See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/282590230

Rapid and efficient image restoration technique based on new adaptive

anisotropic diffusion function

Article  in  Digital Signal Processing · October 2015

DOI: 10.1016/j.dsp.2015.09.013

CITATIONS READS

19 1,760

4 authors, including:

Zouhair Mbarki Hassen Seddik

ENSIT Université de Tunis
18 PUBLICATIONS   58 CITATIONS    83 PUBLICATIONS   304 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

image denoising View project

Deep learning framework for parotid and liver lesions segmentation on CT and MRI data View project

All content following this page was uploaded by Zouhair Mbarki on 02 January 2018.

The user has requested enhancement of the downloaded file.

 Digital Signal Processing 48 (2016) 201–215

Contents lists available at ScienceDirect

Digital Signal Processing

www.elsevier.com/locate/dsp

Rapid and efficient image restoration technique based on new adaptive

anisotropic diffusion function
S. Tebini ∗ , Z. Mbarki, H. Seddik, E. Ben Braiek
University of Tunis, Department of Electrical Engineering, CEREP, ENSIT, 5, Av. Taha Hussein, 1008, Tunis, Tunisia

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

Article history: The anisotropic diffusion is an efficient smoothing process. It is widely used in noise removing and edges
Available online 1 October 2015 preserving via different schemes. In this paper based on a mathematical background and the existing
efficient anisotropic function in the literature we developed a new mathematical anisotropic diffusion
Keywords:
function which is able to overcome the drawbacks of the traditional process such as the details loss
Noise reduction
Image de-noising
and the image blur. The simulations results and the comparative study with other recent techniques
Anisotropic diffusion are conducted and showed that the proposed schema generates a wide improvement in the quality
Conductance function of the restored images. This improvement has been shown subjectively in terms of visual quality, and
Edge preserving objectively with reference to the computation of some criteria. The simulated images are well de-noised
but the most important is that details and structural information are kept intact. In addition to that, the
proposed new function was found very interesting and presents numerous advantages like its similarity
to the conventional model and the importance of the speed hence it converges faster which allows an
opportunity to be well implemented in our de-noising process.
© 2015 Elsevier Inc. All rights reserved.

1. Introduction
Improving the image quality is the most important goal in im-
age processing. So, we are looking for an efficient method of noise
removal, which is able to de-noise efficiently the image. Linear
smoothing technique allows good results but it causes the loss of
the details and the image becomes inexploitable for further pro-
cessing information extraction. They are potentially more powerful
than linear filters because they are able to reduce noise levels
without simultaneously blurring edges however they can be com-
putationally expensive to use as well as they can generate spurious
features and distort existing features in images. Scale space repre-
sentation is based on the convolution with a chosen kernel in the
smoothing operation. The most important parameter supervising
an anisotropic filtering function is its convergence speed. This pa-
rameter is able alone to characterize the filtering process to be real
time running and it allows it to be electronically implemented. The
higher the convergence speed is, the more interesting the filtering
process is especially when the image is composed by many details
and edges or textures. For this reason we are highly interested in
the development of a new anisotropic function which is faster than
* Corresponding author.
E-mail addresses: tebinisondes@yahoo.fr (S. Tebini),
mbarki.zouhair84@gmail.com (Z. Mbarki), Seddik_hassene@yahoo.fr (H. Seddik),
Ezzedine.benbraiek@esstt.rnu.tn (E.B. Braiek).
the existing recent models in the literature. Then, the new model
is chosen such as it behaves in the range of [0 1]. In addition the
proposed model satisfies the convergences requirements and re-
versibility.
Witkin [1] proposed to generate coarser resolution images by
convolving the original image with a Gaussian kernel. This descrip-
tion was developed by Koenderink [2] who motivates the diffusion
equation formulation by announcing two criteria, the first is that
of causality and the second is about the homogeneity and isotropy.
These criteria lead to the diffusion equation formulation. In an
analogous way, anisotropic diffusion is considered as an important
solution for better image de-noising and edges preserving. There-
fore, variational methods for image smoothing have been proposed
to reduce noise and restore the quality of an image for better ob-
servation. The anisotropic diffusion is a nonlinear iterative process
which aims to process separately noise and details by calculating
the gradient of each pixel. The first model of anisotropic diffu-
sion has been proposed by Perona and Malik [3]. This model is
based on a new definition of the scale space representation via
anisotropic diffusion, in which P and M proposed to replace the
heat equation by a nonlinear equation. Many studies based on a
modified anisotropic diffusion have been proposed for an efficient
de-noising technique and edge preserving [4–8]. Anisotropic dif-
fusion algorithms are a powerful tool in image processing. These
processes are based on two constraints such as good edge en-
hancement and fast convergence speed. By this way, this function

http://dx.doi.org/10.1016/j.dsp.2015.09.013
1051-2004/© 2015 Elsevier Inc. All rights reserved.
202 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215

Fig. 1. Spatial neighbors of a selected pixel: (a) input image, (b) zoomed section, (c) first derivative.


2 
will be easier and faster for real time implementation. In this pa-   ∇ I 
g 2 ∇ I  = exp − (4)
per, a new anisotropic adaptive schema which smooths the noisy k
zones at the same time, enhances the edges and preserves the con-
tent is presented. k is the threshold parameter used to adjust the diffusion process.
The remainder of this paper is organized as follows: the basic The gradient ∇ I  serves as the edge detector. Consequently, the
information about Perona & Malik anisotropic diffusion (AD) and diffusion is stopped across edges when the gradient is significant
its regularized models will be given in Section 2. Section 3 de- than k ( g = 0) on the other hand it is encouraged the diffusion is
scribes our contribution to denoised images without hitting edges maximal in the uniform regions (g = 1) when the gradient magni-
and contents. The implementation technique of the new function tude is smaller than k. Perona and Malik sampled the anisotropic
in the denoising process will be held in Section 4. Section 5 ex- diffusion equation to:
poses different evaluation metrics and the comparative study of
the new algorithm with diverse recent works [3,15–17,23–30].
λ   
Conclusion appears in the last section. I t +1 ( s ) = I t ( s ) + gk |∇ I s, p | ∇ I s, p (5)
|ηs | p ∈η
s
2. State of the art and techniques evolution
where I is a discretely sampled image, s denotes the pixel po-
2.1. Perona and Malik model sition in the discrete 2-D grid, t denotes the iteration step,
g is the conductance function and k is the gradient thresh-
The first model of the nonlinear PDE approaches has been pro- old parameter. Constant λ ∈ [0, 1] determines the rate of dif-
posed in [3] by Perona and Malik, as illustrated by the following fusion and ηs represents the 8 neighbors of the pixel s: ηs =
equation: { N , S , E , W , NE, SE, SW , NW } where N , S , E and W are the North,
⎧ South, East and West neighbors of pixel s, respectively. ∇ repre-
⎨ ∂ I (x, y , t )     sents the difference between two adjacent pixels in the different
= div g ∇ I (x, y , t ) .∇ I (x, y , t )
∂ t (1) directions.
⎩ I (x, y , 0) = I_0

where t is the time parameter, I (x, y , t ) is the image at time t ; div ∇ I s, p = I t ( p ) − I t (s), p ∈ ηs = { N , S , E , W , NE, SE, SW , NW }
represents the divergence operator; ∇ I (x, y , t ) is the gradient of
the image, g (.) is the diffusion function and I_0 is the original (6)
image (at t = 0).
The choice of the diffusion coefficient g is very important for Edges are significant local changes of intensity in an image. Of-
an efficient filtering. In fact, the diffusion is higher when a small ten, points that lie on an edge are identified by detecting the
gradient is detected, consecutively, conversely for higher gradi- local maxima or minima of the first derivative or detecting the
ent magnitude. The diffusivity function is chosen to be a non- zero-crossings of the second derivative. The first derivative has a
negative decreasing conductance function which satisfies these peak at the edge while the second derivative has a zero crossing.
conditions: The first derivative is computed based on a finite difference such
⎧ as:
⎪
⎪ lim g (x) = 1
⎨ x→0
and (2) f (x + h, y ) − f (x, y )
⎪
⎪
⎩ lim g (x) = 0 f  (x, y ) = lim (7)
h→0 h
x→∞
f  (x + h, y ) − f  (x, y )
Two different diffusion coefficients are suggested by P&M, in the f  (x, y ) = lim (8)
literature described by Eqs. (3) and (4): h→0 h

  1 Fig. 1 illustrates the 8 neighborhood of a selected pixel in an area

g 1 ∇ I  = ∇ I  2
(3)
1+( ) of the original image, and the first derivative.
k
The differences with the closest neighbors and the conduction
and coefficients are computed as follows:
 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215 203

to avoid the shortcoming of this model developed in [4], Weick-

ert, Ter Harr Romeny, and Viergever [9] proposed an ‘AOS’ scheme
in which they introduced a diffusion stopping time parameter that
controls the restored image quality. In [10] a new well-posed reg-
ularizing method based stable algorithm for image restoration is
proposed.

∂I   ∇I
= g |∇ I σ | |∇ I |div (11)
∂t |∇ I |
where I σ is the smoothed image with a Gaussian kernel of stan-
dard deviation σ and the function g (.) controls the anisotropic
diffusion. The function g has the same role as in the model of
Perona and Malik, namely to limit the spread for strong gradients.
One side effects of this equation is ill-suited to noisy uniform ar-
eas: indeed, the proposed method tries to follow the edges, even
in its absence.
In [11], Whitaker and Pizer mentioned some other illness of the
traditional diffusion model. If the diffusion threshold is high, some-
times, sharp transitions tend to dissociate into shorter and clearer
segments due to the enhancement effect where some transitions
are fractured in half. This phenomenon called staircasing effect.
Another drawback associated with the choice of the parameter k
is the pinhole effect. It occurs in a transition area interrupted by
Fig. 2. Perona & Malik anisotropic diffusion results: (a) noisy image, (b) denoised
image after 10 iterations, (c) denoised image after 30 iterations, (d) denoised image an intermediate zone where the gradient is less than the thresh-
after 60 iterations. old k. In the process, the intermediate zone widens and becomes
⎧   blurred giving this aspect “hole”. This limitation has been intro-
⎪
⎪ ∇ N I i , j = I (i − 1, j ) − I (i , j ); c Ni , j = g |∇ N I i , j | duced and analyzed in [12] and [13]. In [12] a new model of the
⎪
⎪  
⎪
⎪ conduction function is studied in order to reduce these undesirable
⎪
⎪ ∇ S I i , j = I (i + 1, j ) − I (i , j ); c Si , j = g |∇ S I i , j |
⎪
⎪ effects that appear in the model proposed in [3], as illustrated by
⎪
⎪  
⎪
⎪ ∇ E I i , j = I (i , j + 1) − I (i , j ); c Ei , j = g |∇ E I i , j | the following equation:
⎪
⎪
⎪
⎪        
⎨ ∇W I i , j = I (i , j − 1) − I (i , j ); c W i , j = g |∇ W I i , j | g 3 ∇ I  = 0.5 ∗ tanh λ k − ∇ I  +1 (12)
  (9)
⎪
⎪ ∇NE I i , j = I (i − 1, j + 1) − I (i , j ); c NEi , j = g |∇NE I i , j |
⎪
⎪ Many other solutions and models have been mentioned in the lit-
⎪
⎪  
⎪
⎪ ∇SE I i , j = I (i + 1, j + 1) − I (i , j ); c SEi , j = g |∇SE I i , j |
erature which aims to avoid the ill-posedness of the Perona and
⎪
⎪ Malik model such as in [14].
⎪
⎪  
⎪
⎪ ∇SW I i , j = I (i + 1, j − 1) − I (i , j ); c SWi , j = g |∇SW I i , j | In this paper, a new approach based new anisotropic diffusion
⎪
⎪
⎪
⎪   function is developed for better image de-noising and edges and
⎩
∇NW I i , j = I (i − 1, j − 1) − I (i , j ); c NWi , j = g |∇NW I i , j | contents preserving. The novel technique has been compared with
the original anisotropic diffusion filter of Perona and Malik and
The Perona and Malik anisotropic diffusion model presented in
other modified anisotropic diffusion schemes, in terms of the im-
Eq. (1) will be reduced to an isotropic diffusion equation if the
provement of the image quality. Experimental results confirm that
diffusion coefficient g (x, y , t ) is a constant. One of the drawbacks
the proposed approach removes noise and preserves edges and
of this model is to be an inverse heat equation for some gradient
contents more efficiently than other recent techniques proposed in
values.
the literature [15–17,23–30] based on a modified model of Perona
For images corrupted by high densities of noise, the Perona and
and Malik.
Malik model cannot differentiate between edges and noise (Fig. 2).
The noisy image is obtained by adding Gaussian noise of strength
σ = 35. 3. Proposed model for efficient smoothing and edge preserving
Since this first model was proposed by Perona & Malik, many
studies have been developed to prove that the performance of the Anisotropic diffusion is a smoothing process which produces
anisotropic diffusion could be improved to overcome the draw- a family of parameterized images, which aims to remove noise
backs and limitations. while preserving fine details of the studied image, such as edges
and other structures that should be preserved in the image. We
2.2. Improved models propose a new diffusivity function for an adaptive smoothing. Ac-
cording to Perona and Malik [3], the conduction function is chosen
Various modifications of the original scheme of P&M were pre- to be a positive decreasing function that is based on the gradient
sented attempting to overcome its drawbacks. In [4] a new version magnitude. Thus, we tried to develop a new conduction function
of Perona and Malik was proposed by convolving the image with a which favors image smoothing in the homogeneous regions with
Gaussian kernel: small gradient and stop the diffusion across edges at or around
which the image values undergo a sharp variation.
∂I    
= div g |G σ ∗ I | ∇ I (10)
∂t 3.1. Mathematical background development of the new function
where G is a Gaussian of standard deviation σ and ∗ denotes the
convolution operator. If σ tends to 0, the model becomes unsta- The simplest version of anisotropic diffusion proposed in [3]
ble since it is reduced to the model of Perona & Malik. In order aims to remove noise with a constant gradient threshold level and
204 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215

 First of all, it is similar to the Perona and Malik model.

 Second, its speed is very important, hence it converges faster
which allows an opportunity to be well implemented in our
de-noising process aiming at reducing image noise without
hitting significant parts of the image content such as edges,
lines or other details that are important for the image’s inter-
pretation.

3.2. Characteristics of the proposed model

In this section we present a mathematical study of this diffu-

sion function. This study aims to demonstrate the efficiency and
Fig. 3. Curve shape of the tansig function. the highlights and improvement of the new model compared to
the traditional one. It is based on the representation of the den-
a slow speed of convergence. So, we are looking for a new func- sity function, the speed of convergence, starting from the speed in
tion with fast convergence and better performance. Therefore, the each point, as well as from the slope to the tangent. Finally, we
proposed function has to be faster in the convergence speed and will show the necessity to a finite truncation of the function, from
more efficient in noise removal. In addition we look to get the ad- an infinite and continuous representation.
vantage of content and detail preserving. Consequently, the new
model must be a circular model which converges faster than the 3.2.1. Flow function
Perona and Malik model. If we consider the flow function φ modeled by:
Consequently, the new model is based on the tangent sigmoid
(tansig) function. Mathematically, this function is equivalent to the φ(x) = xg (x) (20)
hyperbolic tangent. It differs in that it is faster. Such function is Based on the new diffusion function described in Section 3.1, the
presented by the graph (Fig. 3) and the equation as follows: related flow function is defined by:
2
y = tansig(x) = −1 (13) φ4 (x) = xg 4 (x) (21)
(1 + e (−2x) )
Hence
In our application and according to the mentioned characteristics
by P&M [3], this function must be decreasing positive [0, 1] func-
φ4 (x) = xg 4 (x)
tion. For that, the construction steps of the desired function are
described as follows: 2
=x (22)
1. Develop a function based on the tansig function and depends 1 + exp(2abs(x/k))
on the gradient and the thresholding parameter k (Fig. 4a):
It represents the sum of the brightness flow that is generated. The
  maximum flow is generated at location where ∇ I  = k.
g 41 (x) = tansig abs(x/k) (14)
The behavior of the flux function is illustrated in Fig. 5.
Since As it can be noticed from Fig. 5, the flux function of the pro-
posed model decreases faster. This proves that the processing time
2
y = tansig(n) = −1 (15) of the de-noising process is very rapid and sharp especially when
(1 + e (−2n) ) the function reaches the value of zero which means that we are
Then processing near details and edges. Consequently, the protection of
the edges and the contents of the image during the diffusion pro-
2
g 41 (x) = g ( y ) = −1 (16) cess are assured. Nevertheless, using the Perona and Malik model
(1 + exp(−2abs(x/k)) as well as the model g 3 , the image smoothing still available which
2. The proposed function must be a decreasing function. That’s can blur the edges. In Fig. 6, we show the diffusion coefficients and
why a simple modification has been done. So, we try to inverse the flux function calculated based on the gradient of ‘Cameraman’
the tansig shape as follows: image in 2 directions [N , E]. It’s clear that the coefficients near the
edges are lower than the other which refer to homogeneous areas.
  2 A zoom section is presented to show how vary these coefficients.
g 42 (x) = tansig −abs(x/k) = −1 (17)
1 + exp(2abs(x/k))
The shape of this function is shown in Fig. 4b. 3.2.2. Convergence speed of the proposed function
3. The diffusion function g 4 must tend to [1 0] respectively in As we mentioned previously the speed of convergence is higher
0 and at infinity, with g 4 (0) = 1 and g 4 (x) ≥ 0. Now, let’s make a than the other cited models. Consequently, the processed f can be
very slight change to this formula used for a real time implementation and a rabbit filtering process.
This rapidity is examined in Fig. 7. Fig. 7a illustrates the behavior
  
g 43 (x) = 1 + tanisg − abs(x/k) (18) illustrating the decreasing speed profile of the proposed function
compared with other recent functions [3] and [12]. The new func-
Finally we obtain a model which has the same characteristics of tion converges faster than the traditional models of Perona and
the first model proposed by Perona and Malik. Malik as well as the cited model g 3 .
2 Fig. 7b shows that for the same value of the index in the
g 43 (x) = g 4 (x) = (19) abscissa axis x = 75, the⎧ value reached by the three compared
1 + exp(2abs(x/k))
⎨ g 4 (x) = 0.046
The shape is shown in Fig. 4c. functions are as follows g 1 (x) = 0.1978 which proves that the
⎩
The proposed new function was found very interesting and g 3 (x) = 0.2215
presents several advantages detailed as follows: magnitude of the proposed function is the nearest to zero and then
 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215 205

Fig. 4. Different steps to construct the new model, (a), (b) and (c) are the three steps.

Fig. 5. (a) Behavior of different functions, (b) flow functions variation.

Fig. 6. Diffusion coefficients and flux functions behavior of 2 directions { N , E } and zoomed section.
206 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215

Fig. 7. Waveforms and comparison of various models.

Fig. 8. Speed convergence of different models. Fig. 9. Slope of the model’s tangent.

faster than [3] and [12]. In Fig. 7c, we picked up the same magni- and
tude of the three compared function and we checked the position
−4 ∗ exp((2 ∗ z)/k)
of the corresponding index in the abscissa axis. The value of x de- g 4 ( z) = (26)
termines the graphical position of the function that tells us how k ∗ (exp((2 ∗ z)/k) + 1)2
quickly the new function is ⎧ changing. It is clear from the figure Thus
⎨ x4 = 30 2 4 ∗ exp(2 ∗ ( z/k))
and the following values x = x1 = 54 that the proposed function T1 = − ∗ (x − z )
⎩ exp(2 ∗ ( z/k)) + 1 k ∗ (exp(2 ∗ ( z/k)) + 1)2
x3 = 55
converges faster than the speed of convergence, for one iteration, (27)
is defined as follows:
and the tangent of the Perona & Malik is defined by the following
g ( i ) − g ( i − 1) equation:
v= (23)
i − ( i − 1)
T 2 = g 2 ( z) + g 2 ( z)(x − z)
Fig. 8 illustrates graphically the speed of the proposed function
2

2

compared with those proposed in [3] and [12]. z 2∗z z

This fact is proved mathematically as illustrated by Fig. 9 and
= exp − − 2 ∗ exp − ∗ (x − z ) (28)
k k k
Eqs. (27), (28).
The segments T 1 and T 2 illustrated in Fig. 9 are the tangent g 4 ( z) and g 2 ( z) represent the slopes of the secant of the studied
on the representative curve of the proposed model and Perona & models; the slopes are equal to the difference quotient:
Malik model. Its equations are as follows: ⎧
⎪ dg 4 g 4 (70) − g 4 (50)
⎨ g 4 = = = −0.0047
T 1 = g 4 ( z) + g 4 ( z)(x − z) (24) dx 20
⎩ g  = dg 2 = g 2 (70) − g 2 (50) = −0.0072
⎪
where 2
dx 20
−4 ∗ exp((2 ∗ x)/k) dg 4 (x) dg 2 (x)
g 4 (x) = (25) ⇒ > ⇒ g 4 > g 2 (29)
k ∗ (exp((2 ∗ x)/k) + 1)2 dx dx
 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215 207

Fig. 10. Convergence based Taylor series approximations: (a) Perona & Malik model, (b) proposed model.

Therefore the proposed model’s slope is higher than Perona & Ma- Assume that f has derivatives of all orders on some neighborhood
lik. of a, we have the error goes to zero:

3.2.3. Taylor series approximation εn (x)

lim =0 (35)
A Taylor series is a representation of a function as an infinite x→a (x − a)n
sum of terms that are calculated from the values of the function’s
It is interesting to note that a function may not be equal to its
derivatives at a single point. The Taylor series of a function f (x)
Taylor series, even if its Taylor series converges at every point. And
that is infinitely differentiable at a real or complex number a is
the raising of the Taylor series’s order can let it to diverge, that’s
the power series:
why we need to truncate the function in order to be representable
∞
 through a matrix.
(x − a)n
f (x) = f (n) (a) Thus, in our case the filter has infinite support and it must be
n! truncated to a finite window (3 × 3). That’s why we should approx-
n =0
imate the function by using a finite number of terms of its Taylor
f  (a)(x − a)2 f n (a)(x − a)n
= f (a) + f  (a)(x − a) + + ··· + series in order to better illustrate its convergence. Fig. 10 shows
2! n! the behavior of the Taylor series of the proposed model compared
  
approximation to P&M. It is well observed that the agreement between the Taylor
series and the new function is meaningful (Fig. 10b) in which the
+ εn (x) (30)
   Taylor series follows the innovative model and it diverges after the
remainder point (x = 40.06) but in the case the Perona & Malik function, the
which is equivalent to: Taylor series diverge faster around x = 23.1 (Fig. 10a).

∞
 f (n) (a) 4. Implementation of the proposed diffusion function in image
f (x) = (x − a)n + εn (x) smoothing
n!
n =0
= P n (x) + εn (x) (31) 4.1. Diffusion scheme

where n! denotes the factorial of n and f n (a) denotes the nth

The diffusion process is authorized for low gradient’s values
derivative of f evaluated at the point a. The derivative of order
with a high diffusion coefficient and it is limited for high values by
zero f is defined to be f itself, (x − a)0 and 0! are both defined
a weak coefficient in order to protect the image content. We will
to be 1. When a = 0, the series is also called a Maclaurin series.
precede a new scheme, which we called ‘Successive Anisotropic
P n (x) is an Nth-order Taylor polynomial for f (x) centered at some
Filtering SAF’, for image de-noising using the proposed function.
number a ∈ . The remainder ε after n terms is given by:
In the first step, we applied a diffusion process in which the co-
x efficient diffusion’s computation is based on the image gradient
(x − t )n
εn = f (n+1) (t ) dt (32) magnitude. In the second step, the diffusion coefficients are calcu-
n! lated using the Laplace operator of the image obtained in the first
a
step.
For some ξ ∈ [a, x], therefore, integrating n + 1 times gives the re- Finally, we evaluate the efficiency of the proposed model by
sult: computing some criteria such as the MSE, PSNR and the MSSIM
criteria of the restored images. In addition to these quantitative
(x − a)n+1 (n+1) measures, we illustrate the good performance of the new tech-
εn+1 = f (ξ ) (33)
(n + 1)! nique by displaying many filtered images in order to evaluate the
new method psycho-visually.
For any fixed x ∈ , we have:
The new image filtering process is described in Fig. 11, where
 (n)  n +1 the diffusion coefficients are computed based on these functions
 g (t ) ≤ M ⇒ (x − a) ∗ M ≤ εn (x) g 41 and g 42 which are defined as:
(n + 1)!

(x − a)n+1 εn (x) ∇I
⇒ ≤ (34) g 41 (∇ I ) = 1 + tanisg − abs (36)
(n + 1)! M k
208 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215

Fig. 11. Proposed image smoothing scheme.

I 11
g 42 ( I 11 ) = 1 + tansig − abs (37) Table 1
k Optimization of the threshold parameter.

Iterations kappa_opt
where ∇ and are the gradient and the Laplacian operator, k is
CM Lena House Synthetic image
the threshold parameter for the diffusion process. The Successive
Anisotropic Filtering process ‘SAF’ is based on a double convolu- 10 53 59 53 46
20 52 60 53 37
tion. The first convolution is done in order to compute the gradient 30 52 60 53 33
of the image following the equation: 40 52 60 53 31
 50 52 60 53 29
I 1 ( i , j ) = I ( i , j ) ∗ f 1 (u , v ) = I (i − u , j − v ). f 1 (u , v ) (38) 60 52 60 52 25
70 52 60 53 23
u v 80 53 59 53 14
where I is the image and f 1 is the gradient kernel of size (3 × 3)
because we compute the 8 neighbors of the selected pixel. The sec- 4.2. Threshold parameter optimization
ond convolution is done to determine the Laplacian of the filtered
image based on the gradient coefficients.
The gradient threshold parameter k is an important parame-
ter in which depends the anisotropic diffusion that why it must
I 2 (i , j ) = I 11 (x, y ) ∗ f 2 (α , β) be well selected taking in consideration the best quality of the

= I 11 (x − α , y − β). f 2 (α , β) (39) smoothed image. It is optimized for each studied image. We vary
u v
the value of k and the value of the thresholding parameter leads
to the higher PSNR value.
The Laplacian kernel f 2 is of size (3 × 3). The optimization of k is performed for each iteration number.
The smoothed image is presented as the result of the convolu- After selecting the maximal PSNR, we compare all the PSNR results
tion by an adaptive filter and it is given by: then we select the higher PSNR. Finally we attribute kappa_opt.
 The example in Fig. 12a, illustrates that for one iteration, we have
I 2 (i , j ) = I (i , j ) ∗ f 1 (u , v ) ∗ f 2 (α , β) kappa_opt = 12. The variation of the threshold parameter with the
   number of iterations of a synthetic image is figured in Fig. 12b.
= I (i − u , j − v ). f 1 (u , v ) ∗ f 2 (α , β) For many images, Table 1 shows the optimum value of the
u v threshold parameter related to the iterations.
       The next section is devoted to presenting the results obtained
= I ( i − u ) − α , ( j − v ) − β . f 1 (u , v ) through the proposed algorithm. Comparisons with recent works
α β u v based on the anisotropic diffusion have proved the efficiency of

the new scheme.
. f 2 (α , β)
 5. Experimental results and discussion
= I ( X − α , Y − β). f 2 (α , β) (40)
α β u v In this section, we present and discuss the various results given
by the new anisotropic denoising algorithm. To evaluate its per-
where formance, we tested it using synthetic and natural images and
compared the results with other existing algorithms. This com-
X = (i − u ) and Y = ( j − v) (41) parison is based on the computation of the peak signal to noise
 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215 209

Fig. 12. Non-linear evolution of the threshold parameter.

ratio PSNR, and the structural similarity MSSIM and the correla- Table 2
tion measure. All these used measures are in favor of the proposed Computation of the PSNR, correlation and MSSIM.
anisotropic model in image de-noising and edge preserving. Measures Perona & Malik Proposed technique Error

PSNR (dB) 20.04 29.9 9.86
d2 MSSIM 0.4352 0.56 0.127
PSNR = 10 log10 (42) CORR 0.98 0.99 0.022
MSE
where
Table 3

M −1 N
 −1
1  2  Measure of PSNR and MSSIM.
MSE = I (x, y ) − I d (x, y ) (43) Methods Criteria Gaussian noise variance
M×N
i =0 j =0 V = 10 V = 20 V = 30

I and I d are the original and the recovered images, M × N indicates P&M (1990) [3] PSNR (dB) 27.65 21.07 19.94
MSSIM 0.834 0.78 0.61
the image size and d denotes the maximum value.
RPPM (2010) [15] PSNR (dB) 28.47 26.31
MSSIM 0.815 0.75
1 
MN
DEPM (2011) [16] PSNR (dB) 28.10 25.74
MSSIM( f , r ) = SSIM(xi , y j ) (44) MSSIM 0.79 0.73
MN
i =1 MPME (2013) [17] PSNR (dB) 28.34 26.19
MSSIM 0.82 0.762
where Proposed model PSNR (dB) 29.05 28.69 26.51
MSSIM 0.87 0.795 0.743
(2μx μ y + c 1 )((2cov(x, y ) + c 2 )
SSIM(x, y ) = (45)
(μ2x + μ2y + c 1 )(σx2 + σ y2 + c 2 )
The efficiency of the new proposed technique is shown by the
The MSSIM is computed by choosing the local window pixel by visual observation in Fig. 14, which proves better noise reduction
pixel, in order to get a single overall similarity measure of the two and edges preservation.
images. MSSIM index is in the range [0, 1], with unity for the ideal A very good performance of the proposed algorithm for denois-
structure preservation. μx and μ y are the means of images x and ing synthetic images is obtained compared to the original model
y, respectively, σx and σ y are the variances of images x and y, of Perona and Malik.
cov(x, y ) is the covariance of the two images, and c 1 and c 2 are
two parameters to stabilize the division with small denominators; 5.2. Natural images
the defaults are c 1 = 0.0001, c 2 = 0.0009, xi , y i are the ith local
windows. 5.2.1. Grayscale image denoising
The proposed smoothing technique was also applied to some
5.1. Synthetic images natural images. Fig. 15 demonstrates the results of the de-noising
technique of Perona & Malik and the proposed technique. The
Many experiments are recorded in order to show the important image is corrupted by a Gaussian noise (variance: σ 2 = 10,
contribution of the new anisotropic diffusion function. To evaluate iterations = 19).
the performance of the proposed method, we first experiment with The proposed ‘SAF’ method exhibits greater performance in
synthetic images. We apply the new scheme to a synthetic test both noise removing and contents conservation. The proposed
image illustrated in Fig. 13. technique outperforms the original Perona & Malik anisotropic dif-
The psycho-visual comparison of the restored image shows the fusion. PM suppresses noise but it destroys the most important
efficiency of the proposed anisotropic technique to remove the details in the image (Fig. 15). The high quality of the filtered image
noise while keeping the edges and the borders visible, against the is shown by the measures of the PSNR, correlation and the MSSIM
traditional Perona and Malik model. Table 2 shows clearly the per- (Table 3). These measures are compared with different methods in
formances of the developed algorithm. the literature [3] and [15–17].
Fig. 14 shows the de-noising power of different nonlinear Many other experiments are recorded and show the superiority
anisotropic filters on a synthetic image to which Gaussian noise of the proposed model. In the next figure (Fig. 16), we present the
is added with a standard deviations σ = 0.025. By comparing the filtered images, it can be seen that the developed model is very
contents preservation, it can be seen that the proposed approach efficient than the other, so that intra-regions in the image become
gives better result than P&M which blurs the edges. smoothed while edges of inter-regions are preserved.
210 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215

Fig. 13. (a) Noisy image, (b) restored image based P&M model, (c) restored image based the proposed algorithm, (d) and (e) zoomed section using P&M and the new algorithm
respectively.

Fig. 14. (a) Noisy image, (b) and (c) are the denoised image using P&M Model and a fragment of it, (d) and (e) are the denoised image based on the proposed algorithm and
a fragment of it.

Table 4 Similarly, in Fig. 17, where Gaussian noise was added to the
Results of PSNR and MSSIM of image in Fig. 16. house image, the proposed smoothing technique using the new
Measures Perona & Malik [3] Proposed technique Error diffusion function performs better than the other technique.
PSNR (dB) iter = 10 23.99 25.56 1.57 The results shown in Fig. 17 reveal that the proposed method
iter = 20 21.69 23.21 1.52 can perform the other techniques in terms of the capacity to re-
MSSIM iter = 10 0.71 0.74 0.03 duce noise while keeping the high structures in the image. In
iter = 20 0.63 0.69 0.06
CORR iter = 10 0.95 0.97 0.02
addition to that, the numerical results based on the PSNR and
iter = 20 0.92 0.94 0.02 MSSIM (Table 5), on various images, shows that the new technique
performs well than the Perona & Malik model and the technique
proposed in [18]. The images are degraded by Gaussian noise with
This simulation results indicate that the proposed algorithm zero mean and standard deviation σ2 = 0.025. The quantitative
provides a stable and efficient image processing tool, as it is il- difference between the proposed model and the other models is
lustrated by the evaluation criteria measures in Table 4. represented by the Error1 and Error2 .
 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215 211

Fig. 15. Comparison of the denoising performance: (a) noisy image, (b) Perona and Malik filtering result, (c) the proposed filtering result, (d) and (e) are the zoomed part
successfully.

Fig. 16. Visual comparison of the denoising results: (a) noisy images, (b) P&M results, (c) new algorithm results.

Table 5
Performance comparisons.

Image Measures Perona & Malik [3] Model [18] (2013) Proposed model Error1 Error2
Cameraman PSNR (dB) 26.08 33.68 36.11 10.03 2.43
MSSIM 0.81 0.84 0.96 0.15 0.12
Lena PSNR (dB) 27.05 33.76 36.48 9.43 2.72
MSSIM 0.82 0.82 0.96 0.14 0.14
House PSNR (dB) 29.4 31.72 38.28 8.88 6.56
MSSIM 0.85 0.87 0.94 0.09 0.07
Peppers PSNR (dB) 28.82 35.27 36.32 7.5 1.05
MSSIM 0.66 0.817 0.91 0.25 0.096
212 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215

Fig. 17. Denoising results comparison: (a) noisy image, (b) P&M denoising result, (c) proposed algorithm denoising result.

Fig. 19. Grayscale-image denoising: comparison of the new algorithm with BM3D
algorithm based on the PSNR output for House (256 × 256), C.man (256 × 256), Lena
(512 × 512), Boats (512 × 512), Barbara and Peppers (256 × 256) (images: 1–6).

image denoising; the good performance of the new technique is

well achieved.
In the case of Table 7, where Gaussian noise with standard de-
viation = 30 was added to the images, a comparative study with
recent works TV [24], GVF [25], INGVF [26] and CONVEF [27] mod-
els, was accomplished. This comparison is based on the values of
Fig. 18. Denoising experience: (a) noisy images, (b) restored images. the PSNR, MSSIM, the iteration number as well as the time of the
filtering process. The out performance of the proposed anisotropic
Table 6
model is revealed.
PSNR measures. Table 8 shows the values of PSNR, MSSIM and the iterations
Methods Images (256 × 256)
number of the filtering process. It can be seen that the investi-
gated model has the largest value of PSNR and MSSIM, as well as
Lena Baboon
the least iteration number of the other three models. The noise is
P&M [3] (1990) 26.32 19.45
MSG [19] (2006) 26.53 23.34
removed more rapidly and the details and the content of the image
ALD [20] (2007) 20.7 20.84 are preserved more effectively.
EAD [21] (2010) 24.88 22.47 Fig. 20 and Fig. 21 show the efficiency of the proposed al-
WWBF [22] (2014) 27.12 22.98
gorithm to preserve the edges, texture and the basic details of
Proposed model 29.39 24.22
the image. In this case, we display the image difference between
the original and the filtered image. The proposed algorithm is
The next experiment is done in order to demonstrate the im- applied to the original image Lena. It’s clear that the geometri-
proved performance of the new algorithm. We compared its effi- cal structure are preserved which reveals that the new function
ciency with other existing algorithms [3] and [19–22]. The images tends to disappear close to the edges and details of the image.
are corrupted by additive Gaussian noise with a standard deviation Consequently, the details are preserved and kept intact. Table 9
σ = 25. The results of the PSNR show the improvement of the pro- illustrates the mean square error (MSE) of some denoised im-
posed model. This improvement is inspected from the good quality ages, comparing the new algorithm with different recent tech-
of the smoothed image Fig. 18 which is related to the important nique such as Non-Local-means algorithm (NL-means) [30] and
PSNR values (Table 6). others techniques in use [3,24,28] and [29]. The proposed algo-
In these simulations, higher PSNR value indicates the optimum rithm achieves smaller mean square error than the other tech-
denoising capability of the proposed model compared to the pre- niques which prove the efficiency of the proceeded scheme and the
sented methods. high quality of the recovered image. The energy dispersed by the
In Fig. 19, we expose a comparison based on the PSNR mea- filtering process is lower than this eliminated by the other meth-
sures of the proposed scheme with the BM3D algorithm [23] for ods.
 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215 213

Table 7
Quantitative comparison between different filtering techniques. The experiments presented in this paper are run on 1.83 GHz CPU with 1 G RAM.

Image Method PSNR MSSIM Iteration Time (s)

Pepper (30, 0) (512 × 512) TV model [24] 29.10 0.78 136
GVF model [25] 29.11 0.78 32 35.81
INGVF [26] 29.11 0.78 32 56.49
CONVEF [27] 29.91 0.79 32 1.96
Proposed model 31.33 0.86 10 1.15
Lena (20, 0) (512 × 512) TV model [24] 31.33 0.84 82
GVF model [25] 31.40 0.84 18 35.81
INGVF [26] 31.40 0.84 18 56.49
CONVEF [27] 31.45 0.85 18 1.96
Proposed model 34.25 0.863 8 1.2

Table 8
PSNR, MSSIM and iteration number of the compared methods.

Image (512 × 512) Noise variance Methods PSNR MSSIM Iteration

Lena 0.1 INGVF [26] 32.35 0.88 18
CONVEF [27] 32.39 0.88 18
Proposed model 37.73 0.99 5
0.2 INGVF 30.29 0.85 26
CONVEF 30.35 0.86 26
Proposed model 37.01 0.97 7
Pepper 0.1 INGVF [26] 29.85 0.86 18
CONVEF [27] 29.91 0.87 18
Proposed model 34.88 0.94 15
0.2 INGVF [26] 28.6 0.83 28
CONVEF [27] 28.66 0.84 28
Proposed model 32.7 0.91 18

Fig. 20. Method noise experience: displaying of the image difference. From left to right and from top to bottom: original image, Gauss filtering [28], anisotropic filtering [3],
total variation minimization [24], neighborhood filtering [29] and NL-means algorithm [30], original Lena, convolved image by the new function and obtained image difference.
214 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215

Fig. 21. Cropping section showing the preservation of the borders.

Table 9
Mean square error illustration.

Image Methods
GF [28] AF [3] TVF [24] YNF [29] NL [30] Proposed
algorithm
Lena 120 114 110 129 68 24.49
Baboon 507 418 365 381 292 60.72

Table 10
Color-image denoising: output-PSNR comparison with the C-BM3D algorithm (the
noise was added with sigma = 25).

Images C-BM3D algorithm [23] Proposed algorithm Error

Lena 32.27 33.35 1.08
House 33.03 34.46 1.43
Peppers 31.20 33.58 2.38

6. Conclusion

A new anisotropic diffusion technique is targeted for noise re-

Fig. 22. Color-image denoising: (a) and (c) are the noisy Lena image (σ = 15) and a
fragment of it; (b) and (d) are the filtered image and a fragment of it.
5.2.2. Color image denoising
To show the efficiency of the proposed algorithm based on
the new anisotropic diffusion function, we report in this section
some filtering results related to color images. Fig. 22 and Table 10
summarize these results. It can be observed from Fig. 22 that the
details and the borders are entirely and fully preserved. Table 10
manifests that the filtering process based the new function outper-
form the C-BM3D algorithm.
The whole presented results of synthetic as well as natural
Grayscale or color images prove the efficiency of the proposed
anisotropic diffusion scheme which allows obtaining high quality
of filtered images in a short time.
duction and edges preserving, has been proposed in this paper.
We presented a new scheme based on the gradient and the Lapla-
cian operator in order to keep the high quality of the restored
image. The threshold parameter was optimized in order to pre-
serve the content and the most important details in the image.
Many comparisons are developed to show the efficiency of the
proposed technique not only for synthetic and textured images
but also for natural images including Grayscale and color cate-
gory.
Many cropping and zooms are given in the experimental sec-
tion to provide better judgments. Also, the PSNR, MSE, iteration
number and the computational time are provided in comparison
tables gathering the quantitative values of the proposed technique
compared with many recent techniques in use. For psycho-visual
evaluation, various restored images are given and prove very well
the outperformance of the new technique to remove noise and
preserve edges content and details.
Acknowledgments
We would like to express our gratitude to the editor and
the anonymous reviewers for their constructive comments and
thoughtful suggestions which improved the quality and represen-
tation of our work.
 S. Tebini et al. / Digital Signal Processing 48 (2016) 201–215
Appendix A. Supplementary material
Supplementary material related to this article can be found on-
line at http://dx.doi.org/10.1016/j.dsp.2015.09.013.
References
[1] A.P. Witkin, Scale-space filtering, in: International Joint IEEE Conference on Ar-
tificial Intelligence, Karlsruhe, West Germany, 1983, pp. 1019–1021.
[2] J. Koenderink, The structure of images, Biol. Cybern. 50 (1984) 363–370.
[3] P. Perona, J. Malik, Scale-space and edge detection using anisotropic diffusion,
IEEE Trans. Pattern Anal. Mach. Intell. 12 (1990) 629–639.
[4] F. Catté, P. Lions, J. Morel, T. Coll, Image selective smoothing and edge detection
by nonlinear diffusion, SIAM J. Numer. Anal. 29 (1992) 182–193.
[5] M. Nitzberg, T. Shiota, Nonlinear image filtering with edge and corner enhance-
ment, IEEE Trans. Pattern Anal. Mach. Intell. 14 (8) (1992) 826–833.
[6] M.J. Black, G. Sapiro, D. Marimont, D. Heeger, Robust anisotropic diffusion, IEEE
Trans. Image Process. 7 (3) (March 1998) 421–432.
[7] G. Gilboa, N. Sochen, Y.Y. Zeevi, Image enhancement and denoising by complex
diffusion processes, IEEE Trans. Pattern Anal. Mach. Intell. 26 (8) (August 2004).
[8] H.M. Salinas, D.C. Fernandez, Comparison of PDE-based nonlinear diffusion ap-
proaches for image enhancement and denoising in optical coherence tomogra-
phy, IEEE Trans. Med. Imaging 26 (2007) 761–771.
[9] J. Weickert, B.M. Ter Harr Romeny, M.A. Viergever, Efficient and reliable
schemes for nonlinear diffusion filtering, IEEE Trans. Image Process. 7 (1998)
398–410.
[10] L. Alvarez, P.L. Lions, J.M. Morel, Image selective smoothing and edge detection
by nonlinear diffusion. II, SIAM J. Numer. Anal. 29 (3) (1992) 845–866.
[11] R.T. Whitaker, S.M. Pizer, A multi-scale approach to non-uniform diffusion,
CVGIP, Image Underst. 57 (1) (January 1993) 99–110.
[12] V. Bhateja, G. Singh, A. Srivastava, J. Singh, Speckle reduction in ultrasound im-
ages using an improved conductance function based on anisotropic diffusion,
in: IEEE International Conference on Computing for Sustainable Global Devel-
opment, INDIAcom, 2014.
[13] Y.-L. You, W. Xu, A. Tannenbaum, K. Mostafa, Behavioral analysis of anisotropic
diffusion in image processing, IEEE Trans. Image Process. 5 (11) (November
1996) 1539–1553.
[14] J. Yu, Y. Wang, Y. Shen, Noise reduction and edge detection via kernel
anisotropic diffusion, Pattern Recognit. Lett. 29 (2008) 1496–1503.
[15] Q. Chen, P. Montesinos, Q.S. Sun, Ramp preserving Perona–Malik model, Signal
Process. 90 (2010) 1963–1975.
[16] H.Y. Tian, H.M. Cai, J.H. Lai, Effective image noise removal based on difference
eigenvalue, in: Proc. ICIP, 2011, pp. 3357–3360.
[17] 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.
[18] C. Tsiotsios, M. Petrou, On the choice of the parameters for anisotropic diffusion
in image processing, Pattern Recognit. 46 (2013) 1369–1381.
[19] Y. Shi, Q. Chang, New time dependent model for image restoration, Appl. Math.
Comput. 179 (1) (2006) 121–134.
[20] 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.
[21] V.B.S. Prasath, A. Singh, Well-posed inhomogeneous nonlinear diffusion scheme
for digital image denoising, J. Appl. Math. 2010 (2010) 763847.
View publication stats
215
[22] V.B. Surya Prasath, D. Vorotnikov, Weighted and well-balanced anisotropic dif-
fusion scheme for image denoising and restoration, Nonlinear Anal., Real World
Appl. 17 (2014) 33–46.
[23] K. Dabov, A. Foi, V. Katkovnik, K. Egiazarian, Image denoising by sparse 3D
transform-domain collaborative filtering, IEEE Trans. Image Process. 16 (8) (Au-
gust 2007).
[24] L. Rudin, S. Osher, E. Fatemi, Nonlinear total variation based noise removal al-
gorithms, Physica D 60 (1992) 259–268.
[25] H. Yu, C. Chua, GVF-based anisotropic diffusion models, IEEE Trans. Image Pro-
cess. 15 (6) (2006) 1517–1524.
[26] 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.
[27] Y. Wang, W. Ren, H. Wang, Anisotropic second and fourth order diffusion mod-
els based on convolutional virtual electric field for image de-noising, Comput.
Math. Appl. 66 (10) (December 2013) 1729–1742.
[28] M. Lindenbaum, M. Fischer, A. Bruckstein, On Gabor contribution to image en-
hancement, Pattern Recognit. 27 (1994) 1–8.
[29] L. Yaroslavsky, Digital Picture Processing – An Introduction, Springer Verlag,
1985.
[30] A. Buades, B. Coll, J.M. Morel, A non-local algorithm for image denoising, in:
IEEE Computer Vision and Pattern Recognition, vol. 2, 2005, pp. 60–65.
S. Tebini was born in Tunisia, she obtained Master’s degree in 2012
from the ENSIT School in the field of image processing and specially image
de-noising. Currently, she is a doctorate candidate in the research unit
CEREP. She has published 2 journal papers and 4 conference papers in the
field of image processing.
Z. Mbarki was born in Tunisia, he received the master degree in Auto-
matic from the Higher School of Sciences and Techniques of Tunis, in 2011.
Currently he is in the third year of his Ph.D. Its research interests are fo-
cused on image processing, smoothing, restoration and segmentation.
H. Seddik was born in 15 October 1970 in Tunisia, he has obtained
the electromechanical engineer degree in 1995 and followed by the master
degree in “signal processing: speaker recognition” and the thesis degree in
image processing “watermarking using non-conventional transformations”.
He has over 14 international journals papers and 65 conference papers.
His domain of interest is: Audio-image and video processing applied in
filtering, encryption and watermarking. He belongs to the CEREP research
unit and supervises actually five thesis and 08 masters in the field.
E. Ben Braiek obtained his HDR on 2008 in Electrical engineering from
ENSET Tunisia. He is, presently, full professor in the department of elec-
trical engineering at the National High School of Engineering of Tunis
(ENSIT) and manager of the research group on vision and image processing
at the CEREP. His fields of interest include automatics, electronics, control,
computer vision, image processing and its application in handwritten data
recognition.