0% found this document useful (0 votes)
14 views4 pages

A New Adaptive Switching Median Filter: Smaïl Akkoul, Roger Lédée, Remy Leconge, and Rachid Harba

Download as pdf or txt
Download as pdf or txt
Download as pdf or txt
You are on page 1/ 4

IEEE SIGNAL PROCESSING LETTERS 587

A New Adaptive Switching Median Filter


Smaïl Akkoul, Roger Lédée, Remy Leconge, and Rachid Harba

Abstract—A new Adaptive Switching Median (ASWM) filter for Hence, they cannot preserve the image details and edges,
removing impulse noise from corrupted images is presented. The especially when the noise is high.
originality of ASWM is that no a priori Threshold is needed as in To reduce these effects and improve impulse noise detection,
the case of a classical Switching Median filter. Instead, Threshold
is computed locally from image pixels intensity values in a sliding we propose a new Adaptive Switching Median (ASWM) filter.
window. Results show that ASWM provides better performance in The originality of ASWM is that no a priori Threshold is to
terms of PSNR and MAE than many other median filter variants be given as in the case of a classical SWM filter. Instead, the
for random-valued impulse noise. In addition it can preserve more threshold is computed locally from image pixels intensity values
image details in a high noise environment. in a sliding window using weighted statistics. Thus, it is ex-
Index Terms—Detail-preserving, image restoration, impulse pected that this new strategy will lead to better performances.
noise detection, switching median filter. The ASWM filter will be described and compared to other me-
dian filters using Monte-Carlo simulations.
This letter is organized as follows. In Section II, we review
I. INTRODUCTION
the impulse noise removal principle using SWM. The ASWM
algorithm is described in Section III. Section IV gives simula-
MAGES are frequently corrupted by impulse noise due to
I camera sensors or transmission in noisy channels [1]. It oc-
curs in bioluminescence imaging for which the image acquisi-
tion results using different test images to demonstrate the per-
formances of the new approach. Finally, conclusions are noted
in Section V.
tion is disturbed by the presence of cosmic noise [2]. Cosmic
noise can be considered to be an impulse noise. For this appli- II. SWITCHING MEDIAN FILTER (SWM)
cation in bioluminescence imaging, a step of filtering is needed The principle of the SWM filter is reminded. A SWM filter
as a preprocessing step before the deconvolution task [2]. is a two steps procedure. First, a test decides whether or not a
The main approach for removing impulse noise is to use me- given pixel is contaminated by impulse noise: a pixel is contam-
dian-based filters (see, e.g., [1], [3]). However, since filters are inated if the absolute difference between the median value in its
usually implemented identically across the images, they tend to neighborhood and the value of the current pixel itself is greater
modify both noise and noise-free pixels. Consequently, some than a given Threshold [6]. If contaminated, a classical median
desirable details can be removed [1]. To overcome this problem, fitter is applied; if not, the current pixel is noise free and will be
many modified forms of median filters were proposed among not modified. More precisely, consider an image corrupted
which are the weighted median filter [4] and the center weighted by an impulse noise and the grey level value at position , .
median (CWM) filter [5]. Those filters tend to work well with Let be a square window surrounding this pixel. This window
low noise level, but poorly for highly corrupted images. To solve is of size where is an integer greater than
this problem, the switching median (SWM) filter [6] was intro- zero. The output of the switching median filter is given by
duced. The main idea of the SWM is to use an impulse detector
if ,
before filtering. This detector is based on an a priori Threshold (1)
otherwise
value to decide if a median filter is to be applied or not. Next,
many other approaches were proposed such as tri-state median where is the median value in the window and Threshold
(TSM) filter [7] and more recently, alpha-trimmed mean-based is a fixed parameter.
approach (ATMA) [8], directional weighted median (DWM) The numerical Threshold value is defined a priori or chosen
filter [9], and modified switching median (MSWM) filter [10]. after many data dependant tests. The literature shows that an
All of these filters usually perform well in terms of Peak optimal threshold in the sense of the mean square error can be
Signal-to-Noise Ratio (PSNR) and Mean Absolute Error obtained for most real data [6], [7]. However, Threshold suitable
(MAE). However, the principle drawback is that they have for a particular image is not necessarily adapted to another one.
limited performance in terms of false and missed detections. In addition, an image is often non stationary and statistics in a
region may be different in other part of the image.
Manuscript received February 03, 2010; revised April 02, 2010. Date of pub- The next section presents the new method that does not
lication April 19, 2010; date of current version May 05, 2010. The associate require a priori knowledge and automatically defines a local
editor coordinating the review of this manuscript and approving it for publica- Threshold.
tion was Prof. Carlo S. Regazzoni.
The authors are with the Institut Pluridisciplinaire de Recherche en Ingénierie
des Systèmes, Mécanique, Energétique (PRISME), Université d’Orléans, III. ADAPTIVE SWM FILTER (ASWM)
BP 6744, 45067 Orléans cedex2, France (e-mail: smail.akkoul@univ-or- The proposed method has the same general structure as the
leans.fr; roger.ledee@univ-orleans.fr; remy.leconge@univ-orleans.fr and
rachid.harba@univ-orleans.fr). SWM filter. The difference between the new method and SWM
Digital Object Identifier 10.1109/LSP.2010.2048646 relies on the fact that Threshold is not an a priori choice but is
1070-9908/$26.00 © 2010 IEEE

Authorized licensed use limited to: Constantin Vertan. Downloaded on October 07,2020 at 11:56:07 UTC from IEEE Xplore. Restrictions apply.
588 IEEE SIGNAL PROCESSING LETTERS

computed locally from image pixels. For that reason, we called


it the Adaptive SWM (ASWM) filter.
More precisely, the weighted mean value and the weighted
standard deviation are estimated in the current window. The
weights are the inverse of the distance between the weighted
mean value of pixels in a given window and the considered pixel.
A result is that impulse noise does not corrupt the determination
of these statistics from which the Threshold is derived.
In each window, the weighted mean are first iteratively esti-
mated. Then the weighted standard deviation is calculated and
the Threshold is determined. This procedure is explained in the Fig. 1. Values of  for Lena image degraded respectively by (a) 30% of
following. random-value impulsive noise and (b) 60% of random-value impulsive noise.
Initialization: compute the weighted mean value in a
window surrounding the current where is the initial parameter and the parameter in the
pixel: th step.
In Fig. 1, we show the values of across the “Lena” image
(2) degraded by (a) 30% and (b) 60% of random-value impulsive
noise respectively. Here, we have chosen a 3 3 square window,
, and 6 iterations were used (that means
where is the grey level of the image at pixel location and [see (6)].
and the weights. These weights are all equal to From the images of Fig. 1, we observed that is quite dif-
1 at Initialization. The index variation of and are in ferent depending on its location in the image: while is low
. in uniform regions, it shows high values in textured regions or
Step 1: estimate the weights as in presence of edges. This adaptability is an important point to
preserve image details. Besides, the noise percentage does not
(3) change this general behavior.

, a given small value, avoids possible division by zero. IV. RESULTS


Then, a new weighted mean value is obtained In this section, the restoration property, the noise detection
using (2). capability of ASWM and the visual performances are evalu-
Step 2: if , where is a ated and compared to a number of existing median-based fil-
given small value, then stop, else go to step 1 ( is ters used to remove random-valued impulse noise. Commonly,
the weighted mean value at iteration number ). most authors use the peak signal-to-noise ratio (PSNR) and the
End. mean absolute error (MAE) to quantify the restoration results.
Next, the weighted standard deviation is defined as To complete comparisons, authors of [9] compute the number
of missed noisy pixels and the number of noise-free pixels that
are identified as noise to show the efficiency of their method. In
(4) the same aim, authors of [10] defined a noise detection rate. We
will present such results in following Sections IV-A–IV-C.

A. Restoration Performance Measurements


Finally, ASWM can be summarized as follows:
a) Compute the weighted mean and the weighted Restoration performances are evaluated quantitatively by
standard deviation in the using PSNR and MAE, which are defined as in [3]. We com-
window surrounding the current pixel as described above. pare ASWM to other well known median-based filters, which
b) Use the following rule: include the standard median SM [1] (with a 3 3 filtering
window if noise percentage , and a 5 5 window
if ,
(5) otherwise), CWM filter [5] ( ), SWM filter [6] ( ),
otherwise
TSM filter [7] ( ), MSWM filter [10] ( ,
where is the median in the window , is a given and ), ATMA filter [8] ( , , ,
parameter and represents the local threshold. , , and ), and
ASWM is applied recursively and iteratively. During the iter- DWM filter [9] (a 5 5 filtering window, , and
ations, in each window, the threshold is decreased as proposed ). For ASWM filter, we have
by Dong [9]. This is done by varying the value. From sim- , , and . For
ulations conducted on a broad variety of images, the following all tested methods, a 3 3 filtering window is used, unless
strategy for yields satisfactory results, that is mentioned otherwise.
As test images we adopted the well known images “Lena,”
(6) “Boat,” “Peppers,” “Goldhill,” and “Bridge.” We have applied

Authorized licensed use limited to: Constantin Vertan. Downloaded on October 07,2020 at 11:56:07 UTC from IEEE Xplore. Restrictions apply.
AKKOUL et al.: NEW ADAPTIVE SWITCHING MEDIAN FILTER 589

TABLE I TABLE II
PSNR (DB) AND MAE FOR A PROBABILITY OF 30% RANDOM-VALUED THE NOISE DETECTION COMPARISON RESULTS FOR THE IMAGE
IMPULSIVE NOISEBEST RESULTS ARE BOLD PRINTED. “LENA” CORRUPTED BY RANDOM-VALUED IMPULSIVE NOISE.
BEST RESULTS ARE BOLD PRINTED

“Miss” term means a noisy pixel which is not detected as noise


and “False” term means a noise free pixel detected as noise.
For random-valued impulse noise, the noisy pixel values may
Fig. 2. Performance comparison of different methods for filtering the Peppers not be so different from those of their neighbors. Therefore it is
image degraded by various levels of random-value impulsive noise. more likely for a noise detector to miss a noisy pixel or detect a
noise-free pixel as noise [9], [11]. A good noise detector should
Monte Carlo simulations. Each image is 100 times degraded by be able to identify most of the noisy pixels. Its false alarm rate
a random impulse noise. The obtained mean PSNR and mean should be as small as possible.
MAE are reported in Table I. Results for ASWM are of high quality. ASWM can still dis-
Results show that ASWM performs better than other consid- tinguish most of the noisy pixels, even when the noise level is
ered methods and achieved the image quality with best PSNR as high as 60%.
and MAE for random valued impulse noise.
The performances of ASWM and other considered median- C. Visual Performances
based filters for “Peppers” image in term of PSNR for random As a final illustration and in order to compare subjectively the
valued impulse noise with different noise densities are reported methods, we give in Fig. 3, the “Goldhill” image with a 30%
in Fig. 2. random value impulse noise restored by various methods.
Results show that ASWM performs well for the tested range ASWM exhibit excellent psycho-visual performances com-
data corrupted with various noise percentages up to 60%. Re- pared to other methods. Especially, the sketches of the houses
sults are similar in term of MAE. This confirms that ASWM are well restored using ASWM. This result is of high impor-
achieves good restoration in all range of noise percentage. tance for impulse noise removal.

B. Noise Detection Performance Measurements V. CONCLUSION


Here, we compare ASWM method with five recently pro- This letter proposes a new based switching median filter,
posed methods. Table II lists the number of missed noisy pixels. called adaptive switching median (ASWM) filter. ASWM

Authorized licensed use limited to: Constantin Vertan. Downloaded on October 07,2020 at 11:56:07 UTC from IEEE Xplore. Restrictions apply.
590 IEEE SIGNAL PROCESSING LETTERS

does not need an a priori Threshold as in the case in clas-


sical Switching Median filter to detect noisy pixels. Instead,
Threshold is computed locally from image pixels grey values
in a sliding window.
ASWM has shown high noise detection ability. Extensive
simulations results indicate that ASWM performs significantly
better than many other existing techniques. In addition, psycho-
visual results are of high quality.
Finally, ASWM will be used as pre processing to remove
cosmic noise for an application in bioluminescence imaging [2].

REFERENCES

[1] R. C. Gonzalez and R. E. Woods, Digital Image Processing. Engle-


wood Cliffs, NJ: Prentice-Hall, 2002.
[2] S. Akkoul, R. Lédée, R. Leconge, C. Léger, R. Harba, S. Pesnel, S.
Lerondel, A. Lepape, and L. Vilcahuaman, “Comparison of image
restoration methods for bioluminescence imaging,” in ICISP 08,
Cherbourg, France, 2008, vol. 5099, LNCS, pp. 163–172.
[3] A. Bovik, Handbook of Image and Video Processing. New York:
Academic Press, 2000.
[4] D. Brownrigg, “The weighted median filter,” Commun. Assoc. Comput.
Mach., vol. 27, pp. 807–818, Mar. 1984.
[5] S. J. Ko and Y. H. Lee, “Center weighted median filters and their ap-
plications to image enhancement,” IEEE Trans. Circuits Syst., vol. 38,
pp. 984–993, 1991.
[6] T. Sun and Y. Neuvo, “Detail preserving median based filters in image
processing,” Pattern Recognit. Lett., vol. 15, pp. 341–347, 1994.
[7] T. Chen, K. K. Ma, and L. H. Chen, “Tri-state median filter for image
denoising,” IEEE Trans. Image Process., vol. 8, pp. 1834–1838, Dec.
1999.
[8] W. Luo, “An efficient detail-preserving approach for removing impulse
noise in images,” IEEE Signal Process. Lett., vol. 13, pp. 413–417, Jul.
2006.
[9] Y. Dong and S. Xu, “A new directional weighted median filter for re-
moval of random-value impulse noise,” IEEE Signal Process. Lett., vol.
14, pp. 193–196, Mar. 2007.
[10] C. C. Kang and W. J. Wang, “Modified switching median filter with
one more noise detector for impulse noise removal,” Int. J. Electron.
Commun., no. DOI: 10.1016/j.aeue.2008.08.009, 2008.
[11] Y. Dong, R. H. Chan, and S. Xu, “A detection statistic for
Fig. 3. Restoration performance comparison on the “Goldhill” image degraded random-valued impulse noise,” IEEE Trans. Image Process., vol.
by 30% random-value impulsive noise. 16, pp. 1112–1120, Apr. 2007.

Authorized licensed use limited to: Constantin Vertan. Downloaded on October 07,2020 at 11:56:07 UTC from IEEE Xplore. Restrictions apply.

You might also like