Computers & Mathematics

With Applications
Certificate of publication for the article titled:
"A new rapid auto-adapting diffusion function for adaptive
anisotropique image de-noising and sharp conserved edges "

Authored by:
Zouhair Mbarki *, Hassene Seddik, Sondes Tebini, Ezzedine Ben Braiek

Published in:
Volume 74, Issue 8 (2017), Pages 1751-1768
 Computers and Mathematics with Applications 74 (2017) 1751–1768

Contents lists available at ScienceDirect

Computers and Mathematics with Applications

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

A new rapid auto-adapting diffusion function for adaptive

anisotropic image de-noising and sharply conserved edges
Zouhair Mbarki *, Hassene Seddik, Sondes Tebini, Ezzedine Ben Braiek
University of Tunis, Department of Electrical Engineering, CEREP, ENSIT, 5, Av. Taha, Hussein, 1008, Tunis, Tunisia

article info a b s t r a c t
Article history: Anisotropic diffusion, based on partial differential equation (PDE), is a recent adequate
Received 2 July 2016 solution for the problem of image filtering. The first works in this context are those of
Received in revised form 5 May 2017 Perona and Malik. Recently, several studies have shown the drawbacks of this approach
Accepted 17 June 2017
such as the ’’staircase’’ effect and flow edges caused by the slow convergence of the
Available online 13 July 2017
diffusion function. In this work, we suggest a new diffusion function, which converges
faster than that of Perona and Malik. The suggested function decreases rapidly to disappear
Keywords:
Efficient image de-noising once borders or details are detected. This rapidity to expand and converge to zero allows
Noise reduction us to implement a real time processing device. Moreover, the suggested model is able to
Fast anisotropic diffusion remove the ’’staircase’’ effect, preserve sharp transition and discontinuities and remove
Edge preserving noise efficiently. The diffusion barrier is chosen to get rid of the noise and enhance the
edges. Extensive experiments on several standard test images are conducted to compare
our algorithm with other well-known algorithms. Experimental results are very interesting
and show the efficiency of the suggested method based on a comparison study.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction

Since the image is often infected by noise, the improvement of the image quality has now become a necessity in many
application domains such as the military and medical. The main goal of de-noising is to remove the noise while retaining
as much signal features as possible. A vast literature has emerged recently on image restoration using linear or nonlinear
techniques. For instance, anisotropic diffusion methods have become a very useful tool in image smoothing, edge detection,
image segmentation, and image enhancement [1] and [2]. Anisotropic diffusion filtering can successfully smooth noise while
respecting the region boundaries and small structures within the image. It depends on the norm of the gradient which occurs
in the so-called diffusivity function in the heat equation in order to control the diffusion. This diffusivity function C ∥∇ U ∥
is between [0 1]. If the diffusivity coefficient called C is equal to zero, the distribution will be stopped. Conversely, when C
is equal to ‘‘1’’, it is isotropically scattered by the heat equation. The most well-known diffusion function was suggested by
Perona and Malik [3]. Despite convincing experimental results, several authors have raised the drawbacks of this approach.
In case of a high-level noise, this approach intensifies it. This is due to the non-convergence speed function diffusivity, also
it tends to cause the so called ‘‘staircase’’ effect [4] and [5]. To overcome this problem, a great amount of studies have been
conducted which led to positive results. Catté, Lions, Morel and Coll introduce a convolution of the image with a Gaussian
kernel of standard deviation σ before each iteration [6] and [7], Alvarez, Lion and Morel suggested to solve the following

* Corresponding author.
E-mail addresses: mbarki.zouhair84@gmail.com (Z. Mbarki), Hassene.seddik@esstt.rnu.tn (H. Seddik), tebinisondes@yahoo.fr (S. Tebini),
Ezzedine.benbraiek@esstt.rnu.tn (E.B. Braiek).

http://dx.doi.org/10.1016/j.camwa.2017.06.026
0898-1221/© 2017 Elsevier Ltd. All rights reserved.
1752 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768

PDE [8].
⎨ ∂u
⎧ ( )
∇u
(x, y, t) = g(∥∇ uσ ∥) ∥∇ u∥ div
∂t ∥∇ u∥ (1)
u(x, y, 0) = u0 (x, y)
⎩

where µσ is the smoothed image with a Gaussian kernel of standard deviation σ . Kornprobst suggested a very general
reaction–diffusion model that combines a shock filter to a selective smoothing. The equation of the model contains a diffusion
term associated with the measurement of the average curvature, a feedback term based on the shock filters theory developed
by Osher and Rodin [9] and a term attached to the data to obtain a solution close to the original image [8]. In other work,
You et al. in [10] suggest to solve the following diffusion function which eliminates the ‘‘staircase’’ effect [11].
1
⎧
⎨g(x) = + p(K + ε)p−1 if x < K
⎪
K
(2)
⎩g(x) = 1 + p(x + ε)p−1 if x > K .
⎪
x
In related work [12], authors proposed to remove noise from polluted images such as classical total variation scheme (TV
model) which gives a better effect with tiny staircases to a certain degree through the using of gradients alone. Recently,
an anisotropic diffusion model with a semi-adaptive threshold in diffusion coefficient function (SAT) is applied to get a
restored image [13]. In this work, we suggest a new diffusion function C ∥∇ U ∥ with better mathematical characteristics
which converges faster than the one of the Perona and Malik. This new model can alleviate the ‘‘staircase ’’effect, preserve
sharp discontinuities and remove noise simultaneously. The diffusion barrier is chosen so that noise is removed and edges
are enhanced. The significant advantage is the possibility to implement real time processing based on this algorithm thanks
to its rapidity of convergence. The experimental results are promising and encouraging. A comparative study is also provided
to prove the efficiency of the suggested method.
The paper is organized as follows. In the next section, we first introduce the anisotropic diffusion. An efficient model
proposed by Perona and Malik is then reviewed. In Section 3, we review the limitations of Perona and Malik scheme and we
cite the recent work to overcome this problem, then we present the new model. Section 4 describes the results and Section 5
describes the results evaluation. Section 6 is devoted to the conclusion of this paper.

2. Anisotropic diffusion and Perona and Malik formulation

Anisotropic Diffusion is an example of scale-space image processing [3]. It has transformed the application of the heat
diffusion equation to digital images. Considering an image U, the heat flux φ within the image is given by Fick’s equation [14].

φ = −D · ∇ U (3)

where: D is diffusion tensor in general it is a definite positive, symmetric matrix expresses the relationship between the
gradient ∇ U and the actual flux φ .
Since the diffusion process does not alter the overall energy in the image, its local variation is driven by δt = −divφ ,
where div stands for the divergence operator [14] and [15]. Hence, we have:

δt U = div (D · ∇ U). (4)

In relation to the image, Eq. (2) expresses the energy (heat) variation at each pixel position. The diffusion algorithm
smoothes the image based on partial differential equation ‘PDE’ [16].
Perona and Malik applied an anisotropic process that reduces the diffusivity and consequently the smoothing strength
proportionally to the probability of edge presence.
The method is to evolve an image under the following PDE.
∂U
= div [c(x, y, t)∇ U(x, y, t)] (5)
∂t
withU(x, y, 0) = U0 (x, y). (6)

The conductivity c(x, y) is developed to promote smoothing intra regional and penalize inter regional diffusion through
the use of a decreasing function dependent on the magnitude of the vector gradient.

c(x, y, t) = g (|∇ U |) . (7)

The functions suggested initially by Perona and Malik [17]:

[ ( ) ]
|∇ U | 2
g1 (|∇ U |) = exp − (8)
K
 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768 1753

and
1
g2 (|∇ U |) = [ |∇ U | ]2 . (9)
1+ K

The parameter K is called a gradient threshold parameter; it plays a very important role in the behavior of the diffusion.
In [18], the necessity of this parameter to be a decreasing function of time was first shown.
Perona and Malik sampled their anisotropic diffusion equation to become [1]:
β ∑ (⏐⏐ ⏐)
Ut +1 (x) = Ut (x) + gk ∇ Ux,p ⏐ ∇ Ux,p (10)
|ηx | p∈η
x

where U is a discretely sampled image, x denotes the pixel position in the discrete 2-D grid, t denotes the iteration step, g
is the conductance function and K is the gradient threshold parameter. Constant β ∈ [0 1] determines the rate of diffusion
and ηx = {North, South, East and West} represent the spatial 4-pixels neighborhood of pixel x respectively. Consequently,
ηx is equal to 4 (except for the image borders). The difference between neighboring pixels in each direction:

∇ Ux,p = Ut (p) − Ut (x), p ∈ ηx = {N , S , E , W } . (11)

The differences with the closest neighbors in the four directions (North, South, East and West) and the diffusion coefficient
are calculated as follows:
∇N Ui,j = Ui−1,j − Ui,j ,
(⏐ ⏐)
cNi,j = g ⏐∇N Ui,j ⏐
∇S Ui,j = Ui+1,j − Ui,j ,
(⏐ ⏐)
cSi,j = g ⏐∇S Ui,j ⏐
(12)
∇E Ui,j = Ui,j+1 − Ui,j ,
(⏐ ⏐)
cEi,j = g ⏐∇E Ui,j ⏐
∇w Ui,j = Ui,j−1 − Ui,j , cwi,j = g ⏐∇w Ui,j ⏐ .
(⏐ ⏐)

However, the pixel value can be modified using:

U(x, y) = U(x, y) + λ[cN · ∇N U + cS · ∇S U + cE · ∇E U + cw · ∇w U]x,y , (13)

with 0 ≤ λ ≤ 1/4.

3. Limitation of Perona and Malik scheme and the suggested model

In their work, from the 1D model of the diffusion equation, Whitaker and Pizer showed that for steep profiles, some
choices for the diffusion barrier can introduce instability called ‘‘staircase’’ effect [19]. To study the enhancement, they also
analyze the neighborhood of the inflection point and show that the abrupt edges can turn ramp if the derivative of the profile
slope in a neighborhood x of inflection point x0 verify.
∂ Ux ∂ U x0
≥ . (14)
∂t ∂t
Since the Perona–Malik function g1 (x) does not converge quickly giving blurred edges and fine details removed. In this
section, we present a modified anisotropic diffusion model which converges faster than the Perona and Malik function. This
rapidity to expand and converge to zero allows us to implement a real time processing device. The suggested method enables
to preserve edges, remove the ‘‘staircase’’ effect and keeps all details during the diffusion process. The new diffusion function
is obtained by introducing a speed attractor β (x) that attracts g1 (x) faster to the convergence limits. The new diffusion
functions noted g3 (x) is defined as follows.

g3 (x) = β (x) · g1 (x). (15)

The speed attractor β (x) is a term that must satisfy the mathematical following constraints.

• β (x) must be a decreasing function considering the interval [1, 0] and converges to its limits in both values 0 and ∞.

lim β (x) = 0 (16)

x→∞

• β (x) must converges faster than g1 (x) to attract it to the above defined limits.

These constraints allow the anisotropic function to expand if the processes window is a homogeneous zone and to
decrease or disappear if details or edges and borders are detected. Mathematically, we have to transform these constraints
into a mathematical attractor whose function satisfies the following requirements:
–Continuous on the interval [0, 1] which means that it must satisfy the three following conditions.
1754 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768

2
Fig. 1. Profile 1D of ex and ex .

10

8
functions variation

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
X

2
Fig. 2. Profile 1D of e−x (r = 1, 2, 3, 4, 5 . . . 9).

(i) β (a) is a defined function

(ii) limx→a β (x) exist (i.e. is finite)

(iii) lim β (x) = f (a) (17)

x→a

–Conservative which means that its loss less over its field.

fof −1 (x) = Id (18)

where Id stands for the identity.

–Its derivative must conserve unchanged the behavior of the anisotropic function since the anisotropic process is supervised
by the gradient ‘‘derivative’’ of the image map in each pixel location.
For all these reasons and mathematical requirements, we are forced to choose the exponential ‘‘ex ’’ as the core of this
attractor. Since the exponential evolves over the interval [0 , ∞], we are forced to apply some mathematical transformations
to improve its velocity and carry it to the interval [0, 1].

β1 (x) = ex . (19)
−xr
‘‘ex ’’ is replaced by the ‘‘e ’’ where r belongs to N ∗ to make the function more rapid especially where sudden transitions
are detected as illustrated by the following figure.
2
β2 (x) = ex . (20)
−x −xr
Since the function e is slowly converging and the functions e does not decrease rapidly for r higher than 2 although
they are rapidly converging and since the complexity of calculation which makes real-time implementation difficult. Indeed,
2
the function e−x is in the middle of all curves therefore it is faster in term of convergence than e−x although it is optimal
compared to the other functions since the first phase of convergence is between [0 1] as shown in Fig. 2. For these reasons,
2
we are forced to choose β2 (x) = e−x (see Figs. 1 and 3).
2
The following step consists in removing the imaginary part while conserving the real one by adding e−x .
2
eix = cox(2x) + i sin(2x) (21)
 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768 1755

Fig. 3. Profile 1D of β2 (x).

Fig. 4. Profile 1D of β4 (x).

2
e−ix = cox(2x) − i sin(2x) (22)
ix2 −ix2
e +e = 2cox(2x) (23)
x2 −x2
β3 (x) = e + e (24)
β4 (x) is a function that decreases as illustrated by the following function.
1 1
β4 (x) = = . (25)
β3 (x) 2
ex + e−x
2

β4 (x) is defined over the interval [0, 0.5], we have to normalize it into [0, 1] see Eq. (26) and Fig. 5.
2
β5 (x) = 2 × β4 (x) = 2
. (26)
ex + e−x2
We obtained finally an attractor called β = β5 (x) that is more rapid in convergence and able to attract g1 from 1 to 0 in case
of sharp transitions are detected in the image (see Fig. 4).
Let
2
β (x) = 2 2
. (27)
ex + e−x
|∇ U |
Such as: x = K
with (see Fig. 6):
lim β (x) = 1
{
x→0
(28)
lim β (x) = 0.
x→+∞

To attract g1 (x) to zero when x tends to +∞ more rapidly, we will supervise it by the behavior of β (x) which must behave
as a rapid attractor which conserves the mathematical characteristics of g1 (x). This fact will increase the convergence speed
1756 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768

Fig. 5. Profile 1D of β5 (x).

0.8

0.6
β(x)

0.4

0.2

0
0 5 10 15 20 25 30 35 40 45 50
X

Fig. 6. Profile 1D of β (x).

of g3 (x) and then diminish the processing time and turn to zero the diffusion function near edges, details and borders. This
fact will de-noise rapidly the image and preserve its content and details.
Therefore:
2 2 2 2
ex + e−x ex − e−x
g3 (x) = β (x) · g1 (x) = 2 2
− 2 2
. (29)
ex + e−x ex + e−x
We set y = x2 , then:
ey − e−y
g3 (y) = 1 − = 1 − th(y), (30)
ey + e−y
with th is the hyperbolic tangent of y,
|∇ U |
and since x = K
.
So
( ) (( )2 )
|∇ U | |∇ U |
g3 = 1 − th (31)
K K

where |∇ U | is the gradient magnitude, and g3 (|∇ U |) is an ‘‘edge-stopping’’ function. This function is chosen to satisfy
g(x) → 0 when x → ∞ so that the diffusion is ‘‘stopped rapidly’’ across edges. All edges and details even if they are
thin, they will be kept intact as soon as they are detected based on the adjustment threshold K . When x = 1 i.e., |∇ U | = k
we obtain:

g1 (x) = 2.7183 (32)

β (x) = 0.6481 (33)

g3 (x) = β (x) · g1 (x) = 1.4117. (34)

This shows clearly that in the vicinity of the edges (|∇ U | = k) this function g3 (x) converges more rapidly than the Perona
and Malik function which preserves edges and small details. The function g3 (x) as shown in Fig. 7.
In this work, for each processed pixel eight neighbors are considered. The differences with the closest neighbors in the
eight directions (North, South, East, West, Northeast, Northwest, Southeast, and Southwest) and the diffusion coefficient are
 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768 1757

0.9

0.8

0.7

0.6

g3(x)
0.5

0.4

0.3

0.2

0.1

0
0 50 x 100 150

Fig. 7. Profile 1D of g3 (x).

calculated as follows.
∇N Ui,j = Ui−1,j − Ui,j ,
(⏐ ⏐)
cNi,j = g ⏐∇N Ui,j ⏐
∇S Ui,j = Ui−1,j − Ui,j ,
(⏐ ⏐)
cSi,j = g ⏐∇S Ui,j ⏐
∇E Ui,j = Ui−1,j − Ui,j ,
(⏐ ⏐)
cEi,j = g ⏐∇E Ui,j ⏐
∇W Ui,j = Ui−1,j − Ui,j ,
(⏐ ⏐)
cWi,j = g ⏐∇W Ui,j ⏐
(35)
∇NE Ui,j = Ui−1,j − Ui,j ,
(⏐ ⏐)
cNEi,j = g ⏐∇NE Ui,j ⏐
∇NW Ui,j = Ui−1,j − Ui,j ,
(⏐ ⏐)
cNWi,j = g ⏐∇NW Ui,j ⏐
∇SE Ui,j = Ui−1,j − Ui,j ,
(⏐ ⏐)
cSEi,j = g ⏐∇SE Ui,j ⏐
∇SW Ui,j = Ui−1,j − Ui,j , cSWi,j = g ⏐∇SW Ui,j ⏐ .
(⏐ ⏐)

However, once processed the pixel value is modified as follows:

U(x, y) = U(x, y) + λ[cN · ∇N U + cS · ∇S U + cE · ∇E U + cw · ∇w U + cNE · ∇NE U + cNW · ∇NW U

+ cSE · ∇SE U + cSW · ∇SW U ]x, y. (36)

4. The estimation of gradient threshold parameter

The choice of the threshold is important since it will determine the level of details to retain. This constant serves as a
contrast parameter separating forward low contrast from high contrast. The diffusion barrier controls the manner in which
the contours involved in the diffusion process. If K is high, the results are close to that isotropic diffusion, for low values, the
edges corresponding to a norm gradient bigger than K will be kept and even enhanced. In the literature several methods
have been suggested to precisely estimate the gradient threshold parameter. First, K can be estimated by the user based on
a priori knowledge. For Perona and Malik, K is chosen by the use of the ‘‘noise estimator’’ described by Canny [20], so that
K is set equal to the 90% values of the histogram integral (histogram of the absolute values of the gradient) throughout the
image in every iteration.
In other suggested approaches Black et al. in [20] and [8] defined K as.

k = 1.4826MAD (∇ U ) (37)

where MAD denotes the median absolute variant and is defined as:
MAD = median (∥∇ U − median (∥∇ U ∥)∥) and the term 1.4826 corresponds to a median absolute deviation of a normal
distribution of unit variance.
In this work, we define K as the optimal value of each image, if we smooth each image iteratively with different value of
K , we can obtain a curve as shown in Fig. 7 defining the behavior of the PSNR according to the variation of the threshold. An
optimal value of K can be extracted graphically from this curve by detecting the tangents representing the optimal value as
the maximum of the PSNR.

K = Kopt if PSNR = PK . (38)

Such as (see Fig. 8):

PK = max(PSNR(K )). (39)

K
1758 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768

Fig. 8. PSNR variation depending on the parameter k.

0.9

0.8

0.7
function variation

0.6

0.5

0.4

0.3

0.2

0.1

0
0 50 100 150
x

Fig. 9. Profile 1D of g1 and g3 diffusion functions.

In order to be able to compare the efficiency of the proposed function, we define the flow function φ as:

φ (x) = g(x) · x. (40)

The flow function represents the sum of the brightness flow that is generated. The maximum flow is generated at the
locations where |∇ U | = k. We scaled the conductance functions g1 and g3 and their respective flow functions as shown in
Fig. 9, given by Eq. (40), producing the same amount of brightness flow at the same point x = K .
From Fig. 9, we can see easily that the value of coefficient c ∥∇ u∥ for the proposed method is smaller than that in P&M
method. Therefore, the detail information in edge and boundary areas is preserved and noise is removed.
To compare the convergence speed of the suggested function and the Perona–Malik function, we calculate the coefficient
of both within a constant ε = 10−3 .

g3 (x) = 10−3 i.e., 1 − tanh((x)2 ) = 0.001. (41)

So x = 7.4833

g1 (x1) = 10−3 i.e., exp(−x)2 = 0.001. (42)

So x = 8.6602.
Now we define the error ξ as follow.

ξ = x1 − x = 8.6602 − 7.4833 = 1.1769. (43)

The new diffusion function converges faster than the Perona and Malik function as making it possible to remove noise
and enhance edges better than the Perona and Malik model.

φ1 (x = k) = g1 (x = k) × x (44)

φ1 (x = k) = exp(−1) × k = 0.3679k (45)

φ3 (x = k) = g3 (x = k) × x (46)
φ3 (x = k) = 1 − tanh(1) × k = 0.2384k. (47)
As can be easily seen in Fig. 10, using the Perona and Malik function g1 , the flow φ takes place and smooth the image pro-
gressively, while using the suggested method, the flow drops quickly and prevents a slow diffusion (φ3 (x = k) < φ1 (x = k)),
 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768 1759

Fig. 10. The flow functions φ1 and φ3 .

Fig. 11. 2D Distribution diffusion coefficients and flux functions relative to {N, W} directions.

making it possible to preserve edges and small details. Supposing, in this instance, that an image is characterized only by
strong edges above a given threshold (the point x = K where g3 reaches zero), the function g3 preserves the edges untouched
while the function g1 will not.
The diffusion and flow function calculated in the north and west directions using ‘‘cameraman’’ image is shown in Fig. 11.

We can notice that in homogeneous areas, the diffusion is high (φ (x) = φmax (x)) while in the transition areas (edge) it
decreases to reach rapidly zero especially face to sudden transition (φ (x) ≈ 0) which eliminates noise and highlights the
edges.

5. Experimental results

In this section, we evaluate the suggested method and compare the obtained results with other approaches. We start by
defining the evaluation criteria. The first criterion we used is the PSNR, given by the equation.
d2
( )
PSNR = 10log10 (48)
MSE
with:
M −1 N −1
1 ∑∑(
[f (x, y) − r(x, y)]2 .
)
MSE = (49)
MN
i=0 j=0

The second criterion is the normalized cross-correlation (NCC) and the third one is the SSIM (Structural Similarity Index).
This index is a method for measuring the similarity between two images. It measures the structural change to quantify the
1760 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768

Fig. 12. (a) Noisy image (σ = 0.025), (b) de-noised image using P&M Model, (c) de-noised image based on the proposed algorithm, (d) fragment of image
(b), (e) fragment of image (c).

Fig. 13. (a) Noisy image (σ = 5), (b) de-noised image using P&M Model, (c) de-noised image based on the proposed algorithm, (d) fragment of image (b),
(e) fragment of image (c).

distortion in the content of an image.

(2µx µy + c1 )(2σxy + c2 )
SSIM(x, y) = (50)
(µ + µ2y + c1 )(σx2 + σy2 + c2 )
2
x

where x and y are two windows size N × N, µx is the x average, µy is the y average, σx2 is the x variance, σy2 is the y variance,
σxy is the x and y covariance, c1 = (k1 L)2 , c2 = (k2 L)2 two variables to stabilize the division with weak denominator and L
the dynamic range of the pixel-values. k1 = 0.01 and k2 = 0.03 by default.
The methods described above were applied to synthetic, texture and real images corrupted by Gaussian noise. The results
obtained are compared to those given by the Perona and Malik method and other techniques in use.

5.1. Synthetic images

To show the important contribution of the new anisotropic diffusion function, we apply the new scheme to synthetic test
images corrupted by Gaussian noise. The obtained results are given below (see Figs. 12 and 13).
The psycho-visual examination of the obtained results shows the efficiency of the proposed method to remove the noise
while keeping the edges and the borders visible, against the traditional Perona and Malik model. This is confirmed by the
metrics found in Table 1.

5.2. Texture and real images

The proposed denoising algorithm is tested on the size 256 × 256 Cameraman and House images corrupted by Gaussian
white noise with zero mean and variance equal to 10. The experimental results are measured objectively by PSNR in decibels
(dB), NCC and SSIM which are defined above (see Figs. 14–16).
 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768 1761

Table 1
Computation of the PSNR, NCC and SSIM.

Table 2
Computation of the PSNR, NCC and SSIM.

Fig. 14. (a) Noisy image (σ = 10), (b) de-noised image using P&M Model, (c) de-noised image based on the proposed algorithm, (d) fragment of image (b),
(e) fragment of image (c).

Table 2 summarizes the de-noising results obtained from PM model and the proposed method. It can be seen that the
proposed method is superior to the PM model in all cases.
In order to show the advantage of the proposed method against Gaussian noise, it was also applied to some natural images
with different iterations number (T = 10, 20, 40, 80). The constant parameter K is selected as Kopt for every image in each
iteration. Image are corrupted by additive zero mean Gaussian noise with σ = 10. The obtained results are compared to
those obtained by the Perona and Malik model and given in Figs. 17–25 (see Table 3).
The psycho-visual examination of these results shows the performance of the suggested method against Gaussian noise.
In fact, our method was able to reduce noise and efficiently preserve the details such as (edges, borders and fine details) as
the suggested method converges faster than the P–M model. For T equal 80, we can see that the detail and border obtained
by the proposed method are kept intact compared to the results obtained by P&M method. To confirm these findings, we
used the different objective and structural metrics. The results are significant in terms of PSNR, NCC and SSIM.
1762 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768

Fig. 15. (a) Original image, (b) noisy image (σ = 10), (c) de-noised image using P&M Model, (d) de-noised image based on the proposed algorithm.

Fig. 16. (a) Original image, (b) noisy image (σ = 30), (c) de-noised image using P&M Model, (d) de-noised image based on the proposed algorithm.

5.3. Comparison with state of the art methods

Many experiments are conducted to illustrate the improved performance of the suggested method compared to recent
work in the literature. In Table 4, we show the amelioration in term of PSNR and MSSIM, defined in Eq. (51), of the proposed
method compared to the Perona & Malik model [21], YK model [22], RPPM model [23], TV model [24], DEPM model [25] and
 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768 1763

Fig. 17. Optimization of the threshold parameter.

Fig. 18. Anisotropic diffusion, (a) original image, (b) distorted image, (c) P&M (T = 10), (d) P&M (T = 20), (e) P&M (T = 40), (f) P&M (T = 80), (g)
suggested method (T = 10), (h) suggested method (T = 20), (i) suggested method (T = 40), (j) suggested method (T = 80).

MPMe model [17]. The proposed method was applied to noisy Lena image
M
1 ∑
MSSIM(x, y) = SSIM(xj , yj ) (51)
M
j=1

where x and y are the original and the distorted images, respectively; xj and yj are the image contents at the jth local window;
and M is the number of local windows in the image and the SSIM is the Structural Similarity Index defined in Eq. (50).
The next experiment is done in order to demonstrate the improved performance of the new algorithm. We compared
its efficiency with other existing algorithms [3] and [24–27]. The images are corrupted by additive Gaussian noise with a
standard deviation σ = 25. The results of the PSNR show the improvement of the proposed model. This improvement is
inspected from the good quality of the smoothed image (see Table 5).
In Table 6, we report the PSNR value for boats image corrupted by Gaussian noise with variance equal to 30. The obtained
results are compared respectively to PM [28] method, Catte method [6], TV method [23], EAD method [29], CTD method [2],
MPM method [25] and SAT method [13].
1764 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768

Fig. 19. Anisotropic diffusion, (a) original image, (b) distorted image, (c) P&M (T = 10), (d) P&M (T = 20), (e) P&M (T = 40), (f) P&M (T = 80), (g)
suggested method (T = 10), (h) suggested method (T = 20), (i) suggested method (T = 40), (j) suggested method (T = 80).

32

30

28
PSNR [dB]

26

24

22

20
10 20 30 40 50 60 70 80
iteration T

Fig. 20. Calculation of PSNR (Lena).

32

30

28
PSNR [dB]

26

24

22

20
10 20 30 40 50 60 70 80
iteration T

Fig. 21. Calculation of PSNR (Boats).

 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768 1765

Fig. 22. (a) Noisy images, results by: (b) the YK model, (c) the PM model, (d) the RPPM model, (e) the TV model, (f) the DEPM model, (g) the MPMe model
with α = 0.5 and (h) the proposed model.

Fig. 23. Denoising results, (a) noisy image (V = 25), restored image by the proposed method.

Fig. 24. Denoising results, (a) original image, (b) noisy image (V = 30), (c) restored image by the proposed method.
1766 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768

Table 3
Comparison between P&M method and the suggested method.

Table 4
Performance comparisons.
Variance noise Measures Models
YK [22] PM [21] RPPM [23] TV [24] DEPM [25] MPMe [17] Proposed model
λ = 0.01 α = 0.5
20 Parameter K =3 K = 15 Tau = 0.09 – – – K = 17
PSNR (dB) 29.98 30.62 30.62 31.28 28.60 31.07 32.10
MSSIM 0.804 0.831 0.832 0.843 0.753 0.84 0.826
30 Parameter K =4 K = 18 Tau = 0.08 – – – K = 20
PSNR (dB) 28.09 28.92 28.91 29.63 29.49 29.43 29.85
MSSIM 0.727 0.791 0.795 0.809 0.697 0.806 0.826
40 Parameter K =6 K = 21 Tau = 0.7 – – – K = 23
PSNR (dB) 26.98 27.50 27.82 28.29 25.0 28.09 28.75
MSSIM 0.677 0.759 0.756 0.779 0.650 0.776 0.795
50 Parameter K =6 K = 22 Tau = 0.7 – – – K = 23
PSNR (dB) 25.95 26.41 26.70 27.16 23.88 26.99 27.43
MSSIM 0.640 0.736 0.732 0.756 0.610 0.752 0.789
60 Parameter K =6 K = 22 Tau = 0.7 – – – K = 24
PSNR (dB) 25.23 25.50 25.81 26.26 23.10 26.12 27.02
MSSIM 0.614 0.720 0.713 0.740 0.595 0.736 0.783
70 Parameter K =6 K = 21 Tau = 0.7 – – – K = 26
PSNR (dB) 23.93 24.54 24.86 25.31 23.34 25.10 26. 29
MSSIM 0.547 0.703 0.692 0.721 0.567 0.717 0.760
80 Parameter K =7 K = Tau = 0.8 – – – K = 26
PSNR (dB) 23.44 23.71 23.75 24.29 21.68 24.14 25. 16
MSSIM 0.545 0.691 0.675 0.705 0.550 0.704 0.722
90 Parameter K =8 K = 17 Tau = 0.8 – – – K = 29
PSNR (dB) 22.62 22.93 22.98 23.50 21.15 23.38 24. 56
MSSIM 0.508 0.681 0.664 0.694 0.525 0.692 0.702
100 Parameter K =8 K = 17 Tau = 0.08 – – – K = 32
PSNR (dB) 22.02 22.22 22.21 22.77 20.61 22.64 23.19
MSSIM 0.496 0.670 0.649 0.678 0.503 0.676 0.689

Table 5
Performance comparisons.
Image P&M MSG ALD EAD WWBF Proposed method
Lena 26.32 26.53 20.70 24.88 27.12 29.23
Baboon 19.45 23.34 20.84 22.47 22.98 24.35

From the Table 6, we see that our method has higher PSNR value than the other two methods. Therefore, we conclude

that the proposed method performs better than the tight frame method and the total variation method.
 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768 1767

Table 6
Experimental results of PSNR ‘‘boats image’’.
Method PM Catte TV EAD CTD MPM SAT Proposed method
PSNR (dB) 21.5002 21.9599 25.0609 25.6308 18.9026 25.2560 26.5559 27.2699

Table 7
Quantitative comparison of the results of the second group of experiments. The experiments presented in this paper are run on 2.30 GHz
CPU with 2G RAM.

Fig. 25. (a) Original cameraman image; (b) Gaussian noisy images-N(0, 20); filtered images: (c) P–M equation; (d) TV model (e) GVF-based P–M equation;
(f) INGVF-based P–M equation; (g) CONVEF-based P–M equation; (h) proposed method.

In other experiment, we compared the proposed method with TV model [7], GVF model [30], INGVF-based P–M
equation [7] and CONVEF-based P–M equation [31]. The experimental results are quantified using PSNR and MSSIM. The
results obtained are shown below.
The results shown in Table 7 exhibit that the suggested method can outperform the P–M, TV, GVF, INGVF-based P–M and
CONVEF-based P–M equation in terms of PSNR and MSSIM and in processing time, in fact the iteration number required by
the suggested method is less than this required by other approaches previously cited.
1768 Z. Mbarki et al. / Computers and Mathematics with Applications 74 (2017) 1751–1768

6. Conclusion

In this work, we have reviewed some representative anisotropic diffusion methods in the literature, giving special
attention to the way in which they manage local information. A new mathematical model of diffusion in image selective
smoothing is presented. For this reason, we developed a new diffusion function based on a mathematical attractor called
β (x) that converges faster than the Perona and Malik model and which requires less computational processing time making
possible the implementation in the real time. We proved through experimental results and comparative study that the new
diffusion function is able to efficiently de-noise images and preserve edge and content.

Acknowledgments

We would like to express our gratitude to the editor and anonymous reviewers for their constructive comments that will
lead to this manuscript improvement in terms of quality and representation.

References

[1] Guodong Wang, Nong Sang, Luxin Yan, et al., X-ray angiogram images enhancement by facet-based adaptive anisotropic diffusion, Comput. Med.
Imaging Graph. 33 (2) (1992) 140–147.
[2] S.M. Chao, D.M. Tsai, An improved anisotropic diffusion model for detail and edge preserving smoothing, Pattern Recognit. Lett. 31 (2010) 2012–2023.
[3] P. Perona, J. Malik, Scale-space and edge detection using anisotropic diffusion, IEEE Trans. Pattern Anal. Mach. Intell. 12 (7) (1990) 629–639.
[4] Q. Chen, P. Montesinos, Q.S. Sun, D.S. Xia, Ramp preserving Perona-Malik model, Signal Process. 90 (2010) 1963–1975.
[5] Jiangtao Xu, Yuanyuan Jia, Zaifeng Shi, Ke Pang, An improved anisotropic diffusion filter with semi-adaptive threshold for edge preservation, Signal
Process. 119 (2016) 80–91.
[6] F. Catté, P.L. Lions, J.M. Morel, et al., Image selective smoothing and edge detection, SIAM J. Numer. Anal. 29 (1) (1992) 182–193.
[7] H. Yu, C. Chua, GVF-based anisotropic diffusion models, IEEE Trans. Image Process. 15 (6) (2006) 1517–1524.
[8] L. Alvarez, P.L. Lions, J.M. Morel, Image selective smoothing and edge detection by non linear diffusion II, SIAM J. Numer. Anal. 29 (3) (1992) 845–866.
[9] S. Osher, L. Rudin, Feature-oriented image enhancement using shock filters, SIAM J. Numer. Anal. 27 (4) (1990) 919–940.
[10] Y.L. You, W. Xu, A. Tannenbaum, M. Kaveh, Behavioral analysis of anisotropic diffusion in image processing, IEEE Trans. Image Process. 5 (11) (1996)
1539–1553.
[11] S. Acton, Multigrid anisotropic diffusion, IEEE Trans. Image Process. 7 (3) (1998) 280–291.
[12] L. Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal algorithms, Physica D 60 (1992) 259–268.
[13] Jiangtao Xu, Yuanyuan Jia, Zaifeng Shi, Ke Pang, An improved anisotropic diffusion filter with semi-adaptive threshold for edge preservation, Signal
Process. 119 (2016) 80–91.
[14] Y. Wang, W. Ren, H. Wang, Anisotropic second and fourth order diffusion models based on convolutional virtual electric fieled for image de-noising,
Comput. Math. Appl. 93 (10) (2013) 1729–1742.
[15] C. Lopez-Molina, M. Galar, H. Bustince, et al., On the impact of anisotropic diffusion on edge detection, Pattern Recognit. 47 (1) (2014) 270–281.
[16] G. Papari, N. Petkov, Edge and line oriented contour detection: state of the art, Image Vis. Comput. 29 (2–3) (2011) 79–103.
[17] Guodong Wang, Nong Sang, Luxin Yan, Xubang Shen, On the choice of the parameters for anisotropic diffusion in image processing, Pattern Recognit.
Lett. 46 (5) (2013) 1369–1381.
[18] M. Black, G. Sapiro, D.H. Marimont, D. Heeger, Robust anisotropic diffusion, IEEE Trans. Image Process. 7 (3) (1998) 421–432.
[19] X. Li, T. Chen, Nonlinear diffusion with multiple edginess thresholds, Pattern Recognit. 27 (8) (1994) 1029–1037.
[20] M. Black, G. Sapiro, D. Marimont, et al., Robust anisotropic diffusion and sharpening of scalar and vector images, in: Proceeding of the International
Conference on Image Processing ICIP’97, Santa_Barbara, California, Vol. 1, 1997, pp. 263–266.
[21] Y. You, M. Kaveh, Fourth order partial differential equations for noise removal, IEEE Trans. Image Process. 9 (10) (2000) 1723–1730.
[22] Q. Chen, P. Montesinos, Q.S. Sun, Ramp preserving Perona–Malik model, Signal Process. 90 (2010) 1963–1975.
[23] H. Tian, H. Cai, J. Lai, Effective image noise removal based on difference eigenvalue, in: ICIP, 2011, pp. 3357–3360.
[24] Y.Q. Wang, Jichang Guo, Wufan Chen, Wenxue Zhang, Image denoising using modified Perona–Malik model based on directional Laplacian, Signal
Process. 93 (2013) 2548–2558.
[25] Y. Shi, Q. Chang, New time dependent model for image restoration, Appl. Math. Comput. 179 (1) (2006) 121–134.
[26] W. Kusnezow, W. Horn, R.P. Wurtz, Fast image processing with constraints by solving linear PDEs, in: Partial differential equations methods in graphics
and vision, Electron. Lett. Comput. Vis. Image Anal. 6 (2) (2007) 22–35.
[27] V.B.S. Prasath, A. Singh, Well-posed inhomogeneous nonlinear diffusion scheme for digital image denoising, J. Appl. Math. 2010 (2010) (article ID
763847).
[28] M. Lindenbaum, M. Fischer, A. Bruckstein, On Gabor contribution to image enhancement, Pattern Recognit. 27 (1994) 1–8.
[29] V.B. Surya Prasath, D. Vorotnikov, Weighted and well-balanced anisotropic diffusion scheme for image denoising and restoration, Nonlinear Anal.
RWA 17 (2014) 33–46.
[30] O. Ghita, P.F. Whelan, A new GVF-based image enhancement formulation for use in the presence of mixed noise, Pattern Recognit. 43 (8) (2010)
2646–2658.
[31] Y. Wang, W. Ren, H. Wang, Anisotropic second and fourth order diffusion models based on convolutional virtual electric fieled for image de-noising,
Comput. Math. Appl. 66 (10) (2013) 1729–1742.