IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.

11, November 2009

170

High Density Salt and Pepper Noise Removal in Images

using Improved Adaptive Statistics Estimation Filter
V.Jayaraj ,

D.Ebenezer, K.Aiswarya

Digital Signal Processing Laboratory,

Department of Electronics and Communication Engineering,
Sri Krishna College of Engineering and Technology, kuniamuthur,
Coimbatore- 641008, Tamil nadu, India
Summary
In this paper a Non-linear Adaptive Statistics Estimation Filter to
remove high density Salt and Pepper noise is presented. The
algorithm detects the pixel corrupted by salt and pepper noise
and replaces them with a value estimated using proposed
algorithm. The algorithm detects the corrupted pixel at the initial
stage itself. The performance of proposed algorithm is compared
with various filters and has better image quality than the existing
filters. The proposed method removes noise effectively even at
noise level and preserves the fine details and edges effectively
with reduced streaking at higher noise densities. The proposed
filter has better image quality then existing Non-linear filters of
this type.

Key words:
Salt and Pepper noise, Median filter, Estimation filter, Lorentzian
Estimator.

1. Introduction
During the transmission of images and videos over
channels, images and videos are often corrupted by noise
and degradations due to faulty communication or noisy
channels. Such noises may be introduced due to faulty
camera or the like [1], [2]. In early development of signal
and image processing linear filters were the primary tools.
But linear filters have poor performance in the presence of
noise that is additive in nature. They do not perform well
in the presence of signal dependent noise. In image
processing linear filters tend to blur the edges and do not
remove impulse noise effectively. Non-linear filters are
developed to overcome these limitations. A standard
median filter is a basic non-linear filter that will preserve
the edges and remove impulse noise. Median filter
replaces every pixel by its median value neighborhood.
But this removes some desirable details in the image [3],
[4]. Different remedies of the median filter have been
proposed, e.g. the Standard Median Filter (SMF),
Weighted Median Filter (WMF) [5],[6],[7] and Adaptive
Median Filter [8]. These filters first identify possible noisy
pixels and then replace them by using the median filter or

Manuscript received November 5, 2009

Manuscript revised November 20, 2009

its variants, while leaving all other pixels unchanged. In

these filters more weight is given to some pixels in the
processing window. The main drawback of these filters is
that the noisy pixels are replaced by some median value in
their vicinity without taking into account local features
such as the possible presence of edges. Hence details and
edges are not recovered satisfactorily, especially when the
noise level is high. Decision Based Median Filtering
Algorithm(DBA) [9], Robust Estimation Algorithm(REA)
[10] was proposed to remove high density impulse noise.
Decision Based Algorithm removes high density salt and
pepper noise. The corrupted pixels are replaced by median
or the immediate neighborhood pixel. At higher noise
densities the median may also be a noisy pixel and this
produces streaking at higher noise densities. The major
disadvantage of this method is, the quality of the restored
image degrades as the noise level increases above 50
percentage..

2. Robust Statistics Estimation

Robust estimation is based on the principle that in
robustness safety is more important than efficiency [11].
Consider Median as an estimator. Let x1, x2, x3xn
denote a random sample from a distribution having Pdf
f(x). Let Y1 be the smallest of Xi, Y2 the next Xi in order
of magnitude, and Yn the largest of Xi. That is Y1 < Y2 <
. < Yn. (X1, X2Xn, are arranged in ascending order pf
magnitude). Yi, I = 1, 2n is called the ith order statistic
of the random sample X1, X2,.. .Xn. The median is based
on L1 norm or it is an ML estimate for Laplacian
Distribution. The asymptotic efficiency is greater than one
for long tailed distributions. For median the influence
function is constant. This property makes median a good
and robust estimator. The robust property is shown in
Figure 1[12].An M-estimator which is fairly good for a
distribution having outliers (extreme data value in the
scatter) is called Robust estimator. Median is a robust
estimator. Robust estimators are compared in terms of

IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.11, November 2009

influence function. The influence function for median

estimators is given in equation 1.

171

as a noisy pixel and replaced by a value determined by

the following algorithm using lorentzian estimator. Let X
denote the noise corrupted image and for each pixel X (i, j),
the following algorithm is applied.

3.1. Pseudo code for the Proposed Algorithm

Let X (i, j) = corrupted image
Let output (i, j) = restored image
Wmax = 7 X 7
Window size , W=3;
for i = 1 X no of Rows in a image
for j = 1 X no of Columns in a image
if X (i,j) == 0 or X (i,j) == 255 then
S= pixel elements in the window;
Smed = Median(S);
if Smed == 0 or Smed == 255
then
Fig. 1 Influence function with respect to outlier.

Lorentzian estimator has an Influence function which

tends to zero for increasing estimation distance and
maximum breakdown value; therefore it can be used
to estimate the original image from noise corrupted
image. The Lorentzian estimator and its influence
function are shown in equations (2) and (3).

if W < W max then

W = W + 2;
Corrupted pixel = processed neighbhourhood pixel value;
else
process the next
window;
end
else
p (i,j) = Smed X(i,j);
if p == 0 then
Corrupted pixel = neighbhourhood pixel
value;
e lse
//calculate the robust influence function
(p) = 2p / (2 2 + p2)
(4)
= s/2;
// outlier rejection point
s = * N
; (5) // maximum expected
outlier
N = image standard deviation
= 0.3; // a smoothening factor
calculate

the estimate values from (6) and (7)

Robust estimation is applied to estimate image intensity

values in image denoising. Image model is assumed
non-stationary and, thus, the image pixels are taken from
fixed windows and robust estimation algorithm is applied
to each window.

3. Robust Statistics Estimation Algorithm

In this approach impulses are first detected based on the
minimum (0) and maximum (255) value. If the current
pixel lies inside the dynamic range [0,255] then it is
considered as noise free pixel. Otherwise it is considered

Estimated pixel value = s1 / s2;

end
end
end
end

172

IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.11, November 2009

Fig. 2. Flow chart for the Proposed Algorithm

4. Results and Discussions

e

The proposed algorithm is tested using images such as

Lena and baboon (gray and colour) and shown in Figure
(3) Figure (8). The performance of filters is tested at
different levels of noise densities, and the results are
shown in Tables 1 - 4. The performance of the proposed
algorithm is tested for various levels of noise density and
compared with standard filters namely Standard Median
Filter (SMF), Weighted Median Filter (WMF), Adaptive
Median Filter (AMF), Decision Based Algorithm (DBA),
Robust Estimation Filter(REF) are compared in terms of
Peak Signal-To-Noise Ratio (PSNR)(8), Mean Absolute
Error (MAE)(9) , Mean Square Error (MSE)(10) , Image
Enhancement Factor(IEF) (11) and the results are plotted
in Figure 4.

Fig 3. (a) Original Lena image (b) Noisy image of noise density 70%.
Restoration results of (c)Standard median filter (d) Weighted median
filter (e) Adaptive median filter (f) Decision based algorithm (g)
Robust Estimation algorithm (h) proposed method

IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.11, November 2009

Table 1: Comparative results of various filters in terms of PSNR

for Lena. jpg image

173

NOISE DENSITY VS PSNR

PSNR

50
40

SMF

30

WMF

20

AMF

10

DBA
REA

0
10% 20% 30% 40% 50% 60% 70% 80% 90%

PA

NOISE DENSITY

a
Table 2: Comparative results of various filters in terms of MAE
for Lena. jpg image

M AE

NOISE DENSITY VS MAE

35
30
25
20
15
10
5
0

SMF
WMF
AMF
DBA
REA
10% 20% 30% 40% 50% 60% 70% 80% 90%

PA

NOISE DENSITY

NOISE DENSITY VS MSE

Table 3: Comparative results of various filters in terms of MSE

for Lena. jpg image

6000

M SE

5000

SMF

4000

WMF

3000

AMF

2000

DBA

1000

REA

0
10% 20% 30% 40% 50% 60% 70% 80% 90%

PA

NOISE DENSITY

NOISE DENSITY VS IEF

Table 4: Comparative results of various filters in terms of IEF

for Lena. jpg image
600
500

SMF

IEF

400

WMF

300

AMF

200

DBA

100

REA

0
10% 20% 30% 40% 50% 60% 70% 80% 90%

PA

NOISE DENSITY

d
Fig.4. Comparison graph for PSNR, MAE, MSE and IEF for different
noise densities of Lena.jpg image

174

Corrupted image

IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.11, November 2009

Denoised image

Corrupted image

Denoised image

a. 60 % Noise Density

a. 60 % Noise Density

b. 70 % Noise Density

b. 70 % Noise Density

c. 80 % Noise Density

c. 80 % Noise Density

d. 90 % Noise Density
d. 90 % Noise Density
Fig 5. Corrupted and Denoised images of Baboon.jpg
at different noise densities

Fig. 6. Corrupted and Denoised images of Boat.jpg

at different noise densities

IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.11, November 2009

Corrupted image

Denoised image

a. 60 % Noise Density

b. 70 % Noise Density

c. 80 % Noise Density

d. 90 % Noise Density
Fig. 7. Corrupted and Denoised images of Lenacolour.jpg
at different noise densities

Corrupted image

Denoised image

a. 60 % Noise Density

b. 70 % Noise Density

c. 80 % Noise Density

d. 90 % Noise Density
Fig. 8. Corrupted and Denoised images of Baboonlour.jpg
at different noise densities

175

176

IJCSNS International Journal of Computer Science and Network Security, VOL.9 No.11, November 2009

[5] D.R.K. Brownrigg , The weighted median filters

Communication, ACM, vol. 27, no.8, 1984, pp. 807-818,
[6] S.J. Ko and Y.H.Lee , Center weighted median filters and
their applications to image enhancement IEEE Trans.
Circuits Systems, vol. 38, no. 9,1991, pp. 984 993.
[7] G. Arce and J. Paredes , Recursive Weighted Median Filters
Admit-ting Negative Weights and Their Optimization IEEE
Trans. on Signal Process-ing, vol. 48, no. 3, 2000 ,pp.
768-779.
[8] H. Hwang and R.A. Haddad , Adaptive median filters: new
and results IEEE Trans. on Image Processing, vol. 4, no.4,
1995, pp. 499-502.
[9] K.S. Srinivasan and D. Ebenezer , A New Fast and Efficient
Decision- Based Algorithm for Removal of High-Density
Impulse Noises IEEE signal processing letters, vol. 14, no.
3, 2007, pp.189 -192.
[10]V.R.Vijaykumar, P.T.Vanathi, P.Kanagasabapathy and
D.Ebenezer, High Density Impulse Noise Removal Using
Robust Es-timation Based Fil-ter IAENG International
Journal of Computer Science, August 2008.

Where rij is the original image, xij is the restored

image, nij is the corrupted image.

4. Conclusion
An Adaptive Robust Statistics Estimation Based Filter to
remove low to high-density salt and pepper noise with
edge preservation in digital images is proposed in this
paper. The proposed filter performs well for both gray
scale and color images. For lower noise density up to 30%
almost all the algorithms perform equally well in removing
the salt and pepper noise completely with edge
preservation. For noise densities above 50%, the standard
algorithms such as SMF, WMF, AMF fail to remove the
salt and pepper noise completely. In case of high density
noise, the performance of these methods is very poor in
terms of noise cleaning and edge detail preservation. The
recently proposed algorithm DBA and REA remove noise
at high densities but they produce streaking effect and not
suitable for noise densities above 60%. Experimental
results show that the proposed method restores the original
image much better than standard non-linear median-based
filters. The proposed filter requires less computation time
when compared to other adaptive methods.

V.Jayaraj obtained his Bachelors Degree (BE) in Electronics

and Communication Engineering from Amrita institute of
Technology and Science, Coimbatore and Masters Degree (ME)
in VLSI Design from Kongu Engineering College, Perundurai,
Tamilnadu, India. He is Pursuing his Doctoral Degree in the area
of Non-linear image processing from Anna University
Coimbatore, India. Currently, he is working as a Lecturer in the
department of Electronics and Communication Engineering, Sri
Krishna College of Engineering and Technology, Coimbatore,
India. His areas of interests are Non-linear image processing and
VLSI Signal Processing.

D.Ebenezer
obtained his Bachelors Degree (BE) in
Electronics and Communication Engineering from PSG College
of Technology and Masters Degree(ME) in Applied Electronics
from PSG College of Technology , Coimbatore, India and Ph.D
in Electrical Engineering from Anna University, Chennai. He
served as a member of faculty of the department of Electronics
and Communication Engineering, College of Enginnering, Anna
University, Chennai, India for twenty one years since 1985. At
present he is a professor in the department of Electronics and
Communication Engineering, Sri Krishna College of Engineering
and Technology, Coimbatore, India. He has published Numerous
research papers in the areas of Signal processing and Nonlinear
filtering. He is a referee for journal of Medical Engineering and
physics, IET, IEEE..

References
[1] R.C.Gonzalez and R.E. Woods , Digital Image Processing.
Addison Wesley, 2002.
[2] A.C. Bovik , Handbook of Image and Video Processing.
Academic Press, 2000
[3] I. Pitas and A.N. Venetsanopoulos, Nonlinear Digital
Filters: Principles and applications. Boston, MA: Kluwer
Academic, 1990.
[4] J. Astola and Kuosamanen, Fundementals of Nonlinear
Digital Filtering. CRC Press, 1997.

K.Aiswarya is currently doing her Bachelors Degree (BE) in

Electronics and Communication Engineering from Sri Krishna
College of Engineering and Technology, Coimbatore, India. His
areas of interests are Non-linear image processing and VLSI
Signal Processing.