398 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO.

3, MARCH 1998

Efficient and Reliable Schemes

for Nonlinear Diffusion Filtering
Joachim Weickert, Bart M. ter Haar Romeny, Member, IEEE, and Max A. Viergever

Abstract— Nonlinear diffusion filtering is usually performed a recursive scheme. The presented algorithms are applicable
with explicit schemes. They are only stable for very small time in arbitrary dimensions and their computational and storage
steps, which leads to poor efficiency and limits their practical effort is linear in the image size. This shows their efficiency.
use. Based on a recent discrete nonlinear diffusion scale-space
framework we present semi-implicit schemes which are stable for We prove the reliability of these schemes by verifying that
all time steps. These novel schemes use an additive operator split- they satisfy recently established criteria for nonlinear diffusion
ting (AOS), which guarantees equal treatment of all coordinate scale-spaces [45], [46]. This comes down to checking six sim-
axes. They can be implemented easily in arbitrary dimensions, ple criteria. If these requirements are fulfilled we can be sure
have good rotational invariance and reveal a computational that the scheme preserves the average grey value, satisfies a
complexity and memory requirement which is linear in the
number of pixels. Examples demonstrate that, under typical causality property in terms of a maximum–minimum-principle,
accuracy requirements, AOS schemes are at least ten times more reveals a large class of smoothing Lyapunov functionals, and
efficient than the widely used explicit schemes. converges to a constant steady-state as the time tends to infin-
Index Terms—Absolute stability, nonlinear diffusion, recursive ity. It should be noted that the discrete maximum–minimum
filters. principle is a very restrictive stability criterion (more restric-
tive than the von Neumann stability), since it also takes into
account the boundary conditions and guarantees that over- and
I. INTRODUCTION undershoots cannot appear.

I MPRESSIVE results are the main reason for using nonlin-

ear diffusion filtering in image processing: Unlike linear
diffusion filtering (which is equivalent to convolving with
The goal of this paper is to guide the reader in a system-
atic way to these so-called additive operator splitting (AOS)
schemes. Specific knowledge in numerical analysis is not
a Gaussian), edges remain well localized and can even be necessary, as we shall refer to the required material in the
enhanced. Spatial regularizations of this filter class have a solid literature whenever it is needed. However, the reader who
mathematical foundation as well-posed scale-spaces [12], [44], is interested in a more detailed introduction to the matrix
[46], whose parameter influence is well understood [6], [25]. algebra, which is useful for the present paper, may find this in
Poor efficiency is the main reason for not using nonlinear Ortega [35, ch. 6]. As a prototype of a well-founded nonlinear
diffusion filtering: Most approaches are based on the simplest diffusion filter, we focus on a spatial regularization of the
finite difference discretization by means of a so-called explicit Perona–Malik filter [37] by Catté et al. [12], and Whitaker
or Euler-forward scheme. This scheme requires very small and Pizer [50].
time steps in order to be stable. Hence, the whole filtering The paper is organized as follows. Section II gives a brief
procedure is rather time-consuming. survey on this diffusion model (henceforth, the CLMC equa-
In the present paper, we address this problem. We present tion). In Section III, we review the simplest scheme for the
a novel type of separable schemes that do not suffer from one-dimensional (1-D) CLMC equation: the explicit (Euler
any time step size restriction, since all stability-relevant terms forward) discretization in time. We analyze it by means of
are discretized in an implicit manner. The backbone of these criteria for discrete nonlinear diffusion scale-spaces in order
schemes is a Gaussian algorithm for solving a tridiagonal to explain why it requires rather prohibitive time step sizes.
system of linear equations. It is fast, stable and requires only a As a remedy we study a semi-implicit discretization for
few lines programming work. Its forward and backward substi- which we show that it satisfies all discrete scale-space criteria
tution step can be regarded as a causal and anticausal filter of (including stability) even for arbitrary large time steps. It
requires to solve a tridiagonal linear system of equations,
Manuscript received November 1, 1996; revised March 17, 1997. This work which is easily and efficiently done by a special variant of
was sponsored by the Real World Computing Partnership. The associate editor the well-known Gaussian elimination algorithm. This so-called
coordinating the review of this manuscript and approving it for publication
was Dr. Guillermo Sapiro. Thomas algorithm will be presented in detail, since it forms
J. Weickert was with the RWCP Novel Function SNN Laboratory, the core of the whole scheme.
Image Sciences Institute, Utrecht University Hospital, E01.334, 3584 CX In Section IV we consider the higher dimensional case. It is
Utrecht, The Netherlands. He is now with the Department of Computer
Science, University of Copenhagen, 2100 Copenhagen, Denmark (e-mail: argued that the simple explicit scheme leads to even more
[email protected]). restrictive stability conditions than in the 1-D case, while
B. M. ter Haar Romeny and M. A. Viergever are with the Image Sciences the semi-implicit scheme remains absolutely stable. However,
Institute, Utrecht University Hospital, E01.334, 3584 CX Utrecht, The
Netherlands. solving the -dimensional linear system becomes significantly
Publisher Item Identifier S 1057-7149(98)01786-2. less efficient for dimensions .
1057–7149/98$10.00  1998 IEEE

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.
WEICKERT et al.: NONLINEAR DIFFUSION FILTERING 399

As a remedy we present an alternative scheme which is Let be our image domain and
also semi-implicit, has the same approximation order, and is consider a (scalar) image as a bounded mapping from into
absolutely stable, but it can be separated into 1-D processes. the real numbers . Then the CLMC filter calculates a filtered
Thus, the simple and efficient Thomas algorithm can be image of as a solution of the diffusion equation
applied again. Unlike classical multiplicative splitting schemes
from the mathematical literature, we consider an additive div (1)
operator splitting (AOS). It ensures that all coordinate axes
are treated equally, a very desirable symmetry property in with the original image as initial state
the context of image processing. Furthermore, we shall check
(2)
that the AOS schemes satisfy all criteria for discrete nonlinear
scale-spaces. and reflecting boundary conditions
The section is concluded by proposing a related method for
the regularization step within the CLMC model. Since this on (3)
regularization is based on a Gaussian convolution, it is natural
to regard it as a linear diffusion filter for which one may also where denotes the normal to the image boundary .
apply splitting techniques based on the Thomas algorithm. The “time” is a scale parameter: increasing leads to
Section V presents an -dimensional algorithmic formula- simpler image representations. The whole embedding of the
tion of the AOS schemes and analyzes its complexity. original image into such a one-parameter family of simplified
In Section VI, we evaluate the results by checking the images is called scale-space. The first representative of this
performance of AOS schemes with respect to rotational in- very general and useful image processing concept, namely
variance and accuracy. This allows us to propose reasonable linear diffusion filtering, has been derived in an axiomatic way
time step size and to analyze the accuracy and efficiency by Iijima more than 35 years ago [23], [48].
in comparison to the unsplit semi-implicit scheme and the In order to reduce smoothing at edges, the diffusivity is
widely-used explicit scheme. chosen as a decreasing function of the edge detector .
We conclude the paper with a summary in Section VII. A Here, is the gradient of a smoothed version of which
shortened preliminary version of this paper can be found in is obtained by convolving with a Gaussian of standard
[47]. deviation
Related Work: Our work has been influenced by a number
of related approaches which we would like to mention here. (4)
Implicit splitting-based approaches for linear diffusion fil-
(5)
tering have been proposed in [9] and [20] and also in [2],
[3], and [52], where their realization as recursive filters is
suggested. Impressive results on improved efficiency by means We use the following form for the diffusivity:
of recursive filtering can be found in [14] and [15], and
the close relation between recursive filters and linear scale-
(6)
space approaches has been clarified in [32]. Semidiscrete or .
fully discrete analogs of linear diffusion filtering have been
proposed in [4], [26], [34] and [38]. For such rapidly decreasing diffusivities, smoothing on both
In the nonlinear diffusion field, one can find several ap- sides of an edge is much stronger than smoothing across it.
proaches that aim to be efficient alternatives to the con- As a result, the gradient at edges may even be enhanced (see
ventional two-level explicit finite-difference scheme, for in- [37] for more details). plays the role of a contrast parameter:
stance multigrid methods [1], finite element techniques with Structures with are regarded as edges, where the
adaptive mesh coarsening [5], semi-implicit approaches [12], diffusivity is close to zero, while structures with
three-level methods, numerical schemes with wavelets as trial are considered to belong to the interior of a region. Here the
functions, and pseudospectral methods [18], and multiplica- diffusivity is close to one. In this sense, the CLMC model
tive splittings [43]. Even hardware proposals for nonlinear serves as a selective smoothing, which prefers intraregional
diffusion filtering can be found in the literature [19] and [36]. smoothing to interregional blurring. After some time it leads to
Schemes that inherit a large number of the properties of their segmentationlike results, which are piecewise almost constant.
continuous counterparts have also been proposed in the context The parameter makes the filter insensitive to noise
of curvature-based nonlinear scale-spaces [8], [10], [11], [13]. at scales smaller than . It is also a regularization parameter
Sophisticated algorithms for such processes comprise fast level which guarantees well posedness of the process: Catté et al.
set methods [40], high-order ENO schemes [41], and implicit [12] have shown that their filter has a unique solution which
algorithms for mean curvature motion [2], [31]. is infinitely times differentiable for . Weickert [44], [46]
has proved that it depends continuously on the original image,
satisfies a maximum–minimum principle and reveals a large
II. THE CONTINUOUS FILTER PROCESS family of smoothing Lyapunov functionals which guarantee
In the -dimensional case the filter of Catté et al. [12] has that the solution tends to a constant image for . During
the following structure. the whole evolution, the average grey value remains unaltered.

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.
400 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO. 3, MARCH 1998

Equations of this type have been successfully applied to where is the unit matrix. This scheme is called
process medical images (see e.g., [24], [27], [50]). Never- explicit, since can be directly calculated from without
theless, they are only one representative of a large class of solving a system of equations.
nonlinear scale-spaces. Overviews of other methods can be Such an explicit iteration step is computationally very
found in [16], [21], and [46]. cheap: It requires mainly to calculate the three nonvanishing
matrix entries per row and to perform a matrix–vector multi-
III. 1-D CASE plication. The computational and storage effort is linear in the
pixel number . But does this explicit scheme also create a
good discrete scale-space and how far can we come with one
A. Explicit Scheme
step? We can find an answer to these question by applying a
1) The Scheme: The 1-D CLMC equation is given by framework for discrete nonlinear diffusion scale-spaces, which
we shall review next.
(7)
2) Criteria for Discrete Nonlinear Diffusion Scale-Spaces:
Let us now consider the simplest discrete approximation Recently, a scale-space interpretation for the continuous
of this process. A discrete image can be regarded as a vector CLMC equation and its anisotropic generalizations has been
, whose components , display established [44], [46]. In addition to invariances such as
the grey values at each pixel. Pixel represents some location the preservation of the average grey value, it has been
, and is the grid size. We consider discrete times , shown that—it spite of its contrast-enhancing potential—these
where IN and is the time step size. By we denote equations create smoothing scale-spaces: They obey a
approximations to . maximum–minimum principle, have a large class of smoothing
The simplest discretization of (7) with reflecting boundary Lyapunov functionals, and converge to a constant steady-state.
conditions is given by It would be desirable to ensure that discrete approximations
do also reveal these qualities exactly. Criteria have been
identified under which one can guarantee that a discrete
(8)
scheme of type

where is the set of the two neighbors of pixel (boundary (13)

pixels have only one neighbor). IN (14)
The diffusivities approximate . They can
be obtained as follows. possesses such properties [45], [46]. All one has to check are
In the spatially discrete case the convolution the following criteria for .
comes down to a multiplication of with a suitable
D1) Continuity in Its Argument:
matrix . In Section IV-C we shall present an
efficient way to achieve this in the spatial domain. A gradient
(15)
approximation by central differences gives
D2) Symmetry:
(9)
(16)

for some inner pixel . This expression remains also valid at D3) Unit Row Sum:
the boundary pixels, if we extend the image by reflecting it
at the boundary. (17)
We can write the explicit scheme in matrix–vector notation
as
D4) Nonnegativity:
(10)
(18)
with and
D5) Positive Diagonal:

, (19)

, (11)
D6) Irreducibility:
We can connect any two pixels by a path with nonva-
(else). nishing diffusivities. Formally:
This comes down to the iteration scheme For any there exist with
and such that for
(12) .

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.
WEICKERT et al.: NONLINEAR DIFFUSION FILTERING 401

Under these prerequisites the filtering process is well posed By the construction of it is also evident that the row
and satisfies the following discrete scale-space properties [45], sums of vanish. Hence, all row sums of are one, which
[46]: proves D3.
a) Average Grey-Level Invariance: The average grey Thus, let us investigate the nonnegativity. From for
level is not affected by the discrete , we also have for . Thus, we can focus
diffusion filter: on the diagonal entries. If they are all positive, both D4 and
D5 are satisfied. Since
IN (20)
(27)

This invariance is required in scale-space based segmentation

algorithms such as the hyperstack [33]. and , positive diagonal entries require that
b) Extremum Principle:
(28)
IN (21)

This property is much more than a stability result which In order to show that is irreducible, let us assume that
forbids under- and overshoots. It also ensures that iso-intensity satisfies this restriction and consider two arbitrary pixels
linking toward the original image is possible. Hence, it states and . If then the positivity of implies that
an important causality property, cf., [22].
c) Smoothing Lyapunov Sequences: The process is a (29)
simplifying, information-reducing transform with respect to
If then
the following aspects.
1) The -norms (30)
This establishes D6.
(22) From these considerations we conclude that the explicit
scheme creates a discrete scale-space provided that the time
step size satisfies the restriction (28). In image processing, one
are decreasing in for all .
usually sets . Since the diffusivity is bounded from
2) All even central moments
above by 1, definition (11) allows us to guarantee (28) for
.
IN (23) In practice, this is often a very severe step size restriction.
It means that the use of an explicit scheme is limited rather
are decreasing in . by its stability than its accuracy. For this reason it would be
3) The entropy interesting to look for schemes with better stability properties.
This shall be done next.
(24)
B. Semi-Implicit Scheme
1) The Scheme: We consider a slightly more complicated
a measure of uncertainty and missing information, is discretization of (7), namely
increasing in (if is positive for all ).
d) Convergence to a Constant Steady-State: (31)
(25) which leads to the scheme
Thus, the discrete scale-space evolution tends to the most (32)
global image representation that is possible: a constant image
We observe that this scheme does not give the solution
with the same average grey level as .
directly (explicitly): It requires to solve a linear system first.
3) Does the Explicit Scheme Create a Discrete
For this reason it is called a linear-implicit (semi-implicit)
Scale-Space?: Let us now investigate if the explicit scheme
scheme.
(12) satisfies the criteria D1–D6 for discrete nonlinear
Remark: One may also be interested in studying the (fully)
scale-spaces. Let
implicit scheme
(26)
(33)
By virtue of (11) we observe that the continuity of with
respect to its argument follows directly from the continuity of leading to a nonlinear system of equations. This is more
the diffusivity . complicated to solve. Below we shall see, however, that such
The symmetry of follows from (11) and the symmetry of a high effort is not necessary, since already semi-implicitness
the neighborhood relation [ ]. is sufficient to guarantee absolute stability.

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.
402 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO. 3, MARCH 1998

2) Does the Semi-Implicit Scheme Create a Discrete Step 1) LR Decomposition: We decompose into the
Scale-Space?: In order to establish the semi-implicit scheme product of a lower bidiagonal matrix
(32) as a discrete scale-space we have to check D1–D6 again.
First we have to show that
(34) .. .. (40)
. .
is invertible. This is easily seen, because is strictly diago-
nally dominant and an upper bidiagonal matrix
(35)
.. ..
. . (41)
It is well known from linear algebra that strictly diagonally
dominant matrices are invertible, see e.g., [35, p. 226]. Thus,
(36) Comparing the coefficients shows that for all , and
exists and the continuity of in its argument follows from and can be obtained as follows:
the continuity of . Moreover, the symmetry of carries also
over to and , which establishes D2.
In order to prove D3, consider .
Since has unit row sum, we have . This implies that
(37)
Reading this componentwise shows that has also unit row
sum. Solving for is done in two steps:
D4–D6 can be verified in one step. We already know that Step 2) Forward Substitution: We solve for . This
is strictly diagonally dominant. It is also immediately seen that gives
is irreducible, for , and for all . Then
a theorem by Varga [42, p. 85] tells us that satisfies
(38)
Thus, is nonnegative, has positive diagonal and is irre-
ducible.
From these considerations we observe that the semi-implicit Step 3) Backward Substitution: We solve for .
scheme creates a discrete nonlinear diffusion scale-space for This leads to
arbitrarily large time steps. In particular, it is unconditionally
stable and does not suffer from any time step size restriction.
Unlike the explicit scheme, it can be fully adapted to the
desired accuracy without the need to choose small time steps
for stability reasons.
3) Solving the Tridiagonal Linear System—The Thomas
Algorithm: The semi-implicit scheme requires to solve a This completes the Thomas algorithm. It is stable for
linear system, where the system matrix is tridiagonal and every strictly diagonally dominant system matrix. One may
diagonally dominant. The most efficient way to do this is the also regard it as a recursive filtering: The LR decomposition
so-called Thomas algorithm, a Gaussian elimination algorithm determines the filter coefficients, Step 2 is a causal filter and
for tridiagonal systems. It can be found in many textbooks Step 3 an anticausal one. The whole scheme is very efficient;
on numerical analysis, e.g., [39, pp. 43–45]. However, since it requires only
it builds the backbone of our algorithms and since we want
to keep this paper self contained, we survey its algorithmic (42)
features here.
The principle is as follows. Suppose we want to solve a multiplications/divisions, and
tridiagonal linear system with (43)

subtractions. Hence, the CPU effort is linear in . The same

.. .. .. is true for the storage effort.
. . . (39)
Applying the Thomas algorithm to the semi-implicit scheme
takes almost twice as long as one iteration of the explicit
scheme, but we may use much larger time steps, since the
Then the Thomas algorithm consists of three steps. scheme is absolutely stable.

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.
WEICKERT et al.: NONLINEAR DIFFUSION FILTERING 403

IV. HIGHER-DIMENSIONAL CASE restriction (48) may be replaced by

A. Explicit and Semi-Implicit Schemes

(50)
The -dimensional CLMC equation is given by

Thus, the allowed step size of the explicit scheme becomes

(44)
even smaller for higher dimensions.
However, this does not necessarily imply that the semi-
implicit scheme becomes superior. There appears a new prob-
We can discretize the summands of the right hand side in
lem as well: Although the actual structure of the matrix
the same manner as in the 1-D case. Using only one index
depends on the pixel numbering, it is not possible anymore
for pixel numbering, we may represent the whole image of
to order the pixels in such a way that in the th row all non-
size as a vector of size . In
vanishing matrix elements can be found within the positions
this vector–matrix notation we can write the -dimensional
to ; usually, the matrix reveals a much
explicit scheme as
larger bandwidth. Applying direct algorithms such as Gaussian
elimination would destroy the zeros within the band and would
lead to an immense storage and computation effort. Hence,
(45)
iterative algorithms have to be applied. Classical methods
like Gauss–Seidel or successive overrelaxation (SOR) do not
need additional storage and convergence can be guaranteed
and its semi-implicit counterpart as
for the special structure of . This convergence, however, is
rather slow. Faster iterative methods such as the precondi-
tioned conjugate gradient algorithms [30, pp. 154–161] need
(46) significantly more storage, which can become prohibitive for
large images. A typical problem of iterative methods is also
that their convergence becomes slower for larger , since this
In both cases, the matrix corresponds to increases the condition number of the system matrix. Multigrid
derivatives along the th coordinate axis. Let us also introduce methods [7] appear to be one possibility to circumvent many of
these problems, but their implementation is more complicated.
Recapitulating, we see that for dimensions the semi-
(47) implicit scheme remains absolutely stable, but it is difficult to
take full advantage of this because of the problems to solve
the arising linear system as efficiently as it was possible in the
What about the reliability of both schemes? Checking the 1-D case with the Thomas algorithm.
discrete scale-space requirements, D1–D6 can be done in a
similar way as in the 1-D case, see [46, Th. 8] for more details.
As in the 1-D case one obtains that the explicit scheme creates B. AOS Schemes
a discrete scale-space for In order to address the above-mentioned problem let us
consider a modification of the semi-implicit scheme (46),
namely the additive operator splitting (AOS) scheme
(48)

(51)
and that the semi-implicit scheme satisfies all requirements
unconditionally.
What does this mean regarding efficiency? In the - Several points should be noted, as follows.
dimensional case each inner pixel has neighbors with
• The explicit scheme (45), the semi-implicit scheme (46),
which it is connected via nonvanishing entries in the th row
and the AOS scheme (51) have the same first-order Taylor
of . From (11) we see that we can estimate
expansions in . It is easy to see that all schemes are
approximations to the continuous
equation. From this viewpoint, all schemes are consistent
(49)
to the original equation. One should not make the mistake
to regard the AOS scheme as an algebraically incorrect
reformulation of the semi-implicit scheme: The explicit
where denote the dimensions of an - scheme is also different from the semi-implicit one, but
dimensional pixel. With and , it approximates the same continuous diffusion process.

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.
404 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO. 3, MARCH 1998

• The operators and block irreducible.1 Within each irreducible matrix block,
we have a positive diagonal and nonnegative off-diagonals.
(52) Applying again Varga’s theorem [42, p. 85], we conclude that
the inverse of each block contains only positive elements.
From this it follows that for some implies
describe 1-D diffusion processes along the axes. Thus, that . Thus, the irreducibility of
under a suitable pixel numbering they come down to carries over to , and D6 is satisfied.
strictly diagonally dominant tridiagonal matrices which In particular, since is constructed such that for
can be inverted in an efficient and stable way by the all , it is clear that contains only positive diagonal
Thomas algorithm from Section III-B3. elements. Therefore, D5 is verified as well.
• Since it is an additive splitting, all coordinate axes are These discussions show that the AOS scheme creates a
treated in exactly the same manner. This is in contrast to discrete nonlinear diffusion scale-space for all time step sizes.
conventional splitting techniques from the literature [17],
[28], [30], [51]. They are multiplicative splittings such as
the locally 1-D (LOD) scheme C. Regularization
This section describes a simple method for calculating the
presmoothing in a way which is consistent with
the ideas presented above.
(53)
It is well known that Gaussian convolution with standard
deviation is equivalent to linear diffusion filtering ( ) for
some time . Thus we may use the (semi-)implicit2
Since in the general nonlinear case the split operators do scheme again in order to obtain a stable algorithm. Several
not commute, the result of multiplicative splittings will things make the situation even easier than in the nonlinear
depend on the order of the one-dimensional operators. setting.
This disadvantage will be discussed in Section VI in more • Frequently, is in the order of the pixel size. In this case
detail. we may regularize in a single step by filtering once with
1) Does the AOS Scheme Create a Discrete Scale-Space?: a time step size .
The discussed properties suggest that the AOS scheme is an • The linear diffusion process is separable. Therefore, the
interesting candidate for an efficient discrete diffusion scale- order of the one-dimensional approximations is not of
space. Thus, let us now assess its reliability by checking the importance and we may also use a multiplicative splitting:
criteria D1–D6.
Many reasonings carry over from the 1-D semi-implicit
scheme: First we observe that exist, (55)
since is strictly diagonally dominant. Also the continuity
of
• The system in step can be decomposed into
tridiagonal systems with the same system matrix. Thus,
(54) the LR decomposition needs to be done only once for an
-matrix of type
in its argument is a direct consequence of the continuous
diffusivity and the construction of .
In the same way the symmetry of goes back to the
symmetry of . Note that the symmetry of is independent .. .. .. (56)
. . .
of the pixel numbering: a permutation of their numbering
transforms into for some permutation matrix .
Since and there exists a pixel numbering such that
is transformed into a symmetric tridiagonal matrix just as with . Therefore, the main effort boils down
in the 1-D case, it is clear that has to be symmetric. to performing times the same forward and back-
With the same reasoning as in III-B2, we know that not only ward substitution step from the Thomas algorithm. This
, but also has row sum 1. Thus, has also unit row sum. requires only multiplications/divisions and
To verify D4, we observe that is strictly subtractions. Such an effort is comparable with the
diagonally dominant, for all , and for . recursive filters presented in [2], [3], and [52], but unlike
Under these circumstances we may conclude from [29, p. 192] those Fourier-based methods, the algorithm presented
that is nonnegative in all components. This implies here allows an adequate treatment of the reflecting bound-
the nonnegativity of . ary conditions and preserves the average grey value.
Let us now check D5 and D6 in one step. Since ,
represent 1-D diffusion operators, it follows that 1 Each of the N=N blocks represents the pixels where all components
l
there exist permutation matrices , such that except for the lth are identical.
is not only diagonally dominant, but also tridiagonal 2 Semi-implicit and implicit are identical in the linear case.

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.
WEICKERT et al.: NONLINEAR DIFFUSION FILTERING 405

V. ALGORITHMIC STRUCTURE TABLE I

m
MAIN OPERATIONS FOR ONE -DIMENSIONAL AOS STEP
(M/D: MULTIPLICATIONS OR DIVISIONS; A/S: ADDITIONS OR
A. AOS Algorithm SUBTRACTIONS; LUT: LOOK-UP OPERATIONS IN A TABLE)
We may summarize our considerations in the following
algorithm for one AOS step in dimensions.

A loss of rotational invariance becomes visible as a prefer-

ence of certain directions, while a loss of accuracy becomes
evident in those cases where filtering with time step differs
visually from times filtering with . So let us now check
these approximation effects by applying a 2-D AOS scheme
to two test images.
- -
First we check the rotational invariance. Since the AOS
scheme is consistent to the original equation, we should expect
good rotational invariance for small spatial and temporal steps.
Fig. 1 is used as a test for rotational invariance. It depicts
a Gaussian-like image and its filtered versions. For
B. Complexity both the explicit and AOS scheme are visually indistinguish-
In order to assess the complexity of AOS algorithms, let us able. This step size is also the stability limit for the explicit
consider dimensions and focus on terms of order . scheme, while the AOS scheme allows to increase further.
From the preceding algorithm we recognize that only the We see that for no significant changes appear. Thus,
four vectors , , , and are required. Thus, since all AOS may be used with 20 times larger time steps than the
calculations may be performed in single precision, the main explicit scheme. On the other hand, even for the
storage effort is bytes. This is independent of the deviations from a perfect circular structure are not very severe.
dimension . What about the accuracy? Fig. 2 depicts the filtering of a
Table I summarizes the relevant computational requirements brain image. The situation is similar as in Fig. 1: For
for each step of the AOS algorithm. We observe that the effort the explicit and the AOS scheme are undistinguishable. The
is proportional to the number of pixels and the dimension AOS scheme remains close to these results up to .
. The total effort is only multiplications or divisions, For we get more severe deviations: the filtering
additions or subtractions, and look-ups in effect becomes weaker. This is a typical behavior for implicit
a table. This is less than twice the typical effort needed for schemes with large time steps: implicit techniques always
an explicit scheme, a rather low price for gaining absolute remain on the “safe” side (by orienting the diffusion on the
stability. “smoother” future rather than on the “rougher” past), while
their deviation from the true solution becomes larger with
increasing step size. Thus, their filtering effect on the final
VI. EVALUATION image at a specified time decreases with increasing time step
We have seen that AOS schemes with large time steps size. Again is a good compromise between efficiency
still reveal average grey value invariance, stability based on and accuracy.
extremum principle, Lyapunov functionals, and convergence to After these visual inspections, we shall investigate the accu-
a constant steady-state. Thus, they are legitimate when being racy more quantitatively. To this end we perform a comparison
considered as a pure discrete process which is not intended to between the explicit scheme (45), the semi-implicit scheme
approximate a continuous process. (46), and the AOS scheme (51). Since no analytical solution to
But does this mean that it is recommendable to consider the CLMC equation is known, we have to use a good numerical
arbitrarily large time step sizes? In the extreme case: can one approximation to a test example as a standard for comparison.
filter an image in one step? In our case we took the explicit scheme with the small step
In this case, we should expect problems with those proper- size and applied it 2000 times to the test image from
ties which a naturally linked to continuous ideas and which can Fig. 2.
only be satisfied approximately by discrete schemes: rotational The linear system of the 2-D semi-implicit scheme is solved
invariance and accuracy. by a Gauss–Seidel algorithm. Iterative methods of this type are

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.
406 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO. 3, MARCH 1998

(a) (b)

(a) (b)

(c) (d)

(c) (d)

(e) (f)
=
Fig. 1. Nonlinear diffusion filtering of a Gaussian-like test image ( 8,
 = 1:5). (a) Original image,
= (0; 101)2 . (b) Explicit scheme, 800
iterations,  = 0:25. (c) AOS scheme, 800 iterations,  = 0:25. (d) AOS
scheme, 200 iterations,  = 1. (e) AOS scheme, 40 iterations,  = 5. (f)
AOS scheme, ten iterations,  = 20.
(e) (f)

quite popular for nonlinear PDE’s in image processing [2], Fig. 2. Nonlinear diffusion filtering of a medical image ( = 2,  = 1). (a)
2
Original image,
= (0; 255) (0; 308). (b) Explicit scheme, 800 iterations,
[31], since they are easy to implement and they do not require  = 0:25. (c) AOS scheme, 800 iterations,  = 0:25. (d) AOS scheme, 200
additional memory. Let the diffusion operator iterations,  = 1. (e) AOS scheme, 40 iterations,  = 5. (f) AOS scheme,
ten iterations,  = 20.

(57)
(60)

be decomposed into the strictly lower triangular matrix , the Every second step we calculate the residue
diagonal matrix , and the strictly upper triangular matrix
. Then the Gauss–Seidel method approximates the solution (61)
of the semi-implicit scheme
and we stop the iteration process if its norm
(58) satisfies

by a sequence of vectors with (62)

(59) with some accuracy parameter or .

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.
WEICKERT et al.: NONLINEAR DIFFUSION FILTERING 407

TABLE II Interestingly, for , it is also more accurate than the

COMPARISON OF NONLINEAR DIFFUSION SCHEMES semi-implicit scheme with .
It is worth noticing that there is a fundamental difference
between errors in the AOS scheme and errors that are intro-
duced by an insufficient number of Gauss–Seidel iterations:
Unlike AOS errors which are compatible with the discrete
scale-space framework, Gauss–Seidel errors can violate these
requirements. Thus, properties such as the average grey level
invariance are no more satisfied in an exact manner. In
order to avoid these difficulties, one would have to apply
more Gauss–Seidel iterations, which will finally destroy all
efficiency advantages compared to the explicit scheme; see
also [31].
Fig. 3 gives a graphical representation of Table II, which
allows us to find the most efficient schemes for a desired
accuracy. We observe that for very high accuracy requirements
the explicit scheme is most appropriate.3 This is at the expense
of a height overall computational effort. On the other hand,
even relaxing the accuracy requirements to a relative error
of 1% does not permit to find a more efficient technique.
For errors between 1% and 1.7%, the semi-implicit scheme
with is fastest, and for errors larger than 1.7%, AOS
schemes become rapidly superior. In our previous experiments
we have observed that the accuracy of AOS with
appears to be tolerable for many applications. This corresponds
to an error of about 2.2%. In this case, AOS is almost 2.5 times
more efficient than the semi-implicit scheme with ,
more than 3.5 times faster than the semi-implicit scheme
with , and about 11 times more efficient than the
explicit scheme. Although these relations have been illustrated
by one example only, additional experiments have indicated
that these basic relations between explicit, semi-implicit and
AOS discretizations carry over to a large class of images: The
accuracy requirements of many practical problems allow an
efficiency gain by one order of magnitude. All one has to do
is to replace the explicit scheme by an AOS scheme with 20
Table II compares the explicit scheme, the semi-implicit times larger time step sizes.
scheme with accuracies and , respectively, It should be noted that the AOS schemes calculate the
and the AOS scheme. average of operators of type . They describe
If denotes our reference solution (explicit scheme, 1-D diffusions with a step size . Since multiplicative
), then we calculate the relative error of an approximation splittings such as the LOD scheme (53) use operators of
as type , one can expect that they give even bet-
ter accuracy. However, multiplicative splittings for nonlinear
problems reveal one big disadvantage, which makes their use
(63) in many image processing applications problematic: In the
general nonlinear case the split operators do not commute any
longer. Thus, the result of multiplicative splittings depends
First we observe that the explicit scheme with on the order of the 1-D operators, and the grid axes are
reveals a very small error, while the semi-implicit method with treated differently. In practice, this means that these schemes
is not only less accurate, but also slower for produce different results if the image is rotated by 90 . Such an
and 0.5. For the semi-implicit scheme becomes faster undesirable effect is illustrated in Fig. 4. Since AOS schemes
than the explicit one. On the other hand, the Gauss–Seidel apply the 1-D operators in parallel instead of sequentially, they
algorithm slows down for larger , since this increases the do not suffer from this limitation.
condition number of the system matrix. Hence, the overall Moreover, most multiplicative splittings lead to a nonsym-
CPU time per semi-implicit step increases with increasing . metric system matrix . This violates criterion D2 for
If we relax the accuracy from to , the semi- 3 One can achieve even higher accuracy by methods which are of second
implicit scheme becomes faster, but the error also increases. order in time, for instance predictor–corrector techniques [46]. Such a high
For , the AOS scheme becomes the fastest method. accuracy, however, is rarely required in image processing.

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.
408 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO. 3, MARCH 1998

TABLE III
MEASURED CPU TIMES FOR ONE AOS ITERATION

Fig. 3. Tradeoff between efficiency and accuracy of nonlinear diffusion

solvers. The data were calculated on the test image from Fig. 2, size
2

= (0; 255) (0; 308). Filter parameters:  = 2,  = 1. Stopping time: Three-dimensional data sets from medicine with typical
T = 200. Hardware: one R10000 processor on an SGI Challenge XL. sizes such as 256 256 64 can be processed in less than 1
min per AOS iteration (both on the HP and the SGI). In many
practical applications less than ten iterations are sufficient for
the denoising of such data sets.
Recapping we have observed that—although the desired
approximation quality is of course purpose dependent—under
typical circumstances 20 times larger step sizes than the
stability limit of the explicit scheme appear reasonable. They
give an efficiency gain of a factor ten.4 Especially for large
data sets such as 3-D medical data this is often the difference
between not applicable and applicable. We are currently testing
our schemes for the filtering of 3-D ultrasound images and
preprocessing 3-D MR data for segmentation. In both cases
first results are encouraging.
(a) (b)
Fig. 4. (Non-)commutation of nonlinear diffusion operators. The difference VII. CONCLUSIONS
between filtering prior to rotation by 90 , and rotation prior to filtering is
depicted. Test image: Fig. 2 ( = 2,  = 1,  = 20, ten iterations). We have presented absolutely stable additive operator split-
(a) A multiplicative splitting such as LOD treats x and y axes differently.
(b) Additive operator splitting (AOS) treats all axes equally.
ting (AOS) schemes for the nonlinear diffusion filter of Catté et
al. and Whitaker and Pizer. These schemes satisfy all criteria
for discrete nonlinear diffusion scale-spaces and are easy to
discrete diffusion scale-spaces. For this reason, we have not
implement in any dimension. Both computational and storage
considered these approaches in the present paper.
effort is linear in the number of pixels. Experiments have
Finally we check the relation between the computational
shown that under realistic accuracy requirements one can gain
effort and the number of pixels. Table III shows the measured
an increase of efficiency by a factor of . This makes this
CPU times on a single R10000 processor of an SGI Challenge
type of schemes attractive for applications such as medical
XL and on an HP 900-755, both for 2-D and 3-D images.
3-D data sets.
For small image sizes the computing times reveal good
Implementations of AOS schemes on parallel architectures
proportionality to the overall number of pixel. This is what we
are studied in [49]. These experiments demonstrate that it is
expect from theory. Because of Cache limitations, the CPU
possible to gain a speed-up by another order of magnitude
time per pixel becomes slightly higher for huger data sets:
by exploiting the intrinsic parallelism of AOS schemes. Last
We also observe that this deviation from the linear scaling
behavior is machine dependent. The HP remains closer to the 4 We have seen that an m-dimensional AOS scheme averages 1-D operators

linear scaling behavior than the SGI. On the other hand, with with an effective step size of m . Thus, for higher dimensions m one should
reduce the step size in order to have the same accuracy. However, since
its CPU memory of 1 Gb the SGI permits even to process data explicit schemes also have to decrease the step size for larger m in the same
sets of size 8192 8192 and 512 512 256. way, the factor 10 remains valid for every dimension.

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.
WEICKERT et al.: NONLINEAR DIFFUSION FILTERING 409

but not least, there are also ways to generalize AOS schemes [23] T. Iijima, “Basic theory of pattern normalization (for the case of a typical
to anisotropic diffusion filters with diffusion tensors; a first one-dimensional pattern),” Bull. Electrotech. Lab., vol. 26, pp. 368–388,
1962 (in Japanese).
proposal in this direction can be found in [46, Sect. 4.4.2]. [24] C. Lamberti, M. Sitta, and F. Sgallari, “Improvements to the anisotropic
diffusion model for 2-D echo image processing,” in Proc. Annual Int.
Conf. IEEE Engineering in Medicine and Biology Soc., 1992, vol. 14,
ACKNOWLEDGMENT pp. 1872–1873.
The authors would like to thank R. Maas for careful [25] X. Li and T. Chen, “Nonlinear diffusion with multiple edginess thresh-
olds,” Pattern Recognit., vol. 27, pp. 1029–1037, 1994.
proofreading and helpful hints. [26] T. Lindeberg, “Scale-space for discrete signals,” IEEE Trans. Pattern
Anal. Machine Intell., vol. 12, pp. 234–254, 1990.
[27] D.-S. Luo, M. A. King, and S. Glick, “Local geometry variable
REFERENCES conductance diffusion for post-reconstruction filtering,” IEEE Trans.
Nucl. Sci., vol. 41, pp. 2800–2806, 1994.
[1] S. T. Acton, A. C. Bovik, and M. M. Crawford, “Anisotropic diffusion [28] G. I. Marchuk, “Splitting and alternating direction methods,” Handbook
pyramids for image segmentation,” in Proc. IEEE Int. Conf. Image
of Numerical Analysis, vol. I, P. G. Ciarlet and J.-L. Lions, Eds.
Processing, Austin, TX, Nov. 13–16, 1994, vol. 3, pp. 478–482.
[2] L. Alvarez, “Images and PDE’s,” in ICAOS’96: Images, Wavelets and Amsterdam, The Netherlands: North Holland, 1990, pp. 197–462.
[29] T. Meis and U. Marcowitz, Numerische Behandlung Partieller Differ-
PDE’s, vol. 219., M.-O. Berger, R. Deriche, I. Herlin, J. Jaffré, and
entialgleichungen. New York: Springer-Verlag, 1978.
J.-M. Morel, Eds. New York: Springer, 1996, pp. 3–14.
[3] L. Alvarez and L. Mazorra, “Signal and image restoration using shock [30] A. R. Mitchell and D. F. Griffiths, The Finite Difference Method in
filters and anisotropic diffusion,” SIAM J. Numer. Anal., vol. 31, pp. Partial Differential Equations. New York: Wiley, 1980.
590–605, 1994. [31] E. Nicolet and M. Spühler, “Charakterisierung von unscharfen
[4] K. Åström, and A. Heyden, “Stochastic analysis of image acquisition and blobförmigen 3-D strukturen,” Dipl. thesis, Commun. Technol. Lab.,
scale-space smoothing,” in Gaussian Scale-Space Theory, J. Sporring, Image Sci. Div., ETH-Zentrum, Zürich, Switzerland, 1996.
M. Nielsen, L. Florack, and P. Johansen, Eds. Boston, MA: Kluwer, [32] M. Nielsen, L. Florack, and R. Deriche, “Regularization, scale-space and
1997, pp. 129–136. edge detection filters,” J. Math. Imag. Vis., vol. 7, pp. 291–307, 1997.
[5] E. Bänsch and K. Mikula, “A coarsening finite element strategy in [33] W. J. Niessen, K. L. Vincken, J. Weickert, and M. A. Viergever,
image selective smoothing,” Faculty Math., Univ. Freiburg, Freiburg, “Nonlinear multiscale representations for image segmentation,” Comput.
Germany, preprint no. 18/1996, 1996; to be published in Computation Vis. Image Understand., vol. 66, pp. 233–245, 1997.
and Visualization in Science. [34] E. Norman, “A discrete analogue of the Weierstrass transform,” Proc.
[6] B. Benhamouda, “Parameter adaptation for nonlinear diffusion in im- Amer. Math. Soc., vol. 11, 596–604, 1960.
age processing,” Master’s thesis, Dept. Math., Univ. Kaiserslautern, [35] J. M. Ortega, Matrix Theory. New York: Plenum, 1987.
Kaiserslautern, Germany, 1994. [36] P. Perona and J. Malik, “A network for multiscale image segmentation,”
[7] W. L. Briggs, A Multigrid Tutorial. Philadelphia, PA: SIAM, 1987. in Proc. IEEE Int. Symp. Circuits and Systems, Espoo, Finland, June
[8] A. M. Bruckstein, G. Sapiro, and D. Shaked, “Evolution of planar 7–9, 1988, pp. 2565–2568.
polygons,” Int. J. Pattern Recognit. Artif. Intell., vol. 9, pp. 991–1014, [37] , “Scale space and edge detection using anisotropic diffusion,”
1995. IEEE Trans. Pattern Anal. Machine Intell., vol. 12, pp. 629–639, 1990.
[9] L. D. Cai, “Some notes on repeated averaging smoothing,” in Pattern [38] A. H. Salden, B. M. ter Haar Romeny, and M. A. Viergever, “Linear
Recognition, vol. 301, J. Kittler, Ed. New York: Springer-Verlag, 1988, scale-space theory from physical principles,” J. Math. Imag. Vis., to
pp. 597–605. appear.
[10] F. Catté, “Convergence of iterated affine and morphological filters [39] H. R. Schwarz, Numerische Mathematik. Stuttgart, Germany: Teubner,
by nonlinear semi-group theory,” in ICAOS’96: Images, Wavelets and 1988.
PDE’s, vol. 219, M.-O. Berger, R. Deriche, I. Herlin, J. Jaffré, and J.-M. [40] J. A. Sethian, Level Set Methods. Cambridge, U.K.: Cambridge Univ.
Morel, Eds. New York: Springer-Verlag, 1996, pp. 125–133. Press, 1996.
[11] F. Catté, F. Dibos, and G. Koepfler, “A morphological scheme for [41] K. Siddiqi, B. B. Kimia, and C.-W. Shu, “Geometric shock-capturing
mean curvature motion and applications to anisotropic diffusion and ENO schemes for subpixel interpolation, computation and curve evolu-
motion of level sets,” SIAM J. Numer. Anal., vol. 32, pp. 1895–1909, tion,” Graph. Models Image Process., vol. 59, pp. 278–301, 1997.
1995. [42] R. A. Varga, Matrix Iterative Analysis. Englewood Cliffs, NJ: Prentice-
[12] F. Catté, P.-L. Lions, J.-M. Morel, and T. Coll, “Image selective Hall, 1962.
smoothing and edge detection by nonlinear diffusion,” SIAM J. Numer. [43] J. Weickert, “Zwischenbericht zum projekt ‘nichtlineare diffusionsfil-
Anal., vol. 29, 182–193, 1992. ter’,” Ber. über wissenschaftliche Tätigkeit, Ctr. Appl. Math., Darm-
[13] T. Cohignac, F. Eve, F. Guichard, C. Lopez, and J.-M. Morel, “Numeri- stadt–Kaiserslautern, Germany, pp. 133–142, 1992.
cal analysis of the fundamental equation of image processing,” Preprint [44] , “Theoretical foundations of anisotropic diffusion in image
no. 9254, CEREMADE, Univ. Paris IX–Dauphine, Paris, France, processing,” Computing, suppl. 11, pp. 221–236, 1996.
1992. [45] , “Nonlinear diffusion scale-spaces: From the continuous to the
[14] R. Deriche, “Fast algorithms for low-level vision,” IEEE Trans. Pattern discrete setting,” in Proc. ICAOS’96: Images, Wavelets and PDE’s, vol.
Anal. Machine Intell., vol. 12, pp. 78–87, 1990. 219, M.-O. Berger et al., Eds. New York: Springer, 1996, pp. 111–118.
[15] , “Recursively implementing the Gaussian and its derivatives,” in [46] , Anisotropic Diffusion in Image Processing. Stuttgart, Ger-
Proc. 2nd Int. Singapore Conf. on Image Processing, V. Srinivasan, S. many: Teubner, 1998.
H. Ong, and Y. H. Ang, Eds., Singapore, Sept. 7–11, 1992, pp. 263–267. [47] , “Recursive separable schemes for nonlinear diffusion filters,”
[16] R. Deriche and O. Faugeras, “Les EDP en traitement des images et
Scale-Space Theory in Computer Vision, vol. 1252, B. ter Haar Romeny,
vision par ordinateur,” Traite. Signal, vol. 13, no. 6, 1996.
[17] J. Douglas and J. E. Gunn, “A general formulation of alternating et al., Eds. New York:, Springer-Verlag: 1997, pp. 260–271.
[48] J. Weickert, S. Ishikawa, and A. Imiya, “On the history of Gaussian
direction methods. Part I. Parabolic and hyperbolic problems,” Numer.
Math., vol. 6, pp. 428–453, 1964. scale-space axiomatics,” Gaussian Scale-Space Theory, J. Sporring et
[18] J. Fröhlich and J. Weickert, “Image processing using a wavelet algo- al., Eds. Boston, MA: Kluwer, 1997, pp. 45–59.
rithm for nonlinear diffusion,” Rep. no. 104, Lab. Technomath., Univ. [49] J. Weickert, K. J. Zuiderveld, B. M. ter Haar Romeny, and W. J. Niessen,
Kaiserslautern, Kaiserslautern, Germany, 1994. “Parallel implementations of AOS schemes: A fast way of nonlinear
[19] T. Gijbels et al., “A VLSI-architecture for parallel nonlinear diffusion diffusion filtering,” in Proc. 1997 IEEE Int. Conf. Image Processing,
with applications in vision,” in Proc. IEEE Workshop on VLSI Signal Santa Barbara, CA, Oct. 26–29, 1997.
Processing, 1994, vol. 7, pp. 398–707. [50] R. T. Whitaker and S. M. Pizer, “A multi-scale approach to nonuniform
[20] A. R. Gourlay, “Implicit convolution,” Image Vis. Comput., vol. 3, pp. diffusion,” CVGIP: Image Understand., vol. 57, pp. 99–110, 1993.
15–23, 1985. [51] N. N. Yanenko, The Method of Fractional Steps: The Solution of
[21] B. M. ter Haar Romeny, Ed., Geometry-Driven Diffusion in Computer Problems of Mathematical Physics in Several Variables. New York:
Vision. Boston, MA: Kluwer, 1994. Springer-Verlag, 1971.
[22] R. A. Hummel, “Representations based on zero-crossings in scale [52] D. Zhao and B. Li, “A new implementation of discrete multiscale filter-
space,” in Proc. IEEE Comput. Soc. Conf. Computer Vision and Pattern ing,” in Proc. IEEE Int. Conf. Image Processing, Lausanne, Switzerland,
Recognition, Miami Beach, FL, June 22–26, 1986; pp. 204–209. Sept. 16–19, 1996, vol. 1, pp. 383–386.

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.
410 IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 7, NO. 3, MARCH 1998

Joachim Weickert received the M.Sc. degree in Max A. Viergever received the M.Sc. degree in
industrial mathematics in 1991 and the Ph.D. degree applied mathematics in 1972 and the D.Sc. degree,
in mathematics in 1996, both from Kaiserslautern with a thesis on cochlear mechanics, in 1980,
University, Kaiserslautern, Germany. both from Delft University of Technology, The
After receiving the Ph.D. degree, he worked as Netherlands.
a post-doctoral researcher at the Image Sciences From 1972 to 1988, he was Assistant/Associate
Institute, Utrecht University, The Netherlands. Since Professor of applied mathematics at the same
April 1997, he has been with the computer vision university. Since 1988, he has been a Professor
group, Department of Computer Science, Copen- and Head of the Department of Medical Imaging,
hagen University, Denmark. His current research Utrecht University, The Netherlands, and as of
interests include all aspects of partial differential 1996 Scientific Director of the newly established
equations and scale-space theory in image analysis. Image Sciences Institute, Utrecht University and University Hospital–Utrecht.
Dr. Weickert’s M.Sc. thesis was awarded the Wacker Memorial Prize His research interests comprise all aspects of computer vision and medical
for the best student thesis of the European Consortium for Mathematics in imaging.
Industry. He is author of Anisotropic Diffusion in Image Processing (Stuttgart, Dr. Viergever is co-author of over 200 refereed scientific papers on
Germany: Teubner), and he has given several invited lectures at conferences biophysics and medical image processing, and co-author/editor of 11 books.
and international workshops. He is a Board Member of IPMI and IAPR. He is Editor of the book series
Computational Imaging and Vision (Boston, MA: Kluwer), and serves as
Associate Editor-in-Chief of IEEE TRANSACTIONS ON MEDICAL IMAGING.
He is also Editor of the Journal of Mathematical Imaging and Vision, and
participates on the editorial boards of several journals.
Bart M. ter Haar Romeny (M’91) received the
M.Sc. degree in applied physics from Delft Uni-
versity of Technology, The Netherlands, in 1978,
and Ph.D. degree from Utrecht University, The
Netherlands, in 1983.
After he obtained the Ph.D. degree, he was Princi-
pal Physicist, Utrecht University Hospital Radiology
Department. In 1989, he joined the Image Sciences
Institute, Utrecht University, as an Associate Profes-
sor. His interests are mathematical aspects of front-
end vision, in particular linear and nonlinear scale-
space theory, medical computer vision applications, differential geometry, and
perception.
Dr. ter Haar Romeny is author of a number of papers and book chapters. He
is editor of a recent book on nonlinear diffusion theory in computer vision, and
has initiated a number of international collaborations on these subjects. He is
on the editorial board of the Journal of Mathematical Imaging and Vision, and
was chairman of the First International Conference on Scale-Space Theory in
Computer Vision, Utrecht, 1997.

Authorized licensed use limited to: UNIVERSIDAD DE VALLADOLID. Downloaded on May 17,2010 at 17:26:42 UTC from IEEE Xplore. Restrictions apply.