You are on page 1of 6

J. Vis. Commun. Image R.

23 (2012) 516–521

Contents lists available at SciVerse ScienceDirect

J. Vis. Commun. Image R.


journal homepage: www.elsevier.com/locate/jvci

Anisotropic diffusion for image denoising based on diffusion tensors q


Feng Liu a,⇑, Jingbo Liu b
a
Department of Information Science, School of Science, Xi’an Jiaotong University, Xi’an 710049, PR China
b
Department of Electronic Engineering, Tsinghua University, Beijing 100084, PR China

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, the anisotropic diffusion for image denoising is considered. A new method to construct dif-
Received 12 July 2010 fusion tensors is proposed. The tensors obtained by our approach depend on four directional derivatives
Accepted 17 January 2012 of the intensity of an image, and hence they are adaptively determined by local image structure. It is
Available online 30 January 2012
shown that the proposed diffusion filter is isotropic in the interior of a region, whereas it is anisotropic
at edges. This property of tensors allows us to efficiently remove noise in an image, particularly noise at
Keywords: edges. Several numerical experiments are conducted on both synthetic and real images.
Image denoising
Ó 2012 Elsevier Inc. All rights reserved.
Anisotropic diffusion
Nonlinear diffusion
Diffusion tensor
Diffusion filter
Diffusion direction
Diffusion amount
Multi-scale method

1. Introduction image increases. Hence the diffusion filtering at edges is inhibited.


For this reason, noise at edges cannot be successfully eliminated.
The nonlinear diffusion method for image denoising was first To overcome this problem, Weickert [8] introduced the concept
introduced by Perona and Malik [1]. This method is based on a dif- of diffusion tensors into diffusion models, and established a class of
fusion process governed by a partial differential equation (PDE), the coherence-enhancing diffusion models. The Weickert diffusion
where diffusion amount depends on the gradient of images. With is anisotropic. It can simultaneously enhance edges while smooth-
the Perona-Malik type equation, an image can be selectively ing images, and hence is usually used to enhance an image with
smoothed, which allows us to well preserve edges in an image. flowlike structure. But it is not used for image denoising since
Since the elegant Perona–Malik diffusion model (P–M model), a there exist a lot of the undesired ripples in the filtered image.
wide variety of the theoretical and practical works related to this The objective of this paper is to construct a class of diffusion
topic have been given, such as the scale-space theory [2–4], the var- tensors for image denoising, and generalizes the P–M model to the
iational formulations [5], the well-posed and stable diffusion models anisotropic diffusion filtering. This goal differs from the Weickert
[6–8], and the robust and efficient numerical algorithms [2,9]. Not- model. Here we stress the removal of noise included in images
ing that the diffusion filtering is a kind of multi-scale method, many instead of the enhancing of edges.
authors [10,11] focus their attentions on the connection between Our work is inspired by the Weickert model. But the used
the diffusion filtering and wavelet shrinkage denoising of images. approach to construct diffusion tensor in this paper is distinguished
Chan and Zhou [12] restored an image by wavelet coefficients which from that presented by Weickert [8]. We define a diffusion tensor by
minimize a total variation (TV). Mrazek et al. [13] studied the corre- four directional derivatives of an image. This ensures that the
spondence between diffusivity and shrinkage for Haar wavelet eigenvalues of the diffusion tensor given by our method are adap-
[14,15], and proved that two methods are equivalent. For further tively determined by local image structure.
interesting approaches, we refer to [16–25]. The rest of this paper is organized as follows. We begin in Section
Note that the P–M diffusion is isotropic. It utilizes a scalar-val- 2 with a general review of the P–M model and the Weickert
ued diffusivity that monotonously decreases as the gradient of an approach. In Section 3, we describe in details how to construct a
diffusion tensor, and discuss its basic properties. We in Section 4
q
consider the discretization of the tensor-based diffusion model
This research was supported by the National Natural Science Foundation of
(TDM). In Section 5, the performance of the proposed TDM and the
China under Grant No. 11171270.
⇑ Corresponding author. Fax: +86 029 83237910. C-model is compared. Conclusions and scope for future works are
E-mail address: liuf@mail.xjtu.edu.cn (F. Liu). in Section 6.

1047-3203/$ - see front matter Ó 2012 Elsevier Inc. All rights reserved.
doi:10.1016/j.jvcir.2012.01.012
F. Liu, J. Liu / J. Vis. Commun. Image R. 23 (2012) 516–521 517

2. Nonlinear diffusion filtering

2.1. P–M model

Suppose that the received image is


u0 ¼ u þ n;
where u is an original image, and n is an additive zero-mean Gauss-
ian white noise.
In order to well preserve edges while noise is removed, Perona
and Malik proposed a nonlinear diffusion model [1] (P–M model)
Fig. 1. Filtered result for Lena picture using the Weickert method.
as follows: (
8
l1 ¼ a;
< ut ¼ divðgðjrujÞruÞ ðx; y; tÞ 2 X  ð0; þ1Þ;
> ð5Þ
uðx; y; 0Þ ¼ u0 ðx; yÞ ðx; yÞ 2 X; ð1Þ l2 ¼ a þ ð1  aÞ exp½1=ðk1  k2 Þ2 ;
: @u 
>
@N x2@ X
¼ 0 t 2 ð0; þ1Þ: where k1 and k2 are the eigenvalues of Jq, a is a small positive real
number.
where X and oX denote the support of the image u0(x, y) and its
The Weickert diffusion can well achieve the smoothing and
boundary, respectively, N stands for an unit outward normal to
enhancing of an image with flowlike structure, such as fingerprint,
oX, ru presents the gradient operator, and ‘div’ presents a diver-
the interrupted line can be closed, whereas the semantically impor-
gence operator. g(s) is called a diffusivity or an ‘‘edge-stopping’’
tant singularities are not destroyed. But for a natural image with
function.
texture or corrupted by noise, one often see many undesired ripples
In image processing, the solution u (x, y, t) of the P–M model, for
in the filtered images. As shown in Fig. 1, these ripples lead to the
any fixed t, is regarded as a filtered version of u0(x, y). There exists a
image degradation, where the noisy image (see Fig. 5(a)) is obtained
finite t0 such that the filtered image u (x, y, t0) has the best quality.
from the clear Lena picture (see Fig. 4) by adding an additive
Generally, one assumes that g (s) is a non-negative decreasing
Gaussian white noise with zero mean and standard variance 15.
function satisfying g (0) = 1 and g (s) ? 0 as s ? + 1. With this
assumption, strong edges are less blurred than noise and low-con-
trast details. 3. Tensor-based diffusion models
Hence the edge-stopping function has an important influence
on a diffusion process. This motivates a large number of research- We in this section consider how to construct a diffusion tensor
ers to study the properties of diffusivity, as well as its choice, such for image denoising, and establish a class of the tensor-based
as [2] and the summary of [13]. The two diffusivities suggested by anisotropic diffusion models.
Perona and Malik are of forms as follows:
3.1. Diffusion tensor
gðjrujÞ ¼ 1=ð1 þ jruj2 =k2 Þ; gðjrujÞ ¼ expðjruj2 =k2 Þ; ð2Þ
Since the directional derivative of u along an unit vector n may
where the parameter k plays a role of threshold. The points in an
be written as
image corresponding to jruj > k are regarded as edges, whereas
others points belong to the interior of a region. Depending on the
@u
¼ nT ru;
choice of g (s), the P–M model covers many different nonlinear dif- @n
fusion filters. hence for any edge-stopping function g (), we have
       
Due to the influence of noise on gradients of images, the diffu- @ @u @u @u @u
g   ¼ div g   n
sion filter (1) can often not recognize between the ‘true edges’ and @n @n @n @n @n
   
‘false edges’ caused by noise. To overcome this drawback, Catté @u
et al. [6] considered the following diffusion model (C-model) for ¼ div g   nnT ru : ð6Þ
@n
image denoising, Set D ¼ gðj@u=@njÞnnT . Then D is a diffusion tensor of size 2  2. We
ut ¼ div ðgðjG  rujÞruÞ; ð3Þ see that the vector n is an eigenvector of D, and its corresponding
eigenvalue is g (j@u/@nj).
where G is a Gaussian function. The C-model is well-posed, and makes Generally, a directional derivative can not exactly characterize
the non-linear diffusion filter insensitive to noise. the local image structure since it only represents the rate of change
in u along a given direction. In this sense, D depending on the
2.2. Weickert model directional derivative @u/@n may be not very efficient.
A possible improvement is to use several directional derivatives
The smoothing-enhancing diffusion model based on diffusion to construct diffusion tensors. It is noted that for a digital image,
tensors proposed by Weickert [8] is as follows: there exist four directions inpffiffiffia 3  3 neighborhood, pffiffiffii.e., n1 ¼
ut ¼ div ðDðJ p ðrur ÞÞruÞ; ð4Þ ð1; 0ÞT ; n2 ¼ ð0; 1ÞT ; n3 ¼ ð1; 1ÞT 2=2 and n4 ¼ ð1; 1ÞT 2=2. Thus
we define a diffusion tensor by
with the boundary condition: h D (Jq(r ur))ru, Ni = 0. Here, ur is  
X4  @u 
the convolution of u with a Gaussian kernel kr, and Jq ðrur Þ ¼ kq  D¼ g   nk nT : ð7Þ
rur ruTr . D stands for a mapping belonging to C2 (S2, S2) in which k¼1
@nk  k

S2 denotes the set of symmetric matrices of size 2  2. Thus D


The matrix D may be expressed as
(Jq(rur)) is a matrix, and called a diffusion tensor.  
The Weickert model assumes that two eigenvectors of the diffu-
1 AþC B
D ¼ Dðux ; uy Þ ¼ ; ð8Þ
sion tensor are parallel to gradient and edge, respectively. The cor- 2 B EþC
responding eigenvalues are chosen by the desired goal of filter. For where A = 2g(juxj), B = g(j@u/@n3j)  g(j ou/@n4j), C = g(j@u/@n3j) +
example, if we prefer smoothing along edge to smoothing across it, g(j@u/@n4j) and E = 2g(juyj). It is clear that D is a symmetric, positive
then the corresponding eigenvalues can be chosen as definite matrix.
518 F. Liu, J. Liu / J. Vis. Commun. Image R. 23 (2012) 516–521

3.2. Diffusion models 4. Discretization for the tensor-based diffusion model

By (8), we define a tensor-based diffusion model (TDM) for 4.1. Discretization


image denoising as follows:
For the discretization of nonlinear diffusion PDEs, a variety of
ut ¼ divðDðux ; uy ÞruÞ: ð9Þ methods [2,9] have been proposed. They can be used to solve (9).
But, in order to be convenient for analyses, we use a hybrid scheme
We see that the form of the PDE (9) is the same as the Weickert
that combines forward and backward differences.
model (4). But the used diffusion tensor in (9) is quite different from
By (6) and (7), the PDE (9) can be rewritten as
the Weickert tensor.
Considering the influence of noise on the directional deriva- X4    
@  @u  @u
tives, we modify the TDM (9) by using Gaussian filtering as ut ¼ g   : ð14Þ
k¼1
@nk @nk  @nk
ut ¼ divðDðGr  ux ; Gr  uy ÞruÞ; ð10Þ Without loss of generalization, let the grid size be 1 and the
where Gr is a Gaussian kernel with zero mean and standard varia- time step Dt = s. In the following discussions, uki;j ði ¼ 1; 2; . . . ; M;
tion r. j ¼ 1; 2; . . . ; N; k P 0Þ denotes the value of u(x, y, t) at the point
(x, y) = (i, j) and the time t ¼ ks; dþ 
n and dn represent the forward
3.3. Properties of diffusion tensor and backward differences at the direction n, respectively. Set
m1 = (1, 0), m2 = (0, 1), m3 = (1, 1) and m4 = (1,  1). Then dþ k
nl ui;j ¼
Generally, the eigenvectors of a tensor represent the diffusion ukði;jÞþml  uki;j , and d k k k
nl ui;j ¼ ui;j  uði;jÞml , for l = 1, 2, 3, 4.
directions of a diffusion filter, and the eigenvalues determine the
By (14), the simpler discretization of the PDE (9) with reflecting
diffusion amount along the corresponding eigenvector. Thus in or- boundary conditions is given by
der to better understand the TDM (9) or (10), we in this subsection
discuss the properties of the eigenvectors and eigenvalues of kþ1
ui;j  uki;j X
4

D(ux, uy) specified by (8). ¼ dþnl g jdnl uki;j j dnl uki;j ; for 1 6 i 6 M; 16j
s l¼1
It is easy to show that the eigenvalues of D (ux, uy) are given by
6 N; k P 0:
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
A þ E þ 2C 1
lk ¼  ðA  EÞ2 þ 4B2 ; k ¼ 1; 2: ð11Þ That is
4 4
X
4

Clearly, l1 P l2 P 0, where l1 corresponds to sign ‘‘+’’ in the above kþ1


ui;j ¼ uki;j þ s dþnl g jdnl uki;j j dnl uki;j :
equation. If B = 0, then the corresponding eigenvectors of D(ux, uy) l¼1
are given by

Set g kl;i;j ¼g jd k


nl ui;j j . Then from the above equation, we have

n1 ¼ n1 ; n2 ¼ n2 ; if A P E; 4
X

ð12Þ kþ1
n1 ¼ n2 ; n2 ¼ n1 ; otherwise: ui;j ¼ uki;j þ s g kl;ði;jÞþml ukði;jÞþml  g kl;ði;jÞþml þ g kl;i;j uki;j þ g kl;i;j ukði;jÞml :
l¼1
If B – 0, then ð15Þ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
n1 ¼ ða þ signðBÞ 1 þ a2 ; 1Þ; By the theory of finite difference, the scheme (15) requires small
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi time step s in order to be stable (see [2] for more details).
n2 ¼ ð1; a  signðBÞ 1 þ a2 Þ; ð13Þ
4.2. Property of anisotropic diffusion model
where a = (E  A)/(2B).
Let the image uki;j ð1 6 i 6 M; 1 6 j 6 NÞ for any fixed k P 0 be
Property 1. (i) hn1, n2 i = 0; (ii) l1 = l2 if and only if r u = 0; (iii)
extended by reflecting boundary. Then uk0;j ¼ uk1;j ; ukMþ1;j ¼ ukM;j ;
l1 > l2 if and only if ru – 0.
uki;0 ¼ uki;1 and uki;Nþ1 ¼ uki;N for all i, j and k P 0. Using (15), we can
Property (i) indicates that n1 and n2 are mutually orthogonal, get, for all k P 0
which is important for achieving anisotropic diffusion. Property
1 X M X N
1 X M
s X M X N
(ii) and (iii) show that the filter (9) is isotropic at points belonging ukþ1
i;j ¼ uki;j þ
M  N i¼1 j¼1 M  N i¼1 M  N i¼1 j¼1
to a homogeneous region, but it is anisotropic at edges as expected.
For the anisotropic diffusion at edges, we need to further dis- 4
X

cuss the relations between the eigenvectors and ru so as to under-  g kl;ði;jÞþml ukði;jÞþml  g kl;i;j uki;j
stand the diffusion filter behavior. By the (iii) and (11), ru – 0 if l¼1

and only if (A  E)2 + B2 – 0. With this equivalence, we have the s M X


X N

following results. þ
MN i¼1 j¼1

2 2 4
X

Property 2. Let (A  E) + B – 0. (i) If (A  E)B = 0, then ru is


 g kl;i;j ukði;jÞml  g kl;ði;jÞþml uki;j
parallel to n2, i.e., r ukn2; (ii) If (A  E)B – 0, then j/(ru, n2)j < p/4 as
l¼1
ux P 0, and j/(ru, n2)j < p/4 as ux < 0, where /(x1, x2) denotes the
degree of angle between the vector x1 and x2. 1 X M X N
¼ uk : ð16Þ
M  N i¼1 j¼1 i;j
Property 2 shows that n1 is parallel or close to edge. Since the
corresponding eigenvalue l1 > l2 at edges in images, hence the The above equation indicates that the discrete scheme (15) of the
TDM (9) prefers diffusion along an edge to diffusion along gradient. TDM (9) can preserve the average intensity of an image.
Combining Property 1, Property 2 shows that the TDM is able to To consider maximum-minimum principle for the PDE (9), we
remove noise at edges. rewrite (15) as
F. Liu, J. Liu / J. Vis. Commun. Image R. 23 (2012) 516–521 519

Noting that 0 < g kl;i;j 6 1, it is easy to show that if s < 1/8, then

mini;j uki;j 6 ui;j


kþ1
6 maxi;j uki;j ; for all k P 0: ð17Þ

Eq. (17) indicates that the explicit scheme (15) obeys the maxi-
mum–minimum principle when s < 1/8. This property forbids un-
der-and overshoots in diffusion filtering. It also ensures that iso-
intensity linking toward the original image is possible.

5. Experimental results

We in this section focus our attentions on numerical experi-


ments with image denoising to test the proposed TDM (10) on both
synthetic and real gray level images. Results obtained by the C-
model and the proposed model are displayed for comparisons.
In all experiments, we employ the same explicit scheme like
(15) as an approximation to the used continuous models. The time
step s is taken to be 0.01. The edge-stopping function is chosen as
1
gðsÞ ¼ :
1 þ s2 =k2
It should be pointed that the Weickert filter is a kind of coher-
ence-enhancing anisotropic diffusion, as mentioned in Section 2.2.
It is not used for image denoising. Thus we in this subsection do
not discuss the performance of the Weickert method.

Fig. 2. Filtered results for the synthetic image: (a) original image; (b) noisy image, 5.1. Synthetic image
r = 50, PSNR = 14.17 dB; (c) result of using the C-model, PSNR = 30.15 dB; (d) result
of using the TDM, PSNR = 34.79 dB.
We first test the TDM (10) on a synthetic image to illustrate the
effectiveness of the diffusion filtering at edges. The clear image
Table 1 consists of a triangle and a rectangle, as shown in Fig. 2(a). The
PSNR (dB) of the smoothed images by using the C-model and the proposed model, noisy data (see Fig. 2(b)) is generated from the clear image by add-
where r presents the standard variance of the Gaussian noises, CM and TDM denotes ing an addition Gaussian white noise with zero-mean and standard
the C-model and the proposed model, respectively.
variance r = 50. Fig. 2(c) and (d) depict the denoised results which
r Lena Boat have the largest peak signal-to-noise ratio (PSNR) (in dB). The PSNR
CM TDM CM TDM is defined by
50 25.01 25.15 23.83 24.04
30 26.75 27.29 25.80 26.16 2552
PSNR ¼ 10log10 P ; ð18Þ
25 27.43 28.10 26.29 26.89  j2 =ðM  NÞ
ju  u
20 28.33 29.14 27.44 27.84
15 29.62 30.53 28.50 29.13  denote the clear image and the denoised image,
where u and u
10 31.72 32.65 30.83 31.16
respectively.
We observe that the C-model performs well in the interior of re-
4
gions, but fails to efficiently remove noise at edges. The TDM yields
X
ukþ1 ¼s g kl;ði;jÞþml ukði;jÞþml þ g kl;i;j ukði;jÞml better smoothing results for this example. An improvement in the
i;j
l¼1 result can be perceived by observing the restoration of the edges of
!
4
X
the triangle and rectangle. In addition, the background of Fig. 2(c)
þ 1s g kl;ði;jÞþml þ g kl;i;j uki;j : is slight darker than Fig. 2(d) (in version). This indicates that
l¼1 Fig. 2(c) includes more remained noise.

Fig. 3. Comparisons between the C-model and TDM: (a) result for Lena picture; (b) result for Boat picture.
520 F. Liu, J. Liu / J. Vis. Commun. Image R. 23 (2012) 516–521

Fig. 6. Positions of the vector ru (ux > 0) and n2.

edges in the mirror. The detailed differences are shown in


Fig. 4. From left to right: Boat and Lena picture of size 256  256. Fig. 5(d), where the edges of hat are zoomed. Compared with
Fig. 1, there are not the undesired ripples in Fig. 5(c). This indicates
that the diffusion tensors given by our method are better suited for
the image denoising.

6. Conclusions

We have considered the methods to construct diffusion tensors


for images denoising. The tensor depends on the local image struc-
ture. Its eigenvector corresponding to larger eigenvalue is parallel
or close to the edge. This guards that we obtain a better perfor-
mance of the proposed anisotropic diffusion filtering.
For the TDM, there are a variety of theoretical analyzes and
applications to be done, e.g., the existence of the solutions of the
proposed model, the discrete scheme of the continuous model.
Our future efforts will be focused toward further improving the dif-
fusion tensor such that it can be applied to the speckle removal of
images, such as SAR and medical images.

Appendix A

Proof of Property 1. From (12) and (13), it is easy to check directly


that the equality (i) holds. Thus we only need to verify Property (ii).
Let l1 = l2. By (11), we have A = E and B = 0. Together with the
monotonicity of g(s), it is derived that juxj = juyj and j@u/@n3j = j@u/
@n4j. That is,
pffiffiffi pffiffiffi
2 2
ux ¼  uy ; jux þ uy j ¼ jux  uy j: ð19Þ
2 2
Fig. 5. Denoised results for Lena picture: (a) noisy Lena picture; (b) C-model,
PSNR = 29.60 dB; (c) TDM, PSNR = 30.54 dB; (d) Left picture and right picture are the It follows from (19) that ru = 0. Conversely, it is easy to show that
zoomed edges of hat in figure (b) and (c), respectively. l1 = l2 if ru = 0. The proof of Property 1 is completed. h

5.2. Real image Appendix B

The goal in this application is to evaluate the performance of the


TDM (10). We performed a series of numerical experiments with Proof of Property 2. (i) It is clear that ru – 0. By the condition
image denoising. The original images are all of size 256  256 (Boat (A  E)B = 0, either A  E = 0 or B = 0. If B = 0 and A > E, then we have
picture and Lena picture), as shown in Fig. 4. g(j@u/@n3j) = g(j@u/@n4j) and g(juxj) > g(juyj), i.e., j@u/@n3j = j ou/@n4j
We added several levels of additive zero-mean white Gaussian and juxj < juyj. By the second equality of (19), it is easy to show that
noises to the original images, and used different diffusion methods ux = 0. On the other hand, n2 = n2 by (12). Thus r ukn2. Similarly, it
to search for the optimal solution that maximizes the PSNR of the can be shown that if B = 0 and A < E, then r ukn2.
filtered image. Table 1 summarizes the average optimal PSNR from If B – 0 and A  E = 0, then a = 0. Together with (13), we have
five realizations by the C-model and the TDM, respectively. The
jux j ¼ juy j; n1 ¼ ðsignðBÞ; 1Þ; n2 ¼ ð1; signðBÞÞ: ð20Þ
comparisons between two models are shown in Fig. 3(a) and (b).
We see that the proposed diffusion method in this paper always In the case of B > 0, we p have
ffiffiffi j@u/@n3j < j ou/@n4j and
pffiffiffi n2 = (1,  1).
yields greater PSNR than the Catté model in each case. Since j@u=@n3 j ¼ jux þ uy j 2=2; j@u=@n4 j ¼ jux  uy j 2=2. Hence it
Fig. 5 gives an example of the diffusion denoising using two dif- follows from (20) that ru = (ux,  ux). In the case of B < 0, similarly,
fusion models, where the Lena picture is embedded in an additive it can be shown that ru = (ux, ux) and n2 = (1, 1). To sum up, we can
white Gaussian noise with zero mean and standard variance conclude that rukn2.
r = 15.The parameter k is taken to be 70 for the C-model, whereas (ii) If (A  E)B – 0, then we have a – 0 and B – 0. In this case, n2
k = 8 for the TDM. We see that the TDM exhibits better denoising is given by (13). The orientation of n2 depends on the sign of a and
behavior at edges as expected, such as boundary of shoulder and B.
F. Liu, J. Liu / J. Vis. Commun. Image R. 23 (2012) 516–521 521

Let ux > 0 and h = argru. By the monotonicity of g(s) and [10] Y.L. You, W.Y. Xu, A. Tannenbaum, M. Kaveh, Behavioral analysis of anisotropic
diffusion in image processing, IEEE Trans. Image Process. 5 (11) (1996) 1539–
EA gðjuy jÞ  gðjux jÞ 1553.
a¼ ¼ ; [11] M.J. Black, G. Sapiro, D.H. Marimont, D. Heeger, Robust anisotropic diffusion,
2B gðj@u=@n3 Þ  gðj@u=@n4 jÞ
IEEE Trans. Image Process. 7 (3) (1998) 1539–1553.
[12] T.F. Chan, H.M. Zhou, Total variation improved wavelet thresholding in image
we have the following results, as shown in Fig. 6,
compression, in: Proceedings Seventh International Conference on Image
8 Processing, vol. 2, 2000, pp. 391–394.
> h 2 ðp=4; 0Þ if B > 0 and a > 0;
>
> [13] P. Mrazek, J. Weickert, G. Steidl, Correspondences between wavelet shrinkage
< h 2 ðp=2; p=4Þ if B > 0 and a < 0; and nonlinear diffusion, in: L.D. Griffin, M. Lillholm (Eds.), Scale-Space, LNCS
2695, Springer-Verlag, Berlin, 2003, pp. 101–116.
>
> h 2 ð0; p=4Þ if B < 0 and a < 0;
>
: [14] D.L. Donoho, De-noising by soft-thresholding, IEEE Trans. Inform. Theory. 41
h 2 ðp=4; p=2Þ if B < 0 and a > 0: (3) (1995) 613–627.
[15] D.L. Donoho, I.M. Johnstone, Ideal spatial adaptation via wavelet, shrinkage,
Therefore j/(ru, n2)j < p/4. Similarly, we can verify that j/(ru, Biometrika 81 (1994) 425–455.
[16] M. Nitzberg, T. Shiota, Nonlinear image filtering with edge and corner
n2)j < p/4 when ux < 0. The proof of property 2 is completed. h
enhancement, IEEE Trans. PAMI 14 (1992) 826–833.
[17] J. Weickert, Theoretical foundations of anisotropic diffusion in image
processing, Comput. Suppl. 11 (1996) 221–236.
References [18] U. Clarenz, U. Diewald, M. Rumpf, Processing textured surfaces via anisotropic
geometric diffusion, IEEE Trans. Image Process. 13 (2) (2004) 248–261.
[1] P. Perona, J. Malik, Scale-space and edge detection using anisotropic diffusion, [19] L. Alvarez, P.L. Lions, J.M. Morel, Image selective smoothing and edge detection
IEEE Trans. PAMI 12 (7) (1990) 629–639. by nonlinear diffusion II, SIAM J. Numer. Anal. 29 (3) (1992) 845–866.
[2] G. Aubert, P. Kornprobst, Mathematical Problems in Image Processing: Partial [20] F. Malgouyres, Combining total variation and wavelet packet approaches for
Differential Equations and the Calculus of Variations, Springer, New York, 2002. image debluring, in: Proceedings IEEE Workshop, 2001, pp. 57–64.
[3] L. Alvarez, F. Guichard, P.L. Lions, J.M. Morel, Axioms and fundamental [21] M. Lysaker, A. Lundervold, X.C. Tai, Noise removal using fourth-order partial
equations of image processing, Arch. Ration. Mech. 123 (1993) 200–257. differential equation with applications to medical magnetic resonance images
[4] J. Weickert, Nonlinear diffusion scale-spaces: From the continuous to discrete in space and time, IEEE Trans. Image Process. 12 (12) (2003) 1579–1590.
setting, in: M.O. Berger et al. (Eds.), Proceedings ICAOS’96: Images, Wavelets [22] S. Osher, L.I. Rudin, Feature-oriented image enhancement using shock filters,
and PDE’s 219, Springer, New York, 1996, pp. 111–118. SIAM J. Numer. Anal. 27 (4) (1990) 919–940.
[5] L.I. Rudin, S.Osher.E. Fatemi, Nonlinear total variation based noise removal [23] G. Gilboa, N. Sochen, Y.Y. Zeevi, Forward and backward diffusion processes for
algorithms, Physics D 60 (1992) 259–268. adaptive image enhancement and denoising, IEEE Trans. PAMI 11 (7) (2002)
[6] F. Catté, P.L. Lions, J.M. Morel, T. Coll, Image selective smoothing and edge 689–703.
detection by nonlinear diffusion, SIAM J. Numer. Anal. 29 (1992) 182–193. [24] Y. Wang, L.P. Zhang, P.X. Li, Local variance-controlled forward-and-backward
[7] G. Gilboa, N. Sochen, Y.Y. Zeevi, Image enhancement and denoising by complex diffusion for image enhancement and noise reduction, IEEE Trans. Image
diffusion processes, IEEE Trans. PAMI 26 (8) (2004) 1020–1036. Process. 16 (7) (2007) 1854–1864.
[8] J. Weickert, Anisotropic diffusion in image processing, Stuttgart, Teubner, [25] Y. Yue, M.M. Croitoru, A. Bidani, et al., Nonlinear multiscale wavelet diffusion
Germany, 1998. for speckle suppression and edge enhancement in ultrasound images, IEEE
[9] J. Weickert, Efficient and reliable schemes for nonlinear diffusion filtering, IEEE Trans. Medical Imag. 25 (3) (2006) 297–311.
Trans. Image Process. 7 (3) (1998) 398–410.

You might also like