Computing, Suppl. 11, pp. 221{236, 1996.

THEORETICAL FOUNDATIONS OF
ANISOTROPIC DIFFUSION IN
IMAGE PROCESSING
Joachim Weickert 1
Laboratory of Technomathematics,
University of Kaiserslautern,
P.O. Box 3049,
D-67653 Kaiserslautern, Germany.
weickert@[Link]{[Link]

Abstract
A frequent problem in low-level vision consists of eliminating noise and
small-scale details from an image while still preserving or even enhancing
the edge structure. Nonlinear anisotropic diusion ltering may be one pos-
sibility to achieve these goals. The objective of the present paper is to review
the author's results on a scale-space interpretation of a class of diusion l-
ters which comprises also several nonlinear anisotropic models. It is demon-
strated that these models { which use an adapted diusion tensor instead
of a scalar diusivity { oer advantages over isotropic lters. Most of the
restoration and scale-space properties carry over from the continuous to
the discrete case. Applications are presented ranging from preprocessing of
medical images and postprocessing of uctuating numerical data to visuali-
zing quality relevant features for the grading of wood surfaces and fabrics.
1 Introduction
In recent years, nonlinear diusion methods have proved to be useful in many elds
ranging from medical applications [11, 3, 22] and image-sequence analysis [24, 30]
to computer aided quality control [33] and postprocessing of noisy data [32].
Nevertheless, there seems to remain both practical and theoretical problems:
 In many practical applications, the rst nonlinear diusion technique due
to Perona and Malik [23] is used. This method is well-known to give poor
results for very noisy images. But also more robust regularizations may still
lead to problems at noisy edges, as long as they use scalar-valued diusivities
[33].
1 This work was supported by \Stiftung Volkswagenwerk" and \Stiftung Rheinland{Pfalz f
ur
Innovation"
  For most nonlinear diusion lters, no correct theory is available: questions
of existence and uniqueness of a solution are hardly addressed, almost no
stability analysis is performed, and concerning the steady-state behaviour,
conjectures dominate the eld. As exceptions let us mention the works of
Catte et al. [6], Cottet and Germain [8], and Schnorr [29]. As long as these
questions are unsolved, there is no way for a convincing scale-space interpre-
tation of nonlinear diusion ltering.
The present paper gives a survey of the author's work to address these problems.
To this end, we investigate two regularized, nonlinear anisotropic diusion lters
which use a diusion tensor instead of a scalar diusivity and perform well even
on very noisy images. We shall see that they belong to a general class of diusion
models, which includes certain previous models and for which a correct theory is
available. Finally, we demonstrate their practical use as an image enhancement
tool by discussing many examples.
The outline of the paper is as follows:
Section 2 describes the essential ideas of anisotropic diusion ltering and its
mathematical legitimation. It is illustrated that the proposed models allow contrast
enhancement and oer advantages at noisy edges compared to less sophisticated
diusion techniques. The third section is dedicated to a scale-space interpretation
of anisotropic diusion. We focus especially on the aspects, in which sense a con-
trast enhancing image processing technique can still be regarded as a smoothing
transformation. Section 4 sketches brie y numerical methods for diusion ltering
and in the fth section, we discuss practical applications of this technique. We
conclude with a summary in section 6.
2 Image enhancement by anisotropic diusion
Let the image domain be an open rectangle
:= (0; a1)  (0; a2), ? := @
its
boundary and let an image f (x) be represented by a bounded function f :
! IR.
Then, a ltered version u(x; t) of f (x) with a scale parameter t  0 may be
obtained as the solution of a diusion equation with f as initial condition and
re ecting boundary conditions:
@t u =div (D(ru ) ru) on
 (0; 1) (1)
u(x; 0) = f (x) on
(2)
hD(ru )ru; ni = 0 on ?  (0; 1) (3)
Hereby, n denotes the outer normal and h:; :i the usual inner product. The sym-
metric positive denite diusion tensor D 2 IR22 is chosen to be a function of the
edge estimator ru , where
!
1 jxj2
K (x) :=
22
exp ? 22 ; (4)
u (x; t) := (K  u~(:; t)) (x) ( > 0) (5)
and u~ denotes an extension of u from
to IR2 , which may be obtained by mirroring
at ?. The regularization by convolving with a Gaussian makes the edge detection
insensitive to noise at scales smaller than  and helps to ensure existence and
uniqueness results in a similar way as in [6].
The preceding lter strategy covers a wide class of diusion methods, classic ones
as well as new ones. To see this, we have to specify the diusion tensor. Let us
study three examples:
(a) Linear isotropic diusion ltering
The simplest diusion lter utilizes the unit matrix as diusion tensor:
D (ru ) = I: (6)
This technique goes back to Marr and Hildreth [18], Witkin [37], and Koen-
derink [14]. It is widely used in the image processing community, since it
is equivalent to convolving the original image with Gaussians of increasing
size (when disregarding the boundary conditions). Nevertheless, it smoothes
noise within a region in the same way as it blurs semantically important
structures like edges.
(b) Nonlinear isotropic diusion ltering
In order to avoid blurring of edges, one should construct a diusion l-
ter which reduces the diusivity at those locations which are good can-
didates for being an edge. To this end, consider a decreasing function
g 2 C 1 ([0; 1); (0; 1]), which can be represented on [0; 1) by a convergent
power series and satises g(0) = 1 and g(s) ! 0 for s ! 1. For example,
one may take !
g (s) = exp
?s5
: (7)
55
The theoretically very well-investigated model of Catte, Lions, Morel and
Coll [6] utilizes a diusion tensor of type
D (ru ) = g (jru j)I: (8)
This idea boils down to the linear diusion case (a) in the interior of a region
(jru j ! 0) and inhibits diusion at strong edges (jru j ! 1).
However, the inhibition is isotropic, since all eigenvalues of D are reduced
by the same amount. Thus, noise at edges cannot be removed by permitting
more diusion along the edge than across it. In order to achieve this, we have
to consider anisotropic models.
(c) Nonlinear anisotropic diusion ltering
Anisotropic models do not only take into account the modulus of the edge
detector ru , but also its direction. We construct the orthonormal system
of eigenvectors v1 ,v2 of D such that they re ect the edge structure:
v1 k ru ; v2 ? ru :
 In order to prefer smoothing along the edge to smoothing across it, one
should choose the corresponding eigenvalues 1 and 2 such that
1 (jru j)
2 (jru j)
! 0 for jru j ! 1:
As in the linear and nonlinear isotropic case, we may wish to reduce noise
within a region by means of eigenvalues with
lim 1(s) = slim
s!0
 (s) = 1:
!0 2
Among the numerous ways to construct such a diusion tensor, let us investi-
gate two examples which were proposed by Weickert: Let e' = (cos '; sin ')T ,
and a
b = abT . Then the integration model [31, 33] is given by
D (ru ) :=
2 Z  e
e g(jhru ; e ij) d'; (9)
 ' '  '
 0

and the tensor product model [32] utilizes

q 
r
D ( u ) := g ru
ru : (10)
The main dierence between (9) and (10) is the smoothing behaviour along
edges: for jru j ! 1, (9) gives 2 (jru j) ! 0, whereas (10) yields
2 (jru j) ! 1. Thus, the second proposal is stronger anisotropic. This is not
always an advantage: in some cases, it may also lead to a slightly stronger
rounding of corners [32].
Besides the preceding examples, other anisotropic diusion models can be found in
the literature, see [20, 8, 9]. A related anisotropic image restoration method may
also be constructed using the mean-curvature equation [2].
The lter class (1){(3) with diusion tensors such as (6), (8), (9), (10) is mathe-
matically sound, as it possesses a unique solution in the distributional sense, which
is innitely often dierentiable for t > 0. This distinguishes these lters from non-
linear diusion techniques such as [23] and [21], which are claimed to be ill-posed
[6]. The proof for the case (b) may be found in [6]. It includes (a) and can be
extended in a straightforward way to case (c). For a more general characterization
of the class of diusion lters which satisfy these properties, see [34].
Let us now compare the image restoration properties of the examples (a){(c).
Figure 1(a) consists of a triangle and a rectangle with 70 % of all pixel being
completely degraded by noise. It is taken from the Software package MegaWave
which was developped at the CEREMADE (University Paris IX). All images in the
present paper possess a range within the interval [0; 255], and they are depicted
in such a way that the lowest value is black and the highest one appears white.
In Fig. 1(b) we observe that linear diusion ltering 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, a second
problem appears, the so-called correspondence problem: edges become more and
more dislocated. Thus, once they are identied at a coarse scale, they have to
be traced back in order to nd their true location, a numerically very dicult
problem.
Nonlinear isotropic diusion does not show these correspondence problems, since
edges are hardly aected by the process. On the other hand, they are actually too
less aected: noise at edges remains, as it is to be seen in Fig. 1(c).
Figure 1(d) demonstrates that nonlinear anisotropic ltering shares the advantages
of both beforementioned methods. It combines good noise eliminating properties
of linear diusion with the stable edge structure of nonlinear isotropic diusion.
Anisotropic diusion ltering is an image restoration tool that needs essentially
two natural parameters: a contrast parameter  and a resolution parameter .
Applying the same reasoning as Catte et al. [6], we may relate the stopping time
T to  by choosing T = 0( 2 ).
For fast decreasing diusivities such as (7), an interesting behaviour can be ob-
served: since the diusion at both sides of an edge is much stronger than at the
edge itself, one can investigate contrast enhancement at the edge. Hence, although
the diusivity is always nonnegative, the lter behaves like backward diusion
at edges and like forward diusion within a region. A more detailed analysis of
this phenomenon can be found in [23]. Figure 2 illustrates contrast enhancement
using anisotropic diusion ltering. It depicts a 2D Gaussian-like function with
its isolines before and after processing. It can be observed that two regions with
almost constant grey value evolve which are separated by a fairly steep edge. The
segmentation-like results in Fig. 1(d) and 2(b) indicate also that nonlinear diu-
sion ltering is a useful preprocessing tool making subsequent segmentations very
easy.

3 Scale-space properties
Besides its qualities as an image enhancing tool, anisotropic diusion ltering
can also be regarded as a smoothing transformation which gives a scale-space
representation of the original image. In the following, we will denote by u(x; t) the
unique solution of the lter class (1){(3) with diusion tensors such as (6), (8), (9)
or (10).

3.1 The scale-space concept

In order to get an impression of the concept of scale-spaces, let us recall 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 represent the image at multiple scales. Moreover, by comparing the
 Figure 1: Comparison of dierent diusion lters. (a) Top Left: Test
image,
= (0; 127)2. (b) Top Right: Linear diusion, t = 80. (c)
Bottom Left: Nonlinear isotropic diusion,  = 3:5,  = 3, t = 80.
(d) Bottom Right: Nonlinear anisotropic diusion, tensor product
model,  = 3:5,  = 3, t = 80.

structures at dierent scales, one obtains a hierarchy of image structures, which

eases a subsequent image interpretation.
A scale-space is a representation at a continuum of scales, embedding the original
image f into a family fTtf j t  0g of gradually simplied versions of it. A scale-
space representation has to full several architectural, smoothing (information
reducing) and invariance requirements [1], which we will brie y review now:
The semigroup property is a typical representative for an architectural require-
ment. It states that for t = 0, the scale-space representation gives the original
image f, and the ltering may be split into a sequence of lter banks:
T0 f = f;
Tt+s f = Tt (Ts f ) 8 s; t  0:
 Figure 2: Edge enhancement of anisotropic diusion. (a) Left:
Gaussian-type function,
= (0; 255)2. (b) Right: Filtered, tensor
product model,  = 3,  = 2, t = 1800.

Information reduction arises from the wish that the smoothing transformation
should not create artefacts when passing from ne to coarse representation. Thus,
at a coarse scale, we must not have additional structures which are caused by the
ltering method itself and not by underlying structures at ner scales.
It is desirable that { except for information reduction by smoothing { the scale-
space operator does not alter the image too much, i.e. it should be invariant un-
der many transformations. Examples of such transformations are grey-level shifts,
translations and rotations, but also ane transformations of the space.
The work of Alvarez, Guichard, Lions and Morel [1] shows that every scale-space
fullling some fairly natural architectural, information reducing and invariance
properties is governed by a partial dierential equation (PDE) with the original
image as initial condition.
Examples for scale-space generating PDEs are the linear diusion equation [37, 14]
@t u =
u (11)
and the ane invariant morphological equation [1, 28]
!! 13
@t u = jruj div
ru : (12)
jruj
Morphological transformations possess the property that the ltering result de-
pends only on the level sets of the image and, therefore, they are invariant under
any nondecreasing grey level transformation (grey scale invariance) [1]. On the
other hand, it is evident that a grey scale invariance requirement is not compati-
ble with any contrast dependent image enhancement method. Hence, if one insists
in having scale-spaces that allow contrast enhancement (like in our case), one has
to withdraw morphology.
The goal of this section is to establish the proposed lter class as scale-space trans-
formations. To this end, we shall not focus on further investigations of architectural
requirements, as these qualities are automatically fullled and do not distinguish
nonlinear diusion scale-spaces from other ones. We start with brie y discussing
invariances. Afterwards, we turn to a more crucial task, namely the question, in
which sense our restoration method { which allows edge enhacement { can still
be considered as a smoothing, information reducing image transformation. As this
paper is intended to give a concise survey on this subject, we will focus on the
main ideas. The reader who is interested in the proofs and mathematical details
is referred to a technical report by the author [34].
3.2 Invariances
By the construction of the proposed lters, it is not hard to verify that they are
invariant under grey level shifts and contrast reversions.
If one omits boundary conditions and regards diusion ltering as a pure initial
value problem on an innite domain, it makes also sense to consider translations
and rotations of the image. In this case, anisotropic diusion ltering is also in-
variant under translations and isometric transformations.
For other invariances, the homogeneous Neumann boundary conditions (3) are very
useful: together with the divergence form of the diusion equation, they imply that
the average grey level Z
 :=
1
j
j f (x) dx

(13)
is not aected by nonlinear diusion ltering:
1 Z u(x; t) dx =  8 t > 0:
j
j

This distinguishes diusion scale-spaces from morphological ones. In general, the

latter ones are not of divergence form and do not preserve the average grey level.
3.3 Information reducing properties
3.3.1 Nonenhancement of local extrema
The requirement that a scale-space representation must not amplify local extrema
was rst pointed out by Lindeberg [16] for the linear diusion scale-space. However,
this condition is also satised by nonlinear anisotropic diusion [34]. Let us consider
an arbitrary, but xed time  > 0 and suppose that  2
is a local extremum of
u(:;  ). Then,

@t u(;  )  0; if  is a local maximum,

@t u(;  )  0; if  is a local minimum.
Nonenhancement of local extrema contrasts anisotropic diusion to classical image
enhancing methods such as high-frequency emphasis ([12], pp. 182{183), which do
violate this principle. Although possibly being in the backward diusion region at
edges, nonlinear diusion is always in the forward region at extrema. This ensures
its stability.
3.3.2 Maximum{minimum principle and consequences
An extremum principle is a property which is closely related to the nonenhance-
ment of local extrema. A common way to prove extremum principles for nonlinear
parabolic PDEs is the use of monotony results (see e.g. [25], pp. 186{188). The
monotony condition
f h =) T t f  Tt h 8 t > 0:
has also been proposed as a smoothing requirement for scale-spaces [1]. When
combined with reverse contrast invariance and grey level shift invariance, this
comparison principle implies an extremum principle. However, to establish the
monotony condition for anisotropic diusion ltering, we would need that the
equation is always of forward parabolic type. Since this would forbid contrast
enhancing processes, we will not pursue this idea any further.
But we are not lost: In order to prove a maximum{minimum principle, we may
utilize a technique which does not require monotony, namely Stampacchia's trun-
cation method (cf. [4], p. 211). Suppose that the original image f is bounded by
some a; b 2 IR:
a  f (x)  b 8 x 2
:
Then, Stampacchia's truncation method implies that the ltered image remains
within these bounds for all times [34]:
a  u(x; t)  b on
  (0; 1): (14)
This result is of essential practical importance, as it guarantees for instance that
if we start with an image within a range [0; 255], we will never get a result with a
grey value such as 257.
Hummel [13] shows that, under certain conditions, the extremum principle for
parabolic operators is equivalent to the property that the corresponding scale-space
never creates additional level-crossings for t > 0. This points out the importance
of extremum principles for scale-spaces.
Furthermore, the preceding maximum{minimum principle is a useful tool for es-
tablishing theoretical properties like continuous dependence of the solution on the
initial data. This is an important step for proving the well-posedness of our method.
We denote by Lp(
) with 1  p < 1 the space of functions f :
! IR, for which
jf jp is Lebesgue integrable, and provide it with the norm
Z 1=p
kf kLp(
) :=

jf (x)jp dx :
Then it can be shown that the solution u(x; t) of the diusion lter depends con-
tinuously on the initial image f with respect to the L2 (
) norm [34]. This result is
especially important when applying anisotropic diusion in areas like stereo vision
or analysis of image sequences, as it guarantees that similar images remain similar
after being processed. Although this requirement is hardly mentioned in scale-space
theory, it deserves to be carefully checked. A scale-space violating this property
can be practically completely unstable in spite of having plenty of smoothing and
invariance qualities.

3.3.3 Behaviour for t ! 1

As scale-spaces are intended to simplify an image, it would be desirable, that for
t ! 1, the result converges to the simplest possible image representation, namely
a constant image with the same average grey value as the original one. Unlike
morphological equations such as (12), which may have nontrivial steady-states,
our diusion lter class satises this requirement [34]. The convergence to the
constant image is guaranteed in every Lp norm with 1  p < 1.

3.3.4 Lyapunov functionals

We have seen, that for t ! 1, the diusion ltered image becomes completely
smooth. On the other hand, we obverved that for nite times, locally the opposite
may appear: contrast can be enhanced. Hence the question arises, whether there
exists some quantity which indicates that the image becomes steadily smoother
from a global view.
We have already found such global quantities: The extremum principle tells us
that the global minimum increases and the global maximum decreases with respect
to time. But there exist numerous other quantities as well [34]: all Lp(
) norms
of u(x; t) with 2  p < 1 are decreasing in t. This comprises also the energy
ku(t)k2L2(
) .
Another class of decreasing global qualities is given by the even central moments
[34]
M2n [u(t)] :=
1 Z (u(x; t) ? )2n dx (n 2 IN):
j
j

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 atness of the grey value distribution. Higher moments are
more dicult to interprete, although they do provide important information for
tasks like texture discrimination ([12], pp. 414{415). All decreasing even moments
demonstrate that the image becomes smoother during diusion ltering. Hence,
local eects such as edge enhancement, which object to increase central moments,
are overcompensated by strong smoothing in other areas.
If the initial image f is strictly positive on
, we may regard it also as a two-
dimensional density. (Without loss of generality, we omit the normalization.) Then,
Z
S [u(t)] := ? u(x; t) ln(u(x; t)) dx

is called the entropy of u(t), a measure of uncertainty and missing information

[5]. Anisotropic diusion ltering can be shown to increase the entropy [34], hence
the corresponding scale-space embeds the genuine image f into a family of subse-
quently likelier versions of it which contain less information. Moreover, for t ! 1,
the process reaches the state with the lowest possible information, namely a con-
stant image.
From all the previous considerations, we observe that, in spite of its edge enhancing
properties, anisotropic diusion does really simplify the original image in a steady
way.

4 Numerical approximation
Since a scale-space representation or an enhancement method cannot be better
than its numerical realization, it is of crucial importance to study the discrete case
of diusion ltering.
For diusion problems, numerous numerical methods can be applied:
In [10], a comparison was made between three schemes for a one-dimensional model
of nonlinear diusion ltering: a wavelet method of Petrov{Galerkin type, a spec-
tral method and a nite-dierence (FD) scheme. It turned out that { especially
for large  { all results were fairly similar. Since the computational eort is of a
comparable order of magnitude, it seems to be a matter of taste which scheme
one prefers. Of course, other numerical methods are applicable as well, e.g. nite
elements. Neural network realizations of nonlinear diusion lters were proposed
by Cottet [7, 9] and a multigrid acceleration for a scheme which is related to
nonlinear diusion ltering was studied by Saint-Marc, Chen and Medioni [27].
However, most nonlinear diusion researchers prefer nite dierences, since they
are easy to handle and the pixel structure of a real digital image provides already
a natural discretization on a xed rectangular grid. Thus, in the sequel we restrict
our considerations to the FD case.
Besides low complexity, the requirements for a good numerical scheme for diusion
in image processing dier from requirements in other areas such as computational
uid dynamics: very accurate high-order approximations of the continuous equa-
tion are less important than approximations which exhibit as many qualitative
properties of the continuous equation as possible. So the question arisies which of
the continuous scale-space properties carry over to the FD case.
It is not hard to see that properties like grey level shift invariance or reverse
contrast invariance are easily fullled by every reasonable FD scheme. Translation
invariance and isometry invariance seem to be more problematic, since translations
make only sense for multiples of the grid distance, and isometry invariance can only
be approximatly satised.
In [35], it is investigated under which requirements one can construct semidiscrete
(i.e. continuous in time, discrete in space) and fully discrete anisotropic diusion
lters with the following properties: well-posedness, average grey level invariance,
maximum-minimum principle, existence of Lyapunov functionals, and convergence
to a constant steady-state. It will be shown that it is possible to nd consistent FD
schemes having all these qualities. These comprise also splitting-based, absolutely
stable semiimplicit schemes requiring per time step a computational and storage
eort which is linear in the pixel number.
5 Applications
At the Laboratory of Technomathematics, diusion ltering is currently applied
in the following areas.
(a) Postprocessing of data with numerical uctuations [31]
Particle methods for solving the Boltzmann equation utilize random or pseu-
dorandom processes which lead to numerical uctuations. By means of non-
linear diusion ltering it was possible to smooth \jittery" isolines without
aecting their principal structure. Meanwhile, nonlinear diusion is used as
a tool to process the Bolzmann solution in such a way that it can be coupled
with the smooth solution of the Navier-Stokes equation in adjacent regions.
[17].
(b) Grading of fabrics [33]
The quality of a fabric is determined by two criteria, namely clouds and
stripes. Clouds result from isotropic inhomogenities of the density distri-
bution, whereas stripes are an anisotropic phenomenon caused by adjacent
bers pointing in the same direction. Anisotropic diusion lters are capable
of visualizing both quality-relevant features simultaneously (Fig. 3). For a
suitable parameter choice, they smooth in an isotropic way at clouds and
diuse anisotropically along bres in order to enhance them.
(c) Defect detection of wood surfaces [31, 33]
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 relavant features become better visible und unim-
portant structures disappear. Fig. 4(a) depicts a wood surface possessing one
defect. By means of anisotropic diusion ltering it is possible to visualize
this fault in a much better way (Fig. 4(b)). In [33] it is demonstrated, how a
further modication of the diusion tensor yields even more accurate results
with less roundings at the defect's corners.
(d) Preprocessing of medical images
Fig. 5(a) gives an example for possible medical applications of anisotropic
 Figure 3: Preprocessing of fabric images. (a) Left: Fabric,
=
(0; 256)2. (b) Right: Filtered, tensor product model,  = 4,  = 2,
t = 160.

diusion. It depicts an MR image of the liver. To ease the diagnosis it is

intended to apply segmentation algorithms for classifying the dierent types
of tissue. Fig. 5(c) shows a segmentation according to the Mumford{Shah
energy functional [19] by means of an algorithm due to Koep er, Morel and
Solimini [15]. As it is seen in Fig. 5(d), one can obtain more realistic results
when processing the original image by means of anisotropic diusion (Fig.
5(b)) prior to segmentation. An algorithm for 3D images is currently under
investigation, which allows to lter simultaneously an entire MRI or CT
set [26]. To this end, a version for parallel machines of MIMD type was
developped.

6 Conclusions
We have investigated anisotropic diusion lters which use an adapted diusion
tensor instead of a scalar diusivity. This quality distinguishes them from most
other nonuniform diusion methods such as [23, 21, 6, 36, 22, 29] and oers advan-
tages at noisy edges. The proposed lters belong to a class which is theoretically
well-founded, as it permits existence, uniqueness and regularity results, continuous
dependence of the solution on the initial image, a maximum{minimum principle,
and convergence to a constant steady-state as t tends to innity.
Anisotropic diusion ltering is capable of enhancing images by smoothing isotro-
pically within a region, while diusing in an anisotropic way along edges. Rapidly
decreasing diusivities allow even contrast enhancement at edges.
In spite of this property, anisotropic nonlinear diusion lters turn out to be
smoothing with respect to numerous aspects. Together with certain architectural
 Figure 4: Defect detection in wood. (a) Left: Wood surface,
=
(0; 255)2. (b) Right: Filtered, integration model,  = 4,  = 2:8,
t = 320.

and invariance properties, they induce a scale-space representation of the original

image, which may incorporate user-specic demands on contrast and size of es-
pecially interesting structures. Edges remain well-pronounced and stable across a
wide range of scales.
The presented examples demonstrate that anisotropic diusion is well-suited as a
pre- and postprocessing tool and as a method to visualize quality relevant features
in the area of computer aided quality control (CAQ).
Furthermore, it should be mentioned, that the previous results can be easily ex-
tended to higher dimensions, vector-valued images and more sophisticated descrip-
tors of local structure allowing also corner enhancement [34].

References
[1] L. Alvarez, F. Guichard, P.-L. Lions, and J.-M. Morel. Axioms and fundamental
equations in image processing. Arch. Rational Mech. Anal., Vol. 123: pp. 199{257,
1993.
[2] L. Alvarez, P.-L. Lions, and J.-M. Morel. Image selective smoothing and edge
detection by nonlinear diusion. II. SIAM J. Numer. Anal., Vol. 29: pp. 845{866,
1992.
[3] I. Bajla, M. Marusiak, and M. S ramek. Anisotropic ltering of MRI data based
upon image gradient histogram. In D. Chetverikov, and W.G. Kropatsch (Eds.).
Computer analysis of images and patterns. Springer, Berlin, pp. 90{97, 1993.
[4] H. Brezis. Analyse fonctionelle. Masson, Paris, 1983.
[5] B. Buck, and V. Macaulay (Eds.). Maximum entropy in action. Clarendon, Oxford,
1991.
[6] F. Catte, P.-L. Lions, J.-M. Morel, and T. Coll. Image selective smoothing and edge
detection by nonlinear diusion. SIAM J. Numer. Anal., Vol. 29: pp. 182{193, 1992.
 Figure 5: Preprocessing of MR images. (a) Top Left: Liver,
=
(0; 255)2. (b) Top Right: Filtered, integration model,  = 1,  = 3,
t = 4. (c) Bottom Left: Segmented original image. (d) Bottom
Right: Segmented ltered image.

[7] G.-H. Cottet. Diusion approximation on neural networks and applications for im-
age processing. In F. Hodnett (Ed.). Proc. Sixth European Conference on Mathe-
matics in Industry. Teubner, Stuttgart, pp. 3{9, 1992.
[8] G.-H. Cottet, and L. Germain. Image processing through reaction combined with
nonlinear diusion. Math. Comp., Vol. 61: pp. 659{673, 1993.
[9] G.-H. Cottet. Neural networks: Continuous approach and applications to image
processing. Technical Report No. 113, LMC { IMAG, Universite Joseph Fourier,
B.P. 53, 38041 Grenoble Cedex 9, France, 1994 (submittet to Proc. 2nd Europ.
Conf. on Mathematics applied to Biology and Medicine, Lyon, Dec. 15{18, 1993).
[10] J. Frohlich, and J. Weickert. Image processing using a wavelet algorithm for non-
linear diusion. Report No. 104, Laboratory of Technomathematics, University of
Kaiserslautern, P.O. Box 3049, 67653 Kaiserslautern, Germany, 1994.
[11] G. Gerig, O. Kubler, R. Kikinis and F.A. Jolesz. Nonlinear anisotropic ltering of
MRI data. IEEE Trans. Medical Imaging, Vol. 11: pp. 221{232, 1992.
[12] R.C. Gonzalez, and P. Wintz. Digital image processing. Addison{Wesley, Reading,
1987.
[13] R.A. Hummel. Representations based on zero-crossings in scale space. Proc. IEEE
Computer Society Conf. on Comp. Vision and Pattern Recognition, pp. 204{209,
1986.
[14] J.J. Koenderink. The structure of images. Biol. Cybern., Vol. 50: pp. 363{370,
1984.
[15] G. Koep er, J.-M. Morel, and S. Solimini. Segmentation by minimizing a functional
and the \merging" method. Preprint No. 9022, CEREMADE, Universite Paris IX -
Dauphine, Place du Marechal de Lattre de Tassigny, 75775 Paris Cedex 16, France,
1990.
[16] T. Lindeberg. Scale-space for discrete signals. IEEE Trans. Pattern Anal. Mach.
Intell., Vol. 12: pp. 234{254, 1990.
[17] A. Lukschin, H. Neunzert, and J. Struckmeier. Interim report of the project DPH
6473/91 | Coupling of Navier-Stokes and Boltzmann regions. Internal Report,
Kaiserslautern, 1993.
[18] D. Marr, and E. Hildreth. Theory of edge detection. Proc. R. Soc. Lond. B, Vol.
207: pp. 187{217, 1980.
[19] D. Mumford, and J. Shah. Optimal approximations by piecewise smooth functions
and associated variational problems. Comm. Pure Appl. Math., Vol. 42: pp. 577{
685, 1989.
[20] M. Nitzberg, and T. Shiota. Nonlinear image ltering with edge and corner en-
hancement. IEEE Trans. Pattern Anal. Mach. Intell., Vol. 14: pp. 826{833, 1992.
[21] N. Nordstrom. Biased anisotropic diusion - a unied regularization and diusion
approach to edge detection. Image and Vision Computing, Vol. 8: pp. 318{327,
1990.
[22] K. Ottenberg. Model-based extraction of geometric structure from digital images.
Ph.D. thesis, Utrecht university, The Netherlands, 1993.
[23] P. Perona, and J. Malik. Scale space and edge detection using anisotropic diusion.
IEEE Trans. Pattern Anal. Mach. Intell., Vol. 12: pp. 629{639, 1990.
[24] M. Proesmans, L. Van Gool, E. Pauwels, and A. Oosterlinck. Determination of
optical ow and its discontinuities using non-linear diusion. In J.-O. Eklundh
(Ed.). Computer Vision { ECCV '94. Lecture Notes in Comp. Science, Vol. 801,
Springer, Berlin, pp. 295{304, 1994.
[25] M.H. Protter, and H.F. Weinberger. Maximum principles in dierential equations.
Prentice-Hall, Englewood Clis, 1978.
[26] I. Rambaux, and P. Garcon. Nonlinear anisotropic diusion ltering of 3D images.
Project Work, Departement Genie Mathematique, INSA de Rouen and Laboratory
of Technomathematics, University of Kaiserslautern, 1994.
[27] P. Saint-Marc, J.S. Chen, and G. Medioni. Adaptive smoothing: a general tool for
early vision. IEEE Trans. Pattern Anal. Mach. Intell., Vol. 13: pp. 514{529, 1990.
[28] G. Sapiro, and A. Tannenbaum. Ane invariant scale-space. Intern. J. of Computer
Vision, Vol. 11: pp. 25{44, 1993.
[29] C. Schnorr. Unique reconstruction of piecewise smooth images by minimizing
strictly convex non-quadratic functionals. J. Math. Imag. Vision, Vol. 4: pp. 189{
198, 1994.
[30] C. Schnorr. Bewegungssegmentation von Bildfolgen durch die Minimierung kon-
vexer nicht-quadratischer Funktionale. In W.G. Kropatsch, and H. Bischof (Eds.).
 Tagungsband Mustererkennung 1994. Informatik Xpress 5, Wien, 178{185, 1994.
[31] J. Weickert. Zwischenbericht zum Projekt \Nichtlineare Diusionslter". Bericht
uber die wissenschaftliche Tatigkeit Januar 1991 { Dezember 1991, Center for
Applied Mathematics, Darmstadt { Kaiserslautern, pp. 133{142, 1992.
[32] J. Weickert. Abschlubericht zum Projekt \Nichtlineare Diusionslter". Abschlu-
bericht und Bericht uber die wissenschaftliche Tatigkeit Januar 1992 { Dezember
1993, Center for Applied Mathematics, Darmstadt { Kaiserslautern, pp. 191{209,
1994.
[33] J. Weickert. Anisotropic diusion lters for image processing based quality control.
In A. Fasano, M. Primicerio (Eds.). Proc. Seventh European Conf. on Mathematics
in Industry. Teubner, Stuttgart, 355{362, 1994.
[34] J. Weickert. Scale-space properties of nonlinear diusion ltering with a diusion
tensor. Report No. 110, Laboratory of Technomathematics, University of Kaisers-
lautern, P.O. Box 3049, 67653 Kaiserslautern, Germany, 1994 (submitted).
[35] J. Weickert. Anisotropic diusion in image processing. Ph.D. thesis, Kaiserslautern,
1995 (to be led).
[36] R.T. Whitaker, and S.M. Pizer. A multi-scale approach to nonuniform diusion.
CVGIP: Image Understanding, Vol. 57: pp. 99{110, 1993.
[37] A.P. Witkin. Scale-space ltering. Proc. Eighth Int. Joint Conf. on Articial Intell.,
Karlsruhe, pp. 1019{1022, 1983.