Research Inventy : International Journal of Engineering and Science

Research Inventy: International Journal Of Engineering And Science
Vol.3, Issue 3 (June 2013), PP 13-19
Issn(e): 2278-4721, Issn(p):2319-6483, Www.Researchinventy.Com
13
Restoring Corrupted Images Using Adaptive Fuzzy Filter
Vandana S. Khobragade1
, Sunita Munde2
1,2
Department of Electronics and Telecommunication Lokmanya Tilak College of Engineering
Navi Mumbai , India
ABSTRACT - The proposed histogram adaptive fuzzy (HAF) filter is particularly effective for removing
highly impulsive noise while preserving edge sharpness. This is accomplished through a fuzzy smoothing filter
constructed from a set of fuzzy IF-THEN rules, which alternate adaptively to minimize the output mean squared
error as input histogram statistics change. An algorithm is developed to utilize input histogram to determine
parameters for the fuzzy membership function. As compared to the conventional median filters (MF), the
proposed method has the following merits: it is simple and it has superior performance compared to other
existing ranked-order filters (including MF) for the full range of impulsive noise probability
Keywords - Histogram adaptive fuzzy (HAF) filter, median filters , membership functions.
I. INTRODUCTION
Median filtering (MF) is a nonlinear technique that is known for its effectiveness in removing impulsive
noise while preserving edge sharpness. The 1-D MF is realized by passing a window over the input data and
taking the median value of the data inside the window as the output associated with the center of the window.
In image processing applications, two-dimensional median filters have been used with some success [1]. The
simplest way is to pass a 2-D window, such as a square mask, over the 2-D input image [2]. As with the 1-D MF,
the pixels inside the window are ranked according to their gray intensity values, and the median value is taken as
the output. Although noise suppression is obtainable by using MF, too much signal distortion is introduced, and
features such as sharp corners as well as thin lines are lost. To overcome these problems, several variations of
median filters have been developed specifically the detection –estimation based filter [3],which incorporated a
statistical noise detection algorithm and the median filter for removal impulsive noise. Due to the lack of
adaptability these median filters cannot perform well when NP≤20%. Adaptive systems based on fuzzy or neural
networks with data driven adjustable parameters have emerged as attractive alternatives [4]In this category noise
exclusive adaptive filter [5] have been developed..
Neural networks exploit their frameworks with many theorems and efficient training algorithms. They
embed several input and output mappings on a black box web of connection weights .However, we cannot
directly encode the simple rule of a spatial windowing operation, such as: “If most of the pixels in an input
window are BRIGHT, then assign the output pixel intensity as BRIGHT.” On the other hand, fuzzy systems can
directly encode structured knowledge. Fuzzy systems may invariably store banks of common-sense rules
linguistically articulated by an expert or may adaptively infer and modify their fuzzy rules using representative
symbols (e.g., DARK, BRIGHT) as well as numerical samples. Fuzzy systems and neural networks naturally
combine and this combination produces an adaptive system .The hybrid neuro –fuzzy networks do not represent
general means in restoring images [6].used the adaptive fuzzy median filter(AFMF) with the backpropagation
algorithm [7] to tune a set of randomly given initial membership functions.
In this paper ,we propose a novel adaptive fuzzy filter(HAF) in which a set of memberships is estimated
from the input histogram and used to achieve restoration without any training. In Section II the fuzzy inference
rules related to the task of median filtering and the system architecture of HAF are introduced. In Section III a
systematic algorithm based on conservation in the histogram potential to obtain a set of membership functions is
implemented .In section V experiments are presented to characterize HAF as well as to compare it with and
other existing median filter (MF) and WFM. We also show the generalization capability and adaptive property of
HAF.
II. SYSTEM ARCHITECTURE OF HAF
Here, we will first discuss pre-processing of a 2-D input image. For the problem of interest here, we
assume an input gray image sized 256 x 256 with a pixel intensity between 0 and 255.
Since a noise-corrupted image contains a high level of uncertainty, we can consider it as an array of fuzzy
variables [2][4]. HAF is designed to create three fuzzy membership functions for three fuzzy sets, namely, Dk
(Dark), Md (Medium), and Br (Bright). Therefore, each input pixel intensity p(k, l) is considered as a fuzzy
variable, and the membership degree of three fuzzy sets, Dk, Md, and Br, are calculated, respectively.

14
Since fuzzy systems can directly encode structured knowledge in a numerical framework, in order to be
easily processed by HAF, the intensity of each input pixel at (k, l) is normalized to 0≤p(k, l ) ≤1. In HAF, a square
window of size 3 x 3 is used to scan across the entire image, where the filter output associated with the centre of
the window is denoted as Y. Thus, the elements of the window W ((k, l)) centered at (k, l) are as follows:
x
l
= p(k-1,l-1) , x
2
= p(k-1,l ) ,x
3
= p(k-1,l+1) , x
4
= p(k ,l-1),
x
5
= p(k, l), x
6
= p(k ,l +1), x
7
= p(k+1, l-1), x
8
= p(k+1, l),
and x
9
= p(k+1,l+1).
Each element is considered to be a fuzzy variable, and the membership functions identify the grade of
brightness for each input pixel. Equation (1) gives the bell-shaped membership functions used in HAF:
(1)
i=1,2 …..9 , j=D
k
, M
d
, B
r
Where a
j
, b
j
, and c
j
are adjustable parameters of the bell-shaped membership function. Let p = p(k, l)
denotes the predicted intensity for a pixel at (k, l);
Premise 1:
IF {( p(k-1,l-1)is D
k
) and (p(k-1,l ) is D
k
) and (p(k-1,l+1) is D
k
) and (p(k ,l-1) is D
k
) and (p(k, l) is D
k
) and (p(k, l
+1) is D
k
) and (p(k+1, l-1)is D
k
) and (p(k+1, l) is D
k
) and (p(k+1,l+1) is D
k
)}
THEN {Y is D
k
}
Premise 2:
IF {( p(k-1,l-1)is M
d
) and (p(k-1,l ) is M
d
) and (p(k-1,l+1) is M
d
) and (p(k ,l-1) is M
d
) and (p(k, l) is M
d
) and (p(k, l
+1) is M
d
) and (p(k+1, l-1)is M
d
) and (p(k+1, l) is M
d
) and (p(k+1,l+1) is M
d
)}
THEN {Y is M
d
}
Premise 3:
IF {( p(k-1,l-1)is B
r
) and (p(k-1,l ) is B
r
) and (p(k-1,l+1) is B
r
) and (p(k ,l-1) is B
r
) and (p(k, l) is B
r
) and (p(k, l +1)
is B
r
) and (p(k+1, l-1)is B
r
) and (p(k+1, l) is B
r
) and (p(k+1,l+1) is B
r
)}
THEN {Y is B
r
}
Consequently,
IF { p(k, l ) is closest to D
k
} THEN {Y is D
k
} else
IF { p(k, l) is closest to M
d
} THEN {Y is M
d
} else {Y is B
r
}
We will first consider a 3 x 3 window W(k ,l) that scans the image from left to right and from top to bottom. In
each scan, nine pixels are ranked according to gray intensity.
We further define
,
If p(k ,l ) = Min{W((k, l))} or Max{W((k, l))} or {p(k, l) ≤ T} or {p(k,l ) ≥(1.0-T)}.
Where T is a threshold. The set N
imp
contains pixels that are most likely corrupted by impulses. To explain, we
will use the input sub image as an example.
(x
l
,x
2
,x
3
,x
4
,x
5
,x
6
,x
7
,x
8
,x
9
)=(0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8,0.9)
Given the nine input pixels x
i
of a sub image, HAF performs fuzzification using Eq. (1) i.e., the
membership degree mij= m
j
(x
i
) is calculated. In particular, the slope (specified by b
j
) is made excessively larger
than a
j
and c
j
such that impulsive noise is to be filtered out by the membership functions. Accordingly, in HAF, b
j
is initialized as a large number, say, 15. The second step is to normalize the membership degree of each input
pixel using
(2)
i=1,2 …..9, j=D
k
, M
d
, B
r
After that, HAF calculates the weighted input sum

15
Finally, in the defuzzification step, the predicted intensity y=p is computed by using
Where
x
i
= 0 and I
i
= 0 , if x
i
€ N
imp,
x
i
= xi and I
i
=1 , otherwise .
The final output of HAF can be determined using the following rule:
If
Output Y = Sum
j
that has minimum ,
j=D
k
, M
d
, B
r
Else
Output Y = p (k,l )
End.
That is, if , then the three intermediate weighted results Sum
j
are compared with the predicted
output p(k, l ), from among which the one closest to p(k, l) is chosen to replace the pixel intensity at (k, l). If the
pixel at
p
, then p(k,l ) is used as the final output. With this architecture, the performance of HAF hinges on
the weights w
ij
calculated in Step 2, which in turn is determined by the fuzzy membership functions in Step 1.
III. ESTIMATING MEMBERSHIP FUNCTION
Here we present a systematic algorithm, which can be used to obtain a set of membership functions, by
which we mean that these functions are ideal for directly performing fuzzification without any training. The
proposed approach to obtaining histogram based membership functions(HMF) starts with utilizing the histogram
statistics of the corrupted input image to estimate the histogram of the original image.This is different from the
approach employed by WFM filter [3]. To facilitate the following discussion, important nomenclatures are first
defined as follows:
Membership function of D
K
(4)
i=1,2…….9
Membership function of Md
(5)
i=1,2…….9
Membership function of Br
(6)
i=1,2…….9
IV. CONFIGURATION OF MEMBERSHIP FUNCTION USING HISTOGRAM
Given the histogram H[n], define X(n) = normalized input intensity. Also, b
1
=b
2
=b
3
=15, as reasoned in the
preceding sections. We also denote A as the parameter matrix of HMF:
Where
A (1,i)= parameters a
l
, b
l
, c
l
of membership function Dk,
2
, b
2
, c
2
of membership function M
d
,
3
, b
3
, c
3
of membership function B
r
.
Next, we will use the histogram H[n] shown in Fig. 1. (b) to explain how the initial parameters of HMF are
derived. The histogram is first divided into K equal-length segments, e.g., K=3, and we have D
k
, M
d
, and B
r
.We

16
then define three statistics, pdf (potential density function), Mass, and C, which are useful for describing the
intensity features of the divided histogram segment, namely,
(7)
(8)
(9)
In HAF, the values of Mass and C obtained from the histogram are used as initial values for the parameters a and
c, respectively.
(10)
(11)
(12)
In particular, Mass corresponds to the support length for a fuzzy set. It would be expected that the
support length of Md would be longer than that of Dk and Md because more pixels are located at Md for images
generally encountered. After obtaining the initial parameters, we then apply the conservation in histogram
potential to optimize the membership functions. After tuning is done, the membership function is completely
specified because in order to depict the bell-shaped function in Eq. (1), we only need to know the values of
parameters a, b, and c.
V. EXPERIMENTS
Based on the use of a Flower image as test input, three experiments are presented which explored the
characteristics of HAF. Experiment 1 compared the performance of HAF with MF [1] and WFM. In Experiment
2, we exploited the generalization capability and adaptive property of HAF [2] by using three images having
similar histogram statistics. As a measure of the objective improvement obtained using the restoration techniques
discussed here, we refer to both the input normalized mean square error (NMSEi), and the output summed mean
square error (NMSEo), given by (13) and (14) .
(13)
(14)
Where Tp is the total number of pixels in the image, Xp represents the pixel intensities of the original
(uncorrupted) image, X'p represents the corrupted input pixels intensities, and Yp represents the output intensities
in the filtered image using HAF. Note that 0≤Xp,Yp,X'p≤1. The peak signal-to-noise ratio (PSNR) for a 256 x 256,
8-bits/pixel images is simply written as
(15)
NMSE=NMSEo or NMSEi
Finally, the improvement in PSNR (IPSNR) can be expressed by

17
1. Experiment 1
Based on the results obtained using the Flower
image (256 x 256) as test image, its histogram and corrupted
image using salt and pepper noise, HAF output is shown in
Fig. 1 (a)-(d)
The performance difference in removing noise using MF,
WFM And HAF can be clearly seen by comparing NMSEo
and PSNR between HAF, WFM, and MF are given in Table I
, respectively. Clearly, HAF outperforms other filters for
NP=20% all the way up to 90%. Note that MF performs
better than WFM and MF performs better than HAF only
when NP≤20%.
(a) (b)
(c) (d)
(e) (f) TABLE I. NMSEo and PSNR VALUES
Figure 1. (a) original image , (b) Histogram of Flower ,(c) Corrupted by S&P noise , (d) Median Filter(MF)
output ,(e) Weighted Fuzzy Median Filter(WFM) output , (f) output of Histogram Adaptive Filter(HAF)
2. Experiment 2
Next, we used the HAF, on pictures that had similar histogram statistics. To find out, we used two
other ngc4024 pictures denoted as ngc4024s and ngc4024m, respectively as shown if Fig. 2 (a)-(c)
N
P
HAF WFM MF
NM
SEo
PS
NR
NM
SEo
PS
NR
NM
SEo
PS
NR
10
%
0.00
56
22.
51
0.00
65
21.
87
0.00
30
25.
22
20
%
0.00
63
22.
00
0.00
93
20.
31
0.00
41
23.
87
30
%
0.00
74
21.
30
0.01
53
18.
15
0.00
85
20.
70
40
%
0.00
83
20.
80
0.02
61
15.
83
0.01
85
17.
32
50
%
0.00
99
20.
04
0.04
40
13.
56
0.03
66
14.
36
60
%
0.01
30
18.
86
0.07
34
11.
34
0.06
79
11.
68
70
%
0.02
04
16.
90
0.11
74
9.3
0
0.11
46
9.4
0
80
%
0.03
59
14.
44
0.17
80
7.4
9
0.17
70
7.5
2
90
%
0.07
37
11.
32
0.24
62
6.0
8
0.24
74
6.0
6

18
Figure 1. (a) ngc4024 , (b) ngc4024s ,(c) ngc4024m
NMSE and PSNR results for ngc4024 obtained using HMF are shown in Table II. NMSE and PSNR
results for ngc4024s obtained using HMF estimated by using ngc4024 are shown in Table III. Similarly, results
for ngc4024m obtained using HMF obtained by using ngc4024 are shown in Table IV.
By comparing Tables II,III and IV, we can see that ngc4024s performs better than both ngc4024m and
ngc4024 from NP=0.1 to NP=0.9. Thus, the results from this experiment confirm that HMF calculated from an
arbitrary similar image can be used to restore other images having histogram statistics similar to that of the
image. More significantly, the results from Experiment 2 empirically justify use of the HMF.
TABLE II .Comparisons of NMSEo and PSNR using ngc4024
TABLE III. Comparisons of
NMSEo and PSNR using ngc4024s
TABLE IV. Comparisons of NMSEo and PSNR using ngc4024m
NP MF HAF
NMSEo PSNR NMSEo PSNR
10% 0.000453 33.43 0.0016 27.95
20% 0.0010 30.00 0.0016 27.95
30% 0.0051 22.92 0.0017 27.69
50% 0.0378 14.22 0.0020 26.98
70% 0.1337 8.73 0.0058 22.36
90% 0.2952 5.29 0.0369 14.32
VI. CONCLUSION
A novel adaptive fuzzy filter (HAF), which uses fuzzy spatial filtering optimized via image statistics
rather than a priori knowledge of specific image data, has been presented. Instead of using randomly assumed
membership functions, an effective algorithm based on input histogram statistics has been proposed to obtain a
set of well-conditioned membership functions (i.e., HMF).
For images corrupted by impulsive noise, HAF outperforms MF and WFM filters for the range of
impulsive noise probability. Like MF, HAF shows the ability to remove impulsive noise while preserving edge
sharpness. We have carried out two experiments to illustrate the effectiveness o f HMF and to characterize the
restoration power of HAF. In particular, these extensive results verify that there exists a correlation between input
histogram statistics and fuzzy membership functions. In this paper, this relationship has been proven useful in
(1) Deriving HMF for HAF to achieve near-optimal noise
Suppression power and
(2) Exploiting the generalization capability and adaptive
Property of HAF.
NP MF HAF
10% 0.000623 32.05 0.0171 17.67
20% 0.0012 29.20 0.0251 16.00
30% 0.0049 23.09 0.0308 15.11
50% 0.0392 14.06 0.0357 14.47
70% 0.1262 8.98 0.0501 13.00
90% 0.2770 5.57 0.1106 9.56
NP MF HAF
10% 0.000392 34.05 0.0014 28.53
20% 0.000955 30.19 0.0014 28.53
30% 0.0051 22.92 0.0015 28.23
50% 0.0375 14.25 0.0017 27.69
70% 0.1342 8.722 0.0049 23.09
90% 0.2964 5.28 0.0305 15.15

19
For the later, we have empirically shown that images with similar statistics (in fact, pictures that do not
necessarily look similar) can be successfully restored by the HMF inferred from an arbitrary image chosen
from these similar images. We believe that the generalization capability and adaptive property are useful for
making HAF applicable to video transmission where successive image frames must have similar histograms.
REFERENCES
[1]. G. Arce,R. Foster, “Detail preserving ranked order based filter for image processing,” ASSP-37, pp. 83–98, 1989. (references)
[2]. R.Gonzalez,R.E.Woods,Addison-Wesley,Newyork,NY ,U.S.A ,1992.
[3]. C.S.Lee , Y.H.Kuo , and P.T.Yu”Weighted Fuzzy mean filters for image processing”157-164,1997.
[4]. J.Bezdek”Fuzzy Models:What are they and why?”1-5,1993
[5]. H.Kong and L.Guan, “A noise-exclusive adaptive filtering framework for removing impulse noise in digital image”. IEEE Trans.
Circuits and Systems, 422-428,1998.
[6]. P.Yu and C.S.Lee”Adaptive Fuzzy Median Filter”International symposium on artificial neural networks,25-34,1993.
[7]. E.Rumelhart, G.E.Hinton and R.J.Williams,”Learning internal representation by error propagations”,Parallel distributed
processing,1986.

Research Inventy : International Journal of Engineering and Science

More Related Content

What's hot (20)

Viewers also liked (8)

Similar to Research Inventy : International Journal of Engineering and Science (20)

More from inventy (20)

Recently uploaded (20)

Research Inventy : International Journal of Engineering and Science