IMPULSE NOISE REMOVAL USING IMPROVED PARTICLE SWARM OPTIMIZATION

Full text

International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 4, April 2014
366
ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
IMPULSE NOISE REMOVAL USING
IMPROVED PARTICLE SWARM
OPTIMIZATION
Christo Ananth1, Vivek.T2, Selvakumar.S.3, Sakthi Kannan.S.4, Sankara Narayanan.D5
ABSTRACT— The fuzzy filter based on particle swarm
optimization is used to remove the high density image impulse
noise, which occur during the transmission, data acquisition and
processing. The proposed system has a fuzzy filter which has the
parallel fuzzy inference mechanism, fuzzy mean process, and a
fuzzy composition process. In particular, by using no-reference
Q metric, the particle swarm optimization learning is sufficient
to optimize the parameter necessitated by the particle swarm
optimization based fuzzy filter, therefore the proposed fuzzy
filter can cope with particle situation where the assumption of
existence of “ground-truth” reference does not hold. The
merging of the particle swarm optimization with the fuzzy filter
helps to build an auto tuning mechanism for the fuzzy filter
without any prior knowledge regarding the noise and the true
image. Thus the reference measures are not need for removing
the noise and in restoring the image. The final output image
(Restored image) confirm that the fuzzy filter based on particle
swarm optimization attain the excellent quality of restored
images in term of peak signal-to-noise ratio, mean absolute
error and mean square error even when the noise rate is above
0.5 and without having any reference measures.
Keywords — Fuzzy filter, Particle Swarm Optimization (PSO),
Structural Similarity, Singular Value Decomposition.
I. INTRODUCTION
Digital images are often corrupted by impulsive
noise during data acquisition, transmission, and
processing. Here the turbulent particle swarm
optimization (PSO) (TPSO)-based fuzzy filtering (or
TPFF for short) approach to remove impulse noise from
highly corrupted images.
1 Assistant Professor/ECE, Francis Xavier Engineering College,
Tirunelveli
2 U.G.Scholar/ECE, Francis Xavier Engineering College, Tirunelveli
3 U.G.Scholar/ECE, Francis Xavier Engineering College, Tirunelveli
4 U.G.Scholar/ECE, Francis Xavier Engineering College, Tirunelveli
5 U.G.Scholar/ECE, Francis Xavier Engineering College, Tirunelveli
The proposed fuzzy filter contains a parallel
fuzzy inference mechanism, a fuzzy mean process, and a
fuzzy composition process. To a certain extent, the TPFF
is an improved and online version of those genetic-based
algorithms which had attracted a number of works during
the past years. As the PSO is renowned for its ability of
achieving success rate and solution quality, the superiority
of the TPFF is almost for sure. Therefore, the proposed
fuzzy filter can cope with practical situations where the
assumption of the existence of the “ground-truth”
reference does not hold. The experimental results confirm
that the TPFF attains an excellent quality of restored
images in terms of peak signal-to-noise ratio, mean square
error, and mean absolute error even when the noise rate is
above 0.5 and without the aid of noise-free images.
Fig.1. Image reconstruction process
The Fig.1 shows the process for reconstructing
the noisy image. The process consist of the input image
which is a highly corrupted image, fuzzy filter, the output
NOISY IMAGE
FUZZY FILTER
PARTICLE
SWARM
OPTIMIZATION
FITNESS
EVALUATION
PSNR
CONDITION
RECONSTRUCTED
IMAGE
YES
NO

International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 4, April 2014
367
ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
of the fuzzy filter is given to the PSO, then the PSNR
value is checked, if the PSNR value is satisfied then the
output image(reconstructed image) is obtained. If the
PSNR value is not satisfied then the particles values are
changed by the PSO techniques, this process is repeated
until the expected PSNR value reached, Once the PSNR
value satisfies our expected value then the output image
will obtained.
II MATERIALS AND METHODS
The fuzzy set technique serves as an important
ingredient in the development of information
technologies. In the field of information systems, the
fuzzy set plays a role in the development of intelligent
systems and the storage of imprecise linguistic
information.
Applications can be found in many areas such as
control systems, fuzzy time series forecasting, artificial
intelligence and image processing[7]. In particular, fuzzy
image filters have the advantage of being easy to realize
by simple fuzzy rules that characterize a particular noise.
Over the past years, a huge amount of fuzzy-based noise
reduction methods were developed. The noise adaptive
fuzzy switching median filtering mechanism [1]
employed fuzzy reasoning to handle uncertainty presented
in the extracted local information introduced by noise.
The fuzzy impulse noise detection and reduction method
dealt with the images corrupted by fat-tailed noise.
Particle swarm optimization (PSO) is a decade-
old concept in the global optimization domain. It is a
popular computational technique developed by Eberhart
and Kennedy, based on the social behavior of birds
flocking for food searching. The PSO was generally
found to outperform other evolutionary algorithms (such
as GAs, mimetic algorithms, ant-colony optimization, and
shuffled frog-leaping algorithm) in terms of success rate
and solution quality while being second best in terms of
processing time. To provide quantitative data on the
fidelity of rendered images, quality metrics are of
paramount importance for the PSO.
These metrics can be typically divided into two
main categories, namely, full reference and no reference.
Full-reference metrics need a complete reference image,
and what they calculate is the similarity between the
target and reference images. Metrics such as the classical
peak signal-to-noise ratio (PSNR), MSE, MAE, and the
structural similarity (SSIM) belong to this category.
Unfortunately, the full-reference metrics are hardly
applicable due to the unavailability of the reference
images in most practical applications. To overcome the
deficiency, researchers thus turned their attention to the
parameter optimization problem of the field (i.e., no
reference) using the generalized cross-validation, L-curve
method, Stein’s unbiased risk estimate (SURE), etc. In
recent years, the SURE gradually became popular as it
provided an efficient means for an unbiased estimation of
the MSE without the reference image. However, the
performance of the SURE is barely acceptable for
Gaussian noise, and an accurate estimation of the noise
variance is still necessary.
In fact, an ideal no-reference measure that is
useful for the parameter optimization problem should take
both noise and blur conditions into account. Recently,
Xiang and Milanfar proposed a no-reference image
quality metric based upon the singular value
decomposition (SVD) of the local image gradient matrix
and provided a quantitative measure of true image
content.
III. PROPOSED METHOD
In a scenario where an image (M × N) is
corrupted by impulsive noise, let xij be the 8-b
gray level of the noisy pixel at location (i, j) and
1 ≤ i ≤ M,1 ≤ j ≤ N
Xij= (1)
where φij denotes a noise-free image
pixel and nij denotes a noisy impulse at the
location (i, j) with a noise corruption rate p. The
simplest and most frequently used impulse noise
model is the salt-and-pepper noise, where noisy
pixels take either the minimal or maximal value,
i.e., nij {Lmin, Lmax} with Lmin and Lmax,
respectively representing the lowest and highest
pixel luminance values within the dynamic
range.
Fig.2. Examples of membership function
nij, with probability p
φij,with proability 1-p

International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 4, April 2014
368
ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
Moreover, the uniform impulse noise model can
often be found in the contemporary literature too, where
the noisy pixel can have any value within the dynamic
range with equal probability. Membership functions for
fuzzy sets can be defined in any number of ways as long
as they follow the rules of fuzzy sets. Three examples of
frequently used membership functions for calculating the
fuzzy number are shown in Fig. 2.
However, it is hard to say which membership
function is the best one. The choice of the membership
function depends on the subjective aspect of fuzzy logic,
which allows the desired values to be interpreted
appropriately. Bigand and Colot had demonstrated the
design of a robust fuzzy filter using certain membership
functions, and we also obtain similar results with triangle
and trapezoid membership functions.
It is thus straightforward for us to employ the
triangle membership functions for simplifying the tuning
procedure of filter parameters. The membership function
fA(x) with the fuzzy set A denoted by A = [aA, bA, cA] is
given in Fig.2(c)
fA(x) (2)
A fuzzy set allows its members to have different
degrees of membership. Here, we define S (S ≥ 3) fuzzy
sets for a gray-level image. Suppose that the image can be
classified into “very dark” (VD), “dark” (DK), “medium”
(MD), “bright”(BR), and “very bright” (VB), as shown in
Fig. 3. The membership functions of fuzzy sets VD, DK,
MD, BR, and VB are denoted as VD = [aVD, bVD, cVD],
DK = [aDK, bDK,cDK],M = [aMD, bMD, cMD], B = [aBR, bBR,
cBR], and VB = [aVB, bVB, cVB], respectively.
Fig. 3. Membership function
The architecture of the fuzzy filtering process for
impulsive noise removal is shown in Fig. 4.
Fig. 4. Architecture of fuzzy filtering process
Fig. 5. Linguistic term
A. Parallel Fuzzy Inference Process
Here, the linguistic modifier λ is adopted to
modify the membership functions to fit the linguistic
characterization. Equation defines the computing
mechanism of the fuzzy inference vector through the
linguistic modifier. Two of the most well-known
modifiers are the erosion corresponding to “very” (λ = 2)
and the dilation corresponding to “more-or-less” (λ = 0.5).
Fig. 5 shows the graphic representations of three specific
linguistic terms, including normal bright (λ=1), very
bright (λ=2), and more-or-less bright (λ=0.5).
O3
AS = ( (4 (3)
where 1 ≤ r ≤ n2, 1 ≤ s ≤ S, 1 ≤ l ≤ S. Finally,
the fuzzy inference rules are denoted as follows:
Fuzzy Rule LS :
If (x1 is LS) AND (x2 is LS) AND . . . AND (x9 is LS)
THEN is LS, where LS ∈ {VDK, DK, MD, BR,
VBR}.
((fAS(xrl))2 , if “very”( λ=2)
((fAS(xrl)) , if “normal” (λ=1)
((fAS(xrl))0.5,if “more or less (λ=0.5)

International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 4, April 2014
369
ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
Equation (4.4) shows the computing mechanism
for the output of the aforementioned rules
(4)
B. Fuzzy Mean Process
The fuzzy mean process aims at deriving a fuzzy
mean of the processing windows. The process is carried
out using a trapezoidal membership function of four
parameters, i.e., A = [aA, bA, cA, dA], shown in Fig. 2(c).
Equation specifies the computing process with the fuzzy
interval F_mean for the fuzzy mean process. Moreover,
the fuzzy function F_mean for the fuzzy mean process is
shown in Fig. 6.
(5)
Where
(6)
and the fuzzy set F_mean = [0, α, β, 255].
C. Fuzzy Composition Process
The core concept of the fuzzy filter is to pick out
an adequate inference drawn from different linguistic sets
along with the output of the fuzzy mean process and then
merge these two probable informative sources by fuzzy
weighting.
V.SIMULATION RESULTS
The Simulation results of the proposed method is as given
below.
Fig. 6. Actual image (Before noise)
Fig.6. shows the input image which is the actual image
before adding noise. Fig.7. shows the noisy image with
noise rate of 0.6
Fig. 7. Noisy image (noise rate 0.6)
Fig.8. shows the Restored image which is obtained by
applying the proposed technique to give a noise free
image
Fig. 8. Restored image (noise free image)
TABLE I
COMPARISON OF NOISE AND PSNR
S.No
Noise
Density
Fuzzy
PSNR
Fuzzy-PSO
PSNR
1.
0.4
32.674
33.474
2.
0.5
30.913
31.567
3.
0.6
28.167
29.786
4.
0.7
23.757
24.589
5.
0.8
19.322
21.543
Table I shows the values of Fuzzy PSNR and Fuzzy-PSO
PSNR for different values of Noise Density. The
Proposed Fuzzy-PSO PSNR is able to outsmart the Fuzzy
PSNR with larger values proving its reliability.

International Journal of Advanced Research in Electronics and Communication Engineering (IJARECE)
Volume 3, Issue 4, April 2014
370
ISSN: 2278 – 909X All Rights Reserved © 2014 IJARECE
VI. CONCLUSION
As the PSO has several desirable properties such
as fewer parameters, easier implementation, low
computational cost, and faster convergence to the optimal
solution, the superiority of the fuzzy filter based on PSO
is almost for sure. The no-reference image quality Q
metric adopted here can reflect the extent of both
blurriness and randomness of images without any prior
knowledge.
Using the fuzzy filter based on PSO the high
impulse noise in the image was restored without having
any reference measures, any trial and error.
REFERENCES
[1] Astola, P. Haavisto, Y. Neuvo, Vector median filters,
Proc. IEEE 78 (4) (2011) 678–689.
[2] Bovik, Handbook of Image and Video Processing,
Academic Press, 2010.
[3] C.-H. Chen, C.-Y. Chen, C.-H. Chen, Intelligent
adaptive subband-based multi-state median filter in lowly-
corrupted images, Int. J. Innovative Comput. Inf. Control
5 (9) (2012) 2917–2926.
[4] T. Chen, K. K. Ma, and L. H. Chen. Tri-State Median
Filter for ImageDenoising. IEEE Transactions on Image
Processing, 8(12):1834 – 1838, December 2009.
[5] M. Emre Celebi, Hassan A. Kingravi, Y. Alp
Aslandogan, Nonlinear vector filtering for impulsive
noise removal from color images, J.Electronic imaging
16(3) , 033008 (2010) 1-21.
[6] Fawcett, An introduction to ROC analysis, Pattern
Recognit. Lett. 27 (8) (2006) 861–874.
[7] R. Pandey, An improved switching median filter for
uniformly distributed impulse noise removal, Proc. World
Acad. Sci. Eng. Technol. 38 (2010) 349–351.
[8] N. Suetake, R. Kubota, E. Uchino, Random-valued
impulse noise detection considering color distance and
correlation among RGB components, Int. J. Innovative
Comput. Inf. Control 7 (7(A)) (2011) 3799–3810.
[9] N. Suetake, A new impulse noise detector based on
pattern classification, IEEJ Trans. Electron. Inf. Syst. 122
(5) (2006) 902–903.
[10] T. Sun, Y. Neuvo, Detail-preserving median based
filters in image processing, Pattern Recognit. Lett. 15 (4)
(2004) 341–347.
[11] Z. Wang, D. Zhang, Progressive switching median
filter for the removal of impulse noise from highly
corrupted images, IEEE Trans. Circuits Syst. II: Analog
Digit. Signal Process. 46 (1) (2004) 78–80.
[12] X. Zhang, Z. Yin, Y. Xiong, Adaptive switching
mean filter for impulse noise removal, in: 2008 IEEE
Congress on Image and Signal Processing, 2008, pp. 275–
278.