing, Eur. Phys. J. B 16, 549 (2000). 257. O. Chanal, B. Chabaud, B. Castaing, and B. Hebral, Eur. Phys. J. B 17, 309 (2000). 258. A. Arneodo, S. Manneville, J. F. Muzy, and S. G. Roux, Appl. Comput. Harmonic Anal. 6, 374 (1999). 259. R. Peinke and J. Friedrich, Phys. Rev. Lett. 78, 863 (1997). 260. R. Peinke and J. Friedrich, Physica D 102, 147 (1997). 261. P. O. Amblard and J. M. Brossier, Eur. Phys. J. B 12, 335 (1999). 262. P. Marcq and A. Naert, Phys. Fluids 13, 2590 (2001). 263. J. Davoudi and M. R. R. Tabar, Phys. Rev. Lett. 82, 1680 (1999). 264. J. P. Laval, B. Dubrulle, and S. Nazarenko, Phys. Fluids 13, 1995 (2001). 265. I. Arad, B. Dhruva, S. Kurien, V. S. L’vov, I. Procaccia, and K. R. Sreenivasan, Phys. Rev. Lett. 81, 5330 (1998). 266. I. Arad, V. S. L’vov, and I. Procaccia, Phys. Rev. E 59, 6753 (1999). 267. S. Kurien and K. R. Sreenivasan, Phys. Rev. E 62, 2206 (2000). 268. L. Biferale and F. Toschi, Phys. Rev. Lett. 86, 4831 (2001). 269. A. N. Kolmogorov, J. Fluid Mech. 13, 82 (1962). 270. A. M. Obukhov, J. Fluid Mech. 13, 77 (1962). 271. I. Hosokawa and K. Yamamoto, Phys. Fluids A 4, 457 (1992). 272. A. A. Praskovsky, Phys. Fluids A 4, 2589 (1992). 273. S. T. Thoroddsen and C. W. Van Atta, Phys. Fluids A 4, 2592 (1992). 274. S. Chen, G. D. Doolen, R. H. Kraichnan, and Z. S. She, Phys. Fluids A 5, 458 (1992). 275. G. Stolovitzky, P. Kailasnath, and K. R. Sreenivasan, Phys. Rev. Lett. 69, 1178 (1992). 276. G. Sreenivasan and K. R. Stolovitzky, Rev. Mod. Phys. 66, 229 (1994). 277. A. A. Praskovsky and S. Oncley, Europhys. Lett. 28, 635 (1994). 278. S. T. Thoroddsen, Phys. Fluids 7, 691 (1995). 279. S. Chen, G. D. Doolen, R. H. Kraichnan, and L. P. Wang, Phys. Rev. Lett. 74, 1755 (1995). 280. V. Borue and S. A. Orszag, Phys. Rev. E 53, R21 (1996). 281. R. Benzi, R. Struglia, and R. Tripiccione, Phys. Rev. E 53, R5565 (1996). 282. L. P. Wang, S. Chen, J. G. Brasseur, and J. C. Wyngaard, J. Fluid Mech. 309, 113 (1996). 283. A. Tsinober, E. Kit, and T. Dracos, J. Fluid Mech. 242, 169 (1992). 284. L. Shtilman, M. Spector, and A. Tsinober, J. Fluid Mech. 247, 65 (1993).
90 285. 286. 287. 288. 289. 290. 291. 292. 293. 294. 295. 296. 297. 298. 299. 300. 301. 302. 303. 304. 305. 306. 307. 308. 309. 310. 311. 312. 313. 314. 315. 316. 317. 318. 319. 320. 321. 322. 323. 324. 325. 326. 327. 328. 329. 330.
´ ODO ET AL. ARNE L. Richardson, Proc. R. Soc. London Ser. A 110, 709 (1926). S. Kida, J. Phys. Soc. Jpn. 60, 5 (1990). E. A. Novikov, Physica A 2, 814 (1990). E. A. Novikov, Phys. Rev. E 50, 3303 (1995). B. Dubrulle, Phys. Rev. Lett. 73, 959 (1994). Z. S. She and E. C. Waymire, Phys. Rev. Lett. 74, 262 (1995). B. Dubrulle, J. Phys. II France 6, 1825 (1996). Z. S. She and E. Leveque, Phys. Rev. Lett. 72, 336 (1994). A. B. Chhabra and K. R. Sreenivasan, Phys. Rev. Lett. 68, 2762 (1992). B. Jouault, P. Lipa, and M. Greiner, Phys. Rev E 59, 2451 (1999). D. Sornette, in Scale Invariance and Beyond, edited by B. Dubrulle, F. Graner, and D. Sornette (EDP Sciences, Les Ulis, 1997), p. 235. W. X. Zhou and D. Sornette, Physica D 165, 94 (2002). B. B. Mandelbrot, C. R. Acad. Sci. Paris Ser. A 278, 289, 355 (1974). B. B. Mandelbrot, J. Fluid Mech. 62, 331 (1974). J. P. Kahane and J. Peyrie`re, Adv. Math. 22, 131 (1976). G. M. Molchan, Commun. Math. Phys. 179, 681 (1996). A. Naert, R. Friedrich, and J. Peinke, Phys. Rev. E 56, 6719 (1997). P. Naert and A. Marcq, Physica D 124, 368 (1998). C. Meneveau and K. R. Sreenivasan, Nucl. Phys. B2, 49 (1987). C. Meneveau and K. R. Sreenivasan, Phys. Rev. Lett. 59, 1424 (1987). C. Meneveau and K. R. Sreenivasan, Phys. Lett. A 137, 103 (1989). E. Aurell, U. Frisch, J. Lutsko, and M. Vergassola, J. Fluid Mech. 238, 467 (1992). G. Paladin and A. Vulpiani, Phys. Rev. A 35, 1971 (1987). C. Meneveau and M. Nelkin, Phys. Rev. A 39, 3732 (1989). U. Frisch and M. Vergassola, Europhys. Lett. 14, 439 (1991). W. Van de Water, B. Van der Vorst, and E. Van de Wetering, Europhys. Lett. 16, 443 (1991). J. Molenaar, J. Herweijer, and W. Van de Water, Phys. Rev. E 52, 496 (1995). I. Hosokawa, S. Oide, and K. Yamamoto, Phys. Rev. Lett. 77, 4548 (1996). A. Bershadskii, T. Nakano, D. Fukayama, and T. Gotoh, Eur. Phys. J. B 18, 95 (2000). A. Bershadskii and A. Tsinober, Phys. Rev. E 48, 282 (1993). A. Bershadskii, E. Kit, A. Tsinober, and H. Vaisburd, Fluid Dyn. Res. 14, 71 (1994). R. Badii and P. Talkner, Phys. Rev. E 59, 6715 (1999). R. Badii and P. Talkner, Phys. Rev. E 60, 4138 (1999). V. L’vov and I. Procaccia, Phys. Fluids 8, 2565 (1996). G. He, S. Chen, R. H. Kraichnan, R. Zhang, and Y. Zhou, Phys. Rev. Lett. 81, 4636 (1998). M. Nelkin, Phys. Fluids 11, 2202 (1999). E. Siggia, J. Fluid Mech. 107, 375 (1981). R. Kerr, J. Fluid Mech. 153, 31 (1985). C. Meneveau, K. R. Sreenivasan, G. P. Kailasnath, and M. S. Fan, Phys. Rev. A 41, 894 (1990). M. S. Shafi, Y. Zhu, and R. A. Antonia, Phys. Fluids 8, 2245 (1996). N. Cao, S. Chen, and K. R. Sreenivasan, Phys. Rev. Lett. 76, 616 (1996). S. Chen, K. R. Sreenivasan, and M. Nelkin, Phys. Rev. Lett. 79, 1253 (1997). S. Chen, K. R. Sreenivasan, M. Nelkin, and N. Cao, Phys. Rev. Lett. 79, 2253 (1997). W. Van de Water and J. Herweijer, Bull. Am. Phys. Soc. 41, 1782 (1996). R. Camussi and R. Benzi, Phys. Fluids 9, 257 (1997). O. N. Boratov and R. B. Pelz, Phys. Fluids 9, 1400 (1997).
MULTIFRACTAL IMAGE ANALYSIS 331. 332. 333. 334. 335. 336. 337. 338. 339. 340. 341. 342. 343. 344. 345. 346. 347. 348. 349.
350. 351. 352. 353. 354. 355. 356. 357. 358. 359. 360. 361. 362. 363.
91
S. Grossmann, D. Lohse, and A. Reeh, Phys. Fluids 9, 3817 (1997). R. A. Antonia and B. R. Pearson, Europhys. Lett. 40, 123 (1997). B. Dhruva, Y. Tsuji, and K. R. Sreenivasan, Phys. Rev. E 56, R4928 (1997). M. Meneguzzi, Private communication. E. Ott, Y. Du, K. R. Sreenivasan, A. Juneja, and A. K. Suri, Phys. Rev. Lett. 69, 2654 (1992). Y. Du and E. Ott, Physica D 67, 387 (1993). A. L. Bertozzi and A. B. Chhabra, Phys. Rev. E 49, 4716 (1994). M. J. Broeders and A. L. Verbeek, in Radiological Diagnosis of Breast Diseases, edited by M. Friedrich and E. Sickles (Springer-Verlag, Berlin, 1997), p. 1. M. H. Dilhuydy and B. Barreau, Eur. J. Radiol. 24, 86 (1997). K. Doi, M. L. Giger, R. M. Nishikawa, K. R. Hoffmann, H. MacMahon, R. A. Schmidt, and K. G. Chua, Acta Radiol. 34, 426 (1993). A. F. Laine, S. Schuler, J. Fan, and W. Huda, IEEE Trans. Med. Imaging 13, 725 (1994). W. Qian, M. Kallergi, L. P. Clarke, H.-D. Li, P. Venugopal, D. Song, and R. A. Clark, Med. Phys. 22, 1247 (1995). R. A. Devore, B. Lucier, and Z. Yang, in Wavelets in Medicine and Biology, edited by A. Aldroubi and M. Unser (CRC Press, Boca Raton, FL, 1996), p. 145. J. J. Heine, S. R. Deans, D. K. Cullers, R. Stauduhar, and L. P. Clarke, IEEE Trans. Med. Imaging 16, 503 (1997). H. P. Chan, K. Doi, C. J. Vyborny, R. A. Schmidt, C. E. Metz, K. L. Lam, T. Ogura, Y. Wu, and H. MacMahon, Invest. Radiol. 25, 1102 (1990). D. R. Davies and D. H. Dance, Phys. Med. Biol. 35, 1111 (1990). J. Dengler, S. Behrens, and J. F. Desage, IEEE Trans. Med. Imaging 12, 634 (1993). R. M. Nishikawa, M. L. Giger, K. Doi, C. J. Vyborny, and R. A. Schmidt, Med. Phys. 20, 1161 (1993). A. Bazzani, A. Bevilacqua, D. Bollini, R. Campanini, N. Lanconelli, A. Riccardi, and D. Romani, in Digital Mammography: IWDM 2000, 5th International Workshop, edited by M. Yaffe (Medical Physics Publishing, Madison, 2001). N. Karssemeijer, in Proceedings of the 12th International Conference on Information Processing Medical Imaging, (Springer-Verlag, Berlin, 1991), p. 227. N. Karssemeijer, Int. J. Pattern Recog. Artificial Intell. 7, 1357 (1993). C. E. Priebe, J. L. Solka, R. A. Lorey, G. W. Rogers, W. L. Poston, M. Kallergi, W. Qian, L. P. Clarke, and R. A. Clark, Cancer Lett. 77, 183 (1994). H. D. Li, M. Kallergi, L. P. Clarke, V. K. Jain, and R. A. Clark, IEEE Trans. Med. Imaging 14, 565 (1995). W. J. H. Veldkamp, N. Karssemeijer, J. D. M. Otten, and J. H. C. L. Hendricks, Med. Phys. 27, 2600 (2000). Y. Wu, K. Doi, M. L. Giger, and R. M. Nishikawa, Med. Phys. 19, 555 (1992). Y. Wu, M. L. Giger, K. Doi, C. J. Vyborny, and R. A. Schmidt, Radiology 187, 81 (1993). H. P. Chan, S. C. B. Lo, B. Sahiner, K. L. Lam, and M. A. Helvie, Med. Phys. 22, 1555 (1995). W. Zhang, K. Doi, M. L. Giger, R. M. Nishikawa, and R. A. Schmidt, Med. Phys. 23, 595 (1996). B. Zheng, W. Qian, and L. P. Clarke, IEEE Trans. Med. Imaging 15, 589 (1996). S. Yu L. Guan, IEEE Trans. Med. Imaging 19, 115 (2000). H. Cheng, Y. M. Lui, and R. I. Feiimanis, IEEE Trans. Med. Imaging 17, 442 (1998). M. A. Gavrielides, J. Y. Lo, R. Vargas-Voracek, and C. E. FloydJr., Med. Phys. 27, 13 (2000). P. K. Saha, J. K. Udupa, E. F. Conant, D. P. Chakraborty, and D. Sullivan, IEEE Trans. Med. Imaging 20, 792 (2001).
92
´ ODO ET AL. ARNE
364. H. Yoshida, K. Doi, R. M. Nishikawa, K. Muto, and M. Tsuda, Acad. Rep. Tokyo Inst. Polytech. 17, 24 (1994). 365. H. Yoshida, R. M. Nishikawa, M. L. Giger, and K. Doi, Proc. SPIE. 2825, 805 (1996). 366. H. Yoshida, K. Doi, R. M. Nishikawa, M. L. Giger, and R. A. Schmidt, Acad. Radiol. 3, 621 (1996). 367. R. N. Strickland and H. I. I. Hahn, IEEE Trans. Med. Imaging 15, 218 (1996). 368. W. Zhang, H. Yoshida, R. M. Nishikawa, and K. Doi, Med. Phys. 25, 949 (1998). 369. J. M. Lado, P. G. Tahoces, A. J. Mendez, M. Souto, and J. J. Vidal, Med. Phys. 26, 1294 (1999). 370. T. Netsch and H. O. Peitgen, IEEE Trans. Med. Imaging 18, 774 (1999). 371. W. Qian, L. Li, X. Sun, and R. A. Clark, in Wavelet Applications in Signal and Image Processing VIII, SPIE Conference Proceedings, edited by A. Aldroubi, A. F. Laine, and M. A. Unser , p. 596–604 (2000). 372. J. K. Park and H. W. Kim, IEEE Trans. Med. Imaging 18, 231 (1999). 373. S. K. Lee, C. S. Lo, C. M. Wang, P. C. Chung, C. I. Chang, C. W. Yang, and P. C. Hsu, Int. J. Med. Informatics 60, 29 (2000). 374. C. B. Caldwell, S. J. Stapleton, D. W. Holdsworth, R. A. Jong, W. J. Weiser, G. Cooke, and M. J. Yaffe, Phys. Med. Biol. 35, 235 (1990). 375. F. Lefebvre, H. Benali, R. Gilles, E. Kahn, and R. Di Paola, Med. Phys. 22, 381 (1995). 376. D. L. Thiele, C. Kimme-Smith, T. D. Johnson, M. McCombs, and L. W. Bassett, Med. Phys. 23, 549 (1996). 377. H. Guillemet, H. Benali, E. Kahn, and R. Di Paola, Acta Stereol. 15/2, 125 (1996). 378. V. Velanovich, Am. J. Med. 311, 211 (1996). 379. J. J. Heine, S. R. Deans, R. P. Velthuizen, and L. P. Clarke, Med. Phys. 26, 2254 (1999). 380. M. Heath, K. W. Bowyer, and D. Kopanset al., in Digital Mammography, (Kluwer Academic, Dordrecht, 1998), p. 457. 381. J. N. Wolfe, Cancer. 37, 2486 (1976). 382. A. M. Boyd and N. F. Oza, Epidemiol. Rev. 15, 196 (1993). 383. N. F. Boyd, J. W. Byng, R. A. Jong, E. K. Fishell, L. E. Little, A. B. Miller, G. A. Lockwood, D. L. Tritchler, and M. J. Yaffe, J. Natl. Cancer Inst. 87, 670 (1995). 384. J. J. Heine and R. P. Velthuizen, Med. Phys. 27, 2644 (2000). 385. Universality in Chaos, Universality in Chaos, edited by P. Cvitanovic (Hilger, Bristol, 1984). 386. Chaos, Chaos, edited by B. L. Hao (World Scientific, Singapore, 1984). 387. H. G. Schuster, Deterministic Chaos (Physik Verlag, Weimheim, 1984). 388. P. Berge´, Y. Pomeau, and C. Vidal, Order within Chaos (Wiley, New York, 1986).
ADVANCES IN IMAGING AND ELECTRON PHYSICS, VOL. 126
An Analysis of the Geometric Distortions Produced by Median and Related Image Processing Filters E. R. DAVIES Machine Vision Group, Department of Physics, Royal Holloway College, University of London, Egham, Surrey, TW20 0EX, United Kingdom
I. Introduction. . . . . . . . . . . . . . . . . . . . . . . II. Image Filters . . . . . . . . . . . . . . . . . . . . . . A. Noise Suppression Filters . . . . . . . . . . . . . . . . B. Mode Filters . . . . . . . . . . . . . . . . . . . . . C. Morphological Filters . . . . . . . . . . . . . . . . . . D. In-Depth Study of Median Filters . . . . . . . . . . . . . III. Shifts Produced by Median Filters in Continuous Images . . . . . A. Theory of Edge Shifts Produced by Median Filters in Continuous Binary Images. . . . . . . . . . . . . . . . . . . . . B. Extension to Continuous Gray-Scale Images . . . . . . . . . C. Extension to Discrete Neighborhoods . . . . . . . . . . . . D. Experimental Results for Discrete Binary Images . . . . . . . E. Experimental Results for Discrete Gray-Scale Images . . . . . . F. Edge Shifts Arising with Hybrid Median Filters . . . . . . . . IV. Shifts Produced by Median Filters in Digital Images . . . . . . . A. Using a Discrete Model to Explain Median Shifts . . . . . . . B. Theoretical Shifts for a 3 3 Neighborhood . . . . . . . . . C. More General Calculation of Edge Shifts . . . . . . . . . . D. Experimental Results for a 3 3 Neighborhood . . . . . . . . E. Numerical Computations for 5 5 Neighborhoods . . . . . . F. Numerical Computations for 7 7 Neighborhoods . . . . . . G. Tests of the Theory for 5 5 and 7 7 Neighborhoods . . . . . H. Discussion . . . . . . . . . . . . . . . . . . . . . . I. Trends for Large Neighborhoods . . . . . . . . . . . . . J. Effect of Sampling at the Center of a Pixel . . . . . . . . . . K. Case of Median Filter with Small Circles . . . . . . . . . . V. Shifts Produced by Mean Filters . . . . . . . . . . . . . . . A. Shifts for Step Edges . . . . . . . . . . . . . . . . . . B. Shifts for Linear Slant Edges . . . . . . . . . . . . . . . C. Discussion . . . . . . . . . . . . . . . . . . . . . . VI. Shifts Produced by Mode Filters . . . . . . . . . . . . . . . A. Shifts for Step Edges . . . . . . . . . . . . . . . . . . B. Shifts for Slant Edges . . . . . . . . . . . . . . . . . . C. Discussion . . . . . . . . . . . . . . . . . . . . . . D. Case of Mode Filter with Small Circles . . . . . . . . . . . VII. Shifts Produced by Rank-Order Filters . . . . . . . . . . . . A. Shifts in Rectangular Neighborhoods . . . . . . . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
94 96 98 100 102 104 105
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
105 110 113 113 116 120 122 122 124 128 129 131 135 136 137 137 141 143 146 146 147 149 150 150 150 151 153 156 157
93 Copyright 2003 Elsevier Science (USA). All rights reserved. ISSN 1076-5670/03
94
VIII.
IX.
X.
XI.
XII.
E. R. DAVIES B. Shifts in Circular Neighborhoods . . . . . . . . . . C. Case of High Curvature . . . . . . . . . . . . . D. Test of the Model in a Discrete Case . . . . . . . . E. Mean Distance from Center of Neighborhood to a Tangent F. Discussion . . . . . . . . . . . . . . . . . . . Rank-Order Filters—a Didactic Example . . . . . . . . A. Analysis of the Situation . . . . . . . . . . . . . B. Discussion . . . . . . . . . . . . . . . . . . . A Problem with Closing . . . . . . . . . . . . . . . A. Detailed Analysis . . . . . . . . . . . . . . . . B. Discussion . . . . . . . . . . . . . . . . . . . A Median-Based Corner Detector . . . . . . . . . . . A. Analyzing the Operation of the Median Detector . . . . B. Practical Results . . . . . . . . . . . . . . . . Boundary Length Measurement Problem . . . . . . . . A. Detailed Analysis . . . . . . . . . . . . . . . . B. Discussion . . . . . . . . . . . . . . . . . . . Concluding Remarks . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . .
. . . . . . Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . .
160 163 165 168 169 170 170 173 174 175 177 178 179 181 183 184 187 188 190
I. Introduction Over the past forty years or so, image processing has become a key science that is applied in a great many areas, ranging from photography, cinematography, and television to space and forensic science, medicine, and even the recovery of ancient manuscripts. The general idea of using image processing in these areas is to convert one image into another, with the aim of improving or modifying the output in some way. One example is the elimination of noise from images. Another is the enhancement of images, as in the case of chest radiographs, which might be expected to become easier for a clinician to use for diagnosis. Clearly, modification of images so that they may be viewed and interpreted more easily by human operators constitutes one class of image processing: this might have been expected, and ‘‘improvement’’ is very much a subjective term that pertains naturally to human judgment, though perhaps not so readily to scientific analysis. Further examples of image processing are the restoration of images to some ideal form that they would have had before transmission over some degrading medium. In the case of television, degradation can involve speckling and ghosting, as well as inclusion of the sometimes annoying scan lines that are characteristic of this originally analogue process. In some applications, degradation can take the form of blurring caused by a
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
95
turbulent atmosphere, by motion of an object in the scene being viewed, or by motion of the camera. But there are totally different uses of image processing. Images can be processed to locate particular objects as part of a more general process of image analysis. Here we are less interested in retaining the image representation as such, and far more interested in describing or measuring the content of the images. Indeed, if a suitably detailed description of the image content can be obtained, the image data and format may be thrown away. This happens when a human is driving a car down a motorway, and when a robot vision system is guiding a missile by analyzing a sequence of received images and deducing exactly where the target is. Less mundane tasks involving image analysis include automated visual inspection, where defective products or contaminants have to be located with a view to rejection. There are cogent reasons why locating objects in images is computation intensive. In particular, any object template will have to be applied at every location within an image. It will also have to be applied in all directions and perhaps at many scales. Furthermore, any variations in shade, color, texture, or other characteristics may necessitate many more tests, and the multidimensionality of the search process means that it will involve a combinatorial explosion of possibilities. As a result, it is generally far easier to locate objects from their features, as small features are much easier to locate, not least because they are subject to fewer variations including fewer distinguishable orientations. Proceeding in this way is not without its own problems, as locating the features does not then uniquely locate the objects. In fact, the presence of the objects has to be inferred, and special algorithms are needed to carry out this process—though that is another story (Davies, 1997a). Here the important point is that features have to be located in digital images. Algorithms that perform this sort of task are called filters, as their action is analogous to that of sieves in sorting the various parts of the image and extracting only the relevant parts—in this case the features that are needed for object recognition. Interestingly, the algorithms that are needed for performing many of the other image processing functions mentioned previously are also called filters, typically because they remove noise, scan lines, and speckle, or because they filter out blur or even act so as to locate the most meaningful parts of the image, thereby enhancing it. In the next section we enquire in more detail how the various filtering operations are carried out and how filters may be designed for specific tasks. Meanwhile we note that many of these operations are carried out by wellknown types of filter—mean and Gaussian filters, median filters, mode filters, and the rather large class of morphological filters. Several of these filters have been in wide use for decades, and it might be thought that the
96
E. R. DAVIES
subject of image processing would by now have reached a state of maturity such that the properties of these sorts of filter would be well known and fully documented. However, it has become increasingly clear for some years that although capable of performing certain image processing tasks extremely effectively, several of these filters also inadvertently distort the images they are processing. Furthermore, the extent of these distortions is often not known accurately and may not even be suspected by workers. As a result, it may occasionally be the case (especially when about to make specific measurements from images as part of inspection or related processes) that the best advice to those about to employ image filters is ‘‘Don’t!’’ This article is aimed at the analysis and elucidation of the various distorting processes that occur when applying common image processing filters. The median filter has a special position in the hierarchy, as it is widely used—particularly as it does not blur images in the way that mean or Gaussian filters are known to—and a whole variety of other filters are derived from or closely related to it, so that they tend to have similar properties. Thus the distortions exhibited by median filters apply in one form or another to many other types of filter. In Section II we describe some of the standard image processing filters, and present their basic properties. Then in Section III we consider the origins of the distortions produced by median filters, and make quantitative estimates of them on the assumption of a continuous analogue image space. In Section IV we proceed to extend these estimates by making due allowance for the discrete nature of the digital lattice within the image space. In Section V we apply the continuum model to the mean filter, and in Section VI we apply it to the mode filter. Then in Section VII we generalize the continuum model to cover all types of rank-order filter, with the median filter and certain morphological filters as special cases. This work is extended in Section VIII with a didactic example relating to rank-order filters, in Section IX by study of a morphological problem, in Section X by discussion of a median-based corner detector, and in Section XI by consideration of a boundary length measurement problem. Section XII presents some concluding remarks about a whole range of problems relating to continuous versus discrete representations.
II. Image Filters Perhaps the simplest type of filter is the convolution. In fact, the convolution is the most general spatially invariant linear filter that can be applied to an image. It has many realizations, but all are in the form of a weighted set of mask
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
97
coefficients that is multiplied by corresponding pixel intensities within the filter support region (commonly known as the ‘‘neighborhood’’), and which are then added to produce the final local output value. Not only does this paradigm include spatial matched filtering—a well-known procedure for enhancing desired signals of any known intensity profile—but also it includes local spatial averaging that smoothes the image locally, thereby helping to eliminate noise (Davies, 1997a). Other convolution operations include edge and feature enhancement, thus leading to the possibility of performing edge and feature detection. Although convolution filters form a very important class of filter, they are limited by their very linearity. Nonlinear filters are even more powerful as they embody no such restriction, though arguably this makes them more difficult to design as considerably more freedom is available to the algorithm designer. Nevertheless, a simple technique exists for creating useful nonlinear filters. This is to employ a two-stage procedure in which (1) a convolution is applied to enhance some particular set of features in an image, and (2) the features are located by a process such as thresholding or nonmaximum suppression, or by a combination of these, or by more sophisticated measures. The important point is that the features are ‘‘detected’’ only by the final nonlinear process. Indeed, it is the nonlinear process that provides much of the power: it is also the stage that corresponds to a set of decisions being made about where the features are. Edge detection constitutes a typical instance of the use of convolution followed by a nonlinear decision-making process. Line segment detection and corner detection are further instances in which this technique can be applied (Davies, 1997a). Two further approaches to the design of nonlinear filters are in wide use: one is the rank-order filter, typified by the median filter. The other is the morphological filter. The principle employed in rank-order filters is to take all the intensity values in a given neighborhood, to place these in order of increasing value, and finally to select the rth of the n values and return this value as the filter local output value. Clearly, n rank-order filters can be specified in terms of the value r that is used, but these filters all have the characteristic that they are intrinsically nonlinear, i.e., the output intensity cannot be expressed as a linear sum of the component intensities within the neighborhood. In particular, the median filter (for which r ¼ ðn þ 1Þ=2, and which is defined only if n is odd1) does not normally give the same output image as a mean filter: indeed, it is well known that the mean and median of a distribution are in general only coincident for symmetrical distributions. Note that minimum and 1
If n is even, it is usual to take the mean of the central two values in the distribution as representing the median.
98
E. R. DAVIES
maximum filters (corresponding to r ¼ 1 and r ¼ n respectively) are also often classed as morphological filters (see below and Section II.C). Morphological filters constitute another large class of nonlinear filters. Originally, the basic concept was to analyse and filter object shapes in binary images. However, the mathematical foundation of the subject has been considerably developed in recent years [Haralick et al. (1987) being a landmark paper], and currently the aim is to analyse both intensity variations and shapes in tandem, the mathematics being necessary to understand in a profound way the possible shapes and intensity patterns, and how they are related to each other, and how they may be processed to derive further shapes and intensity patterns.
A. Noise Suppression Filters In this section we consider the nature of the noise removal process, and how an ideal noise removal filter might be constructed. A priori, a good way of removing noise from images is to average a number of nominally identical images. In particular, if a camera is pointing at a still-life scene, averaging a sequence of off-camera images will gradually lead to the signal-to-noise ratio (SNR) being boosted, and by the time N pffiffiffiffiffi images have been averaged, the SNR will be N times larger (the total signal will be multiplied by N, and thepnoise power will be multiplied by N, ffiffiffiffiffi pffiffiffiffi ffi but the SNR will be multiplied by N= N , which equals N ). The problem with this approach is that there must be no camera shake and no motion in the scene. To overcome this problem, we can approximate by averaging locally within a single image, though this introduces a further problem— that the image will become blurred (Davies, 1997a). The blurring problem is also manifest if we consider the action of the filter in the spatial frequency domain, where removal of high spatial frequency noise also suppresses the rapidly varying signal components, and ultimately this must introduce blurring. In fact, optimal low-pass noise suppression filters need to employ mathematically well-behaved functions that are smoothly varying both in the spatial domain and in the spatial frequency domain—and in this sense the Gaussian smoothing kernel is optimal. However, whether the Gaussian smoothing filter or the mean (simple local averaging) filter is used, blurring is bound to occur and the underlying signal will not be exactly preserved. Fortunately, the median filter largely overcomes this problem. At the same time it has quite different noise suppression characteristics. In particular, it preferentially eliminates the outliers in any distribution, and thus has excellent impulse noise elimination characteristics. The validity of
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
99
this statement follows from the fact that the median is the center value of the local intensity distribution. This means that it can be considered to work by eliminating the two most extreme intensities at the lower and upper ends of the distribution, and also the next two, and the next two; and so on until only the central one remains. To understand how the median filter can remove noise without blurring the image, note that any monotonically increasing (or decreasing) function of intensity is unaffected by a median operation. More precisely, if an intensity function is monotonically increasing over the whole neighborhood of a median filter, the median operation will leave it unchanged. This onedimensional (1D) property is found to extend unchanged to twodimensional (2D) images. A further observation about the median is that although it is commonly considered not to blur images, it nonetheless appears to ‘‘soften’’ them. Although this is a subjective impression, a possible scientific explanation is the following: the median will not affect any monotonically increasing signal that increases right across the whole median filter neighborhood, so there is clearly no blurring at this level. However, if any ‘‘texture’’ appears within the neighborhood, i.e., any high-frequency signal components, these will appear like noise and will be suppressed. This will result in a rather ‘‘flat’’ set of image regions, and in this sense the image will appear ‘‘softened.’’ So large-scale edges will be unaffected but any fine alternating edges will be smoothed out of existence. Indeed, there is a general tendency for signals that have been processed by a median filter to have ‘‘runs’’ of pixels of identical intensity. This is illustrated by the following case in which all alternating components have been eliminated by processing with a threeelement 1D median filter2: input : output :
00010111212235445565 00001111122234445555
In a digital lattice of pixels, the median filter holds several surprises, a basic one being that an exactly alternating signal will repeatedly be inverted by multiple application of a three-element median filter: input : output 1 : output 2 :
01010101010101010101 00101010101010101011 00010101010101010111
2 In this and subsequent examples, the problem of neighborhoods that are partially outside the available input region is handled by assuming that the next element is equal to the immediately adjacent element within the neighborhood—a commonly used procedure.
100
E. R. DAVIES
Perhaps more relevant to the main thrust of this article is the fact that not only do median filters eliminate impulse noise highly effectively, but in the process they sometimes introduce shifts—as indicated by the following rather simple 1D example: 00000001011111111111 00000000111111111111
input : output :
In 2D, this would clearly cause a bump to appear on the boundary, adjacent to the position of the original noise spike. This effect has caused worries in the image processing community (Yang and Huang, 1981; Bovik et al., 1987), these worries being exacerbated by the difficulty of mathematically analyzing the properties of the highly nonlinear median operation. However, in this article, we are less worried by the possibility of shifts caused by noise than those that are intrinsic to the median filter. The following example serves to illustrate the problem, and at the same time acts as an existence theorem for the work described in subsequent sections of this article: 0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 1 1 1
0 0 1 1 1
0 0 1 1 1
0 0 1 1 1
!
0 0 0 0 0
0 0 0 0 0
0 0 0 0 0
0 0 0 1 1
0 0 1 1 1
0 0 1 1 1
0 0 1 1 1
Interestingly, this leads to an idea for a median-based corner detector, which is discussed in more detail in Section X (see also Davies, 1997a). A final point worth noting is that in situations of unvarying underlying signal, it is mathematically provable (Davies, 1997a) that the mean is optimal for coping with Gaussian noise, and that the median is optimal for coping with double exponential noise (i.e., noise subject to a distribution of the form expðjrjÞ). As impulse noise is expected to appear in the wings of the distribution rather than near its center, and the double exponential type of distribution has wider wings than a Gaussian; this again helps to clarify the improved capability of the median filter for coping with impulse noise.
B. Mode Filters Although the discussion in the previous section suggests that the median filter will generally be superior to the mean filter, it also leads to the question of whether the mode filter might have even better characteristics. Indeed, an obvious argument is that the mode is the highest point of any
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
101
distribution and represents the most probable value in that distribution. This seems to imply that the mode filter ought to be closer to optimal than the median filter. In fact, tests of this idea (Davies, 1988a) showed that the mode filter is closer to optimal, but only in a particular sense: that the mode is the most likely intensity value of the signal within a neighborhood, but replacing the value at the center of the neighborhood by the mode value tends to widen intensity plateaus in image space. On the other hand, if two intensity plateaus, one dark and one light, are adjacent to each other, the tendency to widen will tend to make the border between them narrower. Clearly, the mode filter is not acting as a pure noise removal filter, and the best way of describing its operation is as an enhancement operator, which tends to make edges crisper by widening adjacent plateau regions. These arguments show that simple analyses based on verbal descriptions are insufficient, as they tend to lead to the idea that the mode filter will generally be superior to the median filter, whereas these filters are actually quite different in what they are able to do. We also note in passing that the mode filter will generally be quite good at eliminating impulse noise, assuming the mode is not in the wings of the local intensity distribution. In fact, it will not be optimal in this respect: on the other hand, the median is optimal in the sense that it can cope with 50 % impulse noise at both ends of the distribution before it starts giving totally erroneous results. For further in-depth analysis of, and interesting insights into, the properties of mode filters, see Griffin (2000). Before leaving this topic, some remarks about the implementation and use of mode filters are appropriate. First, the mode filter has been used relatively little in image processing (see, for example, Coleman and Andrews, 1979; Davies, 1992c; Evans and Nixon, 1995; Griffin, 1997). Part of the reason for this must reside in the fact that the filter is difficult to implement because the sparsity of the local intensity distribution in a small neighborhood makes it difficult to define the mode: indeed, the obvious mode—the highest point in the distribution—may not be statistically significant. Hence special algorithms are required to locate the most accurate ‘‘underlying’’ mode. This problem was tackled by Davies (1988a), and a reasonably accurate and effective solution was obtained by using the median to locate the part of the local intensity distribution corresponding to any minor mode, which could then be eliminated by truncating the distribution. In fact, all the tests on mode filters described in Section VI relate to application of this ‘‘truncated median’’ filter: however, the predicted properties of the mode, such as capability for enhancement and moderate resistance to noise, have been found to apply to the truncated median filter (Davies, 1988a).
102
E. R. DAVIES
Figure 1. Effect of applying a morphological erosion operation to a binary object. The region inside the outer boundary is the original object. The inner dark shaded regions constitute the processed image: note that the erosion operation has broken the single original object into two smaller objects.
C. Morphological Filters The early morphological filters, which operate on binary images, include two particularly important operations—erosion and dilation. In fact, these operations have isotropic and directional variants, but here we shall concentrate on the former. The properties of these operations are defined in terms of structuring elements,3 which in the isotropic case amount to circles of specified radius b: applying such a circle as a structuring element for erosion leads to all objects in the image being eroded in all directions through the distance b (Fig. 1). Naturally, this means that at some points on the boundary, where the object is quite thin, the object will be eroded away completely: this may lead to the object being broken into several parts. Similarly, dilating an object using the same circular structuring element B, it will be expanded in all directions through the distance b, and this may mean that the object will become joined to another part of itself or to another object; it can also mean that any holes or inlets are filled in. Thus erosion and dilation can cause quite serious modifications to object shapes. However, morphology takes these ideas further in that it permits such operations to be used in sequence and combined into more useful operations (Haralick et al., 1987; Bangham and Marshall, 1998). One such operation is closing; another is opening. Closing is defined as dilation by B followed by erosion by B, and opening employs the same two operations in the reverse order. The general tendency is for such well-matched operations to cancel out, the reason being that expanding an object in all directions through a distance b and then contracting it through the same distance should be a null operation. However, although it is null for a shape such as a large circle, for a C shape the ends of the C might become joined by the dilation (or the C could even be filled in), and once this has happened the 3
In morphology the modifying element is known as a ‘‘structuring element’’: the term ‘‘kernel’’ is reserved for convolution.
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
103
erosion will be unable to reverse the situation. We can summarize the overall situation as follows: ðA BÞ B A ðA BÞ B
ð1Þ
where A is the original shape, and ; represent, respectively, dilation and erosion; or alternatively, expressing the combined operations as closing () and opening ( ): ABAA B
ð2Þ
When the opening operation is applied to an image containing convex objects, all objects that are narrower than a certain critical width will disappear, and the remainder will be returned to something approaching their original size and shape. Thus the operator can be envisaged as a filter that filters objects by size, eliminating the small ones. This type of operation can be repeated with opening operations using circular structuring elements of various radii b. This makes it clear that shapes with any range of widths can be eliminated, and any others retained, as long of additional image subtraction operations are permitted (in fact, the required operations are set subtractions in the image space). Clearly, then, morphological operations can be used for filtering. So far we have investigated morphological operations only in binary images. However, there are several ways in which these operations and results can be extended to gray-scale images: perhaps the most obvious is the case when a structuring element B is applied independently to all gray levels of a gray-scale image. In that case it may be shown that dilation corresponds to a local maximum operation, and erosion to a local minimum operation, in each case using a neighborhood of size equal to that of B. Next, it is of some interest that gray-scale edge detection algorithms can be designed by methods such as subtraction of eroded gray-scale images from dilated gray-scale images, the latter being generally larger if image intensity is measured in the positive z-direction: E ¼ ðA BÞ ðA BÞ
ð3Þ
It will also be of interest that there is a whole science of image processing that is based on morphological set operations instead of convolution operations, and that these are very much dual approaches that are able to solve similar sorts of problem, each in its own way, albeit with characteristics that are minutely different. Optimality depends on the individual problem and the criteria that are adopted for judging optimality. Note that the important theorem about matched filtering giving the optimum SNR under conditions of white noise (Davies, 1993) will not be valid for morphological set operations, and it is unknown what alternative
104
E. R. DAVIES
analogous form could apply for such operations. Thus it seems that many feature detectors would be closer to optimal when based on convolution rather than morphology. On the other hand, it is always useful to have alternative tools and algorithms, not least so that increased adaptability can be achieved as the data and the task demand. There are many other morphological filters that involve combinations of these basic operations. Further morphological filters may be designed by generalizing the ways in which gray-scale operations may be made, and by incorporating conditional processes (Haralick and Shapiro, 1992). D. In-Depth Study of Median Filters This article cannot do justice to all the sorts of filter that have been developed, noise-suppression, mode, rank-order, morphological, or otherwise. Instead it will be necessary to concentrate on a few—particularly rankorder filters and filters that are able to remove noise from images—and to explore their properties quite closely. Some of the original motivation for this line of work was the observation that median filters often tend to be used more because they are known to exist (to some extent having a wellknown name tends to encourage use) than because they provide provably optimal means for solving imaging problems; a further part of the motivation arises because they are known not to cause the blurring associated with mean filters, and this leads to the supposition that they do not shift edges—even though quite simple models (see Section II.A) show that this cannot be the case. Accordingly, further in-depth investigation of the properties of median filters and other associated filters seemed to be called for. In particular, (1) other rank-order filters and (2) other noise removal filters needed to be investigated. Table 1 indicates in which sections TABLE 1 The Filters Dealt with in This Article Filter Mean filter Median filter Mode filter Gaussian filter Hybrid median filter Minimum filter Maximum filter Rank-order filters Morphological filters
Described Section Section Section Section Section Section Section Section Section
II.A II.A II.B II.A III.F II.C II.C II II.C
Shifts Analyzed Section V Sections III, IV Section VI Section V Section III.F Section VII Section VII Sections VII, VIII Sections VII, IX
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
105
of the article different filters are first described and also where the shifts and distortions they give rise to are analyzed.
III. Shifts Produced by Median Filters in Continuous Images A. Theory of Edge Shifts Produced by Median Filters in Continuous Binary Images We start by considering a continuous image (i.e., a nondiscrete lattice), assuming first (1) that the image is binary, (2) that the neighborhoods are exactly circular, and (3) that the image is noise free. To proceed we notice that binary edges have symmetrical cross sections, whereas straight edges extend this symmetry into 2D: hence applying a median filter in a (symmetrical) circular neighborhood cannot pull a straight edge to one side or the other. Now consider what happens when the filter is applied to an edge that is not straight. If, for example, the edge is circular, the local intensity distribution will contain two peaks whose relative sizes will vary with the precise position of the neighborhood (Fig. 2). At some position the sizes of the two peaks will be identical. This happens when the center of the neighborhood is at a unique distance from the center of a circular object: this is the position at which the output of the median filter changes from dark to light (or vice versa). It is clear that the median filter produces an inward shift toward the center of a circular object (or the center of curvature), whether the object is dark on a light background or light on a dark background. Next suppose that the edge is irregular with several ‘‘bumps’’ (i.e., prominences or indentations) within the filter neighborhood: clearly, the filter will now tend to average out the bumps and straighten the edge, since it acts in such a way as to form a boundary on which the amounts of dark and light within the neighborhood are equalized (Fig. 3). This means that the edge will be locally biased but only by a reduced amount, since the various bumps will tend to pull the final edge in opposite directions. On the other hand, if there is one gross bump within the neighborhood—i.e., if the curvature has the same sign and is roughly constant at all points on the edge within the filter neighborhood—then all these parts of the edge will act in consort and it will be pulled sideways a significant amount by the filter. Thus a circular section of the boundary constitutes a ‘‘worstcase’’ situation, for which the filter produces the largest bias in the position of the edge. It is clearly worth finding the size of the worst-case shift and for this reason we concentrate attention below on circular objects, in
106
E. R. DAVIES
Figure 2. Variation in local intensity distribution with position of neighborhood. (a) Neighborhood of radius a overlapping a dark circular object of radius b. (b)–(d) Intensity distributions f when the separations of the centers are, respectively, less than, equal to, or greater than the center separation d for which the object bisects the area of the neighborhood. From Davies (1989).
the knowledge that all other shapes will give less serious shifts and distortions. The worst-case calculation is a matter of elementary geometry: we need to find at what distance from the center of a circular object (of radius b) the area of a circular neighborhood (of radius a) is bisected by the object boundary. One way of estimating this is to determine the integrated area within the neighborhood that lies outside the object boundary (Davies, 1989). We start by taking the boundary as being a circle centered at (b, 0) and passing through the origin (Fig. 4):
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
107
Figure 3. Edge smoothing property of the median filter. (a) Original 128 128 pixel image with 6-bit gray scale. (b) Effect of median filter smoothing of irregularities, in particular those around the boundaries, using a 21-element filter operating within a 5 5 neighborhood. Notice how the threads on the screws are virtually eliminated, although detail larger in scale than half the filter area is preserved. (c) Effect of 2LHþ ‘‘detail-preserving’’ filter. From Davies (1989).
ðx bÞ2 þ y2 ¼ b2
The integrated area outside the boundary is now Z a Z a 1 2 1 2 A ¼ a þ xdy ¼ a þ b ðb2 y2 Þ1=2 dy 2 2 a a
ð4Þ ð5Þ
whereas that inside the boundary is a2 A. To equalize these areas we need to institute a boundary shift of D (though this is necessarily an
108
E. R. DAVIES
Figure 4. Geometry for calculating edge shifts. The large circle (radius b) is the boundary of the object. The small circle (radius a) is the neighborhood. The shaded portion corresponds to the integral in Eq. (5).
approximation, as the shift will have a small affect on the intersection points, so the integration limits will no longer be exactly a and a). This gives the following equation for D: A 2aD ða2 AÞ þ 2aD
ð6Þ
1 ; 2aD A a2 2
ð7Þ
giving D
1 2a
Z
a
a
b ðb2 y2 Þ1=2 dy
ð8Þ
To perform the integration, make the substitution This leads to the formula
y ¼ b sin u
ð9Þ
b b2 D b ð1 a2 =b2 Þ1=2 sin1 ða=bÞ ð10Þ 2a 2 To proceed, we use the power series expansion of the inverse sine function: 1 3 sin1 ¼ þ 3 þ 5 þ 6 40
ð11Þ
whence D or
a2 a4 þ 6b 40b3
ð12Þ
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
109
Figure 5. Geometry for calculating neighborhood and object overlap. From Davies (1989).
1 1 ð13Þ D a2 þ 3 a4 6 40 ¼ 1=b being the local curvature. This equation simplifies to the following form for low values of : 1 ð14Þ D a2 6 Although Eq. (14) is often quite useful, Eq. (13) is especially valuable as it turns out to be far more accurate over the range 0 a b than might be expected (see below). However, to find the exact situation, we need a more rigorous theory. This is attained as follows (Davies, 1989). From Figure 5 the area of the sector of angle 2 is b2 , whereas the area of the triangle of angle 2 is b2 sin cos . Hence the area of the segment shown shaded is B ¼ b2 ð sin cos Þ
ð15Þ
Making a similar calculation of the area A of a circular segment of radius a and angle 2 , the area of overlap (Fig. 5) between the circular neighborhood of radius a and the circular object of radius b may be deduced as C ¼AþB
ð16Þ
For a median filter this is equal to a2 =2. Hence F ¼ a2 ð sin cos Þ þ b2 ð sin cos Þ a2 =2 ¼ 0
ð17Þ
where a2 ¼ b2 þ d 2 2bd cos
ð18Þ
110
E. R. DAVIES
and b2 ¼ a2 þ d 2 2ad cos
ð19Þ
To solve this set of equations, we take a given value of d, deduce values of
and , calculate the value of F, and then adjust the value of d until F ¼ 0. Since d is the modified value of b obtained after filtering, the shift produced by the filtering process is D¼bd
ð20Þ
The results of doing this computation numerically have been found by Davies (1989) and are shown in Table 2. As expected, D ! 0 as b ! 1 or as a ! 0. Conversely, the shift becomes very largepas ffiffiffi a first approaches and then exceeds b. Note, however, that when a > 2b the object is ignored, being small enough to be regarded as irrelevant noise by the filter: beyond this point it has no effect on the finalp image. The maximum edge shift before ffiffiffi the object finally disappears is ð2 2Þb 0:586b. B. Extension to Continuous Gray-Scale Images To extend these results to gray-scale images, first consider the effect of applying a median filter near a smooth step edge in 1D. Here the median filter gives zero shift, since for equal distances from the center to either end of the neighborhood there are equal numbers of higher and lower intensity values and hence equal areas under the corresponding portions of the intensity histogram. Clearly this is always valid where the intensity increases monotonically from one end of the neighborhood to the other—a property first pointed out by Gallagher and Wise (1981) [for later discussions on related ‘‘root’’ (invariance) properties of signals under median filtering, see Fitch et al., 1985; Heinonen and Neuvo, 1987]. Next, it is clear that for 2D images, the situation is again unchanged in the vicinity of a straight edge, since the situation remains highly symmetrical. Hence the median filter gives zero shift, as in the binary case. For curved edges, it again turns out that circular boundaries constitute a worst case that should be considered carefully. However, gray-scale edges are unlike binary edges in that they have finite slope. This means that it is necessary to take account of the exact form of the intensity function within the neighborhood. When boundaries are roughly circular, contours of constant intensity often appear as in Figure 6. To find how a median filter acts we merely need to identify the contour of median intensity (in 2D the median intensity value labels a whole contour), which divides the area of the neighborhood into
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
111
TABLE 2 Estimated Edge Shifts for Filtering Circles in a Continuuma a/b
D/b
D0 =b
0.000 0.050 0.100 0.150 0.200 0.250 0.300 0.350 0.400 0.450 0.500 0.550 0.600 0.650 0.700 0.750 0.800 0.850 0.900 0.950 1.000 1.050 1.100 1.150 1.200 1.250 1.300 1.350 1.400 1.414
0.000 0.000 0.002 0.004 0.007 0.011 0.015 0.021 0.027 0.035 0.043 0.053 0.063 0.075 0.088 0.102 0.117 0.133 0.151 0.171 0.192 0.214 0.238 0.264 0.292 0.321 0.353 0.387 0.423 0.433
0.000 0.000 0.002 0.004 0.007 0.010 0.015 0.021 0.027 0.035 0.043 0.052 0.063 0.074 0.087 0.101 0.116 0.132 0.151 0.170 0.192 0.216 0.242 0.272 0.305 0.342 0.387 0.443 0.528 0.586
a
From Davies (1989).
two equal parts. The geometry of the situation is identical to that already examined in Section III.A: the main difference here is that for every position of the neighborhood, there is a corresponding median contour with its own particular value of shift depending on the curvature. Intriguingly, the formulas already deduced may immediately be applied for calculating the shift for each contour. Figure 6 shows an idealized case in which the contours of constant intensity have similar curvature, so that they are all moved inward by similar amounts. This means that to a first approximation, the edges of the object retain their cross-sectional profile as it becomes smaller.
112
E. R. DAVIES
Figure 6. Contours of constant intensity on the edge of a large circular object, as seen within a small circular neighborhood. From Davies (1989).
We next consider the effects of noise. For simplicity we assume that the noise is additive and of symmetrical (nonskew) intensity distribution: this is valid for Gaussian noise and is also likely to be true for many types of impulse noise. Now recall that the median contour divides the neighborhood into two equal parts. Hence, adding noise of symmetrical intensity distribution will on average not change the area on either side of the original median contour: this means that noise will not on average cause edges to shift any differently as a result of applying the median filter—i.e., the shifts of edges caused by noise or by applying a median filter are, to first order, additive. In particular, noise does not affect the general conclusions presented above concerning the shifts of edges introduced by median filters. Though specific experiments have not been performed in this work to introduce noise and check this result quantitatively, it is generally supported by observations on real images containing noise. This section has generalized the results of Section III.A to cover grayscale images. It has also shown that the effects of noise should not materially affect the conclusion that median filters shift edges inward toward local centers of curvature, the worst-case situation arising for circular objects. Finally, it has shown that straight edges remain unshifted for any symmetrically shaped neighborhood. However, the detailed shift calculations (Section III.A) assumed a circular neighborhood: though this is not general, it will produce an ideal isotropic response that could not be guaranteed for any other shape of neighborhood. This fact justifies our concentration on circular neighborhoods in the above analysis. For neighborhoods of other shapes it seems simplest to confirm experimentally that the theory makes substantially correct predictions of edge shifts.
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
113
Figure 7. Edge shift curves for various sizes of neighborhood. This diagram indicates how the edge shift curves would be expected to change as p moves between the two limiting values of 1 and infinity: (a) p ¼ 1; (b) p ¼ 3; (c) p ¼ 5; (d) p ¼ 1. From Davies (1989).
C. Extension to Discrete Neighborhoods In the previous two sections we have developed theory showing how median filters introduce edge shifts in a continuous space. Here we consider the effect of applying median filters in discrete lattices of pixels: specifically, median filters need to be applied in square neighborhoods of p p pixels. In the continuous case covered earlier, p was essentially infinite. Unfortunately, it is difficult to see how to extend the theory accurately to typical cases such as 3 3 and 5 5 neighborhoods. However, it is trivial to cover the case of p ¼ 1, since in this case the median filter leaves the image unchanged. For intermediate values of p we expect that edge shifts will fall between these two cases as upper and lower bounds, and indeed that there will be a steady progression from the one to the other bound as p varies (Fig. 7). As we shall see below, this situation is generally confirmed by the experimental data. D. Experimental Results for Discrete Binary Images In this section we present experimental results for binary images. For reasons outlined earlier, we concentrate on the worst-case situation of small circular objects. Clearly, in a binary image it is only possible to approximate to (filled) circles, and in these tests radii ranged from 0.5 to about 9 pixels. It seemed sufficient to perform experiments using standard median filters in a 3 3 (square) neighborhood. In these experiments, a problem arose since
114
E. R. DAVIES
Figure 8. Binary circles before and after filtering. (a) Set of binary circles of radii ranging from 0.5 to 9 pixels. (b) Result of applying a median filter operating in a 3 3 (square) neighborhood. From Davies (1989).
the effective radius of such a neighborhood is not known accurately: for the present purpose we assume it to be such as to equalize the areas of discrete and idealized neighborhoods. Hence we took the radius as ð9=Þ1=2 ¼ 1:693 pixels. The results are shown in Figures 8 and 9. When compared with the results of Table 2 they show some interesting features. In particular: 1. There is a limited number of possible radius values. 2. For a large proportion of radius values no change in radius occurs on applying a median filter. 3. For very small radius values the circle disappears completely. 4. For other small radius values the circle becomes much smaller. 5. For certain isolated larger radius values there is a reduction in circle size, but (a) the number of instances becomes rarer as radius increases, and (b) the reduction in circle size becomes smaller as radius increases. 6. In general, repeated application of a median filter to circles above a certain critical size must lead to a small reduction in size followed by stability, whereas repeated application to circles below the critical size must lead to their elimination: the critical radius is 2:5 pixels. Ignoring result (1) as obvious, we interpret these results as follows. When applied to discrete binary images, the median filter has the properties predicted in Section III.A. However, median filters in small neighborhoods
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
115
Figure 9. Edge shifts for 3 3 median filter applied to binary circles. In this graph the plots represent the experimental results and the continuous curve is derived from theory (see Section III.A). It is also of interest to compare the experimental plots with the model of Section III.E (see lower curve in Fig. 10). From Davies (1989).
do not have the resolution to detect accurately the curvature of large circles: hence these either become resistant to any change in their size and shape or seem rather unstable and ready to shed their outermost pixels. The latter situation clearly happens for those circles whose boundaries are irregular, since they have some relatively sharp corners that are eliminated by the median filter. Such corners become increasingly rare as circle size increases (see Fig. 8). We return to this point in Section III.E, with the aim of building a more realistic model of the action of the median filter in the discrete case. The stability properties we have observed are related to the root behavior noticed by other workers when median filters are applied repeatedly (mainly to one-dimensional signals) until no further change occurs (Gallagher and Wise, 1981; Fitch et al., 1985; Heinonen and Neuvo, 1987). However, we are here less interested in root behavior than in mean edge shifts, for a single application of a median filter, as curvatures vary. Hence it is instructive to average out the rather random responses that occur for various radii b. It is seen that the resulting curve (Fig. 9) is similar in shape to the theoretical curve of Section III.A, but lies below it and indeed between it and the identity curve corresponding to the null case p ¼ 1, as predicted in Section III.C: in no case does the stability effect cause the predicted change in size of an object to be reversed in sign, though it is frequently reduced to zero.
116
E. R. DAVIES
E. Experimental Results for Discrete Gray-Scale Images The results of the previous section immediately suggest using better approximations to circles, with the jagged binary edges interpolated by appropriate gray-scale values. For each size of circle this was achieved by permitting the intensity to vary linearly from black to white over a range of 1 pixel, and then smoothing the resulting shapes using the following wellknown convolution mask: 2 3 1 2 1 1 4 2 4 25 16 1 2 1
This procedure was successful in giving a realistic approximation to a smoothed step-edge. (Note, however, that other edge models, such as edges that vary linearly over 1 or 2 pixels, seemed to give essentially the same edge shifts—i.e., the edge shift behavior was relatively insensitive to the exact type of edge model chosen.) In the experiments described here, circular objects again varied between 1 and 9 pixels, and pixel intensities were permitted to vary over an 8-bit range. The original and the modified radii were measured by taking the integrated intensities over the circle region and deducing the radii, this approach being used because it overcame problems due to irregularities in circle boundaries. Finally, the experiment was in this case performed for 3 3 and 5 5 neighborhoods, though in the latter case an attempt was made at approximating to the more ideal circular neighborhood by omitting the four corner pixels, and using the pattern:
[Note that in some earlier work on edge detection operators, it was found that this pattern not only required less computation than a full 5 5 operator, but also gave increased edge orientation accuracy (Davies, 1984).] The results of these experiments are shown in Figures 10 and 11. In these cases certain features that were present in the previous experiment on binary circles are now absent. In particular, the variation of edge shift D with initial radius b is very smooth, and there is no evidence of stability of large circles against the median filter: i.e., the median filter is able to reduce
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
117
Figure 10. Edge shifts for 3 3 median filter applied to gray-scale circles. In this graph the plots represent the experimental results and the upper continuous curve is derived from theory (see Section III.A). The lower continuous curve is derived from the model of Section III.E. From Davies (1989).
the size of all circles by an amount that tends to zero much as expected as radius increases. It seems important to analyze this case reasonably thoroughly, as it constitutes a common practical situation. Hence it was compared quantitatively with the theory of Section III.A. Again the problem arose that the effective radius of a discrete median filter in a 3 3 or 5 5 neighborhood is not known accurately: adopting the equal areas strategy of the previous section, we obtain the respective radii as ð9=Þ1=2 ¼ 1:693 and ð21=Þ1=2 ¼ 2:585. The resulting theoretical graphs are shown in Figures 10 and 11. These show that there are some odd results for very low values of circle radius. On the whole these can be explained rather neatly by appealing to the d/a vs. b/a curve (Fig. 12). Here we see that the effect of having an edge profile that varies from black to white over several pixels is to bring in a range of radius values: hence it is necessary to average the graph over a suitable range of values. As a result the experimental curve goes smoothly down to zero below the critical radius, and in the 5 5 case cuts slightly across the theoretical upper bound curve above the critical radius. Other minor inaccuracies can be explained as due to the particular (noncircular) shape of the filter neighborhood and remanent stability effects. Finally, the expected progression from p ¼ 1 through p ¼ 3 and p ¼ 5 to p ¼ 1 is obeyed, though the intermediate edge shift curves both appear to
118
E. R. DAVIES
Figure 11. Edge shifts for 5 5 median filter applied to gray-scale circles. The upper set of plots represent the experimental results and the upper continuous curve is derived from theory (see Section III.A). The lower continuous curve is derived from the model of Section III.E. The lower set of plots represents the much reduced shifts obtained with the detailpreserving type of filter. From Davies (1989).
approach zero more rapidly than would have been predicted on the basis of simple interpolation. We conclude that discrete neighborhoods impart an additional stability on the edge positions. This can be explained in general terms as follows. For a 3 3 median filter, there are essentially only 8 positions around the boundary of a large object that the median filter can erode. [This is approximately correct for binary images where large ‘‘circles’’ tend to be octagonal (see Fig. 8); however, it must also be approximately correct for gray-level images, since we can consider each gray-level outline as a separate binary image that is then eroded on its own.) Hence the effectiveness of the median filter at eroding large objects will in principle fall off by an additional factor proportional to a/b (i.e., relative to the upper bound given by the theory of Section III.A): note, however, that this additional factor will not apply for small objects, so the result can never be larger than the upper bound value of D. For the 5 5 median filter, which is more sensitive to the curvature of large objects, the same general model seems to apply, but with a different constant of proportionality. However, to fully explain the observed variations we need to model the p-variation. Following Section III.C and considering vertical lines through the D/a vs. b/a graph (Fig. 7), we try modeling the vertical variation as (1=a0 1=a), since this quantity
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
119
Figure 12. Method of averaging required for small circles. This diagram shows how averaging over the various contour radii appearing within the neighborhood should be performed. The two main effects are (1) raising of the circle size (reduction of the edge shift) for circular objects below the critical size; (2) lowering of the circle size (increase in the edge shift) for objects above the critical size. From Davies (1989).
automatically approaches a constant value as a tends to infinity, and approaches zero as a approaches a0 . [Here a0 is the effective radius of a 1 1 neighborhood, assumed here to be equal to ð1=Þ1=2 ¼ 0:564.] Taking both variations into account gives the overall model4: D0 ¼ min½D; cDð1=a0 1=aÞða=bÞ ¼ min½D; cDða=a0 1Þ=b
ð21Þ
where the constant c has to be found empirically. As will be clear from Figures 10 and 11, this formula gives very good agreement with the observed results for 3 3 and 5 5 neighborhoods, when c is made equal to 1.0. Thus it is now known with fair accuracy (even if partly on a semiempirical basis) how the upper bound form (Section III.A) adapts to a discrete lattice. At this stage there can be no doubt that the median filter gives definite and measurable edge shifts via a specific type of neighborhood averaging process. Overall, for gray-scale images the shifts predicted by this theory agree with experimental shifts within approximately 10 % for a large range of circle sizes in a discrete lattice (see Figs. 10 and 11). The agreement is less perfect for binary images, since circles of certain sizes show stability effects (akin to median root behavior): these effects tend to average out for gray-scale images, 4 It will be seen in Section IV that the additional factor 1/b that appears here presages a rigorous derivation of an overall variation proportional to 2 at low values of curvature ¼ 1=b: this was not anticipated in Davies (1989).
120
E. R. DAVIES
owing to the presence of many contours of different sizes at different gray levels. Overall, it appears that the edge shifts obtained with median filters are now quite well understood. Figures 3 and 13 give some indication of the magnitudes of these shifts in practical situations. Note that once image detail such as a small hole or screw thread has been eliminated by a filter, it is not possible to apply any edge shift correction formula to recover it, although for larger features such formulas are useful for deducing true edge positions. F. Edge Shifts Arising with Hybrid Median Filters Although median filters preserve edges in digital images, they are also known to remove fine image detail such as lines. For example, 3 3 median filters remove lines one pixel wide, and 5 5 median filters remove lines two pixels wide. In many applications such as remote sensing and X-ray imaging this is exceedingly important and efforts have been made to develop filters that overcome the problem. In 1987 Nieminen et al. reported a new class of ‘‘detail preserving’’ filters: these employ linear subfilters whose outputs are combined by median operations. There is a great variety of such filters, employing different subfilter shapes and having the possibility of several layers of median operations. Hence it is not possible to describe them fully here in the space available. Although these filters are aimed particularly at retention of line detail, and are readily understood in this context, they turn out to have some corner-preserving properties and to be resistant to the edge shifts that arise when the curvature is nonzero. Perhaps the best of the filters in the new class, from the point of view of preserving edge position, is the two-level ‘‘bidirectional’’ linear-median hybrid filter termed ‘‘2LHþ’’ (Nieminen et al., 1987). Its operation in a 5 5 neighborhood may be illustrated as follows. It employs the subfilters AI in the 5 5 region: E
D
C
E
D
C
F
F
A
B
B
G
H
I
G
H
I
pixels marked as being in the same subfilter having their intensities averaged, and dashed pixels being ignored. Nonlinear filtering then proceeds using two levels of median filtering, the final center-pixel intensity being taken as A0 ¼ med½A; medðA; B; D; F ; HÞ; medðA; C; E; G; IÞ
ð22Þ
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
121
Figure 13. Circular holes in metal objects before and after filtering. (a) Original 128 128 pixel image with 6-bit gray scale. (b) 5 5 median-filtered image: the diminution in size of the holes is clearly visible and such distortions would have to be corrected for when taking measurements from real filtered images of this type. (c) Result of using a detail-preserving filter: some distortions are present although the overall result is much better than in (b). From Davies (1989).
122
E. R. DAVIES
We here ignore the line-preserving properties of this filter and concentrate on its corner-preserving, low-edge-shift characteristics. It is quite easy to see that the 5 5 regions 0 0 0 0 0
0 0 0 0 0
0 0 1 1 1
0 0 1 1 1
0 0 1 1 1
0 0 0 0 1
0 0 0 1 1
0 0 1 1 1
0 0 0 1 1
0 0 0 0 1
are preserved by this filter, although these examples represent limiting cases that could be disrupted by minor amounts of noise or slight changes of orientation. Thus the filter seems guaranteed to preserve corners only if the internal angle is greater than 135 . This figure should be compared with the 180 obtained using similar arguments for the normal median filter in 5 5 regions such as 0 0 0 0 0
0 0 0 0 0
0 0 1 1 1
1 1 1 1 1
1 1 1 1 1
Figure 11 shows plots obtained with this filter under the same conditions as for the 5 5 median filters. It always gives at least a 4-fold improvement (reduction) in edge shift over that for the median filter, and this performance improves with increasing radius of curvature b until there is zero shift for b > 4 (note that b ¼ 4 is approximately the figure that would be expected from the corner angle of 135 noted above, within a 5 5 neighborhood). Hence such detail-preserving filters improve the situation dramatically but do not completely overcome the underlying problem described earlier. In addition, this improvement may not have been obtained without cost, since in some cases the filter seems to insert structure where none exists (Davies, 1989). The result is to cast some doubt on the usefulness of this type of filter in all possible situations. Nevertheless, its effect on real images appears to be generally very good (see Figs. 3c and 13c). IV. Shifts Produced by Median Filters in Digital Images A. Using a Discrete Model to Explain Median Shifts To produce a discrete model we need to recognise explicitly the positions of the pixels within an p p neighborhood. We approximate by assuming
GEOMETRIC DISTORTIONS PRODUCED BY IMAGE PROCESSING FILTERS
123
Figure 14. Idealized intensity functions for calculation of contour shifts. This figure shows idealized intensity variations within a circular neighborhood C of radius a. (a) Top: circular contours of radius b; bottom: linear variation in intensity. (b) Single step edge with a circular boundary of radius b. From Davies (1999b).
that the intensity of any pixel is the mean intensity over the whole pixel and is represented by a sample positioned at the center of the pixel. In addition, we take the underlying analogue intensity variation to have contours of curvature , as shown in Figure 14a. Following what happened in the continuum case, it will not matter whether the contours of constant intensity are those of a step edge or those of a slowly varying slant edge: it is what happens at the median contour that determines the shift that arises. The starting point is that zero shift occurs for ¼ 0. Next, if is even minutely greater than zero, the centre pixel will not necessarily be the median pixel. Consider first a situation when the circular median intensity contour does not pass through the center of the central pixel but passes symmetrically through the centers of two other pixels as shown in Figure 15. (With no loss of generality at this stage of the calculation, the two pixels are assumed to lie along the same vertical line.) If the separation of the two pixels is 2, then the geometry of a circle of radius b leads to 2 ¼ D ð2b DÞ
ð23Þ
; D 2 2bD þ 2 ¼ 0
ð24Þ
; D ¼ b ½b2 2 1=2
ð25Þ
where we have to take the minus sign. Approximating for small and large b ð¼ 1 Þ leads to the result
124
E. R. DAVIES
Figure 15. Geometry for calculation of shift when the median contour passes through the centers of two pixels.
h i1=2 D ¼ b b 1