20IO International Conference on Computer Application and System Modeling (ICCASM 2010)

A Detail Preserving Filter for Impulse Noise Removal

Dagao Duan, Qian Mo Yueliang Wan, Zhongming Han

College of Computer and Information Engineering College of Computer and Information Engineering
Beijing Technology and Business University Beijing Technology and Business University
Beijing, China Beijing, China
[email protected] [email protected]

Abstract-It is known that digital images are frequently solutions have been proposed to trade off detail preservation
corrupted by impulse noise due to errors generated in camera against noise reduction, such as the center weighted median
sensors, analog-to-digital conversion and communication (CWM) filter [2] and the soft-switching median (SWM)
channels. Therefore, it is imperative to remove impulse noise in filter[3]. The CWM filter uses weights center pixel of the
order to facilitate the subsequent processing such as edge filtering window to control the filtering behavior preserves
detection, segmentation, analysis, and pattern recognition etc.
the details. It is known that the SWM filter is a powerful
Many linear and nonlinear filtering techniques have been
filter for restoring the corrupted image, which incorporates
proposed earlier to remove impulse noise, however these filter
MED and CWM filter. It uses the sliding window to
often bring along blurred and distorted image of details.
determine whether the current pixel is an impulse noise or
Filtering an image to attenuate noise while keeping the image
not. However, it may miss the noise pixel when the noise's
details preserved is one of the most important issues. In this
paper a detail preserving filter for impulse noise removal is
value is near to its neighbors or the threshold is set too large
proposed, which is based on the soft-switching median (SWM)
for preserving the details, and it may distored some detailed
filter. In the first, the impulse noise candidates are detected by pixels when the threshold is set too small for extracting the
arranging the pixels in order in the sliding window. In the impulse noise rigidly.
second, we analyze the noise candidates and c1assity them into In recent years, many median-based impulse noise filters
noise-free pixels, noisy pixels and detail pixels (edges and proposed in literatures, e.g., the peak-and-valley (PV) filter
smooth changing regions). Finally, the process employed the [4], extremum and median (EM) filter [5], minimum­
rank-ordered mean filter (ROM) to remove the corrupted maximum exclusive mean (MMEM) filter[5], pre-scanned
pixels and the details can be restored and preserved. Extensive minmax center-weighted (PMCW) filter[7], decision-based
experiments indicate that the proposed method can effectively median filters[8], adaptive center weighted median (ACWM)
restore images corrupted with impulse noise and performs filter[9], signal dependent rank order mean (SD-ROM)
significantly better in terms of noise suppression and detail filter[l O], conditional signal-adaptive median (CSAM)
preservation than a number of existing nonlinear techniques. filter[l l ], the pixel-wise MAD (PWMAD) filter[12], and
iterative median filter[13], etc. With the noise detector, these
Keywords: Impulse noise, Median jilter, Noise removel, filters considers only the characteristics of impulse noise, but
Image processing
not effectively integrate the information of the image, so the
edges and fine details of image preservation remains to be
l. INTRODUCTION further improved.
It is known that digital images are frequently corrupted In this paper a detail preserving filter for impulse noise
by impulse noise due to errors generated in camera sensors, removal is proposed, which is based on SWM filter. In the
analog-to-digital conversion and communication channels[l ]. first, the impulse noise candidates are detected by arranging
The intensity of impulse noise has the tendency of being the pixels in order in the sliding window. In the second, we
either relatively high or low. The subsequent image analyze the noise candidates and classifY them into noise­
processing including edge detection, image analysis, and free pixels, noisy pixels and detail pixels (edges and smooth
object tracking may perform worse if the input image has a changing regions). Finally, the process employed a rank­
larger noise ratio. Filtering a digital image to attenuate noise ordered mean (ROM) filter outputs to replace the corrupted
while keeping the image details preserved is an essential part pixels and the details can be restored and preserved.
of image processing. Therefore, how to detect noise pixels This paper is organized as follows. Section II reviews
and then to replace them with appropriate values is one of impulse noise model and indicates the fixed-value impulse
the most important issues. noise be investigated. Section III introduces the proposed
Various filtering techniques have been proposed for algorithm for removing the impulse noise in the corrupted
removing impulse noise in the past. One of the most popular image, Section IV presents the experiments and the
and robust nonlinear filter is the standard median (MED) comparison results between the proposed algorithm and the
filter[l ], which eliminates the noise and performs better in other known methods. Finally we make a brief conclusion in
the smooth region of an image. However, in the detailed Section V.
regions, such as edge and texture, the MED filter might
smear the details. In order to solve this problem, some

978-1-4244-7237-6/10/$26.00 ©201O IEEE V2-265

Authorized licensed use limited to: University of Malaya. Downloaded on March 09,2024 at 15:37:13 UTC from IEEE Xplore. Restrictions apply.
 2010 International Conference on Computer Application and System Modeling (ICCASM 2010)

II. IMPULSE NOISE MODEL A. Impulse noise detection

In this paper, we only focus on the detection and In the proposed scheme, a sliding window centered
denoising of fixed-value impulse noise, namely "salt and around Xu is employed to detect the extrema for the
pepper noise". Consider a corrupted image X, which
corrupted pixels. Let we choose M for the sliding window
containing impulse noise with probability P is always
size, Now there are (M2 -1 ) pixels (excluding Xu itself) in
expressed in the form:
ni p
{ withprobabiliy P the window.
X Let X = represent the input
i, } {Xi-L,J-L,oo.,oo.,Xi+L,i+L}
==
S '
i, } withprobabiliyl- p
(1)
sample in the MXM sliding window where Xi ) is
Xi ) ,
where i=I,2, .....sI and j=I,2, .....s2 and O<p<I. , excluded and M 2L + 1 with L = 2 in this paper. Now, we
represents the intensity of the pixel located at position ( i , j ) re-arrange the input samples X in the ascending order:
==

S { 1 2 1 M'-l 2 M 2_1
i n the image, i,j and n i , j denotes the intensity o f the pixel XI' == X, X , where X ::;; X ::;;
, 00. , X }::;; X • • •

(i, j) in the original image and the noisy image, respectively Usually, the orders of the noise pixels in the sliding window
Xi are either very big or very small, i.e. this kind of impulse
and nI, . a noise value independent from ,} . For gray level noise usually lies either at the beginning or at the end of the
)

images with 8 bits per pixel, when the images are sorted sample data.
contaminated by fixed-value impulse noise, a noise pixel Then we compare Xi ) with each element of XI' . If
,
takes either a large value 255 or a small value O. The
Xi ) is the extreme pixel alone, then it is considered as noisy
n
corrupted pixel ',J is equal to 0 or 255 each with equal ,
probability O.5p. In general, R (Noise Ratio, O<R<J) is pixel, which will be removed. And if Xi ) is the extreme
,
employed to indicate the extent of the image corrupted by pixel but not alone, then it is regarded as noise candidate,
noise. which will be made further judgments. Otherwise, it is a
noise-free pixel, which will be preserved. The extreme test
III. THE PROPOSED ALGORITHM
can be expressed in the form:
In order to preserve the edges and fine details, the edges �".Xij max[Xr]...or",xiJ m n[Xr]
jNN : : � �
and the smooth regions should be detected from the extrema (2)
found by min-max algorithm. These details are considered as xj
i
E if",x
ij
- max[Xr]...or",x'J - mm[Xr],
noise free pixels, and are left unchanged. Only noisy pixel is S else;
replaced by filter output. The proposed algorithm takes these
where, N represents the corrupted pixel, N' denotes the
issues into account and the filtering processing summary is
noisy candidate, and S indicates the noise-free pixel.
shown in Figure 1.
,------,
B. Edge and smooth region detection
When the noise candidates (N) detected by (2), it means
more extreme pixels within the window and some detailed
pixels may be regarded as noise pixels. Consequently, we
must discriminate the detailed pixels from the noise
No candidates.
In order to identify the edges and fine detailed from the
noise candidates, the sub-window is introduced in the
proposed algorithm. In the sliding window can be divided
into four one-dimensional sub-window, i.e., horizontal,
vertical, main diagonal and auxiliary diagonal direction. In
each sub-window, the sum of absolute value of difference
No
between Xi ) and the other pixel is denoted as ml, m2, m3,
,
m4, respectively.
For example, we choose the 3 x 3 sliding window, the
ml, m2, m3, m4 can be expressed as following:
ml =12XiJ. -Xi,Fl -Xi,J+l I
l ; horizontal;

· -x 1+1·1
m2 =12xIJ -x 1,J ,j ; vertical; (3)
/-

m3 =12x . . 1,)+1 l ; main -diagonal;

. -x 1+
IJ. -x l-l,j-l
Figure I. The filtering algorithm processing summary. ; a iliary -diagonal;
. . l ux
. . -x l-l,j+l
m4 =12x IJ. -x l+l,J-I

V2-266

Authorized licensed use limited to: University of Malaya. Downloaded on March 09,2024 at 15:37:13 UTC from IEEE Xplore. Restrictions apply.
 2010 International Conference on Computer Application and System Modeling (ICCASM 2010)

Condition (4) denotes that m represents the minimum of where is the image size, Y i } and Si } are the
IXJ
mI, m2, m3 and m4. , ,
According to the structural characteristics of images, if restored and original pixels on the location (i, j) in the
Xi } is the pixel of the smooth region or edge, m should be image.
, Table I lists the PSNR values of restored images with
very little. Therefore, we employ T as a threshold, if m is p=JO%,20%,30%,40%,or 50% for "Lena" and "baboon"
greater than the threshold, the noise candidate will be images, respectively, where "Lena" is a smooth image and
considered corrupted pixel, as the following expression: "baboon" is a complicated image. The proposed method
{ N i/...m > T; performs the best PSNR in both the corrupted images.
x E (5) Figure 2 shows the corrupted "Lena" image with p=30%
lj S else; impulse noise and the corresponding restored images filtered
where T = 30 is the optimal threshold suggested in [3]. by MED, CWM, SWM, PV, EM, and the proposed
It is noted that, up to now, the impulse noise algorithm, respectively. It is observed from this figure that
classification is just to identifY the noise in the edges or in the proposed method efficiently suppresses the noise with
the smooth regions of the image. preserving image details when compared to other filters.

C. Impulse noise filter V. CONCLUSION

After the step A and B detection, the noisy pixels (N) and This paper has introduced a new image processing
noise-free pixels are identified. The proposed algorithm method to restore the image corrupted by impulse noise,
adopted a rank-ordered mean (ROM) filter[15] to remove the which employs the impulse noise detection, the edge and
noisy pixels. smooth region detection and the rank-ordered mean filter
According to step, Xr is the input samples be re­
A techniques. From the experimental results, we have seen that
arranged in the ascending order. Let ROM represents the the proposed filter can remove impulse noise, also
simultaneously preserve details at higher levels of noise and
median value of Xr and let Yi } denotes the output of
'
, has a much better performance than the previous filters.
filter which can be expressed in the form:
ACKNOWLEDGMENT
= {ROM i/",xi) E N;
Yi) (6) The work is supported by the Scientific Research
xi) 1if ' ' ' Xi) E S', Common Program of Beijing Municipal Commission of
As in the calculation of ROM, the center pixel is not Education, No. KM200810011008, and also supported by
taken into account, so the filter will be more robust. On the Beijing Science New Novel Fund, No. 2006Bl O.
other hand, only a few pixels are performed in order statistics,
REFERENCES
so the scheme reduces the computation greatly.
[ I] RCGonzalez and R.E.Woods, "Digital Image Processing". Prentice
Hall, 2002.
IV. EXPERIMENT AND ANALYSIS
[2] Ko.SJ, and LeeY.H, "Center weighted median filters and their
In this section, extensive experimental results with application to image enhancement". IEEE Trans. Circuits Syst.,
commonly used gray-scale test images of size 512 X 512 are 1991(38), pp:984-993.

presented to assess the performance of the proposed [3] ENG.H.L, and MA K.K, "Noise adaptive soft-switching median
algorithm. These images including "Lena", "Baboon" have filter", IEEE Trans. Image Process., 2001, 10, (2), pp. 242-252.

distinctly different features in terms of details. These images [4] P.S.Windyga, "Fast impulsive noise removal," IEEE Trans. Image
Processing, 2001,vol.l 0, pp. I73�179.
are corrupted by various levels of salt and pepper noise. In
the experimental the sliding window size M is set 3, i.e., 3 x3 [5] XING CJ, WANG SJ., DENG H.l "A New Filtering Algorithm
Based on Extremum and Median Value". Journal of Image and
sliding window for testing the detection. Graphics, 2001(6), pp:533-536.
The performance of the proposed method is compared [6] W.-YHan and l-CLin, "Minimum�maximum exclusive mean
with that of many other well-known algorithms, such as the (MMEM) filter to remove impulse noise from highly corrupted
MED filter[I], the CWM filter[2], the SWM filter[3], the PV images," Electron.Lett.,vo1.33,no.2, 1997, pp:124�125.
filter[4], and the EM filter [5] [7] lH.Wang, "Prescanned minmax center-weighted filters for image
The peak signal-to-noise ratio (PSNR) between the restoration", Proc.Inst.Elect.Eng., vol. I46,no.2,pp.101�107,1999.
restored image and the original image is selected as the [8] D.A.F.Florencio and RW.Schafer, "Decision-based median filter
performance validated in the experiments. The PSNR is using local signal statistics," Proc.SPIE, 1994, vo1.2308, pp.268�275.

defined as [9] T. Chen, H.R Wu, "Adaptive impulse detection using center­
weighted median filters," IEEE Signal Processing Letters, 200 I(1),
2552 (7) vo1.8, pp. I-3.
PSNR = 10 ·log -­
MSE [10] E.Abreu, M.Lightstone, S.K.Mitra, "A new efficient approach for the
Here MSE is the mean squared error defined as removal of impulse noise from highly corrupted images," IEEE Trans.
Image Processing, I 996,vo1.5, pp.1012�1025.
l I .I ( ) (8)
MSE = -LL Yl,j - sl,j 2 [ I I] Pok,l-CLiu and A.S.Nair, "Selective removal of impulse noise based
I . J i=! j=! on homogeneity level information," IEEE Trans. Image Processing,
2003,vol. I2, pp.85�92.

V2-267

Authorized licensed use limited to: University of Malaya. Downloaded on March 09,2024 at 15:37:13 UTC from IEEE Xplore. Restrictions apply.
 2010 International Conference on Computer Application and System Modeling (ICCASM 2010)

[12] V.Cmojevi,V. Senkand and Z.Trpovski, "Advanced impulse [14] Wang Jung-Hua, Lin Lian-Da. "Improved median filter using
detection based on pixel-wise MAD," IEEE Signal Processing Letters, minmax algorithm for image processing" . Electronics Letters. 1997,
2004,vol. l l ,pp.589-592. 33 (16), pp 1362-1363.
[13] W.Luo, "A new efficient impulse detection algorithm for the removal [15] Abreu E, Lightstone M, Mitra S K, et aI, "A new efficient approach
of impulse noise", IEICE Trans. Fundamentals, vo1.E88- for the removal of impulse noise from highly corrupted image", IEEE
A,no.10,pp.2579-2586,October 2005. Trans. Image Processing. 1996, 5(6), pp:1012-1025.

TABLE I. COMPARISONS FOR D IFFERENT F ILTERS

�
Lena Baboon
Algori th 10% 20% 30% 40% 50% 10% 20% 30% 40% 50%
MED[I] 32.05 28.41 23.21 18.87 15.35 25.39 22.52 21.59 20.65 19.36
CWM[2] 32.77 28.63 23.79 19.92 16.41 26.19 22.67 21.69 20.71 19.52
SWM[3] 33.16 29.72 27.97 26.05 23.43 27.87 23.16 22.11 21.35 19.74
PV [4] 32.21 27.29 23.86 19.48 16.36 25.37 22.55 21.64 20.83 19.43
EM[5] 33.93 30.65 28.91 26.98 24.63 28.48 23.43 22.25 21.57 19.96
The Proposed 34.18 30.87 29.22 27.73 25.70 29.25 24.49 23.21 21.92 20.38

Figure 2. Result Compare of Noise Removal with 30% Pepper-Salt Impulse Noise.

V2-268

Authorized licensed use limited to: University of Malaya. Downloaded on March 09,2024 at 15:37:13 UTC from IEEE Xplore. Restrictions apply.