Median Filter For High-Density Salt-And-pepper Noise Removal in Images

New decision-based trimmed median
filter for high-density salt-and-pepper

noise removal in images
Vaithiyam Rengarajan Vijaykumar
Guru Santhanamari
Downloaded From: http://electronicimaging.spiedigitallibrary.org/ on 12/03/2014 Terms of Use: http://spiedl.org/terms
Journal of Electronic Imaging 23(3), 033011 (MayJun 2014)
New decision-based trimmed median filter for

high-density salt-and-pepper noise removal in images
Vaithiyam Rengarajan Vijaykumara,* and Guru Santhanamarib
a
Anna University, Department of ECE, Coimbatore 641047, India

Tamilnadu College of Engineering, Department of ECE, Coimbatore 641659, India
Abstract. A new switching-based trimmed median filter to remove high-density salt-and-pepper noise in digital
images is proposed. Initially, a 3 3 sliding window is applied on each pixel in the noisy image. The minimumand maximum-intensity values are trimmed, and the noisy pixels are detected based on the predefined threshold
value. In the filtering stage, the noisy pixels are replaced by median value of uncorrupted pixels in the trimmed
array. At very high noise density, if all the pixels in the sliding window are corrupted, then the proposed algorithm
replaces noisy pixels by the midpoint of recently processed pixels. The experimental results for various test
images show that the performance of the proposed algorithm is superior to the existing algorithms, namely
SMF, WMF, CWMF, AMF, DBA, and MDBUTMF in terms of visual quality and edge preservation, even at
noise levels as high as 95%. 2014 SPIE and IS&T [DOI: 10.1117/1.JEI.23.3.033011]
Keywords: edge preservation; impulse noise; nonlinear filter; trimmed median filter.
Paper 13352 received Jul. 3, 2013; revised manuscript received Mar. 16, 2014; accepted for publication May 2, 2014; published online
Jun. 9, 2014.
1 Introduction
Images are frequently corrupted by impulse noise due to malfunctioning of camera sensors, faulty memory locations in
hardware, and transmission of images in noisy channels,
which can seriously affect quality of the images.1 Salt-andpepper noise is also called fixed-valued impulse noise, which
can take two extreme-intensity values normally being the
minimum (0) and the maximum (255). In an image, edges
contain essential information and the objective of any filtering technique is to remove the impulses, so that the edge
details should be preserved. In general, linear filtering techniques for image restoration tend to blur the edges. The simplest nonlinear filter to remove the salt-and-pepper noise is
the standard median filter (SMF).2 Median filters are widely
used due to their effective noise suppression capability and
simplicity in implementation. But at higher noise levels, they
do not perform well and tend to remove the image details. In
response to these difficulties, modified median filters, such
as weighted median filter (WMF), center weighted median
filter (CWMF), and multistate median filter,35 are introduced. However, most of the median-based filters are operated uniformly across the image and thus tend to alter both
noisy and noise-free pixels and hence produce blurring on
output images.
In order to overcome the problem of median-based filters,
different kinds of decision-based median filters, such as
progressive switching median filter (PSMF)6 and adaptive
median filter (AMF),7 have been proposed. Since the
AMF uses larger window size and PSMF uses larger numbers of iterations to detect the presence of noise, their computation time is also very high. In addition to that, switching
strategies used in certain switching-based median filters
*Address all correspondence to: Vaithiyam Rengarajan Vijaykumar, E-mail:

vr_vijay@yahoo.com
Journal of Electronic Imaging
cannot differentiate high-frequency edges from high-frequency impulses in a noisy image.

Hence, to recover edges satisfactorily by taking into
account the local features, a SMF incorporated with powerful
noise detection method called boundary discriminative noise
detection (BDND) is proposed.8 In this algorithm, two iterations are invoked to detect and validate the category of current pixel as noisy or noise-free. The pixels detected as noisy
are then restored with noise adaptive switching median
(NASM)9 filter. Though BDND filter provides accurate
noise detection, it is time consuming due to a larger number
of pixels being processed to detect the presence of noise. In
addition to that, it reduces the blurring effect at higher noise
density due to larger filtering window.
In order to overcome the complexity of different switching-based median filters, a new decision-based algorithm
(DBA) is proposed,10 which uses the fixed window with a
size of 3 3 for image denoising. If the center pixel of
the sliding window is either 0 or 255, it is replaced
with median value otherwise retained. At higher noise levels,
all the pixels in the selected window are corrupted, and the
median value may also be a noisy value, in which the previously processed left neighborhood pixel is used to replace
the corrupted center pixel that produces streaking effect.
To address this issue, decision-based unsymmetric
trimmed median filter (DBUTMF) has been proposed.11
In this filter, at higher noise levels, the trimmed median
value cannot be obtained if the selected window contains all
the pixels as noisy pixels. In addition to that, DBUTMF does
not provide better restoration results at higher noise ratios. To
overcome the above drawback, the modified decision-based
unsymmetric trimmed median filter (MDBUTMF) has been
proposed, which takes the mean of all the pixels in a selected
window to replace the noisy center pixel.12 Since this filter
0091-3286/2014/$25.00 2014 SPIE and IS&T
033011-1
MayJun 2014
Vol. 23(3)
Vijaykumar and Santhanamari: New decision-based trimmed median filter for high-density salt. . .
detects the presence of salt-and-pepper noise by maximumintensity value 255 and minimum-intensity value 0, the
detection accuracy is not enough to restore the images corrupted with more than 80% of noise level. In addition to
that, if all the pixels in the window are either 0 or 255,
then the MDBUTMF replaces the noisy pixel by mean value,
which is also either 0 or 255. Unlike the median filter,
mean filters smooth the images. Recently, an improved BDND
filter13 is proposed to overcome the blurring effect of BDND
by considering the spatial correlation between the noisy and
uncorrupted pixels in the filtering window. Though it reduces
the blurring effect, the computation time is still high since it
uses the same detection process as in BDND.
To overcome the above drawbacks in various filters, a new
decision-based trimmed median filter, to remove high-density salt-and-pepper noise with better performance in
terms of qualitative and quantitative results, is proposed in
this paper. In the proposed filter, a fixed 3 3 sliding window is applied to the current pixel being processed, and the
pixels are sorted. A trimmed array is obtained by eliminating
the minimum- and maximum-intensity values in the sorted
array. The noisy pixels are detected by comparing the absolute difference between the center pixel and the median value
of uncorrupted pixels in the trimmed array with a predefined
threshold value T. If it is identified as noisy pixel, then it
is replaced with the median value otherwise left unaltered.
The rest of this paper is organized as follows. In Sec. 2,
the proposed trimmed median filtering algorithm is
described and illustration is given in Sec. 3. The simulation
results of the proposed trimmed median filter in image
denoising application and performance comparisons with
the other existing algorithms are displayed in Sec. 4.
Section 5 concludes the paper.
2 Proposed Decision-Based Trimmed Median Filter
Initially, a 3 3 fixed sliding window is imposed on each
pixel being processed in the noisy image. The minimumintensity (Smin ) and maximum-intensity (Smax ) values
obtained in the selected window are trimmed, and remaining
pixels, which are not equal to Smin and Smax , are collected in
an array. Then, the absolute difference between the center
pixel and the mean value of those pixels collected in the
array is compared with a predefined threshold value T
to detect whether the center pixel is corrupted or not. If it
is detected as noisy pixel, then filtering is applied; otherwise
it is left unaltered. Because of the detection mechanism, the
proposed algorithm works well even above the 50% breakdown limit for a 3 3 fixed window. In the case of 90% and
above noise levels, if all the pixels in the 3 3 filtering window are corrupted, then the proposed algorithm replaces the
center pixel with the midpoint of previously processed four
pixels. The proposed algorithm is explained in the following
steps.
W 3 i; j Xi k; j l;
1 k; l 1:
(1)
Case (i)
Step 2: If Smin 0 and Smax 255 continue; else go to
Step 6.
Step 3: The Smin and Smax values in the array Zi are trimmed
and all the uncorrupted pixels, which are not equal to
Smin and Smax , are retained in the same array. The set
of noise-free coordinates are defined in Eq. (2), in
which f ik;jl is gray level intensity at pixel location
i k, j l. If Zi is empty go to Step 11; else
continue.
Si;j fi k:j 1; i k; j 1
W 3 i; j f ik;j1 Smin f ik;j1 Smax g
(2)
Step 4: Median (MED) of an array Zi is found. If
jX ij MEDj < T, then X ij is noise-free and actual
value is retained; otherwise X ij is replaced with MED,
in which the threshold value T lies in the range
[2730].
Step 5: Repeat the process from Step 1, until all the pixels in
the noisy image are processed.
Case (ii)
Step 8.
Smin , in trimmed array Zi are retained. The set of
noise-free coordinates are defined by Eq. (3), in
which f ik;jl is gray level intensity at pixel location
i k, j l. If Zi is empty go to Step 11; else go
to Step 4.
Si;j fi k; j 1; i k; j 1
W 3 i; j f ik;j1 Smin g:
(3)
Case (iii)
Step 11.
Smax , in trimmed array Zi are retained. The set of
noise-free coordinates is defined by Eq. (4), in which
f ik;jl is gray level intensity at pixel location i k,
j l. If Zi is empty go to Step 11; else go to
Step 4.
Algorithm Steps:
Step 1: Let Xij be a pixel being processed in the noisy
image. A 3 3 fixed window W 3 i; j defined in
Eq. (1) centered about X ij is applied, and elements in
W 3 i; j are collected in an array Ziji19. Smin and
Smax in the window are obtained by sorting the elements
in the array Zi
Si;j fi k; j 1; i k; j 1
W 3 i; j f ik;j1 Smax g:
(4)
Step 10: The current processing pixel X ij is noise-free and

actual value is retained.
033011-2
MayJun 2014
Vol. 23(3)
Step 11: The midpoint of previously processed neighbors

Xi1;j1 , X i1;j X i1;j1 , Xi;j1 is used to replace the
center pixel X ij and go to Step 5.
In the case of 90% and above noise density, if there is no
uncorrupted pixel in the trimmed array Zi, the mean or
median statistics can also be used to replace the corrupted
center pixel instead of midpoint estimator as explained in
Step 11. Nevertheless, there is not much difference among
mean, median, and midpoint estimators; the computation
time for obtaining midpoint of previously processed four
pixels is less compared to mean and median estimators.

Hence, midpoint statistics are chosen in the proposed algorithm. The optimum threshold value T is determined by
trial-and-error approach for various test images and lies
approximately in the range [2730].
3 Illustration of Proposed Filter
This section illustrates the proposed algorithm with four different cases for Lena image corrupted with 70% of salt-andpepper noise.
Table 1 Peak signal-to-noise ratio (PSNR) comparison for Lena and Bridge image.
Lena
ND (%) SMF WMF CWMF PSMF AMF
Bridge
DBA MDBUTMF Proposed SMF WMF CWMF PSMF AMF
DBA MDBUTMF Proposed
10
33.76 34.27 33.62
38.72 38.01 38.06
41.15
39.26
26.16 27.54 27.63
35.22 29.59 29.34
31.16
28.31
20
29.49 27.14 25.29
36.20 34.97 37.23
37.25
37.27
24.40 24.37 23.44
32.77 28.15 28.21
28.55
27.58
30
24.01 21.46 20.02
35.39 32.37 35.68
34.75
35.67
21.40 20.07 19.08
30.71 26.78 26.98
26.64
26.96
40
19.17 17.21 16.14
35.09 29.96 33.76
32.26
34.47
18.07 16.56 15.52
29.49 25.35 25.68
25.08
26.29
50
15.35 14.04 13.11
34.36 28.26 31.65
30.27
33.11
14.76 13.43 12.67
28.40 23.89 24.36
23.53
25.53
60
12.32 11.51
10.8
32.15 26.64 30.10
28.03
31.51
11.86 11.20 10.52
27.25 22.51 23.02
21.85
24.58
70
10.07 9.45
9.05
27.45 24.67 28.20
25.75
29.73
9.77
9.18
8.79
24.57 21.09 21.92
20.16
23.43
80
8.14
7.86
7.56
20.60 21.37 25.75
22.81
27.74
7.85
7.61
7.37
19.32 18.17 20.55
18.20
22.00
90
6.6
6.51
6.36
14.39 14.43 23.14
18.14
24.48
6.42
6.32
6.19
13.79 13.51 18.78
15.45
20.00
Table 2 Mean absolute error (MAE) and image enhancement factor (IEF) for Lena image.
MAE
IEF
10
1.442 1.109 0.913
0.923 0.968 0.973
0.172
0.255
22.83 37.51 56.04
248.0 230.3 236.6
623.97
672.2
20
1.773 1.593 1.650
1.707 1.313 1.281
0.374
0.449
53.98 60.25 58.56
312.0 200.0 339.6
546.08
761.9
30
2.510 2.906 3.347
2.230 1.801 1.760
0.594
0.661
51.33 37.12 31.23
400.6 94.11 368.8
472.18
760.9
40
4.460 5.445 6.570
2.619 2.513 2.366
0.883
0.866
30.29 18.96 14.09
467.5 50.97 321.1
391.03
733.2
50
8.224 10.10 11.84
3.086 3.309 3.214
1.210
1.126
14.02 9.167 6.721
494.5 32.97 263.9
307.31
687.3
60
14.52 16.86 19.04
4.183 4.309 4.124
1.661
1.462
6.630 4.924 3.700
359.6 26.28 205.3
239.86
625.3
70
23.18 25.95 27.98
7.188 5.753 5.439
2.257
1.858
3.554 2.841 2.380
116.0 18.12 152.0
164.66
519.0
80
34.60 36.64 38.77
20.22 8.541 7.614
3.310
2.501
2.068 1.854 1.640
13.37 12.53 102.7
90.632
422.6
90
48.17 45.50 49.80
68.34 21.95 12.79
5.970
3.896
1.357 1.289 1.250
1.232 5.227 54.25
24.676
272.7
033011-3
MayJun 2014
Vol. 23(3)
Case (i)
The case if the center pixel in the selected window is noisy and some neighborhood pixels add salt-and-pepper noise to the image
is illustrated in case (i). The pixels in the selected window are sorted, and Smin and Smax values are trimmed. Then, the uncorrupted
pixels, which are not equal to Smin and Smax in Zi , are retained. The median value of the array Zi is found. Since the absolute
difference between the median value and the center pixel value is greater than 30, the center pixel is noisy and replaced with the
median value as shown below:
0
49
50
Trimmed array Zi = [0, 46, 48, 49, 50, 255, 255]
255
255
Zi = [46, 48, 49, 50] MED = 49
46
255
48
| Xij MED | = 206 > 30, Xij= 49
Table 3 Mean absolute error (MAE) and image enhancement factor (IEF) for Bridge image.
MAE
IEF
10
3.709 2.788 2.321
1.656 2.777 2.947
0.482
1.410
3.865 7.120 9.978
56.96 41.01 52.58
48.28
25.21
20
4.213 3.423 3.175
3.125 3.436 3.543
0.991
1.824
11.39 16.35 18.78
65.00 87.91 133.5
62.80
55.64
30
5.207 4.880 4.942
5.005 4.392 4.436
1.576
2.257
16.15 16.58 14.84
56.91 99.55 204.7
67.90
82.99
40
7.082 7.340 7.994
6.630 5.705 5.674
2.220
2.723
14.14 11.38 9.062
52.24 74.49 275.1
65.54
97.54
50
10.45 11.84 13.05
8.539 7.320 7.186
3.025
3.228
8.935 6.600 5.252
49.18 58.83 291.5
55.65
106.0
60
16.13 17.26 18.97
10.76 9.283 9.090
3.929
3.917
4.982 3.994 3.235
46.51 34.72 298.1
46.19
107.0
70
23.43 25.30 27.01
14.93 11.76 11.21
5.154
4.766
3.054 2.520 2.182
26.76 26.42 240.9
35.62
99.67
80
33.46 34.69 36.08
29.57 15.93 14.06
7.023
5.940
1.903 1.737 1.585
8.104 18.35 157.6
24.89
86.02
90
44.99 45.48 46.13
76.57 30.12 18.62
9.754
7.953
1.342 1.281 1.209
1.113 7.251 78.76
11.21
61.82
Table 4 Structural similarity index (SSIM) for Lena and Bridge image.
Lena
Bridge
10
0.911 0.930 0.934
0.907 0.962 0.964
0.990
0.995
0.780 0.850 0.867
0.903 0.911 0.910
0.976
0.924
20
0.862 0.839 0.788
0.878 0.953 0.957
0.978
0.976
0.733 0.773 0.764
0.886 0.894 0.895
0.950
0.894
30
0.718 0.623 0.525
0.891 0.939 0.940
0.965
0.963
0.629 0.619 0.574
0.834 0.866 0.859
0.917
0.875
40
0.468 0.359 0.282
0.902 0.916 0.921
0.947
0.949
0.468 0.423 0.360
0.789 0.823 0.820
0.877
0.846
50
0.238 0.177 0.129
0.905 0.889 0.887
0.924
0.931
0.284 0.242 0.201
0.731 0.769 0.766
0.820
0.812
60
0.109 0.085 0.065
0.872 0.854 0.836
0.891
0.908
0.148 0.138 0.111
0.661 0.704 0.701
0.751
0.768
70
0.052 0.042 0.036
0.769 0.802 0.774
0.845
0.877
0.080 0.071 0.059
0.550 0.621 0.632
0.653
0.709
80
0.025 0.022 0.018
0.467 0.692 0.691
0.769
0.835
0.037 0.037 0.030
0.326 0.491 0.529
0.514
0.629
90
0.011 0.011 0.009
0.091 0.338 0.537
0.622
0.759
0.015 0.017 0.015
0.077 0.249 0.384
0.333
0.492
033011-4
MayJun 2014
Vol. 23(3)
Case (ii)
The case if the center pixel in the selected window is noisy and some neighborhood pixels add pepper noise to the image is
illustrated in case (ii). Since the absolute difference between the median value of the array Zi and center pixel value is greater
than 30, the center pixel is noisy and replaced with the median value as shown below:
0
205
206
Trimmed array Zi = [0, 0, 205,205,205,206,208]
205
209
Zi = [205, 205, 205, 206, 208] MED = 205
208
205
| Xij MED | = 205 > 30, Xij= 205
Fig. 1 Comparison of (a) PSNR, (b) MAE, (c) IEF, (d) SSIM, (e) run time of various algorithms for Lena
image corrupted with various noise densities.
033011-5
MayJun 2014
Vol. 23(3)
Case (iii)
The case if the center pixel in the selected window is noisy and some neighborhood pixels add salt noise to the image is
illustrated in case (iii). Since the absolute difference between the median value of the array Zi and center pixel value is greater
than 30, the center pixel is noisy and replaced with the median value.
255
49
57
51
255
50
Zi = [49, 50, 51, 52, 57]
48
| Xij MED | = 204 > 30, Xij= 51
52
255
Trimmed array Zi = [49, 50, 51, 52, 57, 255, 255]

MED = 51
Fig. 2 (a)(j) Restoration results for Lena image corrupted with 95% salt-and-pepper noise density.
(a) Original, (b) noisy, (c) SMF, (d) WMF, (e) CWMF, (f) PSMF, (g) AMF, (h) DBA, (i) MDBUTMF, (j) proposed.
033011-6
MayJun 2014
Vol. 23(3)
Case (iv)
If the center pixel value is noise free, then the absolute difference between the median value and center pixel is less than 30,
then the processing pixel is left unaltered.
58
66
56
| Xij MED | = 0 < 30, Xij= 59
73
59
65
Zi = [57, 57, 58, 59, 63, 65, 66] MED=59
58
66
56
| Xij MED | = 0 < 30, Xij= 59
Fig. 3 (a)(j) Restoration results for bridge image corrupted with 80% salt-and-pepper noise density.
(a) Original, (b) noisy, (c) SMF, (d) WMF, (e) CWMF, (f) PSMF, (g) AMF, (h) DBA, (i) MDBUTMF, (j) proposed.
033011-7
MayJun 2014
Vol. 23(3)
4 Results and Discussion

In this section, the performance of the proposed filter is
tested by using standard grayscale test images namely
Lena, Bridge, Elaine, and Peppers of size 512 512,
8 bitspixel and compared with state-of-the-art algorithms,
such as SMF, WMF, CWMF, PSMF, AMF, DBA and
MDBUTMF, by varying the noise level from 10% to 90%
with the incremental step of 10% for every simulation. In
addition to the visual quality, the performance of the

proposed algorithm and the other existing algorithms is
quantitatively measured by the parameters such as peak signal-to-noise ratio (PSNR), mean absolute error (MAE),
structural similarity index (SSIM), and image enhancement
factor (IEF). The performance of restoration and processing
time for proposed filter and the existing filters are analyzed
under the following Secs. 4.1 and 4.2.
Fig. 4 (a)(j) Restoration results for Elaine image corrupted with 70% salt-and-pepper noise density.
(a) Original, (b) noisy, (c) SMF, (d) WMF, (e) CWMF, (f) PSMF, (g) AMF, (h) DBA, (i) MDBUTMF,
(j) proposed.
033011-8
MayJun 2014
Vol. 23(3)
4.1 Quantitative and Visual Results Comparison

The restoration performance of the proposed algorithm is
demonstrated in this section. The quantitative performance
metrics, in terms of PSNR, MAE, IEF, and SSIM for proposed filter and seven competitive filters for the test images
Lena and Bridge at noise level varying from 10% to 90%, are
documented in Tables 14.
The performance of SMF, WMF, and CWMF in terms of
visual and quantitative metrics is poor compared to the
proposed and the other existing filters for low to high

noise densities since there is no noise detection process
involved. Though the PSNR value obtained by the PSMF
algorithm is equally as good as the proposed filter for
noise densities up to 60%, it decays to very low value for
higher noise densities. The AMF algorithm produces better
PSNR up to a noise density level of 40%, and it decreases at
higher noise densities due to larger window size. The fixed
window algorithms, such as DBA and MDBUTMF, produce
Fig. 5 (a)(j) Restoration results for Peppers image corrupted with 60% salt-and-pepper noise density.
(j) proposed.
033011-9
MayJun 2014
Vol. 23(3)
Fig. 6 (a)-(j) Restoration results for gray level variation of 100th row of Lena image corrupted with 90%
noise density. (a) Original, (b) noisy, (c) SMF, (d) WMF, (e) CWMF, (f) PSMF, (g) AMF, (h) DBA,
(i) MDBUTMF, (j) proposed.
033011-10
MayJun 2014
Vol. 23(3)
lower PSNR values than the proposed filter for noise level
more than 60%.
The proposed algorithm produces much lower MAE and
better IEF than the existing filters at higher noise densities,
which are given in Tables 2 and 3. The proposed algorithm
also exhibits improved structural similarity index, which
is displayed in Table 4. The above discussed quantitative
performance comparisons are also presented graphically
in Fig. 1.
The proposed algorithm is analyzed qualitatively by

observing the restoration result of proposed algorithm and
the existing algorithms, for the images Lena corrupted
with 95% of salt-and-pepper noise and Bridge corrupted
with 80% salt-and-pepper noise, visually shown in Figs. 2
and 3, respectively. The algorithms, such as SMF, WMF,
and CWMF, failed to restore the original image for the
noise level greater than 50%, and PSMF leaves some noticeable impulses in the output image at higher noise densities.
Fig. 7 (a)(j) Error images for Lena image corrupted with 90% salt-and-pepper noise density.
(j) proposed.
033011-11
MayJun 2014
Vol. 23(3)
At higher noise densities, DBA produces a streaking effect

due to the repeated replacement of left neighborhood pixel in
the sliding window. The AMF restores the original image
with slight blurring effect, because of the larger window
size. Figures 4 and 5 show the restoration results of
various filters for Elaine image corrupted with 70% saltand-pepper noise and the Peppers image corrupted with
60% salt-and-pepper noise, respectively. Based on the subjective visual comparisons, it is also observed that the proposed filter removes impulse noise effectively and preserves
the image details, such as edges, better than the other existing
filters.
The change in intensity profile for a single row of original, noisy, and restored image for various filters is shown in
Fig. 6. It can be seen from the figure that the intensity variation of the original and restored image remains almost same
for the proposed filter than the existing filters. The error
image of proposed and the other existing filters for Lena corrupted with 90% salt-and-pepper noise density is displayed
in Fig. 7. The error image for the proposed filter looks darker
than other competitive filters, which clearly show the better
edge preservation. Based on the above analysis, it is easy to
see that the proposed filter outperforms the existing switching-based filters in both quantitative and qualitative aspects.
In order to show the visual performance of the proposed
algorithm and the other existing algorithms more clearly,
all four test images are corrupted with 90% salt-and-pepper
noise density, and restoration results are presented in Fig. 8.
The visual results in Fig. 8 clearly indicate the superior performance of the proposed algorithm over the existing
methods.
Fig. 8 Restoration results of (a) Lena, (b) Bridge, (c) Elaine and
(d) Peppers images using various filters, namely SMF, WMF, CWMF,
PSMF, AMF, DBA, MDBUTMF, and proposed algorithm, for 90% saltand-pepper noise densities.
4.2 Run-Time Comparison

In addition to PSNR, MAE, and MSSIM, the performance
of proposed filter is also tested quantitatively in terms of
processing time and compared with the existing filters for
test images, such as Lena, Bridge, Elaine, and Peppers,
with noise level varying from 10% to 90% and listed in
Tables 5 and 6. All the filters are simulated in MATLAB
on a PC equipped with 2.66 GHz operating speed, 2GB
RAM. The computation time for SMF, WMF, and CWMF
is much less because there is no detection mechanism
involved.
From Tables 5 and 6, it is observed that the PSMF
algorithm has taken longer run time than the proposed
algorithm due to processing of more pixels in an iterative
manner. The AMF is executed with longer time than the
proposed filter, especially at higher noise densities, since
it uses a large window size for a higher noise ratio. The
run time for MDBUTMF is less only at lower noise levels
and it takes more time for execution than the proposed
filter at higher noise densities above 70%. It is also inferred
from the tables that the fixed window algorithm DBA has
taken a run time much closer to the proposed filter for noise
levels varying from 10% to 90%, but a streaking effect
dominates in DBA at higher noise ratios. Hence, the proposed filter provides a better tradeoff among run time
and other performance metrics, such as PSNR, and visual
quality than the existing algorithms, especially at higher
noise levels.
033011-12
MayJun 2014
Vol. 23(3)
Table 5 Run time comparison for Lena and Bridge image.
Lena
Bridge
ND (%) SMF WMF CWMF PSMF AMF DBA MDBUTMF Proposed SMF WMF CWMF PSMF AMF DBA MDBUTMF Proposed
10
5.49
6.04
5.65
38.85
7.12
6.51
1.62
7.44
5.5
6.04
5.65
36.75 7.46 6.56
1.60
7.41
20
5.53
6.06
5.68
36.59
7.13
6.54
2.57
7.45
5.54
6.04
5.70
37.01 7.42 6.53
2.59
7.48
30
5.51
6.03
5.62
36.75
7.14
6.50
3.48
7.48
5.51
6.04
5.67
36.62 7.45 6.51
3.56
7.51
40
5.48
6.09
5.71
36.87
7.37
6.54
4.39
7.49
5.48
6.03
5.70
36.68 7.61 6.60
4.43
7.54
50
5.51
6.04
5.70
36.91
7.85
6.51
6.89
7.51
5.51
6.04
5.67
36.78 8.06 6.56
5.32
7.48
60
5.50
6.03
5.75
36.65
8.70
6.56
7.73
7.51
5.50
6.04
5.73
36.73 9.42 6.51
6.23
7.45
70
5.53
6.04
5.69
36.95 10.21 6.53
8.48
7.48
5.53
6.03
5.68
37.01 10.5 6.53
7.09
7.46
80
5.53
6.04
5.70
37.59 13.63 6.50
9.31
7.48
5.53
6.03
5.7
37.82 13.6 6.51
7.98
7.42
90
5.53
6.03
5.70
38.75 20.42 6.54
10.15
7.37
5.53
6.06
5.71
38.96 20.7 6.54
8.78
7.39
Table 6 Run time comparison for Elaine and Pepper image.
Elaine
Peppers
ND (%) SMF WMF CWMF PSMF AMF DBA MDBUTMF Proposed SMF WMF CWMF PSMF
AMF
10
5.54 6.07
5.71
37.26
7.11
6.58
1.57
7.45
5.57 6.12
5.71
38.67
7.14
6.55
1.57
7.45
20
5.56 6.09
5.70
36.38
7.12
6.61
2.54
7.46
5.50 6.07
5.70
36.9
7.25
6.74
2.53
7.46
30
5.57 6.07
5.71
36.86
7.10
6.58
3.48
7.51
5.53 6.09
5.73
37.02
7.18
6.68
3.48
7.48
40
5.56 6.06
5.71
36.73
7.45
6.59
4.42
7.48
5.54 6.07
5.68
36.95
7.51
6.71
4.39
7.5
50
5.56 6.09
5.75
36.94
8.01
6.57
5.34
7.42
5.56 6.07
5.71
36.99
7.85
6.90
5.34
7.53
60
5.53 6.12
5.76
35.94
8.85
6.54
6.25
7.51
5.56 6.07
5.75
37.11
8.89
6.54
6.23
7.5
70
5.56 6.09
5.76
37.88 10.02 6.58
8.54
7.48
5.59 6.09
5.71
37.53 10.630 6.54
8.53
7.54
80
5.53 6.10
5.76
37.54 13.14 6.57
8.10
7.48
5.56 6.09
5.68
37.74
13.48
6.61
9.39
7.65
90
5.53 6.06
5.71
39.01 19.63 6.56
10.2
7.42
5.56 6.06
5.71
37.93
19.41
6.71
10.23
7.65
5 Conclusion
In this paper, a new decision-based trimmed median filter
with fixed window for effective removal of high-density
salt-and-pepper noise in images is proposed. The use of a
small 3 3 fixed window in the proposed algorithm leads
to preservation of fine details, such as edges, satisfactorily.
The better noise-removal capability of the proposed filter
over other fixed window algorithms, such as DBA and
MDBUTMF, is demonstrated using different test images.
Experimental results reveal that the proposed filter outperforms the existing state-of-the-art filters by providing better
PSNR, IEF, SSIM, MAE values and visual quality for the
images corrupted up to 95% of noise level.
References
1. R. C. Gonzalez and R. E. Woods, Digital Image Processing, 2nd ed.,
Prentice Hall, Upper Saddle River, NJ (2002).
2. I. Pitas and A. N. Venetsanopoulos, Order statistics in digital image
processing, Proc. IEEE, 80(12), 18931921 (1992).
3. L. Yin et al., Weighted median filters: a tutorial, IEEE Trans. Circuits
Syst. II: Analog Digital Signal Process. 43(3) 157192 (1996).
4. S. J. Ko Chen and Y. Hoom Lee, Center-weighted median filters and
their applications to image enhancement, IEEE Trans. Circuits Syst.
38(9), 984993 (1991).
033011-13
MayJun 2014
Vol. 23(3)
5. T. Chen and H. R. Wu, Impulse noise removal by multi-state median
filtering, in Proc. 2000 IEEE Int. Conf. on Acoustics, Speech, and
Signal Processing, 2000, ICASSP 00, Istanbul, Vol. 4, pp. 2183
2186 (2000).
6. Z. Wang and 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), 7880
(1999).
7. H. Hwang and R. A. Haddad, Adaptive median filters: new algorithms
and results, IEEE Trans. Image Process. 4(4), 499502 (1995).
8. P. E. Ng and K. K. Ma, A switching median filter with boundary discriminative noise detection for extremely corrupted images, IEEE
Trans. Image Process. 15(6), 15061516 (2006).
9. H. L. Eng and K. K. Ma, Noise adaptive soft-switching median filter,
IEEE Trans. Image Process. 10(2), 242251 (2001).
10. K. S. Srinivasan and D. Ebenezer, A new fast and efficient decision
based algorithm for removal of high density impulse noise, IEEE
Signal Process. Lett. 14(3), 189192 (2007).
11. K. Aiswarya, V. Jayaraj, and D. Ebenezer, A new and efficient
algorithm for the removal of high density salt and pepper noise in
images and videos, in Proc. Second Int. Conf. Computer
Modeling and Simulation, Sanya, Hainan, Vol. 4, pp. 409413,
IEEE (2010).
12. S. Esakkirajan et al., Removal of high density salt and pepper noise
through modified decision based unsymmetric trimmed median filter,
IEEE Signal Process. Lett. 18(5), 287290 (2011).
13. P. E. Ng and K. K. Ma, Efficient improvements on the BDND Filtering

for removal of high density impulse noise, IEEE Trans. Image
Process. 22(3), 12231232 (2013).
Vaithiyam Rengarajan Vijaykumar received his BE degree in electronics and communication engineering from Madras University, and
his ME degree in communication systems at Thiagarajar College of
Engineering Madurai in 1997. He completed his PhD degree in the
area of nonlinear image filtering from Anna University in 2008. He
has 17 years of teaching experience. He has worked at Madras
Institute of Technology, Chennai, and PSG College of Technology
as a teaching faculty. Currently, he is working as associate professor
in the Department of Electronics and Communication Engineering,
Anna University, Coimbatore, TamilNadu, India. His areas of interest
include image processing, signal processing, and digital
communication.
Guru Santhanamari is currently working as an assistant professor
(senior grade) in the Department of Electronics and Communication
Engineering, TamilNadu College of Engineering, Coimbatore, India.
She received her BS degree in Government College of Technology,
Coimbatore, and MS degree in PSG College of Technology, Coimbatore. Her research interests include digital image processing, signal
processing, and embedded systems.
033011-14
MayJun 2014
Vol. 23(3)

Median Filter For High-Density Salt-And-pepper Noise Removal in Images

Diunggah oleh

Informasi Dokumen

Judul Asli

Hak Cipta

Format Tersedia

Bagikan dokumen Ini

Bagikan atau Tanam Dokumen

Opsi Berbagi

Apakah menurut Anda dokumen ini bermanfaat?

Apakah konten ini tidak pantas?

Hak Cipta:

Format Tersedia

Median Filter For High-Density Salt-And-pepper Noise Removal in Images

Diunggah oleh

Hak Cipta:

Format Tersedia

New decision-based trimmed median

filter for high-density salt-and-pepper

Downloaded From: http://electronicimaging.spiedigitallibrary.org/ on 12/03/2014 Terms of Use: http://spiedl.org/terms

Journal of Electronic Imaging 23(3), 033011 (MayJun 2014)

New decision-based trimmed median filter for

Anna University, Department of ECE, Coimbatore 641047, India

*Address all correspondence to: Vaithiyam Rengarajan Vijaykumar, E-mail:

Journal of Electronic Imaging

cannot differentiate high-frequency edges from high-frequency impulses in a noisy image.

Downloaded From: http://electronicimaging.spiedigitallibrary.org/ on 12/03/2014 Terms of Use: http://spiedl.org/terms

Step 10: The current processing pixel X ij is noise-free and

Downloaded From: http://electronicimaging.spiedigitallibrary.org/ on 12/03/2014 Terms of Use: http://spiedl.org/terms

Step 11: The midpoint of previously processed neighbors

pixels is less compared to mean and median estimators.

DBA MDBUTMF Proposed

33.76 34.27 33.62

38.72 38.01 38.06

26.16 27.54 27.63

35.22 29.59 29.34

29.49 27.14 25.29

36.20 34.97 37.23

24.40 24.37 23.44

32.77 28.15 28.21

24.01 21.46 20.02

35.39 32.37 35.68

21.40 20.07 19.08

30.71 26.78 26.98

19.17 17.21 16.14

35.09 29.96 33.76

18.07 16.56 15.52

29.49 25.35 25.68

15.35 14.04 13.11

34.36 28.26 31.65

14.76 13.43 12.67

28.40 23.89 24.36

32.15 26.64 30.10

11.86 11.20 10.52

27.25 22.51 23.02

27.45 24.67 28.20

24.57 21.09 21.92

20.60 21.37 25.75

19.32 18.17 20.55

14.39 14.43 23.14

13.79 13.51 18.78

ND (%) SMF WMF CWMF PSMF AMF

DBA MDBUTMF Proposed SMF WMF CWMF PSMF AMF

DBA MDBUTMF Proposed

1.442 1.109 0.913

0.923 0.968 0.973

22.83 37.51 56.04

248.0 230.3 236.6

1.773 1.593 1.650

1.707 1.313 1.281

53.98 60.25 58.56

312.0 200.0 339.6

2.510 2.906 3.347

2.230 1.801 1.760

51.33 37.12 31.23

400.6 94.11 368.8

4.460 5.445 6.570

2.619 2.513 2.366

30.29 18.96 14.09

467.5 50.97 321.1