SIViP (2012) 6:579–595

DOI 10.1007/s11760-010-0186-4

ORIGINAL PAPER

A new fuzzy-based decision algorithm for high-density impulse

noise removal
Madhu S. Nair · G. Raju

Received: 29 March 2010 / Revised: 23 July 2010 / Accepted: 24 September 2010 / Published online: 16 October 2010
© Springer-Verlag London Limited 2010

Abstract This paper proposes a new efficient fuzzy-based FBDA produces better results in terms of both quantitative
decision algorithm (FBDA) for the restoration of images that measures such as PSNR, SSIM, IEF and qualitative measures
are corrupted with high density of impulse noises. FBDA is such as Image Quality Index (IQI).
a fuzzy-based switching median filter in which the filtering
is applied only to corrupted pixels in the image while the Keywords Fuzzy · Decision-based · Impulse noise ·
uncorrupted pixels are left unchanged. The proposed algo- Median filter · Salt-and-pepper noise
rithm computes the difference measure for each pixel based
on the central pixel (corrupted pixel) in a selected window 1 Introduction
and then calculates the membership value for each pixel
based on the highest difference. The algorithm then elimi- Digital image could be contaminated by impulse noise dur-
nates those pixels from the window with very high and very ing image acquisition or transmission. Two common types of
low membership values, which might represent the impulse impulse noise are the salt-and-pepper noise and the random-
noises. Median filter is then applied to the remaining pix- valued noise. For images corrupted by salt-and-pepper noise
els in the window to get the restored value for the current (respectively, random-valued noise), the noisy pixels can take
pixel position. The proposed algorithm produces excellent only the maximum and the minimum values (respectively,
results compared to conventional method such as standard any random value) in the dynamic range. Thus, it could
median filter (SMF) as well as some advanced techniques severely degrade the image quality and cause some loss of
such as adaptive median filters (AMF), efficient decision- information. Various methods have been proposed for the
based algorithm (EDBA), improved efficient decision-based restoration of images corrupted by impulse noise, and it is
algorithm (IDBA) and boundary discriminative noise detec- well known that linear filters could produce serious image
tion (BDND) switching median filter. The efficiency of the blurring even in low noise density [5].
proposed algorithm is evaluated using different standard Consequently, nonlinear filters have been widely exploited
images. From experimental analysis, it has been found that due to their much improved filtering performance in terms
of impulse noise attenuation. The median filter was once
M. S. Nair
School of Computer Sciences, Mahatma Gandhi University,
the most popular nonlinear filter for removing impulse noise
Kottayam, Kerala, India because of its good denoising power and computational effi-
ciency [2,7]. The median filter replaces the central value of an
M. S. Nair (B) M-by-N neighborhood with its median value. One of the most
Department of Computer Science, University of Kerala,
Kariavattom, Thiruvananthapuram, India
popular and robust nonlinear filters is the standard median fil-
e-mail: [email protected] ter (SMF) [5,8], which exploits the rank-order information
of pixel intensities within a filtering window and replaces the
G. Raju center pixel with the median value. Due to its effectiveness
School of Information Science and Technology,
Kannur University, Mangattuparamba,
in noise suppression and simplicity in implementation, vari-
Kannur 670 567, Kerala, India ous modifications of the SMF have been introduced, such as
e-mail: [email protected] the weighted median (WM) [2] filter and the center weighted

123
580 SIViP (2012) 6:579–595

median (CWM) [10] filter. The major drawback of the SMF and loss of image edge details. The reason for this loss in
is that the filter is effective only for low noise densities, and image edge details is due to the fact that these filters trim the
additionally, exhibits blurring if the window size is large. extreme values even if they are not impulse values [1]. Also
This leads to insufficient noise suppression if the window the mean operation of the filter smoothens the image which
size is small [14]. will cause the edge information to be lost.
Conventional median filtering [2,5,10,20] approaches Srinivasan and Ebenezer [16] proposed an efficient deci-
apply the median operation to each pixel unconditionally, sion-based algorithm (EDBA) in which the corrupted pix-
that is, without considering whether it is uncorrupted or els are replaced by either the median pixel or neighborhood
corrupted. As a result, the image details contributed from pixel by using a fixed window size of 3 × 3 resulting in
the corrupted pixels cause image quality degradation. An lower processing time and good edge preservation. Although
intuitive solution to overcome this problem is to imple- EDBA filter [16] showed promising results, a smooth transi-
ment an impulse-noise detection mechanism prior to filtering. tion between the pixels is lost leading to degradation in the
For this, switching median filters [4,13,18,21] can be used, visual quality of the image in the form of line artifacts, since
which gives significant performance improvement compared it only considers the left neighborhood from the last pro-
to any other existing advanced methods for impulse noise cessed value. To overcome this problem, Madhu et al. [11,12]
removal. In switching median filters, a noise detection mech- proposed an improved decision-based algorithm (IDBA) in
anism has been incorporated so that only those pixels identi- which corrupted pixels can be replaced either by the median
fied as “corrupted” would undergo the filtering process, while pixel or, by the mean of processed pixels in the neighborhood,
those identified as “uncorrupted” would remain intact. which results in a smooth transition between the pixels with
Nonlinear filters such as adaptive median filter (AMF) [6] edge preservation and better visual quality. The drawback of
can be used for discriminating corrupted and uncorrupted this method is that in the case of high-density impulse noise,
pixels and then apply the filtering technique. Noisy pixels will the fixed window size of 3 × 3 will result in image quality
be replaced by the median value, and uncorrupted pixels will degradation due to the presence of corrupted pixels in the
be left unchanged. AMF performs well at low noise densi- neighborhood.
ties since the corrupted pixels that are replaced by the median To address high noise density, a noise adaptive soft-
values are very few. At higher noise densities, window size switching median (NASM) filter was proposed in [4], which
has to be increased to get better noise removal which will consists of a three-level hierarchical soft-switching noise
lead to less correlation between corrupted pixel values and detection process. The NASM achieves a fairly robust per-
replaced median pixel values. In decision-based or switching formance in removing impulse noise, while preserving sig-
median filter, the decision is based on a pre-defined thresh- nal details across a wide range of noise densities. However,
old value. The major drawback of this method is that defining the quality of the recovered image becomes significantly
a robust decision measure is difficult. These filters will not degraded when noise density is greater than 50%. To over-
take into account the local features as a result of which edge come performance degradation at higher noise density, a new
details may not be recovered satisfactorily, especially when efficient method called BDND [13] has been introduced and
the noise is high. it has shown better results. But at high noise density, BDND
Chan et al. [3] proposed an algorithm to overcome this shows higher misdetection and false alarm rate (at random
problem, which consists of two stages. The first stage is to noise). Consequently, it could not preserve the edge details
classify the corrupted and uncorrupted pixels by using AMF, of the recovered image and the quality of the restored image
and in the second stage, regularization method is applied to is reduced. The main drawback of switching median filters
the corrupted pixels to preserve edges and suppress noise. like BDND is that in the case of high-density impulse noises,
The drawback of this method is that for high impulse noise, there is still a chance for good representation of the corrupted
it requires large window size of 39 × 39 and additionally pixels in the selected window to take part in the filtering pro-
requires complex circuitry for the implementation and deter- cess, which may lead to the degradation of image quality.
mination of smoothing factor β to get good results [3]. Another drawback of BDND switching filter is that it first
The alpha-trimmed mean filter (ATMF) [1] and alpha- uses a window size of 21 × 21 to detect whether a pixel is
trimmed-midpoint filter (ATMP) [17] are other types of non- corrupted or not and again a second iteration is performed
linear filters discussed in the literature. These filters are also on a reduced window size of 3 × 3 with same set of steps
used to remove the impulse noise in which the parameter α, to reduce the misclassification of pixels. Actual filtering pro-
called the trimming factor, controls the number of values that cess begins after the two levels of iterations. Consequently,
are trimmed. It can be seen that as the value of the trimming BDND algorithm consumes a lot of time.
factor α increases, the ability of the filter to remove impulse In order to overcome the drawbacks of the above filters,
noise is further increased. However, when the noise density is we propose a new fuzzy-based decision algorithm (FBDA)
as high as 50% and above, there is insufficient noise removal for removing impulse noise at a wide range of noise densities,

123
SIViP (2012) 6:579–595 581

especially for high impulse noise. FBDA is an improved Noise Model 1: Noise is modeled as salt-and-pepper
fuzzy-based switching median filter in which the filtering impulse noise, where pixels are randomly corrupted by two
is applied only to corrupted pixels in the image while the fixed extreme values, 0 and 255 (for 8-bit monochrome
uncorrupted pixels are left unchanged. What makes FBDA image), generated with the same probability. That is, for each
different from other switching median filters such as BDND image pixel at location (i, j) with intensity value Si j , the cor-
is that during the time of filtering process FBDA selects only responding pixel of the noisy image will be xi j , in which the
uncorrupted pixels in the selected window based on a fuzzy probability density function of xi j is
distance membership value. Thus, the advantage of FBDA ⎧p
is that it has both the noise detection power and the power ⎨ 2,
⎪ for x = 0
of eliminating corrupted pixels during the filtering process. f (x) = 1 − p, for x = Si j
⎪
⎩p
2, for x = 255
Another advantage of FBDA is that it uses only one level of
iteration to detect whether a pixel is corrupted or not and it
uses a fixed window size of 3 × 3 or 5 × 5 (based on the where p is the noise density.
noise density) for both the noise detection and the filtering Noise Model 2: Model 2 is similar to Model 1, except that
process. The first step in FBDA is to classify a pixel as either each pixel might be corrupted by either “pepper” noise (i.e.,
corrupted or uncorrupted. For a pixel identified as a corrupted 0) or “salt” noise with unequal probabilities. That is
pixel, FBDA selects a window that consists of neighborhood ⎧
⎪
⎨ p1, for x = 0
pixels. It then computes the difference measure for each pixel
in that selected window based on the central pixel (the cor- f (x) = 1 − p, for x = Si j
⎪
⎩
rupted pixel) and then calculates the membership value for p2, for x = 255
each based on the highest difference. FBDA eliminates those
where p = p1 + p2 is the noise density and p1 = p2.
pixels from the window with very high (close to 1) and very
Noise Model 3: Instead of two fixed values, impulse noise
low (close to 0) membership values, since they may repre-
could be more realistically modeled by two fixed ranges that
sent the impulse noises. Median filter is then applied to the
appear at both ends with a length of m each, respectively. For
remaining pixels in the window to get the new pixel value
example, if m is 10, noise will equal likely be any values in
for the current pixel position.
the range of either [0, 9] or [246, 255]. That is
Proposed FBDA method is a switching median filter which
⎧ p
preserves the edge details in the case of high-density impulse
⎨ 2m ,
⎪ for 0 ≤ x < m
noise situation compared to nonswitching ATMF. In contrast
f (x) = 1 − p, for x = Si j
to ATMF or ATMP, FBDA eliminates the corrupted pixels ⎪
⎩ p
in the window from the filtering process using the fuzzy dis- 2m , for 255 − m < x ≤ 255
tance membership value calculated from the central pixel, where p is the noise density.
which consequently leads to better results. FBDA is thus Noise Model 4: Model 4 is similar to Model 3, except that
more adaptive and efficient compared to ATMF or ATMP. the densities of low-intensity impulse noise and high-inten-
The efficiency of the proposed algorithm is tested using sity impulse noise are unequal. That is
standard images after applying four different noise mod-
⎧ p1
els. The different noise models used are explained in the ⎨ m, for 0 ≤ x < m
Sect. 2. From experimental analysis, it has been found that f (x) = 1 − p, for x = Si j
⎩ p2
m , for 255 − m < x ≤ 255
SMF, AMF, EDBA, IDBA and BDND methods are not giv-
ing convincing results for noise models 2, 3 and 4. But
FBDA gives better visual results for all the different noise where p = p1 + p2 is the noise density and p1 = p2.
models compared to aforementioned filters. It has also been
proved through experimental analysis that FBDA produces
better results in terms of quantitative measures such as PSNR, 3 Fuzzy-based decision algorithm
SSIM, IEF and in terms of qualitative measures such as Image
Quality Index (IQI), compared to other filters. In order to overcome the drawbacks of conventional filters as
well as the advanced filters mentioned in the Sect. 1, we pro-
pose here a new fuzzy-based decision algorithm (FBDA) for
2 Noise models high-density impulse noise removal. FBDA is a novel fuzzy-
based switching median filter that selects only uncorrupted
Four impulse noise models are used for examining the perfor- pixels in the selected window for the filtering process. The
mance of the proposed FBDA filter. Each model is described selection of uncorrupted pixels in a window is made based on
in detail below: a fuzzy distance membership value. Thus, the power of both

123
582 SIViP (2012) 6:579–595

noise detection and corrupted pixel elimination during the fil- Membership
degree
tering process has been incorporated into FBDA. It has been
found from experimental analysis that compared to advanced
filters such as AMF, EDBA, IDBA and BDND, FBDA gives 1
better results for all the different noise models discussed in
Sect. 2.
BDND filter is one of the latest and advanced switching
median filters. A main drawback of BDND switching filter
is that in the first level of iteration, it uses a window size of
21×21 to detect whether a pixel is corrupted or not. The same 0
P Difference
detection process is repeated in the second level of iteration
using a reduced window size of 3 × 3 to avoid the misclassi- Fig. 1 Membership function small (μ S )
fication of pixels. Only after these two levels of iteration, the
filtering process can be applied to the corrupted pixels. This
time-consuming process of BDND makes it less attractive. for each pixel in the selected window based on the central
The proposed FBDA filter is much faster than BDND fil- pixel (the corrupted pixel) and then calculates the member-
ter since it uses only one level of iteration to detect whether ship value μ S for each pixel based on the maximum differ-
a pixel is corrupted or not and it uses a fixed window size ence β in that window.
of 3 × 3 or 5 × 5 (based on the noise density) for both the The difference measure for each pixel in the selected win-
detection and the filtering phase. dow based on the current pixel Ci, j can be calculated as
The first step in FBDA is to identify whether a pixel is follows:
 
corrupted or not. For this, FBDA uses a simple decision rule, i, j,k,l = Ci, j − Ci+k, j+l 
which is being used in conventional filter such as conservative
smoothing filter [2,5]. FBDA procedure first finds the mini- where Ci+k, j+1 represents the neighborhood pixels in the
mum and maximum intensity values of all the pixels within a selected window. If the window size used for FBDA is 3 × 3,
windowed region around the pixel in question. If the intensity then k, l ∈ {−1, 0, 1}. On the other hand, if the window size
of the central pixel lies within the intensity range spread of is 5×5 then k, l ∈ {−2, −1, 0, 1, 2}. Once  has been calcu-
its neighbors, it is passed on to the output image unchanged. lated for each pixel in the window, the maximum difference
Otherwise, the central pixel is detected as corrupted pixel. value β among them can be computed as follows:
 
Let Ci, j represent the central pixel under consideration and β i, j = Max i, j,k,l
let Wmin and Wmax represent the minimum and maximum k,l

intensity values within a selected window around Ci, j . Then, To compute the value that expresses the degree to which the
FBDA decision rule can be written as follows: difference value is small, we will make use of the fuzzy set
I f Wmin < Ci, j < Wmax , then Ci, j is an uncorrupted small. Membership functions are commonly used to repre-
pixel and its value is left unchanged. Otherwise Ci, j is a sent Fuzzy sets, and from such functions, we can derive the
noisy pixel. corresponding membership degrees [9,15,20]. If the differ-
In other switching filters like BDND, once a pixel has been ence value  has a membership degree one (zero) in the fuzzy
identified as a corrupted pixel, then the filtering is applied to set small, it means that this difference is considered as (not)
that pixel by taking care of all the neighboring pixels in the small for sure. Membership degrees between zero and one
selected window. In most of the switching filters, the window indicate that we do not know for sure if such difference is
size for noisy pixel detection may be different from the win- small or not, so there is some kind of uncertainty. The mem-
dow size used for the filtering. Consequently, a lot of time is bership function small is denoted as μ S , and it is shown in
consumed to find out the neighboring pixels for detection as Fig. 1. An alternative notation for μ S is given below:
well as filtering phase. As mentioned earlier, BDND uses a
window size of 21 × 21 and 3 × 3 (two levels of iteration) for
P−x
, if x ≤ P
μ S (x) = P
the detection phase and a 7 × 7 window size for the filtering 0, if x > P
phase, whereas FBDA uses a fixed window size of 3 × 3 or The membership degree for each calculated difference value
5 × 5 for both the noise detection and the filtering phase, i, j,k,l can be computed as follows:
which makes FBDA much faster than BDND. 
βi, j − i, j,k,l
After a pixel has been identified as a corrupted pixel, μ S(i, j,k,l) =
FBDA uses the same window (used for the detection phase) βi, j
of neighborhood pixels around the central pixel Ci, j for the where βi, j > 0. Here, βi, j is substituted for P and i, j,k,l is
filtering process. FBDA computes the difference measure  substituted for x in μ S (x). FBDA then examines the mem-

123
SIViP (2012) 6:579–595 583

bership degree values (μ S(i, j,k,l) ) of each pixel in the selected much better than other algorithms for noise models 2, 3 and
window and eliminates those pixels from the window with 4, we applied our FBDA algorithm as well as other algo-
very high (close to 1) and very low (close to 0) membership rithms on cameraman image corrupted with 75% of noise,
values, since they might represent the impulse noises. That is, based on models 2, 3 and 4. The performance measure values,
the membership values of the pixels selected for the filtering edge maps and restored image results show the superior per-
process should satisfy the following: formance of FBDA over other algorithms. To evaluate the
performance of the FBDA algorithm on color images, we
T1 < μ S(i, j,k,l) < T2 have applied FBDA and other filters on lena color image of
where T1 ∼ = 0 and T2 ∼= 1 (from experimental analysis, we size 256×256 and on peppers color image of size 512×384,
set T1 ∈ [0, 0.05] and T2 ∈ [0.95, 1]. We know that if which are corrupted with 75% of noise based on model 4.
the central pixel is corrupted, then it may hold either a very For the noise model 1, images are corrupted with salt-and-
low-intensity value or a very high-intensity value. So when pepper noise at different noise densities, such as low noise
the absolute difference between central pixel and neighbor- (30%), medium noise (60%) and high noise (90%). For the
ing pixels is computed, those pixels around the central pixel noise models 2, 3 and 4, images are corrupted with 75% of
holding very high-intensity values or very low-intensity val- noise, where 25% are either pepper (value 0) or low-intensity
ues (which may represent impulse noise) will have either very values in the range 0–9 and 50% are either salt (value 255)
high or very low difference value . As a result, those pixels or high-intensity values in the range 246–255.
will have the membership degrees very close to either 0 or The performance of the restoration process is quantified
1. That is the reason why those threshold values are selected using peak signal-to-noise ratio (PSNR), structured simi-
for eliminating the noisy pixels during the filtering process. larity index (SSIM) and image enhancement factor (IEF),
After the elimination of noisy pixels from the window, the defined as follows.
median filter is applied to the remaining pixels in the window  
to get the restored pixel value for the current pixel position. PSNR = 10 ∗ log10 2552 /MSE

The process of elimination of noisy pixels, or in other words, MSE = [O (m, n) − R (m, n)]2 /(M ∗ N )
the selection of uncorrupted pixels from the selected win- m,n
dow and then the process of applying the median filter on the SSIM = L (O, R) ∗ C (O, R) ∗ S(O, R)
selected pixels, both can be represented mathematically as:  
  L (O, R) = (2μ O μ R + C1 ) / μ2O + μ2R + C1
i, j = (k, l)|T1 < μ S(i, j,k,l) < T2
  C (O, R) = (2σ O σ R + C2 )/(σ O2 + σ R2 + C2 )
Ri, j = Median Ci+ p, j+q
S (O, R) = (σ O R + C3 ) / (σ O σ R + C3 )
where Ri, j is the restored value for the current pixel position C1 = (K 1 ∗ G)2 , C2 = (K 2 ∗ G)2 , C3 = C2 /2
and ( p, q) ∈ i, j .
G = 255; K 1 , K 2  1, (K 1 = 0.001, K 2 = 0.002)
 
 
IEF = [P (m, n) − O (m, n)] 2
[R (m, n)
4 Results and discussion m,n m,n
2 ⎞
The performance of the proposed FBDA filter has been eval- −O(m, n) ⎠
uated quantitatively and qualitatively through experimental
analysis. Standard images such as liftingbody image of size
512 × 512 circuit image of size 272 × 280 and lena image where O is the original Image, R is the restored image, P is
of size 512 × 512 have been used to test the performance the corrupted image, MSE is the mean square error, M × N
of the various algorithms for the noise model 1. To evalu- is the size of the image, L is the luminance comparison, C is
ate the performance of the proposed FBDA algorithm, edge the contrast comparison, S is the structure comparison, μ is
maps of restored lena images are also computed, which has the mean and σ is the standard deviation.
been compared with the edge map of the original lena image. In this paper, we also used a qualitative-based perfor-
The superior performance of the FBDA algorithm over other mance measure named image quality index (IQI) to prove
algorithms discussed in the previous sections has been proved the efficiency of our proposed FBDA algorithm. This uni-
through performance measure graphs. To further convince versal objective image quality index was proposed by Wang
the performance of FBDA algorithm, a particular row (one and Bovik [19], which is easy to calculate and applicable to
dimensional signal) of original, corrupted and restored cam- various image processing applications. Instead of using tradi-
eraman image of size 256 × 256 has been shown in the form tional error summation methods, IQI is designed by modeling
of a graph. In order to prove that the FBDA algorithm works any image distortion as a combination of three factors: loss

123
584 SIViP (2012) 6:579–595

of correlation, luminance distortion and contrast distortion.

IQI j = Corr(Ow , Rw ) ∗ Lum(Ow , Rw ) ∗ Cont(Ow , Rw )

Corr(Ow , Rw ) = σ O R /(σ O σ R ),
 
Lum(Ow , Rw ) = (2μ O μ R ) / μ2O + μ2R
Cont(Ow , Rw ) = (2σ O σ R )/(σ O2 + σ R2 )

IQI is first applied to local regions using a sliding window

approach with size 8 × 8. At the jth step, the local quality
index IQI j is computed within the sliding window using the
formula given above. Ow and Rw represent the sliding win-
dow of original and restored images, respectively. If there
are a total of M steps, then the overall image quality index

is given by IQI = (1/M) j IQI j , where j varies from 1 to
M. The dynamic range of IQI is [−1, 1], and the best value Fig. 2 Column a represents original liftingbody, circuit and lena
1 is achieved if and only if restored image R is equal to the images. Columns b, c and d represent liftingbody, circuit and lena
images corrupted with 30, 60 and 90% noise, respectively, based on
original image O. In addition to the IQI value, the image noise model 1
quality map has also been generated to evaluate the perfor-
mance of the different algorithms. Brighter image quality
map (closer to 1) indicates that the restored image is closer Figure 2 shows the original liftingbody, circuit and lena
to the original image, and darker image quality map indi- images and the corresponding corrupted images with 30,
cates that the restored image is more distant from the original 60 and 90% impulse noise based on the noise model 1.
image. Tables 1, 2, 3, 4, 5, 6, 7, 8, 9 show the performance parame-

Table 1 PSNR, SSIM, IEF and IQI values for various filters applied on liftingbody image at low noise density (30%)
Performance measures SMF (5 × 5) AMF (7 × 7) EDBA (3 × 3) IDBA (3 × 3) BDND (21 × 21)(7 × 7) FBDA (3 × 3)

PSNR (dB) 28.12 31.99 33.63 36.43 37.37 38.38

SSIM 0.9507 0.9798 0.9859 0.9926 0.9940 0.9953
IEF 52.03 127.25 185.16 352.63 436.94 553.24
IQI 0.4457 0.8157 0.6160 0.8614 0.8871 0.8740

Table 2 PSNR, SSIM, IEF and IQI values for various filters applied on liftingbody image at high noise density (60%)
Performance measures SMF (7 × 7) AMF (13 × 13) EDBA (3 × 3) IDBA (3 × 3) BDND (21 × 21)(7 × 7) FBDA (5 × 5)

PSNR (dB) 22.34 26.29 29.58 30.63 31.62 34.15

SSIM 0.8321 0.9278 0.9643 0.9718 0.9771 0.9875
IEF 27.64 68.67 146.36 186.46 233.01 419.33
IQI 0.2862 0.6323 0.4876 0.6205 0.6751 0.7157

Table 3 PSNR, SSIM, IEF and IQI values for various filters applied on liftingbody image at high noise density (90%)
Performance measures SMF (23 × 23) AMF (39 × 39) EDBA (3 × 3) IDBA (3 × 3) BDND (21 × 21)(7 × 7) FBDA (5 × 5)

PSNR (dB) 14.77 20.57 23.90 26.06 27.73 28.29

SSIM 0.4066 0.7611 0.8680 0.9170 0.9438 0.9514
IEF 7.23 27.52 59.27 97.46 143.03 162.61
IQI 0.0748 0.2646 0.2741 0.2797 0.3393 0.3754

123
SIViP (2012) 6:579–595 585

Table 4 PSNR, SSIM, IEF and IQI values for various filters applied on circuit image at low noise density (30%)
Performance measures SMF (5 × 5) AMF (7 × 7) EDBA (3 × 3) IDBA (3 × 3) BDND (21 × 21)(7 × 7) FBDA (3 × 3)

PSNR (dB) 27.73 31.81 31.03 33.40 36.78 37.07

SSIM 0.9747 0.9902 0.9882 0.9932 0.9968 0.9971
IEF 56.51 144.73 120.85 208.60 452.94 497.70
IQI 0.7297 0.9260 0.8539 0.9203 0.9543 0.9500

Table 5 PSNR, SSIM, IEF and IQI values for various filters applied on circuit image at high noise density (60%)
Performance measures SMF (7 × 7) AMF (13 × 13) EDBA (3 × 3) IDBA (3 × 3) BDND (21 × 21)(7 × 7) FBDA (5 × 5)

PSNR (dB) 22.31 26.36 26.19 26.91 28.88 31.87

SSIM 0.9141 0.9659 0.9642 0.9695 0.9805 0.9903
IEF 32.99 83.79 80.64 95.06 150.66 297.87
IQI 0.4561 0.7860 0.7023 0.7424 0.7512 0.8759

Table 6 PSNR, SSIM, IEF and IQI values for various filters applied on circuit image at high noise density (90%)
Performance measures SMF (23 × 23) AMF (39 × 39) EDBA (3 × 3) IDBA (3 × 3) BDND (21 × 21)(7 × 7) FBDA (5 × 5)

PSNR (dB) 14.70 19.80 19.53 22.38 24.47 25.19

SSIM 0.5530 0.8488 0.8338 0.9096 0.9459 0.9545
IEF 8.62 27.86 26.20 50.46 81.33 96.35
IQI 0.0734 0.3798 0.3617 0.4199 0.4314 0.5755

Table 7 PSNR, SSIM, IEF and IQI values for various filters applied on lena image at low noise density (30%)
Performance measures SMF (5 × 5) AMF (7 × 7) EDBA (3 × 3) IDBA (3 × 3) BDND (21 × 21) (7 × 7) FBDA (3 × 3)

PSNR (dB) 27.01 30.93 30.70 33.76 35.74 35.77

SSIM 0.9718 0.9886 0.9879 0.9940 0.9962 0.9963
IEF 43.18 106.47 100.84 204.12 321.54 325.04
IQI 0.6240 0.8597 0.7106 0.9003 0.9176 0.9102

Table 8 PSNR, SSIM, IEF and IQI values for various filters applied on lena image at high noise density (60%)
Performance measures SMF (7 × 7) AMF (13 × 13) EDBA (3 × 3) IDBA (3 × 3) BDND (21 × 21) (7 × 7) FBDA (5 × 5)

PSNR (dB) 22.10 25.84 26.15 27.54 29.27 31.49

SSIM 0.9155 0.9636 0.9655 0.9749 0.9831 0.9899
IEF 27.67 65.50 70.24 96.87 145.27 240.21
IQI 0.4682 0.7149 0.5718 0.7122 0.7557 0.7879

Table 9 PSNR, SSIM, IEF and IQI values for various filters applied on lena image at high noise density (90%)
Performance measures SMF (23 × 23) AMF (39 × 39) EDBA (3 × 3) IDBA (3 × 3) BDND (21 × 21)(7 × 7) FBDA (5 × 5)

PSNR (dB) 14.95 19.92 18.81 22.40 25.02 25.46

SSIM 0.6388 0.8637 0.8132 0.9152 0.9546 0.9593
IEF 8.01 25.13 19.45 44.50 81.47 90.14
IQI 0.1405 0.3637 0.2152 0.3674 0.4847 0.5223

123
586 SIViP (2012) 6:579–595

Fig. 3 Column a represents restored images of liftingbody corrupted

Fig. 4 Column a represents the image quality map of the restored
image (30%). Column b represents restored images of circuit corrupted
images of liftingbody shown in the first column of Fig. 3. Column b
image 30%. Column c represents restored images of lena corrupted
represents the image quality map of the restored images of circuit shown
image (30%). Rows 1, 2, 3, 4, 5 and 6 represent the restored images
in the second column of Fig. 3. Column c represents the image quality
after applying the SMF, AMF, EDBA, IDBA, BDND and FBDA filters,
map of the restored images of lena shown in the third column of Fig. 3
respectively

ter values such as PSNR, SSIM, IEF and IQI obtained after 8 show their corresponding image quality maps, and it can be
applying the various filters on the above-mentioned corrupted seen that the quality map of FBDA is brighter compared to
images. From the quantitative values shown in the tables, it is others, for high-density impulse noises such as 60 and 90%.
very clear that FBDA algorithm outperforms all other noise Tables 10, 11, 12, 13 show the quantitative values of the
removal filters. Figures 3, 5 and 7 show the restored images different performance measures obtained after applying the
obtained after applying the various filters, and Figs. 4, 6 and different filters on the corrupted lena image at different noise

123
SIViP (2012) 6:579–595 587

Fig. 6 Column a represents the image quality map of the restored

Fig. 5 Column a represents restored images of liftingbody corrupted images of liftingbody shown in the first column of Fig. 5. Column b
image (60%). Column b represents restored images of circuit corrupted represents the image quality map of the restored images of circuit shown
image (60%). Column c represents restored images of lena corrupted in the second column of Fig. 5. Column c represents the image quality
image (60%). Rows 1, 2, 3, 4, 5 and 6 represent the restored images map of the restored images of lena shown in the third column of Fig. 5
after applying the SMF, AMF, EDBA, IDBA, BDND and FBDA filters,
respectively

Figure 13 shows the edge map of the original lena image

densities based on model 1. From the PSNR, SSIM, IEF and the edge maps of lena images corrupted with 30, 60 and
and IQI performance graphs of Figs. 9, 10, 11 and 12, it 90% of impulse noise, after applying the canny filter. Fig-
can be seen that for high-density noise levels, FBDA gives ure 14 shows the edge map of the restored lena images after
higher values for the performance measures, thereby prov- applying the FBDA and other filters on the corrupted lena
ing its efficiency in removing high-density impulse noises. images.

123
588 SIViP (2012) 6:579–595

Fig. 8 Column a represents the image quality map of the restored

images of liftingbody shown in the first column of Fig. 7. Column b
represents the image quality map of the restored images of circuit shown
in the second column of Fig. 7. Column c represents the image quality
Fig. 7 Column a represents restored images of liftingbody corrupted map of the restored images of lena shown in the third column of Fig. 7
image (90%). Column b represents restored images of circuit corrupted
image (90%). Column c represents restored images of lena corrupted
image (90%). Rows 1, 2, 3, 4, 5 and 6 represent restored images after
applying the SMF, AMF, EDBA, IDBA, BDND and FBDA filters,
respectively rupted lena image, respectively. Similarly Fig. 15j–o rep-
resents 163rd row of the restored cameraman image after
applying filters on the 85% corrupted lena image. From the
Figure 15 shows the 163rd row of the original camera- graphs, it is very evident that FBDA algorithm restores the
man image and the corrupted cameraman image with 60 and corrupted image that is very close to the original image.
85% impulse noise. Figure 15d–i represents 163rd row of The resemblance of the graph shown in Fig. 15a with the
the restored cameraman image after applying SMF, AMF, graphs shown in Fig. 15i and o proves the superior perfor-
EDBA, IDBA, BDND and FBDA filters on the 60% cor- mance of FBDA filter over other filters.

123
SIViP (2012) 6:579–595 589

Table 10 PSNR values for various filters applied on lena image at Table 13 IQI values for various filters applied on lena image at dif-
different noise density levels ferent noise density levels
Noise (%) PSNR Noise (%) IQI

SMF AMF EDBA IDBA BDND FBDA SMF AMF EDBA IDBA BDND FBDA

10 33.11 37.35 34.52 38.11 41.26 38.31 10 0.7559 0.8951 0.8008 0.9492 0.9581 0.9649
20 28.81 33.67 32.19 35.82 38.63 36.85 20 0.6392 0.8855 0.7508 0.9339 0.9495 0.9383
30 27.01 30.93 30.71 33.76 35.74 35.77 30 0.6241 0.8597 0.7106 0.9003 0.9176 0.9102
40 25.61 29.06 29.12 31.58 33.32 34.35 40 0.5936 0.8217 0.6674 0.8537 0.8759 0.8769
50 23.51 27.29 27.63 29.57 31.28 32.96 50 0.5162 0.7751 0.6248 0.7919 0.8235 0.8343
60 22.11 25.84 26.15 27.54 29.27 31.49 60 0.4682 0.7149 0.5718 0.7122 0.7557 0.7879
70 20.08 24.18 24.39 25.96 27.88 29.58 70 0.3719 0.6358 0.4972 0.6137 0.6839 0.7285
80 17.97 22.39 22.24 24.51 26.78 28.01 80 0.2697 0.5292 0.3931 0.5085 0.6036 0.6497
90 14.95 19.92 18.81 22.41 25.02 25.46 90 0.1405 0.3637 0.2152 0.3674 0.4847 0.5223

Table 11 SSIM values for various filters applied on lena image at

different noise density levels
Noise (%) SSIM

SMF AMF EDBA IDBA BDND FBDA

10 0.9931 0.9974 0.9951 0.9978 0.9989 0.9979

20 0.9812 0.9939 0.9914 0.9963 0.9981 0.9971
30 0.9718 0.9886 0.9879 0.9941 0.9962 0.9963
40 0.9614 0.9825 0.9825 0.9901 0.9933 0.9948
50 0.9381 0.9738 0.9755 0.9843 0.9893 0.9928
60 0.9155 0.9636 0.9655 0.9749 0.9831 0.9899
70 0.8684 0.9471 0.9483 0.9638 0.9766 0.9842
80 0.7939 0.9209 0.9154 0.9491 0.9697 0.9774
90 0.6388 0.8637 0.8132 0.9152 0.9546 0.9593

Table 12 IEF values for various filters applied on lena image at dif- Fig. 9 Graph representing the PSNR values obtained after the apply-
ferent noise density levels ing the different filters on the corrupted lena image at difference noise
densities
Noise (%) IEF

SMF AMF EDBA IDBA BDND FBDA

In noise model 3, the image is corrupted with 75% of
10 58.55 155.39 80.88 184.6 384.9 193.73 impulse noise, by taking value of m as 10 to indicate the two
20 43.49 133.34 94.92 219.08 417.31 277.35 range of impulse noise such as 0–9 and 246–255. In noise
30 43.18 106.47 100.84 204.12 321.54 325.04 model 4, the image is corrupted with 75% of impulse noise,
40 41.51 91.73 92.85 163.88 246.06 310.04 in which again 25% is of pepper noise and 50% is of salt noise
50 32.04 76.45 82.74 129.37 192.01 281.23 and at the same time value of m is taken as 10 to indicate the
60 27.67 65.51 70.24 96.87 145.27 240.21 two range of impulse noise such as 0–9 and 246–255. Edge
70 20.41 52.46 55.06 78.89 123.37 181.62 maps as well as the 163rd row of the original and corrupted
80 14.31 39.62 38.31 64.59 108.91 144.51 cameraman image are also shown in Fig. 16. Figures 17, 18
90 8.01 25.13 19.45 44.51 81.47 90.14 and 19 show the restored images of the corrupted cameraman
image based on noise models 2, 3 and 4, respectively.
From the visual results of the restored images as well as
Figure 16 shows the original cameraman image and its from their edge maps, image quality maps and 163rd row
corrupted versions based on the noise model 2, 3 and 4. In graphs, it has been proved that FBDA filter is well suited
noise model 2, the image is corrupted with 75% of impulse for noise models 2, 3 and 4 compared to other filters. The
noise, in which 25% is of pepper noise and 50% is of salt quantitative values for the different performance measures
noise. have been computed for the restored images and shown in

123
590 SIViP (2012) 6:579–595

Fig. 10 Graph representing the SSIM values obtained after the apply-
ing the different filters on the corrupted lena image at difference noise Fig. 12 Graph representing the IQI values obtained after the apply-
densities ing the different filters on the corrupted lena image at difference noise
densities

the Tables 14, 15 and 16. The values indicate that FBDA filter
outperforms all other filters for the noise models 2, 3 and 4.
The FBDA switching median filtering scheme can be
extended to denoise corrupted color images via the scalar
median filtering approach. The scalar approach treats each
color component as an independent entity; that is, the same
filtering scheme will be applied to R-, G- and B-planes inde-
pendently, as if each plane is a separate monochrome image.
The filtered R-, G- and B-planes will be then combined to
form the recovered color image.
Figure 20 shows the original lena and peppers color
images and their corrupted versions. Images are corrupted
with 75% of impulse noise based on model 4 (m = 10),
in which 25% is of pepper noise and 50% is of salt noise.
Fig. 11 Graph representing the IEF values obtained after the apply- The restored images of the corrupted lena and peppers color
ing the different filters on the corrupted lena image at difference noise images are shown in Fig. 21. From the visual results, it is
densities
very evident that FBDA filter gives a better restored image,

Fig. 13 a Edge map of the

original lena image. b, c and d
are the edge maps of the lena
image corrupted with 30, 60 and
90% impulse noise, respectively

123
SIViP (2012) 6:579–595 591

Fig. 15 a, b and c represent 163rd row of the original, 60 and 85%

corrupted cameraman image, respectively. d, e, f, g, h and i represent
163rd row of the restored cameraman image after applying SMF, AMF,
EDBA, IDBA, BDND and FBDA filters on the 60% corrupted cam-
eraman image, respectively. Similarly j–o represents 163rd row of the
restored cameraman image after applying filters on the 85% corrupted
cameraman image

5 Conclusion

Fig. 14 Column a shows the edge map of the restored images of lena In this paper, we proposed a new fast and efficient fuzzy-
image corrupted with 30% of the impulse noise. Column b shows the based decision algorithm (FBDA) for the restoration of
edge map of the restored images of lena image corrupted with 60% images that are corrupted with high density of impulse noises
of the impulse noise. Column c shows the edge map of the restored
images of lena image corrupted with 90% of the impulse noise. Rows based on different noise models. FBDA is an improved
1, 2, 3, 4, 5 and 6 represent the edge map of the restored lena images switching median filter based on fuzzy logic in which the fil-
after applying the SMF, AMF, EDBA, IDBA, BDND and FBDA filters, tering is applied only to corrupted pixels in the image while
respectively the uncorrupted pixels are left unchanged. The elimination
of corrupted pixels from a selected window is done based on
a fuzzy decision. Before applying the median filter FBDA
eliminates those pixels from the window with very high and
which is closer to the original image. Also, from the quan- very low membership values in the fuzzy set small, which
titative values of the different performance measures shown might represent impulse noises. The advantage of FBDA is
in Tables 17 and 18, the superior performance of FBDA filter that it has both the noise detection power and power of elim-
over other filters is very evident. inating corrupted pixels during the filtering process. FBDA

123
592 SIViP (2012) 6:579–595

Fig. 16 a Original cameraman image. b Edge map of the original cam- based on model 3 (m = 10). h Edge map of the corrupted cameraman
eraman image. c 163rd row of the original cameraman image. d camera- image based on model 3. i 163rd row of the corrupted cameraman image
man image corrupted with 75% impulse noise based on model 2 (25% based on model 3. j cameraman image corrupted with 75% impulse
pepper and 50% salt). e Edge map of the corrupted cameraman image noise based on model 4 (25% pepper and 50% salt, m = 10). k Edge
based on model2. f 163rd row of the corrupted cameraman image based map of the corrupted cameraman image based on model 4. l 163rd row
on model 2. g cameraman image corrupted with 75% impulse noise of the corrupted cameraman image based on model 4

Fig. 17 Column a represents restored images of corrupted camera- Fig. 18 Column a represents restored images of corrupted camera-
man image (75%) based on noise model 2 (25% pepper and 50% salt). man image (75%) based on noise model 3 (m = 10). Column b repre-
Column b represents the corresponding image quality maps. Column c sents the corresponding image quality maps. Column c represents edge
represents edge maps of the restored image. Column d represents the maps of the restored image. Column d represents the 163rd row of the
163rd row of the restored images. Rows 1, 2, 3, 4, 5 and 6 represent the restored images. Rows 1, 2, 3, 4, 5 and 6 represent the results obtained
results obtained after applying the SMF, AMF, EDBA, IDBA, BDND after applying the SMF, AMF, EDBA, IDBA, BDND and FBDA filters,
and FBDA filters, respectively respectively

123
SIViP (2012) 6:579–595 593

Table 15 PSNR, SSIM, IEF and IQI values for various filters applied
on cameraman image corrupted at 75% noise density based on noise
model 3 (m = 10)
Performance SMF AMF EDBA IDBA BDND FBDA
Measures (11 × 11) (29 × 29) (3 × 3) (3 × 3) (21 × 21), (5 × 5)
(7 × 7)

PSNR (dB) 15.47 18.79 18.55 18.21 20.59 22.61

SSIM 0.7782 0.8923 0.8852 0.8759 0.9263 0.9537
IEF 7.48 16.09 15.24 14.07 24.40 38.86
IQI 0.1790 0.4634 0.4785 0.4271 0.4649 0.5658

Table 16 PSNR, SSIM, IEF and IQI values for various filters applied
on cameraman image corrupted at 75% noise density based on noise
model 4 (25% pepper and 50% salt, m = 10)
Performance SMF AMF EDBA IDBA BDND FBDA
measures (5 × 5) (29 × 29) (3 × 3) (3 × 3) (21 × 21), (5 × 5)
(7 × 7)

PSNR (dB) 7.97 13.84 18.59 17.51 19.13 22.62

SSIM 0.3289 0.7056 0.8862 0.8495 0.8959 0.9536
IEF 1.37 5.29 15.84 12.33 17.88 39.95
IQI 0.0411 0.2793 0.4635 0.3438 0.4681 0.5754

Fig. 19 Column a represents restored images of corrupted camera-

man image (75%) based on noise model 4 (25% pepper and 50% salt,
m = 10). Column b represents the corresponding image quality maps.
Column c represents edge maps of the restored image. Column d rep-
resents the 163rd row of the restored images. Rows 1, 2, 3, 4, 5 and
6 represent the results obtained after applying the SMF, AMF, EDBA,
IDBA, BDND and FBDA filters, respectively

Table 14 PSNR, SSIM, IEF and IQI values for various filters applied
on cameraman image corrupted at 75% noise density based on noise
model 2 (25% pepper and 50% salt)
Performance SMF AMF EDBA IDBA BDND FBDA
measures (5 × 5) (29 × 29) (3 × 3) (3 × 3) (21 × 21), (5 × 5)
(7 × 7)

PSNR (dB) 7.53 13.78 17.92 18.54 20.18 22.75

SSIM 0.2842 0.7003 0.8652 0.8815 0.9166 0.9549
IEF 1.37 5.76 14.94 17.25 25.18 45.42
IQI 0.0371 0.2949 0.3247 0.3603 0.4743 0.5774 Fig. 20 a Original lena color image b lena image corrupted with 75%
of impulse noise based on model 4 (25% pepper and 50% salt, m = 10).
c Original peppers color image. d Peppers image corrupted with 75%
of impulse noise based on model 4 (25% pepper and 50% salt, m = 10)
is faster compared to BDND since it uses only one level of
iteration and a fixed window size of 3 × 3 or 5 × 5. This
has been proved through experimental analysis, and it shows rithm (IDBA) and boundary discriminative noise detection
that FBDA filter produces better and excellent results com- (BDND) switching median filter. The efficiency of the pro-
pared to conventional method such as standard median filter posed algorithm is proved using standard images and has
(SMF) as well as some advanced promising techniques such been found that FBDA produces better results in terms of
as adaptive median filters (AMF), efficient decision-based both quantitative measures such as PSNR, SSIM, IEF and
algorithm (EDBA), improved efficient decision-based algo- qualitative measures such as image quality index (IQI).

123
594 SIViP (2012) 6:579–595

Fig. 21 a, g Restored images after applying SMF. b, h Restored images after applying AMF. c, i Restored image after applying EDBA. d, j
Restored images after applying IDBA. e, k Restored images after applying BDND. f, l Restored images after applying FBDA

Table 17 PSNR, SSIM, IEF and IQI values for various filters applied on lena color image corrupted at 75% noise density based on noise model
4 (25% pepper and 50% salt, m=10)
Performance measures SMF (13 × 13) AMF (29 × 29) EDBA (3 × 3) IDBA (3 × 3) BDND (21 × 21), (7 × 7) FBDA (5 × 5)

PSNR (dB) 8.49 14.84 19.86 18.71 22.29 25.07

SSIM 0.3343 0.7464 0.9046 0.8744 0.9452 0.9707
IEF 1.48 6.40 20.35 15.59 35.63 67.57
IQI 0.0599 0.3236 0.4631 0.3902 0.5597 0.6580

Table 18 PSNR, SSIM, IEF and IQI values for various filters applied on peppers color image corrupted at 75% noise density based on noise model
4 (25% pepper and 50% salt, m=10)
Performance Measures SMF (13 × 13) AMF (29 × 29) EDBA (3 × 3) IDBA (3 × 3) BDND (21 × 21), (7 × 7) FBDA (5 × 5)

PSNR (dB) 6.35 13.54 25.95 24.11 20.77 29.65

SSIM 0.2509 0.7075 0.9775 0.9665 0.9316 0.9904
IEF 1.16 6.07 105.73 69.19 32.11 247.77
IQI 0.0445 0.3212 0.5947 0.4873 0.5456 0.6972

123
SIViP (2012) 6:579–595 595

Acknowledgments Authors sincerely acknowledge the manage- 2008 International Congress on Image and Signal Processing—
ments of Rajagiri Institutions, Computer Society of India (CSI) and CISP 2008. pp. 426–431. IEEE Computer Society Press, Sanya,
the School of Computer Sciences, Mahatma Gandhi University, for their Hainan, China 1 (2008)
unstinting support and infrastructural facilities provided. This work was 12. Madhu N.S., Revathy K, Tatavarti, R.: Removal of Salt-and-Pepper
supported by Computer Society of India (CSI) under the minor research Noise in Images: A New Decision-Based Algorithm. In: Pro-
project scheme. ceedings of IAENG International Conference on Imaging Engi-
neering—ICIE 2008, IAENG International Multiconference of
Engineers and Computer Scientists—IMECS 2008, pp. 611–616.
Lecture Notes in Engineering and Computer Science 1, Hong Kong
References (2008)
13. Ng, P.-E., Ma, K.-K.: A switching median filter with boundary dis-
1. Astola, J., Kuosmanen, P.: Fundamentals of Non-Linear Digital criminative noise detection for extremely corrupted images. IEEE
Filtering. CRC, BocRaton (1997) Trans. Image Process 15(6), 1506–1516 (2006)
2. Bovik, A.: Handbook of Image and Video Processing. Aca- 14. Pomalaza-Racz, C.A., Macgillem, C.D.: An adaptive non linear
demic, New York (2000) edge preserving filter. IEEE Trans. Acoust. Speech Signal Pro-
3. Chan, R.H., Ho, C.-W., Nikolova, M.: Salt and pepper noise cess ASSP-32, 571–576 (1984)
removal by median type noise detectors and detail—preserving reg- 15. Ross, T.J.: Fuzzy Logic with Engineering Applications. 2nd
ularization. IEEE Trans. Image Process. 14(10), 1479–1485 (2005) edn. Wiley-India, India (2005)
4. Eng, H.-L., Ma, K.-K.: Noise adaptive soft-switching median fil- 16. Srinivasan, K.S., Ebenezer, D.: A new fast and efficient decision-
ter. IEEE Trans. Image Process 10(2), 242–251 (2001) based algorithm for removal of high-density impulse noises. IEEE
5. Gonzalez, R.C., Woods, R.E.: Digital Image Processing. 3rd Signal Process. Lett. 14(3), 189–192 (2007)
edn. Prentice-Hall, Englewood Cliffs NJ (2008) 17. Srinivasan, E., Ebenezer, D.: A new class of midpoint type non-lin-
6. Hwang, H., Haddad, R.A.: Adaptive median filters: new algorithms ear filters for eliminating short and long tailed noise in Images. In:
and results. IEEE Trans. Image Process 4(4), 499–502 (1995) Third International Symposium on Wireless Personal Multimedia
7. Huang, T.S., Yang, G.J., Tang, G.Y.: Fast two-dimensional median Communication, November 2000, Bangkok, Thailand
filtering algorithm. IEEE Trans. Acoust. Speech Signal Pro- 18. Wang, Z., Zhang, D.: Progressive switching median filter for the
cess. ASSP-1(1), 13–18 (1979) removal of impulse noise from highly corrupted images. IEEE
8. Jain, A.K.: Fundamentals of Image Processing. Prentice-Hall, India Trans. Circ. Syst. II 46(1), 78–80 (1999)
(PHI) (1989) 19. Wang, Z., Bovik, A.C.: A universal image quality index. IEEE
9. Klir, G.J., Yuan, B.: Fuzzy Sets and Fuzzy Logic—Theory and Signal Process. Lett. 9(3), 81–84 (2002)
Applications. Prentice-Hall, India (PHI) (1995) 20. Zadeh, L.A.: Fuzzy sets. Inf. Control 8(3), 338–353 (1965)
10. Ko, S.-J., Lee, Y.H.: Center weighted median filters and their appli- 21. Zhang, S., Karim, M.A.: A new impulse detector for switching
cations to image enhancement. IEEE Trans. Circ. Syst 38(9), 984– median filters. IEEE Signal Process. Lett. 9(4), 360–363 (2002)
993 (1991)
11. Madhu, N.S., Revathy, K., Tatavarti, R.: An improved decision-
based algorithm for impulse noise removal. In: Proceedings of

123