This action might not be possible to undo. Are you sure you want to continue?
Anisotropic Diﬀusion
in Image Processing
Joachim Weickert
B.G. Teubner Stuttgart
Anisotropic Diﬀusion
in Image Processing
Joachim Weickert
Department of Computer Science
University of Copenhagen
Copenhagen, Denmark
B.G. Teubner Stuttgart 1998
Dr. rer. nat. Joachim Weickert
Born in 1965 in Ludwigshafen/Germany. Studies in mathematics, physics and com
puter science at the University of Kaiserslautern. 1987 B.Sc. in physics and indus
trial mathematics. 1991 M.Sc. in industrial mathematics. 1996 Ph.D. in mathe
matics. Postdoctoral researcher at the Image Sciences Institute at Utrecht Uni
versity from 2/96 to 3/97. Since then visiting assistant research professor at the
Department of Computer Science, Copenhagen University.
The cover image shows a thresholded nonwoven fabric image which was processed
by applying a coherenceenhancing anisotropic diﬀusion ﬁlter (see Section 5.2 for
more details). The goal was to visualize the quality relevant adjacent ﬁbre struc
tures, socalled stripes. The displayed equations describe the basic structure of
nonlinear diﬀusion ﬁltering in the continuous, semidiscrete, and fully discrete set
ting. Their theoretical foundations are treated in Chapters 2–4.
c ( Copyright 2008 by Joachim Weickert.
All rights reserved. No part of this book may be reproduced by any means, or
transmitted, or translated into a machine language without the written permission
of the author.
This book had been published by B. G. Teubner (Stuttgart) in 1998 and went out
of print in 2001. The copyright has been returned to the author in 2008. In the
current version a few typos and other errors have been corrected.
To my parents Gerda and Norbert
Preface
Partial diﬀerential equations (PDEs) have led to an entire new ﬁeld in image
processing and computer vision. Hundreds of publications have appeared in the last
decade, and PDEbased methods have played a central role at several conferences
and workshops.
The success of these techniques is not really surprising, since PDEs have proved
their usefulness in areas such as physics and engineering sciences for a very long
time. In image processing and computer vision, they oﬀer several advantages:
• Deep mathematical results with respect to wellposedness are available, such
that stable algorithms can be found. PDEbased methods are one of the
mathematically bestfounded techniques in image processing.
• They allow a reinterpretation of several classical methods under a novel uni
fying framework. This includes many wellknown techniques such as Gaussian
convolution, median ﬁltering, dilation or erosion.
• This understanding has also led to the discovery of new methods. They
can oﬀer more invariances than classical techniques, or describe novel ways
of shape simpliﬁcation, structure preserving ﬁltering, and enhancement of
coherent linelike structures.
• The PDE formulation is genuinely continuous. Thus, their approximations
aim to be independent of the underlying grid and may reveal good rotational
invariance.
PDEbased image processing techniques are mainly used for smoothing and
restoration purposes. Many evolution equations for restoring images can be de
rived as gradient descent methods for minimizing a suitable energy functional, and
the restored image is given by the steadystate of this process. Typical PDE tech
niques for image smoothing regard the original image as initial state of a parabolic
(diﬀusionlike) process, and extract ﬁltered versions from its temporal evolution.
The whole evolution can be regarded as a socalled scalespace, an embedding of
the original image into a family of subsequently simpler, more global representa
tions of it. Since this introduces a hierarchy into the image structures, one can use
a scalespace representation for extracting semantically important information.
One of the two goals of this book is to give an overview of the stateoftheart of
PDEbased methods for image enhancement and smoothing. Emphasis is put on a
v
vi PREFACE
uniﬁed description of the underlying ideas, theoretical results, numerical approxi
mations, generalizations and applications, but also historical remarks and pointers
to open questions can be found. Although being concise, this part covers a broad
spectrum: it includes for instance an early Japanese scalespace axiomatic, the
Mumford–Shah functional for image segmentation, continuousscale morphology,
active contour models and shock ﬁlters. Many references are given which point the
reader to useful original literature for a task at hand.
The second goal of this book is to present an indepth treatment of an interest
ing class of parabolic equations which may bridge the gap between scalespace and
restoration ideas: nonlinear diﬀusion ﬁlters. Methods of this type have been pro
posed for the ﬁrst time by Perona and Malik in 1987 [326]. In order to smooth an
image and to simultaneously enhance important features such as edges, they apply
a diﬀusion process whose diﬀusivity is steered by derivatives of the evolving image.
These ﬁlters are diﬃcult to analyse mathematically, as they may act locally like
a backward diﬀusion process. This gives rise to wellposedness questions. On the
other hand, nonlinear diﬀusion ﬁlters are frequently applied with very impressive
results; so there appears the need for a theoretical foundation.
We shall develop results in this direction by investigating a general class of
nonlinear diﬀusion processes. This class comprises linear diﬀusion ﬁlters as well as
spatial regularizations of the Perona–Malik process, but it also allows processes
which replace the scalar diﬀusivity by a diﬀusion tensor. Thus, the diﬀusive ﬂux
does not have to be parallel to the grey value gradient: the ﬁlters may become
anisotropic. Anisotropic diﬀusion ﬁlters can outperform isotropic ones with respect
to certain applications such as denoising of highly degraded edges or enhancing
coherent ﬂowlike images by closing interrupted onedimensional structures. In or
der to establish wellposedness and scalespace properties for this class, we shall
investigate existence, uniqueness, stability, maximum–minimum principles, Lya
punov functionals, and invariances. The proofs present mathematical results from
the nonlinear analysis of partial diﬀerential equations.
Since digital images are always sampled on a pixel grid, it is necessary to know
if the results for the continuous framework carry over to the practically relevant
discrete setting. These questions are an important topic of the present book as
well. A general characterization of semidiscrete and fully discrete ﬁlters, which
reveal similar properties as their continuous diﬀusion counterparts, is presented. It
leads to a semidiscrete and fully discrete scalespace theory for nonlinear diﬀusion
processes. Mathematically, this comes down to the study of nonlinear systems of
ordinary diﬀerential equations and the theory of nonnegative matrices.
Organization of the book. Image processing and computer vision are inter
disciplinary areas, where researchers, practitioners and students may have a very
diﬀerent scientiﬁc background and diﬀering intentions. As a consequence, I have
tried to keep this book as selfcontained as possible, and to include various aspects
PREFACE vii
such that it should contain interesting material for many readers. The prerequisites
are kept to a minimum and can be found in standard textbooks on image process
ing [163], matrix analysis [407], functional analysis [9, 58, 7], ordinary diﬀerential
equations [56, 412], partial diﬀerential equations [185] and their numerical aspects
[293, 286]. The book is organized as follows:
Chapter 1 surveys the fundamental ideas behind PDEbased smoothing and
restoration methods. This general overview sketches their theoretical properties,
numerical methods, applications and generalizations. The discussed methods in
clude linear and nonlinear diﬀusion ﬁltering, coupled diﬀusion–reaction methods,
PDE analogues of classical morphological processes, Euclidean and aﬃne invariant
curve evolutions, and total variation methods.
The subsequent three chapters explore a theoretical framework for anisotropic
diﬀusion ﬁltering. Chapter 2 presents a general model for the continuous setting
where the diﬀusion tensor depends on the structure tensor (interest operator,
secondmoment matrix), a generalization of the Gaussiansmoothed gradient al
lowing a more sophisticated description of local image structure. Existence and
uniqueness are discussed, and stability and an extremum principle are proved.
Scalespace properties are investigated with respect to invariances and information
reducing qualities resulting from associated Lyapunov functionals.
Chapter 3 establishes conditions under which comparable wellposedness and
scalespace results can be proved for the semidiscrete framework. This case takes
into account the spatial discretization which is characteristic for digital images,
but it keeps the scalespace idea of using a continuous scale parameter. It leads
to nonlinear systems of ordinary diﬀerential equations. We shall investigate under
which conditions it is possible to get consistent approximations of the continuous
anisotropic ﬁlter class which satisfy the abovementioned requirements.
In practice, scalespaces can only be calculated for a ﬁnite number of scales,
though. This corresponds to the fully discrete case which is treated in Chapter
4. The investigated discrete ﬁlter class comes down to solving linear systems of
equations which may arise from semiimplicit time discretizations of the semidis
crete ﬁlters. We shall see that many numerical schemes share typical features
with their semidiscrete counterparts, for instance wellposedness results, extremum
principles, Lyapunov functionals, and convergence to a constant steadystate. This
chapter also shows how one can design eﬃcient numerical methods which are in
accordance with the fully discrete scalespace framework and which are based on
an additive operator splitting (AOS).
Chapter 5 is devoted to practical topics such as ﬁlter design, examples and ap
plications of anisotropic diﬀusion ﬁltering. Speciﬁc models are proposed which are
tailored towards smoothing with edge enhancement and multiscale enhancement
of coherent structures. Their qualities are illustrated using images arising from
computer aided quality control and medical applications, but also ﬁngerprint im
viii PREFACE
ages and impressionistic paintings shall be processed. The results are juxtaposed
to related methods from Chapter 1.
Finally, Chapter 6 concludes the book by giving a summary and discussing
possible future perspectives for nonlinear diﬀusion ﬁltering.
Acknowledgments. In writing this book I have been helped and inﬂuenced
by many people, and it is a pleasure to take this opportunity to express my grat
itude to them. The present book is an extended and revised version of my Ph.D.
thesis [416], which was written at the Department of Mathematics at the Uni
versity of Kaiserslautern, Germany. Helmut Neunzert, head of the Laboratory of
Technomathematics, drew my interest to diﬀusion processes in image processing,
and he provided the possibility to carry out this work at his laboratory. I also
thank him and the other editors of the ECMI Series as well as Teubner Verlag for
their interest in publishing this work.
Pierre–Louis Lions (CEREMADE, University Paris IX) invited me to the CERE
MADE, one of the birthplaces of many important ideas in this ﬁeld. He also gave
me the honour to present my results as an invited speaker at the EMS Confer
ence Multiscale Analysis in Image Processing (Lunteren, The Netherlands, October
1994) to an international audience, and he acted as a referee for the Ph.D. thesis.
After the defence of my thesis in Kaiserslautern, I joined the TGV (“tools
for geometry in vision”) group at Utrecht University Hospital for 14 months. In
this young and dynamic group I had the possibility to learn a lot about medical
image analysis, and to experience Bart ter Haar Romeny’s enthusiasm for scale
space. During that time I also met Atsushi Imiya (Chiba University, Japan) at
a workshop in Dagstuhl (Germany). He introduced me into the fascinating world
of early Japanese scalespace research conducted by Taizo Iijima decades before
scalespace became popular in America and Europe.
In the meantime I am with the computer vision group of Peter Johansen and
Jens Arnspang (DIKU, Copenhagen University). The discussions and collabora
tions with the members of this group increased my interest in scalespace related
deep structure analysis and information theory. In the latter ﬁeld I share many
common interests with Jon Sporring.
The proofreading of this book was done by Martin Reißel and Andrea Bechtold
(Kaiserslautern). Martin Reißel undertook the hard job of checking the whole
manuscript for its mathematical correctness, and Andrea Bechtold was a great help
in all kinds of diﬃculties with the English language. Also Robert Maas (Utrecht
University Hospital) contributed several useful hints.
This work has been funded by Stiftung Volkswagenwerk, Stiftung Rheinland–
Pfalz f¨ ur Innovation, the Real World Computing Partnership, the Danish Research
Council, and the EU–TMR Research Network VIRGO.
Copenhagen, October 1997 Joachim Weickert
Contents
1 Image smoothing and restoration by PDEs 1
1.1 Physical background of diﬀusion processes . . . . . . . . . . . . . . 2
1.2 Linear diﬀusion ﬁltering . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Relations to Gaussian smoothing . . . . . . . . . . . . . . . 3
1.2.2 Scalespace properties . . . . . . . . . . . . . . . . . . . . . 6
1.2.3 Numerical aspects . . . . . . . . . . . . . . . . . . . . . . . . 10
1.2.4 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.2.5 Limitations . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.2.6 Generalizations . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3 Nonlinear diﬀusion ﬁltering . . . . . . . . . . . . . . . . . . . . . . 14
1.3.1 The Perona–Malik model . . . . . . . . . . . . . . . . . . . . 15
1.3.2 Regularized nonlinear models . . . . . . . . . . . . . . . . . 20
1.3.3 Anisotropic nonlinear models . . . . . . . . . . . . . . . . . 22
1.3.4 Generalizations . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.3.5 Numerical aspects . . . . . . . . . . . . . . . . . . . . . . . . 25
1.3.6 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
1.4 Methods of diﬀusion–reaction type . . . . . . . . . . . . . . . . . . 27
1.4.1 Single diﬀusion–reaction equations . . . . . . . . . . . . . . 27
1.4.2 Coupled systems of diﬀusion–reaction equations . . . . . . . 29
1.5 Classic morphological processes . . . . . . . . . . . . . . . . . . . . 31
1.5.1 Binary and greyscale morphology . . . . . . . . . . . . . . . 31
1.5.2 Basic operations . . . . . . . . . . . . . . . . . . . . . . . . 32
1.5.3 Continuousscale morphology . . . . . . . . . . . . . . . . . 32
1.5.4 Theoretical results . . . . . . . . . . . . . . . . . . . . . . . 34
1.5.5 Scalespace properties . . . . . . . . . . . . . . . . . . . . . 34
1.5.6 Generalizations . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.5.7 Numerical aspects . . . . . . . . . . . . . . . . . . . . . . . . 36
1.5.8 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
1.6 Curvaturebased morphological processes . . . . . . . . . . . . . . . 37
1.6.1 Meancurvature ﬁltering . . . . . . . . . . . . . . . . . . . . 37
1.6.2 Aﬃne invariant ﬁltering . . . . . . . . . . . . . . . . . . . . 40
1.6.3 Generalizations . . . . . . . . . . . . . . . . . . . . . . . . . 42
1.6.4 Numerical aspects . . . . . . . . . . . . . . . . . . . . . . . . 43
ix
x CONTENTS
1.6.5 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
1.6.6 Active contour models . . . . . . . . . . . . . . . . . . . . . 46
1.7 Total variation methods . . . . . . . . . . . . . . . . . . . . . . . . 49
1.7.1 TVpreserving methods . . . . . . . . . . . . . . . . . . . . . 50
1.7.2 TVminimizing methods . . . . . . . . . . . . . . . . . . . . 50
1.8 Conclusions and further scope of the book . . . . . . . . . . . . . . 53
2 Continuous diﬀusion ﬁltering 55
2.1 Basic ﬁlter structure . . . . . . . . . . . . . . . . . . . . . . . . . . 55
2.2 The structure tensor . . . . . . . . . . . . . . . . . . . . . . . . . . 56
2.3 Theoretical results . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
2.4 Scalespace properties . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.4.1 Invariances . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
2.4.2 Informationreducing properties . . . . . . . . . . . . . . . . 65
3 Semidiscrete diﬀusion ﬁltering 75
3.1 The general model . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.2 Theoretical results . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.3 Scalespace properties . . . . . . . . . . . . . . . . . . . . . . . . . . 81
3.4 Relation to continuous models . . . . . . . . . . . . . . . . . . . . . 86
3.4.1 Isotropic case . . . . . . . . . . . . . . . . . . . . . . . . . . 86
3.4.2 Anisotropic case . . . . . . . . . . . . . . . . . . . . . . . . . 88
4 Discrete diﬀusion ﬁltering 97
4.1 The general model . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.2 Theoretical results . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
4.3 Scalespace properties . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4.4 Relation to semidiscrete models . . . . . . . . . . . . . . . . . . . . 102
4.4.1 Semiimplicit schemes . . . . . . . . . . . . . . . . . . . . . 102
4.4.2 AOS schemes . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5 Examples and applications 113
5.1 Edgeenhancing diﬀusion . . . . . . . . . . . . . . . . . . . . . . . . 114
5.1.1 Filter design . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.1.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
5.2 Coherenceenhancing diﬀusion . . . . . . . . . . . . . . . . . . . . . 127
5.2.1 Filter design . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
5.2.2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
6 Conclusions and perspectives 135
Bibliography 139
Index 165
Chapter 1
Image smoothing and restoration
by PDEs
PDEbased methods appear in a large variety of image processing and computer
vision areas ranging from shapefromshading and histogramme modiﬁcation to
optic ﬂow and stereo vision.
This chapter reviews their main application, namely the smoothing and restora
tion of images. It is written in an informal style and refers to a large amount of
original literature, where proofs and full mathematical details can be found.
The goal is to make the reader sensitive to the similarities, diﬀerences, advan
tages and shortcomings of these techniques, and to point out the main results and
open problems in this rapidly evolving area.
For each class of methods the basic ideas are explained and their theoretical
background, numerical aspects, generalizations, and applications are discussed.
Many of these ideas are borrowed from physical phenomena such as wave prop
agation or transport of heat and mass. Nevertheless, also gas dynamics, crack
propagation, grassﬁre ﬂow, the study of salinity proﬁles in oceanography, or mech
anisms of the retina and the brain are closely related to some of these approaches.
Although a detailed discussion of these connections would be far beyond the scope
of this work, they are mentioned wherever they appear, in order to allow the
interested reader to pursue these ideas. Also many historical notes are added.
The outline of this chapter is as follows: We start with reviewing the physi
cal ideas behind diﬀusion processes. This helps us to better understand the next
sections which are concerned with the properties of linear and nonlinear diﬀusion
ﬁlters in image processing. The subsequent study of image enhancement methods
of diﬀusion–reaction type relates diﬀusion ﬁlters to variational image restoration
techniques. After that we investigate morphological ﬁlters, a topic which looks at
ﬁrst glance fairly diﬀerent to the diﬀusion approach. Nevertheless, it reveals some
interesting relations when it is interpreted within a PDE framework. This becomes
1
2 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
especially evident when considering curvaturebased morphological PDEs. Finally
we shall discuss total variation image restoration techniques which permit discon
tinuous solutions. The last section summarizes the advantages and shortcomings
of the main methods and gives an outline of the questions we are concerned with
in the subsequent chapters.
1.1 Physical background of diﬀusion processes
Most people have an intuitive impression of diﬀusion as a physical process that
equilibrates concentration diﬀerences without creating or destroying mass. This
physical observation can be easily cast in a mathematical formulation.
The equilibration property is expressed by Fick’s law:
j = −D ∇u. (1.1)
This equation states that a concentration gradient ∇u causes a ﬂux j which aims
to compensate for this gradient. The relation between ∇u and j is described by
the diﬀusion tensor D, a positive deﬁnite symmetric matrix. The case where j and
∇u are parallel is called isotropic. Then we may replace the diﬀusion tensor by a
positive scalarvalued diﬀusivity g. In the general anisotropic case, j and ∇u are
not parallel.
The observation that diﬀusion does only transport mass without destroying it
or creating new mass is expressed by the continuity equation
∂
t
u = −div j (1.2)
where t denotes the time.
If we plug in Fick’s law into the continuity equation we end up with the diﬀusion
equation
∂
t
u = div (D ∇u). (1.3)
This equation appears in many physical transport processes. In the context of
heat transfer it is called heat equation. In image processing we may identify the
concentration with the grey value at a certain location. If the diﬀusion tensor is
constant over the whole image domain, one speaks of homogeneous diﬀusion, and
a spacedependent ﬁltering is called inhomogeneous. Often the diﬀusion tensor is a
function of the diﬀerential structure of the evolving image itself. Such a feedback
leads to nonlinear diﬀusion ﬁlters. Diﬀusion which does not depend on the evolving
image is called linear.
Sometimes the computer vision literature deviates from the preceding nota
tions: It can happen that homogeneous ﬁltering is named isotropic, and inhomo
geneous blurring is called anisotropic, even if it uses a scalarvalued diﬀusivity
instead of a diﬀusion tensor.
1.2 LINEAR DIFFUSION FILTERING 3
1.2 Linear diﬀusion ﬁltering
The simplest and best investigated PDE method for smoothing images is to apply
a linear diﬀusion process. We shall focus on the relation between linear diﬀusion
ﬁltering and the convolution with a Gaussian, analyse its smoothing properties for
the image as well as its derivatives, and review the fundamental properties of the
Gaussian scalespace induced by linear diﬀusion ﬁltering. Afterwards a survey on
discrete aspects is given and applications and limitations of the linear diﬀusion
paradigm are discussed. The section is concluded by sketching two linear general
izations which can incorporate apriori knowledge: aﬃne Gaussian scalespace and
directed diﬀusion processes.
1.2.1 Relations to Gaussian smoothing
Gaussian smoothing
Let a greyscale image f be represented by a realvalued mapping f ∈ L
1
(IR
2
). A
widelyused way to smooth f is by calculating the convolution
(K
σ
∗f)(x) :=
IR
2
K
σ
(x−y) f(y) dy (1.4)
where K
σ
denotes the twodimensional Gaussian of width (standard deviation)
σ>0 :
K
σ
(x) :=
1
2πσ
2
exp
−
[x[
2
2σ
2
. (1.5)
There are several reasons for the excellent smoothing properties of this method:
First we observe that since K
σ
∈ C
∞
(IR
2
) we get K
σ
∗f ∈ C
∞
(IR
2
), even if f is
only absolutely integrable.
Next, let us investigate the behaviour in the frequency domain. When deﬁning
the Fourier transformation T by
(Tf)(ω) :=
IR
2
f(x) exp(−i'ω, x`) dx (1.6)
we obtain by the convolution theorem that
(T(K
σ
∗f)) (ω) = (TK
σ
)(ω) (Tf)(ω). (1.7)
Since the Fourier transform of a Gaussian is again Gaussianshaped,
(TK
σ
)(ω) = exp
−
[ω[
2
2/σ
2
, (1.8)
4 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
we observe that (1.4) is a lowpass ﬁlter that attenuates high frequencies in a
monotone way.
Interestingly, the smoothing behaviour can also be understood in the context
of a PDE interpretation.
Equivalence to linear diﬀusion ﬁltering
It is a classical result (cf. e.g. [331, pp. 267–271] and [185, pp. 43–56]) that for any
bounded f ∈C(IR
2
) the linear diﬀusion process
∂
t
u = ∆u, (1.9)
u(x, 0) = f(x) (1.10)
possesses the solution
u(x, t) =
f(x) (t = 0)
(K
√
2t
∗ f)(x) (t > 0).
(1.11)
This solution is unique, provided we restrict ourselves to functions satisfying
[u(x, t)[ ≤ M exp (a[x[
2
) (M, a > 0). (1.12)
It depends continuously on the initial image f with respect to  . 
L
∞
(IR
2
)
, and it
fulﬁls the maximum–minimum principle
inf
IR
2
f ≤ u(x, t) ≤ sup
IR
2
f on IR
2
[0, ∞). (1.13)
From (1.11) we observe that the time t is related to the spatial width σ =
√
2t of
the Gaussian. Hence, smoothing structures of order σ requires to stop the diﬀusion
process at time
T =
1
2
σ
2
. (1.14)
Figure 5.2 (b) and 5.3 (c) in Chapter 5 illustrate the eﬀect of linear diﬀusion
ﬁltering.
Gaussian derivatives
In order to understand the structure of an image we have to analyse grey value
ﬂuctuations within a neighbourhood of each image point, that is to say, we need
information about its derivatives. However, diﬀerentiation is illposed
1
, as small
perturbations in the original image can lead to arbitrarily large ﬂuctuations in
the derivatives. Hence, the need for regularization methods arises. A thorough
1
A problem is called wellposed, if it has a unique solution which depends continuously on
the input data and parameters. If one of these conditions is violated, it is called illposed.
1.2 LINEAR DIFFUSION FILTERING 5
treatment of this mathematical theory can be found in the books of Tikhonov and
Arsenin [402], Louis [266] and Engl et al. [128].
One possibility to regularize is to convolve the image with a Gaussian prior to
diﬀerentiation [404]. By the equality
∂
n
x
1
∂
m
x
2
(K
σ
∗f) = K
σ
∗ (∂
n
x
1
∂
m
x
2
f) = (∂
n
x
1
∂
m
x
2
K
σ
) ∗ f (1.15)
for suﬃciently smooth f, we observe that all derivatives undergo the same Gaussian
smoothing process as the image itself and this process is equivalent to convolving
the image with derivatives of a Gaussian.
Replacing derivatives by these Gaussian derivatives has a strong regularizing
eﬀect. This property has been used to stabilize illposed problems like deblurring
images by solving the heat equation backwards in time
2
[141, 177]. Moreover, Gaus
sian derivatives can be combined to socalled diﬀerential invariants, expressions
that are invariant under transformations such as rotations, for instance [∇K
σ
∗u[
or ∆K
σ
∗u.
Diﬀerential invariants are useful for the detection of features such as edges,
ridges, junctions, and blobs; see [256] for an overview. To illustrate this, we focus
on two applications for detecting edges.
A frequently used method is the Canny edge detector [69]. It is based on calcu
lating the ﬁrst derivatives of the Gaussiansmoothed image. After applying sophis
ticated thinning and linking mechanisms (nonmaxima suppression and hysteresis
thresholding), edges are identiﬁed as locations where the gradient magnitude has a
maximum. This method is often acknowledged to be the best linear edge detector,
and it has become a standard in edge detection.
Another important edge detector is the Marr–Hildreth operator [278], which
uses the LaplacianofGaussian (LoG) ∆K
σ
as convolution kernel. Edges of f are
identiﬁed as zerocrossings of ∆K
σ
∗f. This needs no further postprocessing and
always gives closed contours. There are indications that LoGs and especially their
approximation by diﬀerencesofGaussians (DoGs) play an important role in the
visual system of mammals, see [278] and the references therein. Young developed
this theory further by presenting evidence that the receptive ﬁelds in primate eyes
are shaped like the sum of a Gaussian and its Laplacian [449], and Koenderink
and van Doorn suggested the set of Gaussian derivatives as a general model for
the visual system [242].
If one investigates the temporal evolution of the zerocrossings of an image ﬁl
tered by linear diﬀusion, one observes an interesting phenomenon: When increasing
the smoothing scale σ, no new zerocrossings are created which cannot be traced
back to ﬁner scales [439]. This evolution property is called causality [240]. It is
2
Of course, solutions of the regularization can only approximate the solution of the original
problem (if it exists). In practice, increasing the order of applied Gaussian derivatives or reducing
the kernel size will ﬁnally deteriorate the results of deblurring.
6 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
closely connected to the maximum–minimum principle of certain parabolic opera
tors [189]. Attempts to reconstruct the original image from the temporal evolution
of the zerocrossings of the Laplacian have been carried out by Hummel and Mo
niot [190]. They concluded, however, that this is practically unstable unless very
much additional information is provided.
In the western world the evolution property of the zerocrossings was the key
investigation which has inspired Witkin to the socalled scalespace concept [439].
This shall be discussed next.
1.2.2 Scalespace properties
The general scalespace concept
It is a wellknown fact that images usually contain structures at a large variety of
scales. In those cases where it is not clear in advance which is the right scale for the
depicted information it is desirable to have an image representation at multiple
scales. Moreover, by comparing the structures at diﬀerent scales, one obtains a
hierarchy of image structures which eases a subsequent image interpretation.
A scalespace is an image representation at a continuum of scales, embedding
the image f into a family ¦T
t
f[t ≥ 0¦ of gradually simpliﬁed versions of it, provided
that it fulﬁls certain requirements
3
. Most of these properties can be classiﬁed as
architectural, smoothing (informationreducing) or invariance requirements [12].
An important architectural assumption is recursivity, i.e. for t = 0, the scale
space representation gives the original image f, and the ﬁltering may be split into
a sequence of ﬁlter banks:
T
0
f = f, (1.16)
T
t+s
f = T
t
(T
s
f) ∀ s, t ≥ 0. (1.17)
This property is very often referred to as the semigroup property. Other architec
tural principles comprise for instance regularity properties of T
t
and local behaviour
as t tends to 0.
Smoothing properties and information reduction arise from the wish that the
transformation should not create artifacts when passing from ﬁne to coarse rep
resentation. Thus, at a coarse scale, we should not have additional structures
which are caused by the ﬁltering method itself and not by underlying structures
at ﬁner scales. This simpliﬁcation property is speciﬁed by numerous authors in
diﬀerent ways, using concepts such as no creation of new level curves (causality)
[240, 450, 189, 255], nonenhancement of local extrema [30, 257], decreasing number
3
Recently it has also been proposed to extend the scalespace concept to scale–imprecision
space by taking into account the imprecision of the measurement device [171].
1.2 LINEAR DIFFUSION FILTERING 7
of local extrema [255], maximum loss of ﬁgure impression [196], Tikhonov regular
ization [302, 303], maximum–minimum principle [189, 328], positivity [324, 138],
preservation of positivity [191, 193, 320], comparison principle [12], and Lyapunov
functionals [415, 429]. Especially in the linear setting, many of these properties are
equivalent or closely related; see [426] for more details.
We may regard an image as a representative of an equivalence class containing
all images that depict the same object. Two images of this class diﬀer e.g. by grey
level shifts, translations and rotations or even more complicated transformations
such as aﬃne mappings. This makes the requirement plausible that the scalespace
analysis should be invariant to as many of these transformations as possible, in
order to analyse only the depicted object [196, 16].
The pioneering work of Alvarez, Guichard, Lions and Morel [12] shows that
every scalespace fulﬁlling some fairly natural architectural, informationreducing
and invariance axioms is governed by a PDE with the original image as initial
condition. Thus, PDEs are the suitable framework for scalespaces.
Often these requirements are supplemented with an additional assumption
which is equivalent to the superposition principle, namely linearity:
T
t
(af +bg) = a T
t
f +b T
t
g ∀ t ≥ 0, ∀ a, b ∈ IR. (1.18)
As we shall see below, imposing linearity restricts the scalespace idea to essentially
one representative.
Gaussian scalespace
The historically ﬁrst and best investigated scalespace is the Gaussian scalespace,
which is obtained via convolution with Gaussians of increasing variance, or – equiv
alently – by linear diﬀusion ﬁltering according to (1.9), (1.10).
Usually a 1983 paper by Witkin [439] or a 1980 report by Stansﬁeld [392]
are regarded as the ﬁrst references to the linear scalespace idea. Recent work
by Weickert, Ishikawa and Imiya [426, 427], however, shows that scalespace is
more than 20 years older: An axiomatic derivation of 1D Gaussian scalespace
has already been presented by Taizo Iijima in a technical paper from 1959 [191]
followed by a journal version in 1962 [192]. Both papers are written in Japanese.
In [192] Iijima considers an observation transformation Φ which depends on
a scale parameter σ and which transforms the original image f(x) into a blurred
version
4
Φ[f(x
′
), x, σ]. This class of blurring transformations is called boke (defo
cusing). He assumes that it has the structure
Φ[f(x
′
), x, σ] =
∞
−∞
φ¦f(x
′
), x, x
′
, σ¦ dx
′
, (1.19)
4
The variable x
′
serves as a dummy variable.
8 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
and that it should satisfy ﬁve conditions:
(I) Linearity (with respect to multiplications):
If the intensity of a pattern becomes A times its original intensity, then the
same should happen to the observed pattern:
Φ[Af(x
′
), x, σ] = AΦ[f(x
′
), x, σ]. (1.20)
(II) Translation invariance:
Filtering a translated image is the same as translating the ﬁltered image:
Φ[f(x
′
−a), x, σ] = Φ[f(x
′
), x−a, σ]. (1.21)
(III) Scale invariance:
If a pattern is spatially enlarged by some factor λ, then there exists a σ
′
=
σ
′
(σ, λ) such that
Φ[f(x
′
/λ), x, σ] = Φ[f(x
′
), x/λ, σ
′
]. (1.22)
(IV) (Generalized) semigroup property:
If f is observed under a parameter σ
1
and this observation is observed un
der a parameter σ
2
, then this is equivalent to observing f under a suitable
parameter σ
3
= σ
3
(σ
1
, σ
2
):
Φ
Φ[f(x
′′
), x
′
, σ
1
], x, σ
2
= Φ[f(x
′′
), x, σ
3
]. (1.23)
(V) Preservation of positivity:
If the original image is positive, then the observed image is positive as well:
Φ[f(x
′
), x, σ] > 0 ∀ f(x
′
) > 0, ∀ σ > 0. (1.24)
Under these requirements Iijima derives in a very systematic way that
Φ[f(x
′
), x, σ] =
1
2
√
πσ
∞
−∞
f(x
′
) exp
−(x −x
′
)
2
4σ
2
dx
′
. (1.25)
Thus, Φ[f(x
′
), x, σ] is just the convolution between f and a Gaussian with standard
deviation σ
√
2.
This has been the starting point of an entire world of linear scalespace research
in Japan, which is basically unknown in the western world. Japanese scalespace
theory was wellembedded in a general framework for pattern recognition, feature
extraction and object classiﬁcation [195, 197, 200, 320], and many results have
1.2 LINEAR DIFFUSION FILTERING 9
been established earlier than in the western world. Apart from their historical
merits, these Japanese results reveal many interesting qualities which should induce
everyone who is interested in scalespace theory to have a closer look at them. More
details can be found in [426, 427] as well as in some English scalespace papers
by Iijima such as [195, 197]. In particular, the latter ones show that there is no
justiﬁcation to deny Iijima’s pioneering role in linear scalespace theory because of
language reasons.
Table 1.1: Overview of continuous Gaussian scalespace axiomatics (I1 = Iijima
[191, 192], I2 = Iijima [193, 194], I3 = Iijima [196], O = Otsu [320], K = Koenderink
[240], Y = Yuille/Poggio [450], B = Babaud et al. [30], L1 = Lindeberg [255], F1 =
Florack et al. [140], A = Alvarez et al. [12], P = Pauwels et al. [324], N = Nielsen
et al. [303], L2 = Lindeberg [257], F2 = Florack [138]).
I1 I2 I3 O K Y B L1 F1 A P N L2 F2
convolution kernel • • • • • • • • • • •
semigroup property • • • • • • • • •
locality •
regularity • • • • • • • •
inﬁnetes. generator •
max. loss principle •
causality • • • • •
nonnegativity • • • • • •
Tikhonov regulariz. •
aver. grey level invar. • • • • • •
ﬂat kernel for t →∞ • •
isometry invariance • • • • • • • • • • •
homogen. & isotropy •
separability • •
scale invariance • • • • • • • •
valid for dimension 1 2 2 2 1,2 1,2 1 1 > 1 N 1,2 N N N
Table 1.1 presents an overview of the current Japanese and western Gaussian
scalespace axiomatics (see [426, 427] for detailed explanations). All of these ax
iomatics use explicitly or implicitly
5
a linearity assumption. We observe that –
despite the fact that many axiomatics reveal similar ﬁrst principles – not two of
them are identical. Each of the 14 axiomatics conﬁrms and enhances the evidence
that the others give: that Gaussian scalespace is unique within a linear framework.
A detailed treatment of Gaussian scalespace theory can be found in two
Japanese monographs by Iijima [197, 198], as well as in English books by Lin
deberg [256], Florack [139], and ter Haar Romeny [176]. A collection edited by
5
Often it is assumed that the ﬁlter is a convolution integral. This is equivalent to linearity
and translation invariance.
10 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
Sporring, Nielsen, Florack and Johansen [389] gives an excellent overview of the
various aspects of this theory, and additional material is presented in [211]. Many
relations between Gaussian scalespace and regularization theory have been elab
orated by Nielsen [302], and readers who wish to analyse linear and nonlinear
scalespace concepts in terms of diﬀerential and integral geometry can ﬁnd a lot
of material in the thesis of Salden [351].
1.2.3 Numerical aspects
The preceding theory is entirely continuous. However, in practical problems, the
image is sampled at the nodes (pixels) of a ﬁxed equidistant grid. Thus, the diﬀu
sion ﬁlter has to be discretized.
By virtue of the equivalence of solving the linear diﬀusion equation and con
volving with a Gaussian, we can either approximate the convolution process or the
diﬀusion equation.
When restricting the image to a ﬁnite domain and applying the Fast Fourier
Transformation (FFT), convolution in the spatial domain can be reduced to mul
tiplication in the frequency domain, cf. (1.7). This proceeding requires a ﬁxed
computational eﬀort of order N log N, which depends only on the pixel number
N, but not on the kernel size σ. For large kernels this is faster than most spatial
techniques. Especially for small kernels, however, aliasing eﬀects in the Fourier
domain may create oscillations and over and undershoots [178].
One eﬃcient possibility to approximate Gaussian convolution in the spatial do
main consists of applying recursive ﬁlters [109, 448]. More frequently the Gaussian
kernel is just sampled and truncated at some multiple of its standard deviation
σ. Factorizing a higherdimensional Gaussian into onedimensional Gaussians re
duces the computational eﬀort to O(Nσ). Convolution with a truncated Gaussian,
however, reveals the drawback that it does not preserve the semigroup property of
the continuous Gaussian scalespace [255].
Lindeberg [255] has established a linear scalespace theory for the semidiscrete
6
case. His results are in accordance with those of Norman [312], who proposed in
1960 that the discrete analogue of the Gaussian kernel should be given in terms of
modiﬁed Bessel functions of integer order. Since this scalespace family arises nat
urally from a semidiscretized version of the diﬀusion equation, it has been argued
that approximating the diﬀusion equation should be preferred to discretizing the
convolution integral [255].
Recently, interesting semidiscrete and fully discrete linear scalespace formula
tions have been established utilizing stochastic principles:
˚
Astr¨om and Heyden [27]
study a framework based on stationary random ﬁelds, while the theory by Salden
et al. [353] exploits the relations between diﬀusion and Markov processes.
6
By semidiscrete we mean discrete in space and continuous in time throughout this work.
1.2 LINEAR DIFFUSION FILTERING 11
Among the numerous numerical possibilities to approximate the linear diﬀusion
equation, ﬁnite diﬀerence (FD) schemes dominate the ﬁeld. Apart from some im
plicit approaches [166, 67, 68] allowing realizations as a recursive ﬁlter [14, 10, 451],
explicit schemes are mainly used. A very eﬃcient approximation of the Gaus
sian scalespace results from applying multigrid ideas. The Gaussian pyramid [64]
has the computational complexity O(N) and gives a multilevel representation at
ﬁnitely many scales of diﬀerent resolution. By subsequently smoothing the image
with an explicit scheme for the diﬀusion equation and restricting the result to a
coarser grid, one obtains a simpliﬁed image representation at the next coarser grid.
Due to their simplicity and eﬃciency, pyramid decompositions have become very
popular and have been integrated into commercially available hardware [70, 214].
Pyramids are not invariant under translations, however, and sometimes it is ar
gued that they are undersampled and that the pyramid levels should be closer
7
.
These are the reasons why some people regard pyramids rather as predecessors of
the scalespace idea than as a numerical approximation
8
.
1.2.4 Applications
Due to its equivalence to convolution with a Gaussian, linear diﬀusion ﬁltering has
been applied in numerous ﬁelds of image processing and computer vision. It can
be found in almost every standard textbook in these ﬁelds.
Less frequent are applications which exploit the evolution of an image under
Gaussian scalespace. This deep structure analysis [240] provides useful information
for extracting semantic information from an image, for instance
• for ﬁnding the most relevant scales (scale selection, focusofattention). This
may be done by searching for extrema of (nonlinear) combinations of normal
ized Gaussian derivatives [256] or by analysing information theoretic mea
sures such as the entropy [208, 388] or generalized entropies [390] over scales.
• for multiscale segmentation of images [172, 254, 256, 313, 408]. The idea is
to identify segments at coarse scales and to link backwards to the original
image in order to improve the localization.
In recent years also applications of Gaussian scalespace to stereo, optic ﬂow
and image sequences have become an active research ﬁeld [139, 215, 241, 258, 259,
302, 306, 441]. Several scalespace applications are summarized in a survey paper
by ter Haar Romeny [175].
7
Of course, multiresolution techniques such as pyramids or discrete wavelet transforms [92,
106] are just designed to have few or no redundancies, while scalespace analysis intends to
extract semantical information by tracing signals through a continuum of scales.
8
Historically, this is incorrect: Iijima’s scalespace work [191] is much older than multigrid
ideas in image processing.
12 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
Interesting results arise when one studies linear scalespace on a sphere [236,
353]: while the diﬀusion equation remains the correct concept, Gaussian kernels are
of no use anymore: appropriate kernels have to be expressed in terms of Legendre
functions [236]. This and other results [12, 255] indicate that the PDE formulation
of linear scalespace in terms of a diﬀusion equation is more natural and has a
larger generalization potential than convolution with Gaussians.
1.2.5 Limitations
In spite of several properties that make linear diﬀusion ﬁltering unique and easy
to handle, it reveals some drawbacks as well:
(a) An obvious disadvantage of Gaussian smoothing is the fact that it does not
only smooth noise, but also blurs important features such as edges and, thus,
makes them harder to identify. Since Gaussian smoothing is designed to be
completely uncommitted, it cannot take into account any apriori informa
tion on structures which are worth being preserved (or even enhanced).
(b) Linear diﬀusion ﬁltering dislocates edges when moving from ﬁner to coarser
scales, see e.g. Witkin [439]. So structures which are identiﬁed at a coarse
scale do not give the right location and have to be traced back to the original
image [439, 38, 165]. In practice, relating dislocated information obtained at
diﬀerent scales is diﬃcult and bifurcations may give rise to instabilities. These
coarsetoﬁne tracking diﬃculties are generally denoted as the correspondence
problem.
(c) Some smoothing properties of Gaussian scalespace do not carry over from
the 1D case to higher dimensions: A closed zerocrossing contour can split
into two as the scale increases [450], and it is generally not true that the
number of local extrema is nonincreasing, see [254, 255] for illustrative coun
terexamples. A deep mathematical analysis of such phenomena has been
carried out by Damon [105] and Rieger [342]. It turned out that the pairwise
creation of an extremum and a saddle point is not an exception, but happens
generically.
Regarding (b) and (c), much eﬀorts have been spent in order to understand
the deep structure in Gaussian scalespace, for instance by analysing its toppoints
[210]. There is some evidence that these points, where the gradient vanishes and
the Hessian does not have full rank, carry essential image information [212]. Part
III of the book edited by Sporring et al. [389] and the references therein give an
overview of the stateoftheart in deep structure analysis.
Due to the uniqueness of Gaussian scalespace within a linear framework we
know that any modiﬁcation in order to overcome the problems (a)–(c) will either
1.2 LINEAR DIFFUSION FILTERING 13
renounce linearity or some scalespace properties. We shall see that appropriate
methods to avoid the shortcomings (a) and (b) are nonlinear diﬀusion processes,
while (c) requires morphological equations [206, 207, 218].
1.2.6 Generalizations
Before we turn our attention to nonlinear processes, let us ﬁrst investigate two
linear modiﬁcations which have been introduced in order to address the problems
(a) and (b) from the previous section.
Aﬃne Gaussian scalespace
A straightforward generalization of Gaussian scalespace results from renouncing
invariance under rotations. This leads to the aﬃne Gaussian scalespace
u(x, t) :=
IR
2
1
4π
det(D
t
)
exp
−
(x−y)
⊤
D
−1
t
(x−y)
4
f(y) dy (1.26)
where D
t
:= tD, t > 0, and D ∈ IR
2×2
is symmetric positive deﬁnite
9
. For a ﬁxed
matrix D, calculating the convolution integral (1.26) is equivalent to solving a
linear anisotropic diﬀusion problem with D as diﬀusion tensor:
∂
t
u = div (D ∇u), (1.27)
u(x, 0) = f(x). (1.28)
In [427] it is shown that aﬃne Gaussian scalespace has been axiomatically derived
by Iijima in 1962 [193, 194]. He named u(x, t) the generalized ﬁgure of f, and (1.27)
the basic equation of ﬁgure [196]. In 1971 this concept was realized in hardware in
the optical character reader ASPET/71 [199, 200]. The scalespace part has been
regarded as the reason for its reliability and robustness.
In 1992 Nitzberg and Shiota [310] proposed to adapt the Gaussian kernel shape
to the structure of the original image. By chosing D in (1.26) as a function of the
structure tensor (cf. Section 2.2) of f, they combined nonlinear shape adaptation
with linear smoothing. Later on similar ideas have been developed in [259, 443].
It should be noted that shapeadapted Gaussian smoothing with a spatially
varying D is no longer equivalent to a diﬀusion process of type (1.27). In practice
this can be experienced by the fact that shapeadaptation of Gaussian smoothing
does not preserve the average grey level, while the divergence formulation ensures
that this is still possible for nonuniform diﬀusion ﬁltering; see Section 1.1. Also in
this case the diﬀusion equation seems to be more general. If one wants to relate
9
Isotropic Gaussian scalespace can be recovered using the unit matrix for D.
14 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
shapeadapted Gaussian smoothing to a PDE, one has to carry out sophisticated
scaling limits [310].
Noniterative shapeadapted Gaussian smoothing diﬀers from nonlinear aniso
tropic diﬀusion ﬁltering by the fact that the latter one introduces a feedback into
the process: it adapts the diﬀusion tensor in (1.27) to the diﬀerential structure of
the ﬁltered image instead of the original image. Such concepts will be investigated
in Section 1.3.3 and in the remaining chapters of this book.
Directed diﬀusion
Another method for incorporating apriori knowledge into a linear diﬀusion process
is suggested by Illner and Neunzert [202]. Provided we are given some background
information in form of a smooth image b, they show that under some technical
requirements and suitable boundary conditions the classical solution u of
∂
t
u = b ∆u −u ∆b, (1.29)
u(x, 0) = f(x) (1.30)
converges to b along a path where the relative entropy with respect to b increases in
a monotone way. Numerical experiments have been carried out by Giuliani [159],
and an analysis in terms of nonsmooth b and weak solutions is due to Illner and
Tie [203].
Such a directed diﬀusion process requires to specify an entire image as back
ground information in advance; in many applications it would be desirable to
include a priori knowledge in a less speciﬁc way, e.g. by prescribing that features
within a certain contrast and scale range are considered to be semantically impor
tant and processed diﬀerently. Such demands can be satisﬁed by nonlinear diﬀusion
ﬁlters.
1.3 Nonlinear diﬀusion ﬁltering
Adaptive smoothing methods are based on the idea of applying a process which
itself depends on local properties of the image. Although this concept is well
known in the image processing community (see [349] and the references therein
for an overview), a corresponding PDE formulation was ﬁrst given by Perona and
Malik [326] in 1987. We shall discuss this model in detail, especially its illposedness
aspects. This gives rise to study regularizations. These techniques can be extended
to anisotropic processes which make use of an adapted diﬀusion tensor instead of
a scalar diﬀusivity.
1.3 NONLINEAR DIFFUSION FILTERING 15
1.3.1 The Perona–Malik model
Basic idea
Perona and Malik propose a nonlinear diﬀusion method for avoiding the blurring
and localization problems of linear diﬀusion ﬁltering [326, 328]. They apply an
inhomogeneous process that reduces the diﬀusivity at those locations which have
a larger likelihood to be edges. This likelihood is measured by [∇u[
2
. The Perona–
Malik ﬁlter is based on the equation
∂
t
u = div (g([∇u[
2
) ∇u). (1.31)
and it uses diﬀusivities such as
g(s
2
) =
1
1 +s
2
/λ
2
(λ > 0). (1.32)
Although Perona and Malik name their ﬁlter anisotropic, it should be noted that
– in our terminology – it would be regarded as an isotropic model, since it utilizes
a scalarvalued diﬀusivity and not a diﬀusion tensor.
Interestingly, there exists a relation between (1.31) and the neural dynamics of
brightness perception: In 1984 Cohen and Grossberg [94] proposed a model of the
primary visual cortex with similar inhibition eﬀects as in the Perona–Malik model.
The experiments of Perona and Malik were visually very impressive: edges
remained stable over a very long time. It was demonstrated [328] that edge de
tection based on this process clearly outperforms the linear Canny edge detector,
even without applying nonmaxima suppression and hysteresis thresholding. This
is due to the fact that diﬀusion and edge detection interact in one single process
instead of being treated as two independent processes which are to be applied
subsequently. Moreover, there is another reason for the impressive behaviour at
edges, which we shall discuss next.
Edge enhancement
To study the behaviour of the Perona–Malik ﬁlter at edges, let us for a moment
restrict ourselves to the onedimensional case. This simpliﬁes the notation and
illustrates the main behaviour since near a straight edge a twodimensional image
approximates a function of one variable.
For the diﬀusivity (1.32) it follows that the ﬂux function Φ(s) := sg(s
2
) satisﬁes
Φ
′
(s) ≥ 0 for [s[ ≤ λ, and Φ
′
(s) < 0 for [s[ > λ, see Figure 1.1. Since (1.31) can
be rewritten as
∂
t
u = Φ
′
(u
x
)u
xx
, (1.33)
we observe that – in spite of its nonnegative diﬀusivity – the Perona–Malik model
is of forward parabolic type for [u
x
[ ≤λ, and of backward parabolic type for [u
x
[ >λ.
16 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
0
1
0 lambda
diffusivity
0
0.6
0 lambda
flux function
Figure 1.1: (a) Left: Diﬀusivity g(s
2
)=
1
1+s
2
/λ
2
. (b) Right: Flux function
Φ(s)=
s
1+s
2
/λ
2
.
Hence, λ plays the role of a contrast parameter separating forward (low contrast)
from backward (high contrast) diﬀusion areas.
It is not hard to verify that the Perona–Malik ﬁlter increases the slope at
inﬂection points of edges within a backward area: If there exists a suﬃciently
smooth solution u it satisﬁes
∂
t
(u
2
x
) = 2u
x
∂
x
(u
t
) = 2Φ
′′
(u
x
)u
x
u
2
xx
+ 2Φ
′
(u
x
)u
x
u
xxx
. (1.34)
A location x
0
where u
2
x
is maximal at some time t is characterized by u
x
u
xx
=0
and u
x
u
xxx
≤0. Therefore,
(∂
t
(u
2
x
)) (x
0
, t) ≥ 0 for [u
x
(x
0
, t)[ > λ (1.35)
with strict inequality for u
x
u
xxx
<0.
In the twodimensional case, (1.33) is replaced by [12]
∂
t
u = Φ
′
(∇u)u
ηη
+g([∇u[
2
)u
ξξ
(1.36)
where the gauge coordinates ξ and η denote the directions perpendicular and paral
lel to ∇u, respectively. Hence, we have forward diﬀusion along isophotes (i.e. lines
of constant grey value) combined with forward–backward diﬀusion along ﬂowlines
(lines of maximal grey value variation).
We observe that the forward–backward diﬀusion behaviour is not only restricted
to the special diﬀusivity (1.32), it appears for all diﬀusivities g(s
2
) whose rapid
decay causes nonmonotone ﬂux functions Φ(s) = sg(s
2
). Overviews of several
common diﬀusivities for the Perona–Malik model can be found in [43, 343], and
a family of diﬀusivities with diﬀerent decay rates is investigated in [36]. Rapidly
decreasing diﬀusivities are explicitly intended in the Perona–Malik method as they
1.3 NONLINEAR DIFFUSION FILTERING 17
give the desirable result of blurring small ﬂuctuations and sharpening edges. There
fore, they are the main reason for the visually impressive results of this restoration
technique.
It is evident that the “optimal” value for the contrast parameter λ has to depend
on the problem. Several proposals have been made to facilitate such a choice in
practice, for instance adapting it to a speciﬁed quantile in the cumulative gradient
histogramme [328], using statistical properties of a training set of regions which
are considered as ﬂat [444], or estimating it by means of the local image geometry
[270].
Illposedness
Unfortunately, forward–backward equations of Perona–Malik type cause some the
oretical problems. Although there is no general theory for nonlinear parabolic
processes, there exist certain frameworks which allow to establish wellposedness
results for a large class of equations. Let us recall three examples:
• Let S(N) denote the set of symmetric N N matrices and Hess(u) the
Hessian of u. Classical diﬀerential inequality techniques [411] based on the
Nagumo–Westphal lemma require that the underlying nonlinear evolution
equation
∂
t
u = F(t, x, u, ∇u, Hess(u)) (1.37)
satisﬁes the monotony property
F(t, x, r, p, Y ) ≥ F(t, x, r, p, X) (1.38)
for all X, Y ∈S(2) where Y −X is positive semideﬁnite.
• The same requirement is needed for applying the theory of viscosity solutions.
A detailed introduction into this framework can be found in a paper by
Crandall, Ishii and Lions [103].
• Let H be a Hilbert space with scalar product (., .) and A : H →H . In order
to apply the concept of maximal monotone operators [57] to the problem
du
dt
+Au = 0, (1.39)
u(0) = f (1.40)
one has to ensure that A is monotone, i.e.
(Au−Av, u−v) ≥ 0 ∀ u, v ∈ H. (1.41)
18 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
We observe that the nonmonotone ﬂux function of the Perona–Malik process im
plies that neither (1.38) is satisﬁed nor A deﬁned by Au := −div (g([∇u[
2
) ∇u)
is monotone. Therefore, none of these frameworks is applicable to ensure well
posedness results.
One reason why people became pessimistic about the wellposedness of the
Perona–Malik equation was a result by H¨ollig [187]. He constructed a forward–
backward diﬀusion process which can have inﬁnitely many solutions. Although this
process was diﬀerent from the Perona–Malik process, one was warned what can
happen. In 1994 the general conjecture was that the Perona–Malik ﬁlter might have
weak solutions, but one should neither expect uniqueness nor stability [329]. In the
meantime several theoretical results are available which provide some insights into
the actual degree of illposedness of the Perona–Malik ﬁlter.
Kawohl and Kutev [222] proved that the Perona–Malik process does not have
global (weak) C
1
solutions for intial data that involve backward diﬀusion. The exis
tence of local C
1
solutions remained unproven. If they exist, however, Kawohl and
Kutev showed that these solutions are unique and satisfy a maximumminimum
principle. Moreover, under special assumptions on the initial data, it was possible
to establish a comparison principle.
Kichenassamy [224, 225] proposed a notion of generalized solutions, which are
piecewise linear and contain jumps, and he showed that an analysis of their moving
and merging gives similar eﬀects to those one can observe in practice.
Results of You et al. [446] give evidence that the Perona–Malik process is
unstable with respect to perturbations of the initial image. They showed that the
energy functional leading to the Perona–Malik process as steepest descent method
has an inﬁnite number of global minima which are dense in the image space. Each
of these minima corresponds to a piecewise constant image, and slightly diﬀerent
initial images may end up in diﬀerent minima for t →∞.
Interestingly, forward–backward diﬀusion equations of Perona–Malik type are
not as unnatural as they look at ﬁrst glance: besides their importance in computer
vision they have been proposed as a mathematical model for heat and mass transfer
in a stably stratiﬁed turbulent shear ﬂow. Such a model is used to explain the
evolution of stepwise constant temperature or salinity proﬁles in the ocean. Related
equations also play a role in population dynamics and viscoelasiticity, see [35] and
the references therein.
Numerically, the mainly observable instability is the socalled staircasing eﬀect,
where a sigmoid edge evolves into piecewise linear segments which are separated
by jumps. It has already been observed by Posmentier in 1977 [333]. He used an
equation of Perona–Malik type for numerical simulations of the salinity proﬁles
in oceans. Starting from a smoothly increasing initial distribution he reported the
creation of perturbations which led to a stepwise constant proﬁle after some time.
1.3 NONLINEAR DIFFUSION FILTERING 19
In image processing, numerical studies of the staircasing eﬀect have been carried
out by Nitzberg and Shiota [310], Fr¨ohlich and Weickert [148], and Benhamouda
[36]. All results point in the same direction: the number of created plateaus de
pends strongly on the regularizing eﬀect of the discretization. Finer discretizations
are less regularizing and lead to more “stairs”. Weickert and Benhamouda [425]
showed that the regularizing eﬀect of a standard ﬁnite diﬀerence discretization is
suﬃcient to turn the Perona–Malik ﬁlter into a wellposed initial value problem
for a nonlinear system of ordinary diﬀerential equations. Its global solution satis
ﬁes a maximum–minimum principle and converges to a constant steadystate. The
theoretical framework for this analysis will be presented in Chapter 3.
There exists also a discrete explanation why staircasing is essentially the only
observable instability: In 1D, standard FD discretizations are monotonicity pre
serving, which guarantees that no additional oscillations occur during the evolu
tion. This has been shown by Dzu Magaziewa [123] in the semidiscrete case and
by Benhamouda [36, 425] in the fully discrete case with an explicit time discretiza
tion. Further contributions to the explanation and avoidance of staircasing can be
found in [4, 36, 98, 225, 438].
Scalespace interpretation
Perona and Malik renounced the assumption of Koenderink’s linear scalespace
axiomatic [240] that the smoothing should treat all spatial points and scale levels
equally. Instead of this, they required that region boundaries should be sharp and
should coincide with the semantically meaningful boundaries at each resolution
level (immediate localization), and that intraregion smoothing should be preferred
to interregion smoothing (piecewise smoothing). These properties are of signiﬁcant
practical interest, as they guarantee that structures can be detected easily and
correspondence problems can be neglected. Experiments demonstrated that the
Perona–Malik ﬁlter satisﬁes these requirements fairly well [328].
In order to establish a smoothing scalespace property for this nonlinear dif
fusion process, a natural way would be to prove a maximum–minimum principle,
provided one knows that there exists a suﬃciently smooth solution. Since the ex
istence question used to be the bottleneck in the past, the ﬁrst proof is due to
Kawohl and Kutev who established an extremum principle for their local weak C
1
solution to the Perona–Malik ﬁlter [222]. Of course, this is only partly satisfying,
since in scalespace theory one is interested in having an extremum principle for
the entire time interval [0, ∞).
Nevertheless, also other attempts to apply scalespace frameworks to the Perona–
Malik process have not been more successful yet:
• Salden [350], Florack [143] and Eberly [124] proposed to carry over the linear
scalespace theory to the nonlinear case by considering nonlinear diﬀusion
20 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
processes which result from special rescalings of the linear one. Unfortunately,
the Perona–Malik ﬁlter turned out not to belong to this class [143].
• Alvarez, Guichard, Lions and Morel [12] have developed a nonlinear scale
space axiomatic which comprises the linear scalespace theory as well as
nonlinear morphological processes (which we will discuss in 1.5 and 1.6).
Their smoothing axiom is a monotony assumption (comparison principle)
requiring that the scalespace is orderpreserving:
f ≤ g =⇒ T
t
f ≤ T
t
g ∀ t ≥ 0. (1.42)
This property is closely related to a maximum–minimum principle and to
L
∞
stability of the solution [12, 261]. However, the Perona–Malik model
does not ﬁt into this framework, because its local weak solution satisﬁes a
comparison principle only for some ﬁnite time, but not for all t > 0; see [222].
1.3.2 Regularized nonlinear models
It has already been mentioned that numerical schemes may provide implicit reg
ularizations which stabilize the Perona–Malik process [425]. Hence, it has been
suggested to introduce the regularization directly into the continuous equation in
order to become more independent of the numerical implementation [81, 310].
Since the dynamics of the solution may critically depend on the sort of regu
larization, one should adjust the regularization to the desired goal of the forward–
backward heat equation [35]. One can apply spatial or temporal regularization
(and of course, a combination of both). Below we shall discuss three examples
which illustrate the variety of possibilities and their tailoring towards a speciﬁc
task.
(a) The ﬁrst spatial regularization attempt is probably due to Posmentier who
observed numerically the stabilizing eﬀect of averaging the gradient within
the diﬀusivity [333].
A mathematically sound formulation of this idea is given by Catt´e, Lions,
Morel and Coll [81]. By replacing the diﬀusivity g([∇u[
2
) of the Perona–
Malik model by a Gaussiansmoothed version g([∇u
σ
[
2
) with u
σ
:= K
σ
∗u
they end up with
∂
t
u = div (g([∇u
σ
[
2
) ∇u). (1.43)
In [81] existence, uniqueness and regularity of a solution for σ> 0have been
established.
This process has been analysed and modiﬁed in many ways: Whitaker and
Pizer [438] have suggested that the regularization parameter σ should be
1.3 NONLINEAR DIFFUSION FILTERING 21
a decreasing function in t, and Li and Chen [252] have proposed to subse
quently decrease the contrast parameter λ. A detailed study of the inﬂuence
of the parameters in a regularized Perona–Malik model has been carried out
by Benhamouda [36]. Kaˇcur and Mikula [217] have investigated a modiﬁca
tion which allows to diﬀuse diﬀerently in diﬀerent grey value ranges. Spatial
regularizations of the Perona–Malik process leading to anisotropic diﬀusion
equations have been proposed by Weickert [413, 415] and will be described
in 1.3.3. Torkamani–Azar and Tait [403] suggest to replace the Gaussian
convolution by the exponential ﬁlter of Shen and Castan
10
[381].
In Chapter 2 we shall see that spatial regularizations lead to wellposed scale
spaces with a large class of Lyapunov functionals which guarantee that the
solution converges to a constant steadystate.
From a practical point of view, spatial regularizations oﬀer the advantage
that they make the ﬁlter insensitive to noise at scales smaller than σ. There
fore, when regarding (1.43) as an image restoration equation, it exhibits
besides the contrast parameter λ an additional noise scale σ. This avoids a
shortcoming of the genuine Perona–Malik process which misinterprets strong
oscillations due to noise as edges which should be preserved or even enhanced.
Examples for spatially regularized nonlinear diﬀusion ﬁltering can be found
in Figure 5.2 (c) and 5.4 (a),(b).
(b) P.L. Lions proved in a private communication to Mumford that the one
dimensional process
∂
t
u = ∂
x
(g(v) ∂
x
u), (1.44)
∂
t
v =
1
τ
([∂
x
u[
2
−v) (1.45)
leads to a wellposed ﬁlter (cf. [329]). We observe that v is intended as a
timedelay regularization of [∂
x
u[
2
where the parameter τ > 0 determines
the delay. These equations arise as a special case of the spatiotemporal
regularizations of Nitzberg and Shiota [310] when neglecting any spatial reg
ularization. Mumford conjectures that this model gives piecewise constant
steadystates. In this case, the steadystate solution would solve a segmen
tation problem.
(c) In the context of shear ﬂows, Barenblatt et al. [35] regularized the one
dimensional forward–backward heat equation by considering the thirdorder
equation
∂
t
u = ∂
x
(Φ(u
x
)) +τ∂
xt
(Ψ(u
x
)) (1.46)
10
This renounces invariance under rotation.
22 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
where Ψ is strictly increasing and uniformly bounded in IR, and [Φ
′
(s)[ =
O(Ψ
′
(s)) as s → ±∞. This regularization was physically motivated by in
troducing a relaxation time τ into the diﬀusivity.
For the corresponding initial boundary value problem with homogeneous
Neumann boundary conditions they proved the existence of a unique gen
eralized solution. They also showed that smooth solutions may become dis
continuous within ﬁnite time, before they ﬁnally converge to a piecewise
constant steadystate.
These examples demonstrate that regularization is much more than stabilizing
an illposed process: Regularization is modeling. Appropriately chosen regulariza
tions create the desired ﬁlter features. We observe that spatial regularizations are
closer to scalespace ideas while temporal regularization are more related to image
restoration and segmentation, since they may lead to nontrivial steadystates.
1.3.3 Anisotropic nonlinear models
All nonlinear diﬀusion ﬁlters that we have investigated so far utilize a scalarvalued
diﬀusivity g which is adapted to the underlying image structure. Therefore, they
are isotropic and the ﬂux j = −g∇u is always parallel to ∇u. Nevertheless, in
certain applications it would be desirable to bias the ﬂux towards the orientation
of interesting features. These requirements cannot be satisﬁed by a scalar diﬀu
sivity anymore, a diﬀusion tensor leading to anisotropic diﬀusion ﬁlters has to be
introduced.
First anisotropic ideas in image processing date back to Graham [167] in 1962,
followed by Newman and Dirilten [300], Lev, Zucker and Rosenfeld [250], and
Nagao and Matsuyama [297]. They used convolution masks that depended on
the underlying image structure. Related statistical approaches were proposed by
Knutsson, Wilson and Granlund [237]. These ideas have been further developed
by Nitzberg and Shiota [310], Lindeberg and G˚arding [259], and Yang et al. [443].
Their suggestion to use shapeadapted Gaussian masks has been discussed in Sec
tion 1.2.6.
Anisotropic diﬀusion ﬁlters usually apply spatial regularization strategies
11
. A
general theoretical framework for spatially regularized anisotropic diﬀusion ﬁlters
will be presented in the remaining chapters of this book.
Below we study two representatives of anisotropic diﬀusion processes. The ﬁrst
one oﬀers advantages at noisy edges, whereas the second one is welladapted to the
processing of onedimensional features. They are called edgeenhancing anisotropic
diﬀusion and coherenceenhancing anisotropic diﬀusion, respectively.
11
An exception is the timedelay regularization of Cottet and ElAyyadi [100, 101].
1.3 NONLINEAR DIFFUSION FILTERING 23
(a) Anisotropic regularization of the Perona–Malik process
In the interior of a segment the nonlinear isotropic diﬀusion equation (1.43)
behaves almost like the linear diﬀusion ﬁlter (1.9), but at edges diﬀusion
is inhibited. Therefore, noise at edges cannot be eliminated successfully by
this process. To overcome this problem, a desirable method should prefer
diﬀusion along edges to diﬀusion perpendicular to them.
Anisotropic models do not only take into account the modulus of the edge
detector ∇u
σ
, but also its direction. To this end, we construct the orthonor
mal system of eigenvectors v
1
, v
2
of the diﬀusion tensor D such that they
reﬂect the estimated edge structure:
v
1
 ∇u
σ
, v
2
⊥ ∇u
σ
. (1.47)
In order to prefer smoothing along the edge to smoothing across it, Weickert
[415] proposed to choose the corresponding eigenvalues λ
1
and λ
2
as
λ
1
(∇u
σ
) := g([∇u
σ
[
2
), (1.48)
λ
2
(∇u
σ
) := 1. (1.49)
Section 5.1 presents several examples where this process is applied to test
images.
In general, ∇u does not coincide with one of the eigenvectors of D as long
as σ>0. Hence, this model behaves really anisotropic. If we let the regular
ization parameter σ tend to 0, we end up with the isotropic Perona–Malik
process.
Another anisotropic model which can be regarded as a regularization of an
isotropic nonlinear diﬀusion ﬁlter has been described in [413].
(b) Anisotropic models for smoothing onedimensional objects
A second motivation for introducing anisotropy into diﬀusion processes arises
from the wish to process onedimensional features such as linelike structures.
To this end, Cottet and Germain [102] constructed a diﬀusion tensor with
eigenvectors as in (1.47) and corresponding eigenvalues
λ
1
(∇u
σ
) := 0, (1.50)
λ
2
(∇u
σ
) :=
η[∇u
σ
[
2
1 + ([∇u
σ
[/σ)
2
(η > 0). (1.51)
This is a process diﬀusing solely perpendicular to ∇u
σ
. For σ → 0, we
observe that ∇u becomes an eigenvector of D with corresponding eigenvalue
0. Therefore, the process stops completely. In this sense, it is not intended as
an anisotropic regularization of the Perona–Malik equation. Wellposedness
24 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
results for the Cottet–Germain ﬁlter comprise an existence proof for weak
solutions.
Since the Cottet–Germain model diﬀuses only in one direction, it is clear
that its result depends very much on the smoothing direction. For enhancing
parallel linelike structures, one can improve this model when replacing ∇u
⊥
σ
by a more robust descriptor of local orientation, the structure tensor (cf.
Section 2.2). This leads to coherenceenhancing anisotropic diﬀusion [418],
which shall be discussed in Section 5.2, where also many examples can be
found.
1.3.4 Generalizations
Higher dimensions. It is easily seen that many of the previous results can
be generalized to higher dimensions. This may be useful when considering e.g.
medical image sequences from computerized tomography (CT) or magnetic reso
nance imaging (MRI), or when applying diﬀusion ﬁlters to the postprocessing of
ﬂuctuating higherdimensional numerical data. The ﬁrst threedimensional non
linear diﬀusion ﬁlters have been investigated by Gerig et al. [155] in the isotropic
case and by Rambaux and Gar¸con [339] in the anisotropic case. A generalization
of coherenceenhancing anisotropic diﬀusion to higher dimensions is proposed in
[428], and S´anchez–Ortiz et al. [355] describe nonlinear diﬀusion ﬁltering of 3D
image sequences by treating them as 4D data sets.
More sophisticated structure descriptors. The edge detector ∇u
σ
en
ables us to adapt the diﬀusion to magnitude and direction of edges, but it can
neither distinguish between edges and corners nor does it give a reliable measure
of local orientation. As a remedy, one can steer the smoothing process by more
advanced structure descriptors such as higherorder derivatives [127] or tensor
valued expressions of ﬁrstorder derivatives [414, 418]. The theoretical analysis in
the present work shall comprise the second possibility. It has also been proposed
to replace ∇u
σ
by a Bayesian classiﬁcation result in feature space [26].
Vectorvalued models. Vectorvalued images can arise either from devices
measuring multiple physical properties or from a feature analysis of one single
image. Examples for the ﬁrst category are colour images, multispectral Landsat
exposures and multispin echo MR images, whereas representatives of the second
class are given by statistical moments or the jet space induced by the image itself
and its partial derivatives up to a given order. Feature vectors play an important
role for tasks like texture segmentation.
1.3 NONLINEAR DIFFUSION FILTERING 25
The simplest idea how to apply diﬀusion ﬁltering to multichannel images would
be to diﬀuse all channels separately and independently from each other. This leads
to the undesirable eﬀect that edges may be formed at diﬀerent locations for each
channel. In order to avoid this, one should use a common diﬀusivity which combines
information from all channels. Such isotropic vectorvalued diﬀusion models were
studied by Gerig et al. [155, 156] and Whitaker [433, 434] in the context of medical
imagery. Extensions to anisotropic vectorvalued models with a common tensor
valued structure descriptor for all channels have been investigated by Weickert
[422].
1.3.5 Numerical aspects
For nonlinear diﬀusion ﬁltering numerous numerical methods have been applied:
Finite element techniques are described in [367, 391, 34, 216]. B¨ansch and
Mikula reported a signiﬁcant speedup by supplementing them with an adaptive
mesh coarsening [34]. Neural network approximations to nonlinear diﬀusion ﬁlters
are investigated by Cottet [100, 99] and Fischl and Schwartz [137]. Perona and
Malik [327] propose hardware realizations by means of analogue VLSI networks
with nonlinear resistors. A very detailed VLSI proposal has been developed by
Gijbels et al. [158].
In [148] three schemes for a spatially regularized 1D Perona–Malik ﬁlter are
compared: a wavelet method of Petrov–Galerkin type, a pseudospectral method
and a ﬁnitediﬀerence scheme. It turned out that all results became fairly similar,
when the regularization parameter σ was suﬃciently large. Since the computational
eﬀort is of a comparable order of magnitude, it seems to be a matter of taste which
scheme is preferred.
Most implementations of nonlinear diﬀusion ﬁlters are based on ﬁnite diﬀer
ence methods, since they are easy to handle and the pixel structure of digital
images already provides a natural discretization on a ﬁxed rectangular grid. Ex
plicit schemes are the most simple to code and, therefore, they are used almost
exclusively. Due to their local behaviour, they are wellsuited for parallel architec
tures. Nevertheless, they suﬀer from the fact that fairly small time step sizes are
needed in order to ensure stability. Semiimplicit schemes – which approximate the
diﬀusivity or the diﬀusion tensor in an explicit way and the rest implicitly – are
considered in [81]. They possess much better stability properties. A fast multigrid
technique using a pyramid algorithm for the Perona–Malik ﬁlter has been studied
by Acton et al. [5, 4]; see also [349] for related ideas.
While the preceding techniques are focusing on approximating a continuous
equation, it is often desirable to have a genuinely discrete theory which guarantees
that an algorithm exactly reveals the same qualitative properties as its continuous
counterpart. Such a framework is presented in [420, 421], both for the semidiscrete
26 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
Table 1.2: Requirements for continuous, semidiscrete and fully dis
crete nonlinear diﬀusion scalespace.
requirement continuous semidiscrete discrete
u
t
= div (D∇u)
du
dt
= A(u)u u
0
= f
u(t = 0) = f u(0) = f u
k+1
= Q(u
k
)u
k
'D∇u, n` = 0
smoothness D ∈ C
∞
A Lipschitz Q continuous
continuous
symmetry D symmetric A symmetric Q symmetric
conservation div form; column sums column sums
reﬂective b.c. are 0 are 1
nonnega positive nonnegative nonnegative
tivity semideﬁnite oﬀdiagonals elements
connectivity uniformly pos. irreducible irreducible;
deﬁnite pos. diagonal
and for the fully discrete case. A detailed treatment of this theory can be found
in Chapter 3 and 4, respectively. Table 1.2 gives an overview of the requirements
which are needed in order to prove wellposedness, average grey value invariance,
causality in terms of an extremum principle and Lyapunov functionals, and con
vergence to a constant steadystate [423].
We observe that the requirements belong to ﬁve categories: smoothness, sym
metry, conservation, nonnegativity and connectivity requirements. These criteria
are easy to check for many discretizations. In particular, it turns out that suitable
explicit and semiimplicit ﬁnite diﬀerence discretizations of many discussed models
create discrete scalespaces. The discrete nonlinear scalespace concept has also led
to the development of fast novel schemes, which are based on an additive operator
splitting (AOS) [424, 430]. Under typical accuracy requirements, they are about
10 times more eﬃcient than the widely used explicit schemes, and a speedup by
another order of magnitude can be achieved by a parallel implementation [431]. A
general framework for AOS schemes will be presented in Section 4.4.2.
1.3.6 Applications
Nonlinear diﬀusion ﬁlters have been applied for postprocessing ﬂuctuating data
[269, 415], for visualizing qualityrelevant features in computer aided quality con
trol [299, 413, 418], and for enhancing textures such as ﬁngerprints [418]. They have
proved to be useful for improving subsampling [144] and line detection [156, 418],
for blind image restoration [445], for scalespace based segmentation algorithms
1.4 METHODS OF DIFFUSION–REACTION TYPE 27
[307, 308], for segmentation of textures [433, 437] and remotely sensed data [6, 5],
and for target tracking in infrared images [65]. Most applications, however, are
concerned with the ﬁltering of medical images [26, 28, 29, 155, 244, 248, 264, 270,
308, 321, 355, 386, 393, 431, 434, 437, 444]. Some of these applications will be
investigated in more detail in Chapter 5.
Besides such speciﬁc problem solutions, nonlinear diﬀusion ﬁlters can be found
in commercial software packages such as the medical visualization tool Analyze.
12
1.4 Methods of diﬀusion–reaction type
This section investigates variational frameworks, in which diﬀusion–reaction equa
tions or coupled systems of them are interpreted as steepest descent minimizers of
suitable energy functionals. This idea connects diﬀusion methods to edge detection
and segmentation ideas.
Besides the variational interpretation there exist other interesting theoretical
frameworks for diﬀusion ﬁlters such as the Markov random ﬁeld and mean ﬁeld
annealing context [152, 153, 247, 251, 328, 387], robust statistics [41], and deter
ministic interactive particle models [279]. Their discussion, however, would lead us
beyond the scope of this book.
1.4.1 Single diﬀusion–reaction equations
Nordstr¨om [311] has suggested to obtain a reconstruction u of a degraded image
f by minimizing the energy functional
E
f
(u, w) :=
Ω
β(u−f)
2
+ w[∇u[
2
+ λ
2
(w−lnw)
dx. (1.52)
The parameters β and λ are positive weights and w : Ω →[0, 1] gives a fuzzy edge
representation: in the interior of a region, w approaches 1 while at edges, w is close
to 0 (as we shall see below).
The ﬁrst summand of E punishes deviations of u from f (deviation cost), the
second term detects unsmoothness of u within each region (stabilizing cost), and
the last one measures the extend of edges (edge cost). Cost terms of these three
types are typical for variational image restoration methods.
The corresponding Euler equations to this energy functional are given by
0 = β(u−f) − div (w∇u), (1.53)
0 = λ
2
(1−
1
w
) + [∇u[
2
, (1.54)
12
Analyze is a registered trademark of Mayo Medical Ventures, 200 First Street SW, Rochester,
MN 55905, U.S.A.
28 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
equipped with a homogeneous Neumann boundary condition for u.
Solving (1.54) for w gives
w =
1
1 +[∇u[
2
/λ
2
. (1.55)
We recognize that w is identical with the Perona–Malik diﬀusivity g([∇u[
2
) in
troduced in (1.32). Therefore, (1.53) can be regarded as the steadystate equation
of
∂
t
u = div (g([∇u[
2
) ∇u) + β(f −u). (1.56)
This equation can also be obtained directly as the descent method of the functional
F
f
(u) :=
Ω
β(u−f)
2
+ λ
2
ln
1+
∇u
2
λ
2
dx. (1.57)
The diﬀusion–reaction equation (1.56) consists of the Perona–Malik process
with an additional bias term β(f −u). One of Nordstr¨om’s motivations for intro
ducing this term was to free the user from the diﬃculty of specifying an appropriate
stopping time for the Perona–Malik process.
However, it is evident that the Nordstr¨om model just shifts the problem of
specifying a stopping time T to the problem of determining β. So it seems to
be a matter of taste which formulation is preferred. People interested in image
restoration usually prefer the reaction term, while for scalespace researchers it is
more natural to have a constant steadystate as the simplest image representation.
Nordstr¨om’s method may suﬀer from the same illposedness problems as the
underlying Perona–Malik equation, and it is not hard to verify that the energy
functional (1.57) is nonconvex. Therefore, it can possess numerous local minima,
and the process (1.56) with f as initial condition does not necessarily converge to
a global minimum. Similar diﬃculties may also arise in other diﬀusion–reaction
models, where convergence results have not yet been established [152, 186].
A popular possibility to avoid these illposedness and convergence problems is
to renounce edgeenhancing diﬀusivities in order to end up with (nonquadratic)
convex functionals [43, 88, 110, 367, 391]. In this case the frameworks of convex
optimization and monotone operators are applicable, ensuring wellposedness and
stability of a standard ﬁniteelement approximation [367].
Diﬀusion–reaction approaches have been applied to edge detection [367, 391],
to the restoration of inverse scattering images [263], to SPECT [88] and vascular
reconstruction in medical imaging [102, 325], and to optic ﬂow [368, 111] and
stereo problems [343]. They can be extended to vectorvalued images [369] and
to cornerpreserving smoothing of curves [136, 323]. Diﬀusion–reaction methods
with constant diﬀusivities have also been used for local contrast normalization in
images [330].
1.4 METHODS OF DIFFUSION–REACTION TYPE 29
1.4.2 Coupled systems of diﬀusion–reaction equations
Mumford and Shah [295, 296] have proposed to obtain a segmented image u from
f by minimizing the functional
E
f
(u, K) = β
Ω
(u−f)
2
dx +
Ω\K
[∇u[
2
dx + α[K[ (1.58)
with nonnegative parameters α and β. The discontinuity set K consists of the
edges, and its onedimensional Hausdorﬀ measure [K[ gives the total edge length.
Like the Nordstr¨om functional (1.52), this expression consists of three cost terms:
the ﬁrst one is the deviation cost, the second one gives the stabilizing cost, and
the third one represents the edge cost.
The Mumford–Shah functional can be regarded as a continuous version of the
Markov random ﬁeld method of Geman and Geman [154] and the weak membrane
model of Blake and Zisserman [42]. Related approaches are also used to model
materials with two phases and a free interface.
The fact that (1.58) leads to a free discontinuity problem causes many challeng
ing theoretical questions [249]. The book of Morel and Solimini [292] covers a very
detailed analysis of this functional. Although the existence of a global minimizer
with a closed edge set K has been established [108, 17], uniqueness is in general
not true [292, pp. 197–198]. Regularity results for K in terms of (at least) C
1
arcs
have recently been obtained [18, 19, 20, 48, 49, 107].
The concept of energy functionals for segmenting images oﬀers the practical
advantage that it provides a framework for comparing the quality of two seg
mentations. On the other hand, (1.58) exhibits also some shortcomings, e.g. the
problem that sigmoidlike edges produce multiple segmentation boundaries (over
segmentation, staircasing eﬀect) [377]. Another drawback results from the fact that
the Mumford–Shah functional allows only singularities which are typical for mini
mal surfaces: Corners or Tjunctions are not possible and segments meet at triple
points with 120
o
angle [296]. In order to avoid such problems, modiﬁcations of the
Mumford–Shah functional have been proposed by Shah [379]. An aﬃne invariant
generalization of (1.58) is investigated in [32, 31] and applied to aﬃne invariant
texture segmentation [31, 33], and a Mumford–Shah functional for curves can be
found in [323].
Since many algorithms in image processing can be restated as versions of the
Mumford–Shah functional [292] and since it is a prototype of a free discontinuity
problem it is instructive to study this variational problem in more detail.
Numerical complications arise from the fact that the Mumford–Shah functional
has numerous local minima. Global minimizers such as the simulated annealing
method used by Geman and Geman [154] are extremely slow. Hence, one searches
for fast (suboptimal) deterministic strategies, e.g. pyramidal regiongrowing algo
rithms [3, 239].
30 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
Another important class of numerical methods is based on the idea to approx
imate the discontinuity set K by a smooth function w, which is close to 0 near
edges of u and which approximates 1 elsewhere.
We may for instance study the functional
F
f
(u, w) :=
Ω
β(u−f)
2
+ w
2
[∇u[
2
+ α
c[∇w[
2
+
(1−w)
2
4c
dx (1.59)
with a positive parameter c specifying the “edge width”. Ambrosio and Tortorelli
proved that this functional converges to the Mumford–Shah functional for c → 0
(in the sense of Γconvergence, see [22] for more details).
Minimizing F
f
corresponds to the gradient descent equations
∂
t
u = div (w
2
∇u) + β(f −u), (1.60)
∂
t
w = c∆w −
w
α
[∇u[
2
+
(1−w)
4c
(1.61)
with homogeneous Neumann boundary conditions. Equations of this type are in
vestigated by Richardson and Mitter [341]. Since (1.60) resembles the Nordstr¨om
process (1.56), similar problems can arise: The functional F
f
is not jointly convex
in u and v, so it may have many local minima and a gradient descent algorithm
may get trapped in a poor local minimum. Wellposedness results for this system
have not been obtained up to now, but a maximum–minimum principle and a local
stability proof have been established.
Another diﬀusion–reaction system is studied by Shah [375, 376]. He replaces
the functional (1.58) by two coupled convex energy functionals and applies gradi
ent descent. This results in ﬁnding an equilibrium state between two competing
processes. Experiments indicate that it converges to a stable solution. Proesmans
et al. [337, 336] observed that this solution looks fairly blurred since the equations
contain diﬀusion terms such as ∆u. They obtained pronounced edges by replacing
such a term by its Perona–Malik counterpart div (g([∇u[
2
) ∇u). Related equations
are also studied in [398]. Of course, this approach gives rise to the same theoretical
questions as (1.60), (1.61).
The system of Richardson and Mitter is used for edge detection [341]. Shah in
vestigates diﬀusion–reaction systems for matching stereo images [378], while Proes
mans et al. apply coupled diﬀusionreaction equations to image sequence analysis,
vectorvalued images and stereo vision [336, 338]. Their ﬁnite diﬀerence algorithms
run on a parallel transputer network.
It should also be mentioned that there exist reaction–diﬀusion systems which
have been applied to image restoration [334, 335, 382], texture generation [406, 440]
and halftoning [382], and which are not connected to Perona–Malik or Mumford–
Shah ideas. They are based on Turing’s pattern formation model [405].
1.5 CLASSIC MORPHOLOGICAL PROCESSES 31
1.5 Classic morphological processes
Morphology is an approach to image analysis based on shapes. Its mathematical
formalization goes back to the group around Matheron and Serra, both working
at ENS des Mines de Paris in Fontainebleau. The theory had ﬁrst been developed
for binary images, afterwards it was extended to greyscale images by regarding
level sets as shapes. Its applications cover biology, medical diagnostics, histology,
quality control, radar and remote sensing, science of material, mineralogy, and
many others.
Morphology is usually described in terms of algebraic set theory, see e.g. [280,
371, 181, 184] for an overview. Nevertheless, PDE formulations for classic morpho
logical processes have been discovered recently by Brockett and Maragos [60], van
den Boomgaard [50], Arehart et al. [25] and Alvarez et al. [12].
This section surveys the basic ideas and elementary operations of binary and
greyscale morphology, presents its PDE representations for images and curves,
and summarizes the results concerning wellposedness and scalespace properties.
Afterwards numerical aspects of the PDE formulation of these processes are dis
cussed, and generalizations are sketched which comprise the morphological equiv
alent of Gaussian scalespace.
1.5.1 Binary and greyscale morphology
Binary morphology considers shapes (silhouettes), i.e. closed sets X⊂IR
2
whose
boundaries are Jordan curves [16]. Henceforth, we identify a shape X with its
characteristic function
χ(x) :=
1 if x∈X,
0 else.
(1.62)
Binary morphological operations aﬀect only the boundary curve of the shape and,
therefore, they can be viewed as curve or shape deformations.
Greyscale morphology generalizes these ideas [274] by decomposing an image
f into its level sets ¦X
a
f, a∈IR¦, where
X
a
f := ¦x ∈ IR
2
, f(x) ≥ a¦. (1.63)
A binary morphological operation A can be extended to some greyscale image f
by deﬁning
X
a
(Af) := A(X
a
f) ∀ a ∈ IR. (1.64)
We observe that for this type of morphological operations only greylevel sets
matter. As a consequence, they are invariant under monotone greylevel rescalings.
This morphological invariance (greyscale invariance) is characteristic for all meth
ods we shall study in Section 1.5 and 1.6, except for 1.5.6 and some modiﬁcations
32 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
in 1.6.5. It is a very desirable property in all cases where brightness changes of the
illumination occur or where one wants to be independent of the speciﬁc contrast
range of the camera. On the other hand, for applications like edge detection or
image restoration, contrast provides important information which should be taken
into account. Moreover, in some cases isolines may give inadequate information
about the depicted physical object boundaries.
1.5.2 Basic operations
Classic morphology analyses a shape by matching it with a socalled structuring
element, a bounded set B⊂IR
2
. Typical shapes for B are discs, squares, or ellipses.
The two basic morphological operations, dilation and erosion with a structuring
element B, are deﬁned for a greyscale image f ∈L
∞
(IR
2
) by [60]
dilation: (f ⊕B) (x) := sup ¦f(x−y), y ∈B¦, (1.65)
erosion: (f ⊖B) (x) := inf ¦f(x+y), y ∈B¦. (1.66)
These names can be easily motivated when considering a shape in a binary image
and a discshaped structuring element. In this case dilation blows up its boundaries,
while erosion shrinks them.
Dilation and erosion form the basis for constructing other morphological pro
cesses, for instance opening and closing:
opening: (f ◦ B) (x) := ((f ⊖B) ⊕B) (x), (1.67)
closing: (f • B) (x) := ((f ⊕B) ⊖B) (x). (1.68)
In the preceding shape interpretation opening smoothes the shape by breaking nar
row isthmuses and eliminating small islands, while closing smoothes by eliminating
small holes, fusing narrow breaks and ﬁlling gaps at the contours [181].
1.5.3 Continuousscale morphology
Let us consider a convex structuring element tB with a scaling parameter t >0.
Then, calculating u(t) = f ⊕tB and u(t) = f ⊖tB, respectively
13
, can be shown
to be equivalent to solving
∂
t
u(x, t) = sup
y∈B
'y, ∇u(x, t)`, (1.69)
∂
t
u(x, t) = inf
y∈B
'y, ∇u(x, t)`. (1.70)
with f as initial condition [12, 360].
13
Henceforth, we frequently use the simpliﬁed notation u(t) instead of u(., t)
1.5 CLASSIC MORPHOLOGICAL PROCESSES 33
By choosing e.g. B := ¦y ∈ IR
2
, [y[ ≤ 1¦ one obtains
∂
t
u = [∇u[, (1.71)
∂
t
u = −[∇u[. (1.72)
The solution u(t) is the dilation (resp. erosion) of f with a disc of radius t and
centre 0 as structuring element. Figure 5.5 (a) presents the temporal evolution of
a test image under such a continuousscale dilation.
Connection to curve evolution
Morphological PDEs such as (1.71) or (1.72) are closely related to shape and
curve evolutions. This can be illustrated by considering a smooth Jordan curve
C : [0, 2π] [0, ∞) →IR
2
,
C(p, t) =
x
1
(p, t)
x
2
(p, t)
(1.73)
where p is the parametrization and t is the evolution parameter. We assume that
C evolves in outer normal direction n with speed β, which may be a function of
its curvature κ :=
det(Cp,Cpp)
Cp
3
:
∂
t
C = β(κ) n, (1.74)
C(p, 0) = C
0
(p). (1.75)
One can embed the curve C(p, t) in an image u(x, t) in such a way that C is just
a level curve of u. The corresponding evolution for u is given by [319, 362, 16]
∂
t
u = β(curv(u)) [∇u[. (1.76)
where the curvature of u is
curv(u) := div
∇u
[∇u[
. (1.77)
Sometimes the image evolution (1.76) is called the Eulerian formulation of the
curve evolution (1.74), because it is written in terms of a ﬁxed coordinate system.
We observe that (1.71) and (1.72) correspond to the simple cases β = ±1.
Hence, they describe the curve evolutions
∂
t
C = ±n. (1.78)
This equation moves level sets in normal direction with constant speed. Such a
process is also named grassﬁre ﬂow or prairie ﬂow. It is closely related to the
Huygens principle for wave propagation [25]. Its importance for shape analysis
in biological vision has already been pointed out in the sixties by Blum [47]. He
simulated grassﬁre ﬂow by a selfconstructed opticomechanical device.
34 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
1.5.4 Theoretical results
Equations such as (1.78) may develop singularities and intersections even for
smooth initial data. Hence, concepts of jump conditions, entropy solutions, and
shocks have to be applied to this shape evolution [228].
A suitable framework for the image evolution equation (1.76) is provided by
the theory of viscosity solutions [103]. The advantage of this analysis is that it
allows us to treat shapes with singularities such as corners, where the classical
solution concept does not apply, but a unique weak solution in the viscosity sense
still exists.
It can be shown [90, 129, 103], that for an initial value
f ∈ BUC(IR
2
) := ¦ϕ∈L
∞
(IR
2
) [ ϕ is uniformly continuous on IR
2
¦ (1.79)
the equations (1.69),(1.70) possess a unique global viscosity solution u(x, t) which
fulﬁls the maximum–minimum principle
inf
IR
2
f ≤ u(x, t) ≤ sup
IR
2
f on IR
2
[0, ∞). (1.80)
Moreover, it is L
∞
stable: for two diﬀerent initial images f, g the corresponding
solutions u(t), v(t) satisfy
u(t)−v(t)
L
∞
(IR
2
)
≤ f −g
L
∞
(IR
2
)
. (1.81)
1.5.5 Scalespace properties
Brockett and Maragos [60] pointed out that the convexity of B is suﬃcient to
ensure the semigroup property of the corresponding dilations and erosions. This
establishes an important architectural scalespace property.
Similar results have been found by van den Boomgaard and Smeulders [53].
Moreover, they conjecture a causality property where singularities play a role sim
ilar to zerocrossings in Gaussian scalespace.
Jackway and Deriche [206, 207] prove a causality theorem for the dilation–
erosion scalespace, which is also based on local extrema instead of zerocrossings.
They establish that under erosion the number of local minima is decreasing, while
dilation reduces the number of local maxima. The location of these extrema is
preserved during their whole lifetime.
A complete scalespace interpretation is due to Alvarez, Guichard, Lions and
Morel [12]: They prove that under three architectural assumptions (semigroup
property, locality and regularity), one smoothing axiom (comparison principle) and
additional invariance requirements (greylevel shift invariance, invariance under ro
tations and translations, morphological invariance), a twodimensional scalespace
1.5 CLASSIC MORPHOLOGICAL PROCESSES 35
equation has the following form:
∂
t
u = [∇u[ F(t, curv(u)) (1.82)
Clearly, dilation and erosion belong to the class (1.82), thus being good candi
dates for morphological scalespaces. Indeed, in [12] it is shown that the converse
is true as well: all axioms that lead to (1.82) are fulﬁlled.
14
1.5.6 Generalizations
It is possible to extend morphology with a structuring element to morphology with
nonﬂat structuring functions. In this case we have to renounce invariance under
monotone grey level transformations, but we gain an interesting insight into a
process which has very much in common with Gaussian scalespace.
A dilation of an image f with a structuring function b : IR
2
→IR is deﬁned as
(f ⊕b) (x) := sup
y∈IR
2
¦f(x−y) +b(y)¦. (1.83)
This is a generalization of deﬁnition (1.65), since one can recover dilation with a
structuring element B by considering the ﬂat structuring function
b(x) :=
0 (x∈B),
−∞ (x∈B).
(1.84)
Van den Boomgaard [50, 51] and Jackway [206] proposed to dilate an image f(x)
with quadratic structuring functions of type
b(x, t) = −
[x[
2
4t
(t > 0). (1.85)
It can be shown [50, 53] that the result u(x, t) is a weak solution of
∂
t
u = [∇u[
2
, (1.86)
u(x, 0) = f(x). (1.87)
The temporal evolution of a test image under this process is illustrated in Figure
5.5 (b).
In analogy to the fact that Gaussiantype functions k(x, t) = a exp(
x
2
4t
) are the
only rotationally symmetric kernels which are separable with respect to convolu
tion, van den Boomgaard proves that the quadratic structuring functions b(x, t)
are the only rotationally invariant structuring functions which are separable with
respect to dilation [50, 51].
14
Invariance under rotations is only satisﬁed for a disc centered in 0 as structuring element.
36 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
A useful tool for understanding this similarity and many other analogies be
tween morphology and linear systems theory is the slope transform. This general
ization of the Legendre transform is the morphological equivalent of the Fourier
transform. It has been discovered simultaneously by Dorst and van den Boomgaard
[119] and by Maragos [275] in slightly diﬀering formulations.
The close relation between (1.86) and Gaussian scalespace has also triggered
Florack and Maas [142] to study a oneparameter family of isomorphisms of linear
diﬀusion which reveals (1.86) as limiting case.
1.5.7 Numerical aspects
Dilations or erosions with quadratic structuring functions are separable and, thus,
they can be implemented very eﬃciently by applying onedimensional operations.
A fast algorithm is described by van den Boomgaard [51].
For morphological operations with ﬂat structuring elements, the situation is
more complicated. Schemes for dilation or erosion which are based on curve evo
lution turn out be be diﬃcult to handle: they require prohibitive small time steps,
and suﬀer from the problem of coping with singularities and topological changes
[319, 25, 360].
For this reason it is useful to discretize the corresponding image evolution equa
tions. The widelyused Osher–Sethian schemes [319] are based on the idea to derive
numerical methods for such equations from techniques for hyperbolic conservation
laws. Overviews of these level set approaches and their various applications can be
found in [372, 374].
To illustrate the basic idea with a simple example, let us restrict ourselves to the
onedimensional dilation equation ∂
t
u=[∂
x
u[. A ﬁrstorder upwind Osher–Sethian
scheme for this process is given by
u
n+1
i
−u
n
i
τ
=
min
u
n
i
−u
n
i−1
h
, 0
2
+
max
u
n
i+1
−u
n
i
h
, 0
2
, (1.88)
where h is the pixel size, τ is the time step size, and u
n
i
denotes a discrete
approximation of u(ih, nτ).
Level set methods possess two advantages over classical settheoretic schemes
for dilation/erosion [25, 360, 218, 66]:
(a) They give excellent results for nondigitally scalable structuring elements
whose shapes cannot be represented correctly on a discrete grid, for instance
discs or ellipses.
(b) The time t plays the role of a continuous scale parameter. Therefore, the
size of a structuring element need not be multiples of the pixel size, and it
is possible to get results with subpixel accuracy.
1.6 CURVATUREBASED MORPHOLOGICAL PROCESSES 37
However, they reveal also two disadvantages:
(a) They are slower than settheoretic morphological schemes.
(b) Dissipative eﬀects such as blurring of discontinuities occur.
To address the ﬁrst problem, speedup techniques for shape evolution have been
proposed which use only points close to the curve at every time step [8, 435,
373]. Blurring of discontinuities can be minimized by applying shockcapturing
techniques such as highorder ENO
15
schemes [395, 385].
1.5.8 Applications
Continuousscale morphology has been applied to shapefromshading problems,
gridless image halftoning, distance transformations, and skeletonization. Applica
tions outside the ﬁeld of image processing and computer vision include for instance
shape oﬀsets in CAD and path planning in robotics. Overviews and suitable ref
erences can be found in [233, 276].
1.6 Curvaturebased morphological processes
Besides providing a useful reinterpretation of classic continuousscale morphology,
the PDE approach has led to the discovery of new morphological operators. These
processes are curvaturebased, and – although they cannot be written in conserva
tion form – they reveal interesting relations to diﬀusion processes. Two important
representatives of this class are mean curvature motion and its aﬃne invariant
counterpart. In this subsection we shall discuss these PDEs, possible generaliza
tions, numerical aspects, and applications.
1.6.1 Meancurvature ﬁltering
In order to motivate our ﬁrst curvaturebased morphological PDE, let us recall
that the linear diﬀusion equation (1.9) can be rewritten as
∂
t
u = ∂
ηη
u +∂
ξξ
u, (1.89)
where the unit vectors η and ξ are parallel and perpendicular to ∇u, respectively.
The ﬁrst term on the righthand side of (1.89) describes smoothing along the
ﬂowlines, while the second one smoothes along isophotes. When we want to smooth
15
ENO means essentially nonoscillatory. By adapting the stencil for derivative approximations
to the local smoothness of the solution, ENO schemes obtain both highorder accuracy in smooth
regions and sharp shock transitions [183].
38 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
the image anisotropically along its isophotes, we can neglect the ﬁrst term and end
up with the problem
∂
t
u = ∂
ξξ
u, (1.90)
u(x, 0) = f(x). (1.91)
By straightforward calculations one veriﬁes that (1.90) can also be written as
∂
t
u =
u
2
x
2
u
x
1
x
1
−2u
x
1
u
x
2
u
x
1
x
2
+u
2
x
1
u
x
2
x
2
u
2
x
1
+u
2
x
2
(1.92)
= ∆u −
1
[∇u[
2
'∇u, Hess(u)∇u` (1.93)
= [∇u[ curv(u). (1.94)
Since curv(u)=div
∇u
∇u
is the curvature of u (mean curvature for dimensions ≥
3), equation (1.94) is named (mean) curvature motion (MCM). The corresponding
curve evolution
∂
t
C(p, t) = κ(p, t) n(p, t) (1.95)
shows that (1.90) propagates isophotes in inner normal direction with a velocity
that is given by their curvature κ=
det(Cp,Cpp)
Cp
3
.
Processes of this type have ﬁrst been studied by Brakke in 1978 [54]. They
arise in ﬂame propagation, crystal growth, the derivation of minimal surfaces, grid
generation, and many other applications; see [372, 374] and the references therein
for an overview. The importance of MCM in image processing became only recently
clear: As nicely explained in a paper by Guichard and Morel [173], mean curvature
motion can be regarded as the limit process when classic morphological operators
such as median ﬁltering are iteratively applied.
Figure 5.5 (c) and 5.11 (a) present examples for mean curvature ﬁltering. Equa
tion (1.95) is also called geometric heat equation or Euclidean shortening ﬂow. The
subsequent discussions shall clarify these names.
Intrinsic heat ﬂow
Interestingly, there exists a further connection between linear diﬀusion and motion
by curvature. Let v(p, t) denote the Euclidean arclength of C(p, t), i.e.
v(p, t) :=
p
0
[C
ρ
(ρ, t)[ dρ, (1.96)
where C
ρ
:= ∂
ρ
C. The Euclidean arclength is characterized by [C
v
[ = 1. It is
invariant under Euclidean transformations, i.e. mappings
x → Rx +b (1.97)
1.6 CURVATUREBASED MORPHOLOGICAL PROCESSES 39
where R∈IR
2×2
denotes a rotation matrix and b ∈IR
2
is a translation vector.
Since it is wellknown from diﬀerential geometry (see e.g. [71], p. 14) that
κ(p, t) n(p, t) = ∂
vv
C(p, t), (1.98)
we recognize that curvature motion can be regarded as Euclidean invariant diﬀu
sion of isophotes:
∂
t
C(p, t) = ∂
vv
C(p, t). (1.99)
This geometric heat equation is intrinsic, since it is independent of the curve para
metrization. However, the reader should be aware of the fact that – although this
equation looks like a linear onedimensional heat equation – it is in fact nonlinear,
since the arclength v is again a function of the curve.
Theoretical results
For the evolution of a smooth curve under its curvature, it has been shown in
[188, 151, 169] that a smooth solution exists for some ﬁnite time interval [0, T).
A convex curve remains convex, a nonconvex one becomes convex and, for t →T,
the curve shrinks to a circular point, i.e. a point with a circle as limiting shape.
Moreover, since under all ﬂows C
t
=C
qq
the Euclidean arclength parametrization
q(p):=v(p) is the fastest way to shrink the Euclidean perimeter
[C
p
[ dp, equation
(1.99) is called Euclidean shortening ﬂow [150]. The time for shrinking a circle of
radius σ to a point is given by
T =
1
2
σ
2
. (1.100)
In analogy to the dilation/erosion case it can be shown that, for an initial image
f ∈BUC(IR
2
), equation (1.90) has a unique viscosity solution which is L
∞
stable
and satisﬁes a maximum–minimum principle [12].
Scalespace interpretation
A shape scalespace interpretation for curve evolution under Euclidean heat ﬂow is
studied by Kimia and Siddiqi [227]. It is based on results of Evans and Spruck [129].
They establish the semigroup property as architectural quality, and smoothing
properties follow from the fact that the total curvature decreases. Moreover, the
number of extrema and inﬂection points of the curvature is nonincreasing.
As an image evolution, MCM belongs to the class of morphological scalespaces
which satisfy the general axioms of Alvarez, Guichard, Lions and Morel [12], that
have been mentioned in 1.5.5.
When studying the evolution of isophotes under MCM, it can be shown that,
if one isophote is enclosed by another, this ordering is preserved [129, 227]. Such a
shape inclusion principle implies in connection with (1.100) that it takes the time
T =
1
2
σ
2
to remove all isophotes within a circle of radius σ. This shows that the
40 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
relation between temporal and spatial scale for MCM is the same as for linear
diﬀusion ﬁltering (cf. (1.14)).
Moreover, two level sets cannot move closer to one another than they were
initially [129, 227]. Hence, contrast cannot be enhanced. This property is charac
teristic for all scalespaces of the Alvarez–Guichard–Lions–Morel axiomatic and
distinguishes them from nonlinear diﬀusion ﬁlters.
1.6.2 Aﬃne invariant ﬁltering
Motivation
Although Euclidean invariant smoothing methods are suﬃcient in many applica
tions, there exist certain problems which also require invariance with respect to
aﬃne transformations. A (full) aﬃne transformation is a mapping
x → Ax +b (1.101)
where b ∈ IR
2
denotes a translation vector and the matrix A ∈ IR
2×2
is invertible.
Aﬃne transformations arise as shape distortions of planar objects when being
observed from a large distance under diﬀerent angles.
Aﬃne invariant intrinsic diﬀusion
In analogy to the Euclidean invariant heat ﬂow, Sapiro and Tannenbaum [362,
363] constructed an aﬃne invariant ﬂow by replacing the Euclidean arclength
v(p, t) in (1.99) by an “arclength” s(p, t) that is invariant with respect to aﬃne
transformations with det(A) = 1.
Such an aﬃne arclength was proposed by Blaschke [44, pp. 12–15] in 1923. It
is characterized by det(C
s
, C
ss
)=1, and it can be calculated as
s(p, t) :=
p
0
det
C
ρ
(ρ, t), C
ρρ
(ρ, t)
1
3
dρ. (1.102)
By virtue of
∂
ss
C(p, t) =
κ(p, t)
1
3
n(p, t) (1.103)
we obtain the aﬃne invariant heat ﬂow
∂
t
C(p, t) =
κ(p, t)
1
3
n(p, t), (1.104)
C(p, 0) = C
0
(p). (1.105)
1.6 CURVATUREBASED MORPHOLOGICAL PROCESSES 41
Aﬃne invariant image evolution
When regarding the curve C(p, t) as a levelline of an image u(x, t), we end up
with the evolution equation
∂
t
u = [∇u[
curv(u)
1
3
(1.106)
=
u
2
x
2
u
x
1
x
1
−2u
x
1
u
x
2
u
x
1
x
2
+u
2
x
1
u
x
2
x
2
1
3
(1.107)
= [∇u[
2
3
u
1
3
ξξ
, (1.108)
where ξ is the direction perpendicular to ∇u. The temporal evolution of an image
under such an evolution resembles mean curvature motion; see Fig. 5.6(a).
Besides the name aﬃne invariant heat ﬂow, this equation is also called aﬃne
shortening ﬂow, aﬃne morphological scalespace (AMSS), and fundamental equa
tion in image processing.
This image evolution equation has been discovered independently of and si
multaneously with the curve evolution approach of Sapiro and Tannenbaum by
Alvarez, Guichard, Lions and Morel [12] via an axiomatic scalespace approach.
After having mentioned some theoretical results, we shall brieﬂy sketch this rea
soning below.
Theoretical results
The curve evolution properties of aﬃne invariant heat ﬂow can be shown to be the
same as in the Euclidean invariant case, with three exceptions [363]:
(a) Closed curves shrink to points with an ellipse as limiting shape (elliptical
points).
(b) The name aﬃne shortening ﬂow reﬂects the fact that, under all ﬂows C
t
=
C
qq
, the aﬃne arclength parametrization q(p):=s(p) is the fastest way to
shrink the aﬃne perimeter
L(t) :=
det
C
p
(p, t), C
pp
(p, t)
1
3
dp. (1.109)
(c) The time for shrinking a circle of radius σ to a point is
T =
3
4
σ
4
3
. (1.110)
For the image evolution equation (1.106) we have the same results as for MCM
and dilation/erosion concerning wellposedness of a viscosity solution which satis
ﬁes a maximum–minimum principle [12].
42 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
Scalespace properties
Alvarez, Guichard, Lions and Morel [12] proved that (1.106) is unique (up to tem
poral rescalings) when imposing on the scalespace axioms for (1.82) an additional
aﬃne invariance axiom:
For every invertible A∈IR
2×2
and for all t ≥0, there exists a rescaled
time t
′
(t, A)≥0, such that
T
t
(Af) = A(T
t
′ f) ∀ f ∈ BUC(IR
2
). (1.111)
For this reason they call the AMSS equation also fundamental equation in image
analysis. Simpliﬁcations of this axiomatic and related axioms for shape scalespaces
can be found in [16].
The scalespace reasoning of Sapiro and Tannenbaum investigates properties
of the curve evolution, see [362] and the references therein. Based on results of
[229, 24] they point out that the Euclidean absolute curvature decreases as well
as the number of extrema and inﬂection points of curvature. Moreover, a shape
inclusion principle holds.
1.6.3 Generalizations
In order to analyse planar shapes in a way that does not depend on their location in
IR
3
, one requires a multiscale analysis which is invariant under a general projective
mapping
(x
1
, x
2
)
⊤
→
a
11
x
1
+a
21
x
2
+a
31
a
13
x
1
+a
23
x
2
+a
33
,
a
12
x
1
+a
22
x
2
+a
32
a
13
x
1
+a
23
x
2
+a
33
⊤
(1.112)
with A=(a
ij
) ∈ IR
3×3
and det A=1. Research in this direction has been carried
out by Faugeras and Keriven [130, 131, 133], Bruckstein and Shaked [62], Olver
et al. [314], and Dibos [112, 113]. It turns out that intrinsic heatequationlike
formulations for the projective group are more complicated than the Euclidean
and aﬃne invariant ones, and that there is some evidence that they do not reveal
the same smoothing scalespace properties [314]. A study of heat ﬂows which are
invariant under subgroups of the projective group can be found in [314, 113].
An intrinsic heat ﬂow for images painted on surfaces has been investigated by
Kimmel [232]. It is invariant to the bending (isometric mapping) of the surface.
This geodesic curvature ﬂow and other evolutions, both for scalar and vectorvalued
images, can be regarded as steepest descent methods of energy functionals which
have been proposed by Polyakov in the context of string theory [235].
Euclidean and aﬃne invariant curve evolutions can also be modiﬁed in order
to obtain area or lengthpreserving equations [188, 150, 352, 366].
1.6 CURVATUREBASED MORPHOLOGICAL PROCESSES 43
Recently it has been suggested to modify AMSS such that any evolution at T
or Xjunctions of isophotes is inhibited [75]. Such a ﬁltering intends to simplify
the image while preserving its “semantical atoms” in the sense of Kanizsa [219].
Generalizations of MCM or AMSS to 3D images are investigated in [91, 78, 298,
316]. In this case two principal curvatures occur. Since they can have diﬀerent sign,
the question arises of how to combine them to a process which simpliﬁes arbitrary
surfaces without creating singularities. In contrast to the 2D situation, topological
changes such as the splitting of conconvex structures may occur. This problem is
reminescent of the creation of new extrema in diﬀusion scalespaces when going
from 1D to 2D.
Aﬃne scalespace axiomatics for image sequences (movies) have been estab
lished in [12, 287, 288], and possibilities to generalize the axiomatic of Alvarez,
Guichard, Lions and Morel to colour images are discussed in [82].
1.6.4 Numerical aspects
Due to the equivalence between curve evolution and morphological image pro
cessing PDEs, we have three main classes of numerical methods: curve evolution
schemes, settheoretic morphological schemes and approximation schemes for the
Eulerian formulation. A comparison of diﬀerent methods of these classes can be
found in [95].
Curve evolution schemes are investigated by Mokhtarian and Mackworth [291],
Bruckstein et al. [61], Cohignac et al. [95], Merriman et al. [285], Ruuth [347], and
Moisan [289]. In [61] discrete analogues of MCM and AMSS for the evolution of
planar polygons are introduced. In complete analogy to the behaviour of the con
tinuous equations, convergence to polygones, whose corners belong to circles and
ellipses, respectively, is established. Related discrete curve evolutions are analysed
in [135, 359]. Curve evolution schemes can reveal perfect aﬃne invariance.
Convergent settheoretic morphological schemes for MCM and AMSS have been
proposed by Catt´e et al. [80, 79]. On the one hand, they are very fast and they
are entirely invariant under monotone greyscale transformations, on the other
hand, it is diﬃcult to ﬁnd consistent approximations on a pixel grid which have
good rotational invariance. This is essentially the same tradeoﬀ as for circular
structuring elements in classical settheoretic morphology, cf. 1.5.7.
Most direct approximations of morphological image evolution equations are
based on the Osher–Sethian schemes [319, 372, 374]. In the case of MCM or AMSS,
this leads to an explicit ﬁnite diﬀerence method which approximates the spatial
derivatives by central diﬀerences. Diﬀerent variants of these schemes have been
proposed in order to get better rotational invariance, higher stability or less dissi
pative eﬀects [10, 15, 95, 267, 362]. A comparative evaluation of these approaches
has been carried out by Lucido et al. [267]. Niessen et al. [305, 304] approximate
44 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
the spatial derivatives by Gaussian derivatives which are calculated in the Fourier
domain.
Concerning stability one observes that all these explicit schemes can violate a
discrete maximum–minimum principle and require small time steps to be experi
mentally stable. For AMSS an additional constraint appears: the behaviour of this
equation is highly nonlocal, since aﬃne invariance implies that circles are equiva
lent to ellipses of any arbitrary eccentricity. If one wants to have a good numerical
approximation of aﬃne invariance, one has to decrease the temporal step size sig
niﬁcantly below the step size of experimental stability [16, 218]. Using Gaussian
derivatives for MCM or AMSS permits larger time steps for large kernels [304], but
their calculation in the Fourier domain is computationally expensive and aliasing
may lead to oscillatory solutions, cf. 1.2.3.
One way to achieve unconditional L
∞
stability for MCM is to approximate
u
ξξ
by suitable linear combinations of onedimensional secondorder derivatives
along grid directions and to apply an implicit ﬁnite diﬀerence scheme [95, 13, 146].
Schemes of this type, however, renounce consistency with the original equation as
well as rotational invariance: round shapes evolve into polygonal structures.
A consistent semiimplicit approximation of MCM which discretizes the ﬁrst
order spatial derivatives explicitly and the secondorder derivatives implicitly has
been proposed by Alvarez [10]. In order to solve the resulting linear system of
equations he applies symmetric Gauß–Seidel iterations.
Nicolet and Sp¨ uhler [301] investigate a consistent fully implicit scheme for MCM
leading to a nonlinear system of equations. It is solved by means of nonlinear Gauß–
Seidel iterations. Comparing it with the explicit scheme they report a tradeoﬀ
between the larger time step size and the higher computational eﬀort per step.
An inherent problem of all ﬁnite diﬀerence schemes for morphological image
evolutions are their dissipative eﬀects which create additional blurring of discon
tinuities. As a remedy, one can decompose the image into binary level sets, map
them into Lipschitzcontinuous images by applying a distance transformation, and
run a ﬁnite diﬀerence method on them. Afterwards one extracts the processed level
sets as the zerolevel sets of the evolved images, and assembles the ﬁnal image by
superimposing all evolved level sets [75]. The natural price one has to pay for the
excellent results is a fairly high computational eﬀort.
A software package which contains implementations of MCM, AMSS and many
other modern techniques such as wavelets, MumfordShah segmentation, and ac
tive contour models (cf. 1.6.6) is available under the name MegaWave2.
16
16
MegaWave2 has been developed by Jacques Froment and Sylvain Parrino, CEREMADE,
University Paris IX, 75775 Paris Cedex 16, France. More information can be found under
http://www.ceremade.dauphine.fr.
1.6 CURVATUREBASED MORPHOLOGICAL PROCESSES 45
1.6.5 Applications
The special invariances of AMSS are useful for shape recognition tasks [12, 96, 97,
132], for corner detection [15], and for texture discrimination [265, 16]. MCM and
AMSS have also been applied to denoising [305, 304] and blob detection [157, 301]
in medical images and to terrain matching [268]. The potential of MCM for shape
segmentation [290] has even been used for classifying chrysanthemum leaves [1].
If one aims to use these equations for image restoration one usually modiﬁes
them by multiplying them with a term that reduces smoothing at highcontrast
edges [13, 364, 365, 304]; see also Fig. 5.6 (b). Adapting them to tasks such as tex
ture enhancement requires more sophisticated and more global feature descriptors
than the gradient, for instance an analysis by means of Gabor functions [72, 234].
Introducing a reaction term as in [102] allows to attract the solution to speciﬁed
grey values which can be used as quantization levels [11]. Another modiﬁcation
results from omitting the factor [∇u[ in the mean curvature motion (1.94), see e.g.
[126, 127]. This corresponds to nonlinear diﬀusion ﬁlters and a restoration method
by total variation minimization [345] which shall be described in 1.7.2.
In order to improve images, MCM or AMSS have also been combined with other
processes such as linear diﬀusion [13], shock ﬁltering ([14], cf. 1.7.1) or global PDEs
for histogramme enhancement [358].
Malladi and Sethian [272] propose to replace MCM by a technique in which
the motion of level curves is based on either min(κ, 0) or max(κ, 0), depending on
the average grey value within a certain neighbourhood. This socalled min–max
ﬂow produces a restored image as steadystate solution and reveals good denoising
properties. Wellposedness results are not known up to now, since the theory of
viscosity solutions is no longer applicable.
All these preceding modiﬁcations are at the expense of renouncing the mor
phological invariance of the genuine operators (and also aﬃne invariance in the
case of [364, 365, 304], unless an “aﬃne invariant gradient” [231, 315] is used). If
one wants to stay within the morphological framework one can combine diﬀerent
morphological processes, for instance MCM and dilation/erosion. This leads to a
process which is useful for analysing components of shape [228, 230, 383, 384, 453],
and which is called entropy scalespace or reaction–diﬀusion scalespace.
Recently Steiner et al. proposed a method for caricaturelike shape exaggera
tion [394]. They evolved the boundary curve by means of a backward Euclidean
shortening ﬂow with a stabilizing bias term as in (1.56).
It is interesting to note that, already in 1965, Gabor – the inventor of optical
holography and the socalled Gabor functions – proposed a deblurring algorithm
based on combining MCM with backward smoothing along ﬂowlines [149, 260].
This longtime forgotten method is similar to the Perona–Malik process (1.36) for
large image gradients.
46 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
1.6.6 Active contour models
One of the main applications of curve evolution ideas appears in the context of
active contour models (deformable models). 2D versions are also called snakes,
while 3D active models are sometimes named active surfaces or active blobs. Ac
tive contour models can be used to search image features in an interactive way.
Especially for assisting the segmentation of medical images they have become very
popular [282]: The expert user gives a good initial guess of an interesting contour
(organ, tumour, ...), which will be carried the rest of the way by some energy
minimization. Apart from these practical merits, snake models incorporate many
ideas ranging from energy minimization over curve evolution to the Perona–Malik
ﬁlter and diﬀusion–reaction models. It is therefore not surprising that they play an
important role in several generalization and uniﬁcation attempts [356, 380, 452].
Explicit snakes
Kass, Witkin and Terzopoulos proposed the ﬁrst active contour model in a journal
paper in 1988 [221]. Their snakes can be thought of as an energyminimizing spline,
which is attracted by image features such as edges. For this reason, the energy
functional consists of two parts: an internal energy fraction which controls the
smoothness of the result, and an external energy term attracting the result to
salient image features.
Such a snake is represented by a curve C(s) = (x
1
(s), x
2
(s))
⊤
which minimizes
the functional
E(C(s)) =
C(s)
α
2
[C
s
(s)[
2
+
β
2
[C
ss
(s)[
2
−γ [∇f(C(s))[
2
ds. (1.113)
The ﬁrst summand is a membrane term causing the curve to shrink, the second
one is a rigidity term which encourages a straight contour, and the third term
pushes the contour to high gradients of the image f. We observe that terms 1
and 2 describe the internal energy, while the third one represents the external
(image) energy. The nonnegative parameters α, β and γ serve as weights of these
expressions. Since this model makes direct use of the snake contour, it is also called
an explicit model.
Minimizing the functional (1.113) by steepest descent gives
∂C
∂t
= αC
ss
−βC
ssss
−γ∇([∇f[
2
), (1.114)
which can be approximated by ﬁnite diﬀerences. A 3D version of such an active
contour model is presented in [401].
Usually, the result will depend on the choice of the initial curve and a good
segmentation requires an initial curve which is close to the ﬁnal segment. The main
1.6 CURVATUREBASED MORPHOLOGICAL PROCESSES 47
disadvantage of the preceding model is its topological rigidity: a classical explicit
snake cannot split in order to segment several objects simultaneously
17
. In practice,
it is also sometimes not easy to ﬁnd a good balance between the parameters α, β
and γ.
Implicit snakes
In 1993 some of the inherent diﬃculties of explicit snakes have been solved by
replacing them by a socalled implicit model. It was discovered by Caselles, Catt´e,
Coll and Dibos [73], and later on by Malladi, Sethian and Vemuri [273].
The idea behind implicit snakes is to embed the initial curve C
0
(s) as a zero
level curve into a function u
0
: IR
2
→ IR, for instance by using the distance
transformation. Then u
0
is evolved under a PDE which includes knowledge about
the original image f:
∂
t
u = g([∇f
σ
[
2
) [∇u[
div
∇u
[∇u[
+ ν
. (1.115)
This evolution is stopped at some time T, when the process does hardly alter
anymore, and the ﬁnal contour C is extracted as the zero level curve of u(x, T).
The terms in (1.115) have the following meaning:
• [∇u[ div (∇u/[∇u[) is the curvature term of MCM which smoothes level sets;
see (1.94).
• ν[∇u[ describes motion in normal direction, i.e. dilation or erosion depending
on the sign of ν. This socalled balloon force [93] is required for pushing a
level set into concave regions, a compensation for the property of MCM to
create convex regions.
• g is a stopping function such as the Perona–Malik diﬀusivity (1.32): it be
comes small for large [∇f
σ
[ = [∇K
σ
∗f[. Hence, it slows down the snake as
soon as it approaches signiﬁcant edges of the original image f.
For this model Caselles et al. could prove wellposedness in the viscosity sense
[73]. Whitaker and Chen developed similar implicit active contour models for 3D
images [436, 435], and Caselles and Coll investigated related approaches for image
sequences [74].
An advantage of implicit snake models is their topological ﬂexibility: The con
tour may split. This allows simultaneous segmentation of multiple objects. More
over, they use essentially only one remaining parameter, the balloon force ν. On
17
Recently McInerley and Terzopoulos have proposed modiﬁed explicit deformable models
which allow topological changes [281, 283].
48 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
the other hand, the process does not really stop at the desired result, it only slows
down, and it is diﬃcult to interpret implicit snakes in terms of energy minimiza
tion. In order to address the initialization problem, Tek and Kimia use implicit
active contours starting from randomly chosen seed points, both in 2D [399] and
in 3D [400]. They name this technique reaction–diﬀusion bubbles.
Geodesic snakes
Geodesic snakes make a synthesis of explicit and implicit snake ideas. They have
been proposed simultaneously by Caselles, Kimmel and Sapiro [76] and by Kichenas
samy, Kumar, Olver, Tannenbaum and Yezzi [226]. These snakes replace the con
tour energy (1.113) by
E(C) =
C(s)
α
2
[C
s
(s)[
2
−γ g([∇f
σ
(C)[
2
)
ds (1.116)
where g denotes again a Perona–Malik diﬀusivity of type (1.32). Under some addi
tional assumptions they derive that minimizing (1.116) is equivalent to searching
min
C
C(s)
g
[∇f
σ
(C(s))[
2
[C
s
(s)[ ds. (1.117)
This is nothing else than ﬁnding a curve of minimal distance (geodesic) with respect
to some imageinduced metric. Embedding the initial curve as a level set of some
image u
0
and applying a descent method to the corresponding EulerLagrange
equation leads to the image evolution PDE
∂
t
u = [∇u[ div
g([∇f
σ
[
2
)
∇u
[∇u[
. (1.118)
This active contour model is parameterfree, but often a speed term νg([∇f
σ
[
2
)[∇u[
is added to achieve faster and more stable attraction to edges. A theoretical analysis
of geodesic snakes concerning existence, uniqueness and stability of a viscosity
solution can be found in [76, 226], and extensions to 3D images are studied in [77,
226, 271]. Recently also an aﬃne invariant analogue to geodesic active contours has
been proposed [315]. Techniques which can be related to geodesic active contours
have also been used for solving 3D vision problems such as stereo [134] and motion
analysis [322].
Selfsnakes
The properties of geodesic snakes induced Sapiro to use a related technique for
image enhancement [357]: he replaced g([∇f
σ
[
2
) in (1.118) by g([∇u
σ
[
2
). Then the
1.7 TOTAL VARIATION METHODS 49
snake becomes a selfsnake no longer underlying external image forces. For σ = 0
this gives
∂
t
u = [∇u[ div
g([∇u[
2
)
∇u
[∇u[
(1.119)
= g([∇u[
2
) u
ξξ
+∇
⊤
g([∇u[
2
)
∇u (1.120)
with ξ ⊥ ∇u. Although this equation cannot be cast in divergence form, we observe
striking similarities with the Perona–Malik process from Section 1.3.1: the latter
can be written as
∂
t
u = g([∇u[
2
) ∆u +∇
⊤
g([∇u[
2
)
∇u. (1.121)
Thus, it cannot be excluded that a selfsnake without spatial regularization reveals
the same illposedness problems as the Perona–Malik ﬁlter [356]. For σ > 0, how
ever, Chen, Vemuri and Wong [89] established existence and stability of a unique
viscosity solution to a modiﬁed selfsnake process. Their model contains a reaction
term which inhibits smoothing at edges and keeps the ﬁltered image u close to the
original image f; cf. [380]. The restored image is given by the steadystate of
∂
t
u = [∇u[ div
g([∇u
σ
[
2
)
∇u
[∇u[
+β [∇u[ (f −u) (β > 0). (1.122)
The temporal evolution of a regularized selfsnake without reaction term is de
picted in Fig. 5.6(c). Generalizations of selfsnakes to vectorvalued images [357,
361] can be obtained using Di Zenzo’s ﬁrst fundamental form for colour images
[114]; see also [414, 422] for related ideas.
1.7 Total variation methods
Inspired by observations from ﬂuid dynamics where the total variation (TV)
TV (u) :=
Ω
[∇u[ dx (1.123)
plays an important role for shock calculations, one may ask if it is possible to apply
related ideas to image processing. This would be useful to restore discontinuities
such as edges.
Below we shall focus on two important TVbased image restoration techniques
which have been pioneered by Osher and Rudin: TVpreserving methods and tech
niques which are TVminimizing subject to certain constraints.
18
18
Another image enhancement method that is close in spirit is due to Eidelman, Grossmann
and Friedman [125]. It maps the image grey values to gas dynamical parameters and solves the
compressible Euler equations using shockcapturing total variation diminishing (TVD) techniques
based on Godunov’s method.
50 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
1.7.1 TVpreserving methods
In 1990 Osher and Rudin have proposed to restore blurred images by shock ﬁltering
[317]. These ﬁlters calculate the restored image as the steadystate solution of the
problem
∂
t
u = −[∇u[ F(L(u)), (1.124)
u(x, 0) = f(x). (1.125)
Here, sgn(F(u)) = sgn(u), and L(u) is a secondorder elliptic operator whose zero
crossings correspond to edges, e.g. the Laplacian L(u) = ∆u or the secondorder
directional derivative L(u)=u
ηη
with η ∇u.
By means of our knowledge from morphological processes, we recognize that
this ﬁlter aims to produce a ﬂow ﬁeld that is directed from the interior of a region
towards its edges where it develops shocks. Thus, the goal is to obtain a piecewise
constant steadystate solution with discontinuities only at the edges of the initial
image.
It has been shown that a onedimensional version of this ﬁlter preserves the
total variation and satisﬁes a maximum–minimum principle, both in the continuous
and discrete case. For the twodimensional case not many theoretical results are
available except for a discrete maximum–minimum principle.
Recently van den Boomgaard [52] pointed out that the 1D version of (1.124)
with F(u) := sgn(u) arises as the PDE formulation of a classical image enhance
ment algorithm by Kramer and Bruckner [243]. Kramer and Bruckner proved in
1975 that their Ndimensional discrete scheme converges after a ﬁnite number of
iterations to a state where each point is a local extremum.
Osher and Rudin have also proposed another TVpreserving deblurring tech
nique [318]. It solves the linear diﬀusion equation backwards in time under the
regularizing constraint that the total variation remains constant. This stabiliza
tion can be realized by keeping local extrema ﬁxed during the whole evolution.
From a practical point of view, TVpreserving methods suﬀer from the problem
that ﬂuctuations due to noise do also create shocks. For this reason, Alvarez and
Mazorra [14] replace the operator L(u) =u
ηη
in (1.124) by a Gaussiansmoothed
version L(K
σ
∗u) = K
σ
∗u
ηη
and supplement the resulting equation with a noise
eliminating mean curvature process. They prove that their semiimplicit ﬁnite
diﬀerence scheme has a unique solution which satisﬁes a maximum–minimum prin
ciple.
1.7.2 TVminimizing methods
Total variation is good for quantifying the simplicity of an image since it mea
sures oscillations without unduly punishing discontinuities. For this reason, blocky
1.7 TOTAL VARIATION METHODS 51
images (consisting only of a few almost piecewise constant segments) reveal very
small total variation.
In order to restore noisy blocky images, Rudin, Osher and Fatemi [345] have
proposed to minimize the total variation under constraints which reﬂect assump
tions about noise.
19
To ﬁx ideas, let us study an example. Given an image f with additive noise of
zero mean and known variance σ
2
, we seek a restoration u satisfying
min
u
Ω
[∇u[ dx (1.126)
subject to
Ω
(u−f)
2
dx = σ
2
, (1.127)
Ω
u dx =
Ω
f dx. (1.128)
In order to solve this constrained variational problem, PDE methods can be
applied: A solution of (1.126)–(1.128) veriﬁes necessarily the Euler equation
div
∇u
[∇u[
−µ −λ(u−f) = 0 (1.129)
with homogeneous Neumann boundary conditions. The (unknown) Lagrange mul
tipliers µ and λ have to be determined in such a way that the constraints are
fulﬁlled. Interestingly, (1.129) looks similar to the steadystate equation of the
diﬀusion–reaction equation (1.56), but – in contrast to TV approaches – equa
tion (1.56) is not intended to satisfy the noise constraint exactly [346]. Moreover,
the divergence term in (1.129) is identical with the curvature, which relates TV
minimizing techniques to MCM.
In [345] a gradient descent method is proposed to solve (1.129). It uses an
explicit ﬁnite diﬀerence scheme with central and onesided spatial diﬀerences and
adapts the Lagrange multiplier by means of the gradient projection method of
Rosen.
One may also reformulate the constrained TV minimization as an uncon
strained problem [83]: The penalized least square problem
min
u
1
2
u−f
2
L
2
(Ω)
+α
Ω
[∇u[ dx
(1.130)
is equivalent to the constrained TV minimization, if α is related to the Lagrange
multiplier λ via α =
1
λ
.
19
Related ideas have also been developed by Geman and Reynolds [153].
52 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
In recent years, problems of type (1.130) have attracted much interest from
mathematicians working on inverse problems, optimization, or numerical analysis
[2, 84, 85, 87, 115, 117, 118, 204, 205, 253, 409]. To overcome the problem that
the total variation integral contains the nondiﬀerentiable argument [∇u[, one ap
plies regularization strategies or techniques from nonsmooth optimization. Much
research is done in order to ﬁnd eﬃcient numerical methods for which convergence
can be established.
TVminimizing methods have been generalized in diﬀerent ways:
• The constrained TVminimization idea is frequently adapted to other con
straints such as blur, noise with blur, or other types of noise [262, 344, 346,
116, 253, 410, 83]. Lions et al. [262] and Dobson and Santosa [115] have
shown the existence of BV(Ω)solutions for problems of this type. Recently,
Chambolle and Lions [83] have extended the existence proof to noncompact
operators (which comprises also the situation without blur), and they have
established uniqueness.
• The tendency of TVminimizing to create piecewise constant structures can
cause undesired eﬀects such as the creation of staircases at sigmoidlike edges
[116, 83]. As a remedy, it has been proposed to minimize the L
1
norm of
expressions containing also higherorder derivatives [344, 83]. Another pos
sibility is to consider the constrained minimization of
B(u) :=
Ω
[∇u[
p(∇u)
dx, (1.131)
where p([∇u[) decreases from 2 to 1, as [∇u[ ranges from 0 to ∞; see [46].
• TVminimizing methods have also been used for estimating discontinuous
blurring kernels such as motion or outoffocus blur from a degraded image.
This leads to TVbased blind deconvolution algorithms [86].
• They have been applied to colour images [45], where the generalized TV
norm is chosen as the l
2
norm of the TV norms of the separate channels.
• Strong and Chan have identiﬁed the parameter α in (1.130) as a scale para
meter [396]. By adapting α to the local image structure, they establish re
lations between TVminimizing methods and nonlinear diﬀusion techniques
[397].
Total variation methods have been applied to restoring images of military rele
vance [345, 346, 262, 253], to improving material from criminal and civil investiga
tions as court evidence [344], and to enhancing pictures from confocal microscopy
[409] and tomography [115, 396]. They are useful for enhancing reconstruction al
gorithms for inverse scattering problems [37], and the idea of L
1
norm minimization
has also led to improved optic ﬂow algorithms [245].
1.7 TOTAL VARIATION METHODS 53
1.8 Conclusions and further scope of the book
Although we have seen that there exists a large variety of PDEbased scalespace
and image restoration methods which oﬀer many advantages, we have also become
aware of some limitations. They shall serve as a motivation for the theory which
will be explored in the subsequent chapters.
Linear diﬀusion and morphological scalespaces are wellposed and have a solid
axiomatic foundation. On the other hand, for some applications, they possess the
undesirable property that they do not permit contrast enhancement and that they
may blur and delocalize structures.
Pure restoration methods such as diﬀusion–reaction equations or TVbased
techniques do allow contrast enhancement and lead to stable structures but can
suﬀer from theoretical or practical problems, for instance unsolved wellposedness
questions or the search for eﬃcient minimizers of nonconvex or nondiﬀerentiable
functionals. Moreover, most imageenhancing PDE methods focus on edge detec
tion and segmentation problems. Other interesting image restoration topics have
found less attention.
For both scalespace and restoration methods many questions concerning their
discrete realizations are still open: discrete scalespace results are frequently miss
ing, minimization algorithms can get trapped in a poor local minimum, or the use
of explicit schemes causes restrictive step size limitations.
The goal of the subsequent chapters is to develop a theory for nonlinear aniso
tropic diﬀusion ﬁlters which addresses some of the abovementioned shortcomings.
In particular, we shall see that anisotropic nonlinear diﬀusion processes can share
many advantages of the scalespace and the image enhancement world. A scale
space interpretation is presented which does not exclude contrast enhancement,
and wellposedness results are established. Both scalespace and wellposedness
properties carry over from the continuous to the semidiscrete and discrete setting.
The latter comprises for instance semiimplicit techniques for which unconditional
stability in the L
∞
norm is proved. The general framework, for which the results
hold, includes also linear and isotropic nonlinear diﬀusion ﬁlters. Finally, speciﬁc
anisotropic models are presented which permit applications beyond segmentation
and edge enhancement tasks, for instance enhancement of coherent ﬂowlike struc
tures in textures.
54 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
Chapter 2
Continuous diﬀusion ﬁltering
This chapter presents a general continuous model for anisotropic diﬀusion ﬁlters,
analyses its theoretical properties and gives a scalespace interpretation. To this
end, we adapt the diﬀusion process to the structure tensor, a wellknown tool
for analysing local orientation. Under fairly weak assumptions on the class of ﬁl
ters, it is possible to establish wellposedness and regularity results and to prove a
maximum–minimum principle. Since the proof does not require any monotony as
sumption it applies also to contrastenhancing diﬀusion processes. After sketching
invariances of the resulting scalespace, we focus on analysing its smoothing prop
erties. We shall see that, besides the extremum principle, a large class of associated
Lyapunov functionals plays an important role in this context [414, 415].
2.1 Basic ﬁlter structure
Let us consider a rectangular image domain Ω := (0, a
1
) (0, a
2
) with boundary
Γ := ∂Ω and let an image be represented by a mapping f ∈ L
∞
(Ω). The class
of anisotropic diﬀusion ﬁlters we are concerned with is represented by the initial
boundary value problem
∂
t
u = div (D ∇u) on Ω (0, ∞), (2.1)
u(x, 0) = f(x) on Ω, (2.2)
'D∇u, n` = 0 on Γ (0, ∞). (2.3)
Hereby, n denotes the outer normal and '., .` the Euclidean scalar product on IR
2
.
In order to adapt the diﬀusion tensor D ∈ IR
2×2
to the local image structure, one
would usually let it depend on the edge estimator ∇u
σ
(cf. 1.3.3), where
u
σ
(x, t) := (K
σ
∗ ˜ u(., t)) (x) (σ > 0) (2.4)
and ˜ u denotes an extension of u from Ω to IR
2
, which may be obtained by mirroring
at Γ (cf. [81]).
55
56 CHAPTER 2. CONTINUOUS DIFFUSION FILTERING
However, we shall choose a more general structure descriptor which comprises
the edge detector ∇u
σ
, but also allows to extract more information. This will be
presented next.
2.2 The structure tensor
In order to identify features such as corners or to measure the local coherence of
structures, we need methods which take into account how the orientation of the
(smoothed) gradient changes within the vicinity of any investigated point.
The structure tensor – also called interest operator, scatter matrix or (windowed
second) moment tensor – is an important representative of this class. Matrices of
this type are useful for many diﬀerent tasks, for instance for analysing ﬂowlike
textures [340, 31], corners and Tjunctions [145, 182, 310, 309], shape cues [256,
pp. 349–382] and spatio–temporal image sequences [209, pp. 147–153], [168, pp.
219–258]. Related approaches can also be found in [39, 40, 220]. Let us focus on
some aspects which are of importance in our case.
In order to study orientations instead of directions, we have to identify gra
dients which diﬀer only by their sign: they share the same orientation, but point
in opposite directions. To this end, we reconsider the vectorvalued structure de
scriptor ∇u
σ
within a matrix framework. The matrix J
0
resulting from the tensor
product
J
0
(∇u
σ
) := ∇u
σ
⊗∇u
σ
:= ∇u
σ
∇u
T
σ
(2.5)
has an orthonormal basis of eigenvectors v
1
, v
2
with v
1
 ∇u
σ
and v
2
⊥ ∇u
σ
.
The corresponding eigenvalues [∇u
σ
[
2
and 0 give just the contrast (the squared
gradient) in the eigendirections.
Averaging this orientation information can be accomplished by convolving
J
0
(∇u
σ
) componentwise with a Gaussian K
ρ
. This gives the structure tensor
J
ρ
(∇u
σ
) := K
ρ
∗ (∇u
σ
⊗∇u
σ
) (ρ ≥ 0). (2.6)
It is not hard to verify that the symmetric matrix J
ρ
=
j
11
j
12
j
12
j
22
is positive semidef
inite and possesses orthonormal eigenvectors v
1
, v
2
with
v
1

¸
2j
12
j
22
−j
11
+
(j
11
−j
22
)
2
+ 4j
2
12
¸
. (2.7)
The corresponding eigenvalues µ
1
and µ
2
are given by
µ
1,2
=
1
2
j
11
+j
22
±
(j
11
−j
22
)
2
+ 4j
2
12
, (2.8)
where the + sign belongs to µ
1
. As they integrate the variation of the grey values
within a neighbourhood of size O(ρ), they describe the average contrast in the
2.3. THEORETICAL RESULTS 57
eigendirections. Thus, the integration scale ρ should reﬂect the characteristic win
dow size over which the orientation is to be analysed. Presmoothing in order to
obtain ∇u
σ
makes the structure tensor insensitive to noise and irrelevant details
of scales smaller than O(σ). The parameter σ is called local scale or noise scale.
By virtue of µ
1
≥µ
2
≥0, we observe that v
1
is the orientation with the highest
grey value ﬂuctuations, and v
2
gives the preferred local orientation, the coherence
direction. Furthermore, µ
1
and µ
2
can be used as descriptors of local structure:
Constant areas are characterized by µ
1
= µ
2
= 0, straight edges give µ
1
≫µ
2
= 0,
corners can be identiﬁed by µ
1
≥ µ
2
≫0, and the expression
(µ
1
−µ
2
)
2
= (j
11
−j
22
)
2
+ 4j
2
12
(2.9)
becomes large for anisotropic structures. It is a measure of the local coherence.
An example which illustrates the advantages of the structure tensor for analysing
coherent patterns can be found in Figure 5.10 (d); see Section 5.2.
2.3 Theoretical results
In order to discuss wellposedness results, let us ﬁrst recall some useful notations.
Let H
1
(Ω) be the Sobolev space of functions u(x) ∈ L
2
(Ω) with all distributional
derivatives of ﬁrst order being in L
2
(Ω). We equip H
1
(Ω) with the norm
u
H
1
(Ω)
:=
u
2
L
2
(Ω)
+
2
¸
i=1
∂
x
i
u
2
L
2
(Ω)
1/2
(2.10)
and identify it with its dual space. Let L
2
(0, T; H
1
(Ω)) be the space of functions
u, strongly measurable on [0, T] with range in H
1
(Ω) (for the Lebesgue measure
dt on [0, T]) such that
u
L
2
(0,T;H
1
(Ω))
:=
¸
T
0
u(t)
2
H
1
(Ω)
dt
¸
1/2
< ∞. (2.11)
In a similar way, C([0, T]; L
2
(Ω)) is deﬁned as the space of continuous functions
u : [0, T] →L
2
(Ω) supplemented with the norm
u
C([0,T];L
2
(Ω))
:= max
[0,T]
u(t)
L
2
(Ω)
. (2.12)
As usual, we denote by C
p
(X, Y ) the set of C
p
mappings from X to Y .
Now we can give a precise formulation of the problem we are concerned with.
We need the following prerequisites:
58 CHAPTER 2. CONTINUOUS DIFFUSION FILTERING
Assume that f ∈L
∞
(Ω), ρ≥0, and σ, T >0.
Let a:=ess inf
Ω
f, b := ess sup
Ω
f, and consider the problem
∂
t
u = div (D(J
ρ
(∇u
σ
)) ∇u) on Ω (0, T],
u(x, 0) = f(x) on Ω,
'D(J
ρ
(∇u
σ
))∇u, n` = 0 on Γ (0, T],
where the diﬀusion tensor D = (d
ij
) satisﬁes the following
properties:
(C1) Smoothness:
D ∈ C
∞
(IR
2×2
, IR
2×2
).
(C2) Symmetry:
d
12
(J)=d
21
(J) for all symmetric matrices J ∈IR
2×2
.
(C3) Uniform positive deﬁniteness:
For all w∈L
∞
(Ω, IR
2
) with [w(x)[ ≤ K on
¯
Ω, there
exists a positive lower bound ν(K) for the eigenvalues
of D(J
ρ
(w)).
(P
c
)
Under these assumptions the following theorem, which generalizes and extends
results from [81, 414], can be proved.
Theorem 1 (Wellposedness,
1
regularity, extremum principle)
The problem (P
c
) has a unique solution u(x, t) in the distributional sense which
satisﬁes
u ∈ C([0, T]; L
2
(Ω)) ∩ L
2
(0, T; H
1
(Ω)), (2.13)
∂
t
u ∈ L
2
(0, T; H
1
(Ω)). (2.14)
Moreover, u∈C
∞
(
¯
Ω(0, T]). This solution depends continuously on f with respect
to  . 
L
2
(Ω)
, and it fulﬁls the extremum principle
a ≤ u(x, t) ≤ b on Ω (0, T]. (2.15)
1
For a complete wellposedness proof one also has to establish stability with respect to per
turbations of the diﬀusion equation. This topic will not be addressed here.
2.3. THEORETICAL RESULTS 59
Proof:
(a) Existence, uniqueness and regularity
Existence, uniqueness and regularity are straightforward anisotropic exten
sions of the proof for the isotropic case studied by Catt´e, Lions, Morel and
Coll [81]. Therefore, we just sketch the basic ideas of this proof.
Existence can be proved using Schauder’s ﬁxed point theorem. One considers
the solution U(w) of a distributional linear version of (P
c
) where D depends
on some function w instead of u. Then one shows that U is a weakly contin
uous mapping from a nonempty, convex and weakly compact subset W
0
of
W(0, T) :=
w ∈ L
2
(0, T; H
1
(Ω)),
dw
dt
∈ L
2
(0, T; H
1
(Ω))
¸
into itself. Since
W(0, T) is contained in L
2
(0, T; L
2
(Ω)), with compact inclusion, U reveals a
ﬁxed point u ∈ W
0
, i.e. u = U(u).
Smoothness follows from classical bootstrap arguments and the general the
ory of parabolic equations [246]. Since u(t) ∈ H
1
(Ω) for all t > 0, one deduces
that u(t) ∈ H
2
(Ω) for all t > 0. By iterating, one can establish that u is a
strong solution of (P
c
) and u ∈ C
∞
((0, T]
¯
Ω).
The basic idea of the uniqueness proof consists of using energy estimates for
the diﬀerence of two solutions, such that the Gronwall–Bellman inequality
can be applied. Then, uniqueness follows from the fact that both solutions
start with the same initial values.
Finally an iterative linear scheme is investigated, whose solution is shown to
converge in C([0, T]; L
2
(Ω)) to the strong solution of (P
c
).
(b) Extremum principle
In order to prove a maximum–minimum principle, we utilize Stampacchia’s
truncation method (cf. [58], p. 211).
We restrict ourselves to proving only the maximum principle. The minimum
principle follows from the maximum principle when being applied to the
initial datum −f.
Let G ∈ C
1
(IR) be a function with G(s) = 0 on (−∞, 0] and 0 < G
′
(s) ≤ C
on (0, ∞) for some constant C. Now, we deﬁne
H(s) :=
s
0
G(σ) dσ, s ∈ IR,
ϕ(t) :=
Ω
H(u(x, t) −b) dx, t ∈ [0, T].
By the Cauchy–Schwarz inequality, we have
Ω
[G(u(x, t)−b) ∂
t
u(x, t)[ dx ≤ C u(t)−b
L
2
(Ω)
∂
t
u(t)
L
2
(Ω)
60 CHAPTER 2. CONTINUOUS DIFFUSION FILTERING
and by virtue of (2.13), (2.14) we know that the righthand side of this
estimate exists. Therefore, ϕ is diﬀerentiable for t > 0, and we get
dϕ
dt
=
Ω
G(u−b) ∂
t
u dx
=
Ω
G(u−b) div (D(J
ρ
(∇u
σ
)) ∇u) dx
=
Γ
G(u−b) 'D(J
ρ
(∇u
σ
)) ∇u, n`
. .. .
=0
dS
−
Ω
G
′
(u−b)
. .. .
≥0
'∇u, D(J
ρ
(∇u
σ
)) ∇u`
. .. .
≥0
dx
≤ 0. (2.16)
By means of H(s) ≤
C
2
s
2
, we have
0 ≤ ϕ(t) ≤
Ω
H(u(x, t)−f(x)) dx ≤
C
2
u(t)−f
2
L
2
(Ω)
. (2.17)
Since u ∈ C([0, T]; L
2
(Ω)), the righthand side of (2.17) tends to 0 = ϕ(0)
for t →0
+
which proves the continuity of ϕ(t) in 0. Now from
ϕ ∈ C[0, T], ϕ(0) = 0, ϕ ≥ 0 on [0, T]
and (2.16), it follows that
ϕ ≡ 0 on [0, T].
Hence, for all t ∈ [0, T], we obtain u(x, t)−b ≤ 0 almost everywhere (a.e.) on
Ω. Due to the smoothness of u for t > 0, we ﬁnally end up with the assertion
u(x, t) ≤ b on
¯
Ω (0, T].
(c) Continuous dependence on the initial image
Let f, h ∈ L
∞
(Ω) be two initial values and u, w the corresponding solutions.
In the same way as in the uniqueness proof in [81], one shows that there
exists some constant c > 0 such that
d
dt
u(t)−w(t)
2
L
2
(Ω)
≤ c ∇u(t)
2
L
2
(Ω)
u(t)−w(t)
2
L
2
(Ω)
.
Applying the Gronwall–Bellman lemma [57, pp. 156–137] yields
u(t)−w(t)
2
L
2
(Ω)
≤ f −h
2
L
2
(Ω)
exp
¸
c
t
0
∇u(s)
2
L
2
(Ω)
ds
¸
.
2.3. THEORETICAL RESULTS 61
By means of the extremum principle we know that u is bounded on
¯
Ω[0, T].
Thus, ∇u
σ
is also bounded, and prerequisite (C3) implies that there exists
some constant ν = ν(σ, f
L
∞
(Ω)
) > 0, such that
t
0
∇u(s)
2
L
2
(Ω)
ds
≤
T
0
∇u(s)
2
L
2
(Ω)
ds
≤
1
ν
T
0
Ω
'∇u(x, s), D(J
ρ
(∇u
σ
(x, s)))∇u(x, s)` dx
ds
=
1
ν
T
0
Ω
u(x, s) div
D(J
ρ
(∇u
σ
(x, s)))∇u(x, s)
dx
ds
≤
1
ν
T
0
u(s)
L
2
(Ω)
∂
t
u(s)
L
2
(Ω)
ds
≤
1
ν
u
L
2
(0,T;H
1
(Ω))
∂
t
u
L
2
(0,T;H
1
(Ω))
.
By virtue of (2.13), (2.14), we know that the righthand of this estimate
exists. Now, let ǫ > 0 and choose
δ := ǫ exp
−c
2ν
u
L
2
(0,T;H
1
(Ω))
∂
t
u
L
2
(0,T;H
1
(Ω))
.
Then for f −h
L
2
(Ω)
< δ, the preceding results imply
u(t)−w(t)
L
2
(Ω)
< ǫ ∀ t ∈ [0, T],
which proves the continuous dependence on the initial data. 2
Remarks:
(a) We observe a strong smoothing eﬀect which is characteristic for many dif
fusion processes: under fairly weak assumptions on the initial image (f ∈
L
∞
(Ω)) we obtain an inﬁnitely often diﬀerentiable solution for arbitrary small
positive times. More restrictive requirements – for instance f ∈ BUC(IR
2
)
in order to apply the theory of viscosity solutions – are not necessary in our
case.
(b) Moreover, our proof does not require any monotony assumption. This has the
advantage that contrastenhancing processes are permitted as well. Chapter
5 will illustrate this by presenting examples where contrast is enhanced.
62 CHAPTER 2. CONTINUOUS DIFFUSION FILTERING
(c) The continuous dependence of the solution on the initial image has signif
icant practical impact as it ensures stability with respect to perturbations
of the original image. This is of importance when considering stereo image
pairs, spatiotemporal image sequences or slices from medical CT or MRI
sequences, since we know that similar images remain similar after ﬁltering.
2
(d) The extremum principle oﬀers the practical advantage that, if we start for
instance with an image within the range [0, 255], we will never obtain results
with grey value such as 257. It is also closely related to smoothing scalespace
properties, as we shall see in 2.4.2.
(e) The wellposedness results are essentially based on the fact that the regu
larization by convolution with a Gaussian allows to estimate ∇u
σ

L
∞
(Ω)
by
u
L
∞
(Ω)
. This property is responsible for the uniform positive deﬁniteness
of the diﬀusion tensor.
2.4 Scalespace properties
Let us now investigate scalespace properties of the class (P
c
) and juxtapose the re
sults to other scalespaces. To this end, we shall not focus on further investigations
of architectural requirements like recursivity, regularity and locality, as these qual
ities do not distinguish nonlinear diﬀusion scalespaces from other ones. We start
with brieﬂy discussing invariances. Afterwards, we turn to a more crucial task,
namely the question in which sense our evolution equation – which may allow con
trast enhancement – can still be considered as a smoothing, informationreducing
image transformation.
2.4.1 Invariances
Let u(x, t) be the unique solution of (P
c
) and deﬁne the scalespace operator T
t
by
T
t
f := u(t), t ≥ 0, (2.18)
where u(t) := u(., t).
The properties we discuss now illustrate that an invariance of T
t
with respect to
some image transformation P is characterized by the fact that T
t
and P commute.
Much of the terminology used below is borrowed from [12].
2
This does not contradict contrast enhancement: In the case of two similar images, where
one leads to contrast enhancement and the other not, the regularization damps the enhancement
process in such a way that both images do not diﬀer much after ﬁltering.
2.4. SCALESPACE PROPERTIES 63
Grey level shift invariance
Since the diﬀusion tensor is only a function of J
ρ
(∇u
σ
), but not of u, we may shift
the grey level range by an arbitrary constant C, and the ﬁltered images will also
be shifted by the same constant. Moreover, a constant function is not aﬀected by
diﬀusion ﬁltering. Therefore, we have
T
t
(0) = 0, (2.19)
T
t
(f +C) = T
t
(f) + C ∀ t ≥ 0. (2.20)
Reverse contrast invariance
From D(J
ρ
(−∇u
σ
)) = D(J
ρ
(∇u
σ
)), it follows that
T
t
(−f) = −T
t
(f) ∀ t ≥ 0. (2.21)
This property is not fulﬁlled by classical morphological scalespace equations like
dilation and erosion. When reversing the contrast, the role of dilation and erosion
has to be exchanged as well.
Average grey level invariance
Average grey level invariance is a further property in which diﬀusion scalespaces
diﬀer from morphological scalespaces. In general, the evolution PDEs of the latter
ones are not of divergence form and do not preserve the mean grey value. A con
stant average grey level is essential for scalespace based segmentation algorithms
such as the hyperstack [307, 408]. It is also a desirable quality for applications
in medical imaging where grey values measure physical qualities of the depicted
object, for instance proton densities in MR images.
Proposition 1 (Conservation of average grey value).
The average grey level
µ :=
1
[Ω[
Ω
f(x) dx (2.22)
is not aﬀected by nonlinear diﬀusion ﬁltering:
1
[Ω[
Ω
T
t
f dx = µ ∀ t > 0. (2.23)
Proof:
Deﬁne I(t) :=
Ω
u(x, t) dx for all t ≥ 0. Then the Cauchy–Schwarz inequality
64 CHAPTER 2. CONTINUOUS DIFFUSION FILTERING
implies
[I(t)−I(0)[ =
Ω
u(x, t)−f(x)
dx
≤ [Ω[
1/2
u(t)−f
L
2
(Ω)
.
Since u ∈ C([0, T]; L
2
(Ω)), the preceding inequality gives the continuity of I(t) in
0.
For t > 0, Theorem 1, the divergence theorem and the boundary conditions yield
dI
dt
=
Ω
∂
t
u dx =
Γ
'D(J
ρ
(∇u
σ
))∇u, n` dS = 0.
Hence, I(t) must be constant for all t ≥ 0. 2
Average grey level invariance may be described by commuting operators, when
introducing an averaging operator M : L
1
(Ω) →L
1
(Ω) which maps f to a constant
image with the same mean grey level:
(Mf)(y) :=
1
[Ω[
Ω
f(x) dx ∀ y ∈ Ω. (2.24)
Then Proposition 1 and grey level shift invariance imply that the order of M and
T
t
can be exchanged:
M(T
t
f) = T
t
(Mf) ∀ t ≥ 0. (2.25)
When studying diﬀusion ﬁltering as a pure initial value problem in the domain
IR
2
, it also makes sense to investigate Euclidean transformations of an image. This
leads us to translation and isometry invariance.
Translation invariance
Deﬁne a translation τ
h
by (τ
h
f)(x) := f(x +h). Then diﬀusion ﬁltering fulﬁls
T
t
(τ
h
f) = τ
h
(T
t
f) ∀ t ≥ 0. (2.26)
This is a consequence of the fact that the diﬀusion tensor depends on J
ρ
(∇u
σ
)
solely, but not explicitly on x.
Isometry invariance
Let R ∈ IR
2×2
be an orthogonal transformation. If we apply R to f by deﬁning
Rf(x) := f(Rx), then the eigenvalues of the diﬀusion tensor are unaltered and
any eigenvector v is transformed into Rv. Thus, it makes no diﬀerence whether
the orthogonal transformation is applied before or after diﬀusion ﬁltering:
T
t
(Rf) = R(T
t
f) ∀ t ≥ 0. (2.27)
2.4. SCALESPACE PROPERTIES 65
2.4.2 Informationreducing properties
Nonenhancement of local extrema
Koenderink [240] required that a scalespace evolution should not create new level
curves when increasing the scale parameter. If this is satisﬁed, isointensity linking
through the scales is possible and a structure at a coarse scale can (in principle) be
traced back to the original image (causality). For this reason, he imposed that at
spatial extrema with nonvanishing determinant of the Hessian isophotes in scale
space are upwards convex. He showed that this constraint can be written as
sgn(∂
t
u) = sgn(∆u). (2.28)
A suﬃcient condition for the causality equation (2.28) to hold is requiring that
local extrema with positive or negative deﬁnite Hessians are not enhanced: an
extremum in ξ at time θ satisﬁes ∂
t
u > 0 if ξ is a minimum, and ∂
t
u < 0 if ξ
is a maximum. This implication is easily seen: In the ﬁrst case, for instance, the
eigenvalues η
1
, η
2
of the Hessian Hess(u) are positive. Thus,
∆u = trace(Hess(u)) = η
1
+η
2
> 0, (2.29)
giving immediately the causality requirement (2.28).
Nonenhancement of local extrema has ﬁrst been used by Babaud et al. [30] in
the context of linear diﬀusion ﬁltering. However, it is also satisﬁed by nonlinear
diﬀusion scalespaces, as we shall see now.
3
Theorem 2 (Nonenhancement of local extrema).
Let u be the unique solution of (P
c
) and consider some θ > 0. Suppose that ξ ∈ Ω
is a local extremum of u(., θ) with nonvanishing Hessian. Then,
∂
t
u(ξ, θ) < 0, if ξ is a local maximum, (2.30)
∂
t
u(ξ, θ) > 0, if ξ is a local minimum. (2.31)
Proof:
Let D(J
ρ
(∇u
σ
)) =: (d
ij
(J
ρ
(∇u
σ
))). Then we have
∂
t
u =
2
¸
i=1
2
¸
j=1
∂
x
i
d
ij
(J
ρ
(∇u
σ
))
∂
x
j
u +
2
¸
i=1
2
¸
j=1
d
ij
(J
ρ
(∇u
σ
)) ∂
x
i
x
j
u. (2.32)
Since ∇u(ξ, θ) = 0 and ∂
x
i
d
ij
(J
ρ
(∇u
σ
(ξ, θ))) is bounded, the ﬁrst term of the
righthand side of (2.32) vanishes in (ξ, θ).
3
As in the linear diﬀusion case, nonenhancement of local extrema generally does not imply
that their number is nonincreasing, cf. 1.2.5 and [342].
66 CHAPTER 2. CONTINUOUS DIFFUSION FILTERING
We know that the diﬀusion tensor D := D(J
ρ
(∇u
σ
(ξ, θ))) is positive deﬁnite.
Hence, there exists an orthogonal matrix S ∈ IR
2×2
such that
S
T
DS = diag(λ
1
, λ
2
) =: Λ
with λ
1
, λ
2
being the positive eigenvalues of D.
Now, let us assume that (ξ, θ) is a local maximum where H := Hess(u(ξ, θ))
and, thus, B := (b
ij
) := S
T
HS are negative deﬁnite. Then we have
b
ii
< 0 (i = 1, 2),
and by the invariance of the trace with respect to orthogonal transformations it
follows that
∂
t
u(ξ, θ) = trace (DH)
= trace (S
T
DS S
T
HS)
= trace (ΛB)
=
2
¸
i=1
λ
i
b
ii
< 0.
If ξ is a local minimum of u(x, θ), one proceeds in the same way utilizing the
positive deﬁniteness of the Hessian. 2
Nonenhancement of local extrema distinguishes anisotropic diﬀusion from clas
sical contrast enhancing methods such as highfrequency emphasis [163, pp. 182–
183], which do violate this principle. Although possibly behaving like backward
diﬀusion across edges, nonlinear diﬀusion is always in the forward region at ex
trema. This ensures its stability.
It should be noted that nonenhancement of local extrema is just one possibil
ity to end up with Koenderink’s causality requirement. Another way to establish
causality is via the extremum principle (2.15) following Hummel’s reasoning; see
[189] for more details.
Lyapunov functionals and behaviour for t →∞
Since scalespaces are intended to subsequently simplify an image, it is desirable
that, for t → ∞, we obtain the simplest possible image representation, namely a
constant image with the same average grey value as the original one. The following
theorem states that anisotropic diﬀusion ﬁltering always leads to a constant steady
state. This is due to the class of Lyapunov functionals associated with the diﬀusion
process.
2.4. SCALESPACE PROPERTIES 67
Theorem 3 (Lyapunov functionals and behaviour for t →∞).
Suppose that u is the solution of (P
c
) and let a, b, µ and M be deﬁned as in (P
c
),
(2.22) and (2.24), respectively. Then the following properties are valid:
(a) (Lyapunov functionals)
For all r ∈ C
2
[a, b] with r
′′
≥ 0 on [a, b], the function
V (t) := Φ(u(t)) :=
Ω
r(u(x, t)) dx (2.33)
is a Lyapunov functional:
(i) Φ(u(t)) ≥ Φ(Mf) for all t ≥ 0.
(ii) V ∈ C[0, ∞) ∩ C
1
(0, ∞) and V
′
(t) ≤ 0 for all t > 0.
Moreover, if r
′′
>0 on [a, b], then V (t) = Φ(u(t)) is a strict Lyapunov func
tional:
(iii) Φ(u(t)) = Φ(Mf) ⇐⇒
u(t) = Mf on
¯
Ω (if t > 0)
u(t) = Mf a.e. on Ω (if t = 0)
(iv) If t > 0, then V
′
(t) = 0 if and only if u(t) = Mf on
¯
Ω.
(v) V (0) = V (T) for T > 0 ⇐⇒
f = Mf a.e. on Ω and
u(t) = Mf on
¯
Ω (0, T]
(b) (Convergence)
(i) lim
t→∞
u(t) −Mf
L
p
(Ω)
= 0 for p ∈ [1, ∞).
(ii) In the 1D case, the convergence lim
t→∞
u(x, t) = µ is uniform on
¯
Ω.
Proof:
(a) (i) Since r ∈ C
2
[a, b] with r
′′
≥ 0 on [a, b], we know that r is convex on
[a, b]. Using the average grey level invariance and Jensen’s inequality we
obtain, for all t ≥ 0,
Φ(Mf) =
Ω
r
¸
1
[Ω[
Ω
u(x, t) dx
¸
dy
≤
Ω
¸
1
[Ω[
Ω
r(u(x, t)) dx
¸
dy
=
Ω
r(u(x, t)) dx
= Φ(u(t)). (2.34)
68 CHAPTER 2. CONTINUOUS DIFFUSION FILTERING
(ii) Let us start by proving the continuity of V (t) in 0. Thanks to the
maximum–minimum principle, we may choose a constant
L := max
s∈[a,b]
[r
′
(s)[
such that for all t > 0, the Lipschitz condition
[r(u(x, t))−r(f(x))[ ≤ L [u(x, t)−f(x)[
is veriﬁed a.e. on Ω. From this and the Cauchy–Schwarz inequality, we
get
[V (t)−V (0)[ ≤ [Ω[
1/2
r(u(t))−r(f)
L
2
(Ω)
≤ [Ω[
1/2
L u(t)−f
L
2
(Ω)
,
and by virtue of u ∈ C([0, T]; L
2
(Ω)), the limit t → 0
+
gives the an
nounced continuity in 0.
By Theorem 1 and the boundedness of r
′
on [a, b], we know that V is
diﬀerentiable for t > 0 and V
′
(t) =
Ω
r
′
(u) u
t
dx. Thus, the divergence
theorem yields
V
′
(t) =
Ω
r
′
(u) div (D(J
ρ
(∇u
σ
))∇u) dx
=
Γ
r
′
(u) 'D(J
ρ
(∇u
σ
))∇u, n`
. .. .
=0
dS
−
Ω
r
′′
(u)
. .. .
≥0
'∇u, D(J
ρ
(∇u
σ
))∇u`
. .. .
≥0
dx
≤ 0.
(iii) Let Φ(u(t)) = Φ(Mf).
If t > 0, then u(t) is continuous in
¯
Ω. Let us now show that equality in
the estimate (2.34) implies that u(t) = const. on
¯
Ω. To this end, assume
that u is not constant on
¯
Ω. Then, by the continuity of u, there exists
a partition Ω = Ω
1
∪ Ω
2
with [Ω
1
[, [Ω
2
[ ∈ (0, [Ω[) and
α :=
1
[Ω
1
[
Ω
1
u dx =
1
[Ω
2
[
Ω
2
u dx =: β.
From r
′′
>0 on [a, b] it follows that r is strictly convex on [a, b] and
r
¸
1
[Ω[
Ω
u dx
¸
= r
[Ω
1
[
[Ω[
α +
[Ω
2
[
[Ω[
β
2.4. SCALESPACE PROPERTIES 69
<
[Ω
1
[
[Ω[
r(α) +
[Ω
1
[
[Ω[
r(β)
≤
1
[Ω[
Ω
1
r(u) dx +
1
[Ω[
Ω
2
r(u) dx
=
1
[Ω[
Ω
r(u) dx.
If we utilize this result in the estimate (2.34) we observe that, for t >0,
Φ(u(t)) = Φ(Mf) implies that u(t) = const. on
¯
Ω. Thanks to the
average grey value invariance we ﬁnally obtain u(t) = Mf on
¯
Ω.
So let us turn to the case t = 0. From (i) and (ii), we conclude that
Φ(u(θ)) = Φ(Mf) for all θ > 0. Thus, we have u(θ) = Mf for all θ > 0.
For θ > 0, the Cauchy–Schwarz inequality gives
Ω
[u(x, θ) −µ[ dx ≤ [Ω[
1/2
u(θ) −Mf
L
2
(Ω)
= 0.
Since u ∈ C([0, T]; L
2
(Ω)), the limit θ → 0
+
ﬁnally yields u(0) = Mf
a.e. on Ω.
Conversely, it is obvious that u(t) = Mf (a.e.) on Ω implies Φ(u(t)) =
Φ(Mf).
(iv) Let t > 0 and V
′
(t) = 0. Then from
0 = V
′
(t) = −
Ω
r
′′
(u(x, t))
. .. .
>0
'∇u(x, t), D(J
ρ
(∇u
σ
(x, t)))∇u(x, t)` dx
and the smoothness of u we obtain
'∇u, D(J
ρ
(∇u
σ
))∇u` = 0 on
¯
Ω.
By the uniform boundedness of D, there exists some constant ν > 0,
such that
ν [∇u[
2
≤ '∇u, D(J
ρ
(∇u
σ
))∇u` on
¯
Ω (0, ∞).
Thus, we have ∇u(x, t) = 0 a.e. on Ω. Due to the continuity of ∇u,
this yields u(x, t) = const. for all x ∈ Ω, and the average grey level
invariance ﬁnally gives u(x, t) = µ on Ω.
Conversely, let u(x, t) = µ on Ω. Then,
V
′
(t) = −
Ω
r
′′
(u) '∇u, D(J
ρ
(∇u
σ
))∇u` dx = 0.
70 CHAPTER 2. CONTINUOUS DIFFUSION FILTERING
(v) Suppose that V (T) = V (0). Since V is decreasing, we have
V (t) = const. on [0, T].
Let ǫ > 0. Then for any t ∈ [ǫ, T], we have V
′
(t) = 0, and part (iv)
implies that u(t) = Mf on Ω. Now, the Cauchy–Schwarz inequality
gives
Ω
[f −Mf[ dx ≤ [Ω[
1/2
f −u(t)
L
2
(Ω)
.
As u ∈ C([0, T]; L
2
(Ω)), the limit t →0
+
yields f = Mf a.e. on Ω.
Conversely, if u(t) = Mf (a.e.) on Ω holds for all t ∈ [0, T], it is evident
that V (0) = V (T).
(b) (i) By the grey level shift invariance we know that v := u−Mf satisﬁes the
diﬀusion equation as well. We multiply this equation by v, integrate,
and use the divergence theorem to obtain
Ω
vv
t
dx = −
Ω
'∇v, D(J
ρ
(∇v
σ
))∇v` dx.
Since ∇v
σ
is bounded, we ﬁnd some ν > 0 such that
1
2
d
dt
(v
2
L
2
(Ω)
) ≤ −ν ∇v
2
L
2
(Ω)
.
For t > 0, there exists some x
0
with v(x
0
) = 0. Therefore, Poincar´e’s
inequality (cf. [9, p. 122]) may be applied giving
v
2
L
2
(Ω)
≤ C
0
∇v
2
L
2
(Ω)
with some constant C
0
= C
0
(Ω) > 0. This yields
d
dt
v
2
L
2
(Ω)
≤ −2ν C
0
v
2
L
2
(Ω)
and hence the exponential decay of v
L
2
(Ω)
to 0.
By the maximum principle, we know that v(t)
L
∞
(Ω)
is bounded by
f −Mf
L
∞
(Ω)
. Thus, for q ∈ IN, q ≥ 2, we get
v(t)
q
L
q
(Ω)
≤ f −Mf
q−2
L
∞
(Ω)
v(t)
2
L
2
(Ω)
→ 0,
and H¨older’s inequality gives, for 1 ≤ p < q < ∞,
v(t)
L
p
(Ω)
≤ [Ω[
(1/p)−(1/q)
v(t)
L
q
(Ω)
→ 0.
This proves the assertion.
2.4. SCALESPACE PROPERTIES 71
(ii) To prove uniform convergence in the onedimensional setting, we can
generalize and adapt methods from [202] to our case.
Let Ω = (0, a). From part (a) we know that V (t) :=
a
0
u
2
(x, t) dx is
nonincreasing and bounded from below. Thus, the sequence (V (i))
i∈IN
converges.
Since V ∈ C[0, ∞) ∩ C
1
(0, ∞) the mean value theorem implies
∃ t
i
∈ (i, i+1) : V
′
(t
i
) = V (i + 1) −V (i).
Thus, (t
i
)
i∈IN
→ ∞ and from the convergence of (V (i))
i∈IN
it follows
that
V
′
(t
i
) →0. (2.35)
Thanks to the uniform positive deﬁniteness of D there exists some ν > 0
such that, for t > 0,
V
′
(t) = −2
a
0
u
2
x
D(J
ρ
(∂
x
u
σ
)) dx
≤ −2ν
a
0
u
2
x
dx
≤ 0. (2.36)
Equations (2.35) and (2.36) yield
u
x
(t
i
)
L
2
(Ω)
→0.
Hence, u(t
i
) is a bounded sequence in H
1
(0, a). By virtue of the Rellich–
Kondrachov theorem [7, p. 144] we know that the embedding from
H
1
(0, a) into C
0,α
[0, a], the space of H¨oldercontinuous functions on [0, a]
[7, pp. 9–12], is compact for α∈(0,
1
2
). Therefore, there exists a subse
quence (t
i
j
) →∞ and some ¯ u with
u(t
i
j
) → ¯ u in C
0,α
[0, a].
This also gives u(t
i
j
) → ¯ u in L
2
(0, a). Since we already know from (b)(i)
that u(t
i
j
) →Mf in L
2
(0, a), it follows that ¯ u = Mf. Hence,
lim
j→∞
u(t
i
j
) −Mf
L
∞
(Ω)
= 0. (2.37)
Part (a) tells us that u(t)−Mf
p
L
p
(Ω)
is a Lyapunov function for p ≥ 2.
Thus,
u(t)−Mf
L
∞
(Ω)
= lim
p→∞
u(t)−Mf
L
p
(Ω)
72 CHAPTER 2. CONTINUOUS DIFFUSION FILTERING
is also nonincreasing. Therefore, lim
t→∞
u(t)−Mf
L
∞
(Ω)
exists and from
(2.37) we conclude that
lim
t→∞
u(t)−Mf
L
∞
(Ω)
= 0.
The smoothness of u establishes ﬁnally that the convergence
lim
t→∞
u(x, t) = µ
is uniform on
¯
Ω. 2
Since the class (P
c
) does not forbid contrast enhancement it admits processes
where forward diﬀusion has to compete with backward diﬀusion. Theorem 3 is
of importance as it states that the regularization by convolving with K
σ
tames
the backward diﬀusion in such a way that forward diﬀusion wins in the long run.
Moreover, the competition evolves in a certain direction all the time: although
backward diﬀusion may be locally superior, the global result – denoted by the
Lyapunov functional – becomes permanently better for forward diﬀusion.
Let us have a closer look at what might be the meaning of this global result in
the context of image processing. Considering the Lyapunov functions associated
with r(s) := [s[
p
, r(s) := (s−µ)
2n
and r(s) := s ln s, respectively, the preceding
theorem gives the following corollary.
Corollary 1 (Special Lyapunov functionals).
Let u be the solution of (P
c
) and a and µ be deﬁned as in (P
c
) and (2.22). Then
the following functions are decreasing for t ∈ [0, ∞):
(a) u(t)
L
p
(Ω)
for all p ≥ 2.
(b) M
2n
[u(t)] :=
1
[Ω[
Ω
(u(x, t) −µ)
2n
dx for all n ∈ IN.
(c) H[u(t)] :=
Ω
u(x, t) ln(u(x, t)) dx, if a > 0.
Corollary 1 oﬀers multiple possibilities of how to interpret nonlinear anisotropic
diﬀusion ﬁltering as a smoothing transformation.
As a special case of (a) it follows that the energy u(t)
2
L
2
(Ω)
is reduced by
diﬀusion.
Part (b) gives a probabilistic interpretation of anisotropic diﬀusion ﬁltering.
Consider the intensity in an image f as a random variable Z
f
with distribution
2.4. SCALESPACE PROPERTIES 73
F
f
(z), i.e. F
f
(z) is the probability that an arbitrary grey value Z
f
of f does not
exceed z. By the average grey level invariance, µ is equal to the expected value
EZ
u(t)
:=
IR
z dF
u(t)
(z), (2.38)
and it follows that M
2n
[u(t)] is just the even central moment
IR
z−EZ
u(t)
2n
dF
u(t)
(z). (2.39)
The second central moment (the variance) characterizes the spread of the intensity
about its mean. It is a common tool for constructing measures for the relative
smoothness of the intensity distribution. The fourth moment is frequently used to
describe the relative ﬂatness of the grey value distribution. Higher moments are
more diﬃcult to interpret, although they do provide important information for
tasks like texture discrimination [163, pp. 414–415]. All decreasing even moments
demonstrate that the image becomes smoother during diﬀusion ﬁltering. Hence,
local eﬀects such as edge enhancement, which object to increase central moments,
are overcompensated by smoothing in other areas.
If we choose another probabilistic model of images, then part (c) characterizes
the informationtheoretical side of our scalespace. Provided the initial image f is
strictly positive on Ω, we may regard it also as a twodimensional density.
4
Then,
S[u(t)] := −
Ω
u(x, t) ln(u(x, t)) dx (2.40)
is called the entropy of u(t), a measure of uncertainty and missing information [63].
Since anisotropic diﬀusion ﬁlters increase the entropy the corresponding scalespace
embeds the genuine image f into a family of subsequently likelier versions of it
which contain less information. Moreover, for t →∞, the process reaches the state
with the lowest possible information, namely a constant image. This information
reducing property indicates that anisotropic diﬀusion might be generally useful in
the context of image compression. In particular, it helps to explain the success of
nonlinear diﬀusion ﬁltering as a preprocessing step for subsampling as observed in
[144]. The interpretation of the entropy in terms of Lyapunov functionals carries
also over to generalized entropies; see [390] for more details.
From all the previous considerations, we recognize that, in spite of possible
contrastenhancing properties, anisotropic diﬀusion does really simplify the ori
ginal image in a steady way.
Let us ﬁnally point out another interpretation of the Lyapunov functionals. In
a classic scalespace representation, the time t plays the role of the scale para
meter. By increasing t, one transforms the image from a local to a more global
4
Without loss of generality we omit the normalization.
74 CHAPTER 2. CONTINUOUS DIFFUSION FILTERING
representation. We have seen in Chapter 1 that, for linear diﬀusion scalespaces
and morphological scalespaces, it is possible to associate with the evolution time
a corresponding spatial scale.
In the nonlinear diﬀusion case, however, the situation is more complicated.
Since the smoothing is nonuniform, one can only deﬁne an average measure for
the globality of the representation. This can be achieved by taking some Lyapunov
function Φ(u(t)) and investigating the expression
Ψ(u(t)) :=
Φ(f) −Φ(u(t))
Φ(f) −Φ(Mf)
. (2.41)
We observe that Ψ(t) increases from 0 to 1. It gives the average globality of u(t) and
its value can be used to measure the distance of u(t) from the initial state f and the
ﬁnal state Mf. Prescribing a certain value for Ψ provides us with an aposteriori
criterion for the stopping time of the nonlinear diﬀusion process. Experiments in
this direction can be found in [431, 308].
Chapter 3
Semidiscrete diﬀusion ﬁltering
The goal of this chapter is to study a semidiscrete framework for diﬀusion scale
spaces where the image is sampled on a ﬁnite grid and the scale parameter is
continuous. This leads to a system of nonlinear ordinary diﬀerential equations
(ODEs). We shall investigate conditions under which one can establish similar
properties as in the continuous setting concerning wellposedness, extremum prin
ciples, average grey level invariance, Lyapunov functions, and convergence to a
constant steadystate. Afterwards we shall discuss whether it is possible to obtain
such ﬁlters from spatial discretizations of the continuous models that have been
investigated in Chapter 2. We will see that there exists a ﬁnite stencil on which
a diﬀerence approximation of the spatial derivatives are in accordance with the
semidiscrete scalespace framework.
3.1 The general model
A discrete image can be regarded as a vector f ∈IR
N
, N ≥2, whose components
f
j
, j = 1,...,N represent the grey values at each pixel. We denote the index set
¦1, ..., N¦ by J. In order to specify the requirements for our semidiscrete ﬁlter
class we ﬁrst recall a useful deﬁnition of irreducible matrices [407, pp. 18–20].
Deﬁnition 1 (Irreducibility). A matrix A = (a
ij
) ∈ IR
N×N
is called irreducible
if for any i, j ∈ J there exist k
0
,...,k
r
∈ J with k
0
= i and k
r
= j such that
a
kpk
p+1
= 0 for p = 0,...,r−1.
The semidiscrete problem class (P
s
) we are concerned with is deﬁned in the
following way:
75
76 CHAPTER 3. SEMIDISCRETE DIFFUSION FILTERING
Let f ∈ IR
N
. Find a function u ∈ C
1
([0, ∞), IR
N
) which satisﬁes an
initial value problem of type
du
dt
= A(u) u,
u(0) = f,
where A = (a
ij
) has the following properties:
(S1) Lipschitzcontinuity of A ∈ C(IR
N
, IR
N×N
) for every bounded
subset of IR
N
,
(S2) symmetry: a
ij
(u) = a
ji
(u) ∀ i, j ∈ J, ∀ u ∈ IR
N
,
(S3) vanishing row sums:
¸
j∈J
a
ij
(u) = 0 ∀ i ∈ J, ∀ u ∈ IR
N
,
(S4) nonnegative oﬀdiagonals: a
ij
(u) ≥ 0 ∀ i = j, ∀ u ∈ IR
N
,
(S5) irreducibility for all u ∈ IR
N
.
(P
s
)
Not all of these requirements are necessary for every theoretical result below.
(S1) is needed for wellposedness, the proof of a maximum–minimum principle
involves (S3) and (S4), while average grey value invariance uses (S2) and (S3).
The existence of Lyapunov functions can be established by means of (S2)–(S4),
and strict Lyapunov functions and the convergence to a constant steadystate
require (S5) in addition to (S2)–(S4).
This indicates that these properties reveal some interesting parallels to the
continuous setting from Chapter 2: In both cases we need smoothness assumptions
to ensure wellposedness; (S2) and (S3) correspond to the speciﬁc structure of the
divergence expression with a symmetric diﬀusion tensor D, while (S4) and (S5)
play a similar role as the nonnegativity of the eigenvalues of D and its uniform
positive deﬁniteness, respectively.
3.2 Theoretical results
Before we can establish scalespace results, it is of importance to ensure the ex
istence of a unique solution. This is done in the theorem below which also states
the continuous dependence of the solution and a maximum–minimum principle.
3.2. THEORETICAL RESULTS 77
Theorem 4 (Wellposedness, extremum principle).
For every T > 0 the problem (P
s
) has a unique solution u(t) ∈ C
1
([0, T], IR
N
).
This solution depends continuously on the initial value and the righthand side of
the ODE system, and it satisﬁes the extremum principle
a ≤ u
i
(t) ≤ b ∀ i ∈ J, ∀ t ∈ [0, T], (3.1)
where
a := min
j∈J
f
j
, (3.2)
b := max
j∈J
f
j
. (3.3)
Proof:
(a) Local existence and uniqueness
Local existence and uniqueness are proved by showing that our problem
satisﬁes the requirements of the Picard–Lindel¨of theorem [432, p. 59].
Let t
0
:= 0 and β > 0. Evidently, φ(t, u) := ψ(u) := A(u) u is continuous on
B
0
:= [0, T]
u ∈ IR
N
u
∞
≤ f
∞
+β
¸
,
since it is a composition of continuous functions. Moreover, by the compact
ness of B
0
there exists some c > 0 with
φ(t, u)
∞
≤ c ∀ (t, u) ∈ B
0
.
In order to prove existence and uniqueness of a solution of (P
s
) in
R
0
:=
(t, u)
t ∈ [t
0
, t
0
+min(
β
c
, T)], u−f
∞
≤ β
¸
⊂ B
0
we have to show that φ(t, u) satisﬁes a global Lipschitz condition on R
0
with
respect to u. However, this follows directly from the fact that A is Lipschitz
continuous on ¦u ∈ IR
N
[ u−f
∞
≤ β¦.
(b) Maximum–minimum principle
We prove only the maximum principle, since the proof for the minimum
principle is analogous.
Assume that the problem (P
s
) has a unique solution on [0, θ]. First we show
that the derivative of the largest component of u(t) is nonpositive for every
78 CHAPTER 3. SEMIDISCRETE DIFFUSION FILTERING
t ∈ [0, θ]. Let u
k
(ϑ) := max
j∈J
u
j
(ϑ) for some arbitrary ϑ ∈ [0, θ]. If we keep
this k ﬁxed we obtain, for t = ϑ,
du
k
dt
=
¸
j∈J
a
kj
(u) u
j
= a
kk
(u) u
k
+
¸
j∈J\{k}
a
kj
(u)
. .. .
≥0
u
j
....
≤u
k
≤ u
k
¸
j∈J
a
kj
(u)
(S4)
= 0. (3.4)
Let us now prove that this implies a maximum principle (cf. [201]).
Let ε > 0 and set
u
ε
(t) := u(t) −
¸
¸
¸
εt
.
.
.
εt
¸
.
Moreover, let P := ¦p∈J [ u
εp
(0) = max
j∈J
u
εj
(0)¦. Then, by (3.4),
du
εp
dt
(0) =
du
p
dt
(0)
. .. .
≤0
−ε < 0 ∀ p ∈ P. (3.5)
By means of
max
i∈J\P
u
εi
(0) < max
j∈J
u
εj
(0),
and the continuity of u there exists some t
1
∈ (0, θ) such that
max
i∈J\P
u
εi
(t) < max
j∈J
u
εj
(0) ∀ t ∈ [0, t
1
). (3.6)
Next, let us consider some p ∈ P. Due to (3.5) and the smoothness of u we
may ﬁnd a ϑ
p
∈ (0, θ) with
du
εp
dt
(t) < 0 ∀ t ∈ [0, ϑ
p
).
Thus, we have
u
εp
(t) < u
εp
(0) ∀ t ∈ (0, ϑ
p
)
and, for t
2
:= min
p∈P
ϑ
p
, it follows that
max
p∈P
u
εp
(t) < max
j∈J
u
εj
(0) ∀ t ∈ (0, t
2
). (3.7)
3.2. THEORETICAL RESULTS 79
Hence, for t
0
:= min(t
1
, t
2
), the estimates (3.6) and (3.7) give
max
j∈J
u
εj
(t) < max
j∈J
u
εj
(0) ∀ t ∈ (0, t
0
). (3.8)
Now we prove that this estimate can be extended to the case t ∈ (0, θ). To
this end, assume the opposite is true. Then, by virtue of the intermediate
value theorem, there exists some t
3
which is the smallest time in (0, θ) such
that
max
j∈J
u
εj
(t
3
) = max
j∈J
u
εj
(0).
Let u
εk
:= max
j∈J
u
εj
(t
3
). Then the minimality of t
3
yields
u
εk
(t) < u
εk
(t
3
) ∀ t ∈ (0, t
3
), (3.9)
and inequality (3.4) gives
du
εk
dt
(t
3
) =
du
k
dt
(t
3
)
. .. .
≤0
−ε < 0.
Due to the continuity of
du
dt
there exists some t
4
∈ (0, t
3
) with
du
εk
dt
(t) < 0 ∀ t ∈ (t
4
, t
3
]. (3.10)
The mean value theorem, however, implies that we ﬁnd a t
5
∈ (t
4
, t
3
) with
du
εk
dt
(t
5
) =
u
εk
(t
3
) −u
εk
(t
4
)
t
3
−t
4
(3.9)
> 0,
which contradicts (3.10). Hence, (3.8) must be valid on the entire interval
(0, θ).
Together with u = lim
ε→0
u
ε
and the continuity of u this yields the announced
maximum principle
max
j∈J
u
j
(t) ≤ max
j∈J
u
j
(0) ∀ t ∈ [0, θ].
(c) Global existence and uniqueness
Global existence and uniqueness follow from local existence and uniqueness
when being combined with the extremum principle.
Using the notations and results from (a), we know that the problem (P
s
) has
a unique solution u(t) for t ∈ [t
0
, t
0
+ min(
β
c
, T)].
80 CHAPTER 3. SEMIDISCRETE DIFFUSION FILTERING
Now let t
1
:= t
0
+ min(
β
c
, T), g := u(t
1
), and consider the problem
du
dt
= A(u) u,
u(t
1
) = g.
Clearly, φ(t, u) = A(u)u is continuous on
B
1
:= [0, T]
u ∈ IR
N
u
∞
≤ g
∞
+β
¸
,
and by the extremum principle we know that B
1
⊂ B
0
. Hence,
φ(t, u)
∞
≤ c ∀ (t, u) ∈ B
1
.
with the same c as in (a). Using the same considerations as in (a) one shows
that φ is Lipschitzcontinuous on
R
1
:=
(t, u)
t ∈ [t
1
, t
1
+min(
β
c
, T)], u−g
∞
≤ β
¸
.
Hence, the considered problem has a unique solution on [t
1
, t
1
+ min(
β
c
, T)].
Therefore, (P
s
) reveals a unique solution on [0, min(
2β
c
, T)], and, by iterating
this reasoning, the existence of a unique solution can be extended to the
entire interval [0, T]. As a consequence, the extremum principle is valid on
[0, T] as well.
(d) Continuous dependence
Let u(t) be the solution of
du
dt
= φ(t, u),
u(0) = f
for t ∈[0, T] and φ(u, t) =ψ(u) =A(u)u. In order to show that u(t) depends
continuously on the initial data and the righthand side of the ODE system,
it is suﬃcient to prove that φ(t, u) is continuous, and that there exists some
α > 0 such that φ(t, u) satisﬁes a global Lipschitz condition on
S
α
:=
(t, v)
t ∈ [0, T], v−u
∞
≤ α
¸
.
with respect to its second argument. In this case the results in [412, p. 93]
ensure that for every ε > 0 there exists a δ > 0 such that the solution ˜ u of
the perturbed problem
d˜ u
dt
=
˜
φ(t, ˜ u),
˜ u(0) =
˜
f
3.3. SCALESPACE PROPERTIES 81
with continuous
˜
φ and

˜
f −f
∞
< δ,

˜
φ(t, v) −φ(t, v)
∞
< δ for v −u
∞
< α
exists in [0, T] and satisﬁes the inequality
˜ u(t) −u(t)
∞
< ε.
Similar to the the local existence and uniqueness proof, the global Lipschitz
condition on S
α
follows direcly from the fact that A is Lipschitzcontinuous
on ¦v ∈ IR
N
[ v−u
∞
≤ α¦. 2
3.3 Scalespace properties
It is evident that properties such as grey level shift invariance or reverse contrast
invariance are automatically satisﬁed by every consistent semidiscrete approxima
tion of the continuous ﬁlter class (P
c
). On the other hand, translation invariance
only makes sense for translations in grid direction with multiples of the grid size,
and isometry invariance is satisﬁed for consistent schemes up to an discretization
error. So let us focus on average grey level invariance now.
Proposition 2 (Conservation of average grey value).
The average grey level
µ :=
1
N
¸
j∈J
f
j
(3.11)
is not aﬀected by the semidiscrete diﬀusion ﬁlter:
1
N
¸
j∈J
u
j
(t) = µ ∀ t ≥ 0. (3.12)
Proof:
By virtue of (S2) and (S3) we have
¸
j∈J
a
jk
(u) = 0 for all k ∈ J. Thus, for t ≥ 0,
¸
j∈J
du
j
dt
=
¸
j∈J
¸
k∈J
a
jk
(u) u
k
=
¸
k∈J
¸
j∈J
a
jk
(u)
u
k
= 0,
which shows that
¸
j∈J
u
j
(t) is constant on [0, ∞) and concludes the proof. 2
This property is in complete accordance with the result for the continuous ﬁlter
class.
82 CHAPTER 3. SEMIDISCRETE DIFFUSION FILTERING
Similar to the continuous setting, it is possible to ﬁnd a large class of Lyapunov
functions which establish smoothing scalespace properties and ensure that the
image tends to a constant steadystate with the same average grey level as the
initial image.
Theorem 5 (Lyapunov functions and behaviour for t →∞).
Let u(t) be the solution of (P
s
), let a, b, and µ be deﬁned as in (3.2), (3.3), and
(3.11), respectively, and let c := (µ, µ, ..., µ)
⊤
∈ IR
N
.
Then the following properties are valid:
(a) (Lyapunov functions)
For all r ∈ C
1
[a, b] with increasing r
′
on [a, b], the function
V (t) := Φ(u(t)) :=
¸
i∈J
r(u
i
(t))
is a Lyapunov function:
(i) Φ(u) ≥ Φ(c) for all t ≥ 0.
(ii) V ∈ C
1
[0, ∞) and V
′
(t) ≤ 0 for all t ≥ 0.
Moreover, if r
′
is strictly increasing on [a, b], then V (t) = Φ(u(t)) is a strict
Lyapunov function:
(iii) Φ(u) = Φ(c) ⇐⇒ u = c
(iv) V
′
(t) = 0 ⇐⇒ u = c
(b) (Convergence)
lim
t→∞
u(t) = c.
Proof:
(a) (i) Since r
′
is increasing on [a, b] we know that r is convex on [a, b]. Average
grey level invariance and this convexity yield, for all t ≥ 0,
Φ(c) =
N
¸
i=1
r
¸
N
¸
j=1
1
N
u
j
¸
≤
N
¸
i=1
¸
1
N
N
¸
j=1
r(u
j
)
¸
=
N
¸
j=1
r(u
j
)
= Φ(u). (3.13)
3.3. SCALESPACE PROPERTIES 83
(ii) Since u ∈ C
1
([0, ∞), IR
N
) and r ∈ C
1
[a, b], it follows that V ∈ C
1
[0, ∞).
Using the prerequisites (S2) and (S3) we get
V
′
(t) =
N
¸
i=1
du
i
dt
r
′
(u
i
)
(S3)
=
N
¸
i=1
N
¸
j=1
a
ij
(u) (u
j
−u
i
) r
′
(u
i
)
=
N
¸
i=1
¸
N
¸
j=i+1
+
i−1
¸
j=1
¸
a
ij
(u) (u
j
−u
i
) r
′
(u
i
)
=
N
¸
i=1
N−i
¸
k=1
a
i,i+k
(u) (u
i+k
−u
i
) r
′
(u
i
)
+
N
¸
i=1
N−i
¸
k=1
a
i+k,i
(u) (u
i
−u
i+k
) r
′
(u
i+k
)
(S2)
=
N
¸
i=1
N−i
¸
k=1
a
i,i+k
(u) (u
i+k
−u
i
)
r
′
(u
i
)−r
′
(u
i+k
)
. (3.14)
Since r
′
is increasing, we always have
(u
i+k
−u
i
)
r
′
(u
i
) −r
′
(u
i+k
)
≤ 0.
With this and (S4), equation (3.14) implies that V
′
(t) ≤ 0 for t ≥ 0.
(iii) Let us ﬁrst prove that equality in the estimate (3.13) implies that all
components of u are equal.
To this end, suppose that u
i
0
:= min
i
u
i
< max
j
u
j
=: u
j
0
and let
η :=
N
¸
j=1
j=j
0
1
N
u
j
1 −
1
N
.
Then, η <u
j
0
. Since r
′
is strictly increasing on [a, b], we know that r is
strictly convex. Hence, we get
r
N
¸
i=1
1
N
u
j
= r
1
N
u
j
0
+
1−
1
N
η
<
1
N
r(u
j
0
) +
1−
1
N
r(η)
≤
1
N
r(u
j
0
) +
N
¸
j=1
j=j
0
1
N
r(u
j
)
=
N
¸
j=1
1
N
r(u
j
).
84 CHAPTER 3. SEMIDISCRETE DIFFUSION FILTERING
This shows that equality in (3.13) implies that u
1
= ... = u
N
. By virtue
of the grey level shift invariance we conclude that u = c.
Conversely, it is trivial that Φ(u) = Φ(c) for u = c.
(iv) Let V
′
(t) = 0. From (3.14) we have
0 = V
′
(t) =
N
¸
i=1
N−i
¸
k=1
a
i,i+k
(u) (u
i+k
−u
i
)
r
′
(u
i
)−r
′
(u
i+k
)
. .. .
≤0
,
and by virtue of the symmetry of A(u) it follows that
a
ij
(u) (u
j
−u
i
)
r
′
(u
i
)−r
′
(u
j
)
= 0 ∀ i, j ∈ J. (3.15)
Now consider two arbitrary i
0
, j
0
∈ J. The irreducibility of A(u) implies
that there exist k
0
,...,k
r
∈ J with k
0
= i
0
, k
r
= j
0
, and
a
kpk
p+1
(u) = 0, p = 0, ..., r −1.
As r
′
is strictly increasing we have, for p = 0,...,r −1,
(u
kp
−u
k
p+1
)
r
′
(u
k
p+1
)−r
′
(u
kp
)
= 0 ⇐⇒ u
kp
= u
k
p+1
.
From this and (3.15) we get
u
i
0
= u
k
0
= u
k
1
= ... = u
kr
= u
j
0
.
Since i
0
and j
0
are arbitrary, we obtain u
i
= const. for all i ∈ J,
and the average grey level invariance gives u = c. This proves the ﬁrst
implication.
Conversely, let u
i
= const. for all i ∈ J. Then from the representation
(3.14) we immediately conclude that V
′
(t) = 0.
(b) The convergence proof is based on classical Lyapunov reasonings, see e.g.
[180] for an introduction to these techniques.
Consider the Lyapunov function V (t) := Φ(u(t)) := [u(t)−c[
2
, which results
from the choice r(s) := (s−µ)
2
. Since V (t) is decreasing and bounded from
below by 0, we know that lim
t→∞
V (t) =: η exists and η ≥ 0.
Now assume that η > 0.
Since [u(t)−c[ is bounded from above by α := [f −c[ we have
[u(t)−c[ ≤ α ∀ t ≥ 0. (3.16)
By virtue of Φ(x) = [x −c[
2
we know that, for β ∈ (0,
√
η),
Φ(x) < η ∀ x ∈ IR
N
, [x−c[ < β.
3.3. SCALESPACE PROPERTIES 85
Let w.l.o.g. β < α. Since Φ(u(t)) ≥ η we conclude that
[u(t)−c[ ≥ β ∀ t ≥ 0. (3.17)
So from (3.16) and (3.17) we have
u(t) ∈ ¦x ∈ IR
N
[ β ≤ [x−c[ ≤ α¦ =: S ∀ t ≥ 0.
By (a)(ii),(iv), the compactness of S, and β > 0 there exists some M > 0
such that
V
′
(t) ≤ −M ∀ t ≥ 0.
Therefore, it follows
V (t) = V (0) +
t
0
V
′
(θ) dθ ≤ V (0) −tM
which implies lim
t→∞
V (t) = −∞ and, thus, contradicts (a)(i).
Hence the assumption η > ρ is wrong and we must have η = ρ.
According to (a)(iii) this yields lim
t→∞
u(t) = c. 2
As in the continuous case, we can consider the Lyapunov functions associated
with r(s) := [s[
p
, r(s) := (s−µ)
2n
and r(s) := s ln s, respectively, and obtain the
following corollary.
Corollary 2 (Special Lyapunov functions).
Let u be the solution of (P
s
) and a and µ be deﬁned as in (3.2) and (3.11). Then
the following functions are decreasing for t ∈ [0, ∞):
(a) u(t)
p
for all p ≥ 2.
(b) M
2n
[u(t)] :=
1
N
N
¸
j=1
(u
j
(t) −µ)
2n
for all n ∈ IN.
(c) H[u(t)] :=
N
¸
j=1
u
j
(t) ln(u
j
(t)), if a > 0.
Since all pnorms (p ≥ 2) and all central moments are decreasing, while the
discrete entropy
S[u(t)] := −
N
¸
j=1
u
j
(t) ln(u
j
(t)) (3.18)
is increasing with respect to t, we observe that the semidiscrete setting reveals
smoothing scalespace properties which are closely related to the continuous case.
86 CHAPTER 3. SEMIDISCRETE DIFFUSION FILTERING
3.4 Relation to continuous models
In this section we investigate whether it is possible to use spatial discretizations of
the continuous ﬁlter class (P
c
) in order to construct semidiscrete diﬀusion models
satisfying (S1)–(S5). First we shall verify that this is easily done for isotropic
models. In the anisotropic case, however, the mixed derivative terms make it more
diﬃcult to ensure nonnegative oﬀdiagonal elements. A constructive existence proof
is presented showing that for a suﬃciently large stencil it is always possible to ﬁnd
such a nonnegative discretization. This concept is illustrated by investigating the
situation on a (33)stencil in detail.
3.4.1 Isotropic case
Let the rectangle Ω = (0, a
1
) (0, a
2
) be discretized by a grid of N = n
1
n
2
pixels
such that a pixel (i, j) with 1≤i ≤n
1
and 1≤j ≤n
2
represents the location (x
i
, y
j
)
where
x
i
:= (i −
1
2
) h
1
, (3.19)
y
j
:= (j −
1
2
) h
2
, (3.20)
and the grid sizes h
1
, h
2
are given by h
1
:= a
1
/n
1
and h
2
:= a
2
/n
2
, respectively.
These pixels can be numbered by means of an arbitrary bijection
p : ¦1, ..., n
1
¦ ¦1, ..., n
2
¦ → ¦1, ..., N¦. (3.21)
Thus, pixel (i, j) is represented by a single index p(i, j).
Let us now verify that a standard FD space discretization of an isotropic variant
of (P
c
) leads to a semidiscrete ﬁlter satisfying the requirements (S1)–(S5). To
this end, we may replace the diﬀusion tensor D(J
ρ
(∇u
σ
)) by some scalarvalued
function g(J
ρ
(∇u
σ
)). The structure tensor requires the calculations of convolutions
with ∇K
σ
and K
ρ
, respectively. In the spatially discrete case this comes down to
speciﬁc vector–matrix multiplications. For this reason, we may approximate the
structure tensor by some matrix H(u) = (h
ij
(u)) where H ∈ C
∞
(IR
N
, IR
2×2
).
Next, consider some pixel k = p(i, j). Then a consistent spatial discretization of
the isotropic diﬀusion equation with homogeneous Neumann boundary conditions
can be written as
du
k
dt
=
2
¸
n=1
¸
l∈Nn(k)
g
l
+g
k
2h
2
l
(u
l
−u
k
), (3.22)
where ^
n
(k) consists of the one or two neighbours of pixel k along the nth coor
dinate axis (boundary pixels have only one neighbour) and g
k
:= g ((H(u))
k
).
In vector–matrix notation (3.22) becomes
du
dt
= A(u) u, (3.23)
3.4. RELATION TO CONTINUOUS MODELS 87
where the matrix A(u) = (a
kl
(u))
kl
is given by
a
kl
:=
g
k
+g
l
2h
2
n
(l ∈ ^
n
(k)),
−
2
¸
n=1
¸
l∈Nn(k)
g
k
+g
l
2h
2
n
(l = k),
0 (else).
(3.24)
Let us now verify that (S1)–(S5) are fulﬁlled.
Since H ∈ C
∞
(IR
N
, IR
2×2
) and g ∈ C
∞
(IR
2×2
), we have A ∈ C
∞
(IR
N
, IR
N×N
).
This proves (S1).
The symmetry of A follows directly from (3.24) and the symmetry of the neigh
bourhood relation:
l ∈ ^
n
(k) ⇐⇒ k ∈ ^
n
(l).
By the construction of A it is also evident that all row sums vanish, i.e. (S3)
is satisﬁed. Moreover, since g is positive, it follows that a
kl
≥ 0 for all k = l and,
thus, (S4) holds.
In order to show that A is irreducible, let us consider two arbitrary pixels k and
l. Then we have to ﬁnd k
0
,...,k
r
∈ J with k
0
=k and k
r
=l such that a
kqk
q+1
= 0
for q = 0,...,r−1. If k = l, we already know from (3.24) that a
kk
< 0. In this case
we have the trivial path k = k
0
= k
r
= l. For k = l, we may choose any arbitrary
path k
0
,...,k
r
, such that k
q
and k
q+1
are neighbours for q = 0,...,r−1. Then,
a
kqk
q+1
=
g
kq
+g
k
q+1
2h
2
n
> 0
for some n ∈ ¦1, 2¦. This proves (S5).
Remarks:
(a) We observe that (S1)–(S5) are properties which are valid for all arbitrary
pixel numberings.
(b) The ﬁlter class (P
c
) is not the only family which leads to semidiscrete ﬁl
ters satisfying (S1)–(S5). Interestingly, a semidiscrete version of the Perona–
Malik ﬁlter – which is to a certain degree illposed in the continuous setting
(cf. 1.3.1) – also satisﬁes (S1)–(S5) and, thus, reveals all the discussed well
posedness and scalespace properties [425]. This is due to the fact that the
extremum principle limits the modulus of discrete gradient approximations.
Hence, the spatial discretization implicitly causes a regularization. These
results are also in accordance with a recent paper by Pollak et al. [332].
They study an image evolution under an ODE system with a discontinuous
right hand side, which has some interesting relations to the limit case of a
semidiscrete Perona–Malik model. They also report stable behaviour of their
process.
88 CHAPTER 3. SEMIDISCRETE DIFFUSION FILTERING
3.4.2 Anisotropic case
If one wishes to transfer the results from the isotropic case to the general aniso
tropic setting the main diﬃculty arises from the fact that, due to the mixed deriva
tive expressions, it is not obvious how to ensure (S4), the nonnegativity of all oﬀ
diagonal elements of A(u). The theorem below states that this is always possible
for a suﬃciently large stencil.
Theorem 6 (Existence of a nonnegative discretization).
Let D ∈ IR
2×2
be symmetric positive deﬁnite with a spectral condition number
κ. Then there exists some m(κ) ∈IN such that div (D∇u) reveals a secondorder
nonnegative FD discretization on a (2m+1)(2m+1)stencil.
Proof:
Let us consider some m ∈ IN and the corresponding (2m+1)(2m+1)stencil. The
“boundary pixels” of this stencil deﬁne 4m principal orientations β
i
∈ (−
π
2
,
π
2
],
i = −2m+1, ..., 2m according to
β
i
:=
arctan
ih
2
mh
1
([i[ ≤ m),
arccot
(2m−i)h
1
mh
2
(m < i ≤ 2m),
arccot
(i−2m)h
1
mh
2
(−2m+ 1 ≤ i < −m).
Now let J
m
:= ¦1, ..., 2m−1¦ and deﬁne a partition of (−
π
2
,
π
2
] into 4m−2 subin
tervals I
i
, [i[ ∈J
m
:
(−
π
2
,
π
2
] =
−1
¸
i=−2m+1
(θ
i
, θ
i+1
]
. .. .
=:I
i
∪
2m−1
¸
i=1
(θ
i−1
, θ
i
]
. .. .
=:I
i
,
where
θ
i
:=
0 (i = 0),
1
2
arctan
2
cot β
i
−tan β
i+1
(i ∈ ¦1, ..., 2m−2¦, β
i
+β
i+1
<
π
2
),
π
4
(i ∈ ¦1, ..., 2m−2¦, β
i
+β
i+1
=
π
2
),
π
2
+
1
2
arctan
2
cot β
i
−tan β
i+1
(i ∈ ¦1, ..., 2m−2¦, β
i
+β
i+1
>
π
2
),
π
2
(i = 2m−1),
and
θ
i
:= −θ
−i
(i ∈ ¦−2m+1, ..., −1¦).
It is not hard to verify that β
i
∈I
i
for [i[ ∈J
m
.
Let λ
1
≥ λ
2
> 0 be the eigenvalues of D with corresponding eigenvectors
(cos ψ, sin ψ)
⊤
and (−sin ψ, cos ψ)
⊤
, where ψ ∈ (−
π
2
,
π
2
]. Now we show that for
a suitable m there exists a stencil direction β
k
, [k[ ∈ J
m
such that the splitting
div (D∇u) = ∂
e
β
0
α
0
∂
e
β
0
u
+∂
e
β
k
α
k
∂
e
β
k
u
+∂
e
β
2m
α
2m
∂
e
β
2m
u
(3.25)
3.4. RELATION TO CONTINUOUS MODELS 89
with e
β
i
:= (cos β
i
, sin β
i
)
⊤
reveals nonnegative “directional diﬀusivities” α
0
, α
k
,
α
2m
along the stencil orientations β
0
, β
k
, β
2m
. This can be done by proving the
following properties:
(a) Let ψ ∈ I
k
and D =
a
b
b
c
. Then a nonnegative splitting of type (3.25) is
possible if
min
a −b cot β
k
, c −b tan β
k
≥ 0. (3.26)
(b) Inequality (3.26) is satisﬁed for
λ
1
λ
2
≤ min
cot(ρ
k
−β
k
) tanρ
k
, cot(β
k
−η
k
) cot η
k
=: κ
k,m
(3.27)
with
ρ
k
:=
θ
k
([k[ ∈ ¦1, ..., 2m−2¦),
1
2
(θ
k
+β
k
) ([k[ = 2m−1),
η
k
:=
1
2
β
k
([k[ = 1),
θ
k−1
([k[ ∈ ¦2, ..., 2m−1¦).
(c) lim
m→∞
min
i∈Jm
κ
i,m
= ∞.
Once these assertions are proved a nonnegative secondorder discretization of (3.25)
arises in a natural way, as we shall see at the end of this chapter. So let us now
verify (a)–(c).
(a) In order to use subsequent indices, let ϕ
0
:= 0, ϕ
1
:= β
k
where ψ ∈ I
k
, and
ϕ
2
:=
π
2
. Furthermore, let γ
0
:= α
0
, γ
1
:= α
k
, and γ
2
:= α
2m
. Then (3.25)
requires that
div
a b
b c
∇u
=
2
¸
i=0
∂
∂e
ϕ
i
γ
i
∂u
∂e
ϕ
i
=
∂
∂x
2
¸
i=0
cos ϕ
i
γ
i
(u
x
cos ϕ
i
+u
y
sin ϕ
i
)
+
∂
∂y
2
¸
i=0
sin ϕ
i
γ
i
(u
x
cos ϕ
i
+u
y
sin ϕ
i
)
= div
¸
¸
¸
¸
¸
¸
2
¸
i=0
γ
i
cos
2
ϕ
i
2
¸
i=0
γ
i
sin ϕ
i
cos ϕ
i
2
¸
i=0
γ
i
sin ϕ
i
cos ϕ
i
2
¸
i=0
γ
i
sin
2
ϕ
i
¸
∇u
¸
.
By comparing the coeﬃcients and using the deﬁnition of ϕ
0
, ϕ
1
and ϕ
2
we
obtain the linear system
¸
¸
1 cos
2
β
k
0
0 sin β
k
cos β
k
0
0 sin
2
β
k
1
¸
¸
¸
γ
0
γ
1
γ
2
¸
=
¸
¸
a
b
c
¸
90 CHAPTER 3. SEMIDISCRETE DIFFUSION FILTERING
which has the unique solution
γ
0
= a −b cot β
k
, (3.28)
γ
1
=
b
sin β
k
cos β
k
, (3.29)
γ
2
= c −b tanβ
k
. (3.30)
From the structure of the eigenvalues and eigenvectors of D it is easily seen
that
b = (λ
1
−λ
2
) sin ψ cos ψ.
Now, λ
1
−λ
2
≥ 0, and since ψ, β
k
∈ I
k
we conclude that ψ and β
k
belong to
the same quadrant. Thus, γ
1
is always nonnegative. In order to satisfy the
nonnegativity of γ
0
and γ
2
we need that
min
a −b cot β
k
, c −b tan β
k
≥ 0.
(b) Let
λ
1
λ
2
≤ κ
k,m
and consider the case 0 < β
k
<
π
2
. By deﬁning
B(ϕ) := cos
2
ϕ −sin ϕcos ϕcot β
k
,
C(ϕ) := sin
2
ϕ + sin ϕcos ϕcot β
k
we get
λ
1
λ
2
≤ cot(ρ
k
−β
k
) tanρ
k
= −
C(ρ
k
)
B(ρ
k
)
= min
ϕ∈(β
k
,θ
k
)
−
C(ϕ)
B(ϕ)
.
Since B(ϕ) < 0 on (β
k
,
π
2
) we have
λ
1
B(ϕ) + λ
2
C(ϕ) ≥ 0 ∀ ϕ ∈ (β
k
, θ
k
). (3.31)
Because of
B(ϕ) ≥ 0 ∀ ϕ ∈ [−
π
2
, β
k
],
C(ϕ) ≥ 0 ∀ ϕ ∈ [0,
π
2
],
and the continuity of B(ϕ) and C(ϕ) we may extend (3.31) to the entire
interval I
k
= (θ
k−1
, θ
k
]. In particular, since ψ ∈ I
k
, we have
0 ≤ λ
1
B(ψ) + λ
2
C(ψ)
= (λ
1
cos
2
ψ +λ
2
sin
2
ψ) −(λ
1
−λ
2
) sin ψ cos ψ cot β
k
.
By the representation
a b
b c
=
cos ψ −sin ψ
sin ψ cos ψ
λ
1
0
0 λ
2
cos ψ sin ψ
−sin ψ cos ψ
=
λ
1
cos
2
ψ +λ
2
sin
2
ψ (λ
1
−λ
2
) sin ψ cos ψ
(λ
1
−λ
2
) sin ψ cos ψ λ
1
sin
2
ψ +λ
2
cos
2
ψ
3.4. RELATION TO CONTINUOUS MODELS 91
we recognize that this is just the desired condition
a −b cot β
k
≥ 0. (3.32)
For the case −
π
2
< β
k
< 0 a similar reasoning can be applied leading also to
(3.32).
In an analogous way one veriﬁes that
λ
1
λ
2
≤ cot(β
k
−η
k
) cot η
k
=⇒ c −b tan β
k
≥ 0.
(c) Let us ﬁrst consider the case 1 ≤ i ≤ 2m−2. Then, ρ
i
= θ
i
, and the deﬁnition
of θ
i
implies that
cot ρ
i
−tan ρ
i
= cot β
i
−tanβ
i+1
.
Solving for cot ρ
i
and tan ρ
i
, respectively, yields
cot ρ
i
=
1
2
cot β
i
−tan β
i+1
+
(cot β
i
−tanβ
i+1
)
2
+ 4
,
tan ρ
i
=
1
2
−cot β
i
+ tanβ
i+1
+
(cot β
i
−tan β
i+1
)
2
+ 4
.
By means of these results we obtain
cot(ρ
i
−β
i
) tan ρ
i
=
cot β
i
+ tanρ
i
cot β
i
−cot ρ
i
= 1 +
2
(cot β
i
+tan β
i+1
)
2
(cot β
i
−tan β
i+1
)
2
+4
−1
.
Let us now assume that 1 ≤ i ≤ m−1. Then we have
tanβ
i+1
=
(i + 1) h
2
mh
1
,
cot β
i
=
mh
1
ih
2
.
This gives
(cot β
i
+ tanβ
i+1
)
2
(cot β
i
−tan β
i+1
)
2
+ 4
=
1
1 −
4m
2
i
m
2
h
1
h
2
+i(i+1)
h
2
h
1
2
=:
1
1 −g
m
(i)
=: f
m
(i).
For m >
1
2
h
2
h
1
the function g
m
(x) is bounded and attains its global maximum
in
x
m
:= −
1
6
+
1
6
1 + 12 m
2
h
2
1
h
2
2
.
92 CHAPTER 3. SEMIDISCRETE DIFFUSION FILTERING
Thus, for 1 ≤ i ≤ m−1,
g
m
(i) ≤ g
m
(x
m
) → 0
+
for m →∞,
which yields
f
m
(i) ≤
1
1 −g
m
(x
m
)
→ 1
+
for m →∞.
This gives
min
1≤i≤m−1
cot(ρ
i
−β
i
) tan ρ
i
≥ 1 +
2
f
m
(x
m
) −1
→ ∞ for m →∞.
For m ≤ i ≤ 2m−2 similar calculations show that by means of
tan β
i+1
=
mh
2
(2m−i−1) h
1
,
cot β
i
=
(2m−i) h
1
mh
2
one obtains
min
m≤i≤2m−2
cot(ρ
i
−β
i
) tan ρ
i
→ ∞ for m →∞.
For i = 2m−1 we have
cot
ρ
2m−1
−β
2m−1
tanρ
2m−1
= cot
π
4
−
β
2m−1
2
tan
π
4
+
β
2m−1
2
= tan
2
π
4
+
β
2m−1
2
→ ∞ for m →∞.
It is not hard to verify that for −2m + 1 ≤ i ≤ −1 the preceding results
carry over. Hence,
lim
m→∞
min
i∈Jm
cot(ρ
i
−β
i
) tanρ
i
= ∞. (3.33)
Now, in a similar way as above, one establishes that
lim
m→∞
min
i∈Jm
cot(β
i
−η
i
) cot η
i
= ∞. (3.34)
From (3.33) and (3.34) we ﬁnally end up with the assertion
lim
m→∞
min
i∈Jm
κ
i,m
= ∞.
2
3.4. RELATION TO CONTINUOUS MODELS 93
Remarks:
(a) We observe that the preceding existence proof is constructive. Moreover,
only three directions are suﬃcient to guarantee a nonnegative directional
splitting. Thus, unless m is very small, most of the stencil coeﬃcients can be
set to zero.
(b) Especially for large m, a (2m+1)(2m+1)stencil reveals much more directions
than those 4m that are induced by the 8m “boundary pixels”. Therefore,
even if we use only 3 directions, we may expect to ﬁnd stricter estimates
than those given in the proof. These estimates might be improved further by
admitting more than 3 directions.
(c) For a speciﬁed diﬀusion tensor function D it is possible to give apriori es
timates for the required stencil size: using the extremum principle it is not
hard to show that
[∇u
σ
(x, t)[ = [(∇K
σ
∗u)(x, t)[ ≤
4 f
L
∞
(Ω)
√
2π σ
on
¯
Ω (0, ∞),
where the notations from Chapter 2 have been used. Thanks to the uniform
positive deﬁniteness of D there exists an upper limit for the spectral condition
number of D. This condition limit can be used to ﬁx a suitable stencil size.
(d) The existence of a nonnegative directional splitting distinguishes the ﬁlter
class (P
c
) from morphological anisotropic equations such as mean curvature
motion. In this case it has been proved that it is impossible to ﬁnd a nonneg
ative directional splitting on a ﬁnite stencil [13]. As a remedy, Crandall and
Lions [104] propose to study a convergent sequence of regularizations which
can be approximated on a ﬁnite stencil.
Let us now illustrate the ideas in the proof of Theorem 6 by applying them
to a practical example: We want to ﬁnd a nonnegative spatial discretization of
div (D∇u) on a (33)stencil, where
D =
a
b
b
c
and a, b and c may be functions of J
ρ
(∇u
σ
).
Since m = 1 we have a partition of (−
π
2
,
π
2
] into 4m−2 = 2 subintervals:
(−
π
2
,
π
2
] = (−
π
2
, 0] ∪ (0,
π
2
] =: I
−1
∪ I
1
.
94 CHAPTER 3. SEMIDISCRETE DIFFUSION FILTERING
I
−1
and I
1
belong to the grid angles
β
−1
= arctan
−
h
2
h
1
,
β
1
= arctan
h
2
h
1
=: β.
First we focus on the case ψ ∈ I
1
where (cos ψ, sin ψ) denotes the eigenvector to
the larger eigenvalue λ
1
of D. With the notations from the proof of Theorem 6 we
obtain
θ
1
=
π
2
,
ρ
1
=
θ
1
+β
1
2
=
π
4
+
β
2
,
η
1
=
β
2
.
Therefore, we get
cot(ρ
1
−β
1
) tan ρ
1
= cot
π
4
−
β
2
tan
π
4
+
β
2
=
1 + sin β
1 −sin β
,
cot(β
1
−η
1
) cot η
1
= cot
2
β
2
=
1 + cos β
1 −cos β
,
which restricts the upper condition number for a nonnegative discretization with
ψ ∈ I
1
to
κ
1,1
:= min
1 + sin β
1 −sin β
,
1 + cos β
1 −cos β
. (3.35)
Thanks to the symmetry we obtain the same condition restriction for ψ ∈ I
−1
.
These bounds on the condition number attain their maximal value for h
1
= h
2
. In
this case β =
π
4
gives
κ
1,1
= κ
−1,1
=
1 +
1
2
√
2
1 −
1
2
√
2
= 3 + 2
√
2 ≈ 5.8284. (3.36)
By virtue of (3.28)–(3.30) we obtain as expressions for the directional diﬀusivities
α
−1
=
[b[ −b
h
2
1
+h
2
2
2h
1
h
2
,
α
0
= a −[b[
h
1
h
2
,
α
1
=
[b[ +b
h
2
1
+h
2
2
2h
1
h
2
,
α
2
= c −[b[
h
2
h
1
.
3.4. RELATION TO CONTINUOUS MODELS 95
This induces in a natural way the following secondorder discretization for div (D∇u):
b
i−1,j+1
−b
i−1,j+1
4h
1
h
2
+
b
i,j
−b
i,j
4h
1
h
2
c
i,j+1
+c
i,j
2h
2
2
−
b
i,j+1
+b
i,j

2h
1
h
2
b
i+1,j+1
+b
i+1,j+1
4h
1
h
2
+
b
i,j
+b
i,j
4h
1
h
2
a
i−1,j
+a
i,j
2h
2
1
−
b
i−1,j
+b
i,j

2h
1
h
2
−
a
i−1,j
+2a
i,j
+a
i+1,j
2h
2
1
−
b
i−1,j+1
−b
i−1,j+1
+b
i+1,j+1
+b
i+1,j+1
4h
1
h
2
−
b
i−1,j−1
+b
i−1,j−1
+b
i+1,j−1
−b
i+1,j−1
4h
1
h
2
+
b
i−1,j
+b
i+1,j
+b
i,j−1
+b
i,j+1
+2b
i,j

2h
1
h
2
−
c
i,j−1
+2c
i,j
+c
i,j+1
2h
2
2
a
i+1,j
+a
i,j
2h
2
1
−
b
i+1,j
+b
i,j

2h
1
h
2
b
i−1,j−1
+b
i−1,j−1
4h
1
h
2
+
b
i,j
+b
i,j
4h
1
h
2
c
i,j−1
+c
i,j
2h
2
2
−
b
i,j−1
+b
i,j

2h
1
h
2
b
i+1,j−1
−b
i+1,j−1
4h
1
h
2
+
b
i,j
−b
i,j
4h
1
h
2
All nonvanishing entries of the pth row of A(u) are represented in this stencil,
where p(i, j) is the index of some inner pixel (i, j). Thus, for instance, the upper
left stencil entry gives the element (p(i, j), p(i−1, j+1)) of A(u). The other nota
tions should be clear from the context as well, e.g. b
i,j
denotes a ﬁnite diﬀerence
approximation of b(J
ρ
(∇u
σ
)) at some grid point (x
i
, y
j
).
The problem of ﬁnding nonnegative diﬀerence approximations to elliptic expres
sions with mixed derivatives has a long history; see e.g. [294, 120, 170]. Usually it
is studied for the expression
a(x, y) ∂
xx
u + 2b(x, y) ∂
xy
u +c(x, y) ∂
yy
u.
The approach presented here extends these results to
∂
x
a(x, y) ∂
x
u
+∂
x
b(x, y) ∂
y
u
+∂
y
b(x, y) ∂
x
u
+∂
y
c(x, y) ∂
y
u
and establishes the relation between the condition number of
a
b
b
c
and the non
negativity of the diﬀerence operator. Recently Kocan described an interesting al
ternative to obtain upper bounds for the stencil size as a function of the condition
number [238]. His derivation is based on the diophantine problem of approximating
irrationals by rational numbers.
96 CHAPTER 3. SEMIDISCRETE DIFFUSION FILTERING
Chapter 4
Discrete diﬀusion ﬁltering
This chapter presents a discrete class of diﬀusion processes for which one can estab
lish similar properties as in the semidiscrete case concerning existence, uniqueness,
continuous dependence of the solution on the initial image, maximumminimum
principle, average grey level invariance, Lyapunov sequences and convergence to a
constant steadystate. We shall see that this class comprises αsemiimplicit dis
cretizations of the semidiscrete ﬁlter class (P
s
) as well as certain variants of them
which are based on an additive operator splitting.
4.1 The general model
As in Chapter 3 we regard a discrete image as a vector f ∈IR
N
, N≥2, and denote
the index set ¦1, ..., N¦ by J. We consider the following discrete ﬁlter class (P
d
):
Let f ∈IR
N
. Calculate a sequence (u
(k)
)
k∈IN
0
of processed versions
of f by means of
u
(0)
= f,
u
(k+1)
= Q(u
(k)
) u
(k)
, ∀ k ∈ IN
0
,
where Q = (q
ij
) has the following properties:
(D1) continuity in its argument: Q ∈ C(IR
N
, IR
N×N
),
(D2) symmetry: q
ij
(v) = q
ji
(v) ∀ i, j ∈ J, ∀ v ∈ IR
N
,
(D3) unit row sum:
¸
j∈J
q
ij
(v) = 1 ∀ i ∈ J, ∀ v ∈ IR
N
,
(D4) nonnegativity: q
ij
(v) ≥ 0 ∀ i, j ∈ J, ∀ v ∈ IR
N
,
(D5) irreducibility for all v ∈ IR
N
,
(D6) positive diagonal: q
ii
(v) > 0 ∀ i ∈ J, ∀ v ∈ IR
N
.
(P
d
)
97
98 CHAPTER 4. DISCRETE DIFFUSION FILTERING
Remarks:
(a) Although the basic idea behind scalespaces is to have a continuous scale
parameter, it is evident that fully discrete results are of importance since, in
practice, scalespace evolutions are evaluated exclusively at a ﬁnite number
of scales.
(b) The requirements (D1)–(D5) have a similar meaning as their semidiscrete
counterparts (S1)–(S5). Indeed, (D1) immediately gives wellposedness re
sults, while the proof of the extremum principle requires (D3) and (D4),
and average grey value invariance is based on (D2) and (D3). The existence
of Lyapunov sequences is a consequence of (D2)–(D4), strict Lyapunov se
quences need (D5) and (D6) in addition to (D2)–(D4), and the convergence
to a constant steadystate utilizes (D2)–(D5).
(c) Nonnegative matrices Q = (q
ij
) ∈IR
N×N
satisfying
¸
j∈J
q
ij
=1 for all i ∈J
are also called stochastic matrices. Moreover, if Q is stochastic and
¸
i∈J
q
ij
=
1 for all j ∈J, then Q is doubly stochastic. This indicates that our discrete
diﬀusion processes are related to the theory of Markov chains [370, 223].
4.2 Theoretical results
It is obvious that for a ﬁxed ﬁlter belonging to the class (P
d
) every initial image
f ∈ IR
N
generates a unique sequence (u
(k)
)
k∈IN
0
. Moreover, by means of (D1) we
know that, for every ﬁnite k, u
(k)
depends continuously on f. Therefore, let us now
prove a maximum–minimum principle.
Proposition 3 (Extremum principle).
Let f ∈IR
N
and let (u
(k)
)
k∈IN
0
be the sequence of ﬁltered images according to (P
d
).
Then,
a ≤ u
(k)
i
≤ b ∀ i ∈ J, ∀ k ∈ IN
0
, (4.1)
where
a := min
j∈J
f
j
, (4.2)
b := max
j∈J
f
j
. (4.3)
Proof:
The maximum–minimum principle follows directly from the fact that, for all i ∈J
and k∈IN
0
, the following inequalities hold:
(i) u
(k+1)
i
=
¸
j∈J
q
ij
(u
(k)
)u
(k)
j
(D4)
≤ max
m∈J
u
(k)
m
¸
j∈J
q
ij
(u
(k)
)
(D3)
= max
m∈J
u
(k)
m
.
4.3. SCALESPACE PROPERTIES 99
(ii) u
(k+1)
i
=
¸
j∈J
q
ij
(u
(k)
)u
(k)
j
(D4)
≥ min
m∈J
u
(k)
m
¸
j∈J
q
ij
(u
(k)
)
(D3)
= min
m∈J
u
(k)
m
.
2
4.3 Scalespace properties
All statements from Chapter 3 with respect to invariances are valid in the discrete
framework as well. Below we focus on proving average grey level invariance.
Proposition 4 (Conservation of average grey value).
The average grey level
µ :=
1
N
¸
j∈J
f
j
(4.4)
is not aﬀected by the discrete diﬀusion ﬁlter:
1
N
¸
j∈J
u
(k)
j
= µ ∀ k ∈ IN
0
. (4.5)
Proof:
By virtue of (D2) and (D3) we have
¸
i∈J
q
ij
(u
(k)
)=1 for all j ∈J and k∈IN
0
. This
socalled redistribution property [164] ensures that, for all k ∈ IN
0
,
¸
i∈J
u
(k+1)
i
=
¸
i∈J
¸
j∈J
q
ij
(u
(k)
)u
(k)
j
=
¸
j∈J
¸
i∈J
q
ij
(u
(k)
)
u
(k)
j
=
¸
j∈J
u
(k)
j
,
which proves the proposition. 2
As one might expect, the class (P
d
) allows an interpretation as a transformation
which is smoothing in terms of Lyapunov sequences. These functions ensure that
u
(k)
converges to a constant image as k → ∞. However, we need less regularity
than in the semidiscrete case: The convex function r, which generates the Lyapunov
sequences, needs only to be continuous, but no more diﬀerentiable.
Theorem 7 (Lyapunov sequences and behaviour for k →∞).
Assume that (u
(k)
)
k∈IN
0
satisﬁes the requirements of (P
d
), let a, b, and µ be deﬁned
as in (4.2), (4.3), and (4.4), respectively, and let c := (µ, µ, ..., µ)
⊤
∈ IR
N
.
Then the following properties are fulﬁlled:
(a) (Lyapunov sequences)
For all convex r ∈ C[a, b] the sequence
V
(k)
:= Φ(u
(k)
) :=
¸
i∈J
r(u
(k)
i
), k ∈ IN
0
is a Lyapunov sequence:
100 CHAPTER 4. DISCRETE DIFFUSION FILTERING
(i) Φ(u
(k)
) ≥ Φ(c) ∀ k ∈ IN
0
(ii) V
(k+1)
−V
(k)
≤ 0 ∀ k ∈ IN
0
Moreover, if r is strictly convex, then V
(k)
= Φ(u
(k)
) is a strict Lyapunov
sequence:
(iii) Φ(u
(k)
) = Φ(c) ⇐⇒ u
(k)
= c
(iv) V
(k+1)
−V
(k)
= 0 ⇐⇒ u
(k)
= c
(b) (Convergence)
lim
k→∞
u
(k)
= c.
Proof:
(a) (i) Average grey level invariance and the convexity of r give
Φ(c) =
N
¸
i=1
r
¸
N
¸
j=1
1
N
u
(k)
j
¸
≤
N
¸
i=1
¸
1
N
N
¸
j=1
r(u
(k)
j
)
¸
=
N
¸
j=1
r(u
(k)
j
)
= Φ(u
(k)
). (4.6)
(ii) For i, j ∈ J we deﬁne
a
ij
(u
(k)
) :=
q
ij
(u
(k)
) −1 (i = j)
q
ij
(u
(k)
) (i = j).
(4.7)
Using the convexity of r, the preceding deﬁnition, and the prerequisites
(D2) and (D3) we obtain
V
(k+1)
−V
(k)
=
N
¸
i=1
¸
r
¸
N
¸
j=1
q
ij
(u
(k)
) u
(k)
j
¸
−r(u
(k)
i
)
¸
conv.
≤
N
¸
i=1
¸
N
¸
j=1
q
ij
(u
(k)
) r(u
(k)
j
) −r(u
(k)
i
)
¸
(4.7)
=
N
¸
i=1
N
¸
j=1
a
ij
(u
(k)
) r(u
(k)
j
)
(D3)
=
N
¸
i=1
N
¸
j=1
a
ij
(u
(k)
)
r(u
(k)
j
) −r(u
(k)
i
)
4.3. SCALESPACE PROPERTIES 101
=
N
¸
i=1
N−i
¸
m=1
a
i+m,i
(u
(k)
)
r(u
(k)
i
) −r(u
(k)
i+m
)
+
N
¸
i=1
N−i
¸
m=1
a
i,i+m
(u
(k)
)
r(u
(k)
i+m
) −r(u
(k)
i
)
(D2)
= 0. (4.8)
(iii) This part of the proof can be shown in exactly the same manner as in
the semidiscrete case (Chapter 3, Theorem 5): Equality in the estimate
(4.6) holds due to the strict convexity of r if and only if u
(k)
= c.
(iv) In order to verify the ﬁrst implication, let us start with a proof that
V
(k+1)
= V
(k)
implies u
(k)
1
= ... = u
(k)
N
. To this end, assume that u
(k)
is
not constant:
u
(k)
i
0
:= min
i∈J
u
(k)
i
< max
j∈J
u
(k)
j
=: u
(k)
j
0
.
Then, by the irreducibility of Q(u
(k)
), we ﬁnd l
0
, ..., l
r
∈ J with l
0
= i
0
,
l
r
= j
0
and q
lpl
p+1
= 0 for p = 0, ..., r −1. Hence, there exists some
p
0
∈ ¦0, ..., r −1¦ such that n := l
p
0
, m := l
p
0
+1
, q
nm
(u
(k)
) = 0, and
u
(k)
m
= u
(k)
n
. Moreover, the nonnegativity of Q(u
(k)
) gives q
nm
(u
(k)
) > 0,
and by (D6) we have q
nn
(u
(k)
) > 0. Together with the strict convexity
of r these properties lead to
r
¸
N
¸
j=1
q
nj
(u
(k)
) u
(k)
j
¸
= r
¸
¸
N
¸
j=1
j=n,m
q
nj
(u
(k)
) u
(k)
j
+q
nn
(u
(k)
) u
(k)
n
+q
nm
(u
(k)
) u
(k)
m
¸
<
N
¸
j=1
j=n,m
q
nj
(u
(k)
) r(u
(k)
j
) + q
nn
(u
(k)
) r(u
(k)
n
) + q
nm
(u
(k)
) r(u
(k)
m
)
=
N
¸
j=1
q
nj
(u
(k)
) r(u
(k)
j
).
If we combine this with the results in (4.8), we obtain
V
(k+1)
−V
(k)
=
N
¸
i=1
i=n
¸
r
¸
N
¸
j=1
q
ij
(u
(k)
) u
(k)
j
¸
−r(u
(k)
i
)
¸
+ r
¸
N
¸
j=1
q
nj
(u
(k)
) u
(k)
j
¸
−r(u
(k)
n
)
102 CHAPTER 4. DISCRETE DIFFUSION FILTERING
<
N
¸
i=1
¸
N
¸
j=1
q
ij
(u
(k)
) r(u
(k)
j
) −r(u
(k)
i
)
¸
(4.8)
= 0.
This establishes that V
(k+1)
= V
(k)
implies u
(k)
1
= ... = u
(k)
N
. Then, by
virtue of the grey value invariance, we conclude that u
(k)
= c.
Conversely, let u
(k)
= c. By means of prerequisite (D3) we obtain
V
(k+1)
−V
(k)
=
N
¸
i=1
r
¸
N
¸
j=1
q
ij
(u
(k)
)µ
¸
−
N
¸
i=1
r(µ) = 0.
(c) In order to prove convergence to a constant steadystate, we can argue exactly
in the same way as in the semidiscrete case if we replace Lyapunov functions
by Lyapunov sequences and integrals by sums. See Chapter 3, Theorem 5 for
more details. 2
In analogy to the semidiscrete case the preceding theorem comprises many
Lyapunov functions which demonstrate the informationreducing qualities of our
ﬁlter class. Choosing the convex functions r(s) := [s[
p
, r(s) := (s −µ)
2n
and
r(s) := s ln s, we immediately obtain the following corollary.
Corollary 3 (Special Lyapunov sequences).
Let (u
(k)
)
k∈IN
0
be a diﬀusion sequence according to (P
d
), and let a and µ be deﬁned
as in (4.2) and (4.4). Then the following functions are decreasing in k:
(a) u
(k)

p
for all p ≥ 1.
(b) M
2n
[u
(k)
] :=
1
N
N
¸
j=1
(u
(k)
j
−µ)
2n
for all n ∈ IN.
(c) H[u
(k)
] :=
N
¸
j=1
u
(k)
j
ln(u
(k)
j
), if a > 0.
An interpretation of these results in terms of decreasing energy, decreasing
central moments and increasing entropy is evident.
4.4 Relation to semidiscrete models
4.4.1 Semiimplicit schemes
Let us now investigate in which sense our discrete ﬁlter class covers in a natural
way time discretizations of semidiscrete ﬁlters. To this end, we regard u
(k)
as an
4.4. RELATION TO SEMIDISCRETE MODELS 103
approximation of the solution u of (P
s
) at time t = kτ, where τ denotes the time
step size. We consider a ﬁnite diﬀerence scheme with two time levels where the
operator A – which depends nonlinearly on u – is evaluated in an explicit way,
while the linear remainder is discretized in an αimplicit manner. Such schemes
are called αsemiimplicit. They reveal the advantage that the linear implicit part
ensures good stability properties, while the explicit evaluation of the nonlinear
terms avoids the necessity to solve nonlinear systems of equations. The theorem
below states that this class of schemes is covered by the discrete framework, for
which we have established scalespace results.
Theorem 8 (Scalespace interpretation for αsemiimplicit schemes).
Let α ∈ [0, 1], τ > 0, and let A = (a
ij
) : IR
N
→ IR
N×N
satisfy the requirements
(S1)–(S5) of section 3.1. Then the αsemiimplicit scheme
u
(k+1)
−u
(k)
τ
= A(u
(k)
)
αu
(k+1)
+ (1−α)u
(k)
(4.9)
fulﬁls the prerequisites (D1)–(D6) for discrete diﬀusion models provided that
τ ≤
1
(1−α) max
i∈J
[a
ii
(u
(k)
)[
(4.10)
for α ∈ (0, 1). In the explicit case (α=0) the properties (D1)–(D6) hold for
τ <
1
max
i∈J
[a
ii
(u
(k)
)[
, (4.11)
and the semiimplicit case (α=1) satisﬁes (D1)–(D6) unconditionally.
Proof: Let
B(u
(k)
) := (b
ij
(u
(k)
)) := I −ατA(u
(k)
),
C(u
(k)
) := (c
ij
(u
(k)
)) := I + (1−α)τA(u
(k)
),
where I ∈IR
N
denotes the unit matrix. Since (4.9) can be written as
B(u
(k)
) u
(k+1)
= C(u
(k)
) u
(k)
we ﬁrst have to show that B(u
(k)
) is invertible for all u
(k)
∈ IR
N
. Henceforth, the
argument u
(k)
is suppressed frequently since the considerations below are valid for
all u
(k)
∈ IR
N
.
If α=0, then B=I and hence invertible. Now assume that α > 0. Then B is
strictly diagonally dominant, since
b
ii
= 1 −ατa
ii
(S3)
= 1 +ατ
¸
j∈J
j=i
a
ij
> ατ
¸
j∈J
j=i
a
ij
(S4)
=
¸
j∈J
j=i
[b
ij
[ ∀ i ∈ J.
104 CHAPTER 4. DISCRETE DIFFUSION FILTERING
This also shows that b
ii
> 0 for all i ∈ J, and by the structure of the oﬀdiagonal
elements of B we observe that the irreducibility of A implies the irreducibility of
B. Thanks to the fact that B is irreducibly diagonally dominant, b
ij
≤ 0 for all
i = j, and b
ii
> 0 for all i ∈ J, we know from [407, p. 85] that B
−1
=: H =: (h
ij
)
exists and h
ij
> 0 for all i, j ∈ J. Thus, Q := (q
ij
) := B
−1
C exists and by (S1) it
follows that Q ∈ C(IR
N
, IR
N×N
). This proves (D1).
The requirement (D2) is not hard to satisfy: Since B
−1
and C are symmetric
and reveal the same set of eigenvectors – namely those of A – it follows that
Q = B
−1
C is symmetric as well.
Let us now verify (D3). By means of (S3) we obtain
¸
j∈J
b
ij
= 1 =
¸
j∈J
c
ij
∀ i ∈ J. (4.12)
Let v := (1, ..., 1)
⊤
∈ IR
N
. Then (4.12) is equivalent to
Bv = v = Cv, (4.13)
and the invertibility of B gives
v = B
−1
v = Hv. (4.14)
Therefore, from
Qv = HCv
(4.13)
= Hv
(4.14)
= v
we conclude that
¸
j∈J
q
ij
= 1 for all i ∈ J. This proves (D3).
In order to show that (D4) is fulﬁlled, we ﬁrst check the nonnegativity of C.
For i = j we have
c
ij
= (1−α)τa
ij
(S4)
≥ 0.
The diagonal entries yield
c
ii
= 1 + (1−α)τa
ii
.
If α = 1 we have c
ii
= 1 for all i ∈ J. For 0 ≤ α < 1, however, nonnegativity of C
is not automatically guaranteed: Using (S3)–(S5) we obtain
a
ii
(S3)
= −
¸
j∈J
j=i
a
ij
(S4),(S5)
< 0 ∀ i ∈ J. (4.15)
Hence, C(u
(k)
) is nonnegative if
τ ≤
1
(1−α) max
i∈J
[a
ii
(u
(k)
)[
=: τ
α
(u
(k)
).
4.4. RELATION TO SEMIDISCRETE MODELS 105
Since H is nonnegative, we know that the nonnegativity of C implies the nonneg
ativity of Q=HC.
Now we want to prove (D5). If α = 1, then C = I, and by the positivity of H
we have q
ij
> 0 for all i, j ∈ J. Thus, Q is irreducible.
Next let us consider the case 0 <α <1 and τ ≤τ
α
(u
(k)
). Then we know that
C is nonnegative. Using this information, the positivity of H, the symmetry of C,
and (4.12) we obtain
q
ij
=
¸
k∈J
h
ik
c
kj
≥ min
k∈J
h
ik
. .. .
>0
¸
k∈J
c
kj
. .. .
=1
> 0 ∀ i, j ∈ J,
which establishes the irreducibility of Q.
Finally, for α = 0, we have Q = C. For i, j ∈ J with i = j we know that
a
ij
(u
(k)
) > 0 implies c
ij
(u
(k)
) > 0. Now for
τ <
1
max
i∈J
[a
ii
(u
(k)
)[
it follows that c
ii
(u
(k)
) > 0 for all i ∈ J and, thus, the irreducibility of A(u
(k)
)
carries over to Q(u
(k)
).
In all the abovementioned cases the time step size restrictions for ensuring irre
ducibility imply that all diagonal elements of Q(u
(k)
) are positive. This establishes
(D6). 2
Remarks:
(a) We have seen that the discrete ﬁlter class (P
d
) – although at ﬁrst glance
looking like a pure explicit discretization – covers the αsemiimplicit case as
well. Explicit twolevel schemes are comprised by the choice α=0. Equation
(4.11) shows that they reveal the most prohibitive time step size restrictions.
(b) The conditions (4.10) and (4.11) can be satisﬁed by means of an apriori
estimate. Since the semiimplicit scheme fulﬁls (D1)–(D6) we know by The
orem 3 that the solution obeys an extremum principle. This means that u
(k)
belongs to the compact set ¦v ∈ IR
N
v
∞
≤ f
∞
¦ for all k ∈ IN
0
. By
A∈C(IR
N
, IR
N×N
) it follows that
K
f
:= max
[a
ii
(v)[
i ∈ J, v ∈ IR
N
, v
∞
≤ f
∞
¸
exists, and (4.15) shows that K
f
> 0. Thus, choosing
τ ≤
1
(1−α) K
f
106 CHAPTER 4. DISCRETE DIFFUSION FILTERING
ensures that (4.10) is always satisﬁed, and
τ <
1
K
f
guarantees that (4.11) holds.
(c) If α>0, a large linear system of equations has to be solved. Its system matrix
is symmetric, diagonally dominant, and positive deﬁnite. Usually, it is also
sparse: For instance, if it results from a ﬁnite diﬀerence discretization on a
(2p+1)(2p+1)stencil it contains at most 4p
2
+4p+1 nonvanishing entries
per row. One should not expect, however, that in the ith row these entries
can be found within the positions [i, i−2p
2
−2p] to [i, i+2p
2
+2p]. In general,
the matrix reveals a much larger bandwidth.
Applying standard direct algorithms such as Gaussian elimination would
destroy the zeros within the band and would lead to an immense storage and
computation eﬀort. Modiﬁcations in order to reduce these problems [122] are
quite diﬃcult to implement.
Iterative algorithms appear to be better suited. Classical methods such as
Gauß–Seidel or SOR [447] are easy to code, they do not need additional
storage, and their convergence can be guaranteed for the special structure of
A. Unfortunately, they converge rather slowly. Faster iterative methods such
as preconditioned conjugate gradient algorithms [348] need signiﬁcantly more
storage, which can become prohibitive for large images. A typical problem of
iterative methods is that their convergence slows down for larger τ, since this
increases the condition number of the system matrix. Multigrid methods [59,
179] are one possibility to circumvent these diﬃculties; another possibility
will be studied in Section 4.4.2.
(d) For α = 1 we obtain semiimplicit schemes which do not suﬀer from time
step size restrictions. In spite of the fact that the nonlinearity is discretized
in an explicit way they are absolutely stable in the maximum norm, and
they inherit the scalespace properties from the semidiscrete setting regard
less of the step size. Compared to explicit schemes, this advantage usually
overcompensates for the additional eﬀort of resolving a linear system.
(e) By the explicit discretization of the nonlinear operator A it follows that all
schemes in the preceding theorem are of ﬁrst order in time. This should not
give rise to concern, since in image processing one is in general more inter
ested in maintaining qualitative properties such as maximum principles or
invariances rather than having an accurate approximation of the continu
ous equation. However, if one insists in secondorder schemes, one may for
4.4. RELATION TO SEMIDISCRETE MODELS 107
instance use the predictor–corrector approach by Douglas and Jones [121]:
u
(k+1/2)
−u
(k)
τ/2
= A(u
(k)
) u
(k+1/2)
,
u
(k+1)
−u
(k)
τ
= A(u
(k+1/2)
)
1
2
u
(k+1)
+
1
2
u
(k)
.
This scheme satisﬁes the properties (D1)–(D6) for τ ≤ 2/K
f
.
(f) The assumptions (S1)–(S5) are suﬃcient conditions for the αsemiimplicit
scheme to fulﬁl (D1)–(D6), but they are not necessary. Nonnegativity of
Q(u
(k)
) may also be achieved using spatial discretizations where A(u
(k)
) has
negative oﬀdiagonal elements (see [55] for examples).
4.4.2 AOS schemes
We have seen that, for α > 0, the preceding αsemiimplicit schemes require to
solve a linear system with the system matrix (I − ατA(u
(k)
)). Since this can be
numerically expensive, it would be nice to have an eﬃcient alternative. Suppose
we know a splitting
A(u
(k)
) =
m
¸
l=1
A
l
(u
(k)
), (4.16)
such that the m linear systems with system matrices (I −mατA
l
(u
(k)
)), l = 1,...,m
can be solved more eﬃciently. Then it is advantageous to study instead of the α
semiimplicit scheme
u
(k+1)
=
I −ατ
m
¸
l=1
A
l
(u
(k)
)
−1
I + (1−α)τ
m
¸
l=1
A
l
(u
(k)
)
u
(k)
(4.17)
its additive operator splitting (AOS) variant [424]
u
(k+1)
=
1
m
m
¸
l=1
I −αmτA
l
(u
(k)
)
−1
I + (1−α)mτA
l
(u
(k)
)
u
(k)
. (4.18)
By means of a Taylor expansion one can verify that, although both schemes are not
identical, they have the same approximation order in space and time. Hence, from
a numerical viewpoint, they are both consistent approximations to the semidiscrete
ODE system from (P
s
).
The following theorem clariﬁes the conditions under which AOS schemes create
discrete scalespaces.
108 CHAPTER 4. DISCRETE DIFFUSION FILTERING
Theorem 9 (Scalespace interpretation for AOS schemes).
Let α ∈ (0, 1], τ >0, and let A
l
= (a
ijl
)
ij
: IR
N
→ IR
N×N
, l = 1,...,m satisfy the
requirements (S1)–(S4) of section 3.1. Moreover, assume that A(u) =
¸
m
l=1
A
l
(u)
is irreducible for all u ∈ IR
N
, and that for each A
l
there exists a permutation
matrix P
l
∈ IR
N×N
such that P
l
A
l
P
T
l
is block diagonal and irreducible within each
block. Then the following holds:
For α ∈ (0, 1), the AOS scheme (4.18) fulﬁls the prerequisites (D1)–(D6) of
discrete diﬀusion scalespaces provided that
τ <
1
(1−α) m max
i,l
[a
iil
(u
(k)
)[
. (4.19)
In the semiimplicit case (α = 1), the properties (D1)–(D6) are unconditionally
satisﬁed.
Proof: The reasoning is similar to the proof of Theorem 8. Let
B
l
(u
(k)
) := (b
ijl
(u
(k)
))
ij
:= I −αmτA
l
(u
(k)
),
C
l
(u
(k)
) := (c
ijl
(u
(k)
))
ij
:= I + (1−α)mτA
l
(u
(k)
).
B
l
is invertible because of its strict diagonal dominance:
b
iil
(S3)
= 1 +αmτ
¸
j∈J
j=i
a
ijl
> αmτ
¸
j∈J
j=i
a
ijl
(S4)
=
¸
j∈J
j=i
[b
ijl
[ ∀ i ∈ J.
Since A
l
(u) is continuous in u by virtue of (S1), it follows that
Q(u) :=
1
m
m
¸
l=1
B
−1
l
(u) C
l
(u)
is also continuous in u. This proves (D1).
The symmetry property (D2) of Q results directly from the fact that B
−1
l
and
C
l
are symmetric and share their eigenvectors with those of A
l
.
In the same way as in the proof of Theorem 8 one shows that B
−1
l
C
l
has only
unit row sums for all l. Thus, the row sums of Q are 1 as well, and (D3) is satisﬁed.
To verify (D4), we utilize that B
l
is strictly diagonally dominant, b
iil
> 0 for
all i, and b
ijl
≤ 0 for i = j. Under these circumstances it follows from [284, p. 192]
that H
l
:= B
−1
l
is nonnegative in all components. Thus, a suﬃcient condition for
proving (D4) is to ensure that C
l
is nonnegative for all l. For i = j we have
c
ijl
= (1−α)mτa
ijl
(S4)
≥ 0.
4.4. RELATION TO SEMIDISCRETE MODELS 109
The diagonal entries yield
c
iil
= 1 + (1−α)mτa
iil
.
If α = 1 we have c
ii
= 1 for all i ∈ J. For 0 < α < 1, however, nonnegativity of
C is not automatically guaranteed: Since A
l
satisﬁes (S3) and (S4), we know that
a
iil
≤ 0, for all i. Moreover, by (4.15) it follows that for every i there exists an l
with a
iil
< 0. Thus, requiring
τ <
1
(1−α) m max
i,l
[a
iil
(u
(k)
)[
=: τ
α
(u
(k)
)
guarantees that
c
iil
> 0 ∀ i ∈ J, ∀ l = 1, ..., m. (4.20)
Next we prove (D5), the irreducibility of Q. Suppose that a
i
0
j
0
= 0 for some
i
0
, j
0
∈ J. Then there exists an l
0
∈ ¦1, ..., m¦ such that a
i
0
j
0
l
0
= 0. Denoting B
−1
l
by H
l
= (h
ijl
)
ij
, we show now that a
i
0
j
0
l
0
= 0 implies h
i
0
j
0
l
0
> 0.
This can be seen as follows: There exist permutation matrices P
l
, l = 1, ..., m
such that P
l
B
l
P
T
l
is block diagonal. Each block is irreducible and strictly diagonally
dominant with a positive diagonal and nonpositive oﬀdiagonals. Thus, a theorem
by Varga [407, p. 85] ensures that the inverse of each block contains only positive
elements. As a consequence, a
i
0
j
0
l
0
= 0 implies h
i
0
j
0
l
0
> 0.
Together with (4.20) this yields
q
i
0
j
0
=
1
m
¸
¸
¸
(l,n)=(l
0
,j
0
)
h
i
0
nl
. .. .
≥0
c
nj
0
l
. .. .
≥0
+ h
i
0
j
0
l
0
. .. .
>0
c
j
0
j
0
l
0
. .. .
>0
¸
> 0.
Recapitulating, this means that, for τ < τ
α
(u
(k)
),
a
i
0
j
0
= 0 =⇒ q
i
0
j
0
> 0. (4.21)
Thus, the irreducibility of A carries over to Q, and (D5) is proved.
Moreover, (4.21) also proves (D6): By virtue of (4.15) we have a
ii
< 0 for all
i ∈ J. Therefore, Q must have a positive diagonal. 2
Remarks:
(a) In analogy to the unsplit αsemiimplicit schemes, the case α = 1 is especially
interesting, because no time step size restriction occurs. Again, it is also
possible to construct a predictor–corrector scheme of Douglas–Jones type
[121] within the AOS framework:
u
(k+1/2)
=
1
m
m
¸
l=1
I −
1
2
mτA
l
(u
(k)
)
−1
u
(k)
,
u
(k+1)
=
1
m
m
¸
l=1
I −
1
2
mτA
l
(u
(k+1/2)
)
−1
I +
1
2
mτA
l
(u
(k+1/2)
)
u
(k)
.
110 CHAPTER 4. DISCRETE DIFFUSION FILTERING
It satisﬁes (D1)–(D6) for τ < 2/K
f
, where K
f
is determined by the apriori
estimate
K
f
:= m max
[a
iil
(v)[
i ∈ J, l ∈ ¦1, ..., m¦, v ∈ IR
N
, v
∞
≤ f
∞
¸
.
However, the AOS–Douglas–Jones scheme is only ﬁrst order accurate in time.
(b) The fact that AOS schemes use an additive splitting ensures that all coordi
nate axes are treated in exactly the same manner. This is in contrast to the
various conventional splitting techniques from the literature [120, 277, 286,
354, 442]. They are multiplicative splittings. A typical representative is
u
(k+1)
=
m
¸
l=1
I −τA
l
(u
(k)
)
−1
u
(k)
.
Since in the general nonlinear case the split operators A
l
, l = 1,..., m do not
commute, the result of multiplicative splittings will depend on the order of
the operators. In practice, this means that these schemes produce diﬀerent
results if the image is rotated by 90 degrees. Moreover, most multiplicative
splittings lead to a nonsymmetric matrix Q(u
(k)
). This violates requirement
(D2) for discrete scalespaces.
(c) The result u
(k+1)
of an AOS scheme can be regarded as the average of m
ﬁlters of type
v
(k+1)
l
:=
I −αmτA
l
(u
(k)
)
−1
I +(1−α)mτA
l
(u
(k)
)
u
(k)
(l = 1, ..., m).
Since v
(k+1)
l
, l = 1,...,m can be calculated independently from each other, it
is possible to distribute their computation to diﬀerent processors of a parallel
machine.
(d) AOS schemes with α =1 have been presented in [424, 430] as eﬃcient dis
cretizations of the isotropic nonlinear diﬀusion ﬁlter of Catt´e et al. [81]. In
Section 1.3.2 we have seen that this ﬁlter is based on the PDE
∂
t
u = div (g([∇u
σ
[
2
) ∇u).
In this case, a natural operator splitting A =
¸
m
l=1
A
l
results from a decom
position of the divergence expression into onedimensional terms of type
∂
x
l
(g([∇u
σ
[
2
) ∂
x
l
u) (l = 1, ..., m).
This separation is very eﬃcient: There exist permutation matrices P
l
(pixel
orderings) such that P
l
A
l
P
T
l
is block diagonal and each block is diagonally
dominant and tridiagonal. Hence, the corresponding linear systems can be
4.4. RELATION TO SEMIDISCRETE MODELS 111
solved in linear eﬀort by means of a simple Gaussian algorithm. The resulting
forward substitution and backward elimination can be regarded as a recursive
ﬁlter.
A parallel implementation assigning these tridiagonal subsystems to diﬀerent
processors is described in [431]. The denoising of a medical 3D ultrasound
data set with 138 208 138 voxels on an SGI Power Challenge XL with
eight 195 MHz R10000 processors was possible in less than 1 minute.
(e) The idea to base AOS schemes on decompositions into onedimensional op
erators can also be generalized to anisotropic diﬀusion ﬁlters: Consider for
instance the discretization on a (3 3)stencil at the end of Section 3.4.2. If
it fulﬁls (S1)–(S5), then a splitting of such a 2D ﬁlter into 4 onedimensional
diﬀusion processes acting along the 4 stencil directions satisﬁes all prerequi
sites of Theorem 9.
112 CHAPTER 4. DISCRETE DIFFUSION FILTERING
Chapter 5
Examples and applications
The scalespace theory from Chapters 2–4 also covers methods such as linear or
nonlinear isotropic diﬀusion ﬁltering, for which many interesting applications have
already been mentioned in Chapter 1. Therefore, the goal of the present chapter is
to show that a generalization to anisotropic models with diﬀusion tensors depend
ing on the structure tensor oﬀers novel properties and application ﬁelds. Thus, we
focus mainly on these anisotropic techniques and juxtapose the results to other
methods. In order to demonstrate the ﬂexibility of anisotropic diﬀusion ﬁltering,
we shall pursue two diﬀerent objectives:
• smoothing with simultaneous edgeenhancement,
• smoothing with enhancement of coherent ﬂowlike textures.
All calculations for diﬀusion ﬁltering are performed using semiimplicit FD
schemes with time steps ∆t ∈ [2, 5]. In order to compare anisotropic diﬀusion to
other methods, morphological scalespaces and modiﬁcations of them have been
discretized as well. For MCM and AMSS this is achieved by means of explicit FD
schemes (cf. 1.6.4) with ∆t := 0.1 and ∆t := 0.01, respectively. Dilation with a ﬂat
structuring element is approximated by an OsherSethian scheme of type (1.88)
with ∆t := 0.5, and dilation with a quadratic structuring function is performed
in a noniterative way using van den Boomgaard’s algorithm from [51]. On an
HP 9000/889 workstation it takes less than 0.3 CPU seconds to calculate one
nonlinear diﬀusion step for a 256 256 image, and MCM, AMSS or dilation with
a disc require approximately 0.06 seconds per iteration. Typical dilations with a
quadratic structuring function take less than 0.3 seconds.
113
114 CHAPTER 5. EXAMPLES AND APPLICATIONS
5.1 Edgeenhancing diﬀusion
5.1.1 Filter design
In accordance with the notations in 2.2, let µ
1
, µ
2
with µ
1
≥ µ
2
be the eigenvalues
of the structure tensor J
ρ
, and v
1
, v
2
the corresponding orthonormal eigenvectors.
Since the diﬀusion tensor should reﬂect the local image structure it ought to be
chosen in such a way that it reveals the same set of eigenvectors v
1
, v
2
as J
ρ
. The
choice of the corresponding eigenvalues λ
1
, λ
2
depends on the desired goal of the
ﬁlter.
If one wants to smooth preferably within each region and aims to inhibit dif
fusion across edges, then one can reduce the diﬀusivity λ
1
perpendicular to edges
the more the higher the contrast µ
1
is, see 1.3.3 and [415]. This behaviour may be
accomplished by the following choice (m ∈ IN, C
m
> 0, λ > 0):
λ
1
(µ
1
) := g(µ
1
), (5.1)
λ
2
:= 1 (5.2)
with
g(s) :=
1 (s ≤ 0)
1 −exp
−Cm
(s/λ)
m
(s > 0).
(5.3)
This exponentially decreasing function is chosen in order to fulﬁl the smoothness
requirement stated in (P
c
), cf. 2.3. Since ∇u
σ
remains bounded on Ω[0, ∞) and
µ
1
= [∇u
σ
[
2
, we know that the uniform positive deﬁniteness of D is automatically
satisﬁed by this ﬁlter.
The constant C
m
is calculated in such a way that the ﬂux Φ(s) = sg(s) is
increasing for s ∈ [0, λ] and decreasing for s ∈ (λ, ∞). Thus, the preceding ﬁlter
strategy can be regarded as an anisotropic regularization of the Perona–Malik
model.
The choice m := 4 (which implies C
4
= 3.31488) gives visually good results
and is used exclusively in the examples below. Since in this section we are only
interested in edgeenhancing diﬀusion we may set the integration scale ρ of the
structure tensor equal to 0. Applications which require nonvanishing integration
scales shall be studied in section 5.2.
5.1.2 Applications
Figure 5.1 illustrates that anisotropic diﬀusion ﬁltering is still capable of possessing
the contrastenhancing properties of the Perona–Malik ﬁlter (provided the regu
larization parameter σ is not too large). It depicts the temporal evolution of a
5.1. EDGEENHANCING DIFFUSION 115
Figure 5.1: Anisotropic diﬀusion ﬁltering of a Gaussiantype function,
Ω = (0, 256)
2
, λ = 3.6, σ = 2. From top left to bottom right: t = 0,
125, 625, 3125, 15625, 78125.
116 CHAPTER 5. EXAMPLES AND APPLICATIONS
Gaussianlike function and its isolines.
1
It can be observed that two regions with
almost constant grey value evolve which are separated by a fairly steep edge. Edge
enhancement is caused by the fact that, due to the rapidly decreasing diﬀusivity,
smoothing within each region is strongly preferred to diﬀusion between the two
adjacent regions. The edge location remains stable over a very long time interval.
This indicates that, in practice, the determination of a suitable stopping time is
not a critical problem. After the process of contrast enhancement is concluded, the
steepness of edges decreases very slowly until the gradient reaches a value where no
backward diﬀusion is possible anymore. Then the image converges quickly towards
a constant image.
Let us now compare the denoising properties of diﬀerent diﬀusion ﬁlters. Figure
5.2(a) consists of a triangle and a rectangle with 70 % of all pixels being completely
degraded by noise. This image is taken from the software package MegaWave.
Test images of this type have been used to study the behaviour of ﬁlters such
as [13, 15, 16, 99, 102]. In Fig. 5.2(b) we observe that linear diﬀusion ﬁltering
is capable of removing all noise, but we have to pay a price: the image becomes
completely blurred. Besides the fact that edges get smoothed so that they are
harder to identify, the correspondence problem appears: edges become dislocated.
Thus, once they are identiﬁed at a coarse scale, they have to be traced back in
order to ﬁnd their true location, a theoretically and practically rather diﬃcult
problem.
Fig. 5.2(c) shows the eﬀect when applying the isotropic nonlinear diﬀusion
equation [81]
∂
t
u = div (g([∇u
σ
[
2
)∇u) (5.4)
with g as in (5.3). Since edges are hardly aﬀected by this process, nonlinear
isotropic diﬀusion does not lead to correspondence problems which are charac
teristic for linear ﬁltering. On the other hand, the drastically reduced diﬀusivity
at edges is also responsible for the drawback that noise at edges is preserved.
Figure 5.2(d) demonstrates that nonlinear anisotropic ﬁltering shares the ad
vantages of both methods. It combines the good noise eliminating properties of
linear diﬀusion with the stable edge structure of nonlinear isotropic ﬁltering. Due
to the permitted smoothing along edges, however, corners get more rounded than
in the nonlinear isotropic case.
The scalespace behaviour of diﬀerent PDEbased methods is juxtaposed in
Figures 5.3–5.6, where an MRI slice of a human head is processed [414, 423].
1
Except for Figs. 5.1, 5.3–5.6, where contrast enhancement is to be demonstrated, all images
in the present work are depicted in such a way that the lowest value is black and the highest one
appears white. They reveal a range within the interval [0, 255] and all pixels have unit length in
both directions.
5.1. EDGEENHANCING DIFFUSION 117
Figure 5.2: Restoration properties of diﬀusion ﬁlters. (a) Top Left:
Test image, Ω = (0, 128)
2
. (b) Top Right: Linear diﬀusion, t = 80.
(c) Bottom Left: Nonlinear isotropic diﬀusion, λ = 3.5, σ = 3,
t = 80. (d) Bottom Right: Nonlinear anisotropic diﬀusion, λ = 3.5,
σ = 3, t = 80.
Again we observe that linear diﬀusion (Fig. 5.3(a)) does not only blur all struc
tures in an equal amount but also dislocates them more and more with increasing
scale.
A ﬁrst step to reduce these problems is to adapt the diﬀusivity to the gradient
of the initial image f [147]. Fig. 5.3(b) shows the evolution under
∂
t
u = div (g([∇f[
2
) ∇u), (5.5)
118 CHAPTER 5. EXAMPLES AND APPLICATIONS
where a diﬀusivity of type [88]
g([∇f[
2
) :=
1
1 +[∇f[
2
/λ
2
(λ > 0). (5.6)
is used. Compared with homogeneous linear diﬀusion, edges remain better localized
and their blurring is reduced. On the other hand, for large t the ﬁltered image
reveals some artifacts which reﬂect the diﬀerential structure of the initial image.
A natural idea to reduce the artifacts of inhomogeneous linear diﬀusion ﬁltering
would be to introduce a feedback in the process by adapting the diﬀusivity g to
the gradient of the actual image u(x, t) instead of the original image f(x). This
leads to the nonlinear diﬀusion equation [326]
∂
t
u = div (g([∇u[
2
) ∇u). (5.7)
Figure 5.3(c) shows how such a nonlinear feedback is useful to increase the edge
localization in a signiﬁcant way: Structures remain welllocalized as long as they
can be recognized. Also blurring at edges is reduced very much. The absolute
contrast at edges, however, becomes smaller.
The latter problem can be avoided using a diﬀusivity which decreases faster
than (5.6) and leads to a nonmonotone ﬂux function. This is illustrated in Figure
5.4(a) where the regularized isotropic nonlinear diﬀusion ﬁlter (5.4) with the diﬀu
sivity (5.3) is applied. At the chin we observe that this equation is indeed capable
of enhancing edges. All structures are extremely welllocalized and the results are
segmentationlike. On the other hand, also small structures exist over a long range
of scales if they diﬀer from their vicinity by a suﬃciently large contrast. One can
try to make this ﬁlter faster and more insensitive to smallsize structures by in
creasing the regularizing Gaussian kernel size σ (cf. Fig. 5.4(b)), but this also leads
to stronger blurring of large structures, and it is no longer possible to enhance the
contour of the entire head.
Anisotropic nonlinear diﬀusion (Fig. 5.4(c)) permits diﬀusion along edges and
inhibits smoothing across them. As in Figure 5.2(d), this causes a stronger round
ing of structures, which can be seen at the nose. A positive consequence of this
slight shrinking eﬀect is the fact that small or elongated and thin structures are
better eliminated than in the isotropic case. Thus, we recognize that most of the
depicted “segments” coincide with semantically correct objects that one would ex
pect at these scales. Finally the image turns into a silhouette of the head, before
it converges to a constant image.
The tendency to produce piecewise almost constant regions indicates that dif
fusion scalespaces with nonmonotone ﬂux are ideal preprocessing tools for seg
mentation. Unlike diﬀusion–reaction models aiming to yield one segmentationlike
result for t → ∞ (cf. 1.4), the temporal evolution of these models generates a
5.1. EDGEENHANCING DIFFUSION 119
complete hierarchical family of segmentationlike images. The contrastenhancing
quality distinguishes nonlinear diﬀusion ﬁlters from most other scalespaces. It
should be noted that contrast enhancement is a local phenomenon which cannot
be replaced by simple global rescalings of the grey value range. Therefore, it is
generally not possible to obtain similar segmentationlike results by just rescaling
the grey values from a scalespace which is only contrastreducing.
The contrast and noise parameters λ and σ give the user the liberty to adapt
nonlinear diﬀusion scalespaces to the desired purpose in order to reward inter
esting features with a longer lifetime. Suitable values for them should result in a
natural way from the speciﬁc problem. In this sense, the time t is rather a para
meter of importance, with respect to the speciﬁed task, than a descriptor of spatial
scale. The traditional opinion that the evolution parameter t of scalespaces should
be related to the spatial scale reﬂects the assumption that a scalespace analysis
should be uncommitted. Nonlinear diﬀusion ﬁltering renounces this requirement by
allowing to incorporate apriori information (e.g. about the contrast of semantically
important structures) into the evolution process. The basic idea of scalespaces,
however, is maintained: to provide a family of subsequently simpliﬁed versions of
the original image, which gives a hierarchy of structures and allows to extract the
relevant information from a certain scale.
Besides these speciﬁc features of nonlinear diﬀusion scalespaces it should be
mentioned that, due to the homogeneous Neumann boundary condition and the
divergence form, both linear and nonlinear diﬀusion ﬁlters preserve the average
grey level of the image.
This is not true for the morphological ﬁlters and their modiﬁcations which are
depicted in Fig. 5.5 and 5.6.
Figure 5.5(a) and (b) show the result under continuousscale dilation with a ﬂat
discshaped structuring element and a quadratic structuring function, respectively.
From Section 1.5.3 and 1.5.6 we know that their evolution equations are given by
∂
t
u = [∇u[, (5.8)
for the disc, and by
∂
t
u = [∇u[
2
(5.9)
for the quadratic structuring function. In both cases the number of local maxima
is decreasing, and maxima keep their location in scalespace until they disappear
[206, 207]. The fact that the maximum with the largest grey value will dominate
at the end shows that these processes can be sensitive to noise (maxima might
be caused by noise), and that they usually do not preserve the average grey level.
It is not very diﬃcult to guess the shape of the structuring function from the
scalespace evolution.
120 CHAPTER 5. EXAMPLES AND APPLICATIONS
Figure 5.3: Evolution of an MRI slice under diﬀerent PDEs. Top: Original im
age, Ω = (0, 236)
2
. (a) Left Column: Linear diﬀusion, top to bottom: t = 0,
12.5, 50, 200. (b) Middle Column: Inhomogeneous linear diﬀusion (λ = 8),
t = 0, 70, 200, 600. (c) Right Column: Nonlinear isotropic diﬀusion with the
Charbonnier diﬀusivity (λ = 3), t = 0, 70, 150, 400.
5.1. EDGEENHANCING DIFFUSION 121
Figure 5.4: Evolution of an MRI slice under diﬀerent PDEs. Top: Original image,
Ω = (0, 236)
2
. (a) Left Column: Isotropic nonlinear diﬀusion (λ = 3, σ =
1), t = 0, 25000, 500000, 7000000. (b) Middle Column: Isotropic nonlinear
diﬀusion (λ = 3, σ = 4), t = 0, 40, 400, 1500. (c) Right Column: Edge
enhancing anisotropic diﬀusion (λ = 3, σ = 1), t = 0, 250, 875, 3000.
122 CHAPTER 5. EXAMPLES AND APPLICATIONS
Figure 5.5: Evolution of an MRI slice under diﬀerent PDEs. Top: Original image,
Ω = (0, 236)
2
. (a) Left Column: Dilation with a disc, t = 0, 4, 10, 20. (b)
Middle Column: Dilation with a quadratic structuring function, t = 0, 0.25,
1, 4. (c) Right Column: Mean curvature motion, t = 0, 70, 275, 1275.
5.1. EDGEENHANCING DIFFUSION 123
Figure 5.6: Evolution of an MRI slice under diﬀerent PDEs. Top: Original image,
Ω = (0, 236)
2
. (a) Left Column: Aﬃne morphological scalespace, t = 0, 20,
50, 140. (b) Middle Column: Modiﬁed mean curvature motion (λ = 3, σ = 1),
t = 0, 100, 350, 1500. (c) Right Column: Selfsnakes (λ = 3, σ = 1), t = 0,
600, 5000, 40000.
124 CHAPTER 5. EXAMPLES AND APPLICATIONS
A completely diﬀerent morphological evolution is given by the mean curvature
motion (1.90) depicted in Fig. 5.5(c). Since MCM shrinks level lines with a ve
locity that is proportional to their curvature, lowcurved object boundaries are
less aﬀected by this process, while highcurved structures (e.g. the nose) exhibit
roundings at an earlier stage. This also explains its excellent noise elimination
qualities. After some time, however, the head looks almost like a ball. This is in
accordance with the theory which predicts convergence of all closed level lines to
circular points.
A similar behaviour can be observed for the aﬃne invariant morphological scale
space (1.108) shown in Fig. 5.6(a). Since it takes the time T =
3
4
s
4
3
to remove
all isolines within a circle of radius s – in contrast to T =
1
2
s
2
for MCM – we
see that, for a comparable elimination of small structure, the shrinking eﬀect of
large structures is stronger for AMSS than for MCM. Thus, the correspondence
problem is more severe than for MCM. Nevertheless, the advantage of having aﬃne
invariance may counterbalance the correspondence problem in certain applications.
Since the AMSS involves no additional parameters and oﬀers more invariances
than other scalespaces, it is ideal for uncommitted image analysis and shape
recognition. Both MCM and AMSS give signiﬁcantly sharper edges than linear
diﬀusion ﬁltering, but they are not designed to act contrastenhancing.
One possibility to reduce the correspondence problem of morphological scale
spaces is to attenuate the curve evolution at highcontrast edges. This is at the
expense of withdrawing morphology in terms of invariance under monotone grey
scale transformations.
One possibility is to use the damping function outside the divergence expression.
Processes of this type are studied in [13, 364, 365, 304]. As a simple prototype for
this idea, let us investigate the modiﬁed MCM
∂
t
u = g([∇u
σ
[
2
) [∇u[ div
∇u
[∇u[
(5.10)
with g([∇u
σ
[
2
) as in (1.32). The corresponding evolution is depicted in Fig. 5.6(b).
We observe that structures remain much better localized than in the original MCM.
On the other hand, the experiments give evidence that this process is probably
not contrastenhancing, see e.g. the chin. As a consequence, the results appear less
segmentationlike than those for nonlinear diﬀusion ﬁltering.
Using g([∇u
σ
[
2
) inside the divergence expression leads to
∂
t
u = [∇u[ div
g([∇u
σ
[
2
)
∇u
[∇u[
. (5.11)
In Section 1.6.6 we have seen that processes of this type are called selfsnakes
[357]. Since they diﬀer from isotropic nonlinear diﬀusion ﬁlters by the [∇u[ terms
inside and outside the divergence expression, they will not preserve the average
5.1. EDGEENHANCING DIFFUSION 125
Figure 5.7: Preprocessing of a fabric image. (a) Left: Fabric, Ω =
(0, 257)
2
. (b) Right: Anisotropic diﬀusion, λ = 4, σ = 2, t = 240.
grey value. The evolution in Fig. 5.6(c) indicates that g([∇u
σ
[
2
) gives similar edge
enhancing eﬀects as in a nonlinear diﬀusion ﬁlter, but one can observe a stronger
tendency to create circular structures. This behaviour which resembles MCM is
not surprising if one compares (1.120) with (1.121).
Let us now study two applications of nonlinear diﬀusion ﬁltering in computer
aided quality control (CAQ): the grading of fabrics and wood surfaces (see also
[413]).
The quality of a fabric is determined by two criteria, namely clouds and stripes.
Clouds result from isotropic inhomogeneities of the density distribution, whereas
stripes are an anisotropic phenomenon caused by adjacent ﬁbres pointing in the
same direction. Anisotropic diﬀusion ﬁlters are capable of visualizing both quality
relevant features simultaneously (Fig. 5.7). For a suitable parameter choice, they
perform isotropic smoothing at clouds and diﬀuse in an anisotropic way along ﬁ
bres in order to enhance them. However, if one wants to visualize both features
separately, one can use a fast pyramid algorithm based on linear diﬀusion ﬁlter
ing for the clouds [417], whereas stripes can be enhanced by a special nonlinear
diﬀusion ﬁlter which is designed for closing interrupted lines and which shall be
discussed in Section 5.2.
For furniture production it is of importance to classify the quality of wood
surfaces. If one aims to automize this evaluation, one has to process the image in
such a way that quality relevant features become better visible und unimportant
structures disappear. Fig. 5.8(a) depicts a wood surface possessing one defect. To
visualize this defect, equation (5.4) can be applied with good success (Fig. 5.8(b)).
126 CHAPTER 5. EXAMPLES AND APPLICATIONS
Figure 5.8: Defect detection in wood. (a) Left: Wood surface, Ω =
(0, 256)
2
. (b) Right: Isotropic nonlinear diﬀusion, λ = 4, σ = 2,
t = 2000.
In [413] it is demonstrated how a modiﬁed anisotropic diﬀusion process yields even
more accurate results with less roundings at the corners.
Fig. 5.9(a) gives an example for possible medical applications of nonlinear dif
fusion ﬁltering as a preprocessing tool for segmentation (see also [415] for another
example). It depicts an MRI slice of the human head. For detecting Alzheimer’s
disease one is interested in determining the ratio between the ventricle areas, which
are given by the two white longitudinal objects in the centre, and the entire head
area.
In order to make the diagnosis more objective and reliable, it is intended to
automize this feature extraction step by a segmentation algorithm. Figure 5.9(c)
shows a segmentation according to the following simpliﬁcation of the Mumford–
Shah functional (1.58):
E
f
(u, K) =
Ω
(u−f)
2
dx + α[K[. (5.12)
It has been obtained by a MegaWave programme using a hierarchical region grow
ing algorithm due to Koepﬂer et al. [239]. As is seen in Fig. 5.9(d), one gets a better
segmentation when processing the original image slightly by means of nonlinear
diﬀusion ﬁltering (Fig. 5.9(b)) prior to segmenting it.
5.2. COHERENCEENHANCING DIFFUSION 127
Figure 5.9: Preprocessing of an MRI slice. (a) Top Left: Head, Ω =
(0, 256)
2
. (b) Top Right: Diﬀusionﬁltered, λ = 5, σ = 0.1, t =
2.5. (c) Bottom Left: Segmented original image, α = 8192. (d)
Bottom Right: Segmented ﬁltered image, α = 8192.
5.2 Coherenceenhancing diﬀusion
5.2.1 Filter design
In this section we shall investigate how the structure tensor information can be
used to design anisotropic diﬀusion scalespaces which enhance the coherence of
ﬂowlike textures [418]. This requires a nonvanishing integration scale ρ.
Let again µ
1
, µ
2
with µ
1
≥ µ
2
be the eigenvalues of J
ρ
, and v
1
, v
2
the cor
responding orthonormal eigenvectors. As in 5.1 the diﬀusion tensor D(J
ρ
(∇u
σ
))
ought to possess the same set of eigenvectors as J
ρ
(∇u
σ
).
If one wants to enhance coherent structures, one should smooth preferably
128 CHAPTER 5. EXAMPLES AND APPLICATIONS
Figure 5.10: Local orientation in a ﬁngerprint image. (a) Top Left:
Original ﬁngerprint, Ω = (0, 256)
2
. (b) Top Right: Orientation
of smoothed gradient, σ = 0.5. (c) Bottom Left: Orientation of
smoothed gradient, σ = 5. (d) Bottom Right: Structure tensor
orientation, σ = 0.5, ρ = 4.
along the coherence direction v
2
with a diﬀusivity λ
2
which increases with respect
to the coherence (µ
1
−µ
2
)
2
. This may be achieved by the following choice for the
eigenvalues of the diﬀusion tensor (C > 0, m ∈ IN):
λ
1
:= α,
λ
2
:=
α if µ
1
=µ
2
,
α + (1−α) exp
−C
(µ
1
−µ
2
)
2m
else,
where the exponential function was chosen to ensure the smoothness of D and the
5.2. COHERENCEENHANCING DIFFUSION 129
Figure 5.11: Anisotropic equations applied to the ﬁngerprint image.
(a) Left: Meancurvature motion, t = 5. (b) Right: Coherence
enhancing anisotropic diﬀusion, σ = 0.5, ρ = 4, t = 20.
small positive parameter α∈(0, 1) keeps D(J
ρ
(∇u
σ
)) uniformly positive deﬁnite.
2
All examples below are calculated using C := 1, m := 1, and α := 0.001.
5.2.2 Applications
Figure 5.10 illustrates the advantages of local orientation analysis by means of
the structure tensor. In order to detect the local orientation of the ﬁngerprint
depicted in Fig. 5.10(a), the gradient orientation of a slightly smoothed image has
been calculated (Fig. 5.10(b)). Horizontally oriented structures appear black, while
vertical structures are represented in white. We observe very high ﬂuctuations in
the local orientation. When applying a larger smoothing kernel it is clear that
adjacent gradients having the same orientation but opposite direction cancel out.
Therefore, the results in (c) are much worse than in (b). The structure tensor,
however, averages the gradient orientation instead of its direction. This is the
reason for the reliable estimates of local orientation that can be obtained with this
method (Fig. 5.10(d)).
To illustrate how the result of anisotropic PDE methods depends on the direc
tion in which they smooth, let us recall the example of mean curvature motion (cf.
1.6.1):
∂
t
u = u
ξξ
= [∇u[ curv(u) (5.13)
with ξ being the direction perpendicular to ∇u. Since MCM smoothes by prop
agating level lines in inner normal direction we recognize that its smoothing di
2
Evidently, ﬁlters of this type are not regularizations of the Perona–Malik process: the limit
σ →0, ρ →0 leads to a linear diﬀusion process with constant diﬀusivity α.
130 CHAPTER 5. EXAMPLES AND APPLICATIONS
Figure 5.12: Scalespace behaviour of coherenceenhancing diﬀusion
(σ = 0.5, ρ = 2). (a) Top Left: Original fabric image, Ω = (0, 257)
2
.
(b) Top Right: t = 20. (c) Bottom Left: t = 120. (d) Bottom
Right: t = 640.
rection depends exclusively on ∇u. Thus, although this method is in a complete
anisotropic spirit, we should not expect it to be capable of closing interrupted
linelike structures. The results in Fig. 5.11(a) conﬁrm this impression.
The proposed anisotropic diﬀusion ﬁlter, however, biases the diﬀusive ﬂux to
wards the coherence orientation v
2
and is therefore wellsuited for closing inter
rupted lines in coherent ﬂowlike textures, see Fig. 5.11(b). Due to its reduced
diﬀusivity at noncoherent structures, the locations of the semantically important
singularities in the ﬁngerprint remain the same. This is an important prerequisite
that any image processing method has to satisfy if it is to be applied to ﬁngerprint
analysis.
5.2. COHERENCEENHANCING DIFFUSION 131
Figure 5.13: (a) Top: High resolution slipring CT scan of a femural bone,
Ω = (0, 300) (0, 186). (b) Bottom Left: Filtered by coherenceenhancing
anisotropic diﬀusion, σ = 0.5, ρ = 6, t = 16. (c) Bottom Right: Dito with
t =128.
Figure 5.12 depicts the scalespace behaviour of coherenceenhancing aniso
tropic diﬀusion applied to the fabric image from Fig. 5.7. The temporal behaviour
of this diﬀusion ﬁlter seems to be appropriate for visualizing coherent ﬁbre agglom
erations (stripes) at diﬀerent scales, a diﬃcult problem for the automatic grading
of nonwovens [299].
Figure 5.13 illustrates the potential of CED for medical applications. It depicts
a human bone. Its internal structure has a distinctive texture through the presence
of tiny elongated bony structural elements, the trabeculae. There is evidence that
the trabecular formation is for a great deal determined by the external load. For
this reason the trabecular structure constitutes an important clinical parameter in
orthopedics. Examples are the control of recovery after surgical procedures, such
as the placement or removal of metal implants, quantifying the rate of progression
of rheumatism and osteoporosis, the determination of leftright deviations of sym
metry in the load or establishing optimal load corrections by physiotherapy. From
Figure 5.13(b),(c) we observe that CED is capable of enhancing the trabecular
structures in order to ease their subsequent orientation analysis.
132 CHAPTER 5. EXAMPLES AND APPLICATIONS
Figure 5.14: Image restoration using coherenceenhancing anisotropic
diﬀusion. (a) Left: “Selfportrait” by van Gogh (SaintR´emy, 1889;
Paris, Muse´e d’Orsay), Ω = (0, 215) (0, 275). (b) Right: Filtered,
σ = 0.5, ρ = 4, t = 6.
Let us now investigate the impact of coherenceenhancing diﬀusion on images,
which are not typical texture images, but still reveal a ﬂowlike character. To this
end, we shall process impressionistic paintings by Vincent van Gogh.
Fig. 5.14 shows the restoration properties of coherenceenhancing anisotropic
diﬀusion when being applied to a selfportrait of the artist [161]. We observe that
the diﬀusion ﬁlter can close interrupted lines and enhance the ﬂowlike character
which is typical for van Gogh paintings.
The next painting we are concerned with is called “Road with Cypress and
Star” [162, 429]. It is depicted in Fig. 5.15. In order to demonstrate the inﬂuence
of the integration scale ρ, all ﬁlter parameters are ﬁxed except for ρ. Fig. 5.15(b)
shows that a value for ρ which is too small does not lead to the visually dominant
coherence orientation and creates structures with a lot of undesired ﬂuctuations.
Increasing the value for ρ improves the image signiﬁcantly (Fig. 5.15(c)). Interest
ingly, a further increasing of ρ does hardly alter this result (Fig. 5.15(d)), which
indicates that this van Gogh painting possesses a uniform “texture scale” reﬂecting
the characteristic painting style of the artist.
In a last example the temporal evolution of ﬂowlike images is illustrated by
virtue of the “Starry Night” painting in Fig. 5.16 [160, 419]. Due to the established
5.2. COHERENCEENHANCING DIFFUSION 133
Figure 5.15: Impact of the integration scale on coherenceenhancing
anisotropic diﬀusion (σ = 0.5, t = 8). (a) Top Left: “Road with
Cypress and Star” by van Gogh (AuverssurOise, 1890; Otterlo, Ri
jksmuseum Kr¨oller–M¨ uller), Ω = (0, 203) (0, 290). (b) Top Right:
Filtered with ρ = 1. (c) Bottom Left: ρ = 4. (d) Bottom Right:
ρ = 6.
134 CHAPTER 5. EXAMPLES AND APPLICATIONS
Figure 5.16: Scalespace properties of coherenceenhancing anisotropic diﬀusion
(σ = 0.5, ρ = 4). (a) Top Left: “Starry Night” by van Gogh (SaintR´emy,
1889; New York, The Museum of Modern Art), Ω = (0, 255) (0, 199). (b) Top
Right: t = 8. (c) Bottom Left: t = 64. (d) Bottom Right: t = 512.
scalespace properties, the image becomes gradually simpler in many aspects, be
fore it ﬁnally will tend to its simplest representation, a constant image with the
same average grey value as the original one. The ﬂowlike character, however, is
maintained for a very long time.
3
3
Results for AMSS ﬁltering of this image can be found in [305].
Chapter 6
Conclusions and perspectives
While Chapter 1 has given a general overview of PDEbased smoothing and restora
tion methods, the goal of Chapters 2–5 has been to present a scalespace framework
for nonlinear diﬀusion ﬁltering which does not require any monotony assump
tion (comparison principle). We have seen that, besides the fact that many global
smoothing scalespace properties are maintained, new possibilities with respect to
image restoration appear.
Rather than deducing a unique equation from ﬁrst principles we have ana
lysed wellposedness and scalespace properties of a general family of regularized
anisotropic diﬀusion ﬁlters. Existence and uniqueness results, continuous depen
dence of the solution on the initial image, maximum–minimum principles, invari
ances, Lyapunov functionals, and convergence to a constant steadystate have been
established.
The large class of Lyapunov functionals permits to regard these ﬁlters in many
ways as simplifying, informationreducing transformations. These global smooth
ing properties do not contradict seemingly opposite local eﬀects such as edge en
hancement. For this reason it is possible to design scalespaces with restoration
properties giving segmentationlike results.
Prerequisites have been stated under which one can prove wellposedness and
scalespace results in the continuous, semidiscrete and discrete setting. Each of
these frameworks is selfcontained and does not require the others. On the other
hand, the prerequisites in all three settings reveal many similarities and, as a
consequence, representatives of the semidiscrete class can be obtained by suitable
spatial discretizations of the continuous class, while representatives of the discrete
class may arise from time discretizations of semidiscrete ﬁlters.
The degree of freedom within the proposed class of ﬁlters can be used to tailor
the ﬁlters towards speciﬁc restoration tasks. Therefore, these scalespaces do not
need to be uncommitted; they give the user the liberty to incorporate apriori
knowledge, for instance concerning size and contrast of especially interesting fea
tures.
135
136 CHAPTER 6. CONCLUSIONS AND PERSPECTIVES
The analysed class comprises linear diﬀusion ﬁltering and the nonlinear iso
tropic model of Catt´e et al. [81] and Whitaker and Pizer [438], but also novel ap
proaches have been proposed: The use of diﬀusion tensors instead of scalarvalued
diﬀusivities puts us in a position to design real anisotropic diﬀusion processes
which may reveal advantages at noisy edges. Last but not least, the fact that these
ﬁlters are steered by the structure tensor instead of the regularized gradient allows
to adapt them to more sophisticated tasks such as the enhancement of coherent
ﬂowlike structures.
In view of these results, anisotropic diﬀusion deserves to be regarded as much
more than an adhoc strategy for transforming a degraded image into a more
pleasant looking one. It is a ﬂexible and mathematically sound class of methods
which ties the advantages of two worlds: scalespace analysis and image restoration.
It is clear, however, that nonlinear diﬀusion ﬁltering is a young ﬁeld which has
certainly not reached its ﬁnal state yet. Thus, we can expect a lot of new results
in the near future. Some of its future developments, however, are likely to consist
of straightforward extensions of topics presented in this text:
• While the theory and the examples in the present book focus on 2D grey
scale images, it is evident that most of its results can easily be generalized to
higher dimensions and vectorvalued images. The need for such extensions
grows with the rapid progress in the development of faster computers, the
general availability of aﬀordable colour scanners and printers, and the wish
to integrate information from diﬀerent channels. Some of the references in
Chapter 1 point in directions how this can be accomplished.
• The various possibilities to include semilocal or global information constitute
another future perspective. This could lead to speciﬁcally tuned ﬁlters for
topics such as perceptual grouping. Coherenceenhancing anisotropic diﬀu
sion is only a ﬁrst step in this direction. New ﬁlter models might arise using
other structure descriptors than the (regularized) gradient or the structure
tensor. Interesting candidates could be wavelets, Gabor ﬁlters, or steerable
ﬁlters.
• In contrast to the linear diﬀusion case, the relation between structures at
diﬀerent scales has rarely been exploited in the nonlinear context. Although
this problem is less severe, since the avoidance of correspondence problems
was one of the key motivations to study nonlinear scalespaces, it would cer
tainly be useful to better understand the deep structure in nonlinear diﬀusion
processes. The scalespace stack of these ﬁlters appears to be wellsuited to
extract semantically important information with respect to a speciﬁed task.
This ﬁeld oﬀers a lot of challenging mathematical questions.
137
• Most people working in computer vision do not have a speciﬁc knowledge
on numerical methods for PDEs. As a consequence, the most widelyused
numerical methods for nonlinear diﬀusion ﬁltering are still the simple, but
ineﬃcient explicit (Euler forward) schemes. Novel, more timecritical appli
cation areas could be explored by applying implicit schemes, splitting and
multigrid techniques, or grid adaptation strategies. In this context it would be
helpful to have software packages, where diﬀerent nonlinear diﬀusion ﬁlters
are implemented in an eﬃcient way, and which are easy to use for everyone.
• The price one has to pay for the ﬂexibility of nonlinear diﬀusion ﬁltering is
the speciﬁcation of some parameters. Since these parameters have a rather
natural meaning, this is not a very diﬃcult problem for someone with ex
perience in computer vision. Somebody with another primary interest, for
instance a physician who wants to denoise ultrasound images, may be fright
ened by this perspective. Thus, more research on ﬁnding some guidelines
for automatic parameter determination for a task at hand would encourage
also people without a speciﬁc image processing background to apply nonlin
ear diﬀusion ﬁlters. Several useful suggestions for parameter adaptation can
already be found in [328, 36, 431, 270, 444].
• There are not yet many studies which explore the potential of nonlinear dif
fusion ﬁltering when being combined with other image processing techniques.
Especially combinations with concepts such as data compression, segmenta
tion algorithms, tomographic reconstruction techniques, or neural networks
for learning apriori information might lead to novel application areas for
these techniques.
Thus, there still remains a lot of work to be done. It would be nice if this book has
inspired its readers to contribute to the solution of some of the remaining open
problems.
138 CHAPTER 6. CONCLUSIONS AND PERSPECTIVES
Bibliography
[1] S. Abbasi, F. Mokhtarian, J. Kittler, Reliable classiﬁcation of chrysanthemum leaves
through curvature scale space, B. ter Haar Romeny, L. Florack, J. Koenderink, M.
Viergever (Eds.), Scalespace theory in computer vision, Lecture Notes in Computer Sci
ence, Vol. 1252, Springer, Berlin, 284–295, 1997.
[2] R. Acar, C.R. Vogel Analysis of bounded variation penalty methods for illposed problems,
Inverse Problems, Vol. 10, 1217–1229, 1994.
[3] A. Ackah–Miezan, A. Gagalowicz, Discrete models for energyminimizing segmentation,
Proc. Fourth Int. Conf. on Computer Vision (ICCV ’93, Berlin, May 11–14, 1993), IEEE
Computer Society Press, Los Alamitos, 200–207, 1993.
[4] S.T. Acton, Edge enhancement of infrared imagery by way of the anisotropic diﬀusion
pyramid, Proc. IEEE Int. Conf. Image Processing (ICIP–96, Lausanne, Sept. 16–19, 1996),
Vol. 1, 865–868, 1996.
[5] S.T. Acton, A.C. Bovik, M.M. Crawford, Anisotropic diﬀusion pyramids for image seg
mentation, Proc. IEEE Int. Conf. Image Processing (ICIP–94, Austin, Nov. 13–16, 1994),
Vol. 3, IEEE Computer Society Press, Los Alamitos, 478–482, 1994.
[6] S.T. Acton, M.M. Crawford, A mean ﬁeld solution to anisotropic edge detection of remotely
sensed data, Proc. 12th Int. Geoscience and Remote Sensing Symposium (IGARSS ’92,
Houston, May 26–29, 1992), Vol. 2, 845–847, 1992.
[7] R.A. Adams, Sobolev spaces, Academic Press, New York, 1975.
[8] D. Adalsteinsson, J.A. Sethian, A fast level set method for propagating interfaces, J. Com
put. Phys., Vol. 118, 269–277, 1995.
[9] H.W. Alt, Lineare Funktionalanalysis, Springer, Berlin, 1992.
[10] L. Alvarez, Images and PDE’s, M.O. Berger, R. Deriche, I. Herlin, J. Jaﬀr´e, J.M. Morel
(Eds.), ICAOS ’96: Images, wavelets and PDEs, Lecture Notes in Control and Information
Sciences, Vol. 219, Springer, London, 3–14, 1996.
[11] L. Alvarez, J. Esclarin, Image quantization using reaction–diﬀusion equations, SIAM J.
Appl. Math., Vol. 57, 153–175, 1997.
[12] L. Alvarez, F. Guichard, P.L. Lions, J.M. Morel, Axioms and fundamental equations in
image processing, Arch. Rational Mech. Anal., Vol. 123, 199–257, 1993.
[13] L. Alvarez, P.L. Lions, J.M. Morel, Image selective smoothing and edge detection by
nonlinear diﬀusion. II, SIAM J. Numer. Anal., Vol. 29, 845–866, 1992.
[14] L. Alvarez, L. Mazorra, Signal and image restoration using shock ﬁlters and anisotropic
diﬀusion, SIAM J. Numer. Anal., Vol. 31, 590–605, 1994.
139
140 BIBLIOGRAPHY
[15] L. Alvarez, F. Morales, Aﬃne morphological multiscale analysis of corners and multiple
junctions, Report No. 9402, Departamento de Inform´atica y Sistemas, Universidad de Las
Palmas, Campus Universitario de Taﬁra, 35017 Las Palmas, Spain, 1994; to appear in Int.
J. Comput. Vision.
[16] L. Alvarez, J.M. Morel, Formalization and computational aspects of image analysis, Acta
Numerica, 1–59, 1994.
[17] L. Ambrosio, A compactness theorem for a special class of functions of bounded variations,
Boll. Un. Math. Ital., Vol. 3B, 857–881, 1990.
[18] L. Ambrosio, N. Fusco, D. Pallara, Partial regularity of free discontinuity sets II, Preprint
CVGMT96040106, Math. Dept., Scuola Normale Superiore, Piazza dei Cavalieri, 56126
Pisa, Italy, 1995; submitted to Annali della Scuola Normale Superiore di Pisa.
[19] L. Ambrosio, N. Fusco, D. Pallara, Higher regularity of solutions of free discontinuity
problems, Preprint CVGMT96040107, Math. Dept., Scuola Normale Superiore, Piazza
dei Cavalieri, 56126 Pisa, Italy, 1995.
[20] L. Ambrosio, D. Pallara, Partial regularity of free discontinuity sets I, Preprint CVGMT
96040105, Math. Dept., Scuola Normale Superiore, Piazza dei Cavalieri, 56126 Pisa, Italy,
1995; submitted to Annali della Scuola Normale Superiore di Pisa.
[21] L. Ambrosio, V. Tortorelli, On the approximation of functionals depending on jumps by
elliptic functionals, Comm. Pure Appl. Math., Vol. 43, 999–1036, 1990.
[22] L. Ambrosio, V. Tortorelli, Approximation of functionals depending on jumps by elliptic
functionals via Γconvergence, Boll. Un. Math. Ital., Vol. 7, 105–123, 1992.
[23] S. Angenent, Parabolic equations on curves and surfaces. Part I. Curves with pintegrable
curvature, Annals of Mathematics, Vol. 132, 451–483, 1990.
[24] S. Angenent, Parabolic equations on curves and surfaces. Part II. Intersections, blowup,
and generalized solutions, Annals of Mathematics, Vol. 133, 171–215, 1991.
[25] A.B. Arehart, L. Vincent, B.B. Kimia, Mathematical morphology: The Hamilton–Jacobi
connection, Proc. Fourth Int. Conf. on Computer Vision (ICCV ’93, Berlin, May 11–14,
1993), IEEE Computer Society Press, Los Alamitos, 215–219, 1993.
[26] S.R. Arridge, A. Simmons, Multispectral probabilistic diﬀusion using Bayesian classiﬁca
tion, B. ter Haar Romeny, L. Florack, J. Koenderink, M. Viergever (Eds.), Scalespace
theory in computer vision, Lecture Notes in Computer Science, Vol. 1252, Springer, Berlin,
224–235, 1997.
[27] K.
˚
Astr¨om, A. Heyden, Stochastic analysis of image acquisition and scalespace smoothing,
J. Sporring, M. Nielsen, L. Florack, P. Johansen (Eds.), Gaussian scalespace theory,
Kluwer, Dordrecht, 129–136, 1997.
[28] I. Bajla, I. Holl¨ander, Geometrydriven diﬀusion smoothing of the MR brain images using
a novel variable conductance, Proc. 18th Annual Int. Conf. of the IEEE Engineering in
Medicine and Biology Soc. (EMB ’96, Amsterdam, Oct. 31 – Nov. 3, 1996), CDROM
document N21.
[29] I. Bajla, M. Maruˇsiak, M.
ˇ
Sr´amek, Anisotropic ﬁltering of MRI data based upon image
gradient histogram, D. Chetverikov, W.G. Kropatsch (Eds.), Computer analysis of images
and patterns, Lecture Notes in Computer Science, Vol. 719, Springer, Berlin, 90–97, 1993.
[30] J. Babaud, A.P. Witkin, M. Baudin, R.O. Duda, Uniqueness of the Gaussian kernel for
scale space ﬁltering, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 8, 26–33, 1986.
BIBLIOGRAPHY 141
[31] C. Ballester Nicolau, An aﬃne invariant model for image segmentation: Mathematical
analysis and applications, Ph.D. thesis, University of Illes Baleares, Palma de Mallorca,
Spain, 1995.
[32] C. Ballester, V. Caselles, M. Gonz´alez, Aﬃne invariant segmentation by variational
method, SIAM J. Appl. Math., Vol. 56, 294–325, 1996.
[33] C. Ballester, M. Gonz´alez, Texture segmentation by variational methods, M.O. Berger, R.
Deriche, I. Herlin, J. Jaﬀr´e, J.M. Morel (Eds.), ICAOS ’96: Images, wavelets and PDEs,
Lecture Notes in Control and Information Sciences, Vol. 219, Springer, London, 187–193,
1996.
[34] E. B¨ansch, K. Mikula, A coarsening ﬁnite element strategy in image selective smoothing,
Preprint No. 18/1996, Faculty of Mathematics, University of Freiburg, Hebelstr. 27, 79104
Freiburg, Germany, 1996; to appear in Computation and Visualization in Science.
[35] G.I. Barenblatt, M. Bertsch, R. Dal Passo, M. Ughi, A degenerate pseudoparabolic reg
ularization of a nonlinear forward–backward heat equation arising in the theory of heat
and mass exchange in stably stratiﬁed turbulent shear ﬂow, SIAM J. Math. Anal., Vol. 24,
1414–1439, 1993.
[36] B. Benhamouda, Parameter adaptation for nonlinear diﬀusion in image processing, mas
ter thesis, Dept. of Mathematics, University of Kaiserslautern, P.O. Box 3049, 67653
Kaiserslautern, Germany, 1994.
[37] P.M. van den Berg, R.E. Kleinman, A total variation enhanced modiﬁed gradient algorithm
for proﬁle reconstruction, Physics Express Letters, Inverse Problems, Vol. 11, pp. L5–L10,
1995.
[38] F. Bergholm, Edge focusing, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 9, 726–741,
1987.
[39] J. Big¨ un, G.H. Granlund, Optimal orientation detection of linear symmetry, Proc. First
Int. Conf. on Computer Vision (ICCV ’87, London, June 8–11, 1987), IEEE Computer
Society Press, Washington, 433–438, 1987.
[40] J. Big¨ un, G.H. Granlund, J. Wiklund, Multidimensional orientation estimation with ap
plications to texture analysis and optical ﬂow, IEEE Trans. Pattern Anal. Mach. Intell.,
Vol. 13, 775–790, 1991.
[41] M.J. Black, G. Sapiro, D. Marimont, D. Heeger, Robust anisotropic diﬀusion: connections
between robust statistics, line processing, and anisotropic diﬀusion, B. ter Haar Romeny,
L. Florack, J. Koenderink, M. Viergever (Eds.), Scalespace theory in computer vision,
Lecture Notes in Computer Science, Vol. 1252, Springer, Berlin, 323–326, 1997.
[42] A. Blake, A. Zisserman, Visual reconstruction, MIT Press, Cambridge (Mass.), 1987.
[43] L. Blanc–F´eraud, P. Charbonnier, G. Aubert, M. Barlaud, Nonlinear image processing:
modelling and fast algorithm for regularization with edge detection, Proc. IEEE Int. Conf.
Image Processing (ICIP–95, Washington, Oct. 22–25, 1995), Vol. 1, 474–477, 1995.
[44] W. Blaschke, Vorlesungen ¨ uber Diﬀerentialgeometrie II, Springer, Berlin, 1923.
[45] P. Blomgren, T.F. Chan, Color TV: total variation methods for restoration of vector valued
images, CAM Report 965, Dept. of Mathematics, Univ. of California, Los Angeles, CA
90024, U.S.A., 1996; to appear in IEEE Trans. Image Proc.
142 BIBLIOGRAPHY
[46] P. Blomgren, T.F. Chan, P. Mulet, Extensions to total variation denoising, F.T. Luk
(Ed.), Advanced signal processing: algorithms, architectures, and implementations VII,
SPIE Vol. 3162, 1997, to appear.
[47] H. Blum, A transformation for extracting new descriptors of shape, W. Wathen–Dunn
(Ed.), Models for the perception of speech and visual form, M.I.T. Press, Cambridge,
362–380, 1967.
[48] A. Bonnet, Charact´erisation des minima globaux de la fonctionelle du Mumford–Shah en
segmentation d’images, C. R. Acad. Sci. Paris, t. 321, S´erie I, 1121–1126, 1995.
[49] A. Bonnet, Sur la r´egularit´e des bords des minima de la fonctionelle de Mumford–Shah,
C. R. Acad. Sci. Paris, t. 321, S´erie I, 1275–1279, 1995.
[50] R. van den Boomgaard, Mathematical morphology: Extensions towards computer vision,
Ph.D. thesis, University of Amsterdam, The Netherlands, 1992.
[51] R. van den Boomgaard, The morphological equivalent of the Gauss convolution, Nieuw
Archief Voor Wiskunde, Vierde Serie, Deel 10, 219–236, 1992.
[52] R. van den Boomgaard, Morphology, oral presentation, Ph.D. School on Classical Scale
Space Theory, Copenhagen, May 10–13, 1996.
[53] R. van den Boomgaard, A. Smeulders, The morphological structure of images: The dif
ferential equations of morphological scalespace, IEEE Trans. Pattern Anal. Mach. Intell.,
Vol. 16, 1101–1113, 1994.
[54] K.A. Brakke, The motion of a surface by its mean curvature, Princeton University Press,
Princeton, 1978.
[55] J.H. Bramble, B.E. Hubbard, New monotone type approximations for elliptic problems,
Math. Comp., Vol. 18, 349–367, 1964.
[56] F. Brauer, J.A. Nohel, Ordinary diﬀerential equations: A ﬁrst course, Benjamin, New
York, 1967.
[57] H. Brezis, Op´erateurs maximaux monotones et semigroupes de contractions dans les es
paces de Hilbert, North Holland, Amsterdam, 1973.
[58] H. Brezis, Analyse fonctionelle, Masson, Paris, 1992.
[59] W.L. Briggs, A multigrid tutorial, SIAM, Philadelphia, 1987.
[60] R.W. Brockett, P. Maragos, Evolution equations for continuousscale morphological ﬁlter
ing, IEEE Trans. Signal Processing, Vol. 42, 3377–3386, 1994.
[61] A.M. Bruckstein, G. Sapiro, D. Shaked, Evolution of planar polygons, Int. J. Pattern
Recogn. Artif. Intell., Vol. 9, 991–1014, 1995.
[62] A.M. Bruckstein, D. Shaked, On projective invariant smoothing and evolution of planar
curves and polygons, J. Math. Imag. Vision, Vol. 7, 225–240, 1997.
[63] B. Buck, V. Macaulay (Eds.), Maximum entropy in action, Clarendon, Oxford, 1991.
[64] P.J. Burt, Fast ﬁlter transforms for image processing, Computer Graphics and Image
Processing, Vol. 16, 20–51, 1981.
[65] M. Burton, S. Acton, Target tracking using the anisotropic diﬀusion pyramid, Proc. IEEE
Southwest Symposium on Image Analysis and Interpetation (SSIAI ’96, San Antonio,
April 8–9, 1996).
BIBLIOGRAPHY 143
[66] M.A. Butt, P. Maragos, Comparison of multiscale morphology approaches: PDE imple
mented via curve evolution versus Chamfer distance transform, P. Maragos, R.W. Schafer,
M.A. Butt (Eds.), Mathematical morphology and its applications to image and signal pro
cessing, Kluwer, Boston, 31–40, 1996.
[67] L.D. Cai, Some notes on repeated averaging smoothing, J. Kittler (Ed.), Pattern Recog
nition, Lecture Notes in Computer Science, Vol. 301, Springer, Berlin, 597–605, 1988.
[68] L.D. Cai, A “small leakage” model for diﬀusion smoothing of image data, Proc. 11th
Int. Joint Conf. on Artiﬁcial Intelligence (IJCAI ’89, Detroit, Aug. 20–25, 1989), Vol. 2,
1585–1590, 1989.
[69] J. Canny, A computational approach to edge detection, IEEE Trans. Pattern Anal. Mach.
Intell., Vol. 8, 679–698, 1986.
[70] V. Cantoni, M. Ferretti, Pyramidal architectures for computer vision, Plenum Press, New
York, 1994.
[71] M.P. do Carmo, Diﬀerentialgeometrie von Kurven und Fl¨achen, Vieweg, Braunschweig,
1983.
[72] R. Carmona, S. Zhong, Adaptive image smoothing respecting feature directions, Preprint,
Dept. of Civil Engineering, Princeton University, Princeton, NJ 08544–5263, U.S.A., 1996;
submitted to IEEE Trans. Image Proc.
[73] V. Caselles, F. Catt´e, T. Coll, F. Dibos, A geometric model for active contours in image
processing, Numer. Math., Vol. 66, 1–31, 1993.
[74] V. Caselles, B. Coll, Snakes in movement, SIAM J. Numer. Anal., Vol. 33, 2445–2456,
1996.
[75] V. Caselles, B. Coll, J.M. Morel, Scalespace versus topographic map for natural images,
B. ter Haar Romeny, L. Florack, J. Koenderink, M. Viergever (Eds.), Scalespace theory in
computer vision, Lecture Notes in Computer Science, Vol. 1252, Springer, Berlin, 29–49,
1997.
[76] V. Caselles, R. Kimmel, G. Sapiro, Geodesic active contours, Int. J. Comput. Vision., Vol.
22, 61–79, 1997.
[77] V. Caselles, R. Kimmel, G. Sapiro, C. Sbert, Minimal surfaces: a geometric three dimen
sional segmentation approach, Numer. Math., Vol. 77, 423–451, 1997.
[78] V. Caselles, C. Sbert, What is the best causal scalespace for 3D images?, SIAM J. Appl.
Math., Vol. 56, 1199–1246, 1996.
[79] F. Catt´e, Convergence of iterated aﬃne and morphological ﬁlters by nonlinear semigroup
theory, M.O. Berger, R. Deriche, I. Herlin, J. Jaﬀr´e, J.M. Morel (Eds.), ICAOS ’96:
Images, wavelets and PDEs, Lecture Notes in Control and Information Sciences, Vol. 219,
Springer, London, 125–133, 1996.
[80] F. Catt´e, F. Dibos, G. Koepﬂer, A morphological scheme for mean curvature motion and
applications to anisotropic diﬀusion and motion of level sets, SIAM J. Numer. Anal., Vol.
32, 1895–1909, 1995.
[81] F. Catt´e, P.L. Lions, J.M. Morel, T. Coll, Image selective smoothing and edge detection
by nonlinear diﬀusion, SIAM J. Numer. Anal., Vol. 29, 182–193, 1992.
[82] A. Chambolle, Partial diﬀerential equations and image processing, Proc. IEEE Int. Conf.
Image Processing (ICIP–94, Austin, Nov. 13–16, 1994), Vol. 1, IEEE Computer Society
Press, Los Alamitos, 16–20, 1994.
144 BIBLIOGRAPHY
[83] A. Chambolle, P.L. Lions, Image recovery via total variation minimization and related
problems, Numer. Math., Vol. 76, 167–188, 1997.
[84] R.H. Chan, T.F. Chan, C.K. Wong, Cosine transform based preconditioners for total varia
tion minimization problems in image processing, S. Margenov, P. Vassilevski (Eds.), Proc.
Second IMACS Symposium on Iterative Methods, IMACS Press, 311–329, 1995.
[85] T.F. Chan, G.H. Golub, P. Mulet, A nonlinear primal–dual method for totalvariation
based image restoration, M.O. Berger, R. Deriche, I. Herlin, J. Jaﬀr´e, J.M. Morel (Eds.),
ICAOS ’96: Images, wavelets and PDEs, Lecture Notes in Control and Information Sci
ences, Vol. 219, Springer, London, 241–252, 1996.
[86] T.F. Chan, C.K. Wong, Total variation blind deconvolution, CAM Report 9645, Dept.
of Mathematics, Univ. of California, Los Angeles, CA 90024, U.S.A., 1996; to appear in
IEEE Trans. Image Proc.
[87] T.F. Chan, H.M. Zhou, R.H. Chan, Continuation method for total variation minimizing
problems in image processing, F.T. Luk (Ed.), Advanced signal processing algorithms,
SPIE Vol. 2563, 314–325, 1995.
[88] P. Charbonnier, L. Blanc–F´eraud, G. Aubert, M. Barlaud, Two deterministic half
quadratic regularization algorithms for computed imaging, Proc. IEEE Int. Conf. Image
Processing (ICIP–94, Austin, Nov. 13–16, 1994), Vol. 2, IEEE Computer Society Press,
Los Alamitos, 168–172, 1994.
[89] Y. Chen, B.C. Vemuri, L. Wang, Image denoising and segmentation via nonlinear dif
fusion, Report UFCISETR97015, CISE Dept., University of Florida, Gainesville, FL
32611, U.S.A., 1997; submitted to SIAM J. Appl. Math.
[90] Y.G. Chen, Y. Giga, S. Goto, Uniqueness and existence of viscosity solutions of general
ized mean curvature ﬂow equations, J. Diﬀerential Geometry, Vol. 33, 749–786, 1991.
[91] D.L. Chopp, Computing minimal surfaces via level set curvature ﬂow, Ph.D. thesis LBL
30685, Physics Division, Lawrence Berkeley Laboratory, University of California, Berkeley,
CA 94720, U.S.A., 1991 (see also J. Comput. Phys., Vol. 106, 77–91, 1993).
[92] C.K. Chui, Wavelets: A tutorial in theory and applications, Academic Press, New York,
1992.
[93] L.D. Cohen, On active contour models and balloons, CVGIP: Image Understanding, Vol.
53, 211–218, 1991.
[94] M.A. Cohen, S. Grossberg, Neural dynamics of brightness perception: Features, bound
aries, diﬀusion and resonance, Perception and Psychophysics, Vol. 36, 428–456, 1984.
[95] T. Cohignac, F. Eve, F. Guichard, C. Lopez, J.M. Morel, Numerical analysis of the
fundamental equation of image processing, Preprint No. 9254, CEREMADE, Universit´e
Paris IX – Dauphine, Place du Mar´echal de Lattre de Tassigny, 75775 Paris Cedex 16,
France, 1992.
[96] T. Cohignac, C. Lopez, J.M. Morel, Integral and local aﬃne invariant parameter and
application to shape recognition, Proc. 12th Int. Conf. Pattern Recognition (ICPR 12,
Jerusalem, Oct. 9–13, 1994), Vol. A, IEEE Computer Society Press, Los Alamitos, 164–
168, 1994.
[97] T. Cohignac, J.M. Morel, Scalespace and aﬃne invariant recognition of occluded shapes,
L.I. Rudin, S.K. Bramble (Eds.), Investigative and trial image processing, SPIE Vol. 2567,
214–222, 1995.
BIBLIOGRAPHY 145
[98] G. Cong, S.D. Ma, Nonlinear diﬀusion for early vision, Proc. 13th Int. Conf. Pattern
Recognition (ICPR 13, Vienna, Aug. 25–30, 1996), Vol. A, 403–406, 1996.
[99] G.H. Cottet, Diﬀusion approximation on neural networks and applications for image
processing, F. Hodnett (Ed.), Proc. Sixth European Conf. on Mathematics in Industry,
Teubner, Stuttgart, 3–9, 1992.
[100] G.H. Cottet, Neural networks: Continuous approach and applications to image processing,
J. Biological Systems, Vol. 3, 1131–1139, 1995.
[101] G.H. Cottet, M. El Ayyadi, Nonlinear PDE operators with memory terms for image
processing, Proc. IEEE Int. Conf. Image Processing (ICIP–96, Lausanne, Sept. 16–19,
1996), Vol. 1, 481–483, 1996.
[102] G.H. Cottet, L. Germain, Image processing through reaction combined with nonlinear
diﬀusion, Math. Comp., Vol. 61, 659–673, 1993.
[103] M.G. Crandall, I. Ishii, P.L. Lions, User’s guide to viscosity solutions of second order
partial diﬀerential equations, Bull. Am. Math. Soc., Vol. 27, 1–67, 1992.
[104] M.G. Crandall, P.L. Lions, Convergent diﬀerence schemes for nonlinear parabolic equa
tions and mean curvature motion, Numer. Math., Vol. 75, 17–41, 1996.
[105] J. Damon, Local Morse theory for solutions to the heat equation and Gaussian blurring,
J. Diﬀerential Equations, Vol. 115, 368–401, 1995.
[106] I. Daubechies, Ten lectures on wavelets, SIAM, Philadelphia, 1992.
[107] G. David, C
1
arcs for minimizers of the Mumford–Shah functional, SIAM J. Appl. Math.,
Vol. 56, 783–888, 1996.
[108] E. De Giorgi, M. Carriero, A. Leaci, Existence theorem for a minimum problem with a
free discontinuity set, Arch. Rat. Mech. Anal., Vol. 108, 195–218, 1989.
[109] R. Deriche, Recursively implementing the Gaussian and its derivatives, V. Srinivasan, Ong
S.H., Ang Y.H. (Eds.), Proc. Second Int. Singapore Conf. on Image Proc. (Singapore, Sept.
7–11, 1992), 263–267, 1992.
[110] R. Deriche, O. Faugeras, Les EDP en traitement des images et vision par ordinateur,
Traitement du Signal, Vol. 13, No. 6, 1997.
[111] R. Deriche, P. Kornprobst, G. Aubert, Opticalﬂow estimation while preserving its dis
continuities: A variational approach, Proc. Second Asian Conf. Computer Vision (ACCV
’95, Singapore, December 5–8, 1995), Vol. 2, 290–295, 1995.
[112] F. Dibos, Analyse multi´echelle invariante par transformations projectives, C. R. Acad.
Sci. Paris, t. 320, S´erie I, 917–921, 1995.
[113] F. Dibos, Projective invariant multiscale analysis, Proc. IEEE Int. Conf. Image Processing
(ICIP–96, Lausanne, Sept. 16–19, 1996), Vol. 1, 485–488, 1996.
[114] S. Di Zenzo, A note on the gradient of a multiimage, Computer Vision, Graphics, and
Image Processing, Vol. 33, 116–125, 1986.
[115] D.C. Dobson, F. Santosa, An image enhancement technique for electrical impedance to
mography, Inverse Problems, Vol. 10, 317–334, 1994.
[116] D.C. Dobson, F. Santosa, Recovery of blocky images from noisy and blurred data, SIAM
J. Appl. Math., Vol. 56, 1181–1198, 1996.
146 BIBLIOGRAPHY
[117] D. Dobson, O. Scherzer, Analysis of regularized total variation penalty methods for de
noising, Inverse Problems, Vol. 12, 601–617, 1996.
[118] D.C. Dobson, C.R. Vogel, Convergence of an iterative method for total variation denois
ing, Preprint, Dept. of Mathematical Sciences, Montana State University, Bozeman, MT
597170240, U.S.A., 1995; SIAM J. Numer. Anal., to appear.
[119] L. Dorst, R. van den Boomgaard, Morphological signal processing and the slope transform,
Signal Processing, Vol. 38, 79–98, 1994.
[120] J. Douglas, J.E. Gunn, A general formulation of alternating direction methods. Part I.
Parabolic and hyperbolic problems, Numer. Math., Vol. 6, 428–453, 1964.
[121] J. Douglas, B.F. Jones, On predictor–corrector methods for nonlinear parabolic diﬀerential
equations, J. Soc. Ind. Appl. Math., Vol. 11, 195–204, 1963.
[122] I.S. Duﬀ, A.M. Erisman, J.K. Reid, Direct methods for sparse systems, Clarendon, Oxford,
1986.
[123] S.K. Dzhu Magazieva, Numerical study of a partial diﬀerential equation, U.S.S.R. Comput.
Maths. Math. Phys., Vol. 23, No. 4, 45–49, 1983.
[124] D. Eberly, A diﬀerential geometric approach to anisotropic diﬀusion, B.M. ter Haar
Romeny (Ed.), Geometrydriven diﬀusion in computer vision, Kluwer, Dordrecht, 371–
391, 1994.
[125] S. Eidelman, W. Grossmann, A. Friedman, Nonlinear signal processing using integration
of ﬂuid dynamics equations, A.G. Tescher (Ed.): Applications of digital image processing
XIV, SPIE Vol. 1567, 439–450, 1991.
[126] A.I. ElFallah, G.E. Ford, Nonlinear adaptive image ﬁltering based on inhomogeneous
diﬀusion and diﬀerential geometry, S.A. Rajala, R.L. Stevenson (Eds.), Image and video
processing II, SPIE Vol. 2182, 49–63, 1994.
[127] A.I. ElFallah, G.E. Ford, The evolution of mean curvature in image ﬁltering, Proc. IEEE
Int. Conf. Image Processing (ICIP–94, Austin, Nov. 13–16, 1994), Vol. 1, IEEE Computer
Society Press, Los Alamitos, 298–302, 1994.
[128] H.W. Engl, M. Hanke, A. Neubauer, Regularization of inverse problems, Kluwer, Dor
drecht, 1996.
[129] L.C. Evans, J. Spruck, Motion of level sets by mean curvature I., J. Diﬀerential Geometry,
Vol. 33, 635–681, 1991.
[130] O. Faugeras, Sur l’´evolution de courbes simples du plan projectif r´eel, C. R. Acad. Sci.
Paris, t. 317, S´erie I, 565–570, 1993.
[131] O. Faugeras, Cartan’s moving frame method and its application to the geometry and evolu
tion of curves in the Euclidean, aﬃne and projective planes, J.L. Mundy, A. Zisserman, D.
Forsyth (Eds.), Applications of invariance in computer vision, Lecture Notes in Computer
Science, Vol. 825, Springer, Berlin, 11–46, 1994.
[132] O. Faugeras, R. Keriven, Scalespaces and aﬃne curvature, W. Chenke, R. Mohr (Eds.),
Proc. Workshop on Geometrical Modelling and Invariants for Computer Vision (Xi’an,
April 27–30, 1995), Xidian University Press, Xi’an, 1995.
[133] O. Faugeras, R. Keriven, On projective plane curve evolution, M.O. Berger, R. Deriche,
I. Herlin, J. Jaﬀr´e, J.M. Morel (Eds.), ICAOS ’96: Images, wavelets and PDEs, Lecture
Notes in Control and Information Sciences, Vol. 219, Springer, London, 66–73, 1996.
BIBLIOGRAPHY 147
[134] O. Faugeras, R. Keriven, Level set methods and the stereo problem, B. ter Haar Romeny,
L. Florack, J. Koenderink, M. Viergever (Eds.), Scalespace theory in computer vision,
Lecture Notes in Computer Science, Vol. 1252, Springer, Berlin, 272–283, 1997.
[135] S. Fejes, A. Rosenfeld, Discrete active models and applications, Pattern Recognition, Vol.
30, 817–835, 1997.
[136] P. Fiddelaers, E.J. Pauwels, L.J. Van Gool, Geometrydriven curve evolution, J.O. Ek
lundh (Ed.), Computer Vision – ECCV ’94, Lecture Notes in Computer Science, Vol. 800,
Springer, Berlin, 427–432, 1994.
[137] B. Fischl, E.L. Schwartz, Adaptive nonlinear ﬁltering for nonlinear diﬀusion approxima
tion in image processing, Proc. 13th Int. Conf. Pattern Recognition (ICPR 13, Vienna,
Aug. 25–30, 1996), Vol. D, 276–280, 1996.
[138] L.M.J. Florack, Data, models, and images, Proc. 1996 IEEE Int. Conf. Image Processing
(ICIP–96, Lausanne, Sept. 16–19, 1996), Vol. I, 469–472, 1996.
[139] L.M.J. Florack, Image structure, Kluwer, Dordrecht, 1997.
[140] L.M.J. Florack, B.M. ter Haar Romeny, J.J. Koenderink, M.A. Viergever, Scale and the
diﬀerential structure of images, Image Vision Comput., Vol. 10, 376–388, 1992.
[141] L.M.J. Florack, B.M. ter Haar Romeny, J.J. Koenderink, M.A. Viergever, The Gaussian
scalespace paradigm and the multiscale local jet, Int. J. Comput. Vision, Vol. 18, 61–75,
1996.
[142] L. Florack, R. Maas, Pseudolinear scalespace theory: towards the integration of linear
and morphological scalespace paradigms, Technical Report UUCS199725, Dept. of Com
puter Science, Padualaan 14, 3584 CH Utrecht, The Netherlands, 1997.
[143] L.M.J. Florack, A.H. Salden, B.M. ter Haar Romeny, J.J. Koenderink, M.A. Viergever,
Nonlinear scalespace, Image Vision Comput., Vol. 13, 279–294, 1995.
[144] G.E. Ford, R.R. Estes, H. Chen, Scalespace analysis for image sampling and interpola
tion, Proc. IEEE Int. Conf. Acoustics, Speech and Signal Processing (ICASSP–92, San
Francisco, March 23–26, 1992), Vol. 3, 165–168, 1992.
[145] M.A. F¨orstner, E. G¨ ulch, A fast operator for detection and precise location of distinct
points, corners and centres of circular features, Proc. ISPRS Intercommission Conf. on
Fast Processing of Photogrammetric Data (Interlaken, June 2–4, 1987), 281–305, 1987.
[146] T. Franke, H. Neumann, R. Seydel, Anisotropic diﬀusion based on mean curvature motion:
A computational study, B. J¨ahne, P. Geißler, H. Haußecker, F. Hering (Eds.), Mustererken
nung 1996, Springer, Berlin, 47–54, 1996.
[147] D.S. Fritsch, A medial description of greyscale image structure by gradientlimited dif
fusion, R.A. Robb (Ed.), Visualization in biomedical computing ’92, SPIE Vol. 1808,
105–117, 1992.
[148] J. Fr¨ohlich, J. Weickert, Image processing using a wavelet algorithm for nonlinear diﬀu
sion, Report No. 104, Laboratory of Technomathematics, University of Kaiserslautern,
P.O. Box 3049, 67653 Kaiserslautern, Germany, 1994.
[149] D. Gabor, Information theory in electron microscopy, Laboratory Investigation, Vol. 14,
801–807, 1965.
[150] M. Gage, On areapreserving evolution equation for plane curves, Contemp. Math., Vol.
23, 51–62, 1986.
148 BIBLIOGRAPHY
[151] M. Gage, R.S. Hamilton, The heat equation shrinking convex plane curves, J. Diﬀerential
Geometry, Vol. 23, 69–96, 1986.
[152] D. Geiger, A. Yuille, A common framework for image segmentation, Int. J. Comput.
Vision, Vol. 6, 227–243, 1991.
[153] D. Geman, G. Reynolds, Constrained restoration and the recovery of discontinuities, IEEE
Trans. Pattern Anal. Mach. Intell., Vol. 14, 367–383, 1992.
[154] S. Geman, D. Geman, Stochastic relaxation, Gibbs distributions, and the Bayesian restora
tion of images, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 6, 721–741, 1984.
[155] G. Gerig, O. K¨ ubler, R. Kikinis, F.A. Jolesz, Nonlinear anisotropic ﬁltering of MRI data,
IEEE Trans. Medical Imaging, Vol. 11, 221–232, 1992.
[156] G. Gerig, G. Sz´ekely, Finding line structures by multivalued nonlinear diﬀusion of feature
maps, Report, Communication Technology Laboratory, Image Science Division, ETH–
Zentrum, 8092 Z¨ urich, Switzerland, 1993.
[157] G. Gerig, G. Sz´ekely, G. Israel, M. Berger, Detection and characterization of unsharp blobs
by curve evolution, Y. Bizais, C. Barillot, R. Di Paola (Eds.), Information Processing in
Medical Imaging, Kluwer, Dordrecht, 165–176, 1995.
[158] T. Gijbels, P. Six, L. Van Gool, F. Catthoor, H. De Man, A. Oosterlinck, A VLSI
architecture for parallel nonlinear diﬀusion with applications in vision, Proc. IEEE Work
shop on VLSI Signal Processing, Vol. 7, 398–407, 1994.
[159] D. Giuiliani, Directed diﬀusion equations and edge detection, master thesis, Dept. of
Mathematics, University of Kaiserslautern, P.O. Box 3049, 67653 Kaiserslautern, Ger
many, 1991.
[160] V. van Gogh, Starry night, SaintR´emy, 1889. New York, The Museum of Modern Art.
[161] V. van Gogh, Selfportrait, SaintR´emy, 1889. Paris, Muse´e d’Orsay.
[162] V. van Gogh, Road with cypress and star, AuverssurOise, 1890. Otterlo, Rijksmuseum
Kr¨oller–M¨ uller.
[163] R.C. Gonzalez, P. Wintz, Digital image processing, Addison–Wesley, Reading, 1987.
[164] R. Gorenﬂo, Conservative diﬀerence schemes for diﬀusion problems, Preprint No. 39, FB
Mathematik, FU Berlin, Arnimallee 2–6, W1000 Berlin 33, Germany, 1977.
[165] A. Goshtasby, On edge focusing, Image Vision Comput., Vol. 12, 247–256, 1994.
[166] A.R. Gourlay, Implicit convolution, Image Vision Comput., Vol. 3, 15–23, 1985.
[167] R. Graham, Snow removal: A noisestripping process for TV signals, IRE Trans. Infor
mation Theory, Vol. 8, 129–144, 1962.
[168] G.H. Granlund, H. Knutsson, Signal processing for computer vision, Kluwer, Dordrecht,
1995.
[169] M. Grayson, The heat equation shrinks embedded plane curves to round points, J. Diﬀer
ential Geometry, Vol. 26, 285–314, 1987.
[170] D. Greenspan, P.C. Jain, On nonnegative diﬀerence analogues of elliptic diﬀerential equa
tions, J. Franklin Inst., Vol. 279, 360–365, 1965.
[171] L.D. Griﬃn, Scaleimprecision space, Image Vision Comput., Vol. 15, 369–398, 1997.
[172] L.D. Griﬃn, A.C.F. Colchester, G.P. Robinson, Scale and segmentation of greylevel im
ages using maximum gradient paths, Image Vision Comput., Vol. 10, 389–402, 1992.
BIBLIOGRAPHY 149
[173] F. Guichard, J.M. Morel, Partial diﬀerential equations and image iterative ﬁltering,
Preprint No. 9535, CEREMADE, Universit´e Paris IX – Dauphine, Place du Mar´echal
de Lattre de Tassigny, 75775 Paris Cedex 16, France, 1995; to appear in I.S. Duﬀ, G.A.
Watson (Eds.), The state of the art in numerical analysis, Oxford University Press.
[174] B.M. ter Haar Romeny (Ed.), Geometrydriven diﬀusion in computer vision, Kluwer,
Dordrecht, 1994.
[175] B.M. ter Haar Romeny, Applications of scalespace theory, J. Sporring, M. Nielsen, L.
Florack, P. Johansen (Eds.), Gaussian scalespace theory, Kluwer, Dordrecht, 3–19, 1997.
[176] B.M. ter Haar Romeny, Frontend vision and multiscale image analysis: Introduction to
scalespace theory, Kluwer, Dordrecht, in preparation.
[177] B.M. ter Haar Romeny, L.M.J. Florack, M. de Swart, J. Wilting, M.A. Viergever, Deblur
ring Gaussian blur, F.L. Bookstein, J.S. Duncan, N. Lange, D.C. Wilson (Eds.), Mathe
matical methods in medical imaging III, SPIE Vol. 2299, 139–148, 1994.
[178] B.M. ter Haar Romeny, W.J. Niessen, J. Wilting, L.M.J. Florack, Diﬀerential structure
of images: accuracy of representation, Proc. IEEE Int. Conf. Image Processing (ICIP–94,
Austin, Nov. 13–16, 1994), Vol. 1, IEEE Computer Society Press, Los Alamitos, 21–25,
1994.
[179] W. Hackbusch, Multigrid methods and applications, Springer, New York, 1985.
[180] W. Hahn, Theorie und Anwendung der direkten Methode von Ljapunov, Springer, Berlin,
1959.
[181] R.M. Haralick, S.R. Sternberg, X. Zhuang, Image analysis using mathematical morphology,
IEEE Trans. Pattern Anal. Mach. Intell., Vol. 9, 532–550, 1987.
[182] C.G. Harris, M. Stephens, A combined corner and edge detector, Proc. Fouth Alvey Vision
Conf. (AVC 88, Manchester, Aug. 31 – Sept. 2, 1988), 147–152, 1988.
[183] A. Harten, B. Enquist, S. Osher, S. Chakravarthy, Uniformly high order accurrate essen
tially nonoscillatory schemes, III, J. Comput. Phys., Vol. 71, 231–303, 1987.
[184] H.J.A.M. Heijmans, Mathematical morphology: A modern approach in image processing
based on algebra and geometry, SIAM Review, Vol. 37, 1–36, 1995.
[185] G. Hellwig, Partial diﬀerential equations, Teubner, Stuttgart, 1977.
[186] G. Hewer, C. Kenney, B.S. Manjunath, Variational image segmentation using boundary
functions, Report No. 9607, ECE Dept., University of California, Santa Barbara, CA
93106, U.S.A., 1996; submitted to IEEE Trans. Image Proc.
[187] K. H¨ollig, Existence of inﬁnitely many solutions for a forward–backward heat equation,
Trans. Amer. Math. Soc., Vol. 278, 299–316, 1983.
[188] G. Huisken, Flow by mean curvature of convex surfaces into spheres, J. Diﬀerential Geo
metry, Vol. 20, 237–266, 1984.
[189] R.A. Hummel, Representations based on zerocrossings in scale space, Proc. IEEE Comp.
Soc. Conf. Computer Vision and Pattern Recognition (CVPR ’86, Miami Beach, June
22–26, 1986), IEEE Computer Society Press, Washington, 204–209, 1986.
[190] R. Hummel, R. Moniot, Reconstructions from zerocrossings in scale space, IEEE Trans.
Acoustics, Speech and Signal Processing, Vol. 37, 2111–2130, 1989.
[191] T. Iijima, Basic theory of pattern observation, Papers of Technical Group on Automata
and Automatic Control, IECE, Japan, Dec. 1959 (in Japanese).
150 BIBLIOGRAPHY
[192] T. Iijima, Basic theory on normalization of a pattern (for the case of a typical one
dimensional pattern), Bulletin of the Electrotechnical Laboratory, Vol. 26, 368–388, 1962
(in Japanese).
[193] T. Iijima, Observation theory of twodimensional visual patterns, Papers of Technical
Group on Automata and Automatic Control, IECE, Japan, Oct. 1962 (in Japanese).
[194] T. Iijima, Basic theory on normalization of twodimensional visual pattern, Studies on
Information and Control, Pattern Recognition Issue, IECE, Japan, No. 1, 15–22, 1963 (in
Japanese).
[195] T. Iijima, Theory of pattern recognition, Electronics and Communications in Japan, 123–
134, Nov. 1963 (in English).
[196] T. Iijima, Basic equation of ﬁgure and observational transformation, Systems, Computers,
Controls, Vol. 2, No. 4, 70–77, 1971 (in English).
[197] T. Iijima, Pattern recognition, Coronasha, 1973 (in Japanese).
[198] T. Iijima, Theory of pattern recognition, Series of Basic Information Technology, Vol. 6,
Morishita Publishing, 1989 (in Japanese).
[199] T. Iijima, H. Genchi, K. Mori, A theoretical study of pattern identiﬁcation by matching
method, Proc. First USA–Japan Computer Conference (Tokyo, Oct. 3–5, 1972), 42–48,
1972 (in English).
[200] T. Iijima, H. Genchi, K. Mori, A theory of character recognition by pattern matching
method, Proc. First Int. Joint Conf. Pattern Recognition (IJCPR 1, Washington, Oct.
1973), 50–56, 1973 (in English).
[201] R. Illner, Global existence for twovelocity models of the Boltzmann equation, Math. Meth.
Appl. Sci., Vol. 1, 187–193, 1979.
[202] R. Illner, H. Neunzert, Relative entropy maximization and directed diﬀusion equations,
Math. Meth. Appl. Sci., Vol. 16, 545–554, 1993.
[203] R. Illner, J. Tie, On directed diﬀusion with measurable background, Math. Meth. Appl.
Sci., Vol. 16, 681–690, 1993.
[204] K. Ito, K. Kunisch, Augmented Lagrangian methods for nonsmooth, convex optimization
in Hilbert spaces, Preprint No. 409/1994, Fachbereich Mathematik (3), Technische Uni
versit¨at Berlin, Straße des 17. Juni 136, 10623 Berlin, Germany, 1994.
[205] K. Ito, K. Kunisch, An active set strategy for image restoration based on the augmented
Lagrangian formulation, Preprint No. 410/1994, Fachbereich Mathematik (3), Technische
Universit¨at Berlin, Straße des 17. Juni 136, 10623 Berlin, Germany, 1994.
[206] P.T. Jackway, Morphological scalespace, Proc. 11th Int. Conf. Pattern Recognition (ICPR
11, The Hague, Aug. 30 – Sept. 3, 1992), Vol. C, 252–255, 1992.
[207] P.T. Jackway, M. Deriche, Scalespace properties of the multiscale morphological dilation–
erosion, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 18, 38–51, 1996.
[208] M. J¨agersand, Saliency maps and attention selection in scale and spatial coordinates: an
information theoretic approach, Proc. Fifth Int. Conf. on Computer Vision (ICCV ’95,
Cambridge, June 20–23, 1995), IEEE Computer Society Press, Los Alamitos, 195–202,
1995.
[209] B. J¨ahne, Spatiotemporal image processing, Lecture Notes in Computer Science, Vol. 751,
Springer, Berlin, 1993.
BIBLIOGRAPHY 151
[210] P. Johansen, On the classiﬁcation of toppoints in scalespace, J. Math. Imag. Vision, Vol.
4, 57–67, 1994.
[211] P. Johansen (Ed.), Proceedings of the workshop on Gaussian scalespace theory, Report
DIKU 96/19, Dept. of Computer Science, University of Copenhagen, Universitetsparken
1, 2100 Copenhagen, Denmark, 1996.
[212] P. Johansen, S. Skelboe, K. Grue, J.D. Andersen, Representing signals by their toppoints
in scale space, Proc. Eighth Int. Conf. Pattern Recognition (ICPR 8, Paris, Oct. 27–31,
1986), 215–217, 1986.
[213] F. John, Partial diﬀerential equations, Springer, New York, 1978.
[214] J.M. Jolion, A. Rosenfeld, A pyramid framework for early vision, Kluwer, Dordrecht,
1994.
[215] D.G. Jones, J. Malik, A computational framework for determining stereo correspondence
from a set of linear spatial ﬁlters, G. Sandini (Ed.), Computer vision – ECCV ’92, Lecture
Notes in Computer Science, Vol. 588, Springer, Berlin, 395–410, 1992.
[216] J. Kaˇcur, K. Mikula, Solution of nonlinear diﬀusion appearing in image smoothing and
edge detection, Appl. Num. Math., Vol. 17, 47–59, 1995.
[217] J. Kaˇcur, K. Mikula, Slowed anisotropic diﬀusion, B. ter Haar Romeny, L. Florack, J.
Koenderink, M. Viergever (Eds.), Scalespace theory in computer vision, Lecture Notes
in Computer Science, Vol. 1252, Springer, Berlin, 357–360, 1997.
[218] H.W. Kaﬁtz, Morphologische Diﬀerentialgleichungen in der Bildverarbeitung, diploma
thesis, Dept. of Mathematics, University of Kaiserslautern, P.O. Box 3049, 67653 Kaisers
lautern, Germany, 1996.
[219] G. Kanizsa, Organization in vision, Praeger, New York, 1979.
[220] M. Kass, A. Witkin, Analyzing oriented patterns, Computer Vision, Graphics, and Image
Processing, Vol. 37, 362–385, 1987.
[221] M. Kass, A. Witkin, D. Terzopoulos, Snakes: Active contour models, Int. J. Comput.
Vision, Vol. 1, 321–331, 1988.
[222] B. Kawohl, N. Kutev, Maximum and comparison principles for anisotropic diﬀusion,
Preprint, Mathematical Institute, University of Cologne, 50923 Cologne, Germany, 1997.
[223] J.G. Kemeny, J.L. Snell, Finite Markov chains, D. Van Nostrand, Princeton, 1960.
[224] S. Kichenassamy, Nonlinear diﬀusions and hyperbolic smoothing for edge enhancement,
M.O. Berger, R. Deriche, I. Herlin, J. Jaﬀr´e, J.M. Morel (Eds.), ICAOS ’96: Im
ages, Wavelets and PDEs, Lecture Notes in Control and Information Sciences, Vol. 219,
Springer, London, 119–124, 1996.
[225] S. Kichenassamy, The Perona–Malik paradox, SIAM J. Appl. Math., to appear.
[226] S. Kichenassamy, A. Kumar, P. Olver, A. Tannenbaum, A. Yezzi, Conformal curvature
ﬂows: from phase transitions to active vision, Arch. Rat. Mech. Anal., Vol. 134, 275–301,
1996.
[227] B.B. Kimia, K. Siddiqi, Geometric heat equation and nonlinear diﬀusion of shapes and
images, Computer Vision and Image Understanding, Vol. 64, 305–322, 1996.
[228] B.B. Kimia, A. Tannenbaum, S.W. Zucker, Toward a computational theory of shape: An
overview, O. Faugeras (Ed.), Computer Vision – ECCV ’90, Lecture Notes in Computer
Science, Vol. 427, Springer, Berlin, 402–407, 1990.
152 BIBLIOGRAPHY
[229] B.B. Kimia, A. Tannenbaum, S.W. Zucker, On the evolution of curves via a function of
curvature. I. The classical case, J. Math. Anal. Appl., Vol. 163, 438–458, 1992.
[230] B. Kimia, A. Tannenbaum, S.Zucker, Shapes, shocks, and deformations I: The components
of twodimensional shape and the reaction–diﬀusion space, Int. J. Comput. Vision, Vol.
15, 189–224, 1995.
[231] R. Kimmel, Aﬃne diﬀerential signatures for gray level images of planar shape, Proc. 13th
Int. Conf. Pattern Recognition (ICPR 13, Vienna, Aug. 25–30, 1996), Vol. A, 45–49, 1996.
[232] R. Kimmel, Intrinsic scalespace for images on surfaces: The geodesic curvature ﬂow, B.
ter Haar Romeny, L. Florack, J. Koenderink, M. Viergever (Eds.), Scalespace theory in
computer vision, Lecture Notes in Computer Science, Vol. 1252, Springer, Berlin, 212–223,
1997.
[233] R. Kimmel, N. Kiryati, A.M. Bruckstein, Analyzing and synthesizing images by evolving
curves with the OsherSethian method, Int. J. Comput. Vision, Vol. 24, 37–56, 1997.
[234] R. Kimmel, N. Sochen, R. Malladi, On the geometry of texture, Preprint LBNL39640,
Lawrence Berkeley National Laboratory, University of California, Berkeley, CA 94720,
U.S.A., 1996.
[235] R. Kimmel, N. Sochen, R. Malladi, Images as embedding maps and minimal surfaces:
Movies, color, and volumetric medical images, Proc. IEEE Comp. Soc. Conf. Computer
Vision and Pattern Recognition (CVPR ’97, San Juan, June 17–19, 1997), IEEE Computer
Society Press, Los Alamitos, 350–355, 1997.
[236] G. Kin, M. Sato, Scale space ﬁltering on spherical pattern, Proc. 11th Int. Conf. Pattern
Recognition (ICPR 11, The Hague, Aug. 30 – Sept. 3, 1992), Vol. C, 638–641, 1992.
[237] H.E. Knutsson, R. Wilson, G.H. Granlund, Anisotropic nonstationary image estimation
and its applications: Part I – Restoration of noisy images, IEEE Trans. Comm., Vol. 31,
388–397, 1983.
[238] M. Kocan, Approximation of viscosity solutions of elliptic partial diﬀerential equations on
minimal grids, Numer. Math., Vol. 72, 73–92, 1996.
[239] G. Koepﬂer, C. Lopez, J.M. Morel, A multiscale algorithm for image segmentation by
variational method, SIAM J. Numer. Anal., Vol. 31, 282–299, 1994.
[240] J.J. Koenderink, The structure of images, Biol. Cybern., Vol. 50, 363–370, 1984.
[241] J.J. Koenderink, Scaletime, Biol. Cybern., Vol. 58, 159–162, 1988.
[242] J.J. Koenderink, A.J. van Doorn, Representation of local geometry in the visual system,
Biol. Cybern., Vol. 55, 367–375, 1987.
[243] H.P. Kramer, J.B. Bruckner, Iterations of a nonlinear transformation for enhancement
of digital images, Pattern Recognition, Vol. 7, 53–58, 1975.
[244] K. Krissian, G. Malandain, N. Ayache, Directional anisotropic diﬀusion applied to seg
mentation of vessels in 3D images, B. ter Haar Romeny, L. Florack, J. Koenderink, M.
Viergever (Eds.), Scalespace theory in computer vision, Lecture Notes in Computer Sci
ence, Vol. 1252, Springer, Berlin, 345–348, 1997.
[245] A. Kumar, A.R. Tannenbaum, G.J. Balas, Optic ﬂow: a curve evolution approach, IEEE
Trans. Image Proc., Vol. 5, 598–610, 1996.
[246] O.A. Ladyˇzenskaja, V.A. Solonnikov, N.N. Ural’ceva, Linear and quasilinear equations of
parabolic type, American Mathematical Society, Providence, 1968.
BIBLIOGRAPHY 153
[247] D.S. Lalush, The relationship between diﬀusion ﬁltering and Gibbs smoothing, personal
note, Dept. of Biomedical Engineering, University of North Carolina, Chapel Hill, NC
27599, U.S.A., 1994.
[248] C. Lamberti, M. Sitta, F. Sgallari, Improvements to the anisotropic diﬀusion model for
2D echo image processing, Proc. Annual Int. Conf. of the IEEE Engineering in Medicine
and Biology Society, Vol. 14, 1872–1873, 1992.
[249] A. Leaci, S. Solimini, Variational problems with a free discontinuity set, B.M. ter Haar
Romeny (Ed.), Geometrydriven diﬀusion in computer vision, Kluwer, Dordrecht, 147–
154, 1994.
[250] A. Lev, S. Zucker, A. Rosenfeld, Iterative enhancement of noisy images, IEEE Trans.
Systems, Man and Cybernetics, Vol. 7, 435–442, 1977.
[251] S.Z. Li, On discontinuityadaptive smoothness priors in computer vision, IEEE Trans.
Pattern Anal. Mach. Intell., Vol. 17, 576–586, 1995.
[252] X. Li, T. Chen, Nonlinear diﬀusion with multiple edginess thresholds, Pattern Recognition,
Vol. 27, 1029–1037, 1994.
[253] Y. Li, F. Santosa, A computational algorithm for minimizing total variation in image
enhancement, IEEE Trans. Image Proc., Vol. 5, 987–995, 1996.
[254] L.M. Lifshitz, S.M. Pizer, A multiresolution hierarchical approach to image segmentation
based on intensity extrema, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 12, 529–540,
1990.
[255] T. Lindeberg, Scalespace for discrete signals, IEEE Trans. Pattern Anal. Mach. Intell.,
Vol. 12, 234–254, 1990.
[256] T. Lindeberg, Scalespace theory in computer vision, Kluwer, Boston, 1994.
[257] T. Lindeberg, On the axiomatic foundations of linear scalespace, J. Sporring, M. Nielsen,
L. Florack, P. Johansen (Eds.), Gaussian scalespace theory, Kluwer, Dordrecht, 75–97,
1997.
[258] T. Lindeberg, D. Fagerstr¨om, Scalespace with causal time direction, B. Buxton, R. Cipolla
(Eds.), Computer vision – ECCV ’96, Volume I, Lecture Notes in Computer Science, Vol.
1064, Springer, Berlin, 229–240, 1996.
[259] T. Lindeberg, J. G˚arding, Shapeadapted smoothing in estimation of 3D depth cues from
aﬃne distortions of local 2D brightness structure, Image Vision Comput., Vol. 15, 415–
434, 1997.
[260] M. Lindenbaum, M. Fischer, A. Bruckstein, On Gabor’s contribution to image enhance
ment, Pattern Recognition, Vol. 27, 1–8, 1994.
[261] P.L. Lions, Axiomatic derivation of image processing models, Math. Mod. Meth. Appl.
Sci., Vol. 4, 467–475, 1994.
[262] P.L. Lions, S. Osher, L.I. Rudin, Denoising and deblurring images using constrained
nonlinear partial diﬀerential equations, Preprint, Cognitech, Inc., 2800 – 28 St., Suite 101,
Santa Monica, CA 90405, U.S.A., 1993.
[263] P. Lobel, L. BlancF´eraud, C. Pichot, M. Barlaud, A new regularization scheme for inverse
scattering, Inverse Problems, Vol. 13, 403–410, 1997.
[264] M.H. Loew, J. Rosenman, J. Chen, A clinical tool for enhancement of portal images, M.H.
Loew (Ed.), Image processing, SPIE Vol. 2167, 543–550, 1994.
154 BIBLIOGRAPHY
[265] C. Lopez, J.M. Morel, Axiomatization of shape analysis and applications to texture hyper
discrimination, Proc. IEEE Comp. Soc. Conf. Computer Vision and Pattern Recognition
(CVPR ’93, New York, June 15–17, 1993), IEEE Computer Society Press, Los Alamitos,
646–647, 1993.
[266] A.K. Louis, Inverse und schlecht gestellte Probleme, Teubner, Stuttgart, 1989.
[267] L. Lucido, R. Deriche, L. Alvarez, V. Rigaud, Sur quelques sch´emas num´eriques de
r´esolution d’´equations aux d´eriv´ees partielles pour le traitement d’images, Report No.
3192, ROBOTVIS, INRIA, 2004 route des Lucioles, BP 93, 06902 Sophia–Antipolis Cedex,
France, 1997.
[268] L. Lucido, J. Opderbecke, V. Rigaud, R. Deriche, Z. Zhang, An integrated multiscale
approach for terrain referenced underwater navigation, Proc. IEEE Int. Conf. Image Pro
cessing (ICIP–96, Lausanne, Sept. 16–19, 1996), Vol. 2, 633–636, 1996.
[269] A. Lukschin, H. Neunzert, J. Struckmeier, Interim report of the project DPH 6473/91 –
Coupling of NavierStokes and Boltzmann regions, Internal Report, Laboratory of Tech
nomathematics, Kaiserslautern, 1993.
[270] D.S. Luo, M.A. King, S. Glick, Local geometry variable conductance diﬀusion for post
reconstruction ﬁltering, IEEE Trans. Nuclear Sci., Vol. 41, 2800–2806, 1994.
[271] R. Malladi, R. Kimmel, D. Adalsteinsson, G. Sapiro, V. Caselles, J.A. Sethian, A geometric
approach to segmentation and analysis of 3D medical images, Proc. IEEE Workshop on
Mathematical Models in Biomedical Image Analysis (MMBIA ’96, San Francisco, June
21–22, 1996), IEEE Computer Society Press, Los Alamitos, 244–252, 1996.
[272] R. Malladi, J.A. Sethian, Image processing: ﬂows under min/max curvature and mean
curvature, Graphical Models and Image Processing, Vol. 58, 127–141, 1996.
[273] R. Malladi, J.A. Sethian, B.C. Vemuri, Shape modeling with front propagation: A level set
approach, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 17, 158–175, 1995.
[274] P. Maragos, A representation theory for morphological image and signal processing, IEEE
Trans. Pattern Anal. Mach. Intell., Vol. 11, 586–599, 1989.
[275] P. Maragos, Morphological systems: Slope transforms and max–min diﬀerence and diﬀer
ential equations, Signal Processing, Vol. 38, 57–77, 1994.
[276] P. Maragos, Diﬀerential morphology and image processing, IEEE Trans. Image Proc., Vol.
5, 922–937, 1996.
[277] G.I. Marchuk, Splitting and alternating direction methods, P.G. Ciarlet, J.L. Lions (Eds.),
Handbook of numerical analysis, Vol. I, 197–462, 1990.
[278] D. Marr, E. Hildreth, Theory of edge detection, Proc. R. Soc. London Ser. B, Vol. 207,
187–217, 1980.
[279] J.L. Marroquin, Deterministic interactive particle models for image processing and com
puter graphics, CVGIP: Graphical Models and Image Processing, Vol. 55, 408–417, 1993.
[280] G. Matheron, Random sets and integral geometry, Wiley, New York, 1975.
[281] T. McInerney, D. Terzopoulos, Topologically adaptable snakes, Proc. Fifth Int. Conf. on
Computer Vision (ICCV ’95, Cambridge, June 20–23, 1995), IEEE Computer Society
Press, Los Alamitos, 840–845, 1995.
[282] T. McInerney, D. Terzopoulos, Deformable models in medical image analysis: A survey,
Medical Image Analysis, Vol. 1, 91–108, 1996.
BIBLIOGRAPHY 155
[283] T. McInerney, D. Terzopoulos, Medical image segmentation using topologically adaptable
surfaces, J. Troccaz, E. Grimson, R. M¨osges (Eds.), CVRMed–MRCAS ’97, Lecture Notes
in Computer Science, Vol. 1205, Springer, Berlin, 23–32, 1997.
[284] T. Meis, U. Marcowitz, Numerische Behandlung partieller Diﬀerentialgleichungen,
Springer, Berlin, 1978.
[285] B. Merriman, J. Bence, S. Osher, Motion of multiple junctions: a level set approach, J.
Comput. Phys., Vol. 112, 334–363, 1994.
[286] A.R. Mitchell, D.F. Griﬃths, The ﬁnite diﬀerence method in partial diﬀerential equations,
Wiley, Chichester, 1980.
[287] L. Moisan, Analyse multi´echelle de ﬁlms pour la reconstruction du relief, C. R. Acad. Sci.
Paris, t. 320, S´erie I, 279–284, 1995.
[288] L. Moisan, Perspective invariant movie analysis for depth recovery, L.I. Rudin, S.K. Bram
ble (Eds.), Investigative and trial image processing, SPIE Vol. 2567, 84–94, 1995.
[289] L. Moisan, Aﬃne plane curve evolution: a fully consistent scheme, Preprint No. 9628,
CEREMADE, Universit´e Paris IX – Dauphine, Place du Mar´echal de Lattre de Tassigny,
75775 Paris Cedex 16, France, 1996.
[290] F. Mokhtarian, Multiscale contour segmentation, B. ter Haar Romeny, L. Florack, J.
Koenderink, M. Viergever (Eds.), Scalespace theory in computer vision, Lecture Notes
in Computer Science, Vol. 1252, Springer, Berlin, 296–307, 1997.
[291] F. Mokhtarian, A.K. Mackworth, A theory of multiscale, curvaturebased shape represen
tation for planar curves, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 14, 789–805, 1992.
[292] J.M. Morel, S. Solimini, Variational methods in image segmentation, Birkh¨auser, Boston,
1994.
[293] K.W. Morton, D.F. Mayers, Numerical solution of partial diﬀerential equations, Cam
bridge University Press, Cambridge, 1994.
[294] T.S. Motzkin, W. Wasow, On the approximation of linear elliptic diﬀerential equations
by diﬀerence equations with positive coeﬃcients, J. Mathematics and Physics, Vol. 31,
253–259, 1952.
[295] D. Mumford, J. Shah, Boundary detection by minimizing functionals, I, Proc. IEEE Comp.
Soc. Conf. Computer Vision and Pattern Recognition (CVPR ’85, San Francisco, June
19–23, 1985), IEEE Computer Society Press, Washington, 22–26, 1985.
[296] D. Mumford, J. Shah, Optimal approximations by piecewise smooth functions and associ
ated variational problems, Comm. Pure Appl. Math., Vol. 42, 577–685, 1989.
[297] M. Nagao, T. Matsuyama, Edge preserving smoothing, Computer Graphics and Image
Processing, Vol. 9, 394–407, 1979.
[298] P. Neskovic, B.B. Kimia, Threedimensional shape representation from curvature depen
dent surface evolution, Proc. IEEE Int. Conf. Image Processing (ICIP–94, Austin, Nov.
13–16, 1994), Vol. 1, IEEE Computer Society Press, Los Alamitos, 6–10, 1994.
[299] H. Neunzert, B. Claus, K. Rjasanowa, R. R¨osch, J. Weickert, Mathematische Werkzeuge in
der Bildverarbeitung zur Qualit¨atsbeurteilung von Oberﬂ¨achen, K.H. Hoﬀmann, W. J¨ager,
T. Lohmann, H. Schunck (Eds.), Mathematik – Schl¨ usseltechnologie f¨ ur die Zukunft,
Springer, Berlin, 449–462, 1997.
156 BIBLIOGRAPHY
[300] T.G. Newman, H. Dirilten, A nonlinear transformation for digital picture processing, IEEE
Trans. Computers, Vol. 22, 869–873, 1973.
[301] E. Nicolet, M. Sp¨ uhler, Charakterisierung von unscharfen blobf¨ormigen 3D Strukturen,
diploma thesis, Communication Technology Laboratory, Image Science Division, ETH–
Zentrum, 8092 Z¨ urich, Switzerland, 1996.
[302] M. Nielsen, From paradigm to algorithms in computer vision, Ph.D. thesis, Report DIKU
95/8, Dept. of Computer Science, University of Copenhagen, Universitetsparken 1, 2100
Copenhagen, Denmark, 1995.
[303] M. Nielsen, L. Florack, R. Deriche, Regularization, scalespace and edge detection ﬁlters,
J. Math. Imag. Vision, Vol. 7, 291–307, 1997.
[304] W.J. Niessen, B.M. ter Haar Romeny, L.M.J. Florack, M.A. Viergever, A general frame
work for geometrydriven evolution equations, Int. J. Comput. Vision, Vol. 21, 187–205,
1997.
[305] W.J. Niessen, B.M. ter Haar Romeny, M.A. Viergever, Numerical analysis of geometry
driven diﬀusion equations, B.M. ter Haar Romeny (Ed.), Geometrydriven diﬀusion in
computer vision, Kluwer, Dordrecht, 392–410, 1994.
[306] W.J. Niessen, R. Maas, Multiscale optic ﬂow and stereo, J. Sporring, M. Nielsen, L. Flo
rack, P. Johansen (Eds.), Gaussian scalespace theory, Kluwer, Dordrecht, 31–42, 1997.
[307] W.J. Niessen, K.L. Vincken, J. Weickert, M.A. Viergever, Nonlinear multiscale repre
sentations for image segmentation, Computer Vision and Image Understanding, Vol. 66,
233–245, 1997.
[308] W.J. Niessen, K.L. Vincken, J. Weickert, M.A. Viergever, Threedimensional MR brain
segmentation, Proc. Sixth Int. Conf. on Computer Vision (ICCV ’98, Bombay, Jan. 4–7,
1998), in press.
[309] M. Nitzberg, D. Mumford, T. Shiota, Filtering, segmentation and depth, Lecture Notes in
Computer Science, Vol. 662, Springer, Berlin, 1993.
[310] M. Nitzberg, T. Shiota, Nonlinear image ﬁltering with edge and corner enhancement,
IEEE Trans. Pattern Anal. Mach. Intell., Vol. 14, 826–833, 1992.
[311] N. Nordstr¨om, Biased anisotropic diﬀusion  a uniﬁed regularization and diﬀusion ap
proach to edge detection, Image Vision Comput., Vol. 8, 318–327, 1990.
[312] E. Norman, A discrete analogue of the Weierstrass transform, Proc. Amer. Math. Soc.,
Vol. 11, 596–604, 1960.
[313] O.F. Olsen, Multiscale watershed segmentation, J. Sporring, M. Nielsen, L. Florack, P.
Johansen (Eds.), Gaussian scalespace theory, Kluwer, Dordrecht, 191–200, 1997.
[314] P.J. Olver, G. Sapiro, A. Tannenbaum, Classiﬁcation and uniqueness of invariant geomet
ric ﬂows, C. R. Acad. Sci. Paris, t. 319, S´erie I, 339–344, 1994.
[315] P.J. Olver, G. Sapiro, A. Tannenbaum, Aﬃne invariant detection: edges, active contours,
and segments, Proc. IEEE Comp. Soc. Conf. Computer Vision and Pattern Recognition
(CVPR ’96, San Francisco, June 18–20, 1996), IEEE Computer Society Press, Los Alami
tos, 520–525, 1996.
[316] P.J. Olver, G. Sapiro, A. Tannenbaum, Invariant geometric evolutions of surfaces and
volumetric smoothing, SIAM J. Appl. Math., Vol. 57, 176–194, 1997.
BIBLIOGRAPHY 157
[317] S. Osher, L. Rudin, Featureoriented image enhancement using shock ﬁlters, SIAM J.
Numer. Anal., Vol. 27, 919–949, 1990.
[318] S. Osher, L. Rudin, Shocks and other nonlinear ﬁltering applied to image processing, A.G.
Tescher (Ed.): Applications of digital image processing XIV, SPIE Vol. 1567, 414–431,
1991.
[319] S. Osher, J.A. Sethian, Fronts propagating with curvaturedependent speed: Algorithms
based on Hamilton–Jacobi formulations, J. Comput. Phys., Vol. 79, 12–49, 1988.
[320] N. Otsu, Mathematical studies on feature extraction in pattern recognition, Ph.D. thesis,
Report No. 818, Electrotechnical Laboratory, 114, Umezono, Sakuramura, Niiharigun,
Ibaraki, Japan, 1981 (in Japanese).
[321] K. Ottenberg, Modelbased extraction of geometric structure from digital images, Ph.D.
thesis, Utrecht University, The Netherlands, 1993.
[322] N. Paragios, R. Deriche, A PDEbased levelset approach for detection and tracking of
moving objects, Proc. Sixth Int. Conf. on Computer Vision (ICCV ’98, Bombay, Jan. 4–7,
1998), in press.
[323] E.J. Pauwels, P. Fiddelaers, L.J. Van Gool, Enhancement of planar shape through op
timization of functionals for curves, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 17,
1101–1105, 1995.
[324] E.J. Pauwels, L.J. Van Gool, P. Fiddelaers, T. Moons, An extended class of scaleinvariant
and recursive scale space ﬁlters, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 17, 691–701,
1995.
[325] E. Payot, R. Guillemaud, Y. Trousset, F. Preteux, An adaptive and constrained model for
3D Xray vascular reconstruction, P. Grangeat, J.L. Amans (Eds.), Threedimensional
image reconstruction in radiation and nuclear medicine, Kluwer, Dordrecht, 47–57, 1996.
[326] P. Perona, J. Malik, Scale space and edge detection using anisotropic diﬀusion, Proc. IEEE
Comp. Soc. Workshop on Computer Vision (Miami Beach, Nov. 30 – Dec. 2, 1987), IEEE
Computer Society Press, Washington, 16–22, 1987.
[327] P. Perona, J. Malik, A network for multiscale image segmentation, Proc. IEEE Int. Symp.
Circuits and Systems (ISCAS–88, Espoo, June 7–9, 1988), 2565–2568, 1988.
[328] P. Perona, J. Malik, Scale space and edge detection using anisotropic diﬀusion, IEEE
Trans. Pattern Anal. Mach. Intell., Vol. 12, 629–639, 1990.
[329] P. Perona, T. Shiota, J. Malik, Anisotropic diﬀusion, B.M. ter Haar Romeny (Ed.),
Geometrydriven diﬀusion in computer vision, Kluwer, Dordrecht, 72–92, 1994.
[330] P. Perona, M. Tartagni, Diﬀusion networks for onchip image contrast normalization,
Proc. IEEE Int. Conf. Image Processing (ICIP–94, Austin, Nov. 13–16, 1994), Vol. 1,
IEEE Computer Society Press, Los Alamitos, 1–5, 1994.
[331] I.G. Petrowski, Vorlesungen ¨ uber partielle Diﬀerentialgleichungen, Teubner, Leipzig, 1955.
[332] I. Pollak, A.S. Willsky, H. Krim, Scale space analysis by stabilized inverse diﬀusion equa
tions, B. ter Haar Romeny, L. Florack, J. Koenderink, M. Viergever (Eds.), Scalespace
theory in computer vision, Lecture Notes in Computer Science, Vol. 1252, Springer, Berlin,
200–211, 1997.
[333] E.S. Posmentier, The generation of salinity ﬁnestructure by vertical diﬀusion, J. Phys.
Oceanogr., Vol. 7, 298–300, 1977.
158 BIBLIOGRAPHY
[334] C.B. Price, P. Wambacq, A. Oosterlinck, Applications of reaction–diﬀusion equations to
image processing, IEE Third Int. Conf. Image Processing and its Applications (Warwick,
18–20 July 1989), IEE, London, 49–53, 1989.
[335] C.B. Price, P. Wambacq, A. Oosterlinck, Computing with reaction–diﬀusion systems: ap
plications in image processing, D. Roose et al. (Eds.), Continuation and bifurcations:
Numerical techniques and applications, Kluwer, 379–387, 1990.
[336] M. Proesmans, E.J. Pauwels, L.J. Van Gool, Coupled geometrydriven diﬀusion equations
for lowlevel vision, B.M. ter Haar Romeny (Ed.), Geometrydriven diﬀusion in computer
vision, Kluwer, Dordrecht, 191–228, 1994.
[337] M. Proesmans, E.J. Pauwels, L.J. Van Gool, T. Moons, A. Oosterlinck, Image enhance
ment using nonlinear diﬀusion, Proc. IEEE Comp. Soc. Conf. Computer Vision and
Pattern Recognition (CVPR ’93, New York, June 15–17, 1993), IEEE Computer Society
Press, Los Alamitos, 680–681, 1993.
[338] M. Proesmans, L. Van Gool, E. Pauwels, A. Oosterlinck, Determination of optical ﬂow
and its discontinuities using nonlinear diﬀusion, J.O. Eklundh (Ed.), Computer Vision –
ECCV ’94, Lecture Notes in Computer Science, Vol. 801, Springer, Berlin, 295–304, 1994.
[339] I. Rambaux, P. Gar¸con, Nonlinear anisotropic diﬀusion ﬁltering of 3D images, project
work, D´epartement G´enie Math´ematique, INSA de Rouen and Laboratory of Techno
mathematics, University of Kaiserslautern, 1994.
[340] A.R. Rao, B.G. Schunck, Computing oriented texture ﬁelds, CVGIP: Graphical Models
and Image Processing, Vol. 53, 157–185, 1991.
[341] T. Richardson, S. Mitter, Approximation, computation, and distortion in the variational
formulation, B.M. ter Haar Romeny (Ed.), Geometrydriven diﬀusion in computer vision,
Kluwer, Dordrecht, 169–190, 1994.
[342] J.H. Rieger, Generic evolution of edges on families of diﬀused greyvalue surfaces, J. Math.
Imag. Vision, Vol. 5, 207–217, 1995.
[343] L. Robert, R. Deriche, Dense depth map reconstruction: A minimization and regularization
approach which preserves discontinuities, B. Buxton, R. Cipolla (Eds.), Computer vision
– ECCV ’96, Volume I, Lecture Notes in Computer Science, Vol. 1064, Springer, Berlin,
439–451, 1996.
[344] L.I. Rudin, S. Osher, Total variation based image restoration with free local constraints,
Proc. IEEE Int. Conf. Image Processing (ICIP–94, Austin, Nov. 13–16, 1994), Vol. 1,
IEEE Computer Society Press, Los Alamitos, 31–35, 1994.
[345] L.I. Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal algorithms,
Physica D, Vol. 60, 259–268, 1992.
[346] L.I. Rudin, S. Osher, C. Fu, Total variation based restoration of noisy, blurred images,
Preprint, Cognitech, Inc., 2800 – 28 St., Suite 101, Santa Monica, CA 90405, U.S.A., 1992.
[347] S.J. Ruuth, An algorithm for generating motion by mean curvature, M.O. Berger, R.
Deriche, I. Herlin, J. Jaﬀr´e, J.M. Morel (Eds.), ICAOS ’96: Images, wavelets and PDEs,
Lecture Notes in Control and Information Sciences, Vol. 219, Springer, London, 82–91,
1996.
[348] Y. Saad, Iterative methods for sparse linear systems, PWS Publishing, Boston, 1996.
[349] P. SaintMarc, J.S. Chen, G. Medioni, Adaptive smoothing: a general tool for early vision,
IEEE Trans. Pattern Anal. Mach. Intell., Vol. 13, 514–529, 1990.
BIBLIOGRAPHY 159
[350] A.H. Salden, The nonlinear rescaling process, Report 3DCV 9228, 3D Computer Vision,
AZU, Room E.02.222, Heidelberglaan 100, 3584 CX Utrecht, The Netherlands, 1992.
[351] A.H. Salden, Dynamic scalespace paradigms, Ph.D. thesis, Utrecht University, The
Netherlands, 1996.
[352] A.H. Salden, B.M. ter Haar Romeny, M.A. Viergever, Image structure generating nor
malised geometric scale spaces, Proc. Int. Conf. Volume Image Processing (VIP ’93,
Utrecht, June 2–4, 1993), 141–143, 1993.
[353] A.H. Salden, B.M. ter Haar Romeny, M.A. Viergever, Linear scalespace theory from
physical principles, J. Math. Imag. Vision, to appear.
[354] A.A. Samarskij, Theorie der Diﬀerenzenverfahren, Geest & Portig, Leipzig, 1984.
[355] G.I. S´anchez–Ortiz, D. Rueckert, P. Burger, Knowledgebased anisotropic diﬀusion of
vectorvalued 4dimensional cardiac MR images, R.B. Fisher, E. Trucco (Eds.), Proc.
Seventh British Machine Vision Conference (BMVC 1996, Edinburgh, Sept. 9–12, 1996),
BMVA Press.
[356] G. Sapiro, From active contours to anisotropic diﬀusion: connections between basic PDEs
in image processing, Proc. IEEE Int. Conf. Image Processing (ICIP–96, Lausanne, Sept.
16–19, 1996), Vol. 1, 477–480, 1996.
[357] G. Sapiro, Vector (self ) snakes: a geometric framework for color, texture and multiscale
image segmentation, Proc. IEEE Int. Conf. Image Processing (ICIP–96, Lausanne, Sept.
16–19, 1996), Vol. 1, 817–820, 1996.
[358] G. Sapiro, V. Caselles, Histogram modiﬁcation via diﬀerential equations, J. Diﬀerential
Equations, Vol 135, 238–268, 1997.
[359] G. Sapiro, A. Cohen, A.M. Bruckstein, A subdivision scheme for continuousscale B
splines and aﬃneinvariant progressive smoothing, J. Math. Imag. Vision, Vol. 7, 23–40,
1997.
[360] G. Sapiro, R. Kimmel, D. Shaked, B.B. Kimia, A.M. Bruckstein, Implementing continuous
scale morphology via curve evolution, Pattern Recognition, Vol. 26, 1363–1372, 1993.
[361] G. Sapiro, D.L. Ringach, Anisotropic diﬀusion of multivalued images with applications to
color ﬁltering, IEEE Trans. Image Proc., Vol. 5, 1582–1586, 1996.
[362] G. Sapiro, A. Tannenbaum, Aﬃne invariant scalespace, Int. J. Comput. Vision, Vol. 11,
25–44, 1993.
[363] G. Sapiro, A. Tannenbaum, On aﬃne plane curve evolution, J. Funct. Anal., Vol. 119,
79–120, 1994.
[364] G. Sapiro, A. Tannenbaum, Edge preserving geometric smoothing of MRI data, Technical
Report, Dept. of Electical Engineering, University of Minnesota, Minneapolis, MN 55455,
U.S.A., 1994; submitted to IEEE Trans. Medical Imaging.
[365] G. Sapiro, A. Tannenbaum, Y.L. You, M. Kaveh, Experiments on geometric image en
hancement, Proc. IEEE Int. Conf. Image Processing (ICIP–94, Austin, Nov. 13–16, 1994),
Vol. 2, IEEE Computer Society Press, Los Alamitos, 472–476, 1994.
[366] G. Sapiro, A. Tannenbaum, Area and length preserving geometric invariant scalespaces,
IEEE Trans. Pattern Anal. Mach. Intell., Vol. 17, 67–72, 1995.
[367] C. Schn¨orr, Unique reconstruction of piecewise smooth images by minimizing strictly con
vex nonquadratic functionals, J. Math. Imag. Vision, Vol. 4, 189–198, 1994.
160 BIBLIOGRAPHY
[368] C. Schn¨orr, Segmentation of visual motion by minimizing convex nonquadratic function
als, Proc. 12th Int. Conf. Pattern Recognition (ICPR 12, Jerusalem, Oct. 9–13, 1994),
Vol. A, IEEE Computer Society Press, Los Alamitos, 661–663, 1994.
[369] C. Schn¨orr, Convex variational segmentation of multichannel images, M.O. Berger, R.
Deriche, I. Herlin, J. Jaﬀr´e, J.M. Morel (Eds.), ICAOS ’96: Images, wavelets and PDEs,
Lecture Notes in Control and Information Sciences, Vol. 219, Springer, London, 201–207,
1996.
[370] E. Seneta, Nonnegative matrices and Markov chains, Springer, New York, 1981.
[371] J. Serra, Image analysis and mathematical morphology, Academic Press, London, 1982.
[372] J.A. Sethian, Level set methods, Cambridge University Press, Cambridge, 1996.
[373] J.A. Sethian, A fast marching level set method for monotonically advancing fronts, Proc.
Nat. Acad. Sci., Vol. 93, 1591–1595, 1996.
[374] J.A. Sethian, A review of the theory, algorithms, and applications of level set methods for
propagating interfaces, Acta Numerica, 309–395, 1996.
[375] J. Shah, Segmentation by nonlinear diﬀusion, Proc. IEEE Comp. Soc. Conf. on Computer
Vision and Pattern Recognition (CVPR ’91, Maui, June 3–6, 1992), IEEE Computer
Society Press, Los Alamitos, 202–207, 1991.
[376] J. Shah, Segmentation by nonlinear diﬀusion, II, Proc. IEEE Comp. Soc. Conf. Com
puter Vision and Pattern Recognition (CVPR ’92, Champaign, June 15–18, 1992), IEEE
Computer Society Press, Los Alamitos, 644–647, 1992.
[377] J. Shah, Properties of energyminimizing segmentations, SIAM J. Control and Optimiza
tion, Vol. 30, 99–111, 1992.
[378] J. Shah, A nonlinear diﬀusion model for discontinuous disparity and halfocclusions in
stereo, Proc. IEEE Comp. Soc. Conf. Computer Vision and Pattern Recognition (CVPR
’93, New York, June 15–17, 1993), IEEE Computer Society Press, Los Alamitos, 34–40,
1993.
[379] J. Shah, Piecewise smooth approximations of functions, Calc. Var., Vol. 2, 315–328, 1994.
[380] J. Shah, A common framework for curve evolution, segmentation and anisotropic diﬀusion,
Proc. IEEE Comp. Soc. Conf. Computer Vision and Pattern Recognition (CVPR ’96, San
Francisco, June 18–20, 1996), IEEE Computer Society Press, Los Alamitos, 136–142, 1996.
[381] J. Shen, S. Castan, An optimal linear operator for step edge detection, Graphical Models
and Image Processing, Vol. 54, 112–133, 1992.
[382] A. Sherstinsky, R.W. Picard, Mlattice: From morphogenesis to image processing, IEEE
Trans. Image Proc., Vol. 5, 1137–1150, 1996.
[383] K. Siddiqi, B.B. Kimia, Parts of visual form: computational aspects, IEEE Trans. Pattern
Anal. Mach. Intell., Vol. 17, 239–251, 1995.
[384] K. Siddiqi, B.B. Kimia, A shock grammar for recognition, Proc. IEEE Comp. Soc. Conf.
Computer Vision and Pattern Recognition (CVPR ’96, San Francisco, June 18–20, 1996),
IEEE Computer Society Press, Los Alamitos, 507–513, 1996.
[385] K. Siddiqi, B.B. Kimia, C.W. Shu, Geometric shockcapturing ENO schemes for subpixel
interpolation, computation and curve evolution, Graphical Models and Image Processing,
in press.
BIBLIOGRAPHY 161
[386] J. Sijbers, P. Scheunders, M. Verhoye, A. Van der Linden, D. Van Dyck, E. Raman,
Watershedbased segmentation of 3D MR data for volume quantization, Magnetic Reso
nance Imaging, Vol. 15, 679–688, 1997.
[387] W. Snyder, Y.S. Han, G. Bilbro, R. Whitaker, S. Pizer, Image relaxation: Restoration
and feature extraction, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 17, 620–624, 1995.
[388] J. Sporring, The entropy of scalespace, Proc. 13th Int. Conf. Pattern Recognition (ICPR
13, Vienna, Aug. 25–30, 1996), Vol. A, 900–904, 1996.
[389] J. Sporring, M. Nielsen, L. Florack, P. Johansen (Eds.), Gaussian scalespace theory,
Kluwer, Dordrecht, 1997.
[390] J. Sporring, J. Weickert, On generalized entropies and scalespace, B. ter Haar Romeny,
L. Florack, J. Koenderink, M. Viergever (Eds.), Scalespace theory in computer vision,
Lecture Notes in Computer Science, Vol. 1252, Springer, Berlin, 53–64, 1997.
[391] R. Sprengel, Entwurf und Analyse nichtlinearer Diﬀusionsverfahren f¨ ur die Bildverar
beitung, Ph.D. thesis, Dept. of Computer Science, University of Hamburg, Germany, 1995;
DISKI 123, Inﬁx Verlag, Sankt Augustin, 1996.
[392] J.L. Stansﬁeld, Conclusions from the commodity expert project, A.I. Memo 671, Artiﬁcial
Intelligence Lab., Massachusetts Inst. of Technology, Cambridge, MA, U.S.A., 1980.
[393] E. Steen, B. Olstad, Scalespace and boundary detection in ultrasonic imaging using non
linear signaladaptive anisotropic diﬀusion, M.H. Loew (Ed.), Image processing, SPIE Vol.
2167, 116–127, 1994.
[394] A. Steiner, R. Kimmel, A.M. Bruckstein, Planar shape enhancement and exaggeration,
Proc. 13th Int. Conf. Pattern Recognition (ICPR 13, Vienna, Aug. 25–30, 1996), Vol. A,
523–527, 1996.
[395] P. Stoll, C.W. Shu, B.B. Kimia, Shockcapturing numerical methods for viscosity solutions
of certain PDEs in computer vision: The Godunov, Osher–Sethian and ENO schemes,
Technical Report LEMS132, Division of Engineering, Brown University, Providence, R.I.
02912, U.S.A., 1994.
[396] D.M. Strong, T.F. Chan, Relation of regularization parameter and scale in total variation
based image denoising, CAM Report 967, Dept. of Mathematics, Univ. of California, Los
Angeles, CA 90024, U.S.A., 1996.
[397] D.M. Strong, T.F. Chan, Spatially and scale adaptive total variation based regularization
and anisotropic diﬀusion in image processing, CAM Report 9646, Dept. of Mathematics,
Univ. of California, Los Angeles, CA 90024, U.S.A., 1996.
[398] S. Teboul, L. Blanc–F´eraud, G. Aubert, M. Barlaud, Variational approach for edge
preserving regularization using coupled PDEs, IEEE Trans. Image Proc., March 1998,
in press.
[399] H. Tek, B.B. Kimia, Image segmentation by reaction–diﬀusion bubbles, Proc. Fifth Int.
Conf. on Computer Vision (ICCV ’95, Cambridge, June 20–23, 1995), IEEE Computer
Society Press, Los Alamitos, 156–162, 1995.
[400] H. Tek, B.B. Kimia, Volumetric segmentation of medical images by threedimensional
bubbles, Computer Vision and Image Understanding, Vol. 65, 246–258, 1997.
[401] D. Terzopoulos, A. Witkin, M. Kass, Constraints on deformable models: Recovering 3D
shape and nonrigid motion, Artif. Intell., Vol. 36, 91–123, 1988.
162 BIBLIOGRAPHY
[402] A. Tikhonov, V. Arsenin, Solutions of illposed problems, Wiley, New York, 1977.
[403] F. Torkamani–Azar, K.E. Tait, Image recovery using the anisotropic diﬀusion equation,
IEEE Trans. Image Proc., Vol. 5, 1573–1578, 1996.
[404] V. Torre, T.A. Poggio, On edge detection, IEEE Trans. Pattern Anal. Mach. Intell., Vol.
8, 148–163, 1986.
[405] A.M. Turing, The chemical basis of morphogenesis, Philosoph. Trans. Royal Soc. London
B, Vol. 237, 37–72, 1952.
[406] G. Turk, Generating textures on arbitrary surfaces using reaction–diﬀusion, Computer
Graphics, Vol. 25, No. 4, 289–298, 1991.
[407] R.A. Varga, Matrix iterative analysis, Prentice Hall, Englewood Cliﬀs, 1962.
[408] K.L. Vincken, A.S.E. Koster, M.A. Viergever, Probabilistic multiscale image segmentation,
IEEE Trans. Pattern Anal. Mach. Intell., Vol. 19, 109–120, 1997.
[409] C.R. Vogel, M.E. Oman, Iterative methods for total variation denoising, SIAM J. Sci.
Comput., Vol. 17, 227–238, 1996.
[410] C.R. Vogel, M.E. Oman, Fast numerical methods for total variation minimization in image
reconstruction, F.T. Luk (Ed.), Advanced signal processing algorithms, SPIE Vol. 2563,
359–367, 1995.
[411] W. Walter, Diﬀerential and integral inequalities, Springer, Berlin, 1970.
[412] W. Walter, Gew¨ohnliche Diﬀerentialgleichungen, Springer, Berlin, 1976.
[413] J. Weickert, Anisotropic diﬀusion ﬁlters for image processing based quality control, A.
Fasano, M. Primicerio (Eds.), Proc. Seventh European Conf. on Mathematics in Industry,
Teubner, Stuttgart, 355–362, 1994.
[414] J. Weickert, Scalespace properties of nonlinear diﬀusion ﬁltering with a diﬀusion tensor,
Report No. 110, Laboratory of Technomathematics, University of Kaiserslautern, P.O.
Box 3049, 67653 Kaiserslautern, Germany, 1994.
[415] J. Weickert, Theoretical foundations of anisotropic diﬀusion in image processing, Com
puting, Suppl. 11, 221–236, 1996.
[416] J. Weickert, Anisotropic diﬀusion in image processing, Ph.D. thesis, Dept. of Mathematics,
University of Kaiserslautern, P.O. Box 3049, 67653 Kaiserslautern, Germany, 1996.
[417] J. Weickert, A model for the cloudiness of fabrics, H. Neunzert (Ed.), Progress in industrial
mathematics at ECMI 94, Wiley–Teubner, Chichester, 258–265, 1996.
[418] J. Weickert, Multiscale texture enhancement, V. Hlav´aˇc, R.
ˇ
S´ara (Eds.), Computer analysis
of images and patterns, Lecture Notes in Computer Science, Vol. 970, Springer, Berlin,
230–237, 1995.
[419] J. Weickert, Foundations and applications of nonlinear anisotropic diﬀusion ﬁltering, Z.
Angew. Math. Mech., Vol. 76, Suppl. 1, 283–286, 1996.
[420] J. Weickert, Nonlinear diﬀusion scalespaces: From the continuous to the discrete setting,
M.O. Berger, R. Deriche, I. Herlin, J. Jaﬀr´e, J.M. Morel (Eds.), ICAOS ’96: Images,
wavelets and PDEs, Lecture Notes in Control and Information Sciences, Vol. 219, Springer,
London, 111–118, 1996.
[421] J. Weickert, Nonlinear scalespaces, J. Sporring, M. Nielsen, L. Florack, P. Johansen
(Eds.), Gaussian scalespace theory, Kluwer, Dordrecht, 221–234, 1997.
BIBLIOGRAPHY 163
[422] J. Weickert, Coherenceenhancing diﬀusion of colour images, A. Sanfeliu, J.J. Villanueva,
J. Vitri`a (Eds.), Pattern Recognition and Image Analysis (VII NSPRIA, Barcelona, April
21–25, 1997), Vol. 1, 239–244, 1997.
[423] J. Weickert, A review of nonlinear diﬀusion ﬁltering, B. ter Haar Romeny, L. Florack, J.
Koenderink, M. Viergever (Eds.), Scalespace theory in computer vision, Lecture Notes
in Computer Science, Vol. 1252, Springer, Berlin, 3–28, 1997.
[424] J. Weickert, Recursive separable schemes for nonlinear diﬀusion ﬁlters, B. ter Haar
Romeny, L. Florack, J. Koenderink, M. Viergever (Eds.), Scalespace theory in computer
vision, Lecture Notes in Computer Science, Vol. 1252, Springer, Berlin, 260–271, 1997.
[425] J. Weickert, B. Benhamouda, A semidiscrete nonlinear scalespace theory and its relation
to the Perona–Malik paradox, F. Solina, W.G. Kropatsch, R. Klette, R. Bajcsy (Eds.),
Theoretical foundations of computer vision (TFCV ’96, Dagstuhl, March 18–22, 1996),
Springer, Wien, in press.
[426] J. Weickert, S. Ishikawa, A. Imiya, On the history of Gaussian scalespace axiomatics, J.
Sporring, M. Nielsen, L. Florack, P. Johansen (Eds.), Gaussian scalespace theory, Kluwer,
Dordrecht, 45–59, 1997.
[427] J. Weickert, S. Ishikawa, A. Imiya, Scalespace has been discovered in Japan, Report DIKU
97/18, Dept. of Computer Science, University of Copenhagen, Universitetsparken 1, 2100
Copenhagen, Denmark, 1997, submitted.
[428] J. Weickert, B.M. ter Haar Romeny, A. Lopez, W.J. van Enk, Orientation analysis by
coherenceenhancing diﬀusion, Proc. Symp. Real World Computing (RWC ’97, Tokyo,
Jan. 29–31, 1997), 96–103, 1997.
[429] J. Weickert, B.M. ter Haar Romeny, M.A. Viergever, Conservative image transformations
with restoration and scalespace properties, Proc. IEEE Int. Conf. Image Processing (ICIP–
96, Lausanne, Sept. 16–19, 1996), Vol. 1, 465–468, 1996.
[430] J. Weickert, B.M. ter Haar Romeny, M.A. Viergever, Eﬃcient and reliable schemes for
nonlinear diﬀusion ﬁltering, IEEE Trans. Image Proc., March 1998, to appear.
[431] J. Weickert, K.J. Zuiderveld, B.M. ter Haar Romeny, W.J. Niessen, Parallel implementa
tions of AOS schemes: A fast way of nonlinear diﬀusion ﬁltering, Proc. 1997 IEEE Inter
national Conference on Image Processing (ICIP–97, Santa Barbara, Oct. 26–29, 1997), to
appear.
[432] H. Werner, H. Arndt, Gew¨ohnliche Diﬀerentialgleichungen, Springer, Berlin, 1986.
[433] R.T. Whitaker, Geometry limited diﬀusion in the characterization of geometric patches in
images, CVGIP: Image Understanding, Vol. 57, 111–120, 1993.
[434] R.T. Whitaker, Characterizing ﬁrst and secondorder patches using geometrylimited dif
fusion, H.H. Barrett, A.F. Gmitro (Eds.), Information processing in medical imaging,
Lecture Notes in Computer Science, Vol. 687, Springer, Berlin, 149–167, 1993.
[435] R.T. Whitaker, Algorithms for implicit deformable models, Proc. Fifth Int. Conf. on Com
puter Vision (ICCV ’95, Cambridge, June 20–23, 1995), IEEE Computer Society Press,
Los Alamitos, 822–827, 1995.
[436] R.T. Whitaker, D.T. Chen, Embedded active surfaces for volume visualization, M.H. Loew
(Ed.), Image processing, SPIE Vol. 2167, 340–352, 1994.
[437] R. Whitaker, G. Gerig, Vectorvalued diﬀusion, B.M. ter Haar Romeny (Ed.), Geometry
driven diﬀusion in computer vision, Kluwer, Dordrecht, 93–134, 1994.
164 BIBLIOGRAPHY
[438] R.T. Whitaker, S.M. Pizer, A multiscale approach to nonuniform diﬀusion, CVGIP: Im
age Understanding, Vol. 57, 99–110, 1993.
[439] A.P. Witkin, Scalespace ﬁltering, Proc. Eighth Int. Joint Conf. on Artiﬁcial Intelligence
(IJCAI ’83, Karlsruhe, Aug. 8–12, 1983), Vol. 2, 1019–1022, 1983.
[440] A. Witkin, M. Kass, Reaction–diﬀusion textures, Computer Graphics, Vol. 25, No. 4,
299–308, 1991.
[441] A. Witkin, D. Terzopoulos, M. Kass, Signal matching through scalespace, Int. J. Comput.
Vision, Vol. 1, 134–144, 1987.
[442] N.N. Yanenko, The method of fractional steps: the solution of problems of mathematical
physics in several variables, Springer, New York, 1971.
[443] G.Z. Yang, P. Burger, D.N. Firmin, S.R. Underwood, Structure adaptive anisotropic ﬁl
tering, Image Vision Comput., Vol. 14, 135–145, 1996.
[444] T.S. Yoo, J.M. Coggins, Using statistical pattern recognition techniques to control vari
able conductance diﬀusion, H.H. Barrett, A.F. Gmitro (Eds.), Information processing in
medical imaging, Lecture Notes in Computer Science, Vol. 687, Springer, Berlin, 459–471,
1993.
[445] Y.L. You, M. Kaveh, Anisotropic blind image restoration, Proc. IEEE Int. Conf. Image
Processing (ICIP–96, Lausanne, Sept. 16–19, 1996), Vol. 2, 461–464, 1996.
[446] Y.L. You, W. Xu, A. Tannenbaum, M. Kaveh, Behavioral analysis of anisotropic diﬀusion
in image processing, IEEE Trans. Image Proc., Vol. 5, 1539–1553, 1996.
[447] D.M. Young, Iterative solution of large linear systems, Academic Press, New York, 1971.
[448] I.T. Young, L.J. van Vliet, Recursive implementation of the Gaussian ﬁlter, Signal Pro
cessing, Vol. 44, 139–151, 1995.
[449] R.A. Young, The Gaussian derivative model for spatial vision: I. Retinal mechanisms,
Spatial Vision, Vol. 2, 273–293, 1987.
[450] A.L. Yuille, T.A. Poggio, Scaling theorems for zero crossings, IEEE Trans. Pattern Anal.
Mach. Intell., Vol. 8, 15–25, 1986.
[451] D. Zhao, B. Li, A new implementation of discrete multiscale ﬁltering, Proc. IEEE Int.
Conf. Image Processing (ICIP–96, Lausanne, Sept. 16–19, 1996), Vol. 1, 383–386, 1996.
[452] S.C. Zhu, A.L. Yuille, Region competition: unifying snakes, region growing, en
ergy/Bayes/MDL for multiband image segmentation, IEEE Trans. Pattern Anal. Mach.
Intell., Vol. 18, 884–900, 1996.
[453] S.W. Zucker, A. Dobbins, L. Iverson, B. Kimia, A. Tannenbaum, From curve detection
to shape description: an outline, A. Basu, X. Li (Eds.), Computer vision: systems, theory
and applications, World Scientiﬁc, Singapore, 25–39, 1993.
Index
apriori estimates, 93, 105, 110
apriori knowledge, 12–14, 119, 135
active
blobs, 46
contour models, 44–49
surfaces, 46
adaptive mesh coarsening, 25
adaptive smoothing, 14
additive operator splitting, 26, 107–111
aﬃne
arclength, see arclength
Gaussian scalespace, 13
invariance, see invariance
invariant geodesic snakes, 48
invariant gradient, 45
invariant heat ﬂow, 40
invariant texture segmentation, 29
morphological scalespace, 40–45, 113,
124
Mumford–Shah functional, 29
perimeter, see perimeter
shortening ﬂow, see shortening ﬂow
transformation, see transformation
AMSS, see aﬃne morphological scalespace
anisotropic diﬀusion, see diﬀusion
AOS, see additive operator splitting
applications, 11, 13, 26, 28, 30, 37, 44, 52,
114–126, 129–134
arclength
aﬃne, 40
Euclidean, 38
areapreserving ﬂows, 42
average globality, 74
average grey level invariance, see invariance
axiomatics, see scalespace
balloon force, 47, 48
basic equation of ﬁgure, 13
bias term, 28, 49
blind image restoration, 27, 52
blob detection, 5, 45
BUC, 34, 61
CAD, 37
Canny edge detector, 5, 15
causality, 5, 34, 65
CED, see diﬀusion, coherenceenhancing
central moments, 73, 85, 102
circular point, 39
closing, 32
coherence, 57, 128
coherence direction, 57, 128
coherenceenhancing diﬀusion, see diﬀusion
colour images, see vectorvalued images
commuting operators, 62
comparison principle, 20, 61, 135
computer aided quality control, 26, 125–126,
130
condition number, 88, 95, 106
conjugate gradient methods, 106
continuity, 97
continuity equation, 2
continuous dependence, see wellposedness
continuous diﬀusion, see diﬀusion
contrast enhancement, 15, 40, 53, 61, 62, 72,
113–126
contrast parameter, 16
convergence, 14, 19, 21, 22, 26, 28, 30, 50,
52, 67, 76, 82, 98, 99, 106, 116, 118,
124, 134
convolution theorem, 3, 10
corners, 24, 28, 29, 45, 56, 57, 116
correspondence problem, 12, 19, 53, 116, 124
CPU times, 113
crystal growth, 38
curvature
of a curve, 33
of an image, 33
curve evolution, 33, 38–43, 46, 47
curve evolution schemes, 36, 43
deblurring, 5, 50, 52
165
166 INDEX
deep structure, 11, 136
defect detection, see computer aided quality
control
defocusing, 7
deviation cost, 27, 29
diﬀerencesofGaussians, 5
diﬀerential inequalities, 17
diﬀerential invariants, 5
diﬀusion
anisotropic, 2, 13, 22, 24, 55–74, 87–95,
111, 113–137
backward, 5, 50, 116
coherenceenhancing, 24, 127–134
continuous, 55–74
directed, 14
discrete, 97–111
edgeenhancing, 23, 113–126
equation, 2
forward–backward, 15, 45, 66, 72
homogeneous, 2
inhomogeneous, 2, 117
isotropic, 2–22, 86–87, 110, 116
linear, 2–14, 36, 116, 125
nonlinear, 2, 14–27, 52, 55–137
on a sphere, 12
physical background, 2
semidiscrete, 75–95
tensor, 2, 13, 22, 58, 62, 88, 95, 114, 127,
135
diﬀusion–reaction, 27–30, 118
single equations, 27
systems, 28
diﬀusivity, 2, 15, 16, 47, 48, 116
digitally scalable, 36
dilation, 32, 35, 47, 63, 113, 119
direction, 56
directional diﬀusivities, 89, 94
directional splitting, 88–95, 107–111
discrete analogue
AMSS, MCM, 43
Gaussian, 10
linear diﬀusion, 10
nonlinear diﬀusion, 25, 97–111
discrete diﬀusion, see diﬀusion
dissipative eﬀects, 37, 43, 44
distance transformation, 37, 47
DoGs, see diﬀerencesofGaussians
doubly stochastic matrix, 98
edge cost, 27, 29
edge detection, 5, 15, 24, 28, 30, 50, 53, 55,
57
edge enhancement, see contrast enhancement
edgeenhancing diﬀusion, see diﬀusion
elliptic point, 41
energy, 72, 85, 102
energy functional
explicit snakes, 46
for curves, 28, 29
geodesic snakes, 48
Mumford–Shah, 28, 44, 126
Nordstr¨om, 27, 30, 51
Perona–Malik, 18
Polyakov, 42
TVminimization, 51
ENO schemes, 37
entropy, 11, 73, 85, 102
generalized, 11, 73
entropy scalespace, see scalespace, reaction–
diﬀusion
erosion, 32, 47, 63
Euclidean
arclength, see arclength
perimeter, see perimeter
shortening ﬂow, see shortening ﬂow
transformation, see transformation
Eulerian formulation, 33
existence, see wellposedness
expected value, 73
explicit schemes, 11, 19, 25, 43, 53, 103, 105,
106, 113, 137
external energy, 46
extremum principle, 4, 6, 18–20, 30, 34, 39,
41, 44, 50, 58, 62, 66, 76, 77, 87, 93,
98, 105
fabrics, 125, 130
FD, see ﬁnite diﬀerences
feature vectors, 24
Fick’s law, 2
ﬁlter design, 114, 127
ﬁngerprint enhancement, 26, 129
ﬁnite diﬀerences, 11, 25, 30, 43, 46, 50, 51,
85–95, 102–111, 113
ﬁnite elements, 25, 28
ﬂame propagation, 38
ﬂowline, 16
ﬂux, 2, 15, 22, 114
focusofattention, 11
forensic applications, 52
INDEX 167
Fourier transformation, 3, 10, 35, 44
free discontinuity problems, 29
fundamental equation in image processing,
41, 42
Gabor ﬁlters, 45, 136
Gabor’s restoration method, 45
gas dynamics, 49
gauge coordinates, 16
Gaussian
derivatives, 4, 43, 44
elimination, 106
kernel, 3, 35
pyramid, 11
scalespace, see scalespace
smoothing, 2–14, 56
Gauß–Seidel method, 44, 106
generalized entropy, see entropy
generalized ﬁgure, 13
geodesic, 48
geodesic curvature ﬂow, 42
geometric heat equation, 38, 39
grassﬁre ﬂow, 33
grey level shift invariance, see invariance
greyscale invariance, see invariance
halftoning, 30, 37
hardware realizations, 11, 13, 25, 33
heat equation, 2
histogramme enhancement, 28, 45
Huygens principle, 33
hyperstack, 63
hysteresis thresholding, 5, 15
Iijima’s axiomatic, 7
illposedness, 4, 17, 28–30, 49, 53
image
colour, see vectorvalued images
discrete, 75
greyscale, 3, 55
higher dimensional, see threedimensional
images
vectorvalued, see vectorvalued images
image compression, 73, 137
image enhancement, 14–30, 45, 48–53, 55–
74, 113–134
image restoration, see image enhancement
image sequences, 11, 24, 30, 47, 48, 56, 62
immediate localization, 19
implicit schemes, 11, 44, 137
information theory, 73
integral geometry, 10
integration scale, 57, 114, 127, 132
interest operator, see structure tensor
internal energy, 46
intrinsic heat ﬂows, 39, 40, 42
invariance
aﬃne, 42–45
average grey level, 13, 63, 76, 81, 98, 99,
119
grey level shift, 62, 81
greyscale, 31, 43, 45, 124
isometry, 64, 81
morphological, see invariance, greyscale
reverse contrast, 63, 81
rotational, 13, 21, 35, 43
scale, 8
translation, 8, 64, 81
irreducibility, 75, 76, 97
isometry invariance, see invariance
isophote, 16
Japanese scalespace research, 7, 13
jet space, 24
junctions, 5, 29, 42, 56
Kramer–Bruckner ﬁlter, 50
LaplacianofGaussian, 5
lengthpreserving ﬂows, 42
level set, 31, 44
level set methods, 36
line detection, 26
linear diﬀusion, see diﬀusion
linear systems, 106, 107, 110
linearity, 8
Lipschitzcontinuity, 76
local extrema
creation, 12, 43, 65
noncreation, 34, 39, 119
nonenhancement, 65
local scale, 57
LoG, see LaplacianofGaussian
Lyapunov
functionals, 66–74, 135
functions, 76, 82–85
sequences, 98–102
Markov chains, 98
Markov processes, 11
168 INDEX
Markov random ﬁelds, 27, 29
Marr–Hildreth operator, 5
maximal monotone operator, 17
maximum–minimum principle, see extremum
principle
MCM, see mean curvature motion
mean curvature motion, 37–40, 42–45, 47,
50, 51, 93, 113, 119, 125, 129
backward, 45
mean ﬁeld annealing, 27
median ﬁltering, 38
medical imaging, 27, 28, 45, 46, 52, 62, 63,
111, 116–126, 131
MegaWave, 44, 116, 126
min–max ﬂow, 45
minimal surfaces, 29, 38
monotonicity preservation, 19
morphological invariance, see invariance, grey
scale
morphology, 30–49, 119–124
binary, 31
classical, 30–37
continuousscale, 32
curvaturebased, 37–49
greyscale, 31
multigrid, 11, 25, 106, 137
multiplicative operator splittings, 110
Mumford–Shah functional, see energy func
tional
neural networks, 25, 137
noise elimination, 12, 21, 23, 45, 116, 124
noise scale, 21, 57
nonmaxima suppression, 5, 15
noncreation of local extrema, see local ex
trema
nonenhancement of local extrema, see local
extrema
nonlinear diﬀusion, see diﬀusion
nonnegativity, 76, 88, 97, 107
Nordstr¨om functional, see energy function
als
numerical aspects, 10, 25, 36, 43, 51, 85–95,
102–111, 113
oceanography, 18
ODE, see ordinary diﬀerential equations
opening, 32
optic ﬂow, 11, 28, 52
optical character recognition, 13
optimization
convex, 28
nonconvex, 53
ordinary diﬀerential equations, 75–95
orientation, 56, 129
oscillations, 10, 19, 21, 44, 50
Osher–Sethian schemes, 36, 113
oversegmentation, 29
parallel computing, 25, 26, 30, 110, 111
parameter determination, 17, 46, 114, 119,
129, 137
partial diﬀerential equations, 1–53
particle methods, 27
path planning, 37
PDE, see partial diﬀerential equations
perceptual grouping, 136
perimeter
aﬃne, 41
Euclidean, 39
Perona–Malik ﬁlter, 14–20, 28, 30, 45, 49,
114
basic idea, 15
edge enhancement, 15
illposedness, 17
regularizations, 20–24
scalespace properties, 19
semidiscrete, 87
piecewise smoothing, 19
pixel numbering, 87
positivity, 97
positivity preservation, 8
postprocessing, 26
prairie ﬂow, 33
predictor–corrector schemes, 107, 109
preprocessing, 118, 126
pseudospectral methods, 25
pyramids, 11, 25, 125
quantization, 45
random variable, 72
reaction–diﬀusion bubbles, 48
reaction–diﬀusion scalespace, see scalespace
recursive ﬁlters, 10, 11, 110, 111
recursivity, 6
redistribution property, 99
region growing, 29, 126
regularity, 58, 61
INDEX 169
regularization, 4, 10, 49, 50, 52, 62, 72, 93,
114
remote sensing, 27
rescaling, 20
reverse contrast invariance, see invariance
robust statistics, 27
row sum, 76, 97
scale invariance, see invariance
scale selection, 11
scale–imprecision space, 6
scalespace
aﬃne Gaussian, 13
aﬃne morphological, 41
anisotropic diﬀusion, 62–74
architectural principles, 6
axioms, 6, 34, 42, 53
dilation–erosion, 34
discrete nonlinear diﬀusion, 25, 99–111
for curves, 39, 42
for image sequences, 11, 43
for shapes, 42
Gaussian, 7
general concept, 6
invariance principles, 7
linear diﬀusion, 7, 116
linearity principle, 7
mean curvature, 39
morphological equivalent of Gaussian scale
space, 35
nonlinear diﬀusion, 62–74, 118
projective, 42
reaction–diﬀusion, 45
semidiscrete linear, 10
semidiscrete nonlinear diﬀusion, 25, 81–
95
smoothing principles, 6, 62
scatter matrix, see structure tensor
second moment matrix, see structure tensor
segmentation, 11, 21, 27, 44, 53, 63, 118, 126,
137
selfsnakes, see snakes
semiimplicit schemes, 25, 44, 50, 102–111,
113
semidiscrete, 10
semidiscrete analogue
linear diﬀusion, 10
nonlinear diﬀusion, 25, 75–95
semidiscrete diﬀusion, see diﬀusion
semigroup property, 6, 8, 34
separability, 10, 35, 36, 110
settheoretic schemes, 36, 43
shape, 31
cues, 56
exaggeration, 45
oﬀset, 37
recognition, 45, 124
segmentation, 45
shape inclusion principle, 39
shapeadapted Gaussian smoothing, 13, 22
shapefromshading, 37
shock
capturing schemes, 37, 49
creation, 34, 50
ﬁlters, 50
shortening ﬂow
aﬃne, 41
Euclidean, 38
silhouette, 31
simulated annealing, 29
skeletonization, 37
slope transform, 35
smoothness, 58, 76, 97, 99, 128
snakes, 46
explicit, 46
geodesic, 48
implicit, 47
selfsnakes, 48, 124
software, 27, 44, 116, 137
SOR, 106
stability, 25, 43, 44, 106, 109
stabilizing cost, 27, 29
staircasing eﬀect, 18, 29, 52
Stampacchia’s truncation method, 59
steadystate, see convergence
steerable ﬁlters, 136
stencil size, 88–95
stereo, 11, 28, 30, 48, 62
stochastic matrix, 98
stopping time, 4, 28, 39, 41, 47, 74, 116
string theory, 42
structure tensor, 56, 86, 114, 136
structuring element, 32
structuring function, 35
subpixel accuracy, 36
subsampling, 26, 73
symmetry, 58, 76, 97
target tracking, 27
terrain matching, 45
170 INDEX
texture
analysis, 56
discrimination, 45, 73
enhancement, 26, 45, 127
generation, 30
segmentation, 24, 27, 29
threedimensional images, 24, 43, 46–48, 136
topological changes, 43, 46, 47
toppoint, 12
total variation, 49
minimizing methods, 50
preserving methods, 49
transformation
aﬃne, 40
Euclidean, 38
projective, 42
translation invariance, see invariance
Turing’s pattern formation model, 30
TV, see total variation
uniform positive deﬁniteness, 58, 62, 76, 93,
129
uniqueness, see wellposedness
van Gogh, 131
variance, 73
vectorvalued images, 24, 28, 30, 42, 43, 49,
52, 136
viscosity solution, 17, 34, 41, 45, 47–49, 61
visual system, 5, 15, 33
wavelets, 11, 25, 44, 136
weak membrane model, 29
wellposedness, 4, 20, 23, 28, 29, 34, 39, 41,
47–49, 52, 53, 58, 62, 76, 77, 98, 135
wood surfaces, 125
zerocrossings, 5, 34, 50
Anisotropic Diﬀusion
in Image Processing
Joachim Weickert
University of Copenhagen
Many recent techniques for digital image enhancement and multiscale image rep
resentations are based on nonlinear partial diﬀerential equations.
This book gives an introduction to the main ideas behind these methods, and
it describes in a systematic way their theoretical foundations, numerical aspects,
and applications. A large number of references enables the reader to acquire an
uptodate overview of the original literature.
The central emphasis is on anisotropic nonlinear diﬀusion ﬁlters. Their ﬂexibil
ity allows to combine smoothing properties with image enhancement qualities. A
general framework is explored covering wellposedness and scalespace results not
only for the continuous, but also for the algorithmically important semidiscrete and
fully discrete settings. The presented examples range from applications in medical
image analysis to problems in computer aided quality control.
B.G. Teubner Stuttgart
Anisotropic Diﬀusion in Image Processing
Joachim Weickert
B.G. Teubner Stuttgart
Anisotropic Diﬀusion in Image Processing
Joachim Weickert
Department of Computer Science University of Copenhagen Copenhagen, Denmark
B.G. Teubner Stuttgart 1998
c Copyright 2008 by Joachim Weickert. semidiscrete. The goal was to visualize the quality relevant adjacent ﬁbre structures. . or translated into a machine language without the written permission of the author.2 for more details). socalled stripes. 1991 M. Joachim Weickert Born in 1965 in Ludwigshafen/Germany. in mathematics. nat.D. physics and computer science at the University of Kaiserslautern. The copyright has been returned to the author in 2008. The displayed equations describe the basic structure of nonlinear diﬀusion ﬁltering in the continuous. The cover image shows a thresholded nonwoven fabric image which was processed by applying a coherenceenhancing anisotropic diﬀusion ﬁlter (see Section 5.Sc. Teubner (Stuttgart) in 1998 and went out of print in 2001. No part of this book may be reproduced by any means. in physics and industrial mathematics. 1996 Ph. In the current version a few typos and other errors have been corrected. All rights reserved. rer.Dr. Since then visiting assistant research professor at the Department of Computer Science. Their theoretical foundations are treated in Chapters 2–4. Studies in mathematics.Sc. G. This book had been published by B. 1987 B. in industrial mathematics. or transmitted. and fully discrete setting. Copenhagen University. Postdoctoral researcher at the Image Sciences Institute at Utrecht University from 2/96 to 3/97.
To my parents Gerda and Norbert .
Preface
Partial diﬀerential equations (PDEs) have led to an entire new ﬁeld in image processing and computer vision. Hundreds of publications have appeared in the last decade, and PDEbased methods have played a central role at several conferences and workshops. The success of these techniques is not really surprising, since PDEs have proved their usefulness in areas such as physics and engineering sciences for a very long time. In image processing and computer vision, they oﬀer several advantages: • Deep mathematical results with respect to wellposedness are available, such that stable algorithms can be found. PDEbased methods are one of the mathematically bestfounded techniques in image processing. • They allow a reinterpretation of several classical methods under a novel unifying framework. This includes many wellknown techniques such as Gaussian convolution, median ﬁltering, dilation or erosion. • This understanding has also led to the discovery of new methods. They can oﬀer more invariances than classical techniques, or describe novel ways of shape simpliﬁcation, structure preserving ﬁltering, and enhancement of coherent linelike structures. • The PDE formulation is genuinely continuous. Thus, their approximations aim to be independent of the underlying grid and may reveal good rotational invariance. PDEbased image processing techniques are mainly used for smoothing and restoration purposes. Many evolution equations for restoring images can be derived as gradient descent methods for minimizing a suitable energy functional, and the restored image is given by the steadystate of this process. Typical PDE techniques for image smoothing regard the original image as initial state of a parabolic (diﬀusionlike) process, and extract ﬁltered versions from its temporal evolution. The whole evolution can be regarded as a socalled scalespace, an embedding of the original image into a family of subsequently simpler, more global representations of it. Since this introduces a hierarchy into the image structures, one can use a scalespace representation for extracting semantically important information. One of the two goals of this book is to give an overview of the stateoftheart of PDEbased methods for image enhancement and smoothing. Emphasis is put on a v
vi
PREFACE
uniﬁed description of the underlying ideas, theoretical results, numerical approximations, generalizations and applications, but also historical remarks and pointers to open questions can be found. Although being concise, this part covers a broad spectrum: it includes for instance an early Japanese scalespace axiomatic, the Mumford–Shah functional for image segmentation, continuousscale morphology, active contour models and shock ﬁlters. Many references are given which point the reader to useful original literature for a task at hand. The second goal of this book is to present an indepth treatment of an interesting class of parabolic equations which may bridge the gap between scalespace and restoration ideas: nonlinear diﬀusion ﬁlters. Methods of this type have been proposed for the ﬁrst time by Perona and Malik in 1987 [326]. In order to smooth an image and to simultaneously enhance important features such as edges, they apply a diﬀusion process whose diﬀusivity is steered by derivatives of the evolving image. These ﬁlters are diﬃcult to analyse mathematically, as they may act locally like a backward diﬀusion process. This gives rise to wellposedness questions. On the other hand, nonlinear diﬀusion ﬁlters are frequently applied with very impressive results; so there appears the need for a theoretical foundation. We shall develop results in this direction by investigating a general class of nonlinear diﬀusion processes. This class comprises linear diﬀusion ﬁlters as well as spatial regularizations of the Perona–Malik process, but it also allows processes which replace the scalar diﬀusivity by a diﬀusion tensor. Thus, the diﬀusive ﬂux does not have to be parallel to the grey value gradient: the ﬁlters may become anisotropic. Anisotropic diﬀusion ﬁlters can outperform isotropic ones with respect to certain applications such as denoising of highly degraded edges or enhancing coherent ﬂowlike images by closing interrupted onedimensional structures. In order to establish wellposedness and scalespace properties for this class, we shall investigate existence, uniqueness, stability, maximum–minimum principles, Lyapunov functionals, and invariances. The proofs present mathematical results from the nonlinear analysis of partial diﬀerential equations. Since digital images are always sampled on a pixel grid, it is necessary to know if the results for the continuous framework carry over to the practically relevant discrete setting. These questions are an important topic of the present book as well. A general characterization of semidiscrete and fully discrete ﬁlters, which reveal similar properties as their continuous diﬀusion counterparts, is presented. It leads to a semidiscrete and fully discrete scalespace theory for nonlinear diﬀusion processes. Mathematically, this comes down to the study of nonlinear systems of ordinary diﬀerential equations and the theory of nonnegative matrices. Organization of the book. Image processing and computer vision are interdisciplinary areas, where researchers, practitioners and students may have a very diﬀerent scientiﬁc background and diﬀering intentions. As a consequence, I have tried to keep this book as selfcontained as possible, and to include various aspects
PREFACE
vii
such that it should contain interesting material for many readers. The prerequisites are kept to a minimum and can be found in standard textbooks on image processing [163], matrix analysis [407], functional analysis [9, 58, 7], ordinary diﬀerential equations [56, 412], partial diﬀerential equations [185] and their numerical aspects [293, 286]. The book is organized as follows: Chapter 1 surveys the fundamental ideas behind PDEbased smoothing and restoration methods. This general overview sketches their theoretical properties, numerical methods, applications and generalizations. The discussed methods include linear and nonlinear diﬀusion ﬁltering, coupled diﬀusion–reaction methods, PDE analogues of classical morphological processes, Euclidean and aﬃne invariant curve evolutions, and total variation methods. The subsequent three chapters explore a theoretical framework for anisotropic diﬀusion ﬁltering. Chapter 2 presents a general model for the continuous setting where the diﬀusion tensor depends on the structure tensor (interest operator, secondmoment matrix), a generalization of the Gaussiansmoothed gradient allowing a more sophisticated description of local image structure. Existence and uniqueness are discussed, and stability and an extremum principle are proved. Scalespace properties are investigated with respect to invariances and informationreducing qualities resulting from associated Lyapunov functionals. Chapter 3 establishes conditions under which comparable wellposedness and scalespace results can be proved for the semidiscrete framework. This case takes into account the spatial discretization which is characteristic for digital images, but it keeps the scalespace idea of using a continuous scale parameter. It leads to nonlinear systems of ordinary diﬀerential equations. We shall investigate under which conditions it is possible to get consistent approximations of the continuous anisotropic ﬁlter class which satisfy the abovementioned requirements. In practice, scalespaces can only be calculated for a ﬁnite number of scales, though. This corresponds to the fully discrete case which is treated in Chapter 4. The investigated discrete ﬁlter class comes down to solving linear systems of equations which may arise from semiimplicit time discretizations of the semidiscrete ﬁlters. We shall see that many numerical schemes share typical features with their semidiscrete counterparts, for instance wellposedness results, extremum principles, Lyapunov functionals, and convergence to a constant steadystate. This chapter also shows how one can design eﬃcient numerical methods which are in accordance with the fully discrete scalespace framework and which are based on an additive operator splitting (AOS). Chapter 5 is devoted to practical topics such as ﬁlter design, examples and applications of anisotropic diﬀusion ﬁltering. Speciﬁc models are proposed which are tailored towards smoothing with edge enhancement and multiscale enhancement of coherent structures. Their qualities are illustrated using images arising from computer aided quality control and medical applications, but also ﬁngerprint im
During that time I also met Atsushi Imiya (Chiba University. The present book is an extended and revised version of my Ph.D. Japan) at a workshop in Dagstuhl (Germany).viii PREFACE ages and impressionistic paintings shall be processed. and Andrea Bechtold was a great help in all kinds of diﬃculties with the English language. Pierre–Louis Lions (CEREMADE. The discussions and collaborations with the members of this group increased my interest in scalespace related deep structure analysis and information theory. head of the Laboratory of Technomathematics. thesis. In this young and dynamic group I had the possibility to learn a lot about medical image analysis. Acknowledgments. Copenhagen. Stiftung Rheinland– Pfalz f¨r Innovation. This work has been funded by Stiftung Volkswagenwerk. drew my interest to diﬀusion processes in image processing. Chapter 6 concludes the book by giving a summary and discussing possible future perspectives for nonlinear diﬀusion ﬁltering. October 1994) to an international audience.D. The proofreading of this book was done by Martin Reißel and Andrea Bechtold (Kaiserslautern). Helmut Neunzert. the Real World Computing Partnership. University Paris IX) invited me to the CEREMADE. In the latter ﬁeld I share many common interests with Jon Sporring. and it is a pleasure to take this opportunity to express my gratitude to them. Finally. and he provided the possibility to carry out this work at his laboratory. The results are juxtaposed to related methods from Chapter 1. Germany. After the defence of my thesis in Kaiserslautern. The Netherlands. I also thank him and the other editors of the ECMI Series as well as Teubner Verlag for their interest in publishing this work. He introduced me into the fascinating world of early Japanese scalespace research conducted by Taizo Iijima decades before scalespace became popular in America and Europe. In writing this book I have been helped and inﬂuenced by many people. Martin Reißel undertook the hard job of checking the whole manuscript for its mathematical correctness. I joined the TGV (“tools for geometry in vision”) group at Utrecht University Hospital for 14 months. October 1997 Joachim Weickert . Copenhagen University). one of the birthplaces of many important ideas in this ﬁeld. which was written at the Department of Mathematics at the University of Kaiserslautern. and he acted as a referee for the Ph. and to experience Bart ter Haar Romeny’s enthusiasm for scalespace. Also Robert Maas (Utrecht University Hospital) contributed several useful hints. thesis [416]. and the EU–TMR Research Network VIRGO. In the meantime I am with the computer vision group of Peter Johansen and Jens Arnspang (DIKU. He also gave me the honour to present my results as an invited speaker at the EMS Conference Multiscale Analysis in Image Processing (Lunteren. the Danish Research u Council.
.2 Linear diﬀusion ﬁltering .6 Generalizations . . . . . . . . . . . . . 1. . . . . . . . . . . . . .4 Methods of diﬀusion–reaction type . . . . . . . . . . . . . . . . . . . . . 1.5. . . . . . . . . . . . . . . . . . . . . . 1. 1. . 1 2 3 3 6 10 11 12 13 14 15 20 22 24 25 26 27 27 29 31 31 32 32 34 34 35 36 37 37 37 40 42 43 . . . . . . 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . .2. . . . . . . . .6 Generalizations .2 Coupled systems of diﬀusion–reaction 1. . . . . . . . . . . . . . 1. . . . . .2. . ix . . . . . . . . . . . . . . . . . . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . .8 Applications . . . . . . . . . 1.3. . . . . 1. 1. . . . . . . . . . . . . 1. . . . . . . . . . . . .2 Basic operations . .5. . . 1. . . . . .3. . . . . . . 1. . . .3 Anisotropic nonlinear models . . . .4 Theoretical results .3. 1. . .Contents 1 Image smoothing and restoration by PDEs 1. . . . . . . . . . . . . . . . . . . .2.1 Relations to Gaussian smoothing . .5. . . . . . . . . . . .5. . . . . . . . . . . . . . .6 Curvaturebased morphological processes . 1. . . . 1. . . 1. .5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2. . . . . . . . .6 Applications . . . . . .5 Numerical aspects . . . . . . . . . . . .5 Scalespace properties . . . . . . . . . . . . . . .1 Binary and greyscale morphology . . . . .6. . .7 Numerical aspects . . . . . . .3 Continuousscale morphology . . . . . . . .1 Single diﬀusion–reaction equations . . . .5. . . 1. . . . . . . . . . . . . . . . . .3 Nonlinear diﬀusion ﬁltering .5. . . . . . . . . . . . .5 Classic morphological processes . . .4 Applications . . 1. . . . . .2 Scalespace properties . . . 1. . . . . . . . .2. . . . . .1 Meancurvature ﬁltering . . . . .3 Generalizations . . . . 1. . . . . . . . 1. . . . .3. . . . . .6. . 1. . . . . . . . . . . . . . . . . . . 1. . . . . . . . .2 Regularized nonlinear models . . . . 1. . .3. . . . . . . .3 Numerical aspects . . .6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5 Limitations . . . . . . . .6. . 1. . . . . . . . . . . .3. . 1. . . . . .4. . . .4 Generalizations . . . equations . .2 Aﬃne invariant ﬁltering . . . . . . 1. . . . .4 Numerical aspects .1 The Perona–Malik model . . . . . 1. . . . . . . . . . . . . . . . . . . . .1 Physical background of diﬀusion processes . . . . . . . . . . . . .4. . . . . . .5. . . . . . . . . . . . 1. . . . 1. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 5. . . . . 3. . . . . . . . . . . . .1 The general model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2. . . . . . . . . . . 1. . . . . .1 Edgeenhancing diﬀusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7. . . . .x 1. 3. . . . . . . . .1. . . . . . . CONTENTS . 2. . . . . . . . . . .2 Anisotropic case . . . . . . . . .1 Filter design . . .4. . . . . . . . . . . 3. . . . . . . . 2. . . . . . . . . . 5 Examples and applications 5. . 5. . . . . . . . . . . . . . 1. . . . . . . . . . . . . . .3 Scalespace properties . . . . . . . . .1 Basic ﬁlter structure . . . . .7. . . . 4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4. 5. . . 4. .4. . . 45 46 49 50 50 53 55 55 56 57 62 62 65 75 75 76 81 86 86 88 97 97 98 99 102 102 107 113 114 114 114 127 127 129 135 139 165 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Discrete diﬀusion ﬁltering 4. .4. . . . . .2 The structure tensor . . . . . . . . . . . . . . 2. . . . 2. . . . . . . . . . . .2 AOS schemes . . .1 Isotropic case . . . . .1. . . . .6 Active contour models . .2 Theoretical results . . . . . . . . . . . . .4 Relation to continuous models 3. . . . .5 Applications .2 Theoretical results . . . . . . . . . . .3 Theoretical results . . . . . . . .6. . . . .1 Filter design . . . . . .1 Invariances . . . . . . . . .3 Scalespace properties . . . . . . . . . . . . . . . . .2 Coherenceenhancing diﬀusion 5. . . . . . . . . . . . . . . . . . . 4. . . . . . . . . . . . .4 Relation to semidiscrete models 4. . . .8 2 Continuous diﬀusion ﬁltering 2. Total variation methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4. .7 1. . . . . . .1 TVpreserving methods . .4. . . . . . . . . . . . .2 Informationreducing properties 3 Semidiscrete diﬀusion ﬁltering 3. . . . . . . . . 5. . . . . . . .1 The general model . . . .2. . . . . . . . . . . . . . . . . . . . . . . . 2. . . . 1. . . . . . . . . . . . . .1 Semiimplicit schemes . . . . . . . . . . . . . . 6 Conclusions and perspectives Bibliography Index . . . . . . . . . .4 Scalespace properties . Conclusions and further scope of the book . .6.2 TVminimizing methods . . . . . .4. . . . . . . . . . . . . . . . . . . 3. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Applications .
advantages and shortcomings of these techniques. Nevertheless. Although a detailed discussion of these connections would be far beyond the scope of this work. This chapter reviews their main application. generalizations. they are mentioned wherever they appear. Also many historical notes are added. The goal is to make the reader sensitive to the similarities.Chapter 1 Image smoothing and restoration by PDEs PDEbased methods appear in a large variety of image processing and computer vision areas ranging from shapefromshading and histogramme modiﬁcation to optic ﬂow and stereo vision. or mechanisms of the retina and the brain are closely related to some of these approaches. and applications are discussed. namely the smoothing and restoration of images. the study of salinity proﬁles in oceanography. grassﬁre ﬂow. also gas dynamics. diﬀerences. crack propagation. it reveals some interesting relations when it is interpreted within a PDE framework. The outline of this chapter is as follows: We start with reviewing the physical ideas behind diﬀusion processes. For each class of methods the basic ideas are explained and their theoretical background. Many of these ideas are borrowed from physical phenomena such as wave propagation or transport of heat and mass. This becomes 1 . a topic which looks at ﬁrst glance fairly diﬀerent to the diﬀusion approach. and to point out the main results and open problems in this rapidly evolving area. The subsequent study of image enhancement methods of diﬀusion–reaction type relates diﬀusion ﬁlters to variational image restoration techniques. After that we investigate morphological ﬁlters. Nevertheless. where proofs and full mathematical details can be found. in order to allow the interested reader to pursue these ideas. It is written in an informal style and refers to a large amount of original literature. This helps us to better understand the next sections which are concerned with the properties of linear and nonlinear diﬀusion ﬁlters in image processing. numerical aspects.
(1. The observation that diﬀusion does only transport mass without destroying it or creating new mass is expressed by the continuity equation ∂t u = −div j (1. The case where j and ∇u are parallel is called isotropic.2 CHAPTER 1. and inhomogeneous blurring is called anisotropic. and a spacedependent ﬁltering is called inhomogeneous. If we plug in Fick’s law into the continuity equation we end up with the diﬀusion equation ∂t u = div (D · ∇u). The equilibration property is expressed by Fick’s law: j = −D · ∇u.3) This equation appears in many physical transport processes. Such a feedback leads to nonlinear diﬀusion ﬁlters. . even if it uses a scalarvalued diﬀusivity instead of a diﬀusion tensor. 1. In image processing we may identify the concentration with the grey value at a certain location. Diﬀusion which does not depend on the evolving image is called linear. Often the diﬀusion tensor is a function of the diﬀerential structure of the evolving image itself. one speaks of homogeneous diﬀusion. j and ∇u are not parallel. Sometimes the computer vision literature deviates from the preceding notations: It can happen that homogeneous ﬁltering is named isotropic.2) where t denotes the time.1 Physical background of diﬀusion processes Most people have an intuitive impression of diﬀusion as a physical process that equilibrates concentration diﬀerences without creating or destroying mass. In the general anisotropic case. PARTIAL DIFFERENTIAL EQUATIONS especially evident when considering curvaturebased morphological PDEs. If the diﬀusion tensor is constant over the whole image domain. The relation between ∇u and j is described by the diﬀusion tensor D. In the context of heat transfer it is called heat equation. The last section summarizes the advantages and shortcomings of the main methods and gives an outline of the questions we are concerned with in the subsequent chapters. (1.1) This equation states that a concentration gradient ∇u causes a ﬂux j which aims to compensate for this gradient. Finally we shall discuss total variation image restoration techniques which permit discontinuous solutions. Then we may replace the diﬀusion tensor by a positive scalarvalued diﬀusivity g. a positive deﬁnite symmetric matrix. This physical observation can be easily cast in a mathematical formulation.
(F Kσ )(ω) = exp − ω2 . 1.1 Relations to Gaussian smoothing Gaussian smoothing Let a greyscale image f be represented by a realvalued mapping f ∈ L1 (IR2 ). We shall focus on the relation between linear diﬀusion ﬁltering and the convolution with a Gaussian.1. (1. analyse its smoothing properties for the image as well as its derivatives.4) where Kσ denotes the twodimensional Gaussian of width (standard deviation) σ >0 : x2 1 · exp − 2 . and review the fundamental properties of the Gaussian scalespace induced by linear diﬀusion ﬁltering.6) we obtain by the convolution theorem that (F (Kσ ∗f )) (ω) = (F Kσ )(ω) · (F f )(ω). Since the Fourier transform of a Gaussian is again Gaussianshaped.7) .2.2 Linear diﬀusion ﬁltering The simplest and best investigated PDE method for smoothing images is to apply a linear diﬀusion process. Afterwards a survey on discrete aspects is given and applications and limitations of the linear diﬀusion paradigm are discussed. A widelyused way to smooth f is by calculating the convolution (Kσ ∗f )(x) := Kσ (x−y) f (y) dy IR 2 (1. x ) dx (1. even if f is only absolutely integrable.8) (1. let us investigate the behaviour in the frequency domain. 2/σ 2 (1.2 LINEAR DIFFUSION FILTERING 3 1.5) Kσ (x) := 2πσ 2 2σ There are several reasons for the excellent smoothing properties of this method: First we observe that since Kσ ∈ C ∞ (IR2 ) we get Kσ ∗f ∈ C ∞ (IR2 ). The section is concluded by sketching two linear generalizations which can incorporate apriori knowledge: aﬃne Gaussian scalespace and directed diﬀusion processes. Next. When deﬁning the Fourier transformation F by (F f )(ω) := IR2 f (x) exp(−i ω.
√ From (1. t) ≤ sup f 2 IR IR2 on IR2 × [0.11) we observe that the time t is related to the spatial width σ = 2t of the Gaussian. if it has a unique solution which depends continuously on the input data and parameters. 0) = f (x) possesses the solution u(x.4 CHAPTER 1.3 (c) in Chapter 5 illustrate the eﬀect of linear diﬀusion ﬁltering. PARTIAL DIFFERENTIAL EQUATIONS we observe that (1.2 (b) and 5. Gaussian derivatives In order to understand the structure of an image we have to analyse grey value ﬂuctuations within a neighbourhood of each image point. 43–56]) that for any bounded f ∈ C(IR2 ) the linear diﬀusion process ∂t u = ∆u.14) Figure 5. diﬀerentiation is illposed1 .4) is a lowpass ﬁlter that attenuates high frequencies in a monotone way. provided we restrict ourselves to functions satisfying u(x. pp. L∞ (IR2 ) . u(x. t) ≤ M · exp (ax2 ) (M. Hence. (1. the smoothing behaviour can also be understood in the context of a PDE interpretation. ∞).9) (1. If one of these conditions is violated.12) and it (1. it is called illposed. [331. 1 . as small perturbations in the original image can lead to arbitrarily large ﬂuctuations in the derivatives. (1.10) This solution is unique. Equivalence to linear diﬀusion ﬁltering It is a classical result (cf. that is to say. (1. a > 0). . 267–271] and [185.g. we need information about its derivatives. A thorough A problem is called wellposed. pp.13) It depends continuously on the initial image f with respect to fulﬁls the maximum–minimum principle inf f ≤ u(x. the need for regularization methods arises. smoothing structures of order σ requires to stop the diﬀusion process at time 1 T = 2 σ2 . e. Hence.11) (1. However. Interestingly. t) = f (x) (K√2t ∗ f )(x) (t = 0) (t > 0).
177]. This evolution property is called causality [240]. This property has been used to stabilize illposed problems like deblurring images by solving the heat equation backwards in time2 [141. It is based on calculating the ﬁrst derivatives of the Gaussiansmoothed image. we focus on two applications for detecting edges. Young developed this theory further by presenting evidence that the receptive ﬁelds in primate eyes are shaped like the sum of a Gaussian and its Laplacian [449]. By the equality n m n m n m ∂x1 ∂x2 (Kσ ∗f ) = Kσ ∗ (∂x1 ∂x2 f ) = (∂x1 ∂x2 Kσ ) ∗ f (1. which uses the LaplacianofGaussian (LoG) ∆Kσ as convolution kernel. Moreover. After applying sophisticated thinning and linking mechanisms (nonmaxima suppression and hysteresis thresholding). no new zerocrossings are created which cannot be traced back to ﬁner scales [439].1. This method is often acknowledged to be the best linear edge detector. increasing the order of applied Gaussian derivatives or reducing the kernel size will ﬁnally deteriorate the results of deblurring. expressions that are invariant under transformations such as rotations. Louis [266] and Engl et al. A frequently used method is the Canny edge detector [69]. There are indications that LoGs and especially their approximation by diﬀerencesofGaussians (DoGs) play an important role in the visual system of mammals. It is Of course. junctions. If one investigates the temporal evolution of the zerocrossings of an image ﬁltered by linear diﬀusion. ridges. [128]. 2 . for instance ∇Kσ ∗u or ∆Kσ ∗u.15) for suﬃciently smooth f . To illustrate this. Gaussian derivatives can be combined to socalled diﬀerential invariants. see [278] and the references therein. one observes an interesting phenomenon: When increasing the smoothing scale σ. Another important edge detector is the Marr–Hildreth operator [278]. One possibility to regularize is to convolve the image with a Gaussian prior to diﬀerentiation [404]. Replacing derivatives by these Gaussian derivatives has a strong regularizing eﬀect. Edges of f are identiﬁed as zerocrossings of ∆Kσ ∗f .2 LINEAR DIFFUSION FILTERING 5 treatment of this mathematical theory can be found in the books of Tikhonov and Arsenin [402]. see [256] for an overview. edges are identiﬁed as locations where the gradient magnitude has a maximum. Diﬀerential invariants are useful for the detection of features such as edges. solutions of the regularization can only approximate the solution of the original problem (if it exists). and it has become a standard in edge detection. and blobs. we observe that all derivatives undergo the same Gaussian smoothing process as the image itself and this process is equivalent to convolving the image with derivatives of a Gaussian. This needs no further postprocessing and always gives closed contours. In practice. and Koenderink and van Doorn suggested the set of Gaussian derivatives as a general model for the visual system [242].
embedding the image f into a family {Tt f t ≥ 0} of gradually simpliﬁed versions of it. the scalespace representation gives the original image f . nonenhancement of local extrema [30. This simpliﬁcation property is speciﬁed by numerous authors in diﬀerent ways. (1. using concepts such as no creation of new level curves (causality) [240.e. and the ﬁltering may be split into a sequence of ﬁlter banks: T0 f = f. An important architectural assumption is recursivity. They concluded. by comparing the structures at diﬀerent scales.6 CHAPTER 1. that this is practically unstable unless very much additional information is provided. 3 . decreasing number Recently it has also been proposed to extend the scalespace concept to scale–imprecision space by taking into account the imprecision of the measurement device [171]. Moreover. 1. PARTIAL DIFFERENTIAL EQUATIONS closely connected to the maximum–minimum principle of certain parabolic operators [189].17) This property is very often referred to as the semigroup property. In those cases where it is not clear in advance which is the right scale for the depicted information it is desirable to have an image representation at multiple scales.2 Scalespace properties The general scalespace concept It is a wellknown fact that images usually contain structures at a large variety of scales. t ≥ 0. Other architectural principles comprise for instance regularity properties of Tt and local behaviour as t tends to 0. Most of these properties can be classiﬁed as architectural. we should not have additional structures which are caused by the ﬁltering method itself and not by underlying structures at ﬁner scales. In the western world the evolution property of the zerocrossings was the key investigation which has inspired Witkin to the socalled scalespace concept [439]. one obtains a hierarchy of image structures which eases a subsequent image interpretation.2. 255]. provided that it fulﬁls certain requirements3 . A scalespace is an image representation at a continuum of scales. Smoothing properties and information reduction arise from the wish that the transformation should not create artifacts when passing from ﬁne to coarse representation. smoothing (informationreducing) or invariance requirements [12]. however. Tt+s f = Tt (Ts f ) ∀ s. Thus. Attempts to reconstruct the original image from the temporal evolution of the zerocrossings of the Laplacian have been carried out by Hummel and Moniot [190]. for t = 0. 450. at a coarse scale. i.16) (1. 189. 257]. This shall be discussed next.
preservation of positivity [191. in order to analyse only the depicted object [196.g. 193. and Lyapunov functionals [415. which is obtained via convolution with Gaussians of increasing variance.1. 303]. Thus.18) As we shall see below. Ishikawa and Imiya [426. maximum loss of ﬁgure impression [196]. σ] = 4 ′ ∞ −∞ φ{f (x′ ). x. 320]. This makes the requirement plausible that the scalespace analysis should be invariant to as many of these transformations as possible. In [192] Iijima considers an observation transformation Φ which depends on a scale parameter σ and which transforms the original image f (x) into a blurred version4 Φ[f (x′ ). positivity [324. 138]. (1.2 LINEAR DIFFUSION FILTERING 7 of local extrema [255]. Often these requirements are supplemented with an additional assumption which is equivalent to the superposition principle. imposing linearity restricts the scalespace idea to essentially one representative. x. We may regard an image as a representative of an equivalence class containing all images that depict the same object. 429]. (1. The pioneering work of Alvarez. Especially in the linear setting. 16]. x′ . b ∈ IR. ∀ a. however. σ]. comparison principle [12].19) The variable x′ serves as a dummy variable. namely linearity: Tt (af + bg) = a Tt f + b Tt g ∀ t ≥ 0. maximum–minimum principle [189. many of these properties are equivalent or closely related. translations and rotations or even more complicated transformations such as aﬃne mappings. . Lions and Morel [12] shows that every scalespace fulﬁlling some fairly natural architectural. 427]. informationreducing and invariance axioms is governed by a PDE with the original image as initial condition. shows that scalespace is more than 20 years older: An axiomatic derivation of 1D Gaussian scalespace has already been presented by Taizo Iijima in a technical paper from 1959 [191] followed by a journal version in 1962 [192]. Usually a 1983 paper by Witkin [439] or a 1980 report by Stansﬁeld [392] are regarded as the ﬁrst references to the linear scalespace idea. or – equivalently – by linear diﬀusion ﬁltering according to (1. x. (1. This class of blurring transformations is called boke (defocusing). 328]. Gaussian scalespace The historically ﬁrst and best investigated scalespace is the Gaussian scalespace. Tikhonov regularization [302. σ} dx′ . PDEs are the suitable framework for scalespaces. see [426] for more details.10). by greylevel shifts. He assumes that it has the structure Φ[f (x ). Both papers are written in Japanese. Guichard. Two images of this class diﬀer e.9). Recent work by Weickert.
σ2 = Φ[f (x′′ ). which is basically unknown in the western world. x. then the same should happen to the observed pattern: Φ[Af (x′ ). then this is equivalent to observing f under a suitable parameter σ3 = σ3 (σ1 . 200. σ]. x. σ] = A Φ[f (x′ ). and many results have . 320]. σ3 ]. x−a. σ]. x. x. 197. (1. Japanese scalespace theory was wellembedded in a general framework for pattern recognition. (1. σ] = Φ[f (x′ ). λ) such that Φ[f (x′ /λ). x. PARTIAL DIFFERENTIAL EQUATIONS and that it should satisfy ﬁve conditions: (I) Linearity (with respect to multiplications): If the intensity of a pattern becomes A times its original intensity. σ2 ): Φ Φ[f (x′′ ).25) Thus. x′ .24) Under these requirements Iijima derives in a very systematic way that 1 Φ[f (x ).20) (II) Translation invariance: Filtering a translated image is the same as translating the ﬁltered image: Φ[f (x′ −a). σ] = √ 2 πσ ′ ∞ −∞ f (x′ ) exp −(x − x′ )2 4σ 2 dx′ . σ] > 0 ∀ f (x′ ) > 0. σ] is just the convolution between f and a Gaussian with standard √ deviation σ 2. σ] = Φ[f (x′ ). σ ′ ]. x. x. then the observed image is positive as well: Φ[f (x′ ).8 CHAPTER 1. σ1 ]. ∀ σ > 0. x.21) (III) Scale invariance: If a pattern is spatially enlarged by some factor λ. x/λ. x. (1. then there exists a σ ′ = σ ′ (σ. (1. Φ[f (x′ ). This has been the starting point of an entire world of linear scalespace research in Japan.23) (V) Preservation of positivity: If the original image is positive. feature extraction and object classiﬁcation [195. (1.22) (IV) (Generalized) semigroup property: If f is observed under a parameter σ1 and this observation is observed under a parameter σ2 . (1.
All of these axiomatics use explicitly or implicitly5 a linearity assumption.2 N • L2 • • • • • • • N • F2 • • • • • • • • >1 • • • • • • 1. 427] as well as in some English scalespace papers by Iijima such as [195. [140]. In particular. Apart from their historical merits.1 presents an overview of the current Japanese and western Gaussian scalespace axiomatics (see [426. 198]. & isotropy separability scale invariance valid for dimension I1 • • I2 • • I3 O • K Y • • • • • • • • 1 • 2 • • • • 2 2 • 1.2 • • N 1 N N Table 1. [12]. convolution kernel semigroup property locality regularity inﬁnetes. aver. [324].2 • • B • • • • • • 1 L1 • • • • • • F1 • • • A • • • • P • • • • • • • • 1. I3 = Iijima [196]. 194]. [303]. Table 1. as well as in English books by Lindeberg [256]. L1 = Lindeberg [255]. 5 .1.2 LINEAR DIFFUSION FILTERING 9 been established earlier than in the western world. A collection edited by Often it is assumed that the ﬁlter is a convolution integral. F1 = Florack et al. Y = Yuille/Poggio [450]. [30]. This is equivalent to linearity and translation invariance. Each of the 14 axiomatics conﬁrms and enhances the evidence that the others give: that Gaussian scalespace is unique within a linear framework. generator max.1: Overview of continuous Gaussian scalespace axiomatics (I1 = Iijima [191. Florack [139]. and ter Haar Romeny [176]. O = Otsu [320]. 192]. P = Pauwels et al. grey level invar. More details can be found in [426. A = Alvarez et al. ﬂat kernel for t → ∞ isometry invariance homogen. these Japanese results reveal many interesting qualities which should induce everyone who is interested in scalespace theory to have a closer look at them. 197]. the latter ones show that there is no justiﬁcation to deny Iijima’s pioneering role in linear scalespace theory because of language reasons. 427] for detailed explanations). We observe that – despite the fact that many axiomatics reveal similar ﬁrst principles – not two of them are identical. K = Koenderink [240]. A detailed treatment of Gaussian scalespace theory can be found in two Japanese monographs by Iijima [197. L2 = Lindeberg [257]. I2 = Iijima [193. loss principle causality nonnegativity Tikhonov regulariz. F2 = Florack [138]). N = Nielsen et al. B = Babaud et al.
By virtue of the equivalence of solving the linear diﬀusion equation and convolving with a Gaussian. who proposed in 1960 that the discrete analogue of the Gaussian kernel should be given in terms of modiﬁed Bessel functions of integer order. it has been argued that approximating the diﬀusion equation should be preferred to discretizing the convolution integral [255]. PARTIAL DIFFERENTIAL EQUATIONS Sporring. 448]. (1. More frequently the Gaussian kernel is just sampled and truncated at some multiple of its standard deviation σ. Recently. Since this scalespace family arises naturally from a semidiscretized version of the diﬀusion equation. Thus. 6 By semidiscrete we mean discrete in space and continuous in time throughout this work. which depends only on the pixel number N. but not on the kernel size σ. Many relations between Gaussian scalespace and regularization theory have been elaborated by Nielsen [302]. Florack and Johansen [389] gives an excellent overview of the various aspects of this theory. [353] exploits the relations between diﬀusion and Markov processes. we can either approximate the convolution process or the diﬀusion equation. Lindeberg [255] has established a linear scalespace theory for the semidiscrete6 case.and undershoots [178]. interesting semidiscrete and fully discrete linear scalespace formulations have been established utilizing stochastic principles: ˚str¨m and Heyden [27] A o study a framework based on stationary random ﬁelds.10 CHAPTER 1. however.2. and additional material is presented in [211]. 1. while the theory by Salden et al. His results are in accordance with those of Norman [312]. the image is sampled at the nodes (pixels) of a ﬁxed equidistant grid. in practical problems. however.3 Numerical aspects The preceding theory is entirely continuous. Convolution with a truncated Gaussian. . and readers who wish to analyse linear and nonlinear scalespace concepts in terms of diﬀerential and integral geometry can ﬁnd a lot of material in the thesis of Salden [351]. reveals the drawback that it does not preserve the semigroup property of the continuous Gaussian scalespace [255]. Factorizing a higherdimensional Gaussian into onedimensional Gaussians reduces the computational eﬀort to O(Nσ). convolution in the spatial domain can be reduced to multiplication in the frequency domain. aliasing eﬀects in the Fourier domain may create oscillations and over. Especially for small kernels.7). the diﬀusion ﬁlter has to be discretized. However. When restricting the image to a ﬁnite domain and applying the Fast Fourier Transformation (FFT). cf. This proceeding requires a ﬁxed computational eﬀort of order N log N. Nielsen. One eﬃcient possibility to approximate Gaussian convolution in the spatial domain consists of applying recursive ﬁlters [109. For large kernels this is faster than most spatial techniques.
10. 214]. 7 . The idea is to identify segments at coarse scales and to link backwards to the original image in order to improve the localization. Less frequent are applications which exploit the evolution of an image under Gaussian scalespace. 67. The Gaussian pyramid [64] has the computational complexity O(N) and gives a multilevel representation at ﬁnitely many scales of diﬀerent resolution. Due to their simplicity and eﬃciency. focusofattention).1. 254. 441]. and sometimes it is argued that they are undersampled and that the pyramid levels should be closer7 . explicit schemes are mainly used. 302. linear diﬀusion ﬁltering has been applied in numerous ﬁelds of image processing and computer vision. 1. however. 313. optic ﬂow and image sequences have become an active research ﬁeld [139. Several scalespace applications are summarized in a survey paper by ter Haar Romeny [175]. Of course. 259. This deep structure analysis [240] provides useful information for extracting semantic information from an image.4 Applications Due to its equivalence to convolution with a Gaussian. for instance • for ﬁnding the most relevant scales (scale selection. 306. 256. 8 Historically. 106] are just designed to have few or no redundancies. ﬁnite diﬀerence (FD) schemes dominate the ﬁeld. while scalespace analysis intends to extract semantical information by tracing signals through a continuum of scales. 258. 388] or generalized entropies [390] over scales. A very eﬃcient approximation of the Gaussian scalespace results from applying multigrid ideas. pyramid decompositions have become very popular and have been integrated into commercially available hardware [70. It can be found in almost every standard textbook in these ﬁelds.2 LINEAR DIFFUSION FILTERING 11 Among the numerous numerical possibilities to approximate the linear diﬀusion equation. By subsequently smoothing the image with an explicit scheme for the diﬀusion equation and restricting the result to a coarser grid. In recent years also applications of Gaussian scalespace to stereo. Pyramids are not invariant under translations. 215. • for multiscale segmentation of images [172. this is incorrect: Iijima’s scalespace work [191] is much older than multigrid ideas in image processing. 68] allowing realizations as a recursive ﬁlter [14. multiresolution techniques such as pyramids or discrete wavelet transforms [92. Apart from some implicit approaches [166. one obtains a simpliﬁed image representation at the next coarser grid. 451]. 408].2. These are the reasons why some people regard pyramids rather as predecessors of the scalespace idea than as a numerical approximation8 . This may be done by searching for extrema of (nonlinear) combinations of normalized Gaussian derivatives [256] or by analysing information theoretic measures such as the entropy [208. 241.
So structures which are identiﬁed at a coarse scale do not give the right location and have to be traced back to the original image [439. Part III of the book edited by Sporring et al. (c) Some smoothing properties of Gaussian scalespace do not carry over from the 1D case to higher dimensions: A closed zerocrossing contour can split into two as the scale increases [450]. Witkin [439]. This and other results [12. where the gradient vanishes and the Hessian does not have full rank. relating dislocated information obtained at diﬀerent scales is diﬃcult and bifurcations may give rise to instabilities.12 CHAPTER 1. for instance by analysing its toppoints [210]. Since Gaussian smoothing is designed to be completely uncommitted. A deep mathematical analysis of such phenomena has been carried out by Damon [105] and Rieger [342]. Regarding (b) and (c). 165]. There is some evidence that these points. These coarsetoﬁne tracking diﬃculties are generally denoted as the correspondence problem. and it is generally not true that the number of local extrema is nonincreasing.2. makes them harder to identify. (b) Linear diﬀusion ﬁltering dislocates edges when moving from ﬁner to coarser scales. see [254. 353]: while the diﬀusion equation remains the correct concept. but also blurs important features such as edges and.g. In practice. it reveals some drawbacks as well: (a) An obvious disadvantage of Gaussian smoothing is the fact that it does not only smooth noise. Due to the uniqueness of Gaussian scalespace within a linear framework we know that any modiﬁcation in order to overcome the problems (a)–(c) will either . PARTIAL DIFFERENTIAL EQUATIONS Interesting results arise when one studies linear scalespace on a sphere [236. 38. see e. 1. 255] indicate that the PDE formulation of linear scalespace in terms of a diﬀusion equation is more natural and has a larger generalization potential than convolution with Gaussians. it cannot take into account any apriori information on structures which are worth being preserved (or even enhanced). much eﬀorts have been spent in order to understand the deep structure in Gaussian scalespace.5 Limitations In spite of several properties that make linear diﬀusion ﬁltering unique and easy to handle. but happens generically. thus. Gaussian kernels are of no use anymore: appropriate kernels have to be expressed in terms of Legendre functions [236]. 255] for illustrative counterexamples. It turned out that the pairwise creation of an extremum and a saddle point is not an exception. [389] and the references therein give an overview of the stateoftheart in deep structure analysis. carry essential image information [212].
200]. (1.1. 443].2) of f . Later on similar ideas have been developed in [259. 1. t) := IR 2 −1 (x−y)⊤ Dt (x−y) exp − 4 4π det(Dt ) 1 f (y) dy (1. and D ∈ IR2×2 is symmetric positive deﬁnite9 .26) as a function of the structure tensor (cf. while the divergence formulation ensures that this is still possible for nonuniform diﬀusion ﬁltering.27) the basic equation of ﬁgure [196].28) u(x. For a ﬁxed matrix D. 194]. 207. In 1992 Nitzberg and Shiota [310] proposed to adapt the Gaussian kernel shape to the structure of the original image.6 Generalizations Before we turn our attention to nonlinear processes. In 1971 this concept was realized in hardware in the optical character reader ASPET/71 [199. In [427] it is shown that aﬃne Gaussian scalespace has been axiomatically derived by Iijima in 1962 [193. He named u(x. they combined nonlinear shape adaptation with linear smoothing. 218]. and (1. The scalespace part has been regarded as the reason for its reliability and robustness. We shall see that appropriate methods to avoid the shortcomings (a) and (b) are nonlinear diﬀusion processes.1. Section 2.26) where Dt := tD. Aﬃne Gaussian scalespace A straightforward generalization of Gaussian scalespace results from renouncing invariance under rotations. In practice this can be experienced by the fact that shapeadaptation of Gaussian smoothing does not preserve the average grey level. t) the generalized ﬁgure of f . let us ﬁrst investigate two linear modiﬁcations which have been introduced in order to address the problems (a) and (b) from the previous section. while (c) requires morphological equations [206. t > 0. . This leads to the aﬃne Gaussian scalespace u(x.26) is equivalent to solving a linear anisotropic diﬀusion problem with D as diﬀusion tensor: ∂t u = div (D ∇u). Also in this case the diﬀusion equation seems to be more general.2. By chosing D in (1. see Section 1. It should be noted that shapeadapted Gaussian smoothing with a spatially varying D is no longer equivalent to a diﬀusion process of type (1.27).2 LINEAR DIFFUSION FILTERING 13 renounce linearity or some scalespace properties. 0) = f (x). If one wants to relate 9 Isotropic Gaussian scalespace can be recovered using the unit matrix for D. calculating the convolution integral (1.27) (1.
29) (1.14 CHAPTER 1. one has to carry out sophisticated scaling limits [310]. by prescribing that features within a certain contrast and scale range are considered to be semantically important and processed diﬀerently. a corresponding PDE formulation was ﬁrst given by Perona and Malik [326] in 1987.g.30) u(x.27) to the diﬀerential structure of the ﬁltered image instead of the original image. in many applications it would be desirable to include a priori knowledge in a less speciﬁc way. This gives rise to study regularizations.3 Nonlinear diﬀusion ﬁltering Adaptive smoothing methods are based on the idea of applying a process which itself depends on local properties of the image. These techniques can be extended to anisotropic processes which make use of an adapted diﬀusion tensor instead of a scalar diﬀusivity. (1. Numerical experiments have been carried out by Giuliani [159]. Provided we are given some background information in form of a smooth image b. and an analysis in terms of nonsmooth b and weak solutions is due to Illner and Tie [203].3 and in the remaining chapters of this book. Noniterative shapeadapted Gaussian smoothing diﬀers from nonlinear anisotropic diﬀusion ﬁltering by the fact that the latter one introduces a feedback into the process: it adapts the diﬀusion tensor in (1. especially its illposedness aspects. 1. We shall discuss this model in detail. . they show that under some technical requirements and suitable boundary conditions the classical solution u of ∂t u = b ∆u − u ∆b. Although this concept is wellknown in the image processing community (see [349] and the references therein for an overview). Such concepts will be investigated in Section 1.3. Directed diﬀusion Another method for incorporating apriori knowledge into a linear diﬀusion process is suggested by Illner and Neunzert [202]. Such a directed diﬀusion process requires to specify an entire image as background information in advance. e. 0) = f (x) converges to b along a path where the relative entropy with respect to b increases in a monotone way. PARTIAL DIFFERENTIAL EQUATIONS shapeadapted Gaussian smoothing to a PDE. Such demands can be satisﬁed by nonlinear diﬀusion ﬁlters.
Moreover. This likelihood is measured by ∇u2. and of backward parabolic type for ux  > λ. 328].3 NONLINEAR DIFFUSION FILTERING 15 1. it should be noted that – in our terminology – it would be regarded as an isotropic model. (1.1 The Perona–Malik model Basic idea Perona and Malik propose a nonlinear diﬀusion method for avoiding the blurring and localization problems of linear diﬀusion ﬁltering [326. This is due to the fact that diﬀusion and edge detection interact in one single process instead of being treated as two independent processes which are to be applied subsequently. For the diﬀusivity (1. there exists a relation between (1. This simpliﬁes the notation and illustrates the main behaviour since near a straight edge a twodimensional image approximates a function of one variable. see Figure 1. The experiments of Perona and Malik were visually very impressive: edges remained stable over a very long time. and Φ′ (s) < 0 for s > λ.32) (1. The Perona– Malik ﬁlter is based on the equation ∂t u = div (g(∇u2) ∇u). They apply an inhomogeneous process that reduces the diﬀusivity at those locations which have a larger likelihood to be edges. . since it utilizes a scalarvalued diﬀusivity and not a diﬀusion tensor.31) Although Perona and Malik name their ﬁlter anisotropic. let us for a moment restrict ourselves to the onedimensional case.32) it follows that the ﬂux function Φ(s) := sg(s2 ) satisﬁes ′ Φ (s) ≥ 0 for s ≤ λ. there is another reason for the impressive behaviour at edges. Edge enhancement To study the behaviour of the Perona–Malik ﬁlter at edges. which we shall discuss next. and it uses diﬀusivities such as g(s2) = 1 1 + s2 /λ2 (λ > 0). Since (1.1.31) can be rewritten as ∂t u = Φ′ (ux )uxx . It was demonstrated [328] that edge detection based on this process clearly outperforms the linear Canny edge detector.33) we observe that – in spite of its nonnegative diﬀusivity – the Perona–Malik model is of forward parabolic type for ux  ≤ λ.1. even without applying nonmaxima suppression and hysteresis thresholding. Interestingly.3.31) and the neural dynamics of brightness perception: In 1984 Cohen and Grossberg [94] proposed a model of the primary visual cortex with similar inhibition eﬀects as in the Perona–Malik model. (1.
we have forward diﬀusion along isophotes (i.35) where the gauge coordinates ξ and η denote the directions perpendicular and parallel to ∇u. Therefore. In the twodimensional case.16 1 CHAPTER 1. Overviews of several common diﬀusivities for the Perona–Malik model can be found in [43.34) A location x0 where u2 is maximal at some time t is characterized by ux uxx = 0 x and ux uxxx ≤ 0. PARTIAL DIFFERENTIAL EQUATIONS 0. and a family of diﬀusivities with diﬀerent decay rates is investigated in [36]. respectively. Hence.33) is replaced by [12] ∂t u = Φ′ (∇u)uηη + g(∇u2)uξξ (1. λ plays the role of a contrast parameter separating forward (low contrast) from backward (high contrast) diﬀusion areas.32). lines of constant grey value) combined with forward–backward diﬀusion along ﬂowlines (lines of maximal grey value variation). It is not hard to verify that the Perona–Malik ﬁlter increases the slope at inﬂection points of edges within a backward area: If there exists a suﬃciently smooth solution u it satisﬁes ∂t (u2 ) = 2ux ∂x (ut ) = 2Φ′′ (ux )ux u2 + 2Φ′ (ux )ux uxxx. Rapidly decreasing diﬀusivities are explicitly intended in the Perona–Malik method as they .36) (1. (∂t (u2 )) (x0 .6 diffusivity flux function 0 0 lambda 0 0 lambda Figure 1. 343]. t) ≥ 0 for ux (x0 .1: (a) Left: Diﬀusivity g(s2 ) = 1+s1 /λ2 . (1.e. 2 Hence. (b) Right: Flux function 2 Φ(s) = 1+ss /λ2 . it appears for all diﬀusivities g(s2 ) whose rapid decay causes nonmonotone ﬂux functions Φ(s) = sg(s2 ). x xx (1. We observe that the forward–backward diﬀusion behaviour is not only restricted to the special diﬀusivity (1. t) > λ x with strict inequality for ux uxxx < 0.
∇u. r. v ∈ H. r.e. for instance adapting it to a speciﬁed quantile in the cumulative gradient histogramme [328]. Let us recall three examples: • Let S(N) denote the set of symmetric N × N matrices and Hess(u) the Hessian of u. (Au−Av. It is evident that the “optimal” value for the contrast parameter λ has to depend on the problem. Illposedness Unfortunately.37) satisﬁes the monotony property F (t. Several proposals have been made to facilitate such a choice in practice.39) (1.) and A : H → H . (1. u.40) (1. using statistical properties of a training set of regions which are considered as ﬂat [444]. In order to apply the concept of maximal monotone operators [57] to the problem du + Au = 0. X) for all X. x. i. Therefore.38) . Ishii and Lions [103]. Y ∈ S(2) where Y −X is positive semideﬁnite. • The same requirement is needed for applying the theory of viscosity solutions. x. A detailed introduction into this framework can be found in a paper by Crandall. Classical diﬀerential inequality techniques [411] based on the Nagumo–Westphal lemma require that the underlying nonlinear evolution equation ∂t u = F (t. dt u(0) = f one has to ensure that A is monotone. . x..41) (1.3 NONLINEAR DIFFUSION FILTERING 17 give the desirable result of blurring small ﬂuctuations and sharpening edges. p. they are the main reason for the visually impressive results of this restoration technique. u−v) ≥ 0 ∀ u. or estimating it by means of the local image geometry [270]. forward–backward equations of Perona–Malik type cause some theoretical problems. p. Y ) ≥ F (t.1. • Let H be a Hilbert space with scalar product (. there exist certain frameworks which allow to establish wellposedness results for a large class of equations. Although there is no general theory for nonlinear parabolic processes. Hess(u)) (1.
If they exist. the mainly observable instability is the socalled staircasing eﬀect. however. . Numerically. In the meantime several theoretical results are available which provide some insights into the actual degree of illposedness of the Perona–Malik ﬁlter. In 1994 the general conjecture was that the Perona–Malik ﬁlter might have weak solutions. Kawohl and Kutev [222] proved that the Perona–Malik process does not have global (weak) C 1 solutions for intial data that involve backward diﬀusion. Therefore. none of these frameworks is applicable to ensure wellposedness results. Kawohl and Kutev showed that these solutions are unique and satisfy a maximumminimum principle. 225] proposed a notion of generalized solutions. and he showed that an analysis of their moving and merging gives similar eﬀects to those one can observe in practice. and slightly diﬀerent initial images may end up in diﬀerent minima for t → ∞. Starting from a smoothly increasing initial distribution he reported the creation of perturbations which led to a stepwise constant proﬁle after some time. forward–backward diﬀusion equations of Perona–Malik type are not as unnatural as they look at ﬁrst glance: besides their importance in computer vision they have been proposed as a mathematical model for heat and mass transfer in a stably stratiﬁed turbulent shear ﬂow. The existence of local C 1 solutions remained unproven. which are piecewise linear and contain jumps. He constructed a forward– o backward diﬀusion process which can have inﬁnitely many solutions. Such a model is used to explain the evolution of stepwise constant temperature or salinity proﬁles in the ocean. They showed that the energy functional leading to the Perona–Malik process as steepest descent method has an inﬁnite number of global minima which are dense in the image space. One reason why people became pessimistic about the wellposedness of the Perona–Malik equation was a result by H¨llig [187].18 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS We observe that the nonmonotone ﬂux function of the Perona–Malik process implies that neither (1. Interestingly. Although this process was diﬀerent from the Perona–Malik process. Kichenassamy [224. Related equations also play a role in population dynamics and viscoelasiticity. where a sigmoid edge evolves into piecewise linear segments which are separated by jumps. [446] give evidence that the Perona–Malik process is unstable with respect to perturbations of the initial image. but one should neither expect uniqueness nor stability [329].38) is satisﬁed nor A deﬁned by Au := −div (g(∇u2) ∇u) is monotone. one was warned what can happen. Moreover. He used an equation of Perona–Malik type for numerical simulations of the salinity proﬁles in oceans. see [35] and the references therein. It has already been observed by Posmentier in 1977 [333]. under special assumptions on the initial data. Each of these minima corresponds to a piecewise constant image. it was possible to establish a comparison principle. Results of You et al.
3 NONLINEAR DIFFUSION FILTERING 19 In image processing. Of course. this is only partly satisfying. Florack [143] and Eberly [124] proposed to carry over the linear scalespace theory to the nonlinear case by considering nonlinear diﬀusion . a natural way would be to prove a maximum–minimum principle. Instead of this. standard FD discretizations are monotonicity preserving. Further contributions to the explanation and avoidance of staircasing can be found in [4. Fr¨hlich and Weickert [148]. Its global solution satisﬁes a maximum–minimum principle and converges to a constant steadystate. 225. This has been shown by Dzu Magaziewa [123] in the semidiscrete case and by Benhamouda [36. Since the existence question used to be the bottleneck in the past. the ﬁrst proof is due to Kawohl and Kutev who established an extremum principle for their local weak C 1 solution to the Perona–Malik ﬁlter [222]. The theoretical framework for this analysis will be presented in Chapter 3. All results point in the same direction: the number of created plateaus depends strongly on the regularizing eﬀect of the discretization. 36. as they guarantee that structures can be detected easily and correspondence problems can be neglected. Scalespace interpretation Perona and Malik renounced the assumption of Koenderink’s linear scalespace axiomatic [240] that the smoothing should treat all spatial points and scale levels equally. which guarantees that no additional oscillations occur during the evolution. There exists also a discrete explanation why staircasing is essentially the only observable instability: In 1D. since in scalespace theory one is interested in having an extremum principle for the entire time interval [0. ∞).1. 438]. In order to establish a smoothing scalespace property for this nonlinear diffusion process. 425] in the fully discrete case with an explicit time discretization. also other attempts to apply scalespace frameworks to the Perona– Malik process have not been more successful yet: • Salden [350]. Finer discretizations are less regularizing and lead to more “stairs”. 98. Experiments demonstrated that the Perona–Malik ﬁlter satisﬁes these requirements fairly well [328]. Weickert and Benhamouda [425] showed that the regularizing eﬀect of a standard ﬁnite diﬀerence discretization is suﬃcient to turn the Perona–Malik ﬁlter into a wellposed initial value problem for a nonlinear system of ordinary diﬀerential equations. These properties are of signiﬁcant practical interest. and Benhamouda o [36]. and that intraregion smoothing should be preferred to interregion smoothing (piecewise smoothing). provided one knows that there exists a suﬃciently smooth solution. Nevertheless. numerical studies of the staircasing eﬀect have been carried out by Nitzberg and Shiota [310]. they required that region boundaries should be sharp and should coincide with the semantically meaningful boundaries at each resolution level (immediate localization).
43) In [81] existence. This process has been analysed and modiﬁed in many ways: Whitaker and Pizer [438] have suggested that the regularization parameter σ should be .20 CHAPTER 1. 1. (1. However. A mathematically sound formulation of this idea is given by Catt´. Guichard. Below we shall discuss three examples which illustrate the variety of possibilities and their tailoring towards a speciﬁc task. Hence. e 2 Morel and Coll [81].5 and 1. Unfortunately. but not for all t > 0. (a) The ﬁrst spatial regularization attempt is probably due to Posmentier who observed numerically the stabilizing eﬀect of averaging the gradient within the diﬀusivity [333]. Lions. One can apply spatial or temporal regularization (and of course. the Perona–Malik ﬁlter turned out not to belong to this class [143]. it has been suggested to introduce the regularization directly into the continuous equation in order to become more independent of the numerical implementation [81. a combination of both). PARTIAL DIFFERENTIAL EQUATIONS processes which result from special rescalings of the linear one.3. • Alvarez. 310]. one should adjust the regularization to the desired goal of the forward– backward heat equation [35].2 Regularized nonlinear models It has already been mentioned that numerical schemes may provide implicit regularizations which stabilize the Perona–Malik process [425]. uniqueness and regularity of a solution for σ > 0 have been established.42) This property is closely related to a maximum–minimum principle and to L∞ stability of the solution [12.6). Their smoothing axiom is a monotony assumption (comparison principle) requiring that the scalespace is orderpreserving: f ≤ g =⇒ Tt f ≤ Tt g ∀ t ≥ 0. 261]. the Perona–Malik model does not ﬁt into this framework. because its local weak solution satisﬁes a comparison principle only for some ﬁnite time. Lions and Morel [12] have developed a nonlinear scalespace axiomatic which comprises the linear scalespace theory as well as nonlinear morphological processes (which we will discuss in 1. see [222]. By replacing the diﬀusivity g(∇u ) of the Perona– Malik model by a Gaussiansmoothed version g(∇uσ 2 ) with uσ := Kσ ∗ u they end up with ∂t u = div (g(∇uσ 2 ) ∇u). Since the dynamics of the solution may critically depend on the sort of regularization. (1.
and Li and Chen [252] have proposed to subsequently decrease the contrast parameter λ. Spatial regularizations of the Perona–Malik process leading to anisotropic diﬀusion equations have been proposed by Weickert [413. Therefore. it exhibits besides the contrast parameter λ an additional noise scale σ. These equations arise as a special case of the spatiotemporal regularizations of Nitzberg and Shiota [310] when neglecting any spatial regularization.44) (1.1.4 (a). We observe that v is intended as a timedelay regularization of ∂x u2 where the parameter τ > 0 determines the delay. . Examples for spatially regularized nonlinear diﬀusion ﬁltering can be found in Figure 5. Kaˇur and Mikula [217] have investigated a modiﬁcac tion which allows to diﬀuse diﬀerently in diﬀerent grey value ranges.43) as an image restoration equation.(b). (b) P. This avoids a shortcoming of the genuine Perona–Malik process which misinterprets strong oscillations due to noise as edges which should be preserved or even enhanced. ∂t v = 1 τ (1.3.3 NONLINEAR DIFFUSION FILTERING 21 a decreasing function in t. when regarding (1. 415] and will be described in 1.46) 10 This renounces invariance under rotation. Mumford conjectures that this model gives piecewise constant steadystates. Torkamani–Azar and Tait [403] suggest to replace the Gaussian convolution by the exponential ﬁlter of Shen and Castan10 [381]. A detailed study of the inﬂuence of the parameters in a regularized Perona–Malik model has been carried out by Benhamouda [36].L. the steadystate solution would solve a segmentation problem. (c) In the context of shear ﬂows.3. In Chapter 2 we shall see that spatial regularizations lead to wellposed scalespaces with a large class of Lyapunov functionals which guarantee that the solution converges to a constant steadystate. spatial regularizations oﬀer the advantage that they make the ﬁlter insensitive to noise at scales smaller than σ. Lions proved in a private communication to Mumford that the onedimensional process ∂t u = ∂x (g(v) ∂x u). [329]). Barenblatt et al. [35] regularized the onedimensional forward–backward heat equation by considering the thirdorder equation ∂t u = ∂x (Φ(ux )) + τ ∂xt (Ψ(ux )) (1.2 (c) and 5. In this case.45) (∂x u −v) 2 leads to a wellposed ﬁlter (cf. From a practical point of view.
11 An exception is the timedelay regularization of Cottet and ElAyyadi [100. Anisotropic diﬀusion ﬁlters usually apply spatial regularization strategies11 . 1. First anisotropic ideas in image processing date back to Graham [167] in 1962. Nevertheless. before they ﬁnally converge to a piecewise constant steadystate. The ﬁrst one oﬀers advantages at noisy edges. These requirements cannot be satisﬁed by a scalar diﬀusivity anymore.3. a diﬀusion tensor leading to anisotropic diﬀusion ﬁlters has to be introduced. respectively. This regularization was physically motivated by introducing a relaxation time τ into the diﬀusivity. PARTIAL DIFFERENTIAL EQUATIONS where Ψ is strictly increasing and uniformly bounded in IR. These examples demonstrate that regularization is much more than stabilizing an illposed process: Regularization is modeling. They used convolution masks that depended on the underlying image structure. Zucker and Rosenfeld [250]. . whereas the second one is welladapted to the processing of onedimensional features. They are called edgeenhancing anisotropic diﬀusion and coherenceenhancing anisotropic diﬀusion. Therefore. followed by Newman and Dirilten [300]. A general theoretical framework for spatially regularized anisotropic diﬀusion ﬁlters will be presented in the remaining chapters of this book. 101]. These ideas have been further developed by Nitzberg and Shiota [310].6. Their suggestion to use shapeadapted Gaussian masks has been discussed in Section 1. Appropriately chosen regularizations create the desired ﬁlter features. Lindeberg and G˚ arding [259]. in certain applications it would be desirable to bias the ﬂux towards the orientation of interesting features. Below we study two representatives of anisotropic diﬀusion processes. [443]. We observe that spatial regularizations are closer to scalespace ideas while temporal regularization are more related to image restoration and segmentation. and Nagao and Matsuyama [297]. since they may lead to nontrivial steadystates. they are isotropic and the ﬂux j = −g∇u is always parallel to ∇u. They also showed that smooth solutions may become discontinuous within ﬁnite time.2. and Φ′ (s) = O(Ψ′(s)) as s → ±∞. For the corresponding initial boundary value problem with homogeneous Neumann boundary conditions they proved the existence of a unique generalized solution.22 CHAPTER 1. Wilson and Granlund [237]. and Yang et al. Lev. Related statistical approaches were proposed by Knutsson.3 Anisotropic nonlinear models All nonlinear diﬀusion ﬁlters that we have investigated so far utilize a scalarvalued diﬀusivity g which is adapted to the underlying image structure.
Therefore. we end up with the isotropic Perona–Malik process. the process stops completely. (1. we observe that ∇u becomes an eigenvector of D with corresponding eigenvalue 0. If we let the regularization parameter σ tend to 0. ∇u does not coincide with one of the eigenvectors of D as long as σ > 0. Wellposedness . For σ → 0.43) behaves almost like the linear diﬀusion ﬁlter (1. To this end. Therefore. we construct the orthonormal system of eigenvectors v1 .47) In order to prefer smoothing along the edge to smoothing across it. Another anisotropic model which can be regarded as a regularization of an isotropic nonlinear diﬀusion ﬁlter has been described in [413]. λ2 (∇uσ ) := η∇uσ  1 + (∇uσ /σ)2 2 (1.9). (b) Anisotropic models for smoothing onedimensional objects A second motivation for introducing anisotropy into diﬀusion processes arises from the wish to process onedimensional features such as linelike structures. Hence.1.47) and corresponding eigenvalues λ1 (∇uσ ) := 0. Cottet and Germain [102] constructed a diﬀusion tensor with eigenvectors as in (1.3 NONLINEAR DIFFUSION FILTERING 23 (a) Anisotropic regularization of the Perona–Malik process In the interior of a segment the nonlinear isotropic diﬀusion equation (1.48) (1. but also its direction.50) (η > 0). v2 of the diﬀusion tensor D such that they reﬂect the estimated edge structure: v1 ∇uσ . this model behaves really anisotropic. v2 ⊥ ∇uσ . (1. To overcome this problem. a desirable method should prefer diﬀusion along edges to diﬀusion perpendicular to them. To this end.49) Section 5. noise at edges cannot be eliminated successfully by this process. but at edges diﬀusion is inhibited.1 presents several examples where this process is applied to test images. λ2 (∇uσ ) := 1. In this sense. (1.51) This is a process diﬀusing solely perpendicular to ∇uσ . In general. Anisotropic models do not only take into account the modulus of the edge detector ∇uσ . it is not intended as an anisotropic regularization of the Perona–Malik equation. Weickert [415] proposed to choose the corresponding eigenvalues λ1 and λ2 as λ1 (∇uσ ) := g(∇uσ 2 ).
which shall be discussed in Section 5. The edge detector ∇uσ enables us to adapt the diﬀusion to magnitude and direction of edges. A generalization c of coherenceenhancing anisotropic diﬀusion to higher dimensions is proposed in [428]. Feature vectors play an important role for tasks like texture segmentation. multispectral Landsat exposures and multispin echo MR images. one can improve this model when replacing ∇u⊥ σ by a more robust descriptor of local orientation. .2. the structure tensor (cf. It has also been proposed to replace ∇uσ by a Bayesian classiﬁcation result in feature space [26]. It is easily seen that many of the previous results can be generalized to higher dimensions.g.24 CHAPTER 1. 418]. Examples for the ﬁrst category are colour images. The ﬁrst threedimensional nonlinear diﬀusion ﬁlters have been investigated by Gerig et al.2). one can steer the smoothing process by more advanced structure descriptors such as higherorder derivatives [127] or tensorvalued expressions of ﬁrstorder derivatives [414. Section 2. For enhancing parallel linelike structures. or when applying diﬀusion ﬁlters to the postprocessing of ﬂuctuating higherdimensional numerical data. Vectorvalued images can arise either from devices measuring multiple physical properties or from a feature analysis of one single image.3. Since the Cottet–Germain model diﬀuses only in one direction. medical image sequences from computerized tomography (CT) or magnetic resonance imaging (MRI). PARTIAL DIFFERENTIAL EQUATIONS results for the Cottet–Germain ﬁlter comprise an existence proof for weak solutions. Vectorvalued models. [355] describe nonlinear diﬀusion ﬁltering of 3D a image sequences by treating them as 4D data sets. [155] in the isotropic case and by Rambaux and Gar¸on [339] in the anisotropic case. As a remedy. More sophisticated structure descriptors. The theoretical analysis in the present work shall comprise the second possibility. and S´nchez–Ortiz et al. 1. but it can neither distinguish between edges and corners nor does it give a reliable measure of local orientation.4 Generalizations Higher dimensions. it is clear that its result depends very much on the smoothing direction. This leads to coherenceenhancing anisotropic diﬀusion [418]. where also many examples can be found. whereas representatives of the second class are given by statistical moments or the jet space induced by the image itself and its partial derivatives up to a given order. This may be useful when considering e.
since they are easy to handle and the pixel structure of digital images already provides a natural discretization on a ﬁxed rectangular grid. see also [349] for related ideas. [155. Neural network approximations to nonlinear diﬀusion ﬁlters are investigated by Cottet [100. They possess much better stability properties. B¨nsch and a Mikula reported a signiﬁcant speedup by supplementing them with an adaptive mesh coarsening [34]. [158].3 NONLINEAR DIFFUSION FILTERING 25 The simplest idea how to apply diﬀusion ﬁltering to multichannel images would be to diﬀuse all channels separately and independently from each other. 391. 99] and Fischl and Schwartz [137]. 434] in the context of medical imagery. it seems to be a matter of taste which scheme is preferred. 1. it is often desirable to have a genuinely discrete theory which guarantees that an algorithm exactly reveals the same qualitative properties as its continuous counterpart. a pseudospectral method and a ﬁnitediﬀerence scheme. they are wellsuited for parallel architectures. This leads to the undesirable eﬀect that edges may be formed at diﬀerent locations for each channel. 34.1. 4]. While the preceding techniques are focusing on approximating a continuous equation. Perona and Malik [327] propose hardware realizations by means of analogue VLSI networks with nonlinear resistors. therefore. A fast multigrid technique using a pyramid algorithm for the Perona–Malik ﬁlter has been studied by Acton et al. Most implementations of nonlinear diﬀusion ﬁlters are based on ﬁnite diﬀerence methods. one should use a common diﬀusivity which combines information from all channels. Semiimplicit schemes – which approximate the diﬀusivity or the diﬀusion tensor in an explicit way and the rest implicitly – are considered in [81]. It turned out that all results became fairly similar. 421]. 156] and Whitaker [433. In [148] three schemes for a spatially regularized 1D Perona–Malik ﬁlter are compared: a wavelet method of Petrov–Galerkin type. Due to their local behaviour. when the regularization parameter σ was suﬃciently large. In order to avoid this.3.5 Numerical aspects For nonlinear diﬀusion ﬁltering numerous numerical methods have been applied: Finite element techniques are described in [367. [5. Extensions to anisotropic vectorvalued models with a common tensorvalued structure descriptor for all channels have been investigated by Weickert [422]. Explicit schemes are the most simple to code and. A very detailed VLSI proposal has been developed by Gijbels et al. Nevertheless. both for the semidiscrete . they are used almost exclusively. they suﬀer from the fact that fairly small time step sizes are needed in order to ensure stability. Such isotropic vectorvalued diﬀusion models were studied by Gerig et al. Such a framework is presented in [420. 216]. Since the computational eﬀort is of a comparable order of magnitude.
26 CHAPTER 1. are 0 positive nonnegative semideﬁnite oﬀdiagonals uniformly pos. pos. conservation. 415]. and convergence to a constant steadystate [423]. A detailed treatment of this theory can be found in Chapter 3 and 4. 1. symmetry. they are about 10 times more eﬃcient than the widely used explicit schemes. and a speedup by another order of magnitude can be achieved by a parallel implementation [431]. diagonal smoothness symmetry conservation nonnegativity connectivity and for the fully discrete case. They have proved to be useful for improving subsampling [144] and line detection [156. These criteria are easy to check for many discretizations.2: Requirements for continuous. average grey value invariance. 413. Table 1.4. 418]. In particular. respectively. column sums reﬂective b. n = 0 D ∈ C∞ A Lipschitzcontinuous D symmetric A symmetric div form. and for enhancing textures such as ﬁngerprints [418].3. Under typical accuracy requirements. causality in terms of an extremum principle and Lyapunov functionals. it turns out that suitable explicit and semiimplicit ﬁnite diﬀerence discretizations of many discussed models create discrete scalespaces.6 Applications Nonlinear diﬀusion ﬁlters have been applied for postprocessing ﬂuctuating data [269. 418].2.2 gives an overview of the requirements which are needed in order to prove wellposedness. nonnegativity and connectivity requirements. for visualizing qualityrelevant features in computer aided quality control [299. semidiscrete and fully discrete nonlinear diﬀusion scalespace. requirement continuous semidiscrete ut = div (D∇u) du = A(u)u dt u(t = 0) = f u(0) = f D∇u. irreducible deﬁnite discrete u0 = f uk+1 = Q(uk )uk Q continuous Q symmetric column sums are 1 nonnegative elements irreducible. The discrete nonlinear scalespace concept has also led to the development of fast novel schemes. 430].c. PARTIAL DIFFERENTIAL EQUATIONS Table 1. We observe that the requirements belong to ﬁve categories: smoothness. A general framework for AOS schemes will be presented in Section 4. for blind image restoration [445]. for scalespace based segmentation algorithms . which are based on an additive operator splitting (AOS) [424.
200 First Street SW. This idea connects diﬀusion methods to edge detection and segmentation ideas. 387]. Cost terms of these three types are typical for variational image restoration methods. 270. 1] gives a fuzzy edge representation: in the interior of a region. The corresponding Euler equations to this energy functional are given by 0 = β ·(u−f ) − div (w∇u). 5]. 264.52) The parameters β and λ are positive weights and w : Ω → [0. w is close to 0 (as we shall see below). U.54) + ∇u . 248. 321.1 Single diﬀusion–reaction equations Nordstr¨m [311] has suggested to obtain a reconstruction u of a degraded image o f by minimizing the energy functional Ef (u. 308. 251. Their discussion. 444]. 355. 247. w) := Ω β ·(u−f )2 + w·∇u2 + λ2 ·(w−ln w) dx. in which diﬀusion–reaction equations or coupled systems of them are interpreted as steepest descent minimizers of suitable energy functionals. 2 Analyze is a registered trademark of Mayo Medical Ventures. Besides such speciﬁc problem solutions. MN 55905. 1. 437. and for target tracking in infrared images [65]. Some of these applications will be investigated in more detail in Chapter 5.12 1. 0 = λ 2 1 ·(1− w ) (1. nonlinear diﬀusion ﬁlters can be found in commercial software packages such as the medical visualization tool Analyze. 28. Besides the variational interpretation there exist other interesting theoretical frameworks for diﬀusion ﬁlters such as the Markov random ﬁeld and mean ﬁeld annealing context [152.4. 153. for segmentation of textures [433. 29.4 METHODS OF DIFFUSION–REACTION TYPE 27 [307. Most applications. Rochester. are concerned with the ﬁltering of medical images [26.A.1. would lead us beyond the scope of this book. however. w approaches 1 while at edges.S. 437] and remotely sensed data [6. The ﬁrst summand of E punishes deviations of u from f (deviation cost). however.53) (1. the second term detects unsmoothness of u within each region (stabilizing cost). 328. robust statistics [41]. 155. and deterministic interactive particle models [279]. 308]. and the last one measures the extend of edges (edge cost). 434.4 Methods of diﬀusion–reaction type This section investigates variational frameworks. 393. 431. 12 . (1. 386. 244.
Diﬀusion–reaction methods with constant diﬀusivities have also been used for local contrast normalization in images [330].28 CHAPTER 1. 1 + ∇u2/λ2 (1.32). A popular possibility to avoid these illposedness and convergence problems is to renounce edgeenhancing diﬀusivities in order to end up with (nonquadratic) convex functionals [43. Similar diﬃculties may also arise in other diﬀusion–reaction models. 111] and stereo problems [343]. to SPECT [88] and vascular reconstruction in medical imaging [102. They can be extended to vectorvalued images [369] and to cornerpreserving smoothing of curves [136. where convergence results have not yet been established [152. and the process (1. Nordstr¨m’s method may suﬀer from the same illposedness problems as the o underlying Perona–Malik equation.57) is nonconvex. to the restoration of inverse scattering images [263]. 88. Therefore. Solving (1. it can possess numerous local minima. People interested in image restoration usually prefer the reaction term. 110.54) for w gives w= 1 . (1. 323]. 325]. However.56) This equation can also be obtained directly as the descent method of the functional Ff (u) := Ω β ·(u−f )2 + λ2 ·ln 1+ ∇u λ2 2 dx. (1. ensuring wellposedness and stability of a standard ﬁniteelement approximation [367]. One of Nordstr¨m’s motivations for introo ducing this term was to free the user from the diﬃculty of specifying an appropriate stopping time for the Perona–Malik process. Therefore. In this case the frameworks of convex optimization and monotone operators are applicable.56) consists of the Perona–Malik process with an additional bias term β ·(f −u).55) We recognize that w is identical with the Perona–Malik diﬀusivity g(∇u2) introduced in (1. PARTIAL DIFFERENTIAL EQUATIONS equipped with a homogeneous Neumann boundary condition for u. 186].53) can be regarded as the steadystate equation of ∂t u = div (g(∇u2) ∇u) + β(f −u). 391]. it is evident that the Nordstr¨m model just shifts the problem of o specifying a stopping time T to the problem of determining β. So it seems to be a matter of taste which formulation is preferred. .57) The diﬀusion–reaction equation (1. and to optic ﬂow [368.56) with f as initial condition does not necessarily converge to a global minimum. Diﬀusion–reaction approaches have been applied to edge detection [367. and it is not hard to verify that the energy functional (1. 391]. while for scalespace researchers it is more natural to have a constant steadystate as the simplest image representation. (1. 367.
the problem that sigmoidlike edges produce multiple segmentation boundaries (oversegmentation.58) is investigated in [32.4 METHODS OF DIFFUSION–REACTION TYPE 29 1. The Mumford–Shah functional can be regarded as a continuous version of the Markov random ﬁeld method of Geman and Geman [154] and the weak membrane model of Blake and Zisserman [42]. Although the existence of a global minimizer with a closed edge set K has been established [108.58) with nonnegative parameters α and β. 48. uniqueness is in general not true [292.52). The discontinuity set K consists of the edges.g. Regularity results for K in terms of (at least) C 1 arcs have recently been obtained [18.g. The concept of energy functionals for segmenting images oﬀers the practical advantage that it provides a framework for comparing the quality of two segmentations.4. In order to avoid such problems. 239]. e. 17]. one searches for fast (suboptimal) deterministic strategies.58) exhibits also some shortcomings.1. 20. 19. Numerical complications arise from the fact that the Mumford–Shah functional has numerous local minima. (1. On the other hand. . 107]. modiﬁcations of the Mumford–Shah functional have been proposed by Shah [379]. Global minimizers such as the simulated annealing method used by Geman and Geman [154] are extremely slow. Since many algorithms in image processing can be restated as versions of the Mumford–Shah functional [292] and since it is a prototype of a free discontinuity problem it is instructive to study this variational problem in more detail. and a Mumford–Shah functional for curves can be found in [323].2 Coupled systems of diﬀusion–reaction equations Mumford and Shah [295. the second one gives the stabilizing cost. The book of Morel and Solimini [292] covers a very detailed analysis of this functional. and the third one represents the edge cost. An aﬃne invariant generalization of (1. staircasing eﬀect) [377]. The fact that (1. 296] have proposed to obtain a segmented image u from f by minimizing the functional Ef (u. Hence. pyramidal regiongrowing algorithms [3. 31] and applied to aﬃne invariant texture segmentation [31. this expression consists of three cost terms: o the ﬁrst one is the deviation cost.58) leads to a free discontinuity problem causes many challenging theoretical questions [249]. 197–198]. e. K) = β Ω (u−f )2 dx + Ω\K ∇u2 dx + αK (1. 49. Related approaches are also used to model materials with two phases and a free interface. 33]. Like the Nordstr¨m functional (1. Another drawback results from the fact that the Mumford–Shah functional allows only singularities which are typical for minimal surfaces: Corners or Tjunctions are not possible and segments meet at triple points with 120o angle [296]. and its onedimensional Hausdorﬀ measure K gives the total edge length. pp.
60) resembles the Nordstr¨m o process (1. 335.61) ∂t w = c∆w − + (1−w) 4c with homogeneous Neumann boundary conditions.58) by two coupled convex energy functionals and applies gradient descent.56). Proesmans et al. this approach gives rise to the same theoretical questions as (1. 440] and halftoning [382]. which is close to 0 near edges of u and which approximates 1 elsewhere. w) := Ω β ·(u−f )2 + w 2 ·∇u2 + α· c∇w2 + (1−w) 4c 2 dx (1. Wellposedness results for this system have not been obtained up to now. Since (1.59) with a positive parameter c specifying the “edge width”. Shah investigates diﬀusion–reaction systems for matching stereo images [378]. They are based on Turing’s pattern formation model [405].60). similar problems can arise: The functional Ff is not jointly convex in u and v. Related equations are also studied in [398]. Minimizing Ff corresponds to the gradient descent equations ∂t u = div (w 2 ∇u) + β ·(f −u). vectorvalued images and stereo vision [336. while Proesmans et al.30 CHAPTER 1. This results in ﬁnding an equilibrium state between two competing processes. Another diﬀusion–reaction system is studied by Shah [375. Their ﬁnite diﬀerence algorithms run on a parallel transputer network. Of course. 376]. texture generation [406. He replaces the functional (1. Ambrosio and Tortorelli proved that this functional converges to the Mumford–Shah functional for c → 0 (in the sense of Γconvergence. They obtained pronounced edges by replacing such a term by its Perona–Malik counterpart div (g(∇u2) ∇u).61). We may for instance study the functional Ff (u. 336] observed that this solution looks fairly blurred since the equations contain diﬀusion terms such as ∆u.60) (1. . but a maximum–minimum principle and a local stability proof have been established. PARTIAL DIFFERENTIAL EQUATIONS Another important class of numerical methods is based on the idea to approximate the discontinuity set K by a smooth function w. Equations of this type are investigated by Richardson and Mitter [341]. It should also be mentioned that there exist reaction–diﬀusion systems which have been applied to image restoration [334. apply coupled diﬀusionreaction equations to image sequence analysis. see [22] for more details). The system of Richardson and Mitter is used for edge detection [341]. [337. so it may have many local minima and a gradient descent algorithm may get trapped in a poor local minimum. 382]. and which are not connected to Perona–Malik or Mumford– Shah ideas. Experiments indicate that it converges to a stable solution. w ∇u2 α (1. (1. 338].
χ(x) := (1.63) A binary morphological operation A can be extended to some greyscale image f by deﬁning Xa (Af ) := A(Xa f ) ∀ a ∈ IR. a ∈ IR}. As a consequence. [12]. therefore. The theory had ﬁrst been developed for binary images.6. both working at ENS des Mines de Paris in Fontainebleau.g. Binary morphological operations aﬀect only the boundary curve of the shape and. f (x) ≥ a}.5 CLASSIC MORPHOLOGICAL PROCESSES 31 1. Arehart et al. we identify a shape X with its characteristic function 1 if x ∈ X. [280. 371.5. afterwards it was extended to greyscale images by regarding level sets as shapes. Greyscale morphology generalizes these ideas [274] by decomposing an image f into its level sets {Xa f. radar and remote sensing.5 and 1. Afterwards numerical aspects of the PDE formulation of these processes are discussed. Henceforth. they are invariant under monotone greylevel rescalings. and summarizes the results concerning wellposedness and scalespace properties. see e. mineralogy. 184] for an overview. This section surveys the basic ideas and elementary operations of binary and greyscale morphology. [25] and Alvarez et al. Nevertheless.1. and generalizations are sketched which comprise the morphological equivalent of Gaussian scalespace. except for 1.1 Binary and greyscale morphology Binary morphology considers shapes (silhouettes).5 Classic morphological processes Morphology is an approach to image analysis based on shapes. 1. Morphology is usually described in terms of algebraic set theory. Its mathematical formalization goes back to the group around Matheron and Serra. (1.62) 0 else. where Xa f := {x ∈ IR2 . presents its PDE representations for images and curves.64) We observe that for this type of morphological operations only greylevel sets matter. 181. Its applications cover biology. quality control.6 and some modiﬁcations . i. This morphological invariance (greyscale invariance) is characteristic for all methods we shall study in Section 1. they can be viewed as curve or shape deformations. (1.e. histology. closed sets X ⊂ IR2 whose boundaries are Jordan curves [16].5. and many others. PDE formulations for classic morphological processes have been discovered recently by Brockett and Maragos [60]. medical diagnostics. van den Boomgaard [50]. science of material.
3 Continuousscale morphology Let us consider a convex structuring element tB with a scaling parameter t > 0. 13 Henceforth. (1. dilation and erosion with a structuring element B. t) . PARTIAL DIFFERENTIAL EQUATIONS in 1. respectively13 . for applications like edge detection or image restoration. a bounded set B ⊂ IR2 . 1. (1.2 Basic operations Classic morphology analyses a shape by matching it with a socalled structuring element. y∈B (1.68) (f • B) (x) := ((f ⊕ B) ⊖ B) (x). t) = sup y. Then. are deﬁned for a greyscale image f ∈ L∞ (IR2 ) by [60] dilation: erosion: (f ⊖ B) (x) := inf {f (x+y). y ∈ B}.69) (1.65) (1. t) . The two basic morphological operations. contrast provides important information which should be taken into account. Moreover. can be shown to be equivalent to solving ∂t u(x.5. for instance opening and closing: opening: closing: (f ◦ B) (x) := ((f ⊖ B) ⊕ B) (x).. calculating u(t) = f ⊕ tB and u(t) = f ⊖ tB.67) (1. 360].70) ∂t u(x. Dilation and erosion form the basis for constructing other morphological processes. It is a very desirable property in all cases where brightness changes of the illumination occur or where one wants to be independent of the speciﬁc contrast range of the camera. (f ⊕ B) (x) := sup {f (x−y). we frequently use the simpliﬁed notation u(t) instead of u(. or ellipses. On the other hand. while closing smoothes by eliminating small holes.5.32 CHAPTER 1. ∇u(x. fusing narrow breaks and ﬁlling gaps at the contours [181]. y ∈ B}. Typical shapes for B are discs. while erosion shrinks them.66) These names can be easily motivated when considering a shape in a binary image and a discshaped structuring element.6. y∈B with f as initial condition [12.5. 1. t) = inf y. in some cases isolines may give inadequate information about the depicted physical object boundaries. t) . ∇u(x. squares. In the preceding shape interpretation opening smoothes the shape by breaking narrow isthmuses and eliminating small islands. In this case dilation blows up its boundaries.
y ≤ 1} one obtains ∂t u = ∇u.78) This equation moves level sets in normal direction with constant speed. We observe that (1. (1.74) (1.72) correspond to the simple cases β = ±1.74). ∂t C = β(κ) · n. 0) = C0 (p).Cpp ) : Cp 3 C(p. Connection to curve evolution Morphological PDEs such as (1. Its importance for shape analysis in biological vision has already been pointed out in the sixties by Blum [47]. B := {y ∈ IR2 . 2π] × [0. ∇u (1.76) Sometimes the image evolution (1. where the curvature of u is curv(u) := div ∇u .75) One can embed the curve C(p. It is closely related to the Huygens principle for wave propagation [25]. because it is written in terms of a ﬁxed coordinate system. 33 (1. t) x2 (p.73) where p is the parametrization and t is the evolution parameter.71) (1.71) or (1.72) are closely related to shape and curve evolutions. t) in such a way that C is just a level curve of u.76) is called the Eulerian formulation of the curve evolution (1. erosion) of f with a disc of radius t and centre 0 as structuring element. The solution u(t) is the dilation (resp.77) (1. He simulated grassﬁre ﬂow by a selfconstructed opticomechanical device. Hence. Such a process is also named grassﬁre ﬂow or prairie ﬂow. This can be illustrated by considering a smooth Jordan curve C : [0. (1. . t) (1. C(p. which may be a function of its curvature κ := det(Cp . 16] ∂t u = β(curv(u)) · ∇u. The corresponding evolution for u is given by [319.g.72) ∂t u = −∇u.5 (a) presents the temporal evolution of a test image under such a continuousscale dilation.5 CLASSIC MORPHOLOGICAL PROCESSES By choosing e. ∞) → IR2 . they describe the curve evolutions ∂t C = ± n. Figure 5. 362.71) and (1. t) = x1 (p. t) in an image u(x. We assume that C evolves in outer normal direction n with speed β.1.
Lions and Morel [12]: They prove that under three architectural assumptions (semigroup property. while dilation reduces the number of local maxima. This establishes an important architectural scalespace property. which is also based on local extrema instead of zerocrossings. t) ≤ sup f 2 IR IR 2 on IR2 × [0. t) which fulﬁls the maximum–minimum principle inf f ≤ u(x. A suitable framework for the image evolution equation (1. Jackway and Deriche [206.5 Scalespace properties Brockett and Maragos [60] pointed out that the convexity of B is suﬃcient to ensure the semigroup property of the corresponding dilations and erosions. one smoothing axiom (comparison principle) and additional invariance requirements (greylevel shift invariance. Guichard. Hence.34 CHAPTER 1. entropy solutions. that for an initial value f ∈ BUC(IR2 ) := {ϕ ∈ L∞ (IR2 )  ϕ is uniformly continuous on IR2 } (1. a twodimensional scalespace . it is L∞ stable: for two diﬀerent initial images f .70) possess a unique global viscosity solution u(x. 129.(1.81) 1. Similar results have been found by van den Boomgaard and Smeulders [53].5.78) may develop singularities and intersections even for smooth initial data. morphological invariance). 207] prove a causality theorem for the dilation– erosion scalespace. PARTIAL DIFFERENTIAL EQUATIONS 1. Moreover. 103]. concepts of jump conditions.79) the equations (1. invariance under rotations and translations.69). locality and regularity).76) is provided by the theory of viscosity solutions [103]. but a unique weak solution in the viscosity sense still exists.4 Theoretical results Equations such as (1. where the classical solution concept does not apply.80) Moreover.5. (1. (1. They establish that under erosion the number of local minima is decreasing. and shocks have to be applied to this shape evolution [228]. A complete scalespace interpretation is due to Alvarez. The advantage of this analysis is that it allows us to treat shapes with singularities such as corners. v(t) satisfy u(t)−v(t) L∞ (IR2 ) ≤ f −g L∞ (IR2 ) . g the corresponding solutions u(t). The location of these extrema is preserved during their whole lifetime. they conjecture a causality property where singularities play a role similar to zerocrossings in Gaussian scalespace. It can be shown [90. ∞).
A dilation of an image f with a structuring function b : IR2 → IR is deﬁned as (f ⊕ b) (x) := sup {f (x−y) + b(y)}. 51]. 53] that the result u(x.83) This is a generalization of deﬁnition (1. (1. In this case we have to renounce invariance under monotone grey level transformations.5 (b). but we gain an interesting insight into a process which has very much in common with Gaussian scalespace. since one can recover dilation with a structuring element B by considering the ﬂat structuring function b(x) := 0 −∞ (x ∈ B).84) Van den Boomgaard [50. . y∈IR2 (1. 0) = f (x). (1.87) u(x.5 CLASSIC MORPHOLOGICAL PROCESSES equation has the following form: ∂t u = ∇u F (t. thus being good candidates for morphological scalespaces. t) = a exp( x ) are the 4t only rotationally symmetric kernels which are separable with respect to convolution. 51] and Jackway [206] proposed to dilate an image f (x) with quadratic structuring functions of type b(x. curv(u)) 35 (1.65).86) (1.82) Clearly. van den Boomgaard proves that the quadratic structuring functions b(x. t) are the only rotationally invariant structuring functions which are separable with respect to dilation [50. in [12] it is shown that the converse is true as well: all axioms that lead to (1. (1. 2 In analogy to the fact that Gaussiantype functions k(x. dilation and erosion belong to the class (1. 14 Invariance under rotations is only satisﬁed for a disc centered in 0 as structuring element. Indeed.1. t) is a weak solution of ∂t u = ∇u2 .6 Generalizations It is possible to extend morphology with a structuring element to morphology with nonﬂat structuring functions. The temporal evolution of a test image under this process is illustrated in Figure 5.14 1.85) It can be shown [50.82). t) = − x2 4t (t > 0).82) are fulﬁlled.5. (x ∈ B).
25. nτ ). 360]. let us restrict ourselves to the onedimensional dilation equation ∂t u = ∂x u. A ﬁrstorder upwind Osher–Sethian scheme for this process is given by n un+1 − ui i = τ min un − un i i−1 . and suﬀer from the problem of coping with singularities and topological changes [319. and it is possible to get results with subpixel accuracy. It has been discovered simultaneously by Dorst and van den Boomgaard [119] and by Maragos [275] in slightly diﬀering formulations. Therefore. For this reason it is useful to discretize the corresponding image evolution equations. for instance discs or ellipses. 218. This generalization of the Legendre transform is the morphological equivalent of the Fourier transform.0 h 2 + max un − un i+1 i . 374].36 CHAPTER 1.0 h 2 .5. A fast algorithm is described by van den Boomgaard [51].88) where h is the pixel size. The widelyused Osher–Sethian schemes [319] are based on the idea to derive numerical methods for such equations from techniques for hyperbolic conservation laws. To illustrate the basic idea with a simple example. thus. and un denotes a discrete i approximation of u(ih. (b) The time t plays the role of a continuous scale parameter. The close relation between (1. 66]: (a) They give excellent results for nondigitally scalable structuring elements whose shapes cannot be represented correctly on a discrete grid. . Overviews of these level set approaches and their various applications can be found in [372. PARTIAL DIFFERENTIAL EQUATIONS A useful tool for understanding this similarity and many other analogies between morphology and linear systems theory is the slope transform.86) as limiting case.7 Numerical aspects Dilations or erosions with quadratic structuring functions are separable and. they can be implemented very eﬃciently by applying onedimensional operations. the situation is more complicated. Level set methods possess two advantages over classical settheoretic schemes for dilation/erosion [25. the size of a structuring element need not be multiples of the pixel size. For morphological operations with ﬂat structuring elements.86) and Gaussian scalespace has also triggered Florack and Maas [142] to study a oneparameter family of isomorphisms of linear diﬀusion which reveals (1. (1. 360. τ is the time step size. Schemes for dilation or erosion which are based on curve evolution turn out be be diﬃcult to handle: they require prohibitive small time steps. 1.
numerical aspects.6.6 CURVATUREBASED MORPHOLOGICAL PROCESSES However. while the second one smoothes along isophotes.1. distance transformations. gridless image halftoning.1 Meancurvature ﬁltering In order to motivate our ﬁrst curvaturebased morphological PDE. 1. These processes are curvaturebased. Two important representatives of this class are mean curvature motion and its aﬃne invariant counterpart.6 Curvaturebased morphological processes Besides providing a useful reinterpretation of classic continuousscale morphology. the PDE approach has led to the discovery of new morphological operators. they reveal also two disadvantages: (a) They are slower than settheoretic morphological schemes. 276]. (b) Dissipative eﬀects such as blurring of discontinuities occur. 1. 373]. let us recall that the linear diﬀusion equation (1. In this subsection we shall discuss these PDEs. 15 . 435. The ﬁrst term on the righthand side of (1.8 Applications Continuousscale morphology has been applied to shapefromshading problems.89) describes smoothing along the ﬂowlines. and applications.5.89) where the unit vectors η and ξ are parallel and perpendicular to ∇u.9) can be rewritten as ∂t u = ∂ηη u + ∂ξξ u. 385]. Applications outside the ﬁeld of image processing and computer vision include for instance shape oﬀsets in CAD and path planning in robotics. respectively. speedup techniques for shape evolution have been proposed which use only points close to the curve at every time step [8. and skeletonization. By adapting the stencil for derivative approximations to the local smoothness of the solution. 37 To address the ﬁrst problem. (1. 1. possible generalizations. Blurring of discontinuities can be minimized by applying shockcapturing techniques such as highorder ENO15 schemes [395. When we want to smooth ENO means essentially nonoscillatory. Overviews and suitable references can be found in [233. and – although they cannot be written in conservation form – they reveal interesting relations to diﬀusion processes. ENO schemes obtain both highorder accuracy in smooth regions and sharp shock transitions [183].
5 (c) and 5. They arise in ﬂame propagation. Equation (1.11 (a) present examples for mean curvature ﬁltering. see [372.38 CHAPTER 1. Figure 5.96) where Cρ := ∂ρ C.94) ∇u Since curv(u) = div ∇u is the curvature of u (mean curvature for dimensions ≥ 3).e. It is invariant under Euclidean transformations.95) is also called geometric heat equation or Euclidean shortening ﬂow.90) propagates isophotes in inner normal direction with a velocity that is given by their curvature κ = det(Cp . 0) = f (x).e. p v(p. grid generation. we can neglect the ﬁrst term and end up with the problem ∂t u = ∂ξξ u. equation (1.93) (1.90) can also be written as ∂t u = u2 2 ux1 x1 − 2ux1 ux2 ux1x2 + u2 1 ux2 x2 x x u2 1 + u2 2 x x 1 = ∆u − ∇u. Let v(p. t) denote the Euclidean arclength of C(p. 374] and the references therein for an overview. The importance of MCM in image processing became only recently clear: As nicely explained in a paper by Guichard and Morel [173].90) (1. u(x. Intrinsic heat ﬂow Interestingly. crystal growth. i. there exists a further connection between linear diﬀusion and motion by curvature. (1. t) dρ.97) . The subsequent discussions shall clarify these names. PARTIAL DIFFERENTIAL EQUATIONS the image anisotropically along its isophotes. the derivation of minimal surfaces. The corresponding curve evolution ∂t C(p. (1. mean curvature motion can be regarded as the limit process when classic morphological operators such as median ﬁltering are iteratively applied.94) is named (mean) curvature motion (MCM). t) := 0 Cρ (ρ. and many other applications.Cpp ) . Cp 3 Processes of this type have ﬁrst been studied by Brakke in 1978 [54]. (1. The Euclidean arclength is characterized by Cv  = 1. t). t) · n(p.91) By straightforward calculations one veriﬁes that (1.95) shows that (1. Hess(u)∇u ∇u2 = ∇u curv(u). mappings x → Rx + b (1. i.92) (1. t) (1. t) = κ(p.
a nonconvex one becomes convex and.90) has a unique viscosity solution which is L∞ stable and satisﬁes a maximum–minimum principle [12].5. 227].g. equation (1. t). since it is independent of the curve parametrization. Moreover. a point with a circle as limiting shape. if one isophote is enclosed by another.5. the number of extrema and inﬂection points of the curvature is nonincreasing. equation (1. and smoothing properties follow from the fact that the total curvature decreases. They establish the semigroup property as architectural quality.100) that it takes the time 1 T = 2 σ 2 to remove all isophotes within a circle of radius σ. t). Lions and Morel [12]. Theoretical results For the evolution of a smooth curve under its curvature. t) = ∂vv C(p. This shows that the . In analogy to the dilation/erosion case it can be shown that. The time for shrinking a circle of radius σ to a point is given by 1 (1. Moreover. i. it can be shown that.98) we recognize that curvature motion can be regarded as Euclidean invariant diﬀusion of isophotes: ∂t C(p. t) · n(p. since under all ﬂows Ct = Cqq the Euclidean arclength parametrization q(p) := v(p) is the fastest way to shrink the Euclidean perimeter Cp  dp. However. As an image evolution. 14) that κ(p. When studying the evolution of isophotes under MCM. It is based on results of Evans and Spruck [129]. for an initial image f ∈ BUC(IR2 ). the curve shrinks to a circular point. p.100) T = 2 σ2 .1. [71]. it has been shown in [188. 169] that a smooth solution exists for some ﬁnite time interval [0. A convex curve remains convex. since the arclength v is again a function of the curve. MCM belongs to the class of morphological scalespaces which satisfy the general axioms of Alvarez. (1. that have been mentioned in 1. Scalespace interpretation A shape scalespace interpretation for curve evolution under Euclidean heat ﬂow is studied by Kimia and Siddiqi [227]. 151. Guichard. (1.99) This geometric heat equation is intrinsic. Such a shape inclusion principle implies in connection with (1. Since it is wellknown from diﬀerential geometry (see e. this ordering is preserved [129.99) is called Euclidean shortening ﬂow [150]. t) = ∂vv C(p. T ). for t → T . the reader should be aware of the fact that – although this equation looks like a linear onedimensional heat equation – it is in fact nonlinear.6 CURVATUREBASED MORPHOLOGICAL PROCESSES 39 where R ∈ IR2×2 denotes a rotation matrix and b ∈ IR2 is a translation vector.e.
Aﬃne transformations arise as shape distortions of planar objects when being observed from a large distance under diﬀerent angles. and it can be calculated as p s(p. t) = κ(p. Cρρ (ρ. Css ) = 1.101) where b ∈ IR2 denotes a translation vector and the matrix A ∈ IR2×2 is invertible.104) (1. Such an aﬃne arclength was proposed by Blaschke [44. t) := 0 det Cρ (ρ. 0) = C0 (p). 363] constructed an aﬃne invariant ﬂow by replacing the Euclidean arclength v(p. Hence.40 CHAPTER 1. Sapiro and Tannenbaum [362. This property is characteristic for all scalespaces of the Alvarez–Guichard–Lions–Morel axiomatic and distinguishes them from nonlinear diﬀusion ﬁlters. A (full) aﬃne transformation is a mapping x → Ax + b (1. · n(p.102) By virtue of ∂ss C(p. contrast cannot be enhanced.6. Moreover. (1. t). t) 1 3 dρ. PARTIAL DIFFERENTIAL EQUATIONS relation between temporal and spatial scale for MCM is the same as for linear diﬀusion ﬁltering (cf. 1. two level sets cannot move closer to one another than they were initially [129. t) (1. t) in (1. t) we obtain the aﬃne invariant heat ﬂow ∂t C(p. t) that is invariant with respect to aﬃne transformations with det(A) = 1. (1. there exist certain problems which also require invariance with respect to aﬃne transformations. It is characterized by det(Cs . 12–15] in 1923.99) by an “arclength” s(p. Aﬃne invariant intrinsic diﬀusion In analogy to the Euclidean invariant heat ﬂow. pp.2 Aﬃne invariant ﬁltering Motivation Although Euclidean invariant smoothing methods are suﬃcient in many applications. (1.105) . t).14)). t) = κ(p. t) 1 3 1 3 · n(p.103) C(p. 227].
with three exceptions [363]: (a) Closed curves shrink to points with an ellipse as limiting shape (elliptical points). t) as a levelline of an image u(x. Theoretical results The curve evolution properties of aﬃne invariant heat ﬂow can be shown to be the same as in the Euclidean invariant case. under all ﬂows Ct = Cqq .6 CURVATUREBASED MORPHOLOGICAL PROCESSES Aﬃne invariant image evolution 41 When regarding the curve C(p. . this equation is also called aﬃne shortening ﬂow.106) (1. 5.109) (c) The time for shrinking a circle of radius σ to a point is T = 3 4 σ3. After having mentioned some theoretical results. 2 1 1 3 1 3 (1. we end up with the evolution equation ∂t u = ∇u curv(u) = 3 = ∇u 3 uξξ . Guichard. (b) The name aﬃne shortening ﬂow reﬂects the fact that. t).110) For the image evolution equation (1. 4 (1. Cpp (p. Besides the name aﬃne invariant heat ﬂow. Lions and Morel [12] via an axiomatic scalespace approach. we shall brieﬂy sketch this reasoning below. see Fig. t) 1 3 dp. the aﬃne arclength parametrization q(p) := s(p) is the fastest way to shrink the aﬃne perimeter L(t) := det Cp (p.1. aﬃne morphological scalespace (AMSS). and fundamental equation in image processing. This image evolution equation has been discovered independently of and simultaneously with the curve evolution approach of Sapiro and Tannenbaum by Alvarez.6(a).106) we have the same results as for MCM and dilation/erosion concerning wellposedness of a viscosity solution which satisﬁes a maximum–minimum principle [12]. t). The temporal evolution of an image under such an evolution resembles mean curvature motion.108) u2 2 ux1 x1 − 2ux1 ux2 ux1 x2 + u2 1 ux2 x2 x x where ξ is the direction perpendicular to ∇u.107) (1. (1.
366]. 113].6. . Simpliﬁcations of this axiomatic and related axioms for shape scalespaces can be found in [16]. such that Tt (Af ) = A(Tt′ f ) ∀ f ∈ BUC(IR2 ). The scalespace reasoning of Sapiro and Tannenbaum investigates properties of the curve evolution. 133]. one requires a multiscale analysis which is invariant under a general projective mapping (x1 . can be regarded as steepest descent methods of energy functionals which have been proposed by Polyakov in the context of string theory [235].111) For this reason they call the AMSS equation also fundamental equation in image analysis. A study of heat ﬂows which are invariant under subgroups of the projective group can be found in [314. Research in this direction has been carried out by Faugeras and Keriven [130. there exists a rescaled time t′ (t.112) with A = (aij ) ∈ IR3×3 and det A = 1. Olver et al.3 Generalizations In order to analyse planar shapes in a way that does not depend on their location in IR3 . 113]. It is invariant to the bending (isometric mapping) of the surface.42 CHAPTER 1. This geodesic curvature ﬂow and other evolutions. Guichard. 1. (1. 131. x2 )⊤ → a11 x1 + a21 x2 + a31 a12 x1 + a22 x2 + a32 . and that there is some evidence that they do not reveal the same smoothing scalespace properties [314]. PARTIAL DIFFERENTIAL EQUATIONS Scalespace properties Alvarez.106) is unique (up to temporal rescalings) when imposing on the scalespace axioms for (1. Based on results of [229. Moreover. see [362] and the references therein. A) ≥ 0. [314]. 24] they point out that the Euclidean absolute curvature decreases as well as the number of extrema and inﬂection points of curvature.82) an additional aﬃne invariance axiom: For every invertible A ∈ IR2×2 and for all t ≥ 0. 352. Lions and Morel [12] proved that (1. 150. Euclidean and aﬃne invariant curve evolutions can also be modiﬁed in order to obtain area. It turns out that intrinsic heatequationlike formulations for the projective group are more complicated than the Euclidean and aﬃne invariant ones. An intrinsic heat ﬂow for images painted on surfaces has been investigated by Kimmel [232].or lengthpreserving equations [188. both for scalar and vectorvalued images. and Dibos [112. Bruckstein and Shaked [62]. a shape inclusion principle holds. a13 x1 + a23 x2 + a33 a13 x1 + a23 x2 + a33 ⊤ (1.
[305. the question arises of how to combine them to a process which simpliﬁes arbitrary surfaces without creating singularities. Curve evolution schemes are investigated by Mokhtarian and Mackworth [291]. On the one hand. 95. Such a ﬁltering intends to simplify the image while preserving its “semantical atoms” in the sense of Kanizsa [219]. Bruckstein et al. settheoretic morphological schemes and approximation schemes for the Eulerian formulation. 288]. is established. 362]. 15. [95]. 267. 78. on the other hand.6. 304] approximate . convergence to polygones. Generalizations of MCM or AMSS to 3D images are investigated in [91. 298. This is essentially the same tradeoﬀ as for circular structuring elements in classical settheoretic morphology. 1. it is diﬃcult to ﬁnd consistent approximations on a pixel grid which have good rotational invariance.5. [285]. Guichard.1. Niessen et al. 374]. [80. This problem is reminescent of the creation of new extrema in diﬀusion scalespaces when going from 1D to 2D. Cohignac et al. Merriman et al. higher stability or less dissipative eﬀects [10. we have three main classes of numerical methods: curve evolution schemes. 1. 316]. In complete analogy to the behaviour of the continuous equations. 287. 372. Curve evolution schemes can reveal perfect aﬃne invariance. Related discrete curve evolutions are analysed in [135.6 CURVATUREBASED MORPHOLOGICAL PROCESSES 43 Recently it has been suggested to modify AMSS such that any evolution at Tor Xjunctions of isophotes is inhibited [75]. A comparative evaluation of these approaches has been carried out by Lucido et al. Diﬀerent variants of these schemes have been proposed in order to get better rotational invariance. In contrast to the 2D situation. In [61] discrete analogues of MCM and AMSS for the evolution of planar polygons are introduced. 359]. [267]. this leads to an explicit ﬁnite diﬀerence method which approximates the spatial derivatives by central diﬀerences. Since they can have diﬀerent sign.4 Numerical aspects Due to the equivalence between curve evolution and morphological image processing PDEs. In the case of MCM or AMSS. [61]. whose corners belong to circles and ellipses.7. Ruuth [347]. Lions and Morel to colour images are discussed in [82]. Convergent settheoretic morphological schemes for MCM and AMSS have been proposed by Catt´ et al. A comparison of diﬀerent methods of these classes can be found in [95]. Aﬃne scalespace axiomatics for image sequences (movies) have been established in [12. topological changes such as the splitting of conconvex structures may occur. respectively. they are very fast and they e are entirely invariant under monotone greyscale transformations. 79]. and Moisan [289]. In this case two principal curvatures occur. Most direct approximations of morphological image evolution equations are based on the Osher–Sethian schemes [319. cf. and possibilities to generalize the axiomatic of Alvarez.
ceremade. Comparing it with the explicit scheme they report a tradeoﬀ between the larger time step size and the higher computational eﬀort per step.2.16 MegaWave2 has been developed by Jacques Froment and Sylvain Parrino. 218].dauphine. since aﬃne invariance implies that circles are equivalent to ellipses of any arbitrary eccentricity. 1. In order to solve the resulting linear system of equations he applies symmetric Gauß–Seidel iterations. A software package which contains implementations of MCM. A consistent semiimplicit approximation of MCM which discretizes the ﬁrstorder spatial derivatives explicitly and the secondorder derivatives implicitly has been proposed by Alvarez [10]. however. If one wants to have a good numerical approximation of aﬃne invariance. and assembles the ﬁnal image by superimposing all evolved level sets [75]. 146]. Nicolet and Sp¨ hler [301] investigate a consistent fully implicit scheme for MCM u leading to a nonlinear system of equations. renounce consistency with the original equation as well as rotational invariance: round shapes evolve into polygonal structures. but their calculation in the Fourier domain is computationally expensive and aliasing may lead to oscillatory solutions. One way to achieve unconditional L∞ stability for MCM is to approximate uξξ by suitable linear combinations of onedimensional secondorder derivatives along grid directions and to apply an implicit ﬁnite diﬀerence scheme [95.6) is available under the name MegaWave2. Using Gaussian derivatives for MCM or AMSS permits larger time steps for large kernels [304]. More information can be found under http://www. CEREMADE.44 CHAPTER 1. France. For AMSS an additional constraint appears: the behaviour of this equation is highly nonlocal. AMSS and many other modern techniques such as wavelets. map them into Lipschitzcontinuous images by applying a distance transformation. PARTIAL DIFFERENTIAL EQUATIONS the spatial derivatives by Gaussian derivatives which are calculated in the Fourier domain. 1.fr. 75775 Paris Cedex 16. Afterwards one extracts the processed level sets as the zerolevel sets of the evolved images. As a remedy. cf. and run a ﬁnite diﬀerence method on them. and active contour models (cf. Schemes of this type.3. The natural price one has to pay for the excellent results is a fairly high computational eﬀort. Concerning stability one observes that all these explicit schemes can violate a discrete maximum–minimum principle and require small time steps to be experimentally stable. MumfordShah segmentation.6. 16 . An inherent problem of all ﬁnite diﬀerence schemes for morphological image evolutions are their dissipative eﬀects which create additional blurring of discontinuities. one has to decrease the temporal step size signiﬁcantly below the step size of experimental stability [16. It is solved by means of nonlinear Gauß– Seidel iterations. University Paris IX. one can decompose the image into binary level sets. 13.
260].1) or global PDEs for histogramme enhancement [358]. This leads to a process which is useful for analysing components of shape [228. proposed a method for caricaturelike shape exaggeration [394].6 CURVATUREBASED MORPHOLOGICAL PROCESSES 45 1.7. already in 1965. This socalled min–max ﬂow produces a restored image as steadystate solution and reveals good denoising properties. Wellposedness results are not known up to now. In order to improve images.g. Adapting them to tasks such as texture enhancement requires more sophisticated and more global feature descriptors than the gradient. Malladi and Sethian [272] propose to replace MCM by a technique in which the motion of level curves is based on either min(κ. 234]. cf. It is interesting to note that. Recently Steiner et al. see e.1. 0) or max(κ. This longtime forgotten method is similar to the Perona–Malik process (1. 16]. since the theory of viscosity solutions is no longer applicable. and which is called entropy scalespace or reaction–diﬀusion scalespace. 132]. [126. 453]. Introducing a reaction term as in [102] allows to attract the solution to speciﬁed grey values which can be used as quantization levels [11]. 1. This corresponds to nonlinear diﬀusion ﬁlters and a restoration method by total variation minimization [345] which shall be described in 1. 304]. for instance MCM and dilation/erosion. The potential of MCM for shape segmentation [290] has even been used for classifying chrysanthemum leaves [1]. 364.2.6. 365. All these preceding modiﬁcations are at the expense of renouncing the morphological invariance of the genuine operators (and also aﬃne invariance in the case of [364.56). 127]. 97. 230.36) for large image gradients. 304]. If one wants to stay within the morphological framework one can combine diﬀerent morphological processes. 301] in medical images and to terrain matching [268]. 304] and blob detection [157. 365. They evolved the boundary curve by means of a backward Euclidean shortening ﬂow with a stabilizing bias term as in (1. shock ﬁltering ([14]. 384. 96.94). and for texture discrimination [265. If one aims to use these equations for image restoration one usually modiﬁes them by multiplying them with a term that reduces smoothing at highcontrast edges [13.6 (b). see also Fig. MCM and AMSS have also been applied to denoising [305. 383. depending on the average grey value within a certain neighbourhood. . 5. 0). for instance an analysis by means of Gabor functions [72. MCM or AMSS have also been combined with other processes such as linear diﬀusion [13]. for corner detection [15].7. 315] is used). unless an “aﬃne invariant gradient” [231. Gabor – the inventor of optical holography and the socalled Gabor functions – proposed a deblurring algorithm based on combining MCM with backward smoothing along ﬂowlines [149.5 Applications The special invariances of AMSS are useful for shape recognition tasks [12. Another modiﬁcation results from omitting the factor ∇u in the mean curvature motion (1.
Especially for assisting the segmentation of medical images they have become very popular [282]: The expert user gives a good initial guess of an interesting contour (organ. and an external energy term attracting the result to salient image features. which is attracted by image features such as edges.).. We observe that terms 1 and 2 describe the internal energy. 452]. (1.113) The ﬁrst summand is a membrane term causing the curve to shrink. Minimizing the functional (1. the second one is a rigidity term which encourages a straight contour. while 3D active models are sometimes named active surfaces or active blobs. x2 (s))⊤ which minimizes the functional E(C(s)) = C(s) α 2 Cs (s)2 + β 2 Css (s)2 − γ ∇f (C(s))2 ds. The main . snake models incorporate many ideas ranging from energy minimization over curve evolution to the Perona–Malik ﬁlter and diﬀusion–reaction models. A 3D version of such an active contour model is presented in [401]. the energy functional consists of two parts: an internal energy fraction which controls the smoothness of the result. the result will depend on the choice of the initial curve and a good segmentation requires an initial curve which is close to the ﬁnal segment. and the third term pushes the contour to high gradients of the image f . Since this model makes direct use of the snake contour. while the third one represents the external (image) energy. Witkin and Terzopoulos proposed the ﬁrst active contour model in a journal paper in 1988 [221]. PARTIAL DIFFERENTIAL EQUATIONS 1. . which will be carried the rest of the way by some energy minimization.113) by steepest descent gives ∂C = αCss − βCssss − γ∇(∇f 2 ). tumour..114) which can be approximated by ﬁnite diﬀerences. Such a snake is represented by a curve C(s) = (x1 (s).46 CHAPTER 1.6 Active contour models One of the main applications of curve evolution ideas appears in the context of active contour models (deformable models). it is also called an explicit model.6. Active contour models can be used to search image features in an interactive way. Explicit snakes Kass. The nonnegative parameters α. It is therefore not surprising that they play an important role in several generalization and uniﬁcation attempts [356. For this reason. β and γ serve as weights of these expressions. 2D versions are also called snakes. Usually. Apart from these practical merits. Their snakes can be thought of as an energyminimizing spline. 380. ∂t (1.
could prove wellposedness in the viscosity sense [73]. This allows simultaneous segmentation of multiple objects. for instance by using the distance transformation. a compensation for the property of MCM to create convex regions. It was discovered by Caselles.6 CURVATUREBASED MORPHOLOGICAL PROCESSES 47 disadvantage of the preceding model is its topological rigidity: a classical explicit snake cannot split in order to segment several objects simultaneously17 .32): it becomes small for large ∇fσ  = ∇Kσ ∗f .94). • ν∇u describes motion in normal direction. In practice. and the ﬁnal contour C is extracted as the zero level curve of u(x. 435]. Hence.115) have the following meaning: • ∇u div (∇u/∇u) is the curvature term of MCM which smoothes level sets. Moreover. they use essentially only one remaining parameter. An advantage of implicit snake models is their topological ﬂexibility: The contour may split. e Coll and Dibos [73]. and Caselles and Coll investigated related approaches for image sequences [74]. Catt´. On Recently McInerley and Terzopoulos have proposed modiﬁed explicit deformable models which allow topological changes [281. ∇u (1. The terms in (1. dilation or erosion depending on the sign of ν. T ). i. The idea behind implicit snakes is to embed the initial curve C0 (s) as a zero level curve into a function u0 : IR2 → IR.1. it is also sometimes not easy to ﬁnd a good balance between the parameters α. Whitaker and Chen developed similar implicit active contour models for 3D images [436. when the process does hardly alter anymore. Sethian and Vemuri [273]. it slows down the snake as soon as it approaches signiﬁcant edges of the original image f . For this model Caselles et al.e. • g is a stopping function such as the Perona–Malik diﬀusivity (1. Implicit snakes In 1993 some of the inherent diﬃculties of explicit snakes have been solved by replacing them by a socalled implicit model. see (1. and later on by Malladi. Then u0 is evolved under a PDE which includes knowledge about the original image f : ∂t u = g(∇fσ 2 ) ∇u div ∇u +ν . 17 . This socalled balloon force [93] is required for pushing a level set into concave regions. 283].115) This evolution is stopped at some time T . the balloon force ν. β and γ.
Recently also an aﬃne invariant analogue to geodesic active contours has been proposed [315].48 CHAPTER 1. 226. but often a speed term νg(∇fσ 2 )∇u is added to achieve faster and more stable attraction to edges. (1. Kumar. Techniques which can be related to geodesic active contours have also been used for solving 3D vision problems such as stereo [134] and motion analysis [322]. PARTIAL DIFFERENTIAL EQUATIONS the other hand. Then the .117) This is nothing else than ﬁnding a curve of minimal distance (geodesic) with respect to some imageinduced metric.118) This active contour model is parameterfree. Olver. ∇u (1.32). These snakes replace the contour energy (1.113) by E(C) = C(s) α 2 Cs (s)2 − γ g(∇fσ (C)2 ) ds (1.116) is equivalent to searching min C C(s) g ∇fσ (C(s))2 Cs (s) ds. the process does not really stop at the desired result. They have been proposed simultaneously by Caselles. Tek and Kimia use implicit active contours starting from randomly chosen seed points. it only slows down. both in 2D [399] and in 3D [400]. and it is diﬃcult to interpret implicit snakes in terms of energy minimization. A theoretical analysis of geodesic snakes concerning existence. Tannenbaum and Yezzi [226]. 271]. Kimmel and Sapiro [76] and by Kichenassamy.116) where g denotes again a Perona–Malik diﬀusivity of type (1. Embedding the initial curve as a level set of some image u0 and applying a descent method to the corresponding EulerLagrange equation leads to the image evolution PDE ∂t u = ∇u div g(∇fσ 2 ) ∇u . Under some additional assumptions they derive that minimizing (1.118) by g(∇uσ 2 ). and extensions to 3D images are studied in [77. uniqueness and stability of a viscosity solution can be found in [76. They name this technique reaction–diﬀusion bubbles. Selfsnakes The properties of geodesic snakes induced Sapiro to use a related technique for image enhancement [357]: he replaced g(∇fσ 2 ) in (1. In order to address the initialization problem. Geodesic snakes Geodesic snakes make a synthesis of explicit and implicit snake ideas. 226].
one may ask if it is possible to apply related ideas to image processing. 1.6(c).18 Another image enhancement method that is close in spirit is due to Eidelman. [380]. see also [414.121) Thus. cf. This would be useful to restore discontinuities such as edges. For σ = 0 this gives ∂t u = ∇u div g(∇u2) ∇u ∇u (1.1: the latter can be written as ∂t u = g(∇u2) ∆u + ∇⊤ g(∇u2) ∇u.123) plays an important role for shock calculations. Their model contains a reaction term which inhibits smoothing at edges and keeps the ﬁltered image u close to the original image f . 422] for related ideas. it cannot be excluded that a selfsnake without spatial regularization reveals the same illposedness problems as the Perona–Malik ﬁlter [356].122) The temporal evolution of a regularized selfsnake without reaction term is depicted in Fig. Generalizations of selfsnakes to vectorvalued images [357. however.120) = g(∇u2) uξξ + ∇⊤ g(∇u2) ∇u with ξ ⊥ ∇u. Chen.1. For σ > 0. 361] can be obtained using Di Zenzo’s ﬁrst fundamental form for colour images [114]. we observe striking similarities with the Perona–Malik process from Section 1. Grossmann and Friedman [125].119) (1. 18 . The restored image is given by the steadystate of ∂t u = ∇u div g(∇uσ 2 ) ∇u + β ∇u (f −u) ∇u (β > 0).7 TOTAL VARIATION METHODS 49 snake becomes a selfsnake no longer underlying external image forces. 5. (1. Vemuri and Wong [89] established existence and stability of a unique viscosity solution to a modiﬁed selfsnake process. (1. Although this equation cannot be cast in divergence form.7 Total variation methods T V (u) := Ω Inspired by observations from ﬂuid dynamics where the total variation (TV) ∇u dx (1. Below we shall focus on two important TVbased image restoration techniques which have been pioneered by Osher and Rudin: TVpreserving methods and techniques which are TVminimizing subject to certain constraints.3. It maps the image grey values to gas dynamical parameters and solves the compressible Euler equations using shockcapturing total variation diminishing (TVD) techniques based on Godunov’s method.
Here.124) with F (u) := sgn(u) arises as the PDE formulation of a classical image enhancement algorithm by Kramer and Bruckner [243]. sgn(F (u)) = sgn(u). TVpreserving methods suﬀer from the problem that ﬂuctuations due to noise do also create shocks. Alvarez and Mazorra [14] replace the operator L(u) = uηη in (1. It has been shown that a onedimensional version of this ﬁlter preserves the total variation and satisﬁes a maximum–minimum principle. They prove that their semiimplicit ﬁnitediﬀerence scheme has a unique solution which satisﬁes a maximum–minimum principle.50 CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS 1. It solves the linear diﬀusion equation backwards in time under the regularizing constraint that the total variation remains constant. the Laplacian L(u) = ∆u or the secondorder directional derivative L(u) = uηη with η ∇u.7. By means of our knowledge from morphological processes. Kramer and Bruckner proved in 1975 that their Ndimensional discrete scheme converges after a ﬁnite number of iterations to a state where each point is a local extremum. Recently van den Boomgaard [52] pointed out that the 1D version of (1. 0) = f (x).1 TVpreserving methods In 1990 Osher and Rudin have proposed to restore blurred images by shock ﬁltering [317].124) by a Gaussiansmoothed version L(Kσ ∗u) = Kσ ∗uηη and supplement the resulting equation with a noiseeliminating mean curvature process. (1. This stabilization can be realized by keeping local extrema ﬁxed during the whole evolution. and L(u) is a secondorder elliptic operator whose zerocrossings correspond to edges. For this reason. Osher and Rudin have also proposed another TVpreserving deblurring technique [318]. both in the continuous and discrete case. Thus. blocky . For the twodimensional case not many theoretical results are available except for a discrete maximum–minimum principle. we recognize that this ﬁlter aims to produce a ﬂow ﬁeld that is directed from the interior of a region towards its edges where it develops shocks. For this reason. From a practical point of view.2 TVminimizing methods Total variation is good for quantifying the simplicity of an image since it measures oscillations without unduly punishing discontinuities. These ﬁlters calculate the restored image as the steadystate solution of the problem ∂t u = −∇u F (L(u)).125) u(x.g.124) (1. 1. the goal is to obtain a piecewise constant steadystate solution with discontinuities only at the edges of the initial image.7. e.
but – in contrast to TV approaches – equation (1.56). . Given an image f with additive noise of zero mean and known variance σ 2 . we seek a restoration u satisfying min u Ω ∇u dx (1. It uses an explicit ﬁnite diﬀerence scheme with central and onesided spatial diﬀerences and adapts the Lagrange multiplier by means of the gradient projection method of Rosen. Ω (1. let us study an example. In order to solve this constrained variational problem. Osher and Fatemi [345] have proposed to minimize the total variation under constraints which reﬂect assumptions about noise. 19 Related ideas have also been developed by Geman and Reynolds [153]. In [345] a gradient descent method is proposed to solve (1. the divergence term in (1. Rudin.129).129) looks similar to the steadystate equation of the diﬀusion–reaction equation (1. which relates TVminimizing techniques to MCM.127) (1.56) is not intended to satisfy the noise constraint exactly [346].126) subject to (u−f )2 dx = σ 2 .129) with homogeneous Neumann boundary conditions.129) is identical with the curvature.1. In order to restore noisy blocky images.19 To ﬁx ideas. The (unknown) Lagrange multipliers µ and λ have to be determined in such a way that the constraints are fulﬁlled. One may also reformulate the constrained TV minimization as an unconstrained problem [83]: The penalized least square problem min u 1 u−f 2 2 L2 (Ω) +α Ω ∇u dx (1.130) is equivalent to the constrained TV minimization.128) u dx = Ω Ω f dx. Interestingly. (1.128) veriﬁes necessarily the Euler equation div ∇u − µ − λ(u−f ) = 0 ∇u (1. PDE methods can be applied: A solution of (1. if α is related to the Lagrange 1 multiplier λ via α = λ .126)–(1. Moreover.7 TOTAL VARIATION METHODS 51 images (consisting only of a few almost piecewise constant segments) reveal very small total variation.
and the idea of L1 norm minimization has also led to improved optic ﬂow algorithms [245]. 409]. This leads to TVbased blind deconvolution algorithms [86]. As a remedy. 204. where the generalized TV norm is chosen as the l2 norm of the TV norms of the separate channels. 253. (1. 115. they establish relations between TVminimizing methods and nonlinear diﬀusion techniques [397]. Another possibility is to consider the constrained minimization of B(u) := Ω ∇u p(∇u) dx. To overcome the problem that the total variation integral contains the nondiﬀerentiable argument ∇u. 346. as ∇u ranges from 0 to ∞. [262] and Dobson and Santosa [115] have shown the existence of BV(Ω)solutions for problems of this type. 205. 253. optimization. 83]. and they have established uniqueness. 118. Total variation methods have been applied to restoring images of military relevance [345. 410. Chambolle and Lions [83] have extended the existence proof to noncompact operators (which comprises also the situation without blur). see [46]. it has been proposed to minimize the L1 norm of expressions containing also higherorder derivatives [344. 396]. one applies regularization strategies or techniques from nonsmooth optimization. 83]. or other types of noise [262. 262. 83]. TVminimizing methods have been generalized in diﬀerent ways: • The constrained TVminimization idea is frequently adapted to other constraints such as blur. • They have been applied to colour images [45]. • The tendency of TVminimizing to create piecewise constant structures can cause undesired eﬀects such as the creation of staircases at sigmoidlike edges [116. . Much research is done in order to ﬁnd eﬃcient numerical methods for which convergence can be established. 84. They are useful for enhancing reconstruction algorithms for inverse scattering problems [37]. to improving material from criminal and civil investigations as court evidence [344]. Recently.130) have attracted much interest from mathematicians working on inverse problems.52 CHAPTER 1. 87. 116. or numerical analysis [2. 85. 346. • Strong and Chan have identiﬁed the parameter α in (1. 117. problems of type (1. 344.131) where p(∇u) decreases from 2 to 1. Lions et al. • TVminimizing methods have also been used for estimating discontinuous blurring kernels such as motion or outoffocus blur from a degraded image. 253]. noise with blur.130) as a scale parameter [396]. By adapting α to the local image structure. and to enhancing pictures from confocal microscopy [409] and tomography [115. PARTIAL DIFFERENTIAL EQUATIONS In recent years.
In particular. we shall see that anisotropic nonlinear diﬀusion processes can share many advantages of the scalespace and the image enhancement world. The general framework. On the other hand. we have also become aware of some limitations. for some applications. for which the results hold. Pure restoration methods such as diﬀusion–reaction equations or TVbased techniques do allow contrast enhancement and lead to stable structures but can suﬀer from theoretical or practical problems. minimization algorithms can get trapped in a poor local minimum. Moreover. A scalespace interpretation is presented which does not exclude contrast enhancement. Both scalespace and wellposedness properties carry over from the continuous to the semidiscrete and discrete setting. and wellposedness results are established. The goal of the subsequent chapters is to develop a theory for nonlinear anisotropic diﬀusion ﬁlters which addresses some of the abovementioned shortcomings. or the use of explicit schemes causes restrictive step size limitations. Linear diﬀusion and morphological scalespaces are wellposed and have a solid axiomatic foundation. They shall serve as a motivation for the theory which will be explored in the subsequent chapters. Other interesting image restoration topics have found less attention. speciﬁc anisotropic models are presented which permit applications beyond segmentation and edge enhancement tasks. .1. most imageenhancing PDE methods focus on edge detection and segmentation problems. they possess the undesirable property that they do not permit contrast enhancement and that they may blur and delocalize structures. for instance enhancement of coherent ﬂowlike structures in textures.7 TOTAL VARIATION METHODS 53 1. The latter comprises for instance semiimplicit techniques for which unconditional stability in the L∞ norm is proved. For both scalespace and restoration methods many questions concerning their discrete realizations are still open: discrete scalespace results are frequently missing.8 Conclusions and further scope of the book Although we have seen that there exists a large variety of PDEbased scalespace and image restoration methods which oﬀer many advantages. for instance unsolved wellposedness questions or the search for eﬃcient minimizers of nonconvex or nondiﬀerentiable functionals. Finally. includes also linear and isotropic nonlinear diﬀusion ﬁlters.
54
CHAPTER 1. PARTIAL DIFFERENTIAL EQUATIONS
Chapter 2 Continuous diﬀusion ﬁltering
This chapter presents a general continuous model for anisotropic diﬀusion ﬁlters, analyses its theoretical properties and gives a scalespace interpretation. To this end, we adapt the diﬀusion process to the structure tensor, a wellknown tool for analysing local orientation. Under fairly weak assumptions on the class of ﬁlters, it is possible to establish wellposedness and regularity results and to prove a maximum–minimum principle. Since the proof does not require any monotony assumption it applies also to contrastenhancing diﬀusion processes. After sketching invariances of the resulting scalespace, we focus on analysing its smoothing properties. We shall see that, besides the extremum principle, a large class of associated Lyapunov functionals plays an important role in this context [414, 415].
2.1
Basic ﬁlter structure
Let us consider a rectangular image domain Ω := (0, a1 ) × (0, a2 ) with boundary Γ := ∂Ω and let an image be represented by a mapping f ∈ L∞ (Ω). The class of anisotropic diﬀusion ﬁlters we are concerned with is represented by the initial boundary value problem ∂t u = div (D ∇u) D∇u, n = 0 on on on u(x, 0) = f (x) Ω × (0, ∞), Ω, Γ × (0, ∞). (2.1) (2.2) (2.3)
Hereby, n denotes the outer normal and ., . the Euclidean scalar product on IR2 . In order to adapt the diﬀusion tensor D ∈ IR2×2 to the local image structure, one would usually let it depend on the edge estimator ∇uσ (cf. 1.3.3), where uσ (x, t) := (Kσ ∗ u(., t)) (x) ˜ (σ > 0) (2.4) and u denotes an extension of u from Ω to IR2 , which may be obtained by mirroring ˜ at Γ (cf. [81]). 55
56
CHAPTER 2. CONTINUOUS DIFFUSION FILTERING
However, we shall choose a more general structure descriptor which comprises the edge detector ∇uσ , but also allows to extract more information. This will be presented next.
2.2
The structure tensor
In order to identify features such as corners or to measure the local coherence of structures, we need methods which take into account how the orientation of the (smoothed) gradient changes within the vicinity of any investigated point. The structure tensor – also called interest operator, scatter matrix or (windowed second) moment tensor – is an important representative of this class. Matrices of this type are useful for many diﬀerent tasks, for instance for analysing ﬂowlike textures [340, 31], corners and Tjunctions [145, 182, 310, 309], shape cues [256, pp. 349–382] and spatio–temporal image sequences [209, pp. 147–153], [168, pp. 219–258]. Related approaches can also be found in [39, 40, 220]. Let us focus on some aspects which are of importance in our case. In order to study orientations instead of directions, we have to identify gradients which diﬀer only by their sign: they share the same orientation, but point in opposite directions. To this end, we reconsider the vectorvalued structure descriptor ∇uσ within a matrix framework. The matrix J0 resulting from the tensor product J0 (∇uσ ) := ∇uσ ⊗ ∇uσ := ∇uσ ∇uT (2.5) σ has an orthonormal basis of eigenvectors v1 , v2 with v1 ∇uσ and v2 ⊥ ∇uσ . The corresponding eigenvalues ∇uσ 2 and 0 give just the contrast (the squared gradient) in the eigendirections. Averaging this orientation information can be accomplished by convolving J0 (∇uσ ) componentwise with a Gaussian Kρ . This gives the structure tensor Jρ (∇uσ ) := Kρ ∗ (∇uσ ⊗ ∇uσ ) (ρ ≥ 0).
(2.6)
It is not hard to verify that the symmetric matrix Jρ = j11 j12 is positive semidefj12 j22 inite and possesses orthonormal eigenvectors v1 , v2 with v1
j22 − j11 + 1 j11 +j22 ± 2
The corresponding eigenvalues µ1 and µ2 are given by µ1,2 =
2j12 . 2 (j11 −j22 )2 + 4j12
2 (j11 −j22 )2 + 4j12 ,
(2.7)
(2.8)
where the + sign belongs to µ1 . As they integrate the variation of the grey values within a neighbourhood of size O(ρ), they describe the average contrast in the
10) and identify it with its dual space. the coherence direction. THEORETICAL RESULTS 57 eigendirections. Furthermore.3 Theoretical results In order to discuss wellposedness results. Let L2 (0. and v2 gives the preferred local orientation.9) becomes large for anisotropic structures. and the expression 2 (µ1 − µ2 )2 = (j11 −j22 )2 + 4j12 (2. We need the following prerequisites: . H1(Ω)) be the space of functions u. µ1 and µ2 can be used as descriptors of local structure: Constant areas are characterized by µ1 = µ2 = 0. T ].11) In a similar way.T ]. T ]) such that T 2 H1 (Ω) u L2 (0. (2. corners can be identiﬁed by µ1 ≥ µ2 ≫ 0. (2. C([0. Presmoothing in order to obtain ∇uσ makes the structure tensor insensitive to noise and irrelevant details of scales smaller than O(σ). Now we can give a precise formulation of the problem we are concerned with. let us ﬁrst recall some useful notations. we observe that v1 is the orientation with the highest grey value ﬂuctuations.T . strongly measurable on [0. we denote by Cp (X. Let H1 (Ω) be the Sobolev space of functions u(x) ∈ L2 (Ω) with all distributional derivatives of ﬁrst order being in L2 (Ω). L2 (Ω)) is deﬁned as the space of continuous functions u : [0. 2.L2 (Ω)) := max u(t) [0.10 (d).2. By virtue of µ1 ≥ µ2 ≥ 0. see Section 5. The parameter σ is called local scale or noise scale. T . Thus. We equip H1 (Ω) with the norm u := u 2 L2 (Ω) 2 1/2 H1 (Ω) + i=1 ∂xi u 2 L2 (Ω) (2.T ] L2 (Ω) . the integration scale ρ should reﬂect the characteristic window size over which the orientation is to be analysed.3. T ] → L2 (Ω) supplemented with the norm u C([0. Y ) the set of Cp mappings from X to Y .12) As usual. T ] with range in H1 (Ω) (for the Lebesgue measure dt on [0. An example which illustrates the advantages of the structure tensor for analysing coherent patterns can be found in Figure 5. straight edges give µ1 ≫ µ2 = 0.2. It is a measure of the local coherence.H1 (Ω)) := u(t) 0 dt 1/2 < ∞.
b := ess sup f . T ]. T ]). extremum principle) The problem (Pc ) has a unique solution u(x. L2(Ω) .14) ¯ Moreover. exists a positive lower bound ν(K) for the eigenvalues of D(Jρ (w)). T ].1 regularity. and σ. T ]. T ]. and it fulﬁls the extremum principle a ≤ u(x. (C3) Uniform positive deﬁniteness: 2 ∞ ¯ there For all w ∈ L (Ω.13) (2. 0) = f (x) on Ω. and consider the problem Ω Ω ∂t u = div (D(Jρ(∇uσ )) ∇u) on Ω × (0. IR2×2 ). This solution depends continuously on f with respect to . n = 0 on Γ × (0. u ∈ C∞ (Ω×(0. 2 1 (2. IR ) with w(x) ≤ K on Ω. D(Jρ (∇uσ ))∇u.15) For a complete wellposedness proof one also has to establish stability with respect to perturbations of the diﬀusion equation. CONTINUOUS DIFFUSION FILTERING Assume that f ∈ L∞ (Ω). T > 0. (C2) Symmetry: 2×2 d12 (J) = d21 (J) for all symmetric matrices J ∈ IR . ρ ≥ 0. This topic will not be addressed here. t) in the distributional sense which satisﬁes u ∈ C([0. u(x. Let a := ess inf f . where the diﬀusion tensor D = (dij ) satisﬁes the following properties: (C1) Smoothness: D ∈ C∞ (IR2×2 .58 CHAPTER 2. 1 . ∂t u ∈ L (0. (Pc ) Under these assumptions the following theorem. H1 (Ω)). which generalizes and extends results from [81. (2. can be proved. t) ≤ b on Ω × (0. L2 (Ω)) ∩ L2 (0. T . T . Theorem 1 (Wellposedness. 414]. H (Ω)).
Morel and e Coll [81].e. T . 211). with compact inclusion. convex and weakly compact subset W0 of into itself. Now. H1(Ω)) dt 2 2 W (0. we utilize Stampacchia’s truncation method (cf. ∞) for some constant C. T ) := w ∈ L2 (0. L (Ω)). T ] × Ω). uniqueness and regularity are straightforward anisotropic extensions of the proof for the isotropic case studied by Catt´. Therefore.2. u = U(u). T ]. p. Then one shows that U is a weakly continuous mapping from a nonempty. such that the Gronwall–Bellman inequality can be applied. T . By iterating. t)−b) ∂t u(x. H1(Ω)). 0] and 0 < G′ (s) ≤ C on (0. U reveals a ﬁxed point u ∈ W0 . one deduces that u(t) ∈ H2 (Ω) for all t > 0. t ∈ [0. uniqueness and regularity Existence. Then. T ]. ϕ(t) := Ω H(u(x. we have G(u(x. whose solution is shown to converge in C([0. We restrict ourselves to proving only the maximum principle. uniqueness follows from the fact that both solutions start with the same initial values. Finally an iterative linear scheme is investigated. The basic idea of the uniqueness proof consists of using energy estimates for the diﬀerence of two solutions. THEORETICAL RESULTS Proof: 59 (a) Existence. T . we just sketch the basic ideas of this proof. (b) Extremum principle In order to prove a maximum–minimum principle. Existence can be proved using Schauder’s ﬁxed point theorem. dw ∈ L2 (0. i. Smoothness follows from classical bootstrap arguments and the general theory of parabolic equations [246]. one can establish that u is a ¯ strong solution of (Pc ) and u ∈ C∞ ((0. t) − b) dx. The minimum principle follows from the maximum principle when being applied to the initial datum −f . By the Cauchy–Schwarz inequality. [58]. Lions. s ∈ IR. T ) is contained in L (0.3. One considers the solution U(w) of a distributional linear version of (Pc ) where D depends on some function w instead of u. Since W (0. Since u(t) ∈ H1 (Ω) for all t > 0. Let G ∈ C1 (IR) be a function with G(s) = 0 on (−∞. t) dx ≤ C · u(t)−b L2 (Ω) Ω · ∂t u(t) L2 (Ω) . L2 (Ω)) to the strong solution of (Pc ). we deﬁne s H(s) := 0 G(σ) dσ.
T ].14) we know that the righthand side of this estimate exists. Therefore. T ]. and (2. T ].13). (2. 2 (2. ϕ ≥ 0 on [0. L2 (Ω)). In the same way as in the uniqueness proof in [81]. h ∈ L∞ (Ω) be two initial values and u. T ]. CONTINUOUS DIFFUSION FILTERING and by virtue of (2. t) − b ≤ 0 almost everywhere (a. pp. T ].) on Ω. Hence.16). w the corresponding solutions. we obtain u(x. T ] ≤ c · ∇u(t) 2 L2 (Ω) · u(t)−w(t) 2 L2 (Ω) . n dS =0 = Γ − Ω G (u−b) ∇u. we ﬁnally end up with the assertion ¯ u(x. Applying the Gronwall–Bellman lemma [57. the righthand side of (2. 156–137] yields u(t)−w(t) 2 L2 (Ω) ≤ f −h 2 L2 (Ω) · exp c · t 0 ∇u(s) 2 L2 (Ω) ds .17) Since u ∈ C([0. Now from ϕ ∈ C[0. t) ≤ b on Ω × (0. it follows that ϕ ≡ 0 on [0. H(u(x. t)−f (x)) dx ≤ (2. By means of H(s) ≤ 0 ≤ ϕ(t) ≤ Ω C 2 s. one shows that there exists some constant c > 0 such that d dt u(t)−w(t) 2 L2 (Ω) ϕ(0) = 0.16) we have C u(t)−f 2 2 L2 (Ω) . and we get dϕ = dt = Ω G(u−b) ∂t u dx Ω G(u−b) div (D(Jρ (∇uσ )) ∇u) dx G(u−b) D(Jρ (∇uσ )) ∇u. for all t ∈ [0. Due to the smoothness of u for t > 0. ϕ is diﬀerentiable for t > 0. (c) Continuous dependence on the initial image Let f. .17) tends to 0 = ϕ(0) for t → 0+ which proves the continuity of ϕ(t) in 0.e. D(Jρ(∇uσ )) ∇u dx ≥0 ≥0 ′ ≤ 0.60 CHAPTER 2.
T . Now. < δ. such that t 0 T ∇u(s) ∇u(s) T 2 L2 (Ω) ds ≤ 2 L2 (Ω) ds 0 1 ≤ ν = 1 ν 1 ν 0 Ω T ∇u(x. 2 which proves the continuous dependence on the initial data. Thus. let ǫ > 0 and choose δ := ǫ · exp Then for f −h L2 (Ω) −c u 2ν L2 (0. D(Jρ(∇uσ (x. we know that the righthand of this estimate exists. s). .13). THEORETICAL RESULTS 61 ¯ By means of the extremum principle we know that u is bounded on Ω×[0.T .14).T .H1 (Ω)) . our proof does not require any monotony assumption.3. the preceding results imply L2 (Ω) u(t)−w(t) <ǫ ∀ t ∈ [0. T ]. s) dx ds L2 (Ω) 0 Ω T ≤ ≤ u(s) 0 ∂t u(s) ∂t u L2 (Ω) ds 1 u ν L2 (0. (2. More restrictive requirements – for instance f ∈ BUC(IR2 ) in order to apply the theory of viscosity solutions – are not necessary in our case. T ]. (b) Moreover. s)))∇u(x. s) dx ds u(x.2. Chapter 5 will illustrate this by presenting examples where contrast is enhanced. ∇uσ is also bounded.H1 (Ω)) L2 (0. Remarks: (a) We observe a strong smoothing eﬀect which is characteristic for many diffusion processes: under fairly weak assumptions on the initial image (f ∈ L∞ (Ω)) we obtain an inﬁnitely often diﬀerentiable solution for arbitrary small positive times. and prerequisite (C3) implies that there exists some constant ν = ν(σ.H1 (Ω)) ∂t u L2 (0.T .H1 (Ω)) . s) · div D(Jρ (∇uσ (x. By virtue of (2. s)))∇u(x. f L∞ (Ω) ) > 0. This has the advantage that contrastenhancing processes are permitted as well.
This does not contradict contrast enhancement: In the case of two similar images. Much of the terminology used below is borrowed from [12]. we will never obtain results with grey value such as 257. Afterwards. if we start for instance with an image within the range [0. informationreducing image transformation. 2.18) where u(t) := u(. (2. the regularization damps the enhancement process in such a way that both images do not diﬀer much after ﬁltering.1 Invariances Let u(x.2. we shall not focus on further investigations of architectural requirements like recursivity. spatiotemporal image sequences or slices from medical CT or MRI sequences.. namely the question in which sense our evolution equation – which may allow contrast enhancement – can still be considered as a smoothing. CONTINUOUS DIFFUSION FILTERING (c) The continuous dependence of the solution on the initial image has significant practical impact as it ensures stability with respect to perturbations of the original image. we turn to a more crucial task. 2 . regularity and locality. To this end.4 Scalespace properties Let us now investigate scalespace properties of the class (Pc ) and juxtapose the results to other scalespaces.4.2 (d) The extremum principle oﬀers the practical advantage that. It is also closely related to smoothing scalespace properties. 255]. t ≥ 0. (e) The wellposedness results are essentially based on the fact that the regularization by convolution with a Gaussian allows to estimate ∇uσ L∞ (Ω) by u L∞ (Ω) . t) be the unique solution of (Pc ) and deﬁne the scalespace operator Tt by Tt f := u(t). as these qualities do not distinguish nonlinear diﬀusion scalespaces from other ones. This property is responsible for the uniform positive deﬁniteness of the diﬀusion tensor. We start with brieﬂy discussing invariances. as we shall see in 2. The properties we discuss now illustrate that an invariance of Tt with respect to some image transformation P is characterized by the fact that Tt and P commute. This is of importance when considering stereo image pairs. where one leads to contrast enhancement and the other not.4. since we know that similar images remain similar after ﬁltering.62 CHAPTER 2. t). 2.
Proposition 1 (Conservation of average grey value). and the ﬁltered images will also be shifted by the same constant. (2. for instance proton densities in MR images. 408]. we have Tt (0) = 0. but not of u. Average grey level invariance Average grey level invariance is a further property in which diﬀusion scalespaces diﬀer from morphological scalespaces. A constant average grey level is essential for scalespace based segmentation algorithms such as the hyperstack [307. t) dx for all t ≥ 0. (2. The average grey level 1 f (x) dx µ := Ω Ω (2. Then the Cauchy–Schwarz inequality . SCALESPACE PROPERTIES Grey level shift invariance 63 Since the diﬀusion tensor is only a function of Jρ (∇uσ ).22) is not aﬀected by nonlinear diﬀusion ﬁltering: 1 Ω Proof: Deﬁne I(t) := Ω Tt f dx = µ Ω ∀ t > 0. (2. Therefore.23) u(x. Tt (f + C) = Tt (f ) + C ∀ t ≥ 0.4.20) Reverse contrast invariance From D(Jρ (−∇uσ )) = D(Jρ (∇uσ )). the role of dilation and erosion has to be exchanged as well.21) This property is not fulﬁlled by classical morphological scalespace equations like dilation and erosion. a constant function is not aﬀected by diﬀusion ﬁltering. we may shift the grey level range by an arbitrary constant C. Moreover. It is also a desirable quality for applications in medical imaging where grey values measure physical qualities of the depicted object. it follows that Tt (−f ) = −Tt (f ) ∀ t ≥ 0. the evolution PDEs of the latter ones are not of divergence form and do not preserve the mean grey value.2. When reversing the contrast.19) (2. In general.
27) . (2.64 implies CHAPTER 2.24) Ω Ω Then Proposition 1 and grey level shift invariance imply that the order of M and Tt can be exchanged: M(Tt f ) = Tt (Mf ) ∀ t ≥ 0. (2.25) When studying diﬀusion ﬁltering as a pure initial value problem in the domain IR . For t > 0. Average grey level invariance may be described by commuting operators. Then diﬀusion ﬁltering fulﬁls Tt (τh f ) = τh (Tt f ) ∀ t ≥ 0. Theorem 1. then the eigenvalues of the diﬀusion tensor are unaltered and any eigenvector v is transformed into Rv. t)−f (x) dx ≤ Ω1/2 u(t)−f L2 (Ω) . 2 Translation invariance Deﬁne a translation τh by (τh f )(x) := f (x + h). when introducing an averaging operator M : L1 (Ω) → L1 (Ω) which maps f to a constant image with the same mean grey level: 1 (Mf )(y) := f (x) dx ∀ y ∈ Ω. Isometry invariance Let R ∈ IR2×2 be an orthogonal transformation. the preceding inequality gives the continuity of I(t) in 0. L (Ω)). Thus. n dS = 0. I(t) must be constant for all t ≥ 0. the divergence theorem and the boundary conditions yield dI = dt ∂t u dx = Ω Γ D(Jρ (∇uσ ))∇u. it makes no diﬀerence whether the orthogonal transformation is applied before or after diﬀusion ﬁltering: Tt (Rf ) = R(Tt f ) ∀ t ≥ 0. This leads us to translation and isometry invariance. (2.26) This is a consequence of the fact that the diﬀusion tensor depends on Jρ (∇uσ ) solely. but not explicitly on x. Since u ∈ C([0. CONTINUOUS DIFFUSION FILTERING I(t)−I(0) = 2 Ω u(x. 2 Hence. it also makes sense to investigate Euclidean transformations of an image. If we apply R to f by deﬁning Rf (x) := f (Rx). (2. T ].
η2 of the Hessian Hess(u) are positive.28) A suﬃcient condition for the causality equation (2. if ξ is a local minimum. nonenhancement of local extrema generally does not imply that their number is nonincreasing.4. the ﬁrst term of the righthand side of (2. θ) < 0. 1. isointensity linking through the scales is possible and a structure at a coarse scale can (in principle) be traced back to the original image (causality). For this reason. and ∂t u < 0 if ξ is a maximum.28) to hold is requiring that local extrema with positive or negative deﬁnite Hessians are not enhanced: an extremum in ξ at time θ satisﬁes ∂t u > 0 if ξ is a minimum. ∂t u(ξ. θ))) is bounded. 3 . [30] in the context of linear diﬀusion ﬁltering.28).2. Then.2 Informationreducing properties Nonenhancement of local extrema Koenderink [240] required that a scalespace evolution should not create new level curves when increasing the scale parameter. he imposed that at spatial extrema with nonvanishing determinant of the Hessian isophotes in scalespace are upwards convex. as we shall see now. ∆u = trace(Hess(u)) = η1 + η2 > 0. θ) with nonvanishing Hessian. θ).2.29) giving immediately the causality requirement (2.3 Theorem 2 (Nonenhancement of local extrema). (2. cf. However.30) (2. for instance.5 and [342]. Then we have 2 2 2 2 ∂t u = i=1 j=1 ∂xi dij (Jρ (∇uσ )) ∂xj u + i=1 j=1 dij (Jρ (∇uσ )) ∂xi xj u. (2. Let u be the unique solution of (Pc ) and consider some θ > 0. (2. it is also satisﬁed by nonlinear diﬀusion scalespaces. Nonenhancement of local extrema has ﬁrst been used by Babaud et al. ∂t u(ξ. SCALESPACE PROPERTIES 65 2. if ξ is a local maximum. As in the linear diﬀusion case.32) vanishes in (ξ. Thus. This implication is easily seen: In the ﬁrst case.31) Proof: Let D(Jρ (∇uσ )) =: (dij (Jρ (∇uσ ))). θ) = 0 and ∂xi dij (Jρ (∇uσ (ξ. He showed that this constraint can be written as sgn(∂t u) = sgn(∆u). (2.32) Since ∇u(ξ.. the eigenvalues η1 . If this is satisﬁed. Suppose that ξ ∈ Ω is a local extremum of u(. θ) > 0.4.
there exists an orthogonal matrix S ∈ IR2×2 such that S T DS = diag(λ1 .15) following Hummel’s reasoning. pp. see [189] for more details. θ). B := (bij ) := S T HS are negative deﬁnite. let us assume that (ξ. which do violate this principle. The following theorem states that anisotropic diﬀusion ﬁltering always leads to a constant steadystate. . It should be noted that nonenhancement of local extrema is just one possibility to end up with Koenderink’s causality requirement. Hence. and by the invariance of the trace with respect to orthogonal transformations it follows that ∂t u(ξ. for t → ∞. nonlinear diﬀusion is always in the forward region at extrema. we obtain the simplest possible image representation. θ) = trace (DH) = trace (S T DS S T HS) = trace (ΛB) 2 = i=1 λi bii < 0.66 CHAPTER 2. 2 Nonenhancement of local extrema distinguishes anisotropic diﬀusion from classical contrast enhancing methods such as highfrequency emphasis [163. θ) is a local maximum where H := Hess(u(ξ. it is desirable that. one proceeds in the same way utilizing the positive deﬁniteness of the Hessian. Lyapunov functionals and behaviour for t → ∞ Since scalespaces are intended to subsequently simplify an image. λ2 being the positive eigenvalues of D. 182– 183]. 2). λ2 ) =: Λ with λ1 . θ)) and. θ))) is positive deﬁnite. Now. Then we have bii < 0 (i = 1. namely a constant image with the same average grey value as the original one. thus. CONTINUOUS DIFFUSION FILTERING We know that the diﬀusion tensor D := D(Jρ (∇uσ (ξ. This is due to the class of Lyapunov functionals associated with the diﬀusion process. Although possibly behaving like backward diﬀusion across edges. Another way to establish causality is via the extremum principle (2. This ensures its stability. If ξ is a local minimum of u(x.
e. b. V ∈ C[0. b] with r ′′ ≥ 0 on [a.4. if r ′′ > 0 on [a. respectively. the function V (t) := Φ(u(t)) := Ω r(u(x. for all t ≥ 0. then V (t) = Φ(u(t)) is a strict Lyapunov functional: (iii) (iv) (v) ¯ on Ω a. Suppose that u is the solution of (Pc ) and let a. b] with r ′′ ≥ 0 on [a. µ and M be deﬁned as in (Pc ).2. b]. . on Ω ¯ If t > 0. then V ′ (t) = 0 if and only if u(t) = Mf on Ω. t→∞ lim u(t) − Mf Lp (Ω) =0 for ≤ = Ω 1 Ω 1 r Ω Ω Ω u(x. SCALESPACE PROPERTIES 67 Theorem 3 (Lyapunov functionals and behaviour for t → ∞). b]. t) dx dy r(u(x. t)) dx Ω r(u(x. ∞) ∩ C1 (0. ∞). t)) dx (2.34) = Φ(u(t)).33) is a Lyapunov functional: (i) (ii) Φ(u(t)) ≥ Φ(Mf ) for all t ≥ 0.24). T ] p ∈ [1. Then the following properties are valid: (a) (Lyapunov functionals) For all r ∈ C2 [a. t) = µ is uniform on Ω. Φ(Mf ) = Ω t→∞ ¯ In the 1D case. ∞) and V ′ (t) ≤ 0 for all t > 0.22) and (2. Moreover. the convergence lim u(x. on Ω and ¯ u(t) = Mf on Ω × (0. (b) (Convergence) (i) (ii) Proof: (a) (i) Since r ∈ C2 [a. we know that r is convex on [a.e. Φ(u(t)) = Φ(Mf ) ⇐⇒ u(t) = Mf u(t) = Mf V (0) = V (T ) for T > 0 ⇐⇒ (if t > 0) (if t = 0) f = Mf a. b]. Using the average grey level invariance and Jensen’s inequality we obtain. (2. t)) dx dy (2. b].
D(Jρ(∇uσ ))∇u dx ≥0 ≥0 ≤ 0. the divergence theorem yields V ′ (t) = Ω r ′ (u) div (D(Jρ(∇uσ ))∇u) dx r ′ (u) D(Jρ(∇uσ ))∇u. ¯ If t > 0.68 CHAPTER 2. Ω) and α := 1 Ω1  u dx = Ω1 1 Ω2  u dx =: β. the limit t → 0+ gives the announced continuity in 0. L2 (Ω)). we may choose a constant L := max r ′(s) s∈[a. Ω2  ∈ (0. Then. CONTINUOUS DIFFUSION FILTERING (ii) Let us start by proving the continuity of V (t) in 0. T ]. (iii) Let Φ(u(t)) = Φ(Mf ). there exists a partition Ω = Ω1 ∪ Ω2 with Ω1 .b] such that for all t > 0. the Lipschitz condition r(u(x. Thus.34) implies that u(t) = const. n dS Γ =0 = − Ω r ′′ (u) ∇u. t))−r(f (x)) ≤ L u(x. b] and 1 r Ω Ω u dx = r Ω1  Ω2  α+ β Ω Ω . then u(t) is continuous in Ω. Thanks to the maximum–minimum principle. b] it follows that r is strictly convex on [a. assume ¯ that u is not constant on Ω. b]. Let us now show that equality in ¯ the estimate (2. on Ω. on Ω.e. t)−f (x) is veriﬁed a. From this and the Cauchy–Schwarz inequality. we know that V is diﬀerentiable for t > 0 and V ′ (t) = Ω r ′ (u) ut dx. we get V (t)−V (0) ≤ Ω1/2 r(u(t))−r(f ) ≤ Ω1/2 L u(t)−f L2(Ω) L2 (Ω) . Ω2 From r ′′ > 0 on [a. by the continuity of u. To this end. By Theorem 1 and the boundedness of r ′ on [a. and by virtue of u ∈ C([0.
t). SCALESPACE PROPERTIES Ω1  Ω1  r(α) + r(β) Ω Ω 1 1 r(u) dx + r(u) dx ≤ Ω Ω < Ω1 Ω2 69 = 1 Ω r(u) dx. V ′ (t) = − r ′′ (u) ∇u. θ) − µ dx ≤ Ω1/2 u(θ) − Mf L2 (Ω) = 0. t) dx >0 Ω and the smoothness of u we obtain ¯ ∇u. t) = µ on Ω. on Ω. the Cauchy–Schwarz inequality gives u(x. Ω Since u ∈ C([0. For θ > 0.e. D(Jρ(∇uσ ))∇u ¯ on Ω × (0. Thanks to the ¯ average grey value invariance we ﬁnally obtain u(t) = Mf on Ω. for all x ∈ Ω. this yields u(x. there exists some constant ν > 0. on Ω. D(Jρ(∇uσ (x. we have ∇u(x. let u(x. Thus. we have u(θ) = Mf for all θ > 0. From (i) and (ii). (iv) Let t > 0 and V ′ (t) = 0.34) we observe that.) on Ω implies Φ(u(t)) = Φ(Mf ). D(Jρ(∇uσ ))∇u = 0 on Ω. we conclude that Φ(u(θ)) = Φ(Mf ) for all θ > 0. T ]. ∞). Ω If we utilize this result in the estimate (2. Conversely. t)) ∇u(x.4. So let us turn to the case t = 0.e. Thus. Conversely. By the uniform boundedness of D. the limit θ → 0+ ﬁnally yields u(0) = Mf a. L2 (Ω)). ¯ Φ(u(t)) = Φ(Mf ) implies that u(t) = const. and the average grey level invariance ﬁnally gives u(x. on Ω. D(Jρ(∇uσ ))∇u dx = 0. Then from 0 = V ′ (t) = − r ′′ (u(x. Due to the continuity of ∇u.e. t) = µ on Ω. Ω . t) = 0 a. t) = const.2. Then. t)))∇u(x. for t > 0. such that ν ∇u2 ≤ ∇u. it is obvious that u(t) = Mf (a.
Thus.) on Ω holds for all t ∈ [0. For t > 0. we know that v(t) f −Mf L∞ (Ω) . Therefore. for 1 ≤ p < q < ∞. Since V is decreasing. the limit t → 0+ yields f = Mf a. on Ω. D(Jρ(∇vσ ))∇v dx. Then for any t ∈ [ǫ. if u(t) = Mf (a.e. q ≥ 2. T ]. p. We multiply this equation by v. Now. . we have V (t) = const. T ]. integrate. Ω Ω Since ∇vσ is bounded. and part (iv) implies that u(t) = Mf on Ω. [9. there exists some x0 with v(x0 ) = 0. Ω As u ∈ C([0. o v(t) Lp (Ω) ≤ Ω(1/p)−(1/q) · v(t) Lq (Ω) → 0. This yields d v dt 2 L2 (Ω) ≤ −2ν C0 v 2 L2 (Ω) and hence the exponential decay of v L2 (Ω) to 0. for q ∈ IN. and use the divergence theorem to obtain vvt dx = − ∇v. we ﬁnd some ν > 0 such that 1d ( v 2 dt 2 L2 (Ω) ) ≤ −ν ∇v 2 L2 (Ω) . T ]. (b) (i) By the grey level shift invariance we know that v := u−Mf satisﬁes the diﬀusion equation as well. we get v(t) q Lq (Ω) L∞ (Ω) is bounded by ≤ f −Mf q−2 L∞ (Ω) · v(t) 2 L2 (Ω) → 0.e. it is evident that V (0) = V (T ). Let ǫ > 0. CONTINUOUS DIFFUSION FILTERING (v) Suppose that V (T ) = V (0). T ]. This proves the assertion.70 CHAPTER 2. we have V ′ (t) = 0. Poincar´’s e inequality (cf. L2 (Ω)). the Cauchy–Schwarz inequality gives f − Mf  dx ≤ Ω1/2 f − u(t) L2 (Ω) . and H¨lder’s inequality gives. on [0. Conversely. 122]) may be applied giving v 2 L2 (Ω) ≤ C0 ∇v 2 L2 (Ω) with some constant C0 = C0 (Ω) > 0. By the maximum principle.
a). ¯ This also gives u(tij ) → u in L2 (0. SCALESPACE PROPERTIES 71 (ii) To prove uniform convergence in the onedimensional setting. Hence. ∞) ∩ C1 (0. there exists a subsequence (tij ) → ∞ and some u with ¯ u(tij ) → u in C0.4. (2. a] o 1 [7. (ti )i∈IN → ∞ and from the convergence of (V (i))i∈IN it follows that V ′ (ti ) → 0. Thus.37) Part (a) tells us that u(t)−Mf p p (Ω) is a Lyapunov function for p ≥ 2.2. the space of H¨ldercontinuous functions on [0. is compact for α ∈ (0. a].36) L2 (Ω) → 0. Since we already know from (b)(i) ¯ 2 that u(tij ) → Mf in L (0. 144] we know that the embedding from H1 (0. Thus. 9–12].35) Thanks to the uniform positive deﬁniteness of D there exists some ν > 0 such that. (2. From part (a) we know that V (t) := a 0 u2 (x. Therefore. By virtue of the Rellich– Kondrachov theorem [7. i+1) : V ′ (ti ) = V (i + 1) − V (i). p. pp. Since V ∈ C[0. ∞) the mean value theorem implies ∃ ti ∈ (i. Equations (2. a V (t) = −2 ′ u2 D(Jρ (∂x uσ )) dx x 0 a ≤ −2ν ≤ 0.α [0. Let Ω = (0. Hence. a). the sequence (V (i))i∈IN converges. u(t)−Mf L∞ (Ω) = lim u(t)−Mf Lp (Ω) p→∞ . a). it follows that u = Mf . 2 ). a) into C0. we can generalize and adapt methods from [202] to our case. u(ti ) is a bounded sequence in H1 (0. a].36) yield ux (ti ) u2 dx x 0 (2.35) and (2. a). for t > 0. ¯ j→∞ lim u(tij ) − Mf L∞ (Ω) = 0. t) dx is nonincreasing and bounded from below.α [0. L Thus.
∞): (a) (b) u(t) Lp (Ω) for all p ≥ 2. the competition evolves in a certain direction all the time: although backward diﬀusion may be locally superior. Corollary 1 (Special Lyapunov functionals). Then the following functions are decreasing for t ∈ [0. Therefore. (c) H[u(t)] := Ω u(x. t) − µ)2n dx for all n ∈ IN. Considering the Lyapunov functions associated with r(s) := sp . 1 Ω Ω M2n [u(t)] := (u(x. lim u(t)−Mf t→∞ (2. Let us have a closer look at what might be the meaning of this global result in the context of image processing. Since the class (Pc ) does not forbid contrast enhancement it admits processes where forward diﬀusion has to compete with backward diﬀusion. r(s) := (s−µ)2n and r(s) := s ln s. As a special case of (a) it follows that the energy u(t) 2 2 (Ω) is reduced by L diﬀusion. Theorem 3 is of importance as it states that the regularization by convolving with Kσ tames the backward diﬀusion in such a way that forward diﬀusion wins in the long run. Part (b) gives a probabilistic interpretation of anisotropic diﬀusion ﬁltering. the preceding theorem gives the following corollary. Consider the intensity in an image f as a random variable Zf with distribution . Let u be the solution of (Pc ) and a and µ be deﬁned as in (Pc ) and (2. t)) dx. Moreover.72 CHAPTER 2. Corollary 1 oﬀers multiple possibilities of how to interpret nonlinear anisotropic diﬀusion ﬁltering as a smoothing transformation. The smoothness of u establishes ﬁnally that the convergence t→∞ lim u(x. if a > 0. t) = µ 2 ¯ is uniform on Ω. the global result – denoted by the Lyapunov functional – becomes permanently better for forward diﬀusion. respectively. CONTINUOUS DIFFUSION FILTERING is also nonincreasing. t) ln(u(x.37) we conclude that t→∞ L∞ (Ω) exists and from lim u(t)−Mf L∞ (Ω) = 0.22).
SCALESPACE PROPERTIES 73 Ff (z). t)) dx Ω (2. in spite of possible contrastenhancing properties. are overcompensated by smoothing in other areas.4. The interpretation of the entropy in terms of Lyapunov functionals carries also over to generalized entropies. . Since anisotropic diﬀusion ﬁlters increase the entropy the corresponding scalespace embeds the genuine image f into a family of subsequently likelier versions of it which contain less information. Higher moments are more diﬃcult to interpret. By the average grey level invariance. i. although they do provide important information for tasks like texture discrimination [163. All decreasing even moments demonstrate that the image becomes smoother during diﬀusion ﬁltering. In a classic scalespace representation. In particular. (2. Let us ﬁnally point out another interpretation of the Lyapunov functionals. It is a common tool for constructing measures for the relative smoothness of the intensity distribution. Moreover. pp. one transforms the image from a local to a more global 4 Without loss of generality we omit the normalization.39) The second central moment (the variance) characterizes the spread of the intensity about its mean. By increasing t. the time t plays the role of the scale parameter. a measure of uncertainty and missing information [63]. If we choose another probabilistic model of images. From all the previous considerations. we may regard it also as a twodimensional density. 414–415]. see [390] for more details.40) is called the entropy of u(t).e. (2.38) and it follows that M2n [u(t)] is just the even central moment z−EZu(t) IR 2n dFu(t) (z). then part (c) characterizes the informationtheoretical side of our scalespace. Provided the initial image f is strictly positive on Ω. for t → ∞. local eﬀects such as edge enhancement. anisotropic diﬀusion does really simplify the original image in a steady way.4 Then. µ is equal to the expected value EZu(t) := IR z dFu(t) (z). we recognize that. t) ln(u(x.2. it helps to explain the success of nonlinear diﬀusion ﬁltering as a preprocessing step for subsampling as observed in [144]. namely a constant image. Hence. which object to increase central moments. Ff (z) is the probability that an arbitrary grey value Zf of f does not exceed z. The fourth moment is frequently used to describe the relative ﬂatness of the grey value distribution. S[u(t)] := − u(x. This informationreducing property indicates that anisotropic diﬀusion might be generally useful in the context of image compression. the process reaches the state with the lowest possible information.
for linear diﬀusion scalespaces and morphological scalespaces. This can be achieved by taking some Lyapunov function Φ(u(t)) and investigating the expression Ψ(u(t)) := Φ(f ) − Φ(u(t)) . . It gives the average globality of u(t) and its value can be used to measure the distance of u(t) from the initial state f and the ﬁnal state Mf . We have seen in Chapter 1 that. it is possible to associate with the evolution time a corresponding spatial scale. 308]. one can only deﬁne an average measure for the globality of the representation. the situation is more complicated. Φ(f ) − Φ(Mf ) (2.74 CHAPTER 2.41) We observe that Ψ(t) increases from 0 to 1. CONTINUOUS DIFFUSION FILTERING representation. Experiments in this direction can be found in [431. Prescribing a certain value for Ψ provides us with an aposteriori criterion for the stopping time of the nonlinear diﬀusion process. In the nonlinear diﬀusion case. Since the smoothing is nonuniform. however.
N ≥ 2. whose components fj .. extremum principles. In order to specify the requirements for our semidiscrete ﬁlter class we ﬁrst recall a useful deﬁnition of irreducible matrices [407.Chapter 3 Semidiscrete diﬀusion ﬁltering The goal of this chapter is to study a semidiscrete framework for diﬀusion scalespaces where the image is sampled on a ﬁnite grid and the scale parameter is continuous. average grey level invariance. Afterwards we shall discuss whether it is possible to obtain such ﬁlters from spatial discretizations of the continuous models that have been investigated in Chapter 2. 3. and convergence to a constant steadystate.. We will see that there exists a ﬁnite stencil on which a diﬀerence approximation of the spatial derivatives are in accordance with the semidiscrete scalespace framework.kr ∈ J with k0 = i and kr = j such that akp kp+1 = 0 for p = 0.N represent the grey values at each pixel. 18–20]. We denote the index set {1. Deﬁnition 1 (Irreducibility).....r−1.. Lyapunov functions. j = 1. N} by J.1 The general model A discrete image can be regarded as a vector f ∈ IRN .... A matrix A = (aij ) ∈ IRN ×N is called irreducible if for any i. The semidiscrete problem class (Ps ) we are concerned with is deﬁned in the following way: 75 .. . This leads to a system of nonlinear ordinary diﬀerential equations (ODEs).. j ∈ J there exist k0 . We shall investigate conditions under which one can establish similar properties as in the continuous setting concerning wellposedness.... pp.
while average grey value invariance uses (S2) and (S3).76 CHAPTER 3. while (S4) and (S5) play a similar role as the nonnegativity of the eigenvalues of D and its uniform positive deﬁniteness. aij (u) ≥ 0 ∀ i = j. IRN ×N ) for every bounded (Ps ) subset of IRN .2 Theoretical results Before we can establish scalespace results. (S2) and (S3) correspond to the speciﬁc structure of the divergence expression with a symmetric diﬀusion tensor D. N Not all of these requirements are necessary for every theoretical result below. N ∀ u ∈ IR . dt u(0) = f. respectively. where A = (aij ) has the following properties: (S1) Lipschitzcontinuity of A ∈ C(IRN . Find a function u ∈ C1 ([0. (S5) irreducibility for all u ∈ IRN . ∞). the proof of a maximum–minimum principle involves (S3) and (S4). SEMIDISCRETE DIFFUSION FILTERING Let f ∈ IRN . 3. j∈J (S3) vanishing row sums: (S4) nonnegative oﬀdiagonals: aij (u) = 0 ∀ i ∈ J. N ∀ u ∈ IR . (S2) symmetry: aij (u) = aji (u) ∀ i. (S1) is needed for wellposedness. and strict Lyapunov functions and the convergence to a constant steadystate require (S5) in addition to (S2)–(S4). j ∈ J. . IRN ) which satisﬁes an initial value problem of type du = A(u) u. This is done in the theorem below which also states the continuous dependence of the solution and a maximum–minimum principle. it is of importance to ensure the existence of a unique solution. This indicates that these properties reveal some interesting parallels to the continuous setting from Chapter 2: In both cases we need smoothness assumptions to ensure wellposedness. ∀ u ∈ IR . The existence of Lyapunov functions can be established by means of (S2)–(S4).
u) := ψ(u) := A(u) u is continuous on B0 := [0. In order to prove existence and uniqueness of a solution of (Ps ) in R0 := (t. θ].2) (3. Assume that the problem (Ps ) has a unique solution on [0. j∈J ∀ i ∈ J. T )]. u) satisﬁes a global Lipschitz condition on R0 with respect to u. For every T > 0 the problem (Ps ) has a unique solution u(t) ∈ C1 ([0. since it is a composition of continuous functions. THEORETICAL RESULTS 77 Theorem 4 (Wellposedness. However. IRN ). ∀ t ∈ [0.2. o Let t0 := 0 and β > 0. T ] × u ∈ IRN u ∞ ≤ f ∞ +β . T ]. p.1) (3. This solution depends continuously on the initial value and the righthand side of the ODE system. t0 +min( β . T ]. First we show that the derivative of the largest component of u(t) is nonpositive for every . by the compactness of B0 there exists some c > 0 with φ(t. u) ∞ ≤c ∀ (t. φ(t. j∈J Proof: (a) Local existence and uniqueness Local existence and uniqueness are proved by showing that our problem satisﬁes the requirements of the Picard–Lindel¨f theorem [432. (b) Maximum–minimum principle We prove only the maximum principle.3. extremum principle). this follows directly from the fact that A is Lipschitzcontinuous on {u ∈ IRN  u−f ∞ ≤ β}. u) ∈ B0 . since the proof for the minimum principle is analogous. u) t ∈ [t0 . (3. and it satisﬁes the extremum principle a ≤ ui (t) ≤ b where a := min fj . Evidently. 59]. Moreover.3) b := max fj . u−f c ∞ ≤ β ⊂ B0 we have to show that φ(t.
for t2 := min ϑp . t1 ).4). Let ε > 0 and set εt . duk dt = j∈J akj (u) uj akk (u) uk + j∈J\{k} = ≤ (S4) akj (u) uj ≥0 ≤uk uk · 0. let us consider some p ∈ P . let P := {p ∈ J  uεp (0) = max uεj (0)}.6) Next. akj (u) j∈J = (3. uε (t) := u(t) − . . j∈J duεp (0) = dt By means of i∈J\P dup (0) − ε < 0 dt ≤0 ∀ p ∈ P. (3. it follows that p∈P ∀ t ∈ [0. ϑp ) max uεp(t) < max uεj (0) p∈P j∈J ∀ t ∈ (0. ϑp ). t2 ). θ]. (3. [201]). Let uk (ϑ) := max uj (ϑ) for some arbitrary ϑ ∈ [0. Then. Due to (3. . θ) with duεp (t) < 0 dt Thus.4) Let us now prove that this implies a maximum principle (cf. we have uεp (t) < uεp (0) and.5) max uεi (0) < max uεj (0). for t = ϑ.78 CHAPTER 3. If we keep j∈J this k ﬁxed we obtain. ∀ t ∈ (0. εt Moreover.7) . θ) such that i∈J\P max uεi(t) < max uεj (0) j∈J ∀ t ∈ [0. θ]. j∈J and the continuity of u there exists some t1 ∈ (0. by (3. SEMIDISCRETE DIFFUSION FILTERING t ∈ [0.5) and the smoothness of u we may ﬁnd a ϑp ∈ (0. (3.
T )]. t3 ) with uεk (t3 ) − uεk (t4 ) duεk (t5 ) = dt t3 − t4 (3. (3.10) duεk (t) < 0 dt The mean value theorem.3.9) > 0.8) Now we prove that this estimate can be extended to the case t ∈ (0. Then the minimality of t3 yields j∈J uεk (t) < uεk (t3 ) and inequality (3. (c) Global existence and uniqueness Global existence and uniqueness follow from local existence and uniqueness when being combined with the extremum principle. the estimates (3. (3. we know that the problem (Ps ) has a unique solution u(t) for t ∈ [t0 . θ). implies that we ﬁnd a t5 ∈ (t4 .2. there exists some t3 which is the smallest time in (0.6) and (3.4) gives duεk (t3 ) = dt du dt ∀ t ∈ (0. c . Using the notations and results from (a).8) must be valid on the entire interval (0. Then. however. t0 + min( β . θ). for t0 := min(t1 . θ) such that max uεj (t3 ) = max uεj (0). t3 ) with ∀ t ∈ (t4 . θ]. which contradicts (3. t2 ). Hence. Together with u = lim uε and the continuity of u this yields the announced ε→0 maximum principle max uj (t) ≤ max uj (0) j∈J j∈J ∀ t ∈ [0. by virtue of the intermediate value theorem. THEORETICAL RESULTS Hence. t3 ). (3. t3 ]. t0 ).7) give max uεj (t) < max uεj (0) j∈J j∈J 79 ∀ t ∈ (0. (3. dt ≤0 Due to the continuity of there exists some t4 ∈ (0. assume the opposite is true. To this end.9) duk (t3 ) − ε < 0.10). j∈J j∈J Let uεk := max uεj (t3 ).
In this case the results in [412. with respect to its second argument. As a consequence. u) t ∈ [t1 . u−g c ∞ ≤β . min( 2β . u) satisﬁes a global Lipschitz condition on Sα := (t. p.80 CHAPTER 3. t1 + min( β . (Ps ) reveals a unique solution on [0. 93] ensure that for every ε > 0 there exists a δ > 0 such that the solution u of ˜ the perturbed problem d˜ u ˜ ˜ = φ(t. (d) Continuous dependence Let u(t) be the solution of du = φ(t. T ]. u) ∞ ≤c ∀ (t. dt u(t1 ) = g. u) = A(u)u is continuous on B1 := [0. T )]. Hence. it is suﬃcient to prove that φ(t. t1 +min( β . Clearly. with the same c as in (a). Using the same considerations as in (a) one shows that φ is Lipschitzcontinuous on R1 := (t. v) t ∈ [0. u) is continuous. T )]. t) = ψ(u) = A(u)u. u). by iterating c this reasoning. dt ˜ u(0) = f ˜ . T ]. φ(t. dt u(0) = f for t ∈ [0. u). Hence. u) ∈ B1 . and by the extremum principle we know that B1 ⊂ B0 . the existence of a unique solution can be extended to the entire interval [0. the considered problem has a unique solution on [t1 . and consider the problem c du = A(u) u. φ(t. and. v−u ∞ ≤α . T ] and φ(u. c Therefore. In order to show that u(t) depends continuously on the initial data and the righthand side of the ODE system. T ] × u ∈ IRN u ∞ ≤ g ∞ +β . g := u(t1 ). the extremum principle is valid on [0. T ] as well. T ). SEMIDISCRETE DIFFUSION FILTERING Now let t1 := t0 + min( β . T )]. and that there exists some α > 0 such that φ(t.
Similar to the the local existence and uniqueness proof. 2 3.11) ∀ t ≥ 0. < δ for v−u ∞ < α exists in [0. T ] and satisﬁes the inequality u(t) − u(t) ˜ ∞ < ε. ∞) and concludes the proof. k∈J j∈J j∈J duj = dt j∈J ajk (u) uk = j∈J k∈J which shows that uj (t) is constant on [0. (3. SCALESPACE PROPERTIES ˜ with continuous φ and ˜ φ(t. Thus. So let us focus on average grey level invariance now. v) − φ(t.12) Proof: By virtue of (S2) and (S3) we have j∈J ajk (u) = 0 for all k ∈ J. Proposition 2 (Conservation of average grey value). v) ˜ f −f ∞ ∞ 81 < δ. The average grey level 1 fj µ := N j∈J is not aﬀected by the semidiscrete diﬀusion ﬁlter: 1 N uj (t) = µ j∈J (3. translation invariance only makes sense for translations in grid direction with multiples of the grid size.3.3. the global Lipschitz condition on Sα follows direcly from the fact that A is Lipschitzcontinuous on {v ∈ IRN  v−u ∞ ≤ α}. for t ≥ 0. On the other hand. . 2 This property is in complete accordance with the result for the continuous ﬁlter class.3 Scalespace properties It is evident that properties such as grey level shift invariance or reverse contrast invariance are automatically satisﬁed by every consistent semidiscrete approximation of the continuous ﬁlter class (Pc ). ajk (u) uk = 0. and isometry invariance is satisﬁed for consistent schemes up to an discretization error.
Moreover. b] we know that r is convex on [a. t→∞ Proof: (a) (i) Since r ′ is increasing on [a. ∞) and V ′ (t) ≤ 0 for all t ≥ 0. V ∈ C1 [0. 1 r uj Φ(c) = i=1 j=1 N 1 ≤ N i=1 N N N N N j=1 r(uj ) = j=1 r(uj ) (3.. and let c := (µ. b]. b]. Average grey level invariance and this convexity yield.2). . Then the following properties are valid: (a) (Lyapunov functions) For all r ∈ C1 [a. it is possible to ﬁnd a large class of Lyapunov functions which establish smoothing scalespace properties and ensure that the image tends to a constant steadystate with the same average grey level as the initial image. the function V (t) := Φ(u(t)) := i∈J r(ui (t)) is a Lyapunov function: (i) (ii) Φ(u) ≥ Φ(c) for all t ≥ 0.11). b] with increasing r ′ on [a. SEMIDISCRETE DIFFUSION FILTERING Similar to the continuous setting. if r ′ is strictly increasing on [a. b.3). let a.82 CHAPTER 3. . respectively.. then V (t) = Φ(u(t)) is a strict Lyapunov function: (iii) (iv) Φ(u) = Φ(c) V ′ (t) = 0 ⇐⇒ u=c u=c ⇐⇒ (b) (Convergence) lim u(t) = c. and (3. (3. and µ be deﬁned as in (3. Theorem 5 (Lyapunov functions and behaviour for t → ∞).13) = Φ(u). b]. µ)⊤ ∈ IRN .. µ. for all t ≥ 0. Let u(t) be the solution of (Ps ).
14) = Since r ′ is increasing. IRN ) and r ∈ C1 [a. (iii) Let us ﬁrst prove that equality in the estimate (3. 1− N 1 N Then. we always have (ui+k − ui ) r ′ (ui ) − r ′ (ui+k ) ≤ 0.i(u) (ui −ui+k ) r ′ (ui+k ) ai.i+k (u) (ui+k −ui) r ′ (ui)−r ′ (ui+k ) .3. η := j=1 j=j0 uj 1 .i+k (u) (ui+k −ui) r ′ (ui ) ai+k. b]. b]. Hence. we get r 1 uj i=1 N N = r 1 N 1 ≤ N < N 1 η N 1 r(uj0 ) + 1− r(η) N N 1 r(uj0 ) + r(uj ) j=1 N 1− j=j0 1 uj + N 0 = 1 r(uj ). Using the prerequisites (S2) and (S3) we get V ′ (t) N = (S3) i=1 N N dui ′ r (ui ) dt aij (u) (uj −ui) r ′ (ui) N i−1 = i=1 j=1 N = i=1 + j=i+1 j=1 = + (S2) N N −i i=1 k=1 N N −i i=1 k=1 N N −i i=1 k=1 aij (u) (uj −ui ) r ′ (ui ) ai.14) implies that V ′ (t) ≤ 0 for t ≥ 0. η < uj0 . Since r ′ is strictly increasing on [a. SCALESPACE PROPERTIES 83 (ii) Since u ∈ C1 ([0. ∞). we know that r is strictly convex. ∞).3. it follows that V ∈ C1 [0. To this end. j=1 N . suppose that ui0 := min ui < max uj =: uj0 and let i j N With this and (S4). equation (3. (3.13) implies that all components of u are equal.
. t→∞ Now assume that η > 0. ≤0 and by virtue of the symmetry of A(u) it follows that aij (u) (uj −ui ) r ′ (ui)−r ′ (uj ) = 0 ∀ i.. we know that lim V (t) =: η exists and η ≥ 0. [180] for an introduction to these techniques. Since V (t) is decreasing and bounded from below by 0. Conversely. (3. From (3. The irreducibility of A(u) implies that there exist k0 .. Φ(x) < η ∀ x ∈ IRN . Since i0 and j0 are arbitrary.. . = ukr = uj0 . (3.14) we immediately conclude that V ′ (t) = 0. Conversely.84 CHAPTER 3.15) Now consider two arbitrary i0 . √ η). SEMIDISCRETE DIFFUSION FILTERING This shows that equality in (3.15) we get ui0 = uk0 = uk1 = .. x−c < β.i+k (u) (ui+k −ui ) r ′ (ui)−r ′ (ui+k ) . As r ′ is strictly increasing we have.. we obtain ui = const.. Consider the Lyapunov function V (t) := Φ(u(t)) := u(t)−c2 . for p = 0... By virtue of the grey level shift invariance we conclude that u = c. p = 0.. for all i ∈ J.. for β ∈ (0. . = 0 ⇐⇒ ukp = ukp+1 .kr ∈ J with k0 = i0 . j ∈ J. and the average grey level invariance gives u = c.16) By virtue of Φ(x) = x − c2 we know that. r − 1. Since u(t)−c is bounded from above by α := f −c we have u(t)−c ≤ α ∀ t ≥ 0.. kr = j0 . (b) The convergence proof is based on classical Lyapunov reasonings. which results from the choice r(s) := (s−µ)2 .14) we have 0 = V ′ (t) = N N −i i=1 k=1 ai. see e.g.. it is trivial that Φ(u) = Φ(c) for u = c. = uN ... for all i ∈ J. This proves the ﬁrst implication. and akp kp+1 (u) = 0.r − 1. (iv) Let V ′ (t) = 0. j0 ∈ J. Then from the representation (3.13) implies that u1 = . (ukp −ukp+1 ) r ′ (ukp+1 )−r ′(ukp ) From this and (3. let ui = const.
it follows V (t) = V (0) + t→∞ t 0 V ′ (θ) dθ ≤ V (0) − tM which implies lim V (t) = −∞ and. and obtain the following corollary. Since Φ(u(t)) ≥ η we conclude that u(t)−c ≥ β So from (3.18) is increasing with respect to t. the compactness of S. ∞): (a) (b) (c) u(t) p for all p ≥ 2. if a > 0. 1 N N j=1 M2n [u(t)] := H[u(t)] := N (uj (t) − µ)2n for all n ∈ IN.l. Therefore.3. j=1 uj (t) ln(uj (t)). Then the following functions are decreasing for t ∈ [0.2) and (3. thus. contradicts (a)(i). β < α. According to (a)(iii) this yields lim u(t) = c.g. Since all pnorms (p ≥ 2) and all central moments are decreasing. Hence the assumption η > ρ is wrong and we must have η = ρ. 85 (3. t→∞ 2 As in the continuous case.17) we have u(t) ∈ {x ∈ IRN  β ≤ x−c ≤ α} =: S ∀ t ≥ 0. . we can consider the Lyapunov functions associated with r(s) := sp . while the discrete entropy N S[u(t)] := − uj (t) ln(uj (t)) j=1 (3.17) By (a)(ii). Let u be the solution of (Ps ) and a and µ be deﬁned as in (3.16) and (3. respectively.3.o. SCALESPACE PROPERTIES Let w. ∀ t ≥ 0.11). Corollary 2 (Special Lyapunov functions). r(s) := (s−µ)2n and r(s) := s ln s.(iv). we observe that the semidiscrete setting reveals smoothing scalespace properties which are closely related to the continuous case. and β > 0 there exists some M > 0 such that V ′ (t) ≤ −M ∀ t ≥ 0.
j) is represented by a single index p(i.22) 2 dt n=1 l∈Nn (k) 2hl where Nn (k) consists of the one or two neighbours of pixel k along the nth coordinate axis (boundary pixels have only one neighbour) and gk := g ((H(u))k ).. (3.4... .86 CHAPTER 3. Then a consistent spatial discretization of the isotropic diﬀusion equation with homogeneous Neumann boundary conditions can be written as 2 gl + gk duk = (ul − uk ). consider some pixel k = p(i. h2 are given by h1 := a1 /n1 and h2 := a2 /n2 .. IR2×2 ). The structure tensor requires the calculations of convolutions with ∇Kσ and Kρ .. (3.21) Thus... yj ) where yj := (j − 1 xi := (i − 2 ) h1 . however. dt (3. First we shall verify that this is easily done for isotropic models. N}. we may approximate the structure tensor by some matrix H(u) = (hij (u)) where H ∈ C∞ (IRN .. 3. . n1 } × {1. n2 } → {1. respectively. the mixed derivative terms make it more diﬃcult to ensure nonnegative oﬀdiagonal elements. To this end. j) with 1 ≤ i ≤ n1 and 1 ≤ j ≤ n2 represents the location (xi . j). SEMIDISCRETE DIFFUSION FILTERING 3. a1 ) × (0. a2 ) be discretized by a grid of N = n1 · n2 pixels such that a pixel (i.4 Relation to continuous models In this section we investigate whether it is possible to use spatial discretizations of the continuous ﬁlter class (Pc ) in order to construct semidiscrete diﬀusion models satisfying (S1)–(S5).23) . A constructive existence proof is presented showing that for a suﬃciently large stencil it is always possible to ﬁnd such a nonnegative discretization. (3. we may replace the diﬀusion tensor D(Jρ (∇uσ )) by some scalarvalued function g(Jρ (∇uσ )). Next. . Let us now verify that a standard FD space discretization of an isotropic variant of (Pc ) leads to a semidiscrete ﬁlter satisfying the requirements (S1)–(S5).20) 1 ) h2 . j). In vector–matrix notation (3.19) (3. These pixels can be numbered by means of an arbitrary bijection p: {1. pixel (i. 2 and the grid sizes h1 . This concept is illustrated by investigating the situation on a (3×3)stencil in detail.22) becomes du = A(u) u.1 Isotropic case Let the rectangle Ω = (0. For this reason. In the anisotropic case.. In the spatially discrete case this comes down to speciﬁc vector–matrix multiplications. respectively.
r−1. (l = k).r−1. Then we have to ﬁnd k0 .. Moreover. 2}. Then. since g is positive. In order to show that A is irreducible. which has some interesting relations to the limit case of a semidiscrete Perona–Malik model. such that kq and kq+1 are neighbours for q = 0. let us consider two arbitrary pixels k and l.. If k = l. Since H ∈ C∞ (IRN . By the construction of A it is also evident that all row sums vanish.. This proves (S5). we may choose any arbitrary path k0 . Remarks: (a) We observe that (S1)–(S5) are properties which are valid for all arbitrary pixel numberings.. IR2×2 ) and g ∈ C∞ (IR2×2 ).. IRN ×N ). we have A ∈ C∞ (IRN . (S3) is satisﬁed. thus. it follows that akl ≥ 0 for all k = l and. (b) The ﬁlter class (Pc ) is not the only family which leads to semidiscrete ﬁlters satisfying (S1)–(S5). reveals all the discussed wellposedness and scalespace properties [425].. This is due to the fact that the extremum principle limits the modulus of discrete gradient approximations.3. akq kq+1 = for some n ∈ {1. Hence. This proves (S1). For k = l. [332]. a semidiscrete version of the Perona– Malik ﬁlter – which is to a certain degree illposed in the continuous setting (cf..4.24) that akk < 0.24) akl := Let us now verify that (S1)–(S5) are fulﬁlled..24) and the symmetry of the neighbourhood relation: l ∈ Nn (k) ⇐⇒ k ∈ Nn (l). the spatial discretization implicitly causes a regularization.3. The symmetry of A follows directly from (3.e. (S4) holds.1) – also satisﬁes (S1)–(S5) and... These results are also in accordance with a recent paper by Pollak et al... (3. we already know from (3. thus.kr ∈ J with k0 = k and kr = l such that akq kq+1 = 0 for q = 0. They also report stable behaviour of their process. (else). Interestingly. 1. i. In this case we have the trivial path k = k0 = kr = l... RELATION TO CONTINUOUS MODELS where the matrix A(u) = (akl (u))kl is given by gk +gl 2h2 n 2 n=1 l∈Nn (k) gk +gl 2h2 n 87 (l ∈ Nn (k)). They study an image evolution under an ODE system with a discontinuous right hand side. gkq + gkq+1 > 0 2h2 n 0 − ..kr ..
.. . 2 2 i = −2m+1. θi ]. (−2m + 1 ≤ i < −m). The “boundary pixels” of this stencil deﬁne 4m principal orientations βi ∈ (− π .... 2m according to βi := arctan arccot arccot ih2 mh1 (2m−i)h1 mh2 (i−2m)h1 mh2 (i ≤ m). 2m−2}. θi+1 ] ∪ =:Ii (θi−1 . due to the mixed derivative expressions. where ψ ∈ (− π . βi +βi+1 = π ). . Then there exists some m(κ) ∈ IN such that div (D ∇u) reveals a secondorder nonnegative FD discretization on a (2m+1)×(2m+1)stencil. i=1 =:Ii θi := 1 + 2 arctan (i = 0). . Now let Jm := {1. βi +βi+1 < π ).. sin ψ)⊤ and (− sin ψ.. i ∈ Jm : (− π . 2 (i ∈ {1.. 2 (i ∈ {1... it is not obvious how to ensure (S4). 2 (i = 2m−1)... k ∈ Jm such that the splitting div (D ∇u) = ∂eβ0 α0 ∂eβ0u + ∂eβk αk ∂eβku + ∂eβ2m α2m ∂eβ2mu (3. The theorem below states that this is always possible for a suﬃciently large stencil.. . (m < i ≤ 2m). 2m−2}. π ]. π ] into 4m−2 subin2 2 tervals Ii .. .25) θi := −θ−i (i ∈ {−2m+1.88 CHAPTER 3... −1}). 2m−2}. cos ψ)⊤ . SEMIDISCRETE DIFFUSION FILTERING 3.. Now we show that for 2 2 a suitable m there exists a stencil direction βk . Let D ∈ IR2×2 be symmetric positive deﬁnite with a spectral condition number κ. (i ∈ {1. 2m−1} and deﬁne a partition of (− π .. Let λ1 ≥ λ2 > 0 be the eigenvalues of D with corresponding eigenvectors (cos ψ. βi +βi+1 > π ).2 Anisotropic case If one wishes to transfer the results from the isotropic case to the general anisotropic setting the main diﬃculty arises from the fact that. Theorem 6 (Existence of a nonnegative discretization). . the nonnegativity of all oﬀdiagonal elements of A(u). π ] = 2 2 where 0 1 arctan 2 π 4 π 2 π 2 2 cot βi −tan βi+1 2 cot βi −tan βi+1 −1 i=−2m+1 2m−1 (θi . Proof: Let us consider some m ∈ IN and the corresponding (2m+1)×(2m+1)stencil. π ].4.. and It is not hard to verify that βi ∈ Ii for i ∈ Jm .
. So let us now verify (a)–(c). Once these assertions are proved a nonnegative secondorder discretization of (3. (a) In order to use subsequent indices. Furthermore. This can be done by proving the following properties: (a) Let ψ ∈ Ik and D = a b . Then (3.26) is satisﬁed for λ1 ≤ min cot(ρk −βk ) tan ρk . and γ2 := α2m . 2m−1}). sin βi )⊤ reveals nonnegative “directional diﬀusivities” α0 . c − b tan βk ≥ 0.m θk 1 (θ + βk ) 2 k 1 β 2 k =: κk. α2m along the stencil orientations β0 .. αk .27) (k ∈ {1. m→∞ i∈Jm = ∞. θk−1 (k = 1).3. Then a nonnegative splitting of type (3.m (3. .. and ϕ2 := π . cot(βk −ηk ) cot ηk λ2 with ρk := ηk := (c) lim min κi. (3. let ϕ0 := 0.25) is bc possible if min a − b cot βk . ϕ1 := βk where ψ ∈ Ik .. γ1 := αk . ϕ1 and ϕ2 we obtain the linear system div 2 2 i=0 γi cos2 ϕi i=0 γi sin ϕi cos ϕi γi sin ϕi cos ϕi i=0 2 i=0 2 γi sin2 ϕi ∇u . β2m ... 2m−2}).25) arises in a natural way. (k ∈ {2.26) (b) Inequality (3. as we shall see at the end of this chapter.25) 2 requires that div = + ∂ ∂x ∂ ∂y a b b c 2 ∇u = ∂ i=0 ∂eϕi 2 γi ∂u ∂eϕi cos ϕi γi (ux cos ϕi + uy sin ϕi ) i=0 2 sin ϕi γi(ux cos ϕi + uy sin ϕi ) i=0 = By comparing the coeﬃcients and using the deﬁnition of ϕ0 . let γ0 := α0 .. 1 cos2 βk 0 γ0 a 0 sin βk cos βk 0 γ1 = b 0 sin2 βk 1 γ2 c . βk . (k = 2m−1).4. RELATION TO CONTINUOUS MODELS 89 with eβi := (cos βi .
Now.θk ) min − C(ϕ) .90 CHAPTER 3. γ1 = sin βk cos βk γ2 = c − b tan βk . (b) Let λ1 λ2 ≤ κk. θk ]. By deﬁning 2 B(ϕ) := cos2 ϕ − sin ϕ cos ϕ cot βk . In particular. π ]. λ1 − λ2 ≥ 0. λ1 cos2 ψ + λ2 sin2 ψ (λ1 −λ2 ) sin ψ cos ψ (λ1 −λ2 ) sin ψ cos ψ λ1 sin2 ψ + λ2 cos2 ψ . In order to satisfy the nonnegativity of γ0 and γ2 we need that min a − b cot βk . SEMIDISCRETE DIFFUSION FILTERING which has the unique solution γ0 = a − b cot βk . since ψ ∈ Ik . (3. C(ϕ) := sin2 ϕ + sin ϕ cos ϕ cot βk we get C(ρk ) λ1 ≤ cot(ρk −βk ) tan ρk = − = λ2 B(ρk ) Since B(ϕ) < 0 on (βk .31) ϕ∈(βk . βk ]. (3. c − b tan βk ≥ 0. βk ∈ Ik we conclude that ψ and βk belong to the same quadrant. and since ψ.31) to the entire interval Ik = (θk−1 . 2 ∀ ϕ ∈ (βk . Thus.28) (3. γ1 is always nonnegative.m and consider the case 0 < βk < π .29) (3. b . θk ).30) From the structure of the eigenvalues and eigenvectors of D it is easily seen that b = (λ1 −λ2 ) sin ψ cos ψ. we have 0 ≤ λ1 B(ψ) + λ2 C(ψ) By the representation a b b c = = cos ψ − sin ψ sin ψ cos ψ λ1 0 0 λ2 cos ψ sin ψ − sin ψ cos ψ = (λ1 cos2 ψ + λ2 sin2 ψ) − (λ1 −λ2 ) sin ψ cos ψ cot βk . 2 ∀ ϕ ∈ [0. B(ϕ) and the continuity of B(ϕ) and C(ϕ) we may extend (3. π ) we have 2 λ1 B(ϕ) + λ2 C(ϕ) ≥ 0 Because of B(ϕ) ≥ 0 C(ϕ) ≥ 0 ∀ ϕ ∈ [− π .
1 − gm (i) the function gm (x) is bounded and attains its global maximum 1 xm := − 6 + 1 6 1 + 12 m2 h2 1 . RELATION TO CONTINUOUS MODELS we recognize that this is just the desired condition a − b cot βk ≥ 0. and the deﬁnition of θi implies that cot ρi − tan ρi = cot βi − tan βi+1 . yields cot ρi = tan ρi 1 cot βi − tan βi+1 + (cot βi − tan βi+1 )2 + 4 .4. Then. 91 (3. Solving for cot ρi and tan ρi .3.32). Then we have tan βi+1 = cot βi This gives (cot βi + tan βi+1 )2 = (cot βi − tan βi+1 )2 + 4 1− For m > in 1 h2 2 h1 (i + 1) h2 . ih2 1 4m2 i m2 h1 +i(i+1) h2 2 h h 2 1 =: 1 =: fm (i). respectively. mh1 mh1 = . In an analogous way one veriﬁes that λ1 ≤ cot(βk −ηk ) cot ηk λ2 =⇒ c − b tan βk ≥ 0. ρi = θi .32) For the case − π < βk < 0 a similar reasoning can be applied leading also to 2 (3. −1 Let us now assume that 1 ≤ i ≤ m−1. h2 2 . (c) Let us ﬁrst consider the case 1 ≤ i ≤ 2m−2. 2 By means of these results we obtain cot(ρi −βi ) tan ρi = cot βi + tan ρi = 1+ cot βi − cot ρi 2 (cot βi +tan βi+1 )2 (cot βi −tan βi+1 )2 +4 . 2 1 = − cot βi + tan βi+1 + (cot βi − tan βi+1 )2 + 4 .
92 CHAPTER 3. For m ≤ i ≤ 2m−2 similar calculations show that by means of tan βi+1 = cot βi one obtains m≤i≤2m−2 mh2 . For i = 2m−1 we have cot ρ2m−1 −β2m−1 tan ρ2m−1 π β2m−1 π β2m−1 − tan + 4 2 4 2 π β2m−1 + = tan2 4 2 → ∞ for m → ∞. in a similar way as above. (3. m→∞ lim i∈Jm min cot(ρi −βi ) tan ρi = ∞.34) we ﬁnally end up with the assertion m→∞ lim i∈Jm min κi. one establishes that m→∞ lim i∈Jm min cot(βi −ηi ) cot ηi = ∞. 2 . (3. min cot(ρi −βi ) tan ρi fm (xm ) − 1 → ∞ for m → ∞.m = ∞. SEMIDISCRETE DIFFUSION FILTERING Thus. (2m−i−1) h1 (2m−i) h1 = mh2 min cot(ρi −βi ) tan ρi → ∞ for m → ∞.34) From (3.33) and (3. = cot It is not hard to verify that for −2m + 1 ≤ i ≤ −1 the preceding results carry over. gm (i) ≤ gm (xm ) → 0+ which yields fm (i) ≤ This gives 1≤i≤m−1 for m → ∞. 1 → 1+ 1 − gm (xm ) ≥ 1+ 2 for m → ∞.33) Now. for 1 ≤ i ≤ m−1. Hence.
Thanks to the uniform positive deﬁniteness of D there exists an upper limit for the spectral condition number of D. a (2m+1)×(2m+1)stencil reveals much more directions than those 4m that are induced by the 8m “boundary pixels”. t) ≤ 4 f √ L∞ (Ω) 2π σ ¯ on Ω × (0. Since m = 1 we have a partition of (− π . Therefore.4. π ] = (− π . π ] into 4m−2 = 2 subintervals: 2 2 (− π . Let us now illustrate the ideas in the proof of Theorem 6 by applying them to a practical example: We want to ﬁnd a nonnegative spatial discretization of div (D∇u) on a (3×3)stencil. only three directions are suﬃcient to guarantee a nonnegative directional splitting. where D= a b b c and a. t) = (∇Kσ ∗u)(x.3. Moreover. (b) Especially for large m. Crandall and Lions [104] propose to study a convergent sequence of regularizations which can be approximated on a ﬁnite stencil. where the notations from Chapter 2 have been used. ∞). b and c may be functions of Jρ (∇uσ ). (d) The existence of a nonnegative directional splitting distinguishes the ﬁlter class (Pc ) from morphological anisotropic equations such as mean curvature motion. Thus. we may expect to ﬁnd stricter estimates than those given in the proof. π ] =: I−1 ∪ I1 . most of the stencil coeﬃcients can be set to zero. In this case it has been proved that it is impossible to ﬁnd a nonnegative directional splitting on a ﬁnite stencil [13]. (c) For a speciﬁed diﬀusion tensor function D it is possible to give apriori estimates for the required stencil size: using the extremum principle it is not hard to show that ∇uσ (x. As a remedy. RELATION TO CONTINUOUS MODELS Remarks: 93 (a) We observe that the preceding existence proof is constructive. This condition limit can be used to ﬁx a suitable stencil size. These estimates might be improved further by admitting more than 3 directions. 0] ∪ (0. even if we use only 3 directions. 2 2 2 2 . unless m is very small.
These bounds on the condition number attain their maximal value for h1 = h2 . SEMIDISCRETE DIFFUSION FILTERING I−1 and I1 belong to the grid angles β−1 = arctan − β1 = arctan h2 . 2h1 h2 which restricts the upper condition number for a nonnegative discretization with ψ ∈ I1 to 1 + sin β 1 + cos β κ1. = cot2 2 1 − cos β = 1 + sin β . 2 4 2 β η1 = . h1 =: β. . sin ψ) denotes the eigenvector to the larger eigenvalue λ1 of D. h1 . h2 h2 + h2 2 = b + b · 1 .30) we obtain as expressions for the directional diﬀusivities α−1 = b − b · h2 + h2 1 2 . (3.35) 1 − sin β 1 − cos β α0 = a − b · α1 α2 h1 .8284. 2h1 h2 h2 = c − b · . we get cot(ρ1 −β1 ) tan ρ1 = cot cot(β1 −η1 ) cot η1 π β π β − tan + 4 2 4 2 1 + cos β β = . With the notations from the proof of Theorem 6 we obtain π θ1 = . 2 Therefore. h2 h1 First we focus on the case ψ ∈ I1 where (cos ψ.1 = κ−1. In this case β = π gives 4 √ √ 1+ 1 2 2√ κ1.94 CHAPTER 3.28)–(3. 1 − sin β Thanks to the symmetry we obtain the same condition restriction for ψ ∈ I−1 .1 = = 3 + 2 2 ≈ 5. 2 θ1 + β1 π β ρ1 = = + .1 := min .36) 1 1− 2 2 By virtue of (3. (3.
j 2h2 1 ai−1. [294.j+1 +bi+1.j−1 +bi.j +ci.j  2h1 h2 bi−1. . Recently Kocan described an interesting alternative to obtain upper bounds for the stencil size as a function of the condition number [238]. j+1)) of A(u). yj ).j−1 4h1 h2 + bi. The other notations should be clear from the context as well. 120.j+1 4h1 h2 bi+1. y) ∂y u + ∂y b(x.j−1 4h1 h2 + bi.j +bi+1. the upper left stencil entry gives the element (p(i.j+1 +ci.j +bi.g.j  2h1 h2 + bi.j+1 4h1 h2 + bi.j 2h2 2 − bi.j  2h1 h2 bi+1. bi.j+1 +bi+1. y) ∂x u + ∂x b(x. y) ∂xx u + 2b(x. e. Usually it is studied for the expression a(x.j 2h2 1 − bi−1.j−1 +bi+1. y) ∂x u + ∂y c(x.j  2h1 h2 − ci.j  2h1 h2 − bi−1.j −b2i.j−1 +bi−1.j−1 −bi+1.j 2h2 1 − bi+1.j −b2i. for instance.j+1 −bi−1.j−1 4h1 h2 + bi−1.j 4h1 h ci. p(i−1.j−1 +bi−1.g.j +ai.j +2ai.j +bi. The problem of ﬁnding nonnegative diﬀerence approximations to elliptic expressions with mixed derivatives has a long history.j 2h2 2 − bi. j).j−1 −bi+1.3.j +b2i. see e. y) ∂xy u + c(x. The approach presented here extends these results to ∂x a(x.j−1 +bi.j+1 2h2 2 ai+1.j 4h1 h − ai−1. j) is the index of some inner pixel (i.j−1 +2ci.j +ai.j +bi.j+1 +2bi.j−1 +ci.j+1 +bi.j 4h1 h ci.j +b2i. j).j+1 −bi−1. Thus.j denotes a ﬁnite diﬀerence approximation of b(Jρ (∇uσ )) at some grid point (xi .j 4h1 h All nonvanishing entries of the pth row of A(u) are represented in this stencil.j +ai+1.j+1 4h1 h2 − bi−1. RELATION TO CONTINUOUS MODELS 95 This induces in a natural way the following secondorder discretization for div (D∇u): bi−1. y) ∂yy u.j+1 +bi+1. where p(i. 170]. y) ∂y u and establishes the relation between the condition number of a b and the nonbc negativity of the diﬀerence operator.4. His derivation is based on the diophantine problem of approximating irrationals by rational numbers.
96 CHAPTER 3. SEMIDISCRETE DIFFUSION FILTERING .
(D6) positive diagonal: ∀ v ∈ IR . . N} by J. and denote the index set {1. We consider the following discrete ﬁlter class (Pd ): (0) u = f. IR ).. j∈J (D3) unit row sum: (D4) nonnegativity: qij (v) = 1 ∀ i ∈ J.. j ∈ J. Lyapunov sequences and convergence to a constant steadystate. qii (v) > 0 ∀ i ∈ J. N ∀ v ∈ IR . uniqueness. (Pd ) . (k+1) (k) (k) u = Q(u ) u . where Q = (qij ) has the following properties: N N ×N (D1) continuity in its argument: Q ∈ C(IR . We shall see that this class comprises αsemiimplicit discretizations of the semidiscrete ﬁlter class (Ps ) as well as certain variants of them which are based on an additive operator splitting. N ∀ v ∈ IR .Chapter 4 Discrete diﬀusion ﬁltering This chapter presents a discrete class of diﬀusion processes for which one can establish similar properties as in the semidiscrete case concerning existence.. N ≥ 2. average grey level invariance.1 The general model As in Chapter 3 we regard a discrete image as a vector f ∈ IRN . ∀ k ∈ IN0 . continuous dependence of the solution on the initial image. N ∀ v ∈ IRN . 97 (D5) irreducibility for all v ∈ IRN . Let f ∈ IRN . j ∈ J. qij (v) ≥ 0 ∀ i. maximumminimum principle. Calculate a sequence (u(k) )k∈IN0 of processed versions of f by means of (D2) symmetry: qij (v) = qji (v) ∀ i. 4.
while the proof of the extremum principle requires (D3) and (D4).3) b := max fj . scalespace evolutions are evaluated exclusively at a ﬁnite number of scales. (b) The requirements (D1)–(D5) have a similar meaning as their semidiscrete counterparts (S1)–(S5). u(k) depends continuously on f .1) where a := min fj . for every ﬁnite k. for all i ∈ J and k ∈ IN0 . DISCRETE DIFFUSION FILTERING (a) Although the basic idea behind scalespaces is to have a continuous scale parameter. (4. Moreover. 223]. 4. and the convergence to a constant steadystate utilizes (D2)–(D5). j∈J (4. Moreover. by means of (D1) we know that. let us now prove a maximum–minimum principle.2 Theoretical results It is obvious that for a ﬁxed ﬁlter belonging to the class (Pd ) every initial image f ∈ IRN generates a unique sequence (u(k) )k∈IN0 . and average grey value invariance is based on (D2) and (D3). Proposition 3 (Extremum principle). (c) Nonnegative matrices Q = (qij ) ∈ IRN ×N satisfying j∈J qij = 1 for all i ∈ J are also called stochastic matrices. then Q is doubly stochastic.98 Remarks: CHAPTER 4. in practice. Indeed. the following inequalities hold: (i) ui (k+1) = j∈J qij (u(k) )uj (k) (D4) ≤ max u(k) m m∈J j∈J qij (u(k) ) = max u(k) . This indicates that our discrete diﬀusion processes are related to the theory of Markov chains [370. m m∈J (D3) .2) (4. j∈J Proof: The maximum–minimum principle follows directly from the fact that. Therefore. it is evident that fully discrete results are of importance since. (D1) immediately gives wellposedness results. strict Lyapunov sequences need (D5) and (D6) in addition to (D2)–(D4). Then. (k) a ≤ ui ≤ b ∀ i ∈ J. if Q is stochastic and i∈J qij = 1 for all j ∈ J. Let f ∈ IRN and let (u(k) )k∈IN0 be the sequence of ﬁltered images according to (Pd ). ∀ k ∈ IN0 . The existence of Lyapunov sequences is a consequence of (D2)–(D4).
b. (4. 2 (k) which proves the proposition. m m∈J (D3) = j∈J qij (u(k) )uj (k) (D4) ≥ min u(k) m m∈J j∈J 2 4. Then the following properties are fulﬁlled: (a) (Lyapunov sequences) For all convex r ∈ C[a. As one might expect. The average grey level 1 µ := fj N j∈J is not aﬀected by the discrete diﬀusion ﬁlter: 1 N uj = µ j∈J (k) (4. However. the class (Pd ) allows an interpretation as a transformation which is smoothing in terms of Lyapunov sequences.3). µ)⊤ ∈ IRN . These functions ensure that u(k) converges to a constant image as k → ∞. we need less regularity than in the semidiscrete case: The convex function r. needs only to be continuous. let a. Assume that (u(k) )k∈IN0 satisﬁes the requirements of (Pd ). ui i∈J (k+1) = i∈J j∈J qij (u(k) )uj (k) = j∈J i∈J qij (u(k) ) uj (k) = j∈J uj . and µ be deﬁned as in (4. for all k ∈ IN0 . . This socalled redistribution property [164] ensures that.3.3 Scalespace properties All statements from Chapter 3 with respect to invariances are valid in the discrete framework as well. µ.2).4. SCALESPACE PROPERTIES (ii) ui (k+1) 99 qij (u(k) ) = min u(k) .4). Below we focus on proving average grey level invariance. but no more diﬀerentiable. Proposition 4 (Conservation of average grey value).. b] the sequence V (k) := Φ(u(k) ) := i∈J r(ui ).5) Proof: By virtue of (D2) and (D3) we have i∈J qij (u(k) ) = 1 for all j ∈ J and k ∈ IN0 . (k) k ∈ IN0 is a Lyapunov sequence: . and (4. which generates the Lyapunov sequences. and let c := (µ. respectively. Theorem 7 (Lyapunov sequences and behaviour for k → ∞). (4...4) ∀ k ∈ IN0 .
7) N qij (u(k) ) r(uj ) − r(ui ) (k) aij (u(k) ) r(uj ) i=1 j=1 N N (D3) i=1 j=1 aij (u(k) ) r(uj ) − r(ui ) (k) (k) . the preceding deﬁnition. j ∈ J we deﬁne aij (u(k) ) := qij (u(k) ) − 1 qij (u(k) ) (i = j) (i = j). and the prerequisites (D2) and (D3) we obtain N V (k+1) − V (k) = i=1 conv.6) (k) r(uj ) = Φ(u(k) ). DISCRETE DIFFUSION FILTERING Φ(u(k) ) ≥ Φ(c) ∀ k ∈ IN0 V (k+1) − V (k) ≤ 0 ∀ k ∈ IN0 Moreover. if r is strictly convex. r N j=1 N j=1 N i=1 N (k) (k) qij (u(k) ) uj − r(ui ) (k) (k) ≤ = = (4. then V (k) = Φ(u(k) ) is a strict Lyapunov sequence: (iii) (iv) Φ(u(k) ) = Φ(c) V (k+1) − V (k) = 0 ⇐⇒ u(k) = c u(k) = c ⇐⇒ (b) (Convergence) lim u(k) = c. (ii) For i. (4.7) Using the convexity of r.100 (i) (ii) CHAPTER 4. k→∞ Proof: (a) (i) Average grey level invariance and the convexity of r give N Φ(c) = i=1 N ≤ = i=1 N j=1 r 1 N 1 (k) uj j=1 N N (k) j=1 N r(uj ) (4.
.. (iv) In order to verify the ﬁrst implication. m := lp0 +1 . by the irreducibility of Q(u(k) ). = uN .3. SCALESPACE PROPERTIES = + (D2) N N −i i=1 m=1 N N −i i=1 m=1 101 ai+m. we ﬁnd l0 ..m qnj (u(k) ) r(uj ) + qnn (u(k) ) r(u(k)) + qnm (u(k) ) r(u(k)) n m N = j=1 qnj (u(k) ) r(uj )..m qnj (u(k) ) uj + qnn (u(k) ) u(k) + qnm (u(k) ) u(k) n m (k) < j=1 j=n. assume that u(k) is not constant: ui0 := min ui i∈J (k) (k) < max uj j∈J (k) =: uj0 . .. Theorem 5): Equality in the estimate (4. r − 1} such that n := lp0 .. there exists some p0 ∈ {0. Hence. Moreover.8) (k) (k) (k) (k) = 0... r − 1.4. lr = j0 and qlp lp+1 = 0 for p = 0. . (k) Then.. (k) If we combine this with the results in (4.8).i+m (u(k) ) r(ui+m ) − r(ui ) (4. we obtain V (k+1) − V (k) N = i=1 i=n r N N j=1 + r (k) qij (u(k) ) uj − j=1 (k) qnj (u(k) ) uj (k) r(ui ) − r(u(k)) n . . Together with the strict convexity of r these properties lead to r = N (k) qnj (u(k) ) uj N (k) j=1 r N j=1 j=n. m n and by (D6) we have qnn (u(k) ) > 0... the nonnegativity of Q(u(k) ) gives qnm (u(k) ) > 0. To this end. qnm (u(k) ) = 0.i (u(k) ) r(ui ) − r(ui+m ) ai.6) holds due to the strict convexity of r if and only if u(k) = c. let us start with a proof that (k) (k) V (k+1) = V (k) implies u1 = . (iii) This part of the proof can be shown in exactly the same manner as in the semidiscrete case (Chapter 3. lr ∈ J with l0 = i0 . and u(k) = u(k) .
(k) (k) for all n ∈ IN. Conversely. j=1 uj An interpretation of these results in terms of decreasing energy.1 Relation to semidiscrete models Semiimplicit schemes Let us now investigate in which sense our discrete ﬁlter class covers in a natural way time discretizations of semidiscrete ﬁlters. Then. See Chapter 3. 4.4 4.. by virtue of the grey value invariance. decreasing central moments and increasing entropy is evident. let u(k) = c. we regard u(k) as an . By means of prerequisite (D3) we obtain N V (k+1) − V (k) = i=1 r N j=1 qij (u(k) )µ − N r(µ) = 0. Let (u(k) )k∈IN0 be a diﬀusion sequence according to (Pd ). i=1 (c) In order to prove convergence to a constant steadystate. 2 In analogy to the semidiscrete case the preceding theorem comprises many Lyapunov functions which demonstrate the informationreducing qualities of our ﬁlter class.. DISCRETE DIFFUSION FILTERING N < i=1 (4. if a > 0. 1 N N j=1 (k) M2n [u(k) ] := H[u(k) ] := N (uj − µ)2n ln(uj ). and let a and µ be deﬁned as in (4. Theorem 5 for more details.4). Then the following functions are decreasing in k: (a) (b) (c) u(k) p for all p ≥ 1. Corollary 3 (Special Lyapunov sequences). we can argue exactly in the same way as in the semidiscrete case if we replace Lyapunov functions by Lyapunov sequences and integrals by sums. To this end.4. Choosing the convex functions r(s) := sp . we immediately obtain the following corollary. = uN . r(s) := (s − µ)2n and r(s) := s ln s.2) and (4. qij (u(k) ) r(uj ) − r(ui ) (k) (k) (k) (k) This establishes that V (k+1) = V (k) implies u1 = . we conclude that u(k) = c.102 CHAPTER 4.8) N j=1 = 0.
where τ denotes the time step size. while the explicit evaluation of the nonlinear terms avoids the necessity to solve nonlinear systems of equations. We consider a ﬁnite diﬀerence scheme with two time levels where the operator A – which depends nonlinearly on u – is evaluated in an explicit way.10) for α ∈ (0. . for which we have established scalespace results. The theorem below states that this class of schemes is covered by the discrete framework. and let A = (aij ) : IRN → IRN ×N satisfy the requirements (S1)–(S5) of section 3. 1). the argument u(k) is suppressed frequently since the considerations below are valid for all u(k) ∈ IRN . C(u(k) ) := (cij (u(k) )) := I + (1−α)τ A(u(k) ). If α = 0. Then B is strictly diagonally dominant. RELATION TO SEMIDISCRETE MODELS 103 approximation of the solution u of (Ps ) at time t = kτ . since bii = 1 − ατ aii = 1 + ατ (S3) aij > ατ j∈J j=i j∈J j=i aij = (S4) j∈J j=i bij  ∀ i ∈ J. Theorem 8 (Scalespace interpretation for αsemiimplicit schemes). Then the αsemiimplicit scheme u(k+1) − u(k) = A(u(k) ) αu(k+1) + (1−α)u(k) τ fulﬁls the prerequisites (D1)–(D6) for discrete diﬀusion models provided that τ ≤ 1 (1−α) max aii (u(k) ) i∈J (4. Such schemes are called αsemiimplicit. where I ∈ IRN denotes the unit matrix. Proof: Let B(u(k) ) := (bij (u(k) )) := I − ατ A(u(k) ). 1].11) and the semiimplicit case (α = 1) satisﬁes (D1)–(D6) unconditionally. while the linear remainder is discretized in an αimplicit manner.4. Let α ∈ [0.9) can be written as B(u(k) ) u(k+1) = C(u(k) ) u(k) we ﬁrst have to show that B(u(k) ) is invertible for all u(k) ∈ IRN . max aii (u(k) ) i∈J (4.1. then B = I and hence invertible. They reveal the advantage that the linear implicit part ensures good stability properties. Henceforth.9) (4. Since (4. In the explicit case (α = 0) the properties (D1)–(D6) hold for τ < 1 . Now assume that α > 0. τ > 0.4.
13) (4.14) (4. Then (4. .15) . p. Thanks to the fact that B is irreducibly diagonally dominant. however. we know from [407. IRN ×N ). and by the structure of the oﬀdiagonal elements of B we observe that the irreducibility of A implies the irreducibility of B. By means of (S3) we obtain bij = 1 = j∈J j∈J cij ∀ i ∈ J. nonnegativity of C is not automatically guaranteed: Using (S3)–(S5) we obtain aii = − Hence. and the invertibility of B gives v = B −1 v = Hv.12) is equivalent to Bv = v = Cv. Therefore. Let us now verify (D3). from Qv = HCv we conclude that j∈J (4. we ﬁrst check the nonnegativity of C. Q := (qij ) := B −1 C exists and by (S1) it follows that Q ∈ C(IRN . In order to show that (D4) is fulﬁlled. For 0 ≤ α < 1. The diagonal entries yield cii = 1 + (1−α)τ aii . The requirement (D2) is not hard to satisfy: Since B −1 and C are symmetric and reveal the same set of eigenvectors – namely those of A – it follows that Q = B −1 C is symmetric as well. Thus. For i = j we have cij = (1−α)τ aij ≥ 0. j ∈ J. bij ≤ 0 for all i = j. This proves (D1). (4.13) (4.104 CHAPTER 4.. C(u(k) ) is nonnegative if τ ≤ 1 =: τα (u(k) ). (4. 85] that B −1 =: H =: (hij ) exists and hij > 0 for all i.(S5) < 0 ∀ i ∈ J... DISCRETE DIFFUSION FILTERING This also shows that bii > 0 for all i ∈ J. and bii > 0 for all i ∈ J.12) Let v := (1. This proves (D3). If α = 1 we have cii = 1 for all i ∈ J. (k) ) (1−α) max aii (u i∈J (S3) (S4) aij j∈J j=i (S4).14) = Hv = v qij = 1 for all i ∈ J. 1)⊤ ∈ IRN .
IR Kf := max aii (v) i ∈ J. 2 Remarks: (a) We have seen that the discrete ﬁlter class (Pd ) – although at ﬁrst glance looking like a pure explicit discretization – covers the αsemiimplicit case as well. Thus. and (4. Thus. and (4. Using this information.15) shows that Kf > 0. j ∈ J. v ∞ ≤ f ∞ exists. Equation (4. then C = I. j ∈ J with i = j we know that aij (u(k) ) > 0 implies cij (u(k) ) > 0. Q is irreducible. for α = 0. Now for τ < 1 max aii (u(k) ) i∈J it follows that cii (u(k) ) > 0 for all i ∈ J and. RELATION TO SEMIDISCRETE MODELS 105 Since H is nonnegative.4. For i. By N N ×N ) it follows that A ∈ C(IR . Next let us consider the case 0 < α < 1 and τ ≤ τα (u(k) ).10) and (4.11) can be satisﬁed by means of an apriori estimate. j ∈ J. In all the abovementioned cases the time step size restrictions for ensuring irreducibility imply that all diagonal elements of Q(u(k) ) are positive. If α = 1. the irreducibility of A(u(k) ) carries over to Q(u(k) ). we have Q = C. Explicit twolevel schemes are comprised by the choice α = 0. the positivity of H. Finally. which establishes the irreducibility of Q. and by the positivity of H we have qij > 0 for all i. (b) The conditions (4. Then we know that C is nonnegative. This means that u(k) belongs to the compact set {v ∈ IRN v ∞ ≤ f ∞ } for all k ∈ IN0 . This establishes (D6). we know that the nonnegativity of C implies the nonnegativity of Q = HC.11) shows that they reveal the most prohibitive time step size restrictions. the symmetry of C.4.12) we obtain qij = k∈J hik ckj ≥ min hik · k∈J >0 ckj > 0 k∈J =1 ∀ i. v ∈ IRN . Since the semiimplicit scheme fulﬁls (D1)–(D6) we know by Theorem 3 that the solution obeys an extremum principle. Now we want to prove (D5). thus. choosing τ ≤ 1 (1−α) Kf .
the matrix reveals a much larger bandwidth. DISCRETE DIFFUSION FILTERING ensures that (4. which can become prohibitive for large images. and their convergence can be guaranteed for the special structure of A.2.11) holds. diagonally dominant. and they inherit the scalespace properties from the semidiscrete setting regardless of the step size. However. i−2p2 −2p] to [i. a large linear system of equations has to be solved.106 CHAPTER 4. One should not expect. that in the ith row these entries can be found within the positions [i. i+2p2 +2p]. Classical methods such as Gauß–Seidel or SOR [447] are easy to code. Compared to explicit schemes. Its system matrix is symmetric. (e) By the explicit discretization of the nonlinear operator A it follows that all schemes in the preceding theorem are of ﬁrst order in time. This should not give rise to concern. Unfortunately. if one insists in secondorder schemes. however. and positive deﬁnite. since in image processing one is in general more interested in maintaining qualitative properties such as maximum principles or invariances rather than having an accurate approximation of the continuous equation. Faster iterative methods such as preconditioned conjugate gradient algorithms [348] need signiﬁcantly more storage. 179] are one possibility to circumvent these diﬃculties. (d) For α = 1 we obtain semiimplicit schemes which do not suﬀer from time step size restrictions. In general.4. another possibility will be studied in Section 4. they do not need additional storage. since this increases the condition number of the system matrix. if it results from a ﬁnite diﬀerence discretization on a (2p+1)×(2p+1)stencil it contains at most 4p2 +4p+1 nonvanishing entries per row. Applying standard direct algorithms such as Gaussian elimination would destroy the zeros within the band and would lead to an immense storage and computation eﬀort. this advantage usually overcompensates for the additional eﬀort of resolving a linear system. they converge rather slowly. Iterative algorithms appear to be better suited.10) is always satisﬁed. and τ < guarantees that (4. A typical problem of iterative methods is that their convergence slows down for larger τ . Usually. Multigrid methods [59. 1 Kf (c) If α > 0. In spite of the fact that the nonlinearity is discretized in an explicit way they are absolutely stable in the maximum norm. it is also sparse: For instance. Modiﬁcations in order to reduce these problems [122] are quite diﬃcult to implement. one may for .
2 2 τ This scheme satisﬁes the properties (D1)–(D6) for τ ≤ 2/Kf . The following theorem clariﬁes the conditions under which AOS schemes create discrete scalespaces. (4.4. τ /2 u(k+1) − u(k) = A(u(k+1/2) ) 1 u(k+1) + 1 u(k) . Since this can be numerically expensive.4..17) its additive operator splitting (AOS) variant [424] u(k+1) = 1 m m l=1 I − αmτ Al (u(k) ) −1 I + (1−α)mτ Al (u(k) ) u(k) . (f) The assumptions (S1)–(S5) are suﬃcient conditions for the αsemiimplicit scheme to fulﬁl (D1)–(D6).. Nonnegativity of Q(u(k) ) may also be achieved using spatial discretizations where A(u(k) ) has negative oﬀdiagonal elements (see [55] for examples)..m can be solved more eﬃciently. Suppose we know a splitting m A(u(k) ) = l=1 Al (u(k) ).18) By means of a Taylor expansion one can verify that.16) such that the m linear systems with system matrices (I −mατ Al (u(k) )). they have the same approximation order in space and time. (4. the preceding αsemiimplicit schemes require to solve a linear system with the system matrix (I − ατ A(u(k) )). but they are not necessary..4. from a numerical viewpoint. l = 1. Hence. . RELATION TO SEMIDISCRETE MODELS 107 instance use the predictor–corrector approach by Douglas and Jones [121]: u(k+1/2) − u(k) = A(u(k) ) u(k+1/2) . it would be nice to have an eﬃcient alternative.2 AOS schemes We have seen that. for α > 0. Then it is advantageous to study instead of the αsemiimplicit scheme m u (k+1) = I − ατ Al (u ) l=1 (k) −1 m I + (1−α)τ l=1 Al (u(k) ) u(k) (4. although both schemes are not identical. they are both consistent approximations to the semidiscrete ODE system from (Ps ). 4.
it follows that Q(u) := 1 m m l=1 Bl−1 (u) Cl (u) is also continuous in u. Moreover..108 CHAPTER 4. (S4) . Then the following holds: For α ∈ (0. Under these circumstances it follows from [284. For i = j we have cijl = (1−α)mτ aijl ≥ 0. (1−α) m max aiil (u(k) ) i.18) fulﬁls the prerequisites (D1)–(D6) of discrete diﬀusion scalespaces provided that τ < 1 .1.. The symmetry property (D2) of Q results directly from the fact that Bl−1 and Cl are symmetric and share their eigenvectors with those of Al . 192] that Hl := Bl−1 is nonnegative in all components. p. Since Al (u) is continuous in u by virtue of (S1). To verify (D4). In the same way as in the proof of Theorem 8 one shows that Bl−1 Cl has only unit row sums for all l. and bijl ≤ 0 for i = j. τ > 0. 1]. the row sums of Q are 1 as well. Thus. the AOS scheme (4. Let α ∈ (0. assume that A(u) = m Al (u) l=1 is irreducible for all u ∈ IRN .. Thus. Bl is invertible because of its strict diagonal dominance: biil = 1 + αmτ j∈J j=i Cl (u(k) ) := (cijl (u(k) ))ij := I + (1−α)mτ Al (u(k) ). we utilize that Bl is strictly diagonally dominant.. the properties (D1)–(D6) are unconditionally satisﬁed.l (4. (S3) aijl > αmτ j∈J j=i aijl = (S4) j∈J j=i bijl  ∀ i ∈ J. 1). and that for each Al there exists a permutation matrix Pl ∈ IRN ×N such that Pl Al PlT is block diagonal and irreducible within each block. and (D3) is satisﬁed. and let Al = (aijl )ij : IRN → IRN ×N . Let Bl (u(k) ) := (bijl (u(k) ))ij := I − αmτ Al (u(k) ).19) In the semiimplicit case (α = 1).m satisfy the requirements (S1)–(S4) of section 3. biil > 0 for all i. DISCRETE DIFFUSION FILTERING Theorem 9 (Scalespace interpretation for AOS schemes). l = 1. Proof: The reasoning is similar to the proof of Theorem 8. a suﬃcient condition for proving (D4) is to ensure that Cl is nonnegative for all l. This proves (D1).
2 Remarks: (a) In analogy to the unsplit αsemiimplicit schemes. −1 1 = m I − 1 mτ Al (u(k+1/2) ) 2 I + 1 mτ Al (u(k+1/2) ) u(k) . because no time step size restriction occurs. we show now that ai0 j0 l0 = 0 implies hi0 j0 l0 > 0. we know that aiil ≤ 0. p.l guarantees that ciil > 0 ∀ i ∈ J.15) we have aii < 0 for all i ∈ J. 109 If α = 1 we have cii = 1 for all i ∈ J. Suppose that ai0 j0 = 0 for some i0 . Thus. j0 ∈ J..15) it follows that for every i there exists an l with aiil < 0. m (l.4. Q must have a positive diagonal. ai0 j0 l0 = 0 implies hi0 j0 l0 > 0. ∀ l = 1. (4. Each block is irreducible and strictly diagonally dominant with a positive diagonal and nonpositive oﬀdiagonals. m} such that ai0 j0 l0 = 0.. Therefore. l = 1. and (D5) is proved. the irreducibility of A carries over to Q. however..j0 ) 0 ≥0 ≥0 >0 >0 (k) Recapitulating.. Together with (4. For 0 < α < 1. Again. (4. Then there exists an l0 ∈ {1. for all i. nonnegativity of C is not automatically guaranteed: Since Al satisﬁes (S3) and (S4).20) Next we prove (D5). m.. This can be seen as follows: There exist permutation matrices Pl .. Thus.4. it is also possible to construct a predictor–corrector scheme of Douglas–Jones type [121] within the AOS framework: u(k+1/2) = u(k+1) 1 m m l=1 m l=1 I − 1 mτ Al (u(k) ) 2 −1 u(k) . RELATION TO SEMIDISCRETE MODELS The diagonal entries yield ciil = 1 + (1−α)mτ aiil . ai0 j0 = 0 =⇒ qi0 j0 > 0. 2 .. .n)=(l0 ..21) Thus. .20) this yields qi0 j0 = 1 hi nl cnj0 l + hi0 j0 l0 cj0 j0 l0 > 0. Moreover. . for τ < τα (u ).21) also proves (D6): By virtue of (4. the case α = 1 is especially interesting. the irreducibility of Q. requiring 1 τ < =: τα (u(k) ) (1−α) m max aiil (u(k) ) i. As a consequence. Moreover. a theorem by Varga [407. m such that Pl Bl PlT is block diagonal. (4.. by (4. this means that. Denoting Bl−1 by Hl = (hijl )ij . 85] ensures that the inverse of each block contains only positive elements.
m do not commute. (b) The fact that AOS schemes use an additive splitting ensures that all coordinate axes are treated in exactly the same manner. In practice.2 we have seen that this ﬁlter is based on the PDE ∂t u = div (g(∇uσ 2 ) ∇u). m}.. it is possible to distribute their computation to diﬀerent processors of a parallel machine. v ∞ ≤ f ∞ . the AOS–Douglas–Jones scheme is only ﬁrst order accurate in time.. In e Section 1.m can be calculated independently from each other.110 CHAPTER 4.. this means that these schemes produce diﬀerent results if the image is rotated by 90 degrees. where Kf is determined by the apriori estimate Kf := m max aiil (v) i ∈ J. (c) The result u(k+1) of an AOS scheme can be regarded as the average of m ﬁlters of type vl (k+1) := I − αmτ Al (u(k) ) (k+1) −1 I + (1−α)mτ Al (u(k) ) u(k) (l = 1.. 277. 430] as eﬃcient discretizations of the isotropic nonlinear diﬀusion ﬁlter of Catt´ et al. Hence. They are multiplicative splittings.. Moreover...... (d) AOS schemes with α = 1 have been presented in [424. l = 1. This violates requirement (D2) for discrete scalespaces.... [81]. m). v ∈ IRN . DISCRETE DIFFUSION FILTERING It satisﬁes (D1)–(D6) for τ < 2/Kf . Since vl . the corresponding linear systems can be .. This separation is very eﬃcient: There exist permutation matrices Pl (pixel orderings) such that Pl Al PlT is block diagonal and each block is diagonally dominant and tridiagonal. a natural operator splitting A = m Al results from a decoml=1 position of the divergence expression into onedimensional terms of type ∂xl (g(∇uσ 2 ) ∂xl u) (l = 1.3.. most multiplicative splittings lead to a nonsymmetric matrix Q(u(k) ).. . l ∈ {1. 354. the result of multiplicative splittings will depend on the order of the operators. However. m). This is in contrast to the various conventional splitting techniques from the literature [120. Since in the general nonlinear case the split operators Al .. . 442]. . l = 1. A typical representative is m u (k+1) = l=1 I − τ Al (u(k) ) −1 u(k) . In this case. 286.
The denoising of a medical 3D ultrasound data set with 138 × 208 × 138 voxels on an SGI Power Challenge XL with eight 195 MHz R10000 processors was possible in less than 1 minute. .2.4.4. The resulting forward substitution and backward elimination can be regarded as a recursive ﬁlter. RELATION TO SEMIDISCRETE MODELS 111 solved in linear eﬀort by means of a simple Gaussian algorithm. (e) The idea to base AOS schemes on decompositions into onedimensional operators can also be generalized to anisotropic diﬀusion ﬁlters: Consider for instance the discretization on a (3 × 3)stencil at the end of Section 3. then a splitting of such a 2D ﬁlter into 4 onedimensional diﬀusion processes acting along the 4 stencil directions satisﬁes all prerequisites of Theorem 9. A parallel implementation assigning these tridiagonal subsystems to diﬀerent processors is described in [431].4. If it fulﬁls (S1)–(S5).
DISCRETE DIFFUSION FILTERING .112 CHAPTER 4.
and dilation with a quadratic structuring function is performed in a noniterative way using van den Boomgaard’s algorithm from [51]. 1.4) with ∆t := 0. and MCM.01. respectively. All calculations for diﬀusion ﬁltering are performed using semiimplicit FD schemes with time steps ∆t ∈ [2. we shall pursue two diﬀerent objectives: • smoothing with simultaneous edgeenhancement. For MCM and AMSS this is achieved by means of explicit FD schemes (cf.Chapter 5 Examples and applications The scalespace theory from Chapters 2–4 also covers methods such as linear or nonlinear isotropic diﬀusion ﬁltering.5.3 seconds. 5]. for which many interesting applications have already been mentioned in Chapter 1. the goal of the present chapter is to show that a generalization to anisotropic models with diﬀusion tensors depending on the structure tensor oﬀers novel properties and application ﬁelds. Typical dilations with a quadratic structuring function take less than 0. Therefore. we focus mainly on these anisotropic techniques and juxtapose the results to other methods. 113 . • smoothing with enhancement of coherent ﬂowlike textures. AMSS or dilation with a disc require approximately 0. Dilation with a ﬂat structuring element is approximated by an OsherSethian scheme of type (1.06 seconds per iteration. On an HP 9000/889 workstation it takes less than 0.88) with ∆t := 0. In order to compare anisotropic diﬀusion to other methods.3 CPU seconds to calculate one nonlinear diﬀusion step for a 256 × 256 image. morphological scalespaces and modiﬁcations of them have been discretized as well.1 and ∆t := 0.6. Thus. In order to demonstrate the ﬂexibility of anisotropic diﬀusion ﬁltering.
1) (5. The constant Cm is calculated in such a way that the ﬂux Φ(s) = sg(s) is increasing for s ∈ [0. Since in this section we are only interested in edgeenhancing diﬀusion we may set the integration scale ρ of the structure tensor equal to 0.2. Applications which require nonvanishing integration scales shall be studied in section 5. This behaviour may be accomplished by the following choice (m ∈ IN. If one wants to smooth preferably within each region and aims to inhibit diffusion across edges.3. λ2 := 1 with g(s) := 1 (5. (5. see 1.3 and [415].2) This exponentially decreasing function is chosen in order to fulﬁl the smoothness requirement stated in (Pc ). Cm > 0. v2 the corresponding orthonormal eigenvectors. Since the diﬀusion tensor should reﬂect the local image structure it ought to be chosen in such a way that it reveals the same set of eigenvectors v1 .1. the preceding ﬁlter strategy can be regarded as an anisotropic regularization of the Perona–Malik model.114 CHAPTER 5.1 illustrates that anisotropic diﬀusion ﬁltering is still capable of possessing the contrastenhancing properties of the Perona–Malik ﬁlter (provided the regularization parameter σ is not too large). ∞) and µ1 = ∇uσ 2 . µ2 with µ1 ≥ µ2 be the eigenvalues of the structure tensor Jρ .2.1 Edgeenhancing diﬀusion Filter design In accordance with the notations in 2. we know that the uniform positive deﬁniteness of D is automatically satisﬁed by this ﬁlter. cf. The choice m := 4 (which implies C4 = 3. λ] and decreasing for s ∈ (λ. let µ1 . Thus. 2.1 5. and v1 .3. λ > 0): λ1 (µ1 ) := g(µ1). EXAMPLES AND APPLICATIONS 5. Since ∇uσ remains bounded on Ω × [0.3) 5. 1 − exp −Cm (s/λ)m (s ≤ 0) (s > 0).31488) gives visually good results and is used exclusively in the examples below. ∞). It depicts the temporal evolution of a .1. v2 as Jρ . then one can reduce the diﬀusivity λ1 perpendicular to edges the more the higher the contrast µ1 is. The choice of the corresponding eigenvalues λ1 . λ2 depends on the desired goal of the ﬁlter.2 Applications Figure 5.
5.1. EDGEENHANCING DIFFUSION
115
Figure 5.1: Anisotropic diﬀusion ﬁltering of a Gaussiantype function, Ω = (0, 256)2 , λ = 3.6, σ = 2. From top left to bottom right: t = 0, 125, 625, 3125, 15625, 78125.
116
CHAPTER 5. EXAMPLES AND APPLICATIONS
Gaussianlike function and its isolines.1 It can be observed that two regions with almost constant grey value evolve which are separated by a fairly steep edge. Edge enhancement is caused by the fact that, due to the rapidly decreasing diﬀusivity, smoothing within each region is strongly preferred to diﬀusion between the two adjacent regions. The edge location remains stable over a very long time interval. This indicates that, in practice, the determination of a suitable stopping time is not a critical problem. After the process of contrast enhancement is concluded, the steepness of edges decreases very slowly until the gradient reaches a value where no backward diﬀusion is possible anymore. Then the image converges quickly towards a constant image. Let us now compare the denoising properties of diﬀerent diﬀusion ﬁlters. Figure 5.2(a) consists of a triangle and a rectangle with 70 % of all pixels being completely degraded by noise. This image is taken from the software package MegaWave. Test images of this type have been used to study the behaviour of ﬁlters such as [13, 15, 16, 99, 102]. In Fig. 5.2(b) we observe that linear diﬀusion ﬁltering is capable of removing all noise, but we have to pay a price: the image becomes completely blurred. Besides the fact that edges get smoothed so that they are harder to identify, the correspondence problem appears: edges become dislocated. Thus, once they are identiﬁed at a coarse scale, they have to be traced back in order to ﬁnd their true location, a theoretically and practically rather diﬃcult problem. Fig. 5.2(c) shows the eﬀect when applying the isotropic nonlinear diﬀusion equation [81] ∂t u = div (g(∇uσ 2 )∇u) (5.4)
with g as in (5.3). Since edges are hardly aﬀected by this process, nonlinear isotropic diﬀusion does not lead to correspondence problems which are characteristic for linear ﬁltering. On the other hand, the drastically reduced diﬀusivity at edges is also responsible for the drawback that noise at edges is preserved. Figure 5.2(d) demonstrates that nonlinear anisotropic ﬁltering shares the advantages of both methods. It combines the good noise eliminating properties of linear diﬀusion with the stable edge structure of nonlinear isotropic ﬁltering. Due to the permitted smoothing along edges, however, corners get more rounded than in the nonlinear isotropic case. The scalespace behaviour of diﬀerent PDEbased methods is juxtaposed in Figures 5.3–5.6, where an MRI slice of a human head is processed [414, 423].
Except for Figs. 5.1, 5.3–5.6, where contrast enhancement is to be demonstrated, all images in the present work are depicted in such a way that the lowest value is black and the highest one appears white. They reveal a range within the interval [0, 255] and all pixels have unit length in both directions.
1
5.1. EDGEENHANCING DIFFUSION
117
Figure 5.2: Restoration properties of diﬀusion ﬁlters. (a) Top Left: Test image, Ω = (0, 128)2 . (b) Top Right: Linear diﬀusion, t = 80. (c) Bottom Left: Nonlinear isotropic diﬀusion, λ = 3.5, σ = 3, t = 80. (d) Bottom Right: Nonlinear anisotropic diﬀusion, λ = 3.5, σ = 3, t = 80.
Again we observe that linear diﬀusion (Fig. 5.3(a)) does not only blur all structures in an equal amount but also dislocates them more and more with increasing scale. A ﬁrst step to reduce these problems is to adapt the diﬀusivity to the gradient of the initial image f [147]. Fig. 5.3(b) shows the evolution under
∂t u = div (g(∇f 2) ∇u),
(5.5)
3) is applied. the temporal evolution of these models generates a .2(d). At the chin we observe that this equation is indeed capable of enhancing edges. (5. 5. Fig.4).7) Figure 5. Thus. becomes smaller. also small structures exist over a long range of scales if they diﬀer from their vicinity by a suﬃciently large contrast. On the other hand. The latter problem can be avoided using a diﬀusivity which decreases faster than (5. All structures are extremely welllocalized and the results are segmentationlike. but this also leads to stronger blurring of large structures. As in Figure 5. (5. Finally the image turns into a silhouette of the head. The tendency to produce piecewise almost constant regions indicates that diffusion scalespaces with nonmonotone ﬂux are ideal preprocessing tools for segmentation.4(b)). One can try to make this ﬁlter faster and more insensitive to smallsize structures by increasing the regularizing Gaussian kernel size σ (cf. This leads to the nonlinear diﬀusion equation [326] ∂t u = div (g(∇u2) ∇u). The absolute contrast at edges. A positive consequence of this slight shrinking eﬀect is the fact that small or elongated and thin structures are better eliminated than in the isotropic case. Also blurring at edges is reduced very much. 5. 1. this causes a stronger rounding of structures. EXAMPLES AND APPLICATIONS where a diﬀusivity of type [88] g(∇f 2) := 1 1 + ∇f 2 /λ2 (λ > 0). which can be seen at the nose. we recognize that most of the depicted “segments” coincide with semantically correct objects that one would expect at these scales.4(a) where the regularized isotropic nonlinear diﬀusion ﬁlter (5. for large t the ﬁltered image reveals some artifacts which reﬂect the diﬀerential structure of the initial image. A natural idea to reduce the artifacts of inhomogeneous linear diﬀusion ﬁltering would be to introduce a feedback in the process by adapting the diﬀusivity g to the gradient of the actual image u(x. and it is no longer possible to enhance the contour of the entire head. Compared with homogeneous linear diﬀusion. Unlike diﬀusion–reaction models aiming to yield one segmentationlike result for t → ∞ (cf.6) and leads to a nonmonotone ﬂux function.118 CHAPTER 5.4(c)) permits diﬀusion along edges and inhibits smoothing across them. before it converges to a constant image.6) is used. edges remain better localized and their blurring is reduced. Anisotropic nonlinear diﬀusion (Fig. This is illustrated in Figure 5. On the other hand.4) with the diﬀusivity (5.3(c) shows how such a nonlinear feedback is useful to increase the edge localization in a signiﬁcant way: Structures remain welllocalized as long as they can be recognized. t) instead of the original image f (x). however.
In this sense. This is not true for the morphological ﬁlters and their modiﬁcations which are depicted in Fig.3 and 1. and maxima keep their location in scalespace until they disappear [206. Besides these speciﬁc features of nonlinear diﬀusion scalespaces it should be mentioned that. 5. The traditional opinion that the evolution parameter t of scalespaces should be related to the spatial scale reﬂects the assumption that a scalespace analysis should be uncommitted. Suitable values for them should result in a natural way from the speciﬁc problem. which gives a hierarchy of structures and allows to extract the relevant information from a certain scale.5 and 5. Nonlinear diﬀusion ﬁltering renounces this requirement by allowing to incorporate apriori information (e. The contrastenhancing quality distinguishes nonlinear diﬀusion ﬁlters from most other scalespaces.5. and by ∂t u = ∇u2 (5. The basic idea of scalespaces.5. with respect to the speciﬁed task. Figure 5. than a descriptor of spatial scale. respectively. and that they usually do not preserve the average grey level. for the disc. Therefore. due to the homogeneous Neumann boundary condition and the divergence form. is maintained: to provide a family of subsequently simpliﬁed versions of the original image.g.6. From Section 1. 207].5. The fact that the maximum with the largest grey value will dominate at the end shows that these processes can be sensitive to noise (maxima might be caused by noise). the time t is rather a parameter of importance. it is generally not possible to obtain similar segmentationlike results by just rescaling the grey values from a scalespace which is only contrastreducing. about the contrast of semantically important structures) into the evolution process. It is not very diﬃcult to guess the shape of the structuring function from the scalespace evolution. (5.1. In both cases the number of local maxima is decreasing.9) for the quadratic structuring function.5(a) and (b) show the result under continuousscale dilation with a ﬂat discshaped structuring element and a quadratic structuring function. both linear and nonlinear diﬀusion ﬁlters preserve the average grey level of the image. EDGEENHANCING DIFFUSION 119 complete hierarchical family of segmentationlike images.6 we know that their evolution equations are given by ∂t u = ∇u. It should be noted that contrast enhancement is a local phenomenon which cannot be replaced by simple global rescalings of the grey value range. The contrast and noise parameters λ and σ give the user the liberty to adapt nonlinear diﬀusion scalespaces to the desired purpose in order to reward interesting features with a longer lifetime. however.8) .
(c) Right Column: Nonlinear isotropic diﬀusion with the Charbonnier diﬀusivity (λ = 3). top to bottom: t = 0. 150.120 CHAPTER 5. t = 0.5. 50. EXAMPLES AND APPLICATIONS Figure 5. (a) Left Column: Linear diﬀusion. 70. 200. 236)2. 600. 70. 200. (b) Middle Column: Inhomogeneous linear diﬀusion (λ = 8).3: Evolution of an MRI slice under diﬀerent PDEs. 400. Top: Original image. . 12. Ω = (0. t = 0.
σ = 1). (a) Left Column: Isotropic nonlinear diﬀusion (λ = 3. 25000. 3000. Ω = (0.4: Evolution of an MRI slice under diﬀerent PDEs. (c) Right Column: Edgeenhancing anisotropic diﬀusion (λ = 3. t = 0.5. 40. t = 0. 7000000. 236)2. 875. . 400. EDGEENHANCING DIFFUSION 121 Figure 5.1. σ = 4). 250. t = 0. (b) Middle Column: Isotropic nonlinear diﬀusion (λ = 3. 1500. σ = 1). Top: Original image. 500000.
5: Evolution of an MRI slice under diﬀerent PDEs. Ω = (0.25. (c) Right Column: Mean curvature motion.122 CHAPTER 5. Top: Original image. 0. t = 0. . 236)2. 4. 10. 4. EXAMPLES AND APPLICATIONS Figure 5. t = 0. (a) Left Column: Dilation with a disc. t = 0. 1. (b) Middle Column: Dilation with a quadratic structuring function. 20. 275. 1275. 70.
. 140. t = 0. t = 0. σ = 1). 5000. 50. σ = 1). (b) Middle Column: Modiﬁed mean curvature motion (λ = 3. Top: Original image. Ω = (0. (a) Left Column: Aﬃne morphological scalespace. 600. 350.1. 1500. t = 0. 40000.6: Evolution of an MRI slice under diﬀerent PDEs.5. (c) Right Column: Selfsnakes (λ = 3. 20. EDGEENHANCING DIFFUSION 123 Figure 5. 236)2. 100.
Thus. the nose) exhibit roundings at an earlier stage. 5. 364. the experiments give evidence that this process is probably not contrastenhancing.6(b). Since they diﬀer from isotropic nonlinear diﬀusion ﬁlters by the ∇u terms inside and outside the divergence expression.108) shown in Fig.5(c). After some time. Since it takes the time T = 3 s 3 to remove 4 1 all isolines within a circle of radius s – in contrast to T = 2 s2 for MCM – we see that. Since MCM shrinks level lines with a velocity that is proportional to their curvature. On the other hand. Nevertheless. let us investigate the modiﬁed MCM ∂t u = g(∇uσ 2 ) ∇u div ∇u ∇u (5.g. the correspondence problem is more severe than for MCM.g. lowcurved object boundaries are less aﬀected by this process. One possibility to reduce the correspondence problem of morphological scalespaces is to attenuate the curve evolution at highcontrast edges. As a consequence.124 CHAPTER 5. Processes of this type are studied in [13.11) In Section 1.6 we have seen that processes of this type are called selfsnakes [357]. it is ideal for uncommitted image analysis and shape recognition. The corresponding evolution is depicted in Fig. the chin. This is in accordance with the theory which predicts convergence of all closed level lines to circular points.32). but they are not designed to act contrastenhancing. 5. A similar behaviour can be observed for the aﬃne invariant morphological scale4 space (1. while highcurved structures (e. see e. 5. the advantage of having aﬃne invariance may counterbalance the correspondence problem in certain applications.90) depicted in Fig. Both MCM and AMSS give signiﬁcantly sharper edges than linear diﬀusion ﬁltering.6(a). 304]. We observe that structures remain much better localized than in the original MCM. This is at the expense of withdrawing morphology in terms of invariance under monotone greyscale transformations. the head looks almost like a ball. Since the AMSS involves no additional parameters and oﬀers more invariances than other scalespaces. 365.10) with g(∇uσ 2 ) as in (1. EXAMPLES AND APPLICATIONS A completely diﬀerent morphological evolution is given by the mean curvature motion (1. One possibility is to use the damping function outside the divergence expression. the results appear less segmentationlike than those for nonlinear diﬀusion ﬁltering. the shrinking eﬀect of large structures is stronger for AMSS than for MCM. This also explains its excellent noise elimination qualities. As a simple prototype for this idea. however. ∇u (5. Using g(∇uσ 2 ) inside the divergence expression leads to ∂t u = ∇u div g(∇uσ 2 ) ∇u .6. they will not preserve the average . for a comparable elimination of small structure.
However. The quality of a fabric is determined by two criteria. For a suitable parameter choice. namely clouds and stripes. For furniture production it is of importance to classify the quality of wood surfaces. 257)2. 5. one has to process the image in such a way that quality relevant features become better visible und unimportant structures disappear. .4) can be applied with good success (Fig. 5. Fig. 5. (a) Left: Fabric. This behaviour which resembles MCM is not surprising if one compares (1.120) with (1.6(c) indicates that g(∇uσ 2 ) gives similar edgeenhancing eﬀects as in a nonlinear diﬀusion ﬁlter.121). t = 240. whereas stripes are an anisotropic phenomenon caused by adjacent ﬁbres pointing in the same direction. whereas stripes can be enhanced by a special nonlinear diﬀusion ﬁlter which is designed for closing interrupted lines and which shall be discussed in Section 5.5. Clouds result from isotropic inhomogeneities of the density distribution.2.8(b)). equation (5.1. If one aims to automize this evaluation.7: Preprocessing of a fabric image. To visualize this defect. grey value. but one can observe a stronger tendency to create circular structures. λ = 4. Anisotropic diﬀusion ﬁlters are capable of visualizing both qualityrelevant features simultaneously (Fig. (b) Right: Anisotropic diﬀusion. EDGEENHANCING DIFFUSION 125 Figure 5.7). they perform isotropic smoothing at clouds and diﬀuse in an anisotropic way along ﬁbres in order to enhance them. one can use a fast pyramid algorithm based on linear diﬀusion ﬁltering for the clouds [417]. Let us now study two applications of nonlinear diﬀusion ﬁltering in computer aided quality control (CAQ): the grading of fabrics and wood surfaces (see also [413]). if one wants to visualize both features separately. 5. σ = 2.8(a) depicts a wood surface possessing one defect. The evolution in Fig. Ω = (0.
In order to make the diagnosis more objective and reliable. . and the entire head area. t = 2000. one gets a better segmentation when processing the original image slightly by means of nonlinear diﬀusion ﬁltering (Fig. (5. K) = Ω (u−f )2 dx + αK.9(c) shows a segmentation according to the following simpliﬁcation of the Mumford– Shah functional (1. As is seen in Fig. 256)2.9(d).8: Defect detection in wood. In [413] it is demonstrated how a modiﬁed anisotropic diﬀusion process yields even more accurate results with less roundings at the corners. 5. Fig.12) It has been obtained by a MegaWave programme using a hierarchical region growing algorithm due to Koepﬂer et al. Ω = (0. It depicts an MRI slice of the human head.9(b)) prior to segmenting it. 5. Figure 5. (a) Left: Wood surface.58): Ef (u. EXAMPLES AND APPLICATIONS Figure 5. σ = 2. which are given by the two white longitudinal objects in the centre. For detecting Alzheimer’s disease one is interested in determining the ratio between the ventricle areas. λ = 4. it is intended to automize this feature extraction step by a segmentation algorithm. (b) Right: Isotropic nonlinear diﬀusion.9(a) gives an example for possible medical applications of nonlinear diffusion ﬁltering as a preprocessing tool for segmentation (see also [415] for another example). 5. [239].126 CHAPTER 5.
256)2. If one wants to enhance coherent structures. (b) Top Right: Diﬀusionﬁltered. µ2 with µ1 ≥ µ2 be the eigenvalues of Jρ . (d) Bottom Right: Segmented ﬁltered image.5. σ = 0. α = 8192. v2 the corresponding orthonormal eigenvectors.1 the diﬀusion tensor D(Jρ (∇uσ )) ought to possess the same set of eigenvectors as Jρ (∇uσ ).1. Ω = (0.2 5. and v1 .2. (a) Top Left: Head.1 Coherenceenhancing diﬀusion Filter design In this section we shall investigate how the structure tensor information can be used to design anisotropic diﬀusion scalespaces which enhance the coherence of ﬂowlike textures [418]. 5. λ = 5. one should smooth preferably . (c) Bottom Left: Segmented original image. Let again µ1 . This requires a nonvanishing integration scale ρ. α = 8192.5.2. COHERENCEENHANCING DIFFUSION 127 Figure 5.9: Preprocessing of an MRI slice. As in 5. t = 2.
ρ = 4. 256)2. along the coherence direction v2 with a diﬀusivity λ2 which increases with respect to the coherence (µ1 −µ2 )2 .5. m ∈ IN): λ1 := α. Ω = (0. σ = 0.5. EXAMPLES AND APPLICATIONS Figure 5. (d) Bottom Right: Structure tensor orientation. (c) Bottom Left: Orientation of smoothed gradient. λ2 := α where the exponential function was chosen to ensure the smoothness of D and the α + (1−α) exp −C (µ1−µ2 )2m if µ1 = µ2 . This may be achieved by the following choice for the eigenvalues of the diﬀusion tensor (C > 0. σ = 0. σ = 5. else.10: Local orientation in a ﬁngerprint image. (a) Top Left: Original ﬁngerprint.128 CHAPTER 5. (b) Top Right: Orientation of smoothed gradient. .
(b) Right: Coherenceenhancing anisotropic diﬀusion.1): ∂t u = uξξ = ∇u curv(u) (5. ﬁlters of this type are not regularizations of the Perona–Malik process: the limit σ → 0. 5.10(b)). the results in (c) are much worse than in (b). however. 5.001. In order to detect the local orientation of the ﬁngerprint depicted in Fig.2. Since MCM smoothes by propagating level lines in inner normal direction we recognize that its smoothing di2 Evidently.5. m := 1. t = 5.2 Applications Figure 5. σ = 0.10 illustrates the advantages of local orientation analysis by means of the structure tensor. We observe very high ﬂuctuations in the local orientation. COHERENCEENHANCING DIFFUSION 129 Figure 5. the gradient orientation of a slightly smoothed image has been calculated (Fig. ρ = 4. 5. Therefore. The structure tensor.5. and α := 0. Horizontally oriented structures appear black.2 All examples below are calculated using C := 1. When applying a larger smoothing kernel it is clear that adjacent gradients having the same orientation but opposite direction cancel out. 1. let us recall the example of mean curvature motion (cf.10(a).11: Anisotropic equations applied to the ﬁngerprint image. 1) keeps D(Jρ (∇uσ )) uniformly positive deﬁnite. . small positive parameter α ∈ (0. This is the reason for the reliable estimates of local orientation that can be obtained with this method (Fig.10(d)).6. while vertical structures are represented in white. t = 20. 5.2. ρ → 0 leads to a linear diﬀusion process with constant diﬀusivity α. (a) Left: Meancurvature motion.13) with ξ being the direction perpendicular to ∇u. To illustrate how the result of anisotropic PDE methods depends on the direction in which they smooth. averages the gradient orientation instead of its direction.
Thus. 5. although this method is in a complete anisotropic spirit. see Fig. (d) Bottom Right: t = 640. (a) Top Left: Original fabric image. EXAMPLES AND APPLICATIONS Figure 5. (c) Bottom Left: t = 120. biases the diﬀusive ﬂux towards the coherence orientation v2 and is therefore wellsuited for closing interrupted lines in coherent ﬂowlike textures. The results in Fig.11(a) conﬁrm this impression. Due to its reduced diﬀusivity at noncoherent structures.5.130 CHAPTER 5. 257)2. ρ = 2). Ω = (0. however. (b) Top Right: t = 20.12: Scalespace behaviour of coherenceenhancing diﬀusion (σ = 0.11(b). The proposed anisotropic diﬀusion ﬁlter. the locations of the semantically important singularities in the ﬁngerprint remain the same. we should not expect it to be capable of closing interrupted linelike structures. 5. rection depends exclusively on ∇u. . This is an important prerequisite that any image processing method has to satisfy if it is to be applied to ﬁngerprint analysis.
(c) we observe that CED is capable of enhancing the trabecular structures in order to ease their subsequent orientation analysis.13 illustrates the potential of CED for medical applications. t = 16. Figure 5. (c) Bottom Right: Dito with t = 128.12 depicts the scalespace behaviour of coherenceenhancing anisotropic diﬀusion applied to the fabric image from Fig. . the trabeculae.2. such as the placement or removal of metal implants.13: (a) Top: High resolution slipring CT scan of a femural bone. (b) Bottom Left: Filtered by coherenceenhancing anisotropic diﬀusion. The temporal behaviour of this diﬀusion ﬁlter seems to be appropriate for visualizing coherent ﬁbre agglomerations (stripes) at diﬀerent scales. Its internal structure has a distinctive texture through the presence of tiny elongated bony structural elements.7. quantifying the rate of progression of rheumatism and osteoporosis. a diﬃcult problem for the automatic grading of nonwovens [299]. 186). COHERENCEENHANCING DIFFUSION 131 Figure 5. 5. Ω = (0. 300) × (0. There is evidence that the trabecular formation is for a great deal determined by the external load. ρ = 6. Examples are the control of recovery after surgical procedures.13(b). For this reason the trabecular structure constitutes an important clinical parameter in orthopedics. It depicts a human bone.5. σ = 0. Figure 5. From Figure 5. the determination of leftright deviations of symmetry in the load or establishing optimal load corrections by physiotherapy.5.
In order to demonstrate the inﬂuence of the integration scale ρ. (a) Left: “Selfportrait” by van Gogh (SaintR´my. t = 6. Fig. all ﬁlter parameters are ﬁxed except for ρ. Interestingly. 215) × (0. In a last example the temporal evolution of ﬂowlike images is illustrated by virtue of the “Starry Night” painting in Fig. 5. 1889. ρ = 4. a further increasing of ρ does hardly alter this result (Fig. The next painting we are concerned with is called “Road with Cypress and Star” [162. 5. It is depicted in Fig. Ω = (0.15(c)). Let us now investigate the impact of coherenceenhancing diﬀusion on images. (b) Right: Filtered. Due to the established .5. 5. 5. Muse´ d’Orsay). which are not typical texture images. but still reveal a ﬂowlike character. which indicates that this van Gogh painting possesses a uniform “texture scale” reﬂecting the characteristic painting style of the artist. To this end.132 CHAPTER 5. 275).14: Image restoration using coherenceenhancing anisotropic diﬀusion.15(b) shows that a value for ρ which is too small does not lead to the visually dominant coherence orientation and creates structures with a lot of undesired ﬂuctuations. We observe that the diﬀusion ﬁlter can close interrupted lines and enhance the ﬂowlike character which is typical for van Gogh paintings.15.16 [160. Increasing the value for ρ improves the image signiﬁcantly (Fig. e Paris.14 shows the restoration properties of coherenceenhancing anisotropic diﬀusion when being applied to a selfportrait of the artist [161]. we shall process impressionistic paintings by Vincent van Gogh. e σ = 0. 5. 5. Fig. 419]. EXAMPLES AND APPLICATIONS Figure 5.15(d)). 429].
(d) Bottom Right: ρ = 6.15: Impact of the integration scale on coherenceenhancing anisotropic diﬀusion (σ = 0. t = 8).5. Rijksmuseum Kr¨ller–M¨ ller). 203) × (0. . (a) Top Left: “Road with Cypress and Star” by van Gogh (AuverssurOise. (c) Bottom Left: ρ = 4. Ω = (0. 1890. 290).2. COHERENCEENHANCING DIFFUSION 133 Figure 5. Otterlo. (b) Top Right: o u Filtered with ρ = 1.5.
The Museum of Modern Art). Ω = (0. (b) Top Right: t = 8. scalespace properties. 199). 255) × (0. e 1889. EXAMPLES AND APPLICATIONS Figure 5. (d) Bottom Right: t = 512. a constant image with the same average grey value as the original one. the image becomes gradually simpler in many aspects. New York. The ﬂowlike character. before it ﬁnally will tend to its simplest representation. is maintained for a very long time.5. however.16: Scalespace properties of coherenceenhancing anisotropic diﬀusion (σ = 0. ρ = 4).3 3 Results for AMSS ﬁltering of this image can be found in [305]. (a) Top Left: “Starry Night” by van Gogh (SaintR´my. (c) Bottom Left: t = 64. .134 CHAPTER 5.
new possibilities with respect to image restoration appear. while representatives of the discrete class may arise from time discretizations of semidiscrete ﬁlters. maximum–minimum principles. Lyapunov functionals. Therefore. invariances. informationreducing transformations. These global smoothing properties do not contradict seemingly opposite local eﬀects such as edge enhancement. For this reason it is possible to design scalespaces with restoration properties giving segmentationlike results. continuous dependence of the solution on the initial image. the goal of Chapters 2–5 has been to present a scalespace framework for nonlinear diﬀusion ﬁltering which does not require any monotony assumption (comparison principle). as a consequence. Rather than deducing a unique equation from ﬁrst principles we have analysed wellposedness and scalespace properties of a general family of regularized anisotropic diﬀusion ﬁlters. We have seen that. they give the user the liberty to incorporate apriori knowledge. The degree of freedom within the proposed class of ﬁlters can be used to tailor the ﬁlters towards speciﬁc restoration tasks. On the other hand.Chapter 6 Conclusions and perspectives While Chapter 1 has given a general overview of PDEbased smoothing and restoration methods. Existence and uniqueness results. besides the fact that many global smoothing scalespace properties are maintained. these scalespaces do not need to be uncommitted. Each of these frameworks is selfcontained and does not require the others. semidiscrete and discrete setting. for instance concerning size and contrast of especially interesting features. representatives of the semidiscrete class can be obtained by suitable spatial discretizations of the continuous class. and convergence to a constant steadystate have been established. The large class of Lyapunov functionals permits to regard these ﬁlters in many ways as simplifying. Prerequisites have been stated under which one can prove wellposedness and scalespace results in the continuous. the prerequisites in all three settings reveal many similarities and. 135 .
however. are likely to consist of straightforward extensions of topics presented in this text: • While the theory and the examples in the present book focus on 2D greyscale images. however. Thus. Coherenceenhancing anisotropic diﬀusion is only a ﬁrst step in this direction. In view of these results. it is evident that most of its results can easily be generalized to higher dimensions and vectorvalued images. Last but not least. . The scalespace stack of these ﬁlters appears to be wellsuited to extract semantically important information with respect to a speciﬁed task. and the wish to integrate information from diﬀerent channels. This could lead to speciﬁcally tuned ﬁlters for topics such as perceptual grouping. • The various possibilities to include semilocal or global information constitute another future perspective. or steerable ﬁlters. It is a ﬂexible and mathematically sound class of methods which ties the advantages of two worlds: scalespace analysis and image restoration. we can expect a lot of new results in the near future. Gabor ﬁlters. It is clear. Although this problem is less severe. the general availability of aﬀordable colour scanners and printers. This ﬁeld oﬀers a lot of challenging mathematical questions. the relation between structures at diﬀerent scales has rarely been exploited in the nonlinear context. it would certainly be useful to better understand the deep structure in nonlinear diﬀusion processes. Some of the references in Chapter 1 point in directions how this can be accomplished.136 CHAPTER 6. New ﬁlter models might arise using other structure descriptors than the (regularized) gradient or the structure tensor. the fact that these ﬁlters are steered by the structure tensor instead of the regularized gradient allows to adapt them to more sophisticated tasks such as the enhancement of coherent ﬂowlike structures. that nonlinear diﬀusion ﬁltering is a young ﬁeld which has certainly not reached its ﬁnal state yet. The need for such extensions grows with the rapid progress in the development of faster computers. anisotropic diﬀusion deserves to be regarded as much more than an adhoc strategy for transforming a degraded image into a more pleasant looking one. but also novel ape proaches have been proposed: The use of diﬀusion tensors instead of scalarvalued diﬀusivities puts us in a position to design real anisotropic diﬀusion processes which may reveal advantages at noisy edges. since the avoidance of correspondence problems was one of the key motivations to study nonlinear scalespaces. Interesting candidates could be wavelets. [81] and Whitaker and Pizer [438]. • In contrast to the linear diﬀusion case. CONCLUSIONS AND PERSPECTIVES The analysed class comprises linear diﬀusion ﬁltering and the nonlinear isotropic model of Catt´ et al. Some of its future developments.
Especially combinations with concepts such as data compression. splitting and multigrid techniques. Since these parameters have a rather natural meaning. As a consequence. more timecritical application areas could be explored by applying implicit schemes. In this context it would be helpful to have software packages. may be frightened by this perspective. . 444].137 • Most people working in computer vision do not have a speciﬁc knowledge on numerical methods for PDEs. 431. Thus. the most widelyused numerical methods for nonlinear diﬀusion ﬁltering are still the simple. 36. or neural networks for learning apriori information might lead to novel application areas for these techniques. tomographic reconstruction techniques. and which are easy to use for everyone. Novel. Somebody with another primary interest. Several useful suggestions for parameter adaptation can already be found in [328. there still remains a lot of work to be done. more research on ﬁnding some guidelines for automatic parameter determination for a task at hand would encourage also people without a speciﬁc image processing background to apply nonlinear diﬀusion ﬁlters. for instance a physician who wants to denoise ultrasound images. but ineﬃcient explicit (Euler forward) schemes. • There are not yet many studies which explore the potential of nonlinear diffusion ﬁltering when being combined with other image processing techniques. or grid adaptation strategies. • The price one has to pay for the ﬂexibility of nonlinear diﬀusion ﬁltering is the speciﬁcation of some parameters. It would be nice if this book has inspired its readers to contribute to the solution of some of the remaining open problems. where diﬀerent nonlinear diﬀusion ﬁlters are implemented in an eﬃcient way. 270. Thus. segmentation algorithms. this is not a very diﬃcult problem for someone with experience in computer vision.
CONCLUSIONS AND PERSPECTIVES .138 CHAPTER 6.
M.T. [7] R. Acton. [12] L. Image Processing (ICIP–94. London. Crawford. 1992). 2. Lions. B.Bibliography [1] S. Geoscience and Remote Sensing Symposium (IGARSS ’92. Numer. A fast level set method for propagating interfaces. Math. Alvarez. 1993.T. Acton. Berlin. Sethian. Berger. SIAM J. Numer. SIAM J. 478–482. wavelets and PDEs. 1997. Kittler. Image quantization using reaction–diﬀusion equations. Image selective smoothing and edge detection by nonlinear diﬀusion. Scalespace theory in computer vision. Alvarez. Ackah–Miezan. [2] R. [6] S. J. P.L.A. Springer. 1992. J. 1996). Vol. Conf. Anal. 1996. Phys. [14] L. Proc.M. [13] L. Abbasi. 31. Morel.M. R. Comput. J. A. Proc. J. 284–295. Vol. 1994.). 845–866. Acton. [8] D. 13–16. 1997. M. M. Images and PDE’s. 1995. M. Lecture Notes in Control and Information Sciences. 845–847. 139 . Vol. Deriche. Adams. Gagalowicz. Anal. Springer. Mokhtarian.. Jaﬀr´. Crawford. L. Los Alamitos. May 26–29. Morel. 219. J. Adalsteinsson. Berlin. Image Processing (ICIP–96. on Computer Vision (ICCV ’93. Reliable classiﬁcation of chrysanthemum leaves through curvature scale space. Vol. 269–277. Vol. Vol. 1975. [4] S.T.. Alvarez. Esclarin. 1994.R. Vol. Lineare Funktionalanalysis. Springer.L. 1994. Axioms and fundamental equations in image processing. ICAOS ’96: Images. May 11–14.. [5] S. A mean ﬁeld solution to anisotropic edge detection of remotely sensed data. 16–19. Alt. IEEE Int. 57. New York. Signal and image restoration using shock ﬁlters and anisotropic diﬀusion. SIAM J. Koenderink. J. Discrete models for energyminimizing segmentation. 1993). Florack. Proc. 3. Lecture Notes in Computer Science. IEEE Int. Academic Press. Anisotropic diﬀusion pyramids for image segmentation. Sept.C. Conf. 1992. Alvarez.M. Anal. Inverse Problems.). J. 1993. Vol.A. Berlin. 199–257. Morel e (Eds. 123. ter Haar Romeny. [9] H. 12th Int. J. Los Alamitos. Fourth Int. J. 200–207. L. [10] L. Viergever (Eds. Vol. 29. Guichard. 590–605.W.O. I. 1. IEEE Computer Society Press. 153–175. Vol. 1252. IEEE Computer Society Press. 1994). M. F. Bovik. Lausanne. Mazorra. A. 865–868.M. 1217–1229. Houston. Austin. Alvarez. F. Vol. Rational Mech. 3–14. Edge enhancement of infrared imagery by way of the anisotropic diﬀusion pyramid. Herlin. Arch. Vogel Analysis of bounded variation penalty methods for illposed problems. Proc.. 10.. 118. II. [11] L. 1992. Appl. Acar. Conf. Nov. [3] A. 1996. P. Lions. Sobolev spaces. C.
Dept. A o J. Annals of Mathematics. Parabolic equations on curves and surfaces. P. Partial regularity of free discontinuity sets I. M. J. 857–881. R.. 8. Alvarez. Piazza dei Cavalieri. 1997. Mathematical morphology: The Hamilton–Jacobi connection. 1991. D. Lecture Notes in Computer Science. Preprint CVGMT96040106. Math. 719. Scuola Normale Superiore. IEEE Trans. Anisotropic ﬁltering of MRI data based upon image s gradient histogram. A. Spain. Pallara. J. 31 – Nov.B.R. 451–483. 129–136. On the approximation of functionals depending on jumps by elliptic functionals. Math. 1986.). Comput. Simmons. Kluwer. Springer.O. Bajla. [24] S. Preprint CVGMT96040105. Intersections. 3. Pallara. Duda. Preprint CVGMT96040107. Amsterdam. IEEE Computer Society Press. CDROM document N21. 56126 Pisa. and generalized solutions. A compactness theorem for a special class of functions of bounded variations. Arridge.). 26–33. Campus Universitario de Taﬁra. Sporring. Angenent. 1995. 1992. 1–59. Berlin. 35017 Las Palmas. Tortorelli. [26] S. Parabolic equations on curves and surfaces. 18th Annual Int. 56126 Pisa. Ambrosio. I. Pattern Anal. Babaud. Fourth Int. Gaussian scalespace theory. Berlin. Bajla. Part I. Johansen (Eds. Math. Annals of Mathematics. Italy. 1997. L. 105–123. Uniqueness of the Gaussian kernel for scale space ﬁltering.G. submitted to Annali della Scuola Normale Superiore di Pisa. ter Haar Romeny. L. 215–219. 1995.. Mach. Dordrecht. J. Italy. Acta Numerica. Universidad de Las a Palmas. (EMB ’96. D. A. Un. blowup. Boll. Ambrosio. 999–1036. Pure Appl. Vol. of the IEEE Engineering in Medicine and Biology Soc. L.. Ambrosio. [17] L. Berlin. M. Vol. [25] A. W. 7. Italy. D. Tortorelli. Geometrydriven diﬀusion smoothing of the MR brain images using a a novel variable conductance. 90–97. Vol. 3B. 43. Alvarez. Scalespace theory in computer vision. Ital. Vol. 1995. Scuola Normale Superiore. Vol. V. Kimia. Un.. Maruˇiak. M. Arehart. Conf. Ambrosio. Springer. Ambrosio.P. Chetverikov. Piazza dei Cavalieri. 1994. Vol. Viergever (Eds. Kropatsch (Eds. 1993). D. Piazza dei Cavalieri. Fusco.. Dept. 133. B. Aﬃne morphological multiscale analysis of corners and multiple junctions. [19] L. [27] K.. Math. Vision. V.B. Fusco. Scuola Normale Superiore. [21] L. Boll. Intell. 1993. Florack. Formalization and computational aspects of image analysis. B. Morel. Morales. Vincent. Nielsen. Holl¨nder. Math. Part II. 1990. Koenderink. 1990. . ˇ a [29] I. Los Alamitos. Dept. [20] L. 56126 Pisa. [18] L. Florack. Proc. A. M. Departamento de Inform´tica y Sistemas. 1996). Computer analysis of images and patterns. Partial regularity of free discontinuity sets II. to appear in Int. submitted to Annali della Scuola Normale Superiore di Pisa. [22] L. [28] I.140 BIBLIOGRAPHY [15] L. May 11–14. Baudin. [16] L. Vol. Math. 171–215. Sr´mek. Lecture Notes in Computer Science. Ambrosio. Report No. Higher regularity of solutions of free discontinuity problems. Comm. 132. 1993. Vol. Oct. Heyden. [30] J. 1990.M. Conf. Pallara. M. 9402. Approximation of functionals depending on jumps by elliptic functionals via Γconvergence. N. 224–235.). 1994. N. 1252. Proc. on Computer Vision (ICCV ’93. Multispectral probabilistic diﬀusion using Bayesian classiﬁcation. Angenent. F. Stochastic analysis of image acquisition and scalespace smoothing.. Curves with pintegrable curvature. ˚str¨m. [23] S. Ital. Witkin.
IEEE Trans. Mikula.. Univ. 1993. Berlin. Ph. Sapiro. Robust anisotropic diﬀusion: connections between robust statistics. Aﬃne invariant segmentation by variational a method.D. Dept. Vol. B. a Preprint No. Blomgren. 775–790. 1414–1439. Granlund. An aﬃne invariant model for image segmentation: Mathematical analysis and applications. 24. Vol. Big¨ n. 1995. 79104 Freiburg. London. Lecture Notes in Computer Science. Granlund. 474–477. Texture segmentation by variational methods. 1996. Caselles. 1996. 1987). Dept. Physics Express Letters. G. 13. 27. Spain. A coarsening ﬁnite element strategy in image selective smoothing. Intell. [39] J. pp.M. 1991. CA 90024. Color TV: total variation methods for restoration of vector valued images. [35] G. Springer. 1987. Ballester. [36] B. ¨ [45] P. [32] C. R. M. Image Processing (ICIP–95. . Intell. thesis. J. of Mathematics. June 8–11. 219. Florack. Springer. B¨nsch. to appear in Computation and Visualization in Science. Washington. L.. Barenblatt.. Big¨ n. Ballester. Image Proc.H. Pattern Anal. Jaﬀr´. Palma de Mallorca. 726–741. [44] W. Gonz´lez. ter Haar Romeny. Visual reconstruction. Blanc–F´raud. Vol. Inverse Problems. 56. M. M. of California. M. 1923.O. wavelets and PDEs. Anal. and anisotropic diﬀusion. A. IEEE Trans. 1996. Optimal orientation detection of linear symmetry. Multidimensional orientation estimation with apu plications to texture analysis and optical ﬂow. [34] E. [42] A. Proc.S. Vol. M.F. IEEE Int.). M. Germany. Benhamouda.J. line processing. Kleinman. Aubert. IEEE Computer Society Press. Black. L5–L10. Washington. University of Illes Baleares. D. 1995. Box 3049.. Blake. G. Oct. van den Berg. G. J. master thesis. [43] L. 1995). Chan. 1994. 22–25.). Koenderink. Mach.A.). Barlaud. Charbonnier. [33] C.BIBLIOGRAPHY 141 [31] C. M. Blaschke. J.. 1995. [41] M. 18/1996. Edge focusing. Heeger. [38] F. Conf.O. MIT Press. Appl. 1. Wiklund. SIAM J. 187–193. Nonlinear image processing: e modelling and fast algorithm for regularization with edge detection. Germany. University of Kaiserslautern. Morel (Eds. Pattern Anal. 323–326. London. ICAOS ’96: Images. 1997. a Deriche. Bertsch. T. Dal Passo. CAM Report 965. Cambridge (Mass. 433–438. Faculty of Mathematics. R. Springer. I. P. Hebelstr. [37] P. Math. 9. Zisserman. Vol.H. Math. 11. Gonz´lez. [40] J.M. 1996. Berger. 294–325. of Mathematics. Viergever (Eds. 1987. K. Vorlesungen uber Diﬀerentialgeometrie II.I. Vol. A degenerate pseudoparabolic regularization of a nonlinear forward–backward heat equation arising in the theory of heat and mass exchange in stably stratiﬁed turbulent shear ﬂow. Berlin. e Lecture Notes in Control and Information Sciences. P. to appear in IEEE Trans. SIAM J. Ballester Nicolau. D. Herlin. Parameter adaptation for nonlinear diﬀusion in image processing. Los Angeles. University of Freiburg.E. J. Ughi. Scalespace theory in computer vision. Proc. on Computer Vision (ICCV ’87. U. First u Int. A total variation enhanced modiﬁed gradient algorithm for proﬁle reconstruction. 67653 Kaiserslautern. V. G. Bergholm. Conf. 1252. Vol. Vol. R. 1987. Marimont. Mach.
Mathematical morphology: Extensions towards computer vision. 1995. [60] R. Shaked.D. Bruckstein. [56] F.D. Pattern Anal. Clarendon.M. 42. 1978. Models for the perception of speech and visual form. 1992. 1101–1113. [62] A. Vierde Serie. Vision. 1995. e [50] R. Advanced signal processing: algorithms. Signal Processing. 362–380. Ordinary diﬀerential equations: A ﬁrst course. [52] R. F. 1987. C.. 20–51. [59] W. Int. Bonnet. On projective invariant smoothing and evolution of planar curves and polygons. Burton. Bonnet. 1992. 16. Sur la r´gularit´ des bords des minima de la fonctionelle de Mumford–Shah. P. 1967. Benjamin. 1995. 225–240. New York. Briggs. [57] H. [47] H. Paris. Imag.. Vol.T. Acad. University of Amsterdam. Copenhagen. 1992. van den Boomgaard.). Vol. t. May 10–13. Fast ﬁlter transforms for image processing. School on Classical ScaleSpace Theory. P. D.H. Brauer. 991–1014. 1996). 1275–1279. Analyse fonctionelle. 9. Shaked. The motion of a surface by its mean curvature. T.L.). Ph.T. 1997. 321. Chan. Hubbard. The morphological equivalent of the Gauss convolution. [63] B. Amsterdam. S.J. A multigrid tutorial. [53] R. thesis. 3162. oral presentation.A.142 BIBLIOGRAPHY [46] P. Blum.E. 1973. R. Blomgren. 1981. Computer Graphics and Image Processing. Morphology. Sci. J. Ph. A. van den Boomgaard. The Netherlands. Bramble. San Antonio. G. Op´rateurs maximaux monotones et semigroupes de contractions dans les ese paces de Hilbert. IEEE Southwest Symposium on Image Analysis and Interpetation (SSIAI ’96. Brockett. [61] A. 3377–3386. [58] H.. Luk (Ed. Mulet. Vol. Vol. 1994. van den Boomgaard. [48] A. Brezis. 1994. S´rie I. Comp. IEEE Trans. Vol. Maragos. Maximum entropy in action. J. Sapiro. .A. Press. Mach. architectures. 1997. [55] J.I. M. e e C. Nohel. Paris. 16. Deel 10. [51] R. Nieuw Archief Voor Wiskunde. New monotone type approximations for elliptic problems. to appear. 1964. Vol. 18. 321. 1121–1126. R. Math. SIAM. van den Boomgaard. 349–367. D. IEEE Trans. April 8–9. Burt. S´rie I. Cambridge. Macaulay (Eds. Bruckstein. Intell. [65] M. B. Target tracking using the anisotropic diﬀusion pyramid.W. Brakke. Acton. e [49] A. 1991. J. and implementations VII. A transformation for extracting new descriptors of shape.). t. SPIE Vol. 1996.F. 1967. Oxford. Acad. Princeton University Press. Extensions to total variation denoising. Intell. Philadelphia. Pattern Recogn. Smeulders. Artif. Brezis. Evolution equations for continuousscale morphological ﬁltering. Masson. 219–236. Math. Buck. Charact´risation des minima globaux de la fonctionelle du Mumford–Shah en e segmentation d’images. The morphological structure of images: The differential equations of morphological scalespace. Paris. Princeton. W. North Holland. Sci. Wathen–Dunn (Ed. 7. [54] K. V. [64] P. Proc. Evolution of planar polygons.M.
Canny. G. Morel (Eds. Herlin. Vol.BIBLIOGRAPHY 143 [66] M.). Int. Berlin. Coll. M. 1996.S. Numer.). Vol. M. A “small leakage” model for diﬀusion smoothing of image data... 1988. Butt. [80] F. Sbert. Austin. Partial diﬀerential equations and image processing.. Caselles. 1992.L. Anal.D. 11th Int.. J.. 301. G. Pyramidal architectures for computer vision. Coll. J. 597–605. Braunschweig. SIAM J. Sapiro. 182–193. Joint Conf. Vol. C. 1994). Maragos. Caselles. Dept. [69] J. 1997. U. Image Processing (ICIP–94. 1895–1909. A geometric model for active contours in image e processing.). 1997. Cai. [76] V. J. Vol. Numer. Comparison of multiscale morphology approaches: PDE implemented via curve evolution versus Chamfer distance transform. 33. Springer. 423–451. 1996. NJ 08544–5263. 8.W. ter Haar Romeny. Minimal surfaces: a geometric three dimensional segmentation approach. A morphological scheme for mean curvature motion and e applications to anisotropic diﬀusion and motion of level sets. Adaptive image smoothing respecting feature directions. 29–49. B.A. 77. Springer. 1989). J. Lecture Notes in Control and Information Sciences. 125–133. IEEE Trans. 29. Nov. [75] V. wavelets and PDEs. Viergever (Eds. Kluwer. R. Pattern Recognition. Vol. F. 66. Math. 1996. SIAM J. 1989. J. Vol. Vol. 22. 1585–1590. 1995. Berlin. Morel. A computational approach to edge detection. Cantoni. Caselles. [77] V. Preprint. 1997. Math. Los Alamitos. 679–698. Numer. a 1983. Image selective smoothing and edge detection e by nonlinear diﬀusion. Vol. [68] L. [74] V. [79] F. [78] V. [70] V. Lions. [71] M. F.P. Appl. Kimmel.A. 1996. Some notes on repeated averaging smoothing. Mach. Coll.M. Koepﬂer. [82] A. T. Princeton University. Cai. 20–25. B. [73] V. Princeton. Kimmel. What is the best causal scalespace for 3D images?. I. Snakes in movement. Caselles. Pattern Anal. IEEE Computer Society Press. Sapiro. P. Morel. Proc. Lecture Notes in Computer Science. 1252.O. Anal. 31–40. Schafer. Comput. Caselles. Ferretti. Image Proc.M. R. Carmona. Mathematical morphology and its applications to image and signal processing. Proc. Math. Scalespace theory in computer vision. J. Vision. Berger. 61–79. 1993. Maragos. Anal. Zhong. Convergence of iterated aﬃne and morphological ﬁlters by nonlinear semigroup e theory. Kittler (Ed. Dibos. M. Scalespace versus topographic map for natural images. Vol. SIAM J. 13–16.A. Numer. T.. C. Conf. New York. 2445–2456. 1996. Intell. B. Geodesic active contours. 16–20. Dibos. 219. Chambolle. Aug. Koenderink. G. Catt´. 1–31. [72] R.D. Catt´. 1994. 1986. of Civil Engineering. Vol. P. Butt (Eds. Plenum Press.. submitted to IEEE Trans. R. Deriche. J. Diﬀerentialgeometrie von Kurven und Fl¨chen. Springer. ICAOS ’96: e Images. [67] L. 1. Catt´. do Carmo. Vol. Lecture Notes in Computer Science. Caselles. Florack. 1994. F.. on Artiﬁcial Intelligence (IJCAI ’89.M. Boston. SIAM J. 1199–1246. R.. Vieweg.). Catt´. [81] F. Sbert. Detroit. 56. P. S. 2. Vol. Jaﬀr´. L. 32. London. M. Vol. . Coll. IEEE Int. Numer.
S.O. submitted to SIAM J. Ph. Neural dynamics of brightness perception: Features. Charbonnier. Los Alamitos. Vol. On active contour models and balloons.D. 1992.A. Advanced signal processing algorithms. [84] R. Wang. 53. P. Chan. Numer. Gainesville. 164– 168. 314–325. 36. 1997.I. Zhou. J. Lopez. 2567. 214–222. Dept. M. [85] T. e France. 1995. 33. Chopp. CA 94720. CAM Report 9645.). Wavelets: A tutorial in theory and applications. 211–218. Cohen. J. J. Chambolle. New York.). IEEE Computer Society Press. Second IMACS Symposium on Iterative Methods. Comput. CISE Dept. Chan. of California. R. Lawrence Berkeley Laboratory. 75775 Paris Cedex 16. Lopez. IEEE Int. Vemuri. F. Vassilevski (Eds. 311–329. C..A. [89] Y. University of Florida.F.L. [90] Y. Berger. Vol. [86] T.K. Place du Mar´chal de Lattre de Tassigny. Appl. 76. Proc. FL 32611. SPIE Vol.M.. Diﬀerential Geometry. Jaﬀr´.K.L. Image Processing (ICIP–94. Universit´ e Paris IX – Dauphine. Preprint No. Vol.F. Herlin. J. CA 90024. 1994. Morel. wavelets and PDEs. [88] P. . Berkeley. 1991. [91] D. R. U. P. Lecture Notes in Control and Information Sciences. B. C. [97] T. Vol. Nov. Chan. Rudin. diﬀusion and resonance. Math.K.H. boundaries. Proc. 106.S. Goto. 1992. U. 167–188. 1991.M.H. London. 1994). [94] M. Austin. Giga. Chen. Springer. CEREMADE.G. [92] C. Perception and Psychophysics. 1996. Physics Division. Vol. Aubert. Continuation method for total variation minimizing problems in image processing. thesis LBL30685. 1994. Univ. IMACS Press.. Vol. University of California.. IEEE Computer Society Press. Vol. Guichard. A nonlinear primal–dual method for totalvariation based image restoration. Morel. Phys. G..T. of Mathematics.M. Conf. J. 77–91. S. Morel (Eds. Grossberg. Image Proc.).S.M. Two deterministic halfe quadratic regularization algorithms for computed imaging. F. Eve. G. 241–252. 2. Chen. Oct. Integral and local aﬃne invariant parameter and application to shape recognition. C. Chan. Luk (Ed.F. P. Los Alamitos. S. Deriche. H. L. Chan. 9–13. 13–16. Computing minimal surfaces via level set curvature ﬂow. Total variation blind deconvolution.D.F.). 1991 (see also J. Margenov. 2563. Scalespace and aﬃne invariant recognition of occluded shapes. to appear in IEEE Trans. J. Chui. C. 749–786. [87] T. Bramble (Eds. [93] L. 428–456. Chan. M.A. 1994). Barlaud.S. 219.A. 1993). 1997. Cohignac. Cohignac.. A. Cosine transform based preconditioners for total variation minimization problems in image processing. Report UFCISETR97015.144 BIBLIOGRAPHY [83] A. Vol. Lions. Pattern Recognition (ICPR 12. Cohen. Image denoising and segmentation via nonlinear diffusion. Numerical analysis of the fundamental equation of image processing. CVGIP: Image Understanding. Wong. SPIE Vol. Image recovery via total variation minimization and related problems. Jerusalem. U.H. 168–172. 1995. Conf. Cohignac. I.C. L. Golub.M. F. Proc. Math. Morel. S. Blanc–F´raud. [95] T. Academic Press. T. Mulet. 9254. 12th Int. Wong. Y. e ICAOS ’96: Images. Investigative and trial image processing. 1995.K. [96] T. 1984. Los Angeles. L. Uniqueness and existence of viscosity solutions of generalized mean curvature ﬂow equations. 1996.
Bull. Math. O. 13th Int. Kornprobst. e [113] F. Proc. Vol. Di Zenzo.D. [100] G. Cottet. Anal. [105] J. Second Int. Santosa. Diﬀerential Equations. [111] R. 1996. Vol. De Giorgi.H. on Mathematics in Industry. [114] S. Deriche. 1995. 1996). Ong S. 783–888. 16–19. Proc. An image enhancement technique for electrical impedance tomography.. 1–67. S. 403–406. 195–218. Arch. Conf. Second Asian Conf. F. 1996. 75. Am. Soc. Acad. 7–11. t. [108] E. [104] M. IEEE Int. 6. Dibos. Vol. Daubechies. M. Cottet. 17–41. 290–295. Crandall. Deriche. Diﬀusion approximation on neural networks and applications for image processing. Singapore. C. Sept. V. Aubert. 1992. Image processing through reaction combined with nonlinear diﬀusion. A.C. 263–267. [107] G. 1997. A note on the gradient of a multiimage. Germain. Proc. Sept.. SIAM.H. P.H. Recovery of blocky images from noisy and blurred data. Analyse multi´chelle invariante par transformations projectives. Local Morse theory for solutions to the heat equation and Gaussian blurring. 56. P. 3–9. 1.G. 1992. Vol. S´rie I. Lions. 108. I. 61. Neural networks: Continuous approach and applications to image processing. Biological Systems.L. . Traitement du Signal. [115] D. 1996. 1992. Hodnett (Ed. Lausanne. 115. 2. Proc.BIBLIOGRAPHY 145 [98] G. Ma. [102] G. 1995. [110] R. Dobson. Vol. Convergent diﬀerence schemes for nonlinear parabolic equations and mean curvature motion. [106] I. 1995. [101] G. 1989. SIAM J. G. Crandall. Proc. Aug. 1995). Ang Y. Vol. 116–125. Vol. Appl. Lions. Vol. Damon.. 1994. Graphics.H. No.H. P. Conf. Cong. Math. 1996). 10. 317–334. F. Santosa. (Eds. Mech. Vol. Math. J. Image Processing (ICIP–96. L. Math. F. Philadelphia. El Ayyadi. December 5–8. Srinivasan. M. 3.L. 1993. Inverse Problems. [112] F. Dibos. Sept. 1996). (Singapore. SIAM J.. Math. Vol. Ten lectures on wavelets. Projective invariant multiscale analysis. Nonlinear diﬀusion for early vision. 27. Vol. Deriche.). 485–488. 320. Nonlinear PDE operators with memory terms for image processing. Leaci. Carriero. Appl. Pattern Recognition (ICPR 13. 16–19. 1995. 56. 481–483.. Comp. Conf. 917–921. Paris. 1986. Image Processing (ICIP–96. Teubner.). David. User’s guide to viscosity solutions of second order partial diﬀerential equations. 368–401. 1996. Existence theorem for a minimum problem with a free discontinuity set. Computer Vision. Stuttgart. Ishii. 1996. [99] G. Singapore Conf. Numer.. Lausanne. Computer Vision (ACCV ’95. 1992). Dobson. Cottet. J. Proc. A. Opticalﬂow estimation while preserving its discontinuities: A variational approach. [116] D. 1. Vol. Cottet. on Image Proc. Les EDP en traitement des images et vision par ordinateur. Vol.C. and Image Processing. e Sci. [103] M. 25–30. 1992. Sixth European Conf. 33. Vol.. 1181–1198. 1996. Vol.G. 1131–1139. Recursively implementing the Gaussian and its derivatives. [109] R. C 1 arcs for minimizers of the Mumford–Shah functional.H. Faugeras. Rat. 13. Vienna. IEEE Int. R. 659–673.
M. Lecture Notes in Computer Science. [126] A.. 1995. A diﬀerential geometric approach to anisotropic diﬀusion. 1995). Acad.E. Springer. . 1996. Montana State University.). Dorst.F. Dordrecht. Vol. [127] A.A. A. wavelets and PDEs.G. Deriche. Faugeras. G. G.C. U. Nov. van den Boomgaard. Math. Xidian University Press.S. Keriven. 371– 391. R. 1994. Clarendon. Numer. 635–681. Chenke. Jones. 1963. Convergence of an iterative method for total variation denoising. Applications of invariance in computer vision. 1991. Parabolic and hyperbolic problems. Faugeras. Math. Preprint. Part I. Douglas. Spruck. Workshop on Geometrical Modelling and Invariants for Computer Vision (Xi’an.M. Vol. 79–98. Dobson. R.. Dept. Signal Processing. 825. [123] S. C. Kluwer. On predictor–corrector methods for nonlinear parabolic diﬀerential equations. Appl. London. SPIE Vol. Diﬀerential Geometry. Tescher (Ed. B. Berlin. 23. Analysis of regularized total variation penalty methods for denoising. R. Hanke. 6. Numerical study of a partial diﬀerential equation. A general formulation of alternating direction methods. 4. Geometrydriven diﬀusion in computer vision. Los Alamitos. Vol. W. 1995. [124] D. ElFallah. Neubauer. On projective plane curve evolution. e e Paris. 1983. Soc. W. SPIE Vol.146 BIBLIOGRAPHY [117] D. Cartan’s moving frame method and its application to the geometry and evolution of curves in the Euclidean. M. J. J. Numer. Vol. Motion of level sets by mean curvature I. Lecture e Notes in Control and Information Sciences. 45–49. [129] L. Vol. Proc. J. Gunn. 219.R. [122] I. 601–617. Springer. 1. Jaﬀr´. Phys. [125] S. Vol. 1993. Stevenson (Eds. Dobson. [130] O.). 428–453. J. 11. Regularization of inverse problems. 66–73. Vol. R. Mohr (Eds. S´rie I. Conf. Duﬀ. 565–570.I. IEEE Int. Maths. 1994. 439–450. MT 597170240. Proc. J.C. ElFallah. 1567. SIAM J. O. Image and video processing II. Ind. D. 49–63. 13–16.S. Morel (Eds. J. Morphological signal processing and the slope transform. Nonlinear adaptive image ﬁltering based on inhomogeneous diﬀusion and diﬀerential geometry. Xi’an. [120] J.). Vol. Reid.R. R. Douglas.I.. [119] L. Image Processing (ICIP–94. Faugeras. e [131] O. Anal.). Mundy. April 27–30. Nonlinear signal processing using integration of ﬂuid dynamics equations. aﬃne and projective planes. J. C. ICAOS ’96: Images. 317. ter Haar Romeny (Ed.K.S. t. [121] J. 2182. 298–302. Bozeman. Erisman. Sci. 1986. M. of Mathematical Sciences.E. Scherzer. A. Evans.). 38. Eidelman. R. Dzhu Magazieva.S. Sur l’´volution de courbes simples du plan projectif r´el. Forsyth (Eds. Austin. The evolution of mean curvature in image ﬁltering. Grossmann. Keriven. Scalespaces and aﬃne curvature. 195–204. U. [133] O. Vol. A. Math. 1994. Ford. Eberly. 1994).L. [128] H. S. Vogel. Berger.. Rajala.A. 1994. R. Direct methods for sparse systems. 1964. Faugeras.O.. A. 33. No.K. to appear. 11–46. 1991. Engl.W.L. Friedman. 1994. Zisserman. Kluwer. 12. Ford..): Applications of digital image processing XIV. I. [118] D. B. A. Inverse Problems.M. Oxford. Herlin. J. 1996. IEEE Computer Society Press. 1996. Comput.E. [132] O. Dordrecht.
Proc.J. Ford.BIBLIOGRAPHY 147 [134] O. D. models. B.. M. 1997. J. Schwartz. 279–294. Germany. Florack. Scalespace theory in computer vision. 1808. corners and centres of circular features. B. Viergever (Eds. Contemp. 16–19. ter Haar Romeny. Image Vision Comput. F¨rstner. B. M. Florack. J. 30. 276–280. 1992. IEEE Int. Image Vision Comput. 18. Koenderink. Pauwels. [145] M. Geißler. Adaptive nonlinear ﬁltering for nonlinear diﬀusion approximation in image processing. Chen. 13. Koenderink. Dept. Pattern Recognition. Technical Report UUCS199725. Springer. E. 13th Int. Image Processing (ICIP–96.J. Maas. 817–835.J. Estes. Florack. 1997. 427–432. Vol. Box 3049. San Francisco. Scale and the diﬀerential structure of images. Florack. 1997. 25–30. [148] J. [150] M. On areapreserving evolution equation for plane curves.M. Nonlinear scalespace.. . SPIE Vol. Neumann. Robb (Ed. Mustererkena nung 1996. R. Springer.M. [138] L. 1996 IEEE Int. Springer. H. [141] L. 1996. Proc. R.J. Math. Report No. Conf. Seydel. Vol. 801–807. Conf. 61–75. H. 1995. Gabor. Vol.A. of Computer Science. 1986.A. Fiddelaers. [136] P. on Fast Processing of Photogrammetric Data (Interlaken. Vol. 51–62.L. Proc. E. 1252. 1992. 1987.S. Franke.M.A. 10. F. Image processing using a wavelet algorithm for nonlinear diﬀuo sion. 67653 Kaiserslautern. [140] L. H. [135] S. Dordrecht. Fejes. Vol. [143] L. R.M. Koenderink. [149] D. Fritsch. 1994. Faugeras. J. 1965. Speech and Signal Processing (ICASSP–92. 1996. M.J. R. J.O.J. Aug. Discrete active models and applications. Florack. Hering (Eds. 1994. R. Lecture Notes in Computer Science. 3584 CH Utrecht. 165–168.M. Pseudolinear scalespace theory: towards the integration of linear and morphological scalespace paradigms. 376–388. Visualization in biomedical computing ’92. 104. J.). [142] L. 105–117.M. Padualaan 14. and images. University of Kaiserslautern. Berlin. A medial description of greyscale image structure by gradientlimited diffusion. 47–54. ter Haar Romeny. Florack. Vol. P. Lecture Notes in Computer Science. Kluwer. ter Haar Romeny. 281–305. Keriven. Image structure. [147] D. 1992. 1996. Level set methods and the stereo problem. Haußecker.A. June 2–4. Fischl. 14. Van Gool. [137] B. Proc. E. Information theory in electron microscopy. P. J. Viergever. Conf. I. Vol.H. L. Salden. Comput. Gage.E. Fr¨hlich. Scalespace analysis for image sampling and interpolation. J.J. [139] L.).J. Weickert. Vol. Anisotropic diﬀusion based on mean curvature motion: A computational study.M. 1992). Laboratory of Technomathematics. M. B. March 23–26. Acoustics. Rosenfeld. 1996). 1997. Vision. 1996. The Netherlands. 23. Viergever. J¨hne.M. ter Haar Romeny.). Koenderink. Data. Vienna. [144] G.). The Gaussian scalespace paradigm and the multiscale local jet. Vol. 3. [146] T.J. A. G¨ lch. Vol. B. A fast operator for detection and precise location of distinct o u points. Lausanne.. 1987). 800. ISPRS Intercommission Conf. Berlin. Sept. 1996). 272–283. Eklundh (Ed. Int. 469–472.O.R. Florack. Vol. Geometrydriven curve evolution. A. Berlin. Viergever.J. L. Pattern Recognition (ICPR 13. Laboratory Investigation. Computer Vision – ECCV ’94.A.
. 7. A. Proc. Giuiliani.. J. Digital image processing. [153] D. Vol. H. D. 1992. Catthoor. 1985. H. Gerig. Vol. Jain. Jolesz. Switzerland. De Man. Selfportrait. Reynolds. van Gogh. and the Bayesian restoration of images. IRE Trans. ETH– Zentrum. University of Kaiserslautern. Granlund. [169] M. 15. 8092 Z¨ rich. Gerig. 1991. Wintz. [158] T.C. Vol. AuverssurOise. Medical Imaging. Int. Communication Technology Laboratory. Muse´ d’Orsay. Gerig. 1962. Diﬀerential Geometry. [166] A. Di Paola (Eds.O. Starry night. Robinson. Comput. Image Vision Comput. Kluwer.S.C. Gonzalez. IEEE Trans. A common framework for image segmentation. [171] L. Diﬀerential Geometry. Finding line structures by multivalued nonlinear diﬀusion of feature e maps. 1986.D. [159] D. 8. 69–96. 721–741. Vol. Grayson. C. of Mathematics. 1889. 15–23. Information Processing in Medical Imaging. 279. P. Otterlo. Addison–Wesley.D. . 1995. 247–256. SaintR´my. Germany. 1992. R. FB Mathematik. 285–314. 367–383.. Arnimallee 2–6. New York. Intell. G. [152] D. Vol. 23. Dordrecht. The heat equation shrinks embedded plane curves to round points. W1000 Berlin 33. Image Vision Comput. Kluwer. Gage. P. Intell. Six. o u [163] R.. 10. Stochastic relaxation. Germany. SaintR´my. [167] R. [172] L. R. R. Vol. 39. Kikinis. 67653 Kaiserslautern. Greenspan. Preprint No. 26. 1890. G. Vol. Box 3049. 369–398. Graham. e e [162] V. FU Berlin. Road with cypress and star. On edge focusing.R. Van Gool. Image Vision Comput. [164] R. J. Gijbels. Goshtasby. P. 129–144. G. A. Hamilton.. Vision.. Report. 14. u [157] G.). L. Y. Knutsson. Vol. 1995. Barillot. G. Franklin Inst. Vol. J. Snow removal: A noisestripping process for TV signals. Israel. Sz´kely. 1992.C. Scaleimprecision space. Gibbs distributions. Reading. Yuille. Bizais.P.H. Vol. Colchester. Paris. [154] S. 12. [165] A. Rijksmuseum Kr¨ller–M¨ ller. Vol. 1994. 6. IEEE Workshop on VLSI Signal Processing. The heat equation shrinking convex plane curves. A. van Gogh. master thesis. Griﬃn. Gorenﬂo. Directed diﬀusion equations and edge detection. 227–243. Dept. Scale and segmentation of greylevel images using maximum gradient paths. 1965. 165–176. Griﬃn. 1889. van Gogh. Constrained restoration and the recovery of discontinuities. Geman. Information Theory..A. Sz´kely. [168] G. K¨ bler. Geman. e [161] V. 11. Geman. Geiger. A VLSIarchitecture for parallel nonlinear diﬀusion with applications in vision. [170] D. Image Science Division. On nonnegative diﬀerence analogues of elliptic diﬀerential equations. M. Berger.148 BIBLIOGRAPHY [151] M. Nonlinear anisotropic ﬁltering of MRI data. IEEE Trans. Image Vision Comput. Dordrecht. 3. Pattern Anal. 1984. Signal processing for computer vision. 1987. Implicit convolution. O. Gourlay.F. Vol. 1977. Mach. 360–365. Detection and characterization of unsharp blobs e by curve evolution. Conservative diﬀerence schemes for diﬀusion problems. 398–407. G. Pattern Anal. 1994. [160] V. 6. 1997. The Museum of Modern Art. u IEEE Trans. P. 1993. F. Vol. [156] G. 1987. Mach. J. F. 389–402. Oosterlinck. [155] G. 1991. 221–232.
Proc. J. Viergever.J. J. Diﬀerential structure of images: accuracy of representation.M. [178] B. 1984. Manjunath. Uniformly high order accurrate essentially nonoscillatory schemes. Dordrecht. F.). [189] R.J. Mathematical morphology: A modern approach in image processing based on algebra and geometry. Kluwer. Aug. Math. Mathematical methods in medical imaging III. [181] R. Reconstructions from zerocrossings in scale space. 71. Heijmans. Dordrecht. Huisken. L. Variational image segmentation using boundary functions. Watson (Eds. Harris. SIAM Review.. S. Vol. 1959. 2299. Miami Beach. Bookstein. Image Proc. 20. N. The state of the art in numerical analysis. C. Manchester.M. University of California. 2. Duncan. 2111–2130. Preprint No. Florack. 299–316. Soc. France. June 22–26. Santa Barbara. 231–303. 1959 (in Japanese). 3–19. [190] R. [188] G. Hellwig. U. S.M. Wilting.S.. 532–550. J. R. M. [176] B. Nielsen. Hewer. . 1996. Japan.M. 1. Moniot. Enquist. to appear in I.R. [184] H. [180] W..A. o Trans. [175] B. 31 – Sept. Applications of scalespace theory. 147–152. Vol. J. Niessen. Vol. Stuttgart. IEEE Int. IEEE Computer Society Press. 37. Vol. Hackbusch. 1994). M.M. 1977. Phys. J. 278. Oxford University Press. Basic theory of pattern observation. Proc. Report No.C. CEREMADE. Stephens. L. CA 93106. Guichard. Harten. 1994. J. ter Haar Romeny. X. Frontend vision and multiscale image analysis: Introduction to scalespace theory. Morel. Computer Vision and Pattern Recognition (CVPR ’86. Amer. [174] B. Duﬀ. SPIE Vol.M. Osher. New York. B. Partial diﬀerential equations and image iterative ﬁltering. IEEE Comp. Gaussian scalespace theory. Universit´ Paris IX – Dauphine. Image analysis using mathematical morphology. Speech and Signal Processing. 1989. Hummel. [182] C. Sporring. IEEE Trans. Springer. Pattern Anal. H¨llig. Hummel. III. IEEE Computer Society Press. Dordrecht. S. [185] G. Kenney. Diﬀerential Geometry. [179] W.). Chakravarthy. Comput.A. Conf. 1995. [191] T. Conf. Deblurring Gaussian blur. P. [177] B. Fouth Alvey Vision Conf. Theorie und Anwendung der direkten Methode von Ljapunov. 204–209. A combined corner and edge detector. 13–16. 1986). 1994. Zhuang.A.G. [187] K. 9607.J. Hahn. Proc. 1995. Intell. Vol. L. Place du Mar´chal e e de Lattre de Tassigny. Iijima. 1985. [186] G. M. Los Alamitos. Springer. Wilson (Eds. Berlin. ter Haar Romeny (Ed. Florack. IEEE Trans. Wilting. 1997. J. Image Processing (ICIP–94. 9535. B. Washington. Haralick.). Lange. (AVC 88.M. 1987. Partial diﬀerential equations. Existence of inﬁnitely many solutions for a forward–backward heat equation. submitted to IEEE Trans. ter Haar Romeny.S. Sternberg. Geometrydriven diﬀusion in computer vision. Mach.M.L. D. 9. 1983. in preparation. 1987.J. M. [183] A.S.M.A. ter Haar Romeny. Johansen (Eds. 1988. 1986.A. de Swart. Papers of Technical Group on Automata and Automatic Control. W.BIBLIOGRAPHY 149 [173] F. Soc. 75775 Paris Cedex 16. 37. Nov. 1994. Flow by mean curvature of convex surfaces into spheres. Acoustics. Representations based on zerocrossings in scale space. 237–266. ter Haar Romeny. Teubner. Kluwer. 21–25. Florack. Multigrid methods and applications.. 1988).. Vol. ECE Dept. Vol. 1–36.). Dec. G. 139–148.M. Kluwer. IECE. Austin.S.
Tie. The Hague. No. Iijima. 10623 Berlin. 16. Japan. Systems. 1992. Basic theory on normalization of a pattern (for the case of a typical onedimensional pattern). 545–554. 1989 (in Japanese). 1973 (in English). Aug. [203] R. 11th Int. Juni 136. Iijima. Meth. 16. H. [208] M. K. Vol. Basic equation of ﬁgure and observational transformation.. 1993. Germany.T. 1993. 1962 (in Japanese). Pattern Recognition (ICPR 11. Appl. Deriche. Ito. Vol. Fachbereich Mathematik (3). [198] T. [194] T. 1994. First Int. 252–255. C. Los Alamitos.. 123– 134. Spatiotemporal image processing. Augmented Lagrangian methods for nonsmooth. Jackway. J. An active set strategy for image restoration based on the augmented Lagrangian formulation. Oct. Vol. Japan. Theory of pattern recognition. Math. on Computer Vision (ICCV ’95. Global existence for twovelocity models of the Boltzmann equation. Proc. J¨gersand. Math. Series of Basic Information Technology. [199] T. Nov. Straße des 17. 1996. a [205] K. Electronics and Communications in Japan. [209] B. Neunzert. 26. Morphological scalespace. June 20–23. Sci. H. Kunisch. Fachbereich Mathematik (3). Illner. Oct. Proc. 1973 (in Japanese). M. [204] K. Berlin. 1993. 368–388.T. Conf. 1971 (in English). J¨hne. Mori. Juni 136. Iijima. Math. Relative entropy maximization and directed diﬀusion equations. Genchi. IEEE Trans. a [206] P. [201] R. 4. Illner. Sci. 1992). 1972). Germany. [193] T. 195–202. Vol. A theory of character recognition by pattern matching method. 1. 50–56. [202] R. 681–690. Iijima. [207] P. Basic theory on normalization of twodimensional visual pattern. 38–51. Sci. Intell. 3. Theory of pattern recognition. Mori. Iijima. Kunisch. 751. First USA–Japan Computer Conference (Tokyo. H. Fifth Int. Pattern Anal. Cambridge. Washington. K. 1963 (in English). 1994. 10623 Berlin. [196] T. 1979. Pattern Recognition (IJCPR 1. Vol. Vol. Coronasha. 70–77. 1963 (in Japanese). K. 409/1994. 6. IECE. Pattern recognition. Preprint No.. 30 – Sept. Ito. 410/1994. Joint Conf. Illner. Iijima. IECE. 18. Saliency maps and attention selection in scale and spatial coordinates: an a information theoretic approach. Studies on Information and Control. 187–193. A theoretical study of pattern identiﬁcation by matching method. convex optimization in Hilbert spaces. Morishita Publishing. Pattern Recognition Issue. Proc. K. Technische Universit¨t Berlin. Iijima. No. Vol. Conf. 2. Vol. Proc. Jackway. 1962 (in Japanese). 1972 (in English). 1995. [197] T. 1995). Straße des 17. Controls. Vol. Meth. Technische Universit¨t Berlin. IEEE Computer Society Press. Mach. Bulletin of the Electrotechnical Laboratory.150 BIBLIOGRAPHY [192] T. Genchi. Preprint No. Papers of Technical Group on Automata and Automatic Control. Oct. Iijima. On directed diﬀusion with measurable background. [200] T. [195] T. 3–5. Lecture Notes in Computer Science. 42–48. 1. Appl. Observation theory of twodimensional visual patterns. a Springer. 1973). Scalespace properties of the multiscale morphological dilation– erosion. Iijima. Appl. Meth. Computers.. 15–22. .
1997. 275–301. Vol. [216] J. Computer Vision. J. to appear. [218] H. A. Johansen. Vision. B. Kimia. Kass. [225] S. A. c Koenderink. Springer. 1986. Germany. Praeger. Berger. Witkin. Wavelets and PDEs.)..). D. [214] J. Tannenbaum. Num. [212] P. Rosenfeld. Lecture Notes in Control and Information Sciences. Scalespace theory in computer vision. Vol. Vol. 1986). Zucker. K. London. University of Kaiserslautern. [219] G. Computer vision – ECCV ’92. 1996. Springer. Appl. J. Kichenassamy. A. 305–322. [228] B. 321–331. Vol. Computer Vision and Image Understanding. J. [213] F.O.W. 395–410.B. Appl. 1. Mikula. 1988. Sandini (Ed. Maximum and comparison principles for anisotropic diﬀusion.). diploma thesis. Imag. 1997. Terzopoulos. SIAM J. J. Vol. Proceedings of the workshop on Gaussian scalespace theory.). Mikula. and Image Processing. 119–124. Kaˇur. 1987.G. A computational framework for determining stereo correspondence from a set of linear spatial ﬁlters. [221] M. Math. M. Slowed anisotropic diﬀusion.. Faugeras (Ed. 219.BIBLIOGRAPHY 151 [210] P. A. Viergever (Eds. Malik. Kumar. Jones. Kluwer. ter Haar Romeny. L. Jolion. Computer Vision – ECCV ’90. [217] J. 64. Kichenassamy. Report DIKU 96/19. Johansen. Herlin. Tannenbaum. A pyramid framework for early vision. University of Cologne. Toward a computational theory of shape: An overview. Lecture Notes in Computer Science. Kawohl. Morphologische Diﬀerentialgleichungen in der Bildverarbeitung. Princeton. 57–67. Berlin. A. Partial diﬀerential equations. Deriche.W. 2100 Copenhagen. 357–360. Siddiqi. 1992. 17. Skelboe. Organization in vision. 27–31. Dept. Van Nostrand. Springer. Kimia. Vol. 215–217. J. Springer. S. [223] J. Germany. Comput. Kaﬁtz. [211] P. P. [224] S. Lecture Notes in Computer Science. 1252. P.G. 362–385. K. 4. Eighth Int. Olver. Nonlinear diﬀusions and hyperbolic smoothing for edge enhancement. Preprint. D. Kaˇur. Kass. Vol. Anal. O. of Computer Science. New York. Morel (Eds. Mech. . 1995. Johansen (Ed. 1994.. Math. Vision. Kutev.O. 1960. 1996. 50923 Cologne. Vol. I. Yezzi. Analyzing oriented patterns. On the classiﬁcation of toppoints in scalespace. Kemeny.).D. 47–59. of Mathematics. Kanizsa. [226] S. Kichenassamy. Dordrecht. J. 1990. A. [227] B. Vol. Finite Markov chains. Conf. Universitetsparken 1. G. 1996. [215] D. 427. Witkin. Paris. Pattern Recognition (ICPR 8. Dept. Arch. Mathematical Institute. Denmark.M. 1979. Berlin. Int. Math. 67653 Kaiserslautern. 134. Snell. J.L. Snakes: Active contour models. M. [222] B. N. Grue. Jaﬀr´. Proc. The Perona–Malik paradox. Springer. Berlin. J. New York. Conformal curvature ﬂows: from phase transitions to active vision. A. Rat. Vol. [220] M.B. Solution of nonlinear diﬀusion appearing in image smoothing and c edge detection. University of Copenhagen. Oct. 1996. K. R. S. Andersen. 1978. ICAOS ’96: Ime ages. Box 3049. Florack. 1994. 37. K. 588. 1996. Representing signals by their toppoints in scale space. 402–407. Lecture Notes in Computer Science. Geometric heat equation and nonlinear diﬀusion of shapes and images. John.M. Graphics.
15. Iterations of a nonlinear transformation for enhancement of digital images. 598–610. 282–299.. Comput. [246] O. Kimmel. Vol. [244] K. 1252. S.R. . 1968. I. R. 438–458. 1996. CA 94720. The structure of images. 45–49. 30 – Sept. 5. Conf. 50. Vol. [235] R. N. University of California. 1987.. June 17–19. and volumetric medical images. Anal. [239] G. 189–224. R. 1975.M. [238] M. Koenderink. 1992.152 BIBLIOGRAPHY [229] B. Anal. N. 159–162. 1997. Granlund. B.A. Aug. C. Scale space ﬁltering on spherical pattern. 55. Bruckner. ter Haar Romeny.J. Florack. Ayache.M. Viergever (Eds. 1997). Kimmel. Math. Vol.. M. Sochen. On the evolution of curves via a function of curvature. Morel. Vol. Cybern. 31. 1996). Koenderink. Images as embedding maps and minimal surfaces: Movies. Representation of local geometry in the visual system. Approximation of viscosity solutions of elliptic partial diﬀerential equations on minimal grids. Comm. V. Springer. IEEE Computer Society Press.. Vol. G. M. Vienna. [237] H. Vol. Kimmel. Proc. A. 1996. N. Springer. Berlin.J. 638–641. Numer. Vol. Scalespace theory in computer vision. Optic ﬂow: a curve evolution approach. [236] G. Linear and quasilinear equations of z parabolic type. 350–355. L. Proc. IEEE Trans.J. G. Aﬃne diﬀerential signatures for gray level images of planar shape. Vol. Koenderink. 212–223. Sochen. 345–348. The Hague. Int. Vol.. Shapes. Soc. J. Koenderink. Cybern. 1988. Biol. 1996. SIAM J.B. Vision. C. [245] A. Pattern Recognition (ICPR 11. 1984. Malandain. J. 1997. Kin.Zucker.. On the geometry of texture. Krissian. J. 31. Kimia. Sato. 1995. Directional anisotropic diﬀusion applied to segmentation of vessels in 3D images. S. U.W. A. 1997. N. Lawrence Berkeley National Laboratory.S. A. Solonnikov. Vision. Cybern. and deformations I: The components of twodimensional shape and the reaction–diﬀusion space. N. Kramer. 363–370. 367–375. 24. [243] H.). Koepﬂer. 7. Viergever (Eds. 3. Tannenbaum. Int. The classical case. Anisotropic nonstationary image estimation and its applications: Part I – Restoration of noisy images. Comput. 11th Int. Kiryati.E. M. Wilson.. G. Computer Vision and Pattern Recognition (CVPR ’97. [231] R. Preprint LBNL39640. 1992. Pattern Recognition (ICPR 13. 1997. 72. 1994.N. [234] R. Math.J. Appl. 37–56. Berlin. van Doorn. Lecture Notes in Computer Science. Pattern Recognition. San Juan. L. [233] R. 25–30. 73–92. J. Vol. Vol. 1252. J. Kimmel. Numer. Kumar. Conf.H.P. IEEE Comp.B. Tannenbaum.. ter Haar Romeny. Kocan. Bruckstein.). 53–58. Koenderink. Zucker. 1983. Scalespace theory in computer vision. color. R. 1992). Ural’ceva. 388–397. A. 163. Tannenbaum. Florack. J. A multiscale algorithm for image segmentation by variational method. B. Vol.. Lecture Notes in Computer Science. [232] R. J. Malladi. Malladi. Proc.A. 13th Int. Vol. Lopez. A. 58. Scaletime. Kimia. 1996.J. Knutsson. shocks. A. Berkeley. [240] J. Intrinsic scalespace for images on surfaces: The geodesic curvature ﬂow. Los Alamitos.A. Biol. Vol. [241] J. [230] B. Aug. Ladyˇenskaja. [242] J. Image Proc. Biol. Analyzing and synthesizing images by evolving curves with the OsherSethian method. Conf. American Mathematical Society. Balas. Kimmel. Vol. IEEE Trans. Providence.
Bruckstein. 1990. Image Proc. [263] P. 1977. Kluwer. 1–8. Pichot.). F. R. 987–995. Intell. Rosenfeld. Scalespace for discrete signals. D. Boston. Sporring.H. Vol. 1872–1873. Mach. 75–97. Lifshitz. Pattern Anal.S. The relationship between diﬀusion ﬁltering and Gibbs smoothing. Mach.. A multiresolution hierarchical approach to image segmentation based on intensity extrema. 27. . Springer. Fagerstr¨m. 7. IEEE Trans.H. 1029–1037. 1990. IEEE Trans.M. [260] M.A. [253] Y. Chen. 14. Lindeberg. Nielsen. SPIE Vol. CA 90405. Vol.. Vol. 529–540. ter Haar Romeny (Ed. C. IEEE Trans. Loew (Ed. 1995. B. A computational algorithm for minimizing total variation in image enhancement. P. 1064. Shapeadapted smoothing in estimation of 3D depth cues from aﬃne distortions of local 2D brightness structure. Dept.L. G˚ arding. Intell. 1994. 1992.BIBLIOGRAPHY 153 [247] D.. Axiomatic derivation of image processing models. Kluwer. 1996. Image Vision Comput. Vol. Vol. S. Geometrydriven diﬀusion in computer vision. IEEE Trans. 1994.). Rudin. [255] T. Annual Int. A.M.. Scalespace theory in computer vision. NC 27599. U. M. 229–240.A. B. S.). IEEE Trans. [256] T. S. [257] T. Sgallari. Zucker. L. Chapel Hill. L. [264] M. Sitta.. 12. On Gabor’s contribution to image enhancement. Li. Lobel. Image processing. Chen. On discontinuityadaptive smoothness priors in computer vision. Meth. of Biomedical Engineering. Improvements to the anisotropic diﬀusion model for 2D echo image processing. Nonlinear diﬀusion with multiple edginess thresholds. Osher. Buxton. Math. 27. 1997. BlancF´raud. Rosenman. Vol. Lions. Volume I. J. 147– 154. Variational problems with a free discontinuity set. Fischer. J. Santosa. Scalespace with causal time direction. of the IEEE Engineering in Medicine and Biology Society. 1997. [249] A. Pattern Recognition. 576–586. [254] L.. Suite 101. 2167. M.Z. Vol. Leaci. 1997. Gaussian scalespace theory. M. 543–550. Mach. L. [259] T. [258] T. S. 13.I. Inverse Problems. 1994. Iterative enhancement of noisy images.L. A clinical tool for enhancement of portal images. [252] X. Lev. University of North Carolina. Lindeberg. 1994. Cognitech. 415– 434. J. Computer vision – ECCV ’96. Lalush. Denoising and deblurring images using constrained nonlinear partial diﬀerential equations. Berlin. Lecture Notes in Computer Science.. Dordrecht. Sci. 1993. On the axiomatic foundations of linear scalespace. 2800 – 28 St. Vol. [262] P. Solimini. Man and Cybernetics. [261] P. personal note. 234–254. Santa Monica. Johansen (Eds. Pattern Recognition.). Cipolla o (Eds. Barlaud. Lamberti. Vol. Vol. 1994. 1994.. Pattern Anal. Conf. J.S. Pizer. Florack. [248] C. M. Vol. M. 435–442. Lindeberg.. Systems. 403–410.S.. Intell. Dordrecht. 15. Vol. Lindeberg. A new regularization scheme for inverse e scattering. Inc. Lindeberg. F. Li. Lindenbaum. Pattern Anal. 4. Preprint. Loew. Proc. 5. 12. [250] A. A. 17. Kluwer. Lions. Li.M. T. Mod. U. 1996. 1994. 467–475. [251] S. Appl.
Fifth Int. B. An integrated multiscale approach for terrain referenced underwater navigation. Sethian. Graphical Models and Image Processing. Intell. A representation theory for morphological image and signal processing. IEEE Trans. 16–19. V. Medical Image Analysis. 646–647. Stuttgart.C. Random sets and integral geometry. June 20–23. IEEE Comp. 5. 1. J. Kimmel. Vol. Interim report of the project DPH 6473/91 – Coupling of NavierStokes and Boltzmann regions. [269] A. on Computer Vision (ICCV ’95. Matheron. Terzopoulos. Computer Vision and Pattern Recognition (CVPR ’93. Proc. 187–217. IEEE Trans. Conf. [270] D. New York. Image processing: ﬂows under min/max curvature and mean curvature. Image Proc. Los Alamitos. Proc.. 1993. 197–462. King. 11. .A.S. M. INRIA. 38. Vol. IEEE Trans.). Vol. Marr. Zhang. Lausanne. 1989.A. San Francisco. Pattern Anal. Neunzert.I. Inverse und schlecht gestellte Probleme. Maragos. Conf. R. ROBOTVIS. D. Caselles. Terzopoulos. Deformable models in medical image analysis: A survey. [277] G. Vol. 1990. Struckmeier. Vol. 41. 1993). Proc. IEEE Workshop on Mathematical Models in Biomedical Image Analysis (MMBIA ’96. 1995). Alvarez. 1996. 1996. Sethian. S. 2800–2806. 57–77. IEEE Computer Society Press. Glick. P. Lions (Eds. D. Laboratory of Technomathematics. Deriche. [267] L. Theory of edge detection. 91–108. 06902 Sophia–Antipolis Cedex. Kaiserslautern. June 15–17. Marroquin. Z. 2. Mach. J. 1993. 207. 922–937.L.. B. Malladi. 1993. V. Deriche. 55. IEEE Computer Society Press. [271] R. 244–252. A geometric approach to segmentation and analysis of 3D medical images. McInerney. Vemuri. [273] R. 1994. 1995. Maragos. Los Alamitos. Adalsteinsson. Sept. [275] P. [279] J. Cambridge. Lopez. J. Axiomatization of shape analysis and applications to texture hyperdiscrimination.. 127–141. France. Sapiro. Rigaud. Signal Processing.154 BIBLIOGRAPHY [265] C. IEEE Computer Society Press. Vol. [280] G. 586–599. 633–636. 1996. 17. [268] L. Louis. Ciarlet. Proc. BP 93. Conf. Mach. Malladi.K. 1989. Morel. J. 2004 route des Lucioles. [266] A. Splitting and alternating direction methods. London Ser.G. McInerney. H. 1995. Maragos.A. Internal Report. [276] P. 58. June 21–22. Pattern Anal. L. 1994. 1980. J. IEEE Trans. [278] D. Vol. R.. Vol. [272] R. Hildreth. Malladi. J. Morphological systems: Slope transforms and max–min diﬀerence and diﬀerential equations. Vol. E. Vol. I. 408–417. Luo. R. e e e e 3192. G. Opderbecke. Sur quelques sch´mas num´riques de e e r´solution d’´quations aux d´riv´es partielles pour le traitement d’images. [281] T. R. Local geometry variable conductance diﬀusion for postreconstruction ﬁltering. 1975.L. [274] P. Rigaud. Vol. [282] T. 158–175. Wiley. Marchuk. 1996). CVGIP: Graphical Models and Image Processing. Shape modeling with front propagation: A level set approach. Nuclear Sci. Los Alamitos. Report No.M. 1996. 1996. Lucido. J. Teubner. 840–845. New York. IEEE Int. Lucido. Soc. Intell. Handbook of numerical analysis. V. Proc. Diﬀerential morphology and image processing. 1996). D. 1997.A. Sethian. Soc. Deterministic interactive particle models for image processing and computer graphics. Image Processing (ICIP–96. Topologically adaptable snakes. Lukschin.
BIBLIOGRAPHY
155
[283] T. McInerney, D. Terzopoulos, Medical image segmentation using topologically adaptable surfaces, J. Troccaz, E. Grimson, R. M¨sges (Eds.), CVRMed–MRCAS ’97, Lecture Notes o in Computer Science, Vol. 1205, Springer, Berlin, 23–32, 1997. [284] T. Meis, U. Marcowitz, Numerische Behandlung partieller Diﬀerentialgleichungen, Springer, Berlin, 1978. [285] B. Merriman, J. Bence, S. Osher, Motion of multiple junctions: a level set approach, J. Comput. Phys., Vol. 112, 334–363, 1994. [286] A.R. Mitchell, D.F. Griﬃths, The ﬁnite diﬀerence method in partial diﬀerential equations, Wiley, Chichester, 1980. e [287] L. Moisan, Analyse multi´chelle de ﬁlms pour la reconstruction du relief, C. R. Acad. Sci. Paris, t. 320, S´rie I, 279–284, 1995. e [288] L. Moisan, Perspective invariant movie analysis for depth recovery, L.I. Rudin, S.K. Bramble (Eds.), Investigative and trial image processing, SPIE Vol. 2567, 84–94, 1995. [289] L. Moisan, Aﬃne plane curve evolution: a fully consistent scheme, Preprint No. 9628, CEREMADE, Universit´ Paris IX – Dauphine, Place du Mar´chal de Lattre de Tassigny, e e 75775 Paris Cedex 16, France, 1996. [290] F. Mokhtarian, Multiscale contour segmentation, B. ter Haar Romeny, L. Florack, J. Koenderink, M. Viergever (Eds.), Scalespace theory in computer vision, Lecture Notes in Computer Science, Vol. 1252, Springer, Berlin, 296–307, 1997. [291] F. Mokhtarian, A.K. Mackworth, A theory of multiscale, curvaturebased shape representation for planar curves, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 14, 789–805, 1992. [292] J.M. Morel, S. Solimini, Variational methods in image segmentation, Birkh¨user, Boston, a 1994. [293] K.W. Morton, D.F. Mayers, Numerical solution of partial diﬀerential equations, Cambridge University Press, Cambridge, 1994. [294] T.S. Motzkin, W. Wasow, On the approximation of linear elliptic diﬀerential equations by diﬀerence equations with positive coeﬃcients, J. Mathematics and Physics, Vol. 31, 253–259, 1952. [295] D. Mumford, J. Shah, Boundary detection by minimizing functionals, I, Proc. IEEE Comp. Soc. Conf. Computer Vision and Pattern Recognition (CVPR ’85, San Francisco, June 19–23, 1985), IEEE Computer Society Press, Washington, 22–26, 1985. [296] D. Mumford, J. Shah, Optimal approximations by piecewise smooth functions and associated variational problems, Comm. Pure Appl. Math., Vol. 42, 577–685, 1989. [297] M. Nagao, T. Matsuyama, Edge preserving smoothing, Computer Graphics and Image Processing, Vol. 9, 394–407, 1979. [298] P. Neskovic, B.B. Kimia, Threedimensional shape representation from curvature dependent surface evolution, Proc. IEEE Int. Conf. Image Processing (ICIP–94, Austin, Nov. 13–16, 1994), Vol. 1, IEEE Computer Society Press, Los Alamitos, 6–10, 1994. [299] H. Neunzert, B. Claus, K. Rjasanowa, R. R¨sch, J. Weickert, Mathematische Werkzeuge in o der Bildverarbeitung zur Qualit¨tsbeurteilung von Oberﬂ¨chen, K.H. Hoﬀmann, W. J¨ger, a a a T. Lohmann, H. Schunck (Eds.), Mathematik – Schl¨ sseltechnologie f¨ r die Zukunft, u u Springer, Berlin, 449–462, 1997.
156
BIBLIOGRAPHY
[300] T.G. Newman, H. Dirilten, A nonlinear transformation for digital picture processing, IEEE Trans. Computers, Vol. 22, 869–873, 1973. [301] E. Nicolet, M. Sp¨ hler, Charakterisierung von unscharfen blobf¨rmigen 3D Strukturen, u o diploma thesis, Communication Technology Laboratory, Image Science Division, ETH– Zentrum, 8092 Z¨ rich, Switzerland, 1996. u [302] M. Nielsen, From paradigm to algorithms in computer vision, Ph.D. thesis, Report DIKU 95/8, Dept. of Computer Science, University of Copenhagen, Universitetsparken 1, 2100 Copenhagen, Denmark, 1995. [303] M. Nielsen, L. Florack, R. Deriche, Regularization, scalespace and edge detection ﬁlters, J. Math. Imag. Vision, Vol. 7, 291–307, 1997. [304] W.J. Niessen, B.M. ter Haar Romeny, L.M.J. Florack, M.A. Viergever, A general framework for geometrydriven evolution equations, Int. J. Comput. Vision, Vol. 21, 187–205, 1997. [305] W.J. Niessen, B.M. ter Haar Romeny, M.A. Viergever, Numerical analysis of geometrydriven diﬀusion equations, B.M. ter Haar Romeny (Ed.), Geometrydriven diﬀusion in computer vision, Kluwer, Dordrecht, 392–410, 1994. [306] W.J. Niessen, R. Maas, Multiscale optic ﬂow and stereo, J. Sporring, M. Nielsen, L. Florack, P. Johansen (Eds.), Gaussian scalespace theory, Kluwer, Dordrecht, 31–42, 1997. [307] W.J. Niessen, K.L. Vincken, J. Weickert, M.A. Viergever, Nonlinear multiscale representations for image segmentation, Computer Vision and Image Understanding, Vol. 66, 233–245, 1997. [308] W.J. Niessen, K.L. Vincken, J. Weickert, M.A. Viergever, Threedimensional MR brain segmentation, Proc. Sixth Int. Conf. on Computer Vision (ICCV ’98, Bombay, Jan. 4–7, 1998), in press. [309] M. Nitzberg, D. Mumford, T. Shiota, Filtering, segmentation and depth, Lecture Notes in Computer Science, Vol. 662, Springer, Berlin, 1993. [310] M. Nitzberg, T. Shiota, Nonlinear image ﬁltering with edge and corner enhancement, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 14, 826–833, 1992. [311] N. Nordstr¨m, Biased anisotropic diﬀusion  a uniﬁed regularization and diﬀusion apo proach to edge detection, Image Vision Comput., Vol. 8, 318–327, 1990. [312] E. Norman, A discrete analogue of the Weierstrass transform, Proc. Amer. Math. Soc., Vol. 11, 596–604, 1960. [313] O.F. Olsen, Multiscale watershed segmentation, J. Sporring, M. Nielsen, L. Florack, P. Johansen (Eds.), Gaussian scalespace theory, Kluwer, Dordrecht, 191–200, 1997. [314] P.J. Olver, G. Sapiro, A. Tannenbaum, Classiﬁcation and uniqueness of invariant geometric ﬂows, C. R. Acad. Sci. Paris, t. 319, S´rie I, 339–344, 1994. e [315] P.J. Olver, G. Sapiro, A. Tannenbaum, Aﬃne invariant detection: edges, active contours, and segments, Proc. IEEE Comp. Soc. Conf. Computer Vision and Pattern Recognition (CVPR ’96, San Francisco, June 18–20, 1996), IEEE Computer Society Press, Los Alamitos, 520–525, 1996. [316] P.J. Olver, G. Sapiro, A. Tannenbaum, Invariant geometric evolutions of surfaces and volumetric smoothing, SIAM J. Appl. Math., Vol. 57, 176–194, 1997.
BIBLIOGRAPHY
157
[317] S. Osher, L. Rudin, Featureoriented image enhancement using shock ﬁlters, SIAM J. Numer. Anal., Vol. 27, 919–949, 1990. [318] S. Osher, L. Rudin, Shocks and other nonlinear ﬁltering applied to image processing, A.G. Tescher (Ed.): Applications of digital image processing XIV, SPIE Vol. 1567, 414–431, 1991. [319] S. Osher, J.A. Sethian, Fronts propagating with curvaturedependent speed: Algorithms based on Hamilton–Jacobi formulations, J. Comput. Phys., Vol. 79, 12–49, 1988. [320] N. Otsu, Mathematical studies on feature extraction in pattern recognition, Ph.D. thesis, Report No. 818, Electrotechnical Laboratory, 114, Umezono, Sakuramura, Niiharigun, Ibaraki, Japan, 1981 (in Japanese). [321] K. Ottenberg, Modelbased extraction of geometric structure from digital images, Ph.D. thesis, Utrecht University, The Netherlands, 1993. [322] N. Paragios, R. Deriche, A PDEbased levelset approach for detection and tracking of moving objects, Proc. Sixth Int. Conf. on Computer Vision (ICCV ’98, Bombay, Jan. 4–7, 1998), in press. [323] E.J. Pauwels, P. Fiddelaers, L.J. Van Gool, Enhancement of planar shape through optimization of functionals for curves, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 17, 1101–1105, 1995. [324] E.J. Pauwels, L.J. Van Gool, P. Fiddelaers, T. Moons, An extended class of scaleinvariant and recursive scale space ﬁlters, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 17, 691–701, 1995. [325] E. Payot, R. Guillemaud, Y. Trousset, F. Preteux, An adaptive and constrained model for 3D Xray vascular reconstruction, P. Grangeat, J.L. Amans (Eds.), Threedimensional image reconstruction in radiation and nuclear medicine, Kluwer, Dordrecht, 47–57, 1996. [326] P. Perona, J. Malik, Scale space and edge detection using anisotropic diﬀusion, Proc. IEEE Comp. Soc. Workshop on Computer Vision (Miami Beach, Nov. 30 – Dec. 2, 1987), IEEE Computer Society Press, Washington, 16–22, 1987. [327] P. Perona, J. Malik, A network for multiscale image segmentation, Proc. IEEE Int. Symp. Circuits and Systems (ISCAS–88, Espoo, June 7–9, 1988), 2565–2568, 1988. [328] P. Perona, J. Malik, Scale space and edge detection using anisotropic diﬀusion, IEEE Trans. Pattern Anal. Mach. Intell., Vol. 12, 629–639, 1990. [329] P. Perona, T. Shiota, J. Malik, Anisotropic diﬀusion, B.M. ter Haar Romeny (Ed.), Geometrydriven diﬀusion in computer vision, Kluwer, Dordrecht, 72–92, 1994. [330] P. Perona, M. Tartagni, Diﬀusion networks for onchip image contrast normalization, Proc. IEEE Int. Conf. Image Processing (ICIP–94, Austin, Nov. 13–16, 1994), Vol. 1, IEEE Computer Society Press, Los Alamitos, 1–5, 1994. [331] I.G. Petrowski, Vorlesungen uber partielle Diﬀerentialgleichungen, Teubner, Leipzig, 1955. ¨ [332] I. Pollak, A.S. Willsky, H. Krim, Scale space analysis by stabilized inverse diﬀusion equations, B. ter Haar Romeny, L. Florack, J. Koenderink, M. Viergever (Eds.), Scalespace theory in computer vision, Lecture Notes in Computer Science, Vol. 1252, Springer, Berlin, 200–211, 1997. [333] E.S. Posmentier, The generation of salinity ﬁnestructure by vertical diﬀusion, J. Phys. Oceanogr., Vol. 7, 298–300, 1977.
L. Boston.A. Dordrecht. R. Oosterlinck. 1993.B. Schunck. 1992. 439–451. Mitter. Math. CVGIP: Graphical Models and Image Processing. Physica D. Conf. Computing oriented texture ﬁelds. Computing with reaction–diﬀusion systems: applications in image processing. Geometrydriven diﬀusion in computer vision. R. Rudin. Pauwels. R. Santa Monica. [345] L. B. Ruuth. Total variation based image restoration with free local constraints. P.). Price. blurred images. E. T. C. 680–681. Osher. e Lecture Notes in Control and Information Sciences. P. 2800 – 28 St.. Gar¸on. University of Kaiserslautern. L. London. B. J. Los Alamitos. Intell. Proc. 1996. IEEE Int. [337] M. [336] M. Wambacq. New York. [341] T. Pauwels. Oosterlinck. project c work. Continuation and bifurcations: Numerical techniques and applications.H. [343] L. 1992. [335] C. Cipolla (Eds. 1. Los Alamitos. U.I. Conf. Adaptive smoothing: a general tool for early vision. Berlin. Osher. Pattern Anal. Soc. 207–217.. Image Processing (ICIP–94.I. 1994. and distortion in the variational formulation.). Jaﬀr´. 13–16.I.M. IEEE Comp. Morel (Eds. Van Gool. E. Image Processing and its Applications (Warwick. Vol. June 15–17. Deriche. Deriche. [349] P. Buxton. 379–387. [346] L.. 1994. Preprint. Kluwer. Cognitech. Vol.J. IEE. Wambacq.M. G. B. Oosterlinck. I. Coupled geometrydriven diﬀusion equations for lowlevel vision. IEE Third Int. 31–35. ICAOS ’96: Images. S. Fatemi.158 BIBLIOGRAPHY [334] C. Herlin. IEEE Trans. 53. 169–190. B. Chen. Geometrydriven diﬀusion in computer vision. Inc. (Eds. 295–304.O. Austin. 1996. Saad. IEEE Computer Society Press. Fu. 13. Computer Vision and Pattern Recognition (CVPR ’93.J. 514–529. [342] J. 1994. 1990. 1989. 801. E. CA 90405. A. PWS Publishing. Suite 101. Rieger. Proesmans. Nov. Applications of reaction–diﬀusion equations to image processing. 1994). Mach. 1990. computation. An algorithm for generating motion by mean curvature. Determination of optical ﬂow and its discontinuities using nonlinear diﬀusion. London. M. Dordrecht.R. 1994. IEEE Computer Society Press.J.J. Nonlinear anisotropic diﬀusion ﬁltering of 3D images. A. J. 1996. Kluwer. Medioni. Rudin. Lecture Notes in Computer Science. 82–91. Image enhancement using nonlinear diﬀusion.J.G. Springer. Berlin.B. Rao. 1991. Springer. 1993). ter Haar Romeny (Ed. Pauwels. Volume I.O. E. Proesmans. J. D. Vol. Vol. S. Generic evolution of edges on families of diﬀused greyvalue surfaces. S. 18–20 July 1989). Proc. Vol. 49–53.).M. [340] A. Kluwer.. Moons. J. INSA de Rouen and Laboratory of Technoe e e mathematics. 191–228. [347] S. Van Gool. Computer vision – ECCV ’96. Nonlinear total variation based noise removal algorithms. L. Rambaux. Vol. ter Haar Romeny (Ed. . Springer. 219. Oosterlinck. 157–185.S. Lecture Notes in Computer Science. Roose et al.). A. Approximation. [344] L. [339] I. [348] Y. Vol. 5. Total variation based restoration of noisy. Dense depth map reconstruction: A minimization and regularization approach which preserves discontinuities. J.). [338] M.). 259–268. wavelets and PDEs.S. Richardson. Vision. D´partement G´nie Math´matique. 60. 1064. A. Osher. Eklundh (Ed. Computer Vision – ECCV ’94. Rudin. Berger. Imag. 1995. P. Price. Robert. Proesmans. 1994. Vol. S. Conf. Iterative methods for sparse linear systems. Van Gool. SaintMarc.
Vol. [354] A. [355] G. Tannenbaum. A subdivision scheme for continuousscale Bsplines and aﬃneinvariant progressive smoothing. Salden. Vision. Sapiro. 13–16. Vol. Anal. Vision. [363] G. Sept. Kimmel.M.H. [357] G. [351] A. 238–268. Sapiro. Comput. Sapiro. Nov. Tannenbaum. to appear. 1363–1372. Dept. Vol 135. [365] G. 4. . Caselles. 25–44. University of Minnesota. 472–476. Pattern Anal. Heidelberglaan 100. 9–12. Image Processing (ICIP–96. J. 1996. 1993). Viergever. A. Vol. Math. Bruckstein. 1984. Vol. [352] A. IEEE Trans. Lausanne. Viergever.BIBLIOGRAPHY 159 [350] A. 67–72. IEEE Int. 7. Rueckert.. Dynamic scalespace paradigms. Edge preserving geometric smoothing of MRI data.02.D.A. 1996). A. [361] G. Aﬃne invariant scalespace.. 1994). 1997. J. Sapiro. Proc. Anisotropic diﬀusion of multivalued images with applications to color ﬁltering. Vol. Medical Imaging. Knowledgebased anisotropic diﬀusion of a vectorvalued 4dimensional cardiac MR images. 1996). Vision. Report 3DCV 9228. 1582–1586. 189–198. Tannenbaum. [359] G. MN 55455. D. V. A. ter Haar Romeny. 3D Computer Vision. J. J. Salden. Kimia. Seventh British Machine Vision Conference (BMVC 1996. Sapiro. M.. Y. 1997. [360] G.L. A. Image Proc. D. 3584 CX Utrecht. Proc. [358] G. Sapiro. Minneapolis. 23–40. Diﬀerential Equations. The nonlinear rescaling process. [367] C. Burger. Sapiro.A. Proc. A. 16–19. A. M. Cohen.S. Image Processing (ICIP–94. Conf. Volume Image Processing (VIP ’93. 79–120.). Experiments on geometric image enhancement. 1993. AZU.M. 17. Technical Report. Histogram modiﬁcation via diﬀerential equations. thesis. texture and multiscale image segmentation. Edinburgh. Utrecht. Kaveh. Math. [362] G. Sapiro. Salden.H. 119. Vol. IEEE Int. ter Haar Romeny. Sapiro. Imag. [356] G. Area and length preserving geometric invariant scalespaces. 1994. The Netherlands. S´nchez–Ortiz. Vector (self ) snakes: a geometric framework for color. A. 477–480. 1996).H. 1996. Int. Math. Image structure generating normalised geometric scale spaces. Trucco (Eds. 2. Sept. Funct. Samarskij. Schn¨rr. Mach. Imag. Conf. 1. Utrecht University. Tannenbaum. Implementing continuousscale morphology via curve evolution. M. submitted to IEEE Trans.M. 1994. Linear scalespace theory from physical principles. Room E. [366] G. Shaked. P. [353] A. [364] G. Geest & Portig. Los Alamitos. Ph. Intell. IEEE Computer Society Press. 11. E.I.. 5. Sapiro. 1995. Sapiro. Vol. IEEE Trans. Theorie der Diﬀerenzenverfahren. Vision. Proc. Conf. From active contours to anisotropic diﬀusion: connections between basic PDEs in image processing. Sept. A. 1992. B. June 2–4. You.B. Proc. B.222. J. Imag. J.A. D. of Electical Engineering. Image Processing (ICIP–96. Salden. Tannenbaum. 26. Vol. Bruckstein. 1. U. 817–820. 1994.A. 1996. Int. BMVA Press. 16–19. 1994. Vol. R. Conf.L. Leipzig. R. 1993. Austin. Unique reconstruction of piecewise smooth images by minimizing strictly cono vex nonquadratic functionals.M.H. Vol. B. On aﬃne plane curve evolution. Pattern Recognition.B. 1996. 1993. Lausanne. 141–143. IEEE Int. The Netherlands. Fisher. Ringach.
Soc. A. 54. 1591–1595. 201–207. June 18–20. Sethian. computation and curve evolution. IEEE Comp. June 15–18.). 136–142. 1995. B. Properties of energyminimizing segmentations. 239–251. 1996). Springer. Herlin. Conf. IEEE Computer Society Press. B. Segmentation by nonlinear diﬀusion. [383] K. R. Vol. Vol. e Lecture Notes in Control and Information Sciences. Morel (Eds. June 3–6. 219. I. Picard. 315–328. M. Sethian.. Image analysis and mathematical morphology. Berger. Proc. New York. Jaﬀr´. C. 30. 1981. R. New York. Parts of visual form: computational aspects. Computer Vision and Pattern Recognition (CVPR ’93. segmentation and anisotropic diﬀusion. Segmentation of visual motion by minimizing convex nonquadratic functiono als. 34–40. Los Alamitos. Maui. Sci. Shah. 93. 99–111.B. Graphical Models and Image Processing. 1994. 1992). Vol. 1996. Acta Numerica. Vol. Intell. Sethian. IEEE Computer Society Press.160 BIBLIOGRAPHY [368] C. 1992). Shen. 309–395. [382] A. B.. IEEE Comp. Shah. [375] J. 507–513.A. IEEE Comp. Schn¨rr. Kimia. Level set methods. II. IEEE Computer Society Press. Vol. 1996. Shah. [381] J. June 18–20. [385] K. Seneta. Los Alamitos. Acad. 1992. An optimal linear operator for step edge detection. [380] J. Nonnegative matrices and Markov chains. in press.B. IEEE Comp. Mach. 1996. Siddiqi. Cambridge. Los Alamitos.A. Jerusalem. Soc. Los Alamitos.. Soc. Vol. S. J. Sherstinsky. 1993. 112–133. J. Proc. Pattern Recognition (ICPR 12. [370] E.. Kimia. [377] J. IEEE Trans. . Control and Optimization.A. Conf. IEEE Computer Society Press. Shu. 2. 12th Int. Springer. 1996. Conf. Proc. Champaign. [376] J. Conf. Shah. 1982. Convex variational segmentation of multichannel images. Serra. Kimia. San Francisco. IEEE Trans. 661–663. [379] J. Proc. Soc. 17. Mlattice: From morphogenesis to image processing. Siddiqi. Proc. Castan. and applications of level set methods for propagating interfaces. A review of the theory. SIAM J. Computer Vision and Pattern Recognition (CVPR ’92. 1994). [372] J. Proc. Pattern Anal. A nonlinear diﬀusion model for discontinuous disparity and halfocclusions in stereo. Shah. Var. 9–13. 644–647. Los Alamitos. Piecewise smooth approximations of functions. [371] J. ICAOS ’96: Images. 1993). A shock grammar for recognition. algorithms. [374] J. [378] J. Computer Vision and Pattern Recognition (CVPR ’96. IEEE Computer Society Press. 1996).O. o Deriche. 1992. Geometric shockcapturing ENO schemes for subpixel interpolation. Vol. A fast marching level set method for monotonically advancing fronts. 1137–1150. 1992. Soc. London. Siddiqi. 202–207. 1994. Los Alamitos. IEEE Computer Society Press. June 15–17. on Computer Vision and Pattern Recognition (CVPR ’91. [384] K.W. Academic Press. Segmentation by nonlinear diﬀusion. Oct. Shah. Cambridge University Press. Calc. Schn¨rr.W. Computer Vision and Pattern Recognition (CVPR ’96. San Francisco. 1996. 1996. Conf. 1991. Conf.M. 5. IEEE Comp. Image Proc. Vol. [369] C. Nat. [373] J. Graphical Models and Image Processing. wavelets and PDEs. London. A common framework for curve evolution.B. Proc. 1996.
Strong. 1996).A.. B.. R. Blanc–F´raud. 02912. 679–688. [399] H.B. March 1998. 1995. [401] D. Gaussian scalespace theory. M.D. 1252..S. Vol. A. Conclusions from the commodity expert project. 1996. [394] A.B. Scheunders. Vol. Nielsen. [393] E. Vienna.. T. Shu. 1995. M. Weickert. Chan. Tek. Vol. Cambridge. G. R. Computer Vision and Image Understanding. Shockcapturing numerical methods for viscosity solutions of certain PDEs in computer vision: The Godunov. L. [398] S. Raman. Proc. 15. Los Angeles.W. Stansﬁeld. MA. Conf. Bilbro. Vol. Berlin. J. Intell. Sprengel. A. Bruckstein. D.). Cambridge. Sporring. Mach. 523–527. Olstad. 156–162. University of Hamburg. Relation of regularization parameter and scale in total variation based image denoising. J. Proc. 13th Int. 1994. Y. 36. L. P. 91–123. thesis. Whitaker.S. M. Pattern Anal. Inﬁx Verlag. Pizer. Univ. U.M.F. Osher–Sethian and ENO schemes.A. of Computer Science. Germany. 1997. Proc. [396] D.S. [395] P. 1997. The entropy of scalespace. Aug.BIBLIOGRAPHY 161 [386] J. CA 90024. IEEE Trans. Steiner. Artiﬁcial Intelligence Lab.A. Vol. 65. 17.). Snyder. B. Kimia. Witkin. Aubert. Memo 671. 900–904. 1997. Vol. M.. L. Division of Engineering. Pattern Recognition (ICPR 13.). Florack. Kimmel. B. A. 1988. 1997. C. Kimia. of Mathematics.B. 116–127. 25–30.. Van Dyck. Sijbers. M. U. 13th Int. U.H. [387] W.L. Strong. Johansen (Eds. 2167. R. [397] D. T. Univ. 246–258. Vienna. [400] H. Stoll. on Computer Vision (ICCV ’95.. Intell. of Mathematics. A. Image processing. Los Angeles. Han. Verhoye. 1980. Viergever (Eds. G. Kimia. E. Planar shape enhancement and exaggeration. DISKI 123. Dept. Brown University. Sporring. 1996. U. Terzopoulos. Van der Linden. 1994. Dordrecht.M. Image Proc. Dept. A. Sankt Augustin. IEEE Computer Society Press. Spatially and scale adaptive total variation based regularization and anisotropic diﬀusion in image processing. Kass. June 20–23. IEEE Trans. S.F. [391] R. On generalized entropies and scalespace. [388] J. Barlaud. CAM Report 9646. Steen.S. Koenderink. Fifth Int. Technical Report LEMS132. 620–624. [390] J. 25–30. Los Alamitos. CA 90024. Watershedbased segmentation of 3D MR data for volume quantization. Pattern Recognition (ICPR 13. Dept.I. of Technology. Massachusetts Inst. ter Haar Romeny. 1996). Kluwer. in press. Image segmentation by reaction–diﬀusion bubbles. 53–64. Florack. M. Artif. 1996. . Constraints on deformable models: Recovering 3D shape and nonrigid motion. Scalespace theory in computer vision.. Tek. Variational approach for edgee preserving regularization using coupled PDEs. Image relaxation: Restoration and feature extraction. [389] J. Volumetric segmentation of medical images by threedimensional bubbles. 1995). 1995. Chan.I. Entwurf und Analyse nichtlinearer Diﬀusionsverfahren f¨r die Bildveraru beitung. of California. Providence. of California. Aug.M.S. B. CAM Report 967. Conf. P.A. SPIE Vol. Vol. A. Lecture Notes in Computer Science. Springer. Teboul. 1996. Sporring. [392] J. B. Ph. Scalespace and boundary detection in ultrasonic imaging using nonlinear signaladaptive anisotropic diﬀusion. 1996. Magnetic Resonance Imaging. Loew (Ed. Conf.
19. IEEE Trans. Lecture Notes in Control and Information Sciences. M. Tikhonov. IEEE Trans. [404] V. Kluwer. 1994.R. Springer. 37–72. Prentice Hall.O.). ICAOS ’96: Images. [416] J. Berlin. Computer Graphics. 1970.. 1573–1578.). 1996. 1962. Vincken. F. M. Royal Soc. 355–362. Deriche. Florack. Vol. M. [408] K. Intell. Turk. [406] G. Berlin. Sporring. Englewood Cliﬀs. London.. Dordrecht. P. 221–236. Suppl.). 230–237. Hlav´ˇ. Vol. Primicerio (Eds. 148–163. 1997. Poggio. 67653 Kaiserslautern. 25. M. Jaﬀr´. Weickert. 11. Nonlinear diﬀusion scalespaces: From the continuous to the discrete setting. Fasano. Springer. 2563. Pattern Anal. Gew¨hnliche Diﬀerentialgleichungen. Box 3049. Fast numerical methods for total variation minimization in image reconstruction. Neunzert (Ed. Berlin. A. 1995. Vol.D. [414] J. o [413] J. A model for the cloudiness of fabrics. University of Kaiserslautern. 1995. Vogel. Weickert. 4. 221–234. No. Viergever. [420] J. Luk (Ed. Computing. Johansen (Eds. Germany. e wavelets and PDEs. A. Iterative methods for total variation denoising. New York.O. Chichester. 219.M. Varga. Dept. Z. Springer. Gaussian scalespace theory. 1977.A. Vol.R. H. Mach. Pattern Anal. 1952. 1997.E. The chemical basis of morphogenesis. thesis. Weickert. Koster. Walter.E. R. V. J. Vol. Weickert. [417] J. Herlin. Vol. Generating textures on arbitrary surfaces using reaction–diﬀusion. Morel (Eds. S´ra (Eds. Vol.S. Advanced signal processing algorithms. 111–118. Weickert. 1996. on Mathematics in Industry.M. M. . 1976. Seventh European Conf. Trans. Foundations and applications of nonlinear anisotropic diﬀusion ﬁltering. [405] A. SIAM J. J. Torre. T.T. ˇa [418] J. Turing. Intell. Wiley. Wiley–Teubner. Image Proc. Weickert. Germany. 5. 109–120. 17. R. 1996. 970. [419] J. of Mathematics. Anisotropic diﬀusion ﬁlters for image processing based quality control. Scalespace properties of nonlinear diﬀusion ﬁltering with a diﬀusion tensor. Weickert. [412] W. K. Nonlinear scalespaces.). 283–286. Nielsen. 1996.E. [410] C. Walter. Solutions of illposed problems. [421] J. University of Kaiserslautern. I. 8. 289–298.).A. L. Theoretical foundations of anisotropic diﬀusion in image processing. Vogel.. 227–238. J. P.. [415] J. Philosoph. Tait. Probabilistic multiscale image segmentation. Springer. 258–265. Box 3049. On edge detection. 237. Mach. Laboratory of Technomathematics. Proc. 1994. 1996. Weickert. Weickert. 67653 Kaiserslautern.O. Diﬀerential and integral inequalities. Stuttgart. London B. Image recovery using the anisotropic diﬀusion equation. Computer analysis ac of images and patterns. M. Lecture Notes in Computer Science.A. [403] F. [411] W. V. Mech. SPIE Vol. Vol. Report No. Arsenin. P.). Progress in industrial mathematics at ECMI 94. Suppl. Ph. Oman. Vol. Berger. 110. 1. Sci. 1986. Angew. 1996.162 BIBLIOGRAPHY [402] A. 76. Comput. Teubner.L. Torkamani–Azar. IEEE Trans. Anisotropic diﬀusion in image processing.E. Oman. 359–367. 1991.. Multiscale texture enhancement. Math. [409] C. 1996. Matrix iterative analysis. [407] R.
M. ter Haar Romeny.H. B. Weickert. ter Haar Romeny (Ed. 1997). Viergever. 1997). B.). S. on Computer Vision (ICCV ’95. ter Haar Romeny. Kluwer. Sporring. Dordrecht. of Computer Science. 29–31. 93–134. F. ter Haar Romeny. Cambridge.J. [425] J. 16–19. Whitaker. Scalespace theory in computer vision. [424] J. Viergever (Eds. Whitaker. 1997. J. University of Copenhagen. Berlin. Florack. in press. 1993. Weickert. Lecture Notes in Computer Science. to appear. 1997. Vitri` (Eds. H.). Whitaker. [428] J. Weickert. Weickert. Nielsen. Bajcsy (Eds. M. Lecture Notes in Computer Science. Proc. 1995).F. Wien. S. W. Viergever (Eds. W. [436] R. 1994.H. 3–28. 822–827. B. Lausanne. B. L. Imiya. Information processing in medical imaging. 1996. Villanueva. Algorithms for implicit deformable models. Embedded active surfaces for volume visualization. Theoretical foundations of computer vision (TFCV ’96. M. L. Weickert. [435] R.).M. Proc. A. M. Sept.J. Gaussian scalespace theory. Real World Computing (RWC ’97. A. Coherenceenhancing diﬀusion of colour images. [432] H. Berlin. Pattern Recognition and Image Analysis (VII NSPRIA. Sanfeliu. Springer. Weickert. Scalespace has been discovered in Japan. 2167.). Kropatsch. ter Haar Romeny. Weickert. Springer.T.T. Springer. Universitetsparken 1. Conf. Berlin. Solina. Kluwer. Ishikawa.M. 2100 Copenhagen. 1252. Dagstuhl. Berlin. Characterizing ﬁrst and secondorder patches using geometrylimited diffusion. Vectorvalued diﬀusion. 96–103. Florack. Lecture Notes in Computer Science. Geometrydriven diﬀusion in computer vision. Vol.A.T. 1997 IEEE International Conference on Image Processing (ICIP–97. Report DIKU 97/18. Vol. SPIE Vol. Dept. June 20–23. On the history of Gaussian scalespace axiomatics. Klette. Conf.J. W. 1997. Los Alamitos. Symp. Vol. Barcelona. Image Processing (ICIP– 96. Barrett. Geometry limited diﬀusion in the characterization of geometric patches in images. [437] R. IEEE Int. Orientation analysis by coherenceenhancing diﬀusion. Niessen. A. [427] J. Viergever. 1993.BIBLIOGRAPHY 163 [422] J.T. Eﬃcient and reliable schemes for nonlinear diﬀusion ﬁltering. 1994.T. Arndt. 149–167. 260–271. R. B. A. IEEE Trans. Proc. Parallel implementations of AOS schemes: A fast way of nonlinear diﬀusion ﬁltering. 1997. Denmark. March 1998. R. . Lopez. B. 57. H. Ishikawa. Vol.). CVGIP: Image Understanding. Chen.). van Enk. A. submitted. J. G. Vol. 26–29.A. ter Haar Romeny. Whitaker. Weickert. Florack. 111–120. Werner. Image processing. Conservative image transformations with restoration and scalespace properties. March 18–22.M. Tokyo. 1986. M. Zuiderveld. 1997. L.M. to appear. 239–244. Gerig. Springer.). 45–59. Benhamouda.G. Oct. Springer. Recursive separable schemes for nonlinear diﬀusion ﬁlters. Koenderink.M. 1. J. [434] R. 1996). Scalespace theory in computer vision. 1995.). Santa Barbara. Weickert. 1. J. [431] J. B. Image Proc. J. Weickert. [430] J. A semidiscrete nonlinear scalespace theory and its relation to the Perona–Malik paradox. 1997. 1996).J. IEEE Computer Society Press. 687. Proc. Gew¨hnliche Diﬀerentialgleichungen. o [433] R. A review of nonlinear diﬀusion ﬁltering. Dordrecht. Koenderink. D. Johansen (Eds. ter Haar Romeny. [426] J. 1252. Imiya. 340–352. P. Gmitro (Eds. Jan. Loew (Ed. [423] J. 465–468. 1997). Vol. [429] J. Whitaker.. Fifth Int. M. April a 21–25. K. B.
Yanenko. Signal Processing. 1986. Conf. Eighth Int. Firmin. [445] Y. A. Image Processing (ICIP–96. [450] A.L. 2. 139–151. M. 299–308. Computer Graphics. Young. A.M. energy/Bayes/MDL for multiband image segmentation. 1019–1022.M.F.L. 57. Pizer. Tannenbaum. J. Iverson. A. 273–293.W. 1971. Burger. B. Witkin. Vol. 8–12. Witkin. Vision. 1993. [443] G. Anisotropic blind image restoration. 459–471. Signal matching through scalespace.. Young.). Joint Conf. [446] Y. Vol. Vol. Yoo. [441] A. A. 1971. Singapore. Pattern Anal. [447] D. 4.J.T. 2. 1993. Aug. on Artiﬁcial Intelligence (IJCAI ’83. Poggio. Kaveh. D.R. M.S. 1995. Proc. Young. Dobbins. Berlin. Vol. Scaling theorems for zero crossings. 14. Witkin. Lausanne. IEEE Trans. Comput. [439] A. L.. You. A multiscale approach to nonuniform diﬀusion.P. 1993.. Mach.L. Scalespace ﬁltering. Yuille. Springer. region growing. From curve detection to shape description: an outline. CVGIP: Image Understanding. Tannenbaum. 383–386. 99–110.M. Li (Eds.). 44. 1996. Using statistical pattern recognition techniques to control variable conductance diﬀusion.T. A new implementation of discrete multiscale ﬁltering.A. Zhu.N. World Scientiﬁc. Whitaker. 18. Retinal mechanisms. Sept. van Vliet. IEEE Trans.C. B. 135–145. Image Processing (ICIP–96. D. 1983). 1996. Vol. Kimia. Spatial Vision. Vol. Lecture Notes in Computer Science. Coggins. 1996). 134–144. Terzopoulos. Lausanne. W. 1996. theory and applications. [442] N.. M.H. L. Conf. Kass. Image Proc. Yang. 1996. Underwood. [451] D. 1987. [449] R. [452] S. Xu. New York. Gmitro (Eds. Information processing in medical imaging. Vol. S. Vol. Mach. Barrett. Springer. Karlsruhe. 16–19. Intell. Vol. Kass. Vol. 15–25. . Proc. A. The Gaussian derivative model for spatial vision: I. Proc. IEEE Int. Reaction–diﬀusion textures. [448] I. Zhao. 687. 2. Region competition: unifying snakes. J. [453] S. Li. 1996). 1991. Image Vision Comput. H. 25. [440] A. 1. Academic Press. New York. Behavioral analysis of anisotropic diﬀusion in image processing. P. T.A. [444] T. Vol. IEEE Trans. X. 1987.N. Basu. 16–19. 8. 25–39. 884–900. 1539–1553.Z. Int. Structure adaptive anisotropic ﬁltering. No. Kaveh. Vol. The method of fractional steps: the solution of problems of mathematical physics in several variables. Yuille. Pattern Anal. 461–464.L. 5. Sept. IEEE Int. 1. Zucker. M. Intell. A. Iterative solution of large linear systems. S. Vol. Computer vision: systems. Recursive implementation of the Gaussian ﬁlter. You. 1996.164 BIBLIOGRAPHY [438] R. 1983.
34. 3. 53. see invariance axiomatics. 21. 14. 50. 11. 5. 57. 49 blind image restoration. 27. 28. 67. see diﬀusion. 97 continuity equation. 45. 38 areapreserving ﬂows. 44–49 surfaces. 48 basic equation of ﬁgure. see additive operator splitting applications. 26. 129–134 arclength aﬃne. 125–126. 19. 29 morphological scalespace. 110 apriori knowledge. 116. 113. 28. 61. 130 condition number. 124 CPU times. see perimeter shortening ﬂow. 82. 53. 52 blob detection. 95. 45 invariant heat ﬂow. 14 additive operator splitting. 61 CAD. 116. 62. 42 average globality. 124 Mumford–Shah functional. see vectorvalued images commuting operators. 88. 128 coherence direction. 47. see arclength Gaussian scalespace. 65 CED. 33. 56. 12–14. 15. 40 invariant texture segmentation. 5. 73. 30. 119. 28. 30. 33 curve evolution. see diﬀusion colour images. 57. 99. 15 causality. 37.Index apriori estimates. 5. see wellposedness continuous diﬀusion. 48 invariant gradient. 85. 105. 52. 20. 124. 134 convolution theorem. 13. 93. 57. 107–111 aﬃne arclength. 2 continuous dependence. see scalespace balloon force. 36. 46 adaptive mesh coarsening. 25 adaptive smoothing. see diﬀusion contrast enhancement. 61. 12. 45 BUC. 40. 106. see diﬀusion AOS. 19. 76. 106 conjugate gradient methods. 46 contour models. 113 crystal growth. 28. 52 165 . 29 perimeter. 135 active blobs. 26. 102 circular point. 40–45. 52. 32 coherence. 118. 5. 39 closing. 46. 22. 72. 40 Euclidean. 135 computer aided quality control. 38–43. 114–126. 38 curvature of a curve. 44. see shortening ﬂow transformation. 33 of an image. 13 invariance. see transformation AMSS. 106 continuity. 10 corners. 24. 113–126 contrast parameter. 13 bias term. 43 deblurring. 37 Canny edge detector. 128 coherenceenhancing diﬀusion. see aﬃne morphological scalespace anisotropic diﬀusion. see invariance invariant geodesic snakes. 50. 116 correspondence problem. 26. 98. 74 average grey level invariance. 26. 47 curve evolution schemes. 34. 16 convergence. 29. coherenceenhancing central moments. 62 comparison principle.
see diﬀusion elliptic point. 89. 75–95 tensor. 103. 114 focusofattention. 48. 127. 51 o Perona–Malik. 93. 28. 37. 5 diﬀerential inequalities. 39. 36 dilation. 16 ﬂux. 24. 25. 50. 2. 55–137 on a sphere. 117 isotropic. 130 FD. 126 Nordstr¨m. 72. 19. 110. 86–87. 28 ﬂame propagation. 14–27. 87. 127 ﬁngerprint enhancement. 2. 127–134 continuous. 66. 28 diﬀusivity. 16. 44 distance transformation. 44. 30. see ﬁnite diﬀerences feature vectors. 55. 114. 76. 25. 105. 37. 114. 11. 113–137 backward. 2. 58. 11. reaction– diﬀusion erosion. 4. 135 diﬀusion–reaction. 24 Fick’s law. 32. 116 coherenceenhancing. 46 extremum principle. 50. 18–20. 41. 88–95. 116 linear. 11. 106. 107–111 discrete analogue AMSS. 2 forward–backward. 53. 11. 29 geodesic snakes. 37 entropy. 11. 2 semidiscrete. 23. 94 directional splitting. 72 homogeneous. 48 Mumford–Shah. 12 physical background. 113–126 equation. 66. 15. 51 ENO schemes. 98. 50. 2 inhomogeneous. 33 existence. 52. 5 diﬀusion anisotropic. 11 forensic applications. 10 linear diﬀusion. 113. 41 energy. 42 TVminimization. see wellposedness expected value. 27 systems. 29 INDEX edge detection. see computer aided quality control defocusing. 2. 102 energy functional explicit snakes. see perimeter shortening ﬂow. 35. 43. 27. 38 ﬂowline. 14 discrete. 47. 88. 2–22. 113. 30. 27. 73 explicit schemes. 22. 6. 22. 62. 97–111 discrete diﬀusion. 137 external energy. 24. 43. 73. 13. 25. 5. 97–111 edgeenhancing. 87–95. 29 diﬀerencesofGaussians. 47. 85. 2–14. 118 single equations. 58. 136 defect detection. 15. 28. see transformation Eulerian formulation. 32. 27–30. 30. 28.166 deep structure. 102–111. 56 directional diﬀusivities. 10 nonlinear diﬀusion. 13. 73 entropy scalespace. 22. 119 direction. 2. 105 fabrics. 95. 52 . 46. 44. MCM. 102 generalized. 53. 36. 125. see scalespace. 18 Polyakov. see arclength perimeter. 113 ﬁnite elements. 98 edge cost. 55–74. 63. 2. 26. 116. 7 deviation cost. 43 Gaussian. 85. 50. 116 digitally scalable. 5. 57 edge enhancement. 17 diﬀerential invariants. 129 ﬁnite diﬀerences. see shortening ﬂow transformation. 47. see contrast enhancement edgeenhancing diﬀusion. 25. 45. 47 DoGs. 15. see diﬀerencesofGaussians doubly stochastic matrix. see diﬀusion dissipative eﬀects. 27. 34. 77. 63 Euclidean arclength. 51. 111. 24. 30. 62. 55–74 directed. 2 ﬁlter design. 46 for curves. 15. 125 nonlinear. 85–95. 43.
137 image enhancement. see invariance greyscale invariance. 132 interest operator. 63 hysteresis thresholding. 127. 136 Gabor’s restoration method. 40. 62 immediate localization. 57. see invariance halftoning. 35. 81. see invariance isophote. 44 level set methods. 81 irreducibility. 13. 75 greyscale. 31. 26 linear diﬀusion. 45. 25. 48 geodesic curvature ﬂow. 35 pyramid. 124 isometry. 42 invariance aﬃne. 13 jet space. 76. 82–85 sequences. 35. 44 elimination. see image enhancement image sequences. see diﬀusion linear systems. 81 rotational. see LaplacianofGaussian Lyapunov functionals. 8. 45 Huygens principle. 106 generalized entropy. 36 line detection. 3. 16 Japanese scalespace research. see invariance. 34. 29. 65 noncreation. 81 morphological. 119 grey level shift. 53 image colour. 64. 64. 42 Gabor ﬁlters. 43. 76 local extrema creation. 76. 2 histogramme enhancement. 75. 30. 110 linearity. 21. 63. 42 level set. 24. 44 free discontinuity problems. 113–134 image restoration. 98 Markov processes. 11. 65 local scale. 106. 4. 12. 114. 5. 57 LoG. 24 junctions. 41. 37 hardware realizations. 7. 4. 19 implicit schemes. see vectorvalued images image compression. 137 167 information theory. 45 gas dynamics. 99. 3. 13 geodesic. see scalespace smoothing. 81 greyscale. 44. 135 functions. 97 isometry invariance. greyscale reverse contrast. 39. 33 grey level shift invariance. 46 intrinsic heat ﬂows. 56. see entropy generalized ﬁgure. 11 scalespace. 45. 49. 107. 13. 5 lengthpreserving ﬂows. 43 scale. 47. 39. 56 Gauß–Seidel method. 15 Iijima’s axiomatic. 76. 48. 49 gauge coordinates. 42 geometric heat equation. 42–45 average grey level. 13. 28. 10 integration scale. 42. 43. 17. 16 Gaussian derivatives. 33 hyperstack. 11 . 11. 29 fundamental equation in image processing. 98. 10. 43. 38. 55 higher dimensional. 7 illposedness. 66–74. see threedimensional images vectorvalued. 44. 45. 8 Lipschitzcontinuity. see vectorvalued images discrete. 28–30. 50 LaplacianofGaussian. 106 kernel. 8 translation. 48–53. 11. 31. 56 Kramer–Bruckner ﬁlter. 14–30. 33 heat equation. 73 integral geometry. 5. 73. 2–14.INDEX Fourier transformation. 55– 74. 98–102 Markov chains. 63. 3. 30. 62. 119 nonenhancement. see structure tensor internal energy. 39 grassﬁre ﬂow.
131 MegaWave. 45 random variable. 6 redistribution property. 20–24 scalespace properties. 25. 107 Nordstr¨m functional. 102–111. 31 classical. see local extrema nonenhancement of local extrema. 30. see ordinary diﬀerential equations opening. 44. 129 backward. 17 regularizations. 45. 46. 30–37 continuousscale. see diﬀusion nonnegativity. greyscale morphology. 85–95. 136 perimeter aﬃne. 125. 29. 87 positivity. 45. 137 noise elimination. 43. 47. 36. 113 oversegmentation. 113 oceanography. 27. 21. 114. 11. 99 region growing. 126 regularity. 10. 45 minimal surfaces. 106. 45. 37 PDE. 19 morphological invariance. 32 curvaturebased. 111 parameter determination. see scalespace recursive ﬁlters. 107. 61 . 28 nonconvex. 137 partial diﬀerential equations. 15 edge enhancement. 119. 25 pyramids. 8 postprocessing. 53 ordinary diﬀerential equations. 21. 41 Euclidean. 28. see partial diﬀerential equations perceptual grouping. 124 noise scale. 48 reaction–diﬀusion scalespace. 58. 15 illposedness. 29 Marr–Hildreth operator. 19 pixel numbering. 18 ODE. 25. 52. 114 basic idea. 33 predictor–corrector schemes. 93. 32 optic ﬂow. 37–40. 52 optical character recognition. 17. 29 parallel computing. 36. 26. 10. see energy functiono als numerical aspects. see mean curvature motion mean curvature motion. 21. 116–126. 75–95 orientation. 126 min–max ﬂow. 45 mean ﬁeld annealing. 50 Osher–Sethian schemes. 110. 49. 11. see invariance. 11. 109 preprocessing. 25. 72 reaction–diﬀusion bubbles. 23. see local extrema nonlinear diﬀusion. 29. 5 maximal monotone operator. 26 prairie ﬂow. 31 multigrid.168 Markov random ﬁelds. 88. 113. 39 Perona–Malik ﬁlter. 10. 28. 28. 38 monotonicity preservation. 51. 5. 97 positivity preservation. 118. 30–49. 116. 125 quantization. 57 nonmaxima suppression. 30. see extremum principle MCM. 111 recursivity. 62. 119–124 binary. 17 maximum–minimum principle. 27 median ﬁltering. 11. 19 semidiscrete. 97. 44. 129 oscillations. 56. 51. 25. 12. 116. 13 INDEX optimization convex. 76. 37–49 greyscale. 87 piecewise smoothing. 38 medical imaging. see energy functional neural networks. 19. 63. 111. 110 Mumford–Shah functional. 46. 129. 15 noncreation of local extrema. 119. 27 path planning. 14–20. 25. 1–53 particle methods. 126 pseudospectral methods. 137 multiplicative operator splittings. 27. 42–45. 50. 110.
6. 97 scale invariance. 113 semidiscrete. 99. 76. 81– 95 smoothing principles. 99–111 for curves. 116 linearity principle. 13. 47. 49 creation. 35. 56 exaggeration. 29 skeletonization. 128 snakes. 34. 41. 43. 136 stencil size. 44. 30. 47 selfsnakes. 45 . 58. 97 target tracking. 118. 106 stability. 34 discrete nonlinear diﬀusion. 11. 45 shape inclusion principle. 124 software. 39 shapeadapted Gaussian smoothing. see snakes semiimplicit schemes. 35 smoothness. 26. 126. 31 simulated annealing. 4. 11 scale–imprecision space. see invariance scale selection. 6 axioms. 37. 75–95 semidiscrete diﬀusion. 62 stochastic matrix. 29. 62. 41 anisotropic diﬀusion. 58. 25. 42 reaction–diﬀusion. 43 shape. 10 nonlinear diﬀusion. 6. 72. 7. 116 string theory. 53. 50. 44. 46 explicit. 49. see diﬀusion semigroup property. see structure tensor second moment matrix. 110 settheoretic schemes. 88–95 stereo. 48. 106. 35 subpixel accuracy. 42 structure tensor. 41 Euclidean. 76. 20 reverse contrast invariance. 45 oﬀset. see structure tensor segmentation. 34. 56. 137 selfsnakes. 37 slope transform. 10. 27 terrain matching. 50. 52. 59 steadystate. 116. 10 semidiscrete analogue linear diﬀusion. 39. 10. 38 silhouette. 42 for image sequences. 25. 45 semidiscrete linear. 27 row sum. 7 linear diﬀusion. 50 ﬁlters. 25. 27. 4. 97. 29 staircasing eﬀect. 8. 21. 53 dilation–erosion. 6 scalespace aﬃne Gaussian. 62–74 architectural principles. 76. 109 stabilizing cost. 74. 11. 27 rescaling. 7 mean curvature. 6. 48. 63. 44. 22 shapefromshading. 136 structuring element. 52 Stampacchia’s truncation method. 28. 118 projective. 42 Gaussian. 114 remote sensing. 25. 39 morphological equivalent of Gaussian scalespace. 27. 37 recognition. 44.INDEX regularization. 93. 11. 48 implicit. 43 for shapes. 35 nonlinear diﬀusion. 27. 46 geodesic. 34 169 separability. 36 subsampling. see convergence steerable ﬁlters. 13 aﬃne morphological. 98 stopping time. 18. 39. 6 invariance principles. 73 symmetry. 114. 31 cues. 36. 25. 62–74. 7 general concept. 50 shortening ﬂow aﬃne. 42. 102–111. 32 structuring function. 86. 124 segmentation. 36. 137 SOR. 62 scatter matrix. 28. 45. 37 shock capturing schemes. see invariance robust statistics. 10 semidiscrete nonlinear diﬀusion.
76. 49. 125 zerocrossings. 34. 73 vectorvalued images. 58. 61 visual system. 49 minimizing methods. 131 variance. see total variation uniform positive deﬁniteness. 43. 28. 62. 42. 34. 29. 30 segmentation. 30 TV. 27. 44. 53. 17. 42 translation invariance. 136 topological changes. 23. 39. 76. 29 threedimensional images. 135 wood surfaces. 52. see wellposedness van Gogh. 98. 129 uniqueness. 45. 136 viscosity solution. 41. 41. 28. 56 discrimination. 24. 29 wellposedness. 11. 34. 47–49. 43. 77. 47 toppoint. 33 wavelets. 45. 12 total variation. 15. 5. 136 weak membrane model. 25. 58. 4. 47–49. 26. 52. see invariance Turing’s pattern formation model. 49 transformation aﬃne. 24. 40 Euclidean. 127 generation. 50 preserving methods. 30. 46. 38 projective. 45. 20. 62. 43. 93. 50 INDEX .170 texture analysis. 24. 5. 73 enhancement. 46–48.
numerical aspects. A large number of references enables the reader to acquire an uptodate overview of the original literature. A general framework is explored covering wellposedness and scalespace results not only for the continuous.Anisotropic Diﬀusion in Image Processing Joachim Weickert University of Copenhagen Many recent techniques for digital image enhancement and multiscale image representations are based on nonlinear partial diﬀerential equations. Their ﬂexibility allows to combine smoothing properties with image enhancement qualities. B.G. This book gives an introduction to the main ideas behind these methods. but also for the algorithmically important semidiscrete and fully discrete settings. and applications. Teubner Stuttgart . The central emphasis is on anisotropic nonlinear diﬀusion ﬁlters. The presented examples range from applications in medical image analysis to problems in computer aided quality control. and it describes in a systematic way their theoretical foundations.