IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 12, NO.

1, JANUARY 2003

85

Selective Removal of Impulse Noise Based on

Homogeneity Level Information
Gouchol Pok, Jyh-Charn Liu, and Attoor Sanju Nair

AbstractIn this paper, we propose a decision-based, signaladaptive median filtering algorithm for removal of impulse noise.
Our algorithm achieves accurate noise detection and high SNR
measures without smearing the fine details and edges in the image.
The notion of homogeneity level is defined for pixel values based
on their global and local statistical properties. The cooccurrence
matrix technique is used to represent the correlations between a
pixel and its neighbors, and to derive the upper and lower bound of
the homogeneity level. Noise detection is performed at two stages:
noise candidates are first selected using the homogeneity level, and
then a refining process follows to eliminate false detections. The
noise detection scheme does not use a quantitative decision measure, but uses qualitative structural information, and it is not subject to burdensome computations for optimization of the threshold
values. Empirical results indicate that our scheme performs significantly better than other median filters, in terms of noise suppression and detail preservation.

Fig. 1.

System flowchart of CSAM filtering.

TABLE I
IDENTIFICATION OF NOISE FOR THE LENA IMAGE

Index TermsAdaptive median filter, cooccurrence matrix, homogeneity level, impulse noise.

I. INTRODUCTION

EDIAN filter is a nonlinear filtering technique widely

used for removal of impulse noise [2], [6], [10]. Despite
its effectiveness in smoothing noise, the median filter tends to
remove fine details when it is applied to an image uniformly.
To address this drawback, a number of modified median filters have been proposed, e.g., minimummaximum exclusive
mean (MMEM) filter [7], prescanned minmax center-weighted
(PMCW) filter [14], and decision-based median filters [4], [5],
[8], [13]. In these methods, the filtering operation adapts to
the local properties and structures in the image. In the decision-based filtering, for example, image pixels are first classified as corrupted and uncorrupted, and then passed through the
median and identity filters, respectively. The main issue of the
decision-based filter lies in building a decision rule, or a noise
measure, that can discriminate the uncorrupted pixels from the
corrupted ones as precisely as possible.
In the method proposed by Han et al. [7], pixels that have
values close to the maximum and minimum in a filter window
of remaining pixels in the
are discarded, and the average
Manuscript received February 13, 2001; revised May 1, 2002. The associate
editor coordinating the review of this manuscript and approving it for publication was Prof. Uday B. Desai.
G. Pok is with the Department of Computer Science, Yanbian University,
Yanji, China (e-mail: gcpok@yahoo.co.kr).
J.-C. Liu is with the Department of Computer Science, Texas A&M University, College Station, TX 77843 USA (e-mail: liu@cs.tamu.edu).
A. S. Nair is with the Department of Electrical Engineering, Texas A&M
University, College Station, TX 77843 USA (e-mail: sanju@ee.tamu.edu).
Digital Object Identifier 10.1109/TIP.2002.804278

TABLE II
IDENTIFICATION OF NOISE FOR THE BRIDGE IMAGE

window are computed. If the difference between the center pixel

exceeds a threshold, the center pixel is replaced by ;
and
otherwise, unchanged. Florencio et al. [5] proposed a decision
measure, based on a second order statistic called normalized
deviation
DEV

(1)

where is the gray level of the center pixel , and and respectively are the mean and the standard deviation of gray levels
in the horizontal ( ) and vertical ( ) nine-point window around
. Without explicit classification of pixels, Wang [14] presented
an iterative, center-weighted filtering method [9], in which the
center weights and the window size adaptively increase with filtering iterations.

1057-7149/03$17.00 2003 IEEE

Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on November 9, 2009 at 02:20 from IEEE Xplore. Restrictions apply.

86

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 12, NO. 1, JANUARY 2003

(a)
Fig. 2.

(b)

Graphical representation of (a) cooccurrence matrix and (b) histogram P (I = 100; 1).

Fig. 3. Upper and lower bound of the PWD matrix shown in Fig. 2(a).

Kasparis et al. [8] presented an adaptive threshold selection method in which the threshold varies with the local
average intensities. Two thresholds, corresponding to the salt
and pepper components, are experimentally determined
to optimize the trade-off between noise removal and detail
preservation. In the adaptive scheme proposed by Sawant et al.
[13], the difference of the noisy and the median-filtered image,
and the difference of the noisy and the Gaussian-smoothed
image were computed. The ratio of the two differences was
used as a decision measure. Signal-dependent rank-order mean
(SDROM) filter [1] uses a framework in which the filtering
operation is conditioned on a state variable. The output is
defined as a weighted sum of the signal value and the ROM
value; the weighting coefficients being a function of the state
variable are optimized on image training data. In the two-state
approach, SDROM merely switches between the signal and
the ROM values: 1 or 0. The rank-ordered mean is computed

by rank ordering pixels in a 3

3 window and computing
where
is the set of
rank-ordered pixels.
Two issues are involved in developing a decision process.
Firstly, a decision measure should be defined, as a statistical parameter, to capture and represent the local property of the region.
Secondly, a mechanism to compute a threshold value should be
determined. If a decision measure does not reflect the local SNR
measure, the performance of the decision-based filter depends
on the values of a threshold, and thus one needs to optimize
the thresholding. As observed in existing decision-based filters,
however, threshold values optimized for high PSNR measures
can lead to poor noise detection.
In this paper, we propose a decision-based, signal-adaptive
median filter that achieves nearly perfect detection of impulse
noise, and superb visual quality for the restoration results.
For identification of the noise, we propose the notion of
homogeneity level, which is defined on the basis of the pairwise
correlations between neighboring image pixels. The pairwise
correlations are obtained by the cooccurrence matrix. Given a
and
pixel value , we compute the upper and lower bounds,
for acceptable signal deviations from . In a filter window,
the number of pixels that have the same and different homogeneity level classifications as the center pixel are denoted as
and , respectively. The center pixel is considered a noise
and , and
.
candidate if it is not in the range of
Further steps will be applied to noise candidates to differentiate
noise and signal using the local image structure. Finally, the
median filter is applied only to the noisy pixels. Our scheme
is highly accurate and robust in noise detection. Experimental
results show that our scheme significantly outperforms other
techniques for both objective (SNR) and subjective (visual
perception) performance measures. The proposed scheme does
not amplify the noise through iterations, making it a highly

Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on November 9, 2009 at 02:20 from IEEE Xplore. Restrictions apply.

POK et al.: SELECTIVE REMOVAL OF IMPULSE NOISE BASED ON HOMOGENEITY LEVEL INFORMATION

Fig. 4.

87

(a)

(b)

(c)

(d)

Noise identification. (a) Lena image corrupted by 10% impulse noise; false detections at (b) first iteration; (c) second iteration; and (d) third iteration.

stable solution approach for noise filtering, and other similar

applications. In the rest of this paper, we present the proposed
algorithm in Section II, the experimental results in Section III,
and summarize the paper in Section IV.
II. CONDITIONAL SIGNAL-ADAPTIVE MEDIAN (CSAM) FILTER
A decision-based median filter consists of two major functions: decision making and noise filtering. The first function
is to determine presence of noise in a filter window, and the
second to smooth the noise pixels. When necessary, as in our
scheme, detection and filtering can be iteratively applied to the
input image. The operation of median filters is straight-forward,
that is, noise is replaced with the median of the pixels in the
filter window. On the other hand, design of the decision-making
process is much more complicated, since two pertinent issues,

defining a decision measure and searching for an optimal

threshold value, must be adequately addressed to achieve
robust noise detection. If a decision measure is defined so that
the computation of a threshold does not involve the center pixel
value, we have experimentally observed that a lot of noise were
not detected with a high rate of false detection (see Tables I
and II). The conditional signal-adaptive median (CSAM) filter,
shown in Fig. 1, addresses these problems with a simple and
robust decision-making scheme that can iteratively detect
corrupted image pixels, with little visual quality loss, and at
very high SNRs.
A. Filter Design
A signal-adaptive filter adapts its operation in accordance
with local image characteristics. To represent local image char-

Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on November 9, 2009 at 02:20 from IEEE Xplore. Restrictions apply.

88

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 12, NO. 1, JANUARY 2003

(a)

(b)

Fig. 5. False detection of noise for the Lena image with Florencios method [5]. (a) 20% and (b) 30% impulse noise.

acteristics, we introduce the notion of homogeneity level that

defines the acceptable range of deviations from a pixel value.
Noise candidates are first determined using the homogeneity
level, and then the candidates are further classified into signal
and noise based on the structural information. We present details of the CSAM filtering algorithm next as shown in Fig. 1.
Step 1) Calculation of the bounds of homogeneity level:
for small
For a pixel value , some deviations from (i.e.,
) can be considered as homogeneous with . Homogeneity
level associated with is represented by an interval with the
and lower bound , so that any in
upper bound
is considered homogeneous with . More details on calculating
the bounds are described in the next subsection.
Step 2) Detection of impulse noise candidates: In a 3 3
window, we denote the center pixel as , and its eight neighbors
. Let the value of
be . For
, let
be the number of s that are homogeneous with , that is,
, and be the number of the ones that are not. The
is simple: if
impulse noise detection decision rule for
, is counted as signal, and otherwise as noise candidate. Let
denote the set of noise candidate pixels, and be the set of
,
remaining pixels, which are mostly signals, we have
where is the corrupted input image.
Step 3) Refined selection of impulse noise: usually contains some falsely detected signals. To minimize the false detections, we use the following step to remove noise-free pixels from
the noise candidate set , using a different filtering rule. These
falsely detected image pixels mostly locate around the edges and
fine details, e.g., see Figs. 4(b) and 6(b), and they can be characterized as high-frequency signals with low SNR measure. Based
on this observation, we augment the analysis window to 5 5,
so that edge information can be better presented in the analysis
window. Then, pixels are divided into two groups; those homogeneous with , denoted by , and the remaining pixels. Now,
, let be the number of s that belong to , and
for all
be the number of those that belong to . If
,
is

counted as signal and its membership is changed from to .

does not
This step is repeated until the number of pixels in
decrease further. We note that this step is highly computational
efficient, because it only applies to pixels in , which constitute of a small portion of the total image pixels.
Step 4) Median filtering: We define the notion of an -neighborhood isolated pixel, which is essentially a noisy pixel with
noisy pixels in its 3 3 kernel. To remove noise effectively,
, 1, 2, are identified
all -neighborhood isolated pixels, for
and a median filter of window size 3 is applied to these pixels. A
median filter of window size 5 is applied to the remaining noisy
pixels. The choice of a window of size 5 for the later stage of
median filtering is based on the fact that a median filter of size
3 will not be effective when there are more noisy pixels in the
3-neighborhood of the pixel in question.
CSAM has highly robust, accurate performance in impulse
noise detection. In our experiments, whose results are illustrated
in Tables I and II, noise detection is nearly perfect for images
corrupted by up to 20% to 30% of impulse noise. We further
note that perfect identification of noise is essentially to an iterative filtering process. Otherwise, iterative filtering tends to
amplify the noise effect and result in smeared image details
[14]. Some intuitive observations may help understanding the
behavior of CSAM. The first step is to calculate the acceptable
natural/normal dynamic ranges of image pixels of natural images, i.e., the upper and lower bounds of the homogeneity levels
for each pixel value of the image. We note that this is the global
distribution of acceptable pixel values for any pixel location.
Step 2) not only picks up isolated impulse noise, but also likely
some edge points. Then, in the Step 3), by using a larger window
we have a much better chance to identify the pixels located near
or on the edges. Obviously, we will need to use a new decision
rule to differentiate noise from normal signal. The iterative algorithm helps to correct false detections in each run. Such an
observation, although not a formal proof, could be supported by
the simulation results shown in Fig. 4(b) and (c).

Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on November 9, 2009 at 02:20 from IEEE Xplore. Restrictions apply.

POK et al.: SELECTIVE REMOVAL OF IMPULSE NOISE BASED ON HOMOGENEITY LEVEL INFORMATION

Fig. 6.

89

(a)

(b)

(c)

(d)

Noise identification. (a) Lena image corrupted by 30% impulse noise; false detections at (b) first iteration; (c) second iteration; and (d) third iteration.

B. Computing the Homogeneity Levels

The homogeneity level of a pixel value is defined as the acceptable upper and lower bounds of deviations in a noise-free
region. This is based on the simple observation on the fact that
natural images can be well characterized by the Gaussian distribution. Impulse noise can be characterized as extremely large
or small pixel values at isolated random locations, and thus they
likely would locate beyond the homogeneity levels, except for
certain sharp image features like edges. The two bounds are estimated from the global statistics of pairwise correlations between every pixel and its neighbors, by using the cooccurrence
image under consideration has
matrix. We assume that
. In a 3
8-bit gray scale pixel resolution, that is,
3 window
at
, the center pixel is denoted as
,

and its neighbors as

is defined as
with respect to

where

if

. The cooccurrence matrix

and

otherwise.

(2)

Note that and take integer values ranging from 0 to 255,

represents a histogram. The histogram
and for each ,
contains the global statistics about the pairwise correlations between a pixel and its neighbors, and from these statistics we can
and
for
estimate the homogeneity levels. To determine

Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on November 9, 2009 at 02:20 from IEEE Xplore. Restrictions apply.

90

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 12, NO. 1, JANUARY 2003

(a)

(b)

Fig. 7. False detection of noise for the Lena image with Kaspariss method (CMF) [8]. (a) 20% and (b) 30% impulse noise.

each , we first replace histogram entries for

and
,
to avoid biased estimation of the global statistics. Then, we employ a simple technique that thresholds the moving averaged
values in the tail regions as follows:

TABLE III
COMPARATIVE RESTORATION RESULTS IN PSNR

(3)

(4)

where is a threshold. These equations simply search for the

histogram slot where the 3-point moving average is less than ,
which indicates that most slots around it have small values. For
most cases, a moving average window size of 3 and a threshold
gave good results. The cooccurrence matrix obsize of
tained from the 10% noisy Lena image is graphically shown
in Fig. 2(a), in which the grayscale intensity of the each pixel
location ( , ) represents the normalized occurrence count for
, and its neighbors being . By normalization
the value
we mean that we divided each count value by the maximum
of all, and multiple the result by 255. The calculated value is
, we
then printed as a pixel point on this graph. For
to 255 in
further plotted the cooccurrence counts for
Fig. 2(b). One can see from these graphs that natural images
have high locality, making it possible to differentiate them from
and
for
the sharply isolated impulse noise. The bounds
corresponding to the cooccurrence matrix in
all
Fig. 2(a) is shown in Fig. 3.

performance measures: false noise detection rate, quantitative

measure such as PSNR and MAE, and perceptual gains.
For the experiments, 8-bit 512 512 images, that are corrupted by 10%, 20%, and 30% of impulse noise, are used. The
impulse noise is randomly generated as 0 or 255 with equal
probability for all the pixels. A number of simulations were performed by varying the values of the parameters of each method,
and only the best result of each compared method is described
below. As performance measures, we used the peak signal-tonoise ratio (PSNR) and the mean absolute error (MAE) defined
as

PSNR

MAE

III. EXPERIMENTAL RESULTS

In this section, the proposed algorithm is evaluated and compared with other filtering techniques in terms of a number of

where

is the total number of pixels in an image.

Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on November 9, 2009 at 02:20 from IEEE Xplore. Restrictions apply.

POK et al.: SELECTIVE REMOVAL OF IMPULSE NOISE BASED ON HOMOGENEITY LEVEL INFORMATION

91

(a)

(b)

(c)

(d)

Fig. 8. Restoration results for (a) 10% and (b) 30% impulse noise Lena image, and (c) 10% and (d) 30% impulse noise bridge image.

One important characteristic of the CSAM filter is that it can

detect almost all noise while false noise detection is very low,
as illustrated in Figs. 47.
In Tables I and II, miss refers to number of noisy pixels that
were not detected, and false-hit refers to the number of signal
pixels misclassified as noise. Compared with other techniques,
the CSAM filter performs better in identifying the noisy pixels.
It makes an iterative filtering scheme operate properly without
the noise amplification effect.
The CSAM filter also outperforms other techniques for the
performance measured in PSNR and the perceptual quality. Referring to Table III, the proposed CSAM filter achieved signif-

icant improvement over other techniques, ranging from 2.3 dB

to 8.3 dB in PSNR.
IV. CONCLUSION
In this paper, we presented a new framework for the removal
of impulse noise in which filtering operation is selectively applied to the pixels that are classified as corrupted. The novel
feature of the algorithm lies in introducing the notion of homogeneity level that is used for computing the signal-adaptive and
flexible thresholds. In addition to that, structural information is
also applied to maximize the true detection rate. We have shown
experimentally that the proposed algorithm significantly outper-

Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on November 9, 2009 at 02:20 from IEEE Xplore. Restrictions apply.

92

IEEE TRANSACTIONS ON IMAGE PROCESSING, VOL. 12, NO. 1, JANUARY 2003

(a)

(b)

Fig. 9. Restoration results for (a) 10% and (b) 30% impulse noise Lena image for SDROM filter.

form a number of existing techniques. The gains are obtained in

terms of the perceptual quality and the PSNR of the restoration
results (see Figs. 8 and 9).

Gouchol Pok received the B.Sc. degree in mathematics from Yonsei University, Korea, in 1981, and
the Ph.D. degree in computer science from Texas
A&M University, College Station, in 1999.
During 1984-1995, he was with Korea Development Institute (KDI), and during 1999-2000 he was a
Postdoctoral Research Associate at Texas A&M University. Since 2001, he has been with Yanbian University of Science and Technology, Yanji, China, where
he is now an Associate Professor. His research interests include image processing, pattern recognition,

REFERENCES
[1] E. Abreu, M. Lightstone, S. K. Mitra, and K. Arakawa, A new efficient approach for the removal of impulse noise from highly corrupted
images, IEEE Trans. Image Processing, vol. 5, pp. 10121025, June
1996.
[2] G. R. Arce and M. P. McLoughlin, Theoretical analysis of the
MAX/Median filter, IEEE Trans. Acoust., Speech, Signal Processing,
vol. ASSP-35, pp. 6069, Jan. 1987.
[3] G. R. Arce and R. E. Foster, Detail-preserving ranked-order based filters for image processing, IEEE Trans. Acoust., Speech, Signal Processing, vol. ASSP-37, pp. 8398, Jan. 1987.
[4] C.-T. Chen and L.-G. Chen, A self-adjusting weighted median filter for
removing impulse noise in image, in Proc. IEEE Int. Conf. on Image
Processing, 1998, pp. 419422.
[5] D. A. F. Florencio and R. W. Schafer, Decision-based median filter
using local signal statistics, Proc. SPIE, vol. 2308, pp. 268275, 1994.
[6] M. Gabbouj, E. J. Coyle, J. Neal, and C. Gallagher, An overview of
median and stack filtering, Circuits Syst. Signal Process., vol. 11, pp.
745, 1992.
[7] W.-Y. Han and J.-C. Lin, Minimummaximum exclusive mean
(MMEM) filter to remove impulse noise from highly corrupted
images, Electron. Lett., vol. 33, no. 2, pp. 124125, 1997.
[8] T. Kasparis, N. S. Tzannes, and Q. Chen, Detail-preserving adaptive
conditional median filters, J. Electron. Imag., vol. 1, no. 14, pp.
358364, 1992.
[9] S. J. Ko and Y. H. Lee, Center weighted median filters and their applications to image enhancement, IEEE Trans. Circuits Syst., vol. 38, pp.
984993, Sept. 1991.
[10] Y. H. Lee and S. Tantaratana, Decision-based order statistics filters,
IEEE Trans. Acoust., Speech, Signal Processing, vol. 38, pp. 406420,
Mar. 1990.
[11] G. Qiu, Functional optimization properties of median filtering, IEEE
Signal Processing Lett., vol. 1, pp. 6465, Apr. 1994.
[12]
, An improved recursive median filtering scheme for image processing, IEEE Trans. Image Processing, vol. 5, pp. 646648, Apr. 1996.
[13] A. Sawant, H. Zeman, D. Muratore, S. Samant, and F. DiBianka, An
adaptive median filter algorithm to remove impulse noise in X-ray and
CT images and speckle in ultrasound images, Proc. SPIE, vol. 3661,
pp. 12631274, Feb. 1999.
[14] J.-H. Wang, Prescanned minmax center-weighted filters for image
restoration, Proc. Inst. Elect. Eng., vol. 146, no. 2, pp. 101107, 1999.

and neural networks.

Jyh-Charn Liu received thes Ph.D. degree from

the Electrical Engineering and Computer Science
Department, University of Michigan, Ann Arbor, in
1989.
He is Associate Professor with the Computer
Science Department, Texas A&M University,
College Station. He joined Texas A&M University
as Assistant Professor in 1989, and became tenured
Associate Professor 1995. He specializes in the area
of image processing, real-time computing, and network security. He is also involved with development
of large scale information systems, telemedicine technologies, and imagery
database. He is actively involved in different professional activities, conference
organizations, and he has extensive outreach to professional communities.

Attoor Sanju Nair received the undergraduate degree in electronics and communication engineering
from PSG College of Technology, India, and is currently pursuing the M.S. degree in electrical engineering at Texas A&M University, College Station.
His research interests include pattern recognition
and image processing.

Authorized licensed use limited to: MALAVIYA NATIONAL INSTITUTE OF TECHNOLOGY. Downloaded on November 9, 2009 at 02:20 from IEEE Xplore. Restrictions apply.