Received September 25, 2019, accepted October 18, 2019, date of publication October 29, 2019,

date of current version November 12, 2019.

Digital Object Identifier 10.1109/ACCESS.2019.2950348

Adaptive Sequentially Weighted Median Filter for

Image Highly Corrupted by Impulse Noise
JIAYI CHEN 1, YINWEI ZHAN2 , (Member, IEEE), AND HUIYING CAO1
1 School of Information Engineering, Guangdong Medical University, Zhanjiang 524023, China
2 School of Computer Science and Technology, Guangdong University of Technology, Guangzhou 510006, China

Corresponding author: Yinwei Zhan ([email protected])

This work was supported in part by the National Natural Science Foundation of China under Grant 61170320, in part by the Guangdong
Provincial Science and Technology Program in China under Grant 2017B010110015, in part by the Project of Science and Technology
Program of Guangzhou in China under Grant 201604016034, and in part by the Medical Scientific Research Foundation of Guangdong
Province in China under Grant B2018190.

ABSTRACT To tackle the difficulties in the detection and removal of impulse noise faced by the existing
filters, and to further improve the denoising performance, we propose an adaptive sequentially weighted
median filter for image corrupted by impulse noise. In the proposed method, a noise detector employing the
3σ principle of normal distribution and the local intensity statistics, is proposed; and a sequentially weighted
median filter with a neighborhood of adaptive size, is proposed for noise removal, in which the weighted
operator is derived in reference to the spatial distances from central noisy pixel, i.e., the weighting coefficients
are sequentially inversely proportional to the spatial distances. The experimental results confirm that the
proposed method outperforms the existing filters, excelling in the capability of noise removal, structure and
edge information preservation.

INDEX TERMS Image denoising, median filter, noise detection, noise removal, sequentially weighted
median filter, 3σ principle.

I. INTRODUCTION switching median filters [4]–[7] that integrate the noise

An image is often corrupted by impulse noise in the process removal processing with a noise detector so as to make the
of acquisition and transmission; and there are two types removal processing imposed only on the detected noisy pixels
of impulse noise: fixed-valued impulse noise and random- so that the performance of median filter was improved consid-
valued impulse noise [1]. Fixed-valued impulse noise is erably. And in the wake of development of image processing,
also called salt and pepper noise, one most common noise analysis, and application, the better denoising performance
in images; it severely impacts the image processing and of filters is highly demanded; thus, various improved filters
analysis, such as image recognition, segmentation, and so integrated with various strategies were proposed. However,
on. Therefore, effective removal of impulse noise is highly the existing filters inevitably have inherent shortcomings,
needed. For removal of fixed-valued impulse noise, the mean and are not necessarily effective, especially for high density
filter [2] and median filter [3] were originally proposed. noise: they either overly smooth the image, or are unable to
However, mean filter was found unable to preserve the struc- restore effectively the structure and edge information, so that
ture and edge information of image, while median filter is they still could not satisfy the high requirements of image
preferred because of its simple processing and good per- analysis and application. To tackle this problem and provide
formance. But thereafter, the traditional median filter was high quality image for analysis and application, we proposed
found unable to obtain a thorough noise removal and structure an adaptive sequentially weighted median filter (ASWMF)
information preservation simultaneously, especially for high for image highly corrupted by impulse noise; the contri-
density noise, because it processes all pixels regardless of butions of the proposed ASWMF are briefly described as
whether they are noisy or not, destroying the noise free pixels. follows.
To address this problem, some researchers initially proposed (i) A noise detector employing the 3σ principle of normal
distribution and the local intensity statistics based on the
The associate editor coordinating the review of this manuscript and intensity distribution of natural image, taking full advantage
approving it for publication was Halil Ersin Soken . of intensity distribution features to discriminate accurately

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see http://creativecommons.org/licenses/by/4.0/
VOLUME 7, 2019 158545
 J. Chen et al.: Adaptive Sequentially Weighted Median Filter for Image Highly Corrupted by Impulse Noise

the noisy pixels from the noise free ones having the same results, especially for high density noise, such as [14], [15].
intensity. In [15], the intensity estimation of noisy pixel is performed
(ii) A noise removal method employing sequentially by an adaptive weighted mean based on Euler distance; it is
weighted median of neighborhood of adaptive size; the claimed by the authors that this filter can achieve excellent
weighted operator employed is derived in reference to the noise removal and good edge preservation. Apart from this,
spatial distances from central noisy pixel, in which the mean filter integrated with median filter may achieve better
weighting coefficients are sequentially inversely proportional performance, such as [16]–[18]. In [18], the filter using pixel-
to the spatial distances, distinguishing accurately the various variation gain factors (PVGF) groups the neighbor pixels
contributions and impacts of neighbor pixels on the central having non-extreme intensity according to the intensity vari-
noisy pixel according to the distances. ation, and then, the distribution ratio and pixel variation level
The rest of this paper is organized by several parts as of each group are employed to determine the gain factors;
follows. Section II states the related works, followed by the thereafter, the value obtained by the gain factors multiplied
proposed method detailed in section III; the experiments and with the median of each group is taken as the estimated
result analyses are conducted in section IV; and section V intensity of noisy pixel.
concludes this paper. As improved versions of switching median or mean filter,
weighted filters [19]–[24] remove impulse noise by taking
II. RELATED WORKS the weighted median or mean of neighbor noise free pix-
For improving the standard median filter, researchers pro- els with a weighted operator, differentiating the contribu-
posed switching median filters [4]–[7], which discriminate tions and impacts of neighbor pixels on the central pixel by
the noisy pixels from the noise free ones prior to the noise weighted processing so as to achieve a better denoising result.
removal processing, so as to only make the detected noisy pix- The adaptive dynamically weighted median filter (ADWMF)
els undergo noise removal processing and keep the noise free [22] estimates the intensity of noisy pixel by employing
ones unchanged. Comparatively, the switching median filters the weighted median of a neighborhood of adaptive size;
protect the original information of noise free pixels. However, the weighted operator employed is derived from Gaussian
switching median filters are unable to handle various density surface. The filter using radial basis functions interpolation
of noise. Considering this issue, adaptive switching median (RBFI) in [24] estimates the intensity of noisy pixel with
filters [8]–[11] were proposed; they are robust for various radial basis functions interpolation, and then, refines the
density of impulse noise. Wang et al. [10] proposed a novel recovery image using a distance inversely weighted mean
learning-based switching median filter which detects noise by filter.
a learning-based method, and by an iterative manner, takes In view of the wide application of fuzzy theory, it was
the median of noise free pixels surrounding noisy pixel as introduced into the image processing for improving the
the estimated intensity of noisy pixel. However, its improved denoising performance [25]–[28]. The SVM classification
performance is achieved at the expense of computational based fuzzy filter (SVMFF) [25], aiming at performance
complexity. Erkan et al. [11] proposed a different applied improvement irrespective of noise density, employs a support
median filter (DAMF) to protect the thin lines and edges vector machine classification for noise detection, along with
of image, and avoid smearing image details while removing a histogram based fuzzy filtering for noise removal. Inspired
noise; the DAMF employs the noise free pixels in a neighbor- by the fuzzy switching median filters and the works on the
hood of adaptive size for noise removal, and uses previously concept of information sets, a noise adaptive information
processed pixels to remove the residual noises. set based switching median filter (NAISM) is proposed in
In order to further improve the performance with various [27]; the information sets are derived from fuzzy sets to
strategies, decision based filters [12], [13] were proposed; deal with the uncertainty, and by virtue of the switching
their robust strategies improve the noise detection accuracy criterion and the local effective information surrounding the
and the performance of noise removal. A modified decision noisy pixel, the best calculated value replaces the noisy pixel.
based unsymmetric trimmed median filter [12] proposed Generally, existing techniques strongly rely on exploiting
a noise detector identifying impulse noise simply by the the neighbor information of noisy pixel for noise removal;
extreme intensity; the detected noisy pixels are replaced however, considering the diversity singularity and non-
by the unsymmetric trimmed median of a neighborhood of stationary feature of image signal in local neighborhood,
fixed size. A neighborhood decision based impulse noise the estimation result could easily diverge from the true value
filter (NDBINF) in [13] employs a neighborhood decision and cause ugly visual effects in textures and edge regions.
approach to protect the noise free pixels having extreme These inspired some researchers to expect better denoising
intensity while detecting noise, and a first-order neighbor- performance by exploiting the nonlocal information during
hood decision approach to restore the detected noisy pixels. the noise removal procedure. However, initially, the nonlo-
It is generally believed that mean filters are unable to cal techniques were only for removal of additive Guassian
achieve good performance, for they go without the ability noise and random-valued impulse noise [29]; and subse-
of structure and edge information preservation; however, quently, some researchers ingeniously proposed the improved
the improved variants of mean filter may get remarkable versions of nonlocal techniques for fixed-valued impulse

158546 VOLUME 7, 2019

J. Chen et al.: Adaptive Sequentially Weighted Median Filter for Image Highly Corrupted by Impulse Noise

noise removal. Wang et al. [30] proposed an iterative nonlocal

means filter (INLM); the concept of nonlocal means filter is
based on the fact that there exist lots of similar patches with
repeat patterns in natural image, and the central pixels of these
similar patches share the same intensity value distribution;
the central noisy pixel under processing is thus replaced by
the weighted mean of central pixels of all similar patterns.
FIGURE 1. Zoomed-in local part of Lenna and its corrupted versions.
In general, the nonlocal filters show high computational (a) A local part of original image Lenna; corrupted versions by (b) 10%
complexity, and performance improvements for fixed-valued noises; (c) 50% noises; (d) 90% noises.

impulse noise are not necessarily conspicuous.

Originally, the deep learning technique based filters versions carefully, several important observations can be
focused on Gaussian noise removal; thereafter, inspired obtained as follows.
by the emerging deep learning theories, some researchers (i)Fixed-valued impulse noise takes extreme intensity,
attempted to use them for fixed-valued impulse noise, such i.e., 0 and 255.
as [31], [32]. When the noise is additive Gaussian, the noisy (ii)Noises with minimum intensity and noises with max-
image value is still correlated to the original value, hence, imum intensity distribute randomly but local unevenly with
deep learning based filter can achieve a good performance equal probability.
for Gaussian noise removal. However, unlike Gaussian noise, (iii)Local noise free pixels vary smoothly, and smooth
impulse noise is not correlated with the original image data, regions are separated by edges.
each pixel has a probability of being corrupted; thus, the deep (iv)For noisy pixel, its intensity differs greatly from those
learning based filters are not fit for impulse noise removal, of neighbor noise free pixels surrounding it, it is thus in
also, by observing their experimental results in [31], [32], isolation.
their performances are not conspicuous. (v)There are often some noise blocks in highly corrupted
image.
III. PROPOSED METHOD
A. MODEL OF FIXED-VALUED IMPULSE NOISE 2) NOISE DETECTOR BASED ON 3σ PRINCIPLE
AND LOCAL STATISTICS
Generally, the fixed-valued impulse noise can be modeled by
its intensity and distribution. (i) Impulse noise takes extreme For a natural image, it can be noted from its intensity his-
intensity in the image intensity range; in an 8-bit gray image togram that generally, the intensities of pixels approximately
whose intensity ranges from 0 to 255, the impulse noise takes obey normal distribution, and so does the intensities of local
intensity 0 and 255. (ii) When corrupts an image, the impulse pixels. The local noise free pixels have high similarity, and are
noise distributes randomly and evenly with a certain proba- highly correlated with each other, so that they approximately
bility, being independent from neighbor noise free pixels; the obey local normal distribution. And the noisy pixels take
noises with minimum intensity and the ones with maximum fixed extreme intensity, and their locations obey random dis-
intensity distribute with equal probability. Correspondingly, tribution; thus, they are lonely, deviating considerably from
the impulse noise can be mathematically modeled as the neighbor noise free pixels.
 Naturally, based on the intensity feature of impulse noise,
min d/2
 assuming all pixels having extreme intensity to be noisy,
f (p) = max d/2 (1) may not be valid, as the noise free pixels having extreme
 intensity are definitely taken as noisy pixels by this assump-
o 1−d

tion. To address this problem, we seek help from the normal
Here, f (p) is the intensity of pixel p in corrupted image with distribution.
a noise density d, extreme values min and max denote the Based on the just above analyses, we employ the 3σ
intensity of noisy pixel, and o denotes the intensity of noise principle of normal distribution for further detection, with
free pixel. the expectation that the noise detector has good ability to
discriminate noise free pixels from the noisy ones having the
B. NOISE DETECTION BY 3σ PRINCIPLE same intensity value. As shown in FIGURE 2, denote by u the
AND LOCAL STATISTICS mean of a set, and σ the standard deviation, if the individuals
1) FEATURES OF FIXED-VALUED IMPULSE NOISE X of the set obey normal distribution, then 68.26% individu-
To reveal the features of impulse noise as well as the dis- als would locate in interval (u−σ , u+σ ), 95.44% individuals
tinction between noisy pixel and noise free pixel, a sim- locate in interval (u − 2σ , u + 2σ ), and 99.74% individuals
ple experiment is carried out, and the results are shown locate in interval (u − 3σ , u + 3σ ); they can be defined by
in FIGURE 1. FIGURE 1 shows a zoomed-in local part
P{µ − σ < X < µ + σ } = 68.26% (2)
of image Lenna, which locates at 135∼154 rows and
220∼239 columns, and its corrupted versions by 10%, 50%, P{µ − 2σ < X < µ + 2σ } = 95.44% (3)
and 90% noises, respectively. By observing the corrupted P{µ − 3σ < X < µ + 3σ } = 99.74% (4)

VOLUME 7, 2019 158547

 J. Chen et al.: Adaptive Sequentially Weighted Median Filter for Image Highly Corrupted by Impulse Noise

FIGURE 2. 3σ principle of normal distribution.

Note that one pixel with extreme intensity means high proba-
bility of being noisy, because of the extreme intensity feature
of impulse noise; but if it locates in the 1σ interval shown
in FIGURE 2 and formula (2), we take it as noise free, for
it shows high correlation with the neighbor pixels having
non-extreme intensity. FIGURE 3. PSNR of ASWMF with various T on BSD68 dataset corrupted
However, in a black or white region, noise detection by the with various density of impulse noise.
extreme intensity and the 3σ principle of normal distribution
fails; but fortunately, the features of black or white region can For the optimal value of T , by using vast trial and error
be inferred and utilized for noise detection. For example, for methods, we conducted an experiment with ASWMF on
a black region, two important observations can be made as BSD68 dataset corrupted with 20% noises, 50% noises, and
follows. 80% noises, respectively, the result of which is shown in
(i) Originally, most or all the pixels in this black region take FIGURE 3. The result shows how the denoising performance
minimum intensity. depends on the noise detector with various values of T . Based
(ii) When impulse noise corrupts this region, the noises on the result, we prefer to take T = 20, which means the
with minimum intensity disappear for they are assimilated, following formula should be approximately satisfied
only the noises with maximum intensity exist. Accordingly, ∂PSNR(T ) 

it is noteworthy that most of pixels in this corrupted region =0 (5)
∂T 
T =20
are of minimum intensity.
Therefore, in a black image region, we employ local The performance of an impulse noise detector depends on
intensity statistics for noise detection based on these two its capability to detect all the true positives effectively in
observations, that is, if a pixel takes minimum intensity, the presence of false positives as less as possible. And the
and the minimum intensity accounts for the majority in its proposed noise detector is able to achieve this so that the edge
neighborhood, this pixel is labeled as noise free, otherwise as and structure information can thus be better preserved and
noisy. The noise detection strategy can be similarly made for restored in the noise removal processing.
a white region.
In addition, by reference to the literatures, and based on C. NOISE REMOVAL BY ADAPTIVE SEQUENTIALLY
the approximate local symmetry of the pixels in spatial dis- WEIGHTED MEDIAN
tribution, we take square neighborhood, which is symmetrical In our method, only the noisy pixels undergo a noise removal
about the center, for noise detection and removal; we denote processing, and noise free pixels are left unchanged so as to
by Np (k) the neighborhood of size k × k centered at pixel p. protect the original image information as much as possible;
Specifically, the proposed noise detector based on 3σ prin- in addition, for one noisy pixel, only the noise free pixels
ciple and local intensity statistics is defined as follows. surrounding itself are employed in its intensity estimation.
(i) Take the pixels having extreme intensity as noise candi- Besides, we employ a neighborhood of adaptive size for noise
dates, because of the intensity feature of impulse noise. removal processing, i.e., if no noise free pixels are available in
(ii)For each noise candidate p, in Np (7), if more than two small neighborhood for noise removal processing, the neigh-
non-extreme values are available, turn to (iii), or else, turn to borhood is enlarged so as to contain noise free pixels.
(iv). In our noise removal processing, noise free pixels only on
(iii)Compute the mean u and standard deviation σ of the the border of one neighborhood are employed. Take FIGURE
non-extreme intensity in Np (7); if f (p) falls into interval (u- 4 for an example. For the central noisy pixel p, initially, its
σ , u + σ ), label p as noise free, otherwise as noisy. Np (3) neighborhood is employed, and thus, noise free pixels
(iv)In Np (5), if f (p) = 0 and the number of intensity only on the border of Np (3) are employed; and if no noise free
0 (denoted by n0 ) accounts for the majority, here we set pixels are available in the Np (3) for noise removal, the Np (5)
n0 > T , label p as noise free, otherwise as noisy. This strategy is employed, in this case, equally, noise free pixels only on
is similar for f (p) = 255. the border of Np (5) are employed, for no noise free pixels

158548 VOLUME 7, 2019

J. Chen et al.: Adaptive Sequentially Weighted Median Filter for Image Highly Corrupted by Impulse Noise

In summary, the adaptive sequentially weighted median

filter is set forth as follows. Each detected noisy pixel p under-
goes the following adaptive weighted median processing.
(i) If noise free pixels are available on the border of Np (k)
(initially, k = 3 is taken) under consideration, turn to (ii), or
else, turn to (iii).
(ii) Conduct weighted processing on the noise free pixels
by the weighted operator W (k), and take the median of
weighted ones as the intensity of p, which can be formulated
as

f (p) = median(Np (k)  W (k)) (6)

FIGURE 4. Neighborhood of adaptive size.
where the symbol  is the pixel-wise duplication operator; for
are available in the Np (3) which is also included in Np (5). example, for a noise free pixel p with weighted coefficient n,
However, the noise free pixels on the border of one neigh- the weighted processing is defined by
borhood are not equidistant, and therefore, their contributions
n times
and impacts on the central pixel are unequal so that directly z }| {
taking the median of them as the estimated intensity of central f (p)  n = f (p), f (p), f (p), · · ·, f (p) (7)
noisy pixel is partial and inadvisable. To differentiate the
contributions and impacts of neighbor noise free pixels on the (iii) Enlarge the Np (k), i.e. set k = k+2; if k ≤MaxN, turn
central pixel, the noise free pixels available on the border of to (i), or else, let p unprocessed.
one neighborhood, are weighted by a weighted operator, and (iv) After processing all the detected noisy pixels by (i) and
then, the median of the weighted ones is taken as the intensity (ii), replace each unprocessed noisy pixel p with the median
of noisy pixel under consideration. of its Np (5) including the pixels of previously processed and
Our weighted operator is innovatively defined in unprocessed.
FIGURE 5, in which denoted by W (k) the weighted operator The adaptive sequentially weighted median processing,
of size k × k. Note that in W (k), the non-zero coefficients, which differentiates accurately the contributions and impacts
by which the corresponding pixels are duplicated so as to of neighbor pixels on the central pixel with a sequentially
control their probability of becoming the median, are sequen- weighted operator, can achieves a better recovery result, and
tially inversely proportional to their spatial distances from can be capable of restoring the edge and structure information
central pixel, and are signed only to the noise free pixels very well.
at the corresponding locations, representing accurately the
contributions and impacts of neighbor pixels on the cen- IV. EXPERIMENTS
tral pixel. In addition, it is noteworthy that the coefficients By running Matlab R2019a on a machine with Intel(R)
assigned to noisy pixels are zero, for noisy pixels could Core(TM) i7-7700 CPU at 3.60 GHZ, equipped with 8 GB
not participate in the intensity estimation of central noisy RAM, we conduct the experiments on datasets SET12,
pixel. BSD68, and medical images shown in FIGURE 6. The empir-
As mentioned above, if no noise free pixels are available, ical validation for the proposed ASWMF is conducted by
the neighborhood is enlarged until at least one noise free performing thorough comparative analyses with the state-
pixel is found or the neighborhood size has reached the of-the-art filters proposed recently in literatures, which are
predefined maximum. At this point, if the neighborhood has DAMF [11], NDBINF [13], PVGF [18], ADWMF [22], RBFI
reached the maximum size, and still no noise free pixels are [24], SVMFF [25], NAISM [27], and INLM [30], in terms of
available, the noisy pixel under consideration is replaced by noise detection accuracy, peak signal to noise ratio (PSNR),
the median of all pixels in its neighborhood N (5), including structural similarity index (SSIM) [33], edge preservation
the pixels of previously processed, unprocessed, and noise index (EPI) [34], image entropy H [35], visual perception,
free. As to the maximum size of neighborhood employed for and computational time. The PSNR, SSIM, EPI, and H are
noise removal, denote it by MaxN; it could not be too large, defined by
for pixels at far distance show weak correlation and impact,
also, it could not be too small, for it shows no robustness for m × n × 2552
PSNR = 10 × log10 m P
n (8)
high density noise. By conducting experiments on various P
(f (i, j) − g(i, j))2
images and using trial and error method, we see that our i=1 j=1
proposed method with MaxN=9 and MaxN=11 shows almost (2uf ug + C1 )(2σfg + C2 )
the same best denoising performance; thus, we prefer to take SSIM = ,
(u2f + u2g + C1 )(σf2 + σg2 + C2 )
9 as the optimal value of MaxN, because of computational  
complexity. C1 = (K1 L)2 , C2 = (K2 L)2 (9)

VOLUME 7, 2019 158549

 J. Chen et al.: Adaptive Sequentially Weighted Median Filter for Image Highly Corrupted by Impulse Noise

FIGURE 5. Sequentially weighted operator W (k) of (a) size 3 × 3; (b) size 5 × 5; (c) size 7 × 7; (d) size 9 × 9.

TABLE 1. Performance of the state-of-the-art filters and ASWMF in noise

detection on image Chest_Xray.

FIGURE 6. Medical images. (a) Chest_Xray; (b) Abdomen_CT.

m P
P n


fhp (i, j)−ufhp × ghp (i, j)−ughp
i=1 j=1
EPI = s the texture of image; larger entropy implies coarser texture
m P
n m P
n
P
2 P
2 and better texture preservation ability [36].
fhp (i, j)−ufhp × ghp (i, j)−ughp
i=1 j=1 i=1j=1
A. NOISE DETECTION PERFORMANCE OF FILTERS
(10)
255 X
255

255 X
255
 Among DAMF, NAISM, PVGF, and RBFI, we only take
X X DAMF for evaluation, for they adopt the same noise detection
H=− Pij ln Pij ,  Pij = 1 (11)
strategy, i.e., min-max noise detector. In terms of missing
i=0 j=0 i=0 j=0
detection rate (MDR) and false detection rate (FDR), the
Here, f and g are the original image and recovery image, detected results of all filters on image Chest_Xray are shown
respectively; m and n are the height and width of image, in TABLE 1; and the MDR and FDR are given by
respectively. Denoted by uf and σf the mean and standard (Rr (p) = 1 ∧ Rd (p) = 0)
P
deviation of f , respectively, σfg the covariance of f and g. p∈f
MDR = 100 × (%) (12)
The stabilizing constant C1 and C2 are calculated with the Rr (p)
P
dynamic range, L = 255, K1 and K2 , by default K1 and K2 p∈f
(Rr (p) = 0 ∧ Rd (p) = 1)
P
are selected as 0.01 and 0.03, respectively [33]. fhp and ghp
are respectively the high-pass filtering images of f and g p∈f
FDR = 100 × (%) (13)
Rr (p)
P
with Laplacian filter, with uf hp and ughp being their mean
values, respectively [34]. Pij reflects the comprehensive char- p∈f

acteristics of the intensity distribution of one pixel and its Here Rr denotes the matrix for indicating the true noises,
surrounding pixels [35]. which indicate a pixel p as noise free with Rr (p) = 0 or
PSNR measures the similarity of original image and recov- noisy with Rr (p) = 1; Rd denotes the matrix for indicating
ery image; the higher the PSNR is, the better the noise the detected noises.
removal capability is. SSIM measures the structure infor- By comparing the statistics in TABLE 1, two points can
mation preservation capability; higher SSIM signifies better be concluded. (i) At various noise densities, some existing
preservation capability. EPI measures the edge preservation filters have a very small number in MDR, show superiority
capability; the higher the EPI is, the better the edge preser- over the proposed ASWMF; however, they have a very large
vation capability is. Entropy H is a statistical measure of number in FDR, such as DAMF and SVMFF. (ii) Although
randomness and variability that can be used to characterize the ASWMF shows no superiority in MDR, it achieves

158550 VOLUME 7, 2019

J. Chen et al.: Adaptive Sequentially Weighted Median Filter for Image Highly Corrupted by Impulse Noise

TABLE 2. PSNR and SSIM of the state-of-the-art filters and ASWMF on TABLE 3. PSNR and SSIM of the state-of-the-art filters and ASWMF on
image Lenna. image Mandrill.

a far smaller number in FDR. These imply that ASWMF can TABLE 4. PSNR and EPI of the state-of-the-art filters and ASWMF on
achieve better noise detection overall comparing to the other image Man.

filters at various noise densities.

B. PERFORMANCE OF FILTERS ON SET12 DATASET

Here, the PSNR, SSIM, and EPI of the recovery results of
the state-of-the-art filters and ASWMF on images Lenna,
Mandrill, Man, and Boat from SET12 dataset, which are cor-
rupted by various density of impulse noise, are tabulated; and
the best results are indicated in bold. Applying the proposed
ASWMF brings about a considerable increase in the PSNR,
SSIM and EPI compared with the state-of-the-art filters with
respect to various noise densities.
The results of all filters on image Lenna are tabulated as
TABLE 2. It can be inferred from numerical results that the
ASWMF gives a considerable improvement in noise removal
and structure information preservation at all noise densities;
in PSNR, the ASWMF consistently achieves superior results
to those of the other filters, and in SSIM, the superiority of
ASWMF to the other filters at high noise density is more sig- the ASWMF gives more promising results; its higher
nificant than that at low noise density. These can be attributed PSNR values imply better noise removal capability, and its
to the intelligent use of the 3σ principle of normal distribution higher EPI values signify better edge preservation capability.
and the local intensity statistics in noise detection, as well as In PSNR, the superiority of ASWMF to the other filters
the sequentially weighted median processing. is almost approximate along various noise densities; and in
The restored results of all filters for image Mandrill are EPI, the superiority of ASWMF grows, as the noise density
revealed in TABLE 3. From the PSNR and SSIM values increases. This is primarily due to the fact that ASWMF can
in TABLE 3, it can be easily grasped that the performance accurately discriminate the noisy pixels from the noise free
of ASWMF is much better than those of the other compet- ones, and effectively estimate the intensity of noisy pixel,
ing state-of-the-art filters. ASWMF is the most successful thereby making it more able to restore the image from high
method than the others even at high noise density. It is density noise.
noteworthy that the superiority of ASWMF in SSIM over In terms of PSNR and EPI, TABLE 5 shows the superior
the other filters grows, as the noise density increases. These results of ASWMF to those of the other filters for image
imply that the ASWMF is successfully devoted to improving Boat. By observing carefully TABLE 5, we arrived at two
the noise detection, noise removal and structural information conclusions. (i) The ASWMF improve the performance of
preservation. the existing filters by removing noise thoroughly and ren-
In terms of PSNR and EPI, the superior performance dering the restored image free from blur effect, while the
of ASWMF to the other filters for image Man is demon- edge information is well preserved. (ii) The superiority of
strated in TABLE 4. Compared to the state-of-the-art filters, ASWMF to the other filters in EPI increases with the noise

VOLUME 7, 2019 158551

 J. Chen et al.: Adaptive Sequentially Weighted Median Filter for Image Highly Corrupted by Impulse Noise

FIGURE 7. PSNR and EPI of the state-of-the-art filters and ASWMF on image chest_Xray. (a)PSNR; (b)EPI.

TABLE 5. PSNR and EPI of the state-of-the-art filters and ASWMF on by EPI curves, the conclusion derived from EPI is similar
image boat.
as that derived from PSNR; along various noise densities,
the ASWMF shows excellent performance in edge preser-
vation. The reason behind these improvements is that the
ASWMF can accurately discriminate the noisy pixels from
the noise free ones, so as to protect the noise free pixels,
and effectively remove the noise while preserving the edge
information.
FIGURE 8 clearly demonstrates that ASWMF gives
definite improvements in PSNR and SSIM over those of
the other filters for image Abdomen_CT. Compared to the
other filters, the ASWMF shows more promise for pro-
cessing medical images; in noise removal processing by
ASWMF, more structure information can be preserved and
better denoising performance can be achieved. Many pix-
els having extreme intensity are often available in medical
images; hence, the superiority of ASWMF over the other
filters in medical image processing highly depends on its
density increases; thus, the edge information preserved by noise detector.
ASWMF is remarkably better than those preserved through
the other filters. These imply that noise removal and edge D. PERFORMANCE OF FILTERS ON BSD68 DATASET
information preservation can be achieved more effectively by We plot in FIGURE 9 and 10 the average PSNR, SSIM, EPI,
ASWMF. and entropy values of the state-of-the-art filters and ASWMF
on BSD68 dataset with various density of impulse noise,
C. PERFORMANCE OF FILTERS ON MEDICAL IMAGES visualizing their denoising performance.
Here, we investigate the performance of ASWMF on med- As expected, the results in FIGURE 9 confirm the supe-
ical images; and the superiority of ASWMF is revealed by riority of ASWMF again, and are consistent with the results
the results plotted in FIGURE 7 and 8. The PSNR and reported above. Three important observations can be made
EPI of the recovery results of the state-of-the-art filters from FIGURE 9. (i) As the noise density increases, consis-
and ASWMF on image Chest_Xray with various density of tently for all filters, the restoration quality degenerates, but
impulse noise, are plotted in FIGURE 7. From FIGURE 7, our method consistently gives the best performance. (ii) The
it is noticed that compared to the other filters, substantial per- PSNR and SSIM curves of ASWMF are above those of the
formance improvements can be obtained through ASWMF; other filters, keeping a gap with them along various noise
the ASWMF outperforms the state-of-the-art filters in noise densities. (iii) Interestingly and apparently, the performance
removal thoroughly which is revealed by PSNR curves. gap between ASWMF and the other filters in SSIM grows
As to the edge information preservation, which is revealed larger, as the noise density increases.

158552 VOLUME 7, 2019

J. Chen et al.: Adaptive Sequentially Weighted Median Filter for Image Highly Corrupted by Impulse Noise

FIGURE 8. PSNR and SSIM of the state-of-the-art filters and ASWMF on image Abdomen_CT. (a)PSNR; (b)SSIM.

FIGURE 9. Average PSNR and SSIM of the state-of-the-art filters and ASWMF on BSD68 dataset. (a)PSNR; (b)SSIM.

The results in FIGURE 10 confirm the better capability of at low noise density is often difficult to evaluate by visual
proposed ASWMF in edge preservation, and keeping the uni- perception, we focus on evaluating it at high noise density.
formity of image intensity distribution and the image textures. Visual analyses on the capability of noise removal, structure
It can be seen obviously that the EPI values of ASWMF are and edge information preservation of the proposed ASWMF
larger than those of other filters along various noise densities, against the state-of-the-art filters in the recovery results for
showing better edge preservation capability; as to the entropy image Barbara with 80% impulse noises and Jetplane with
of recovery image, the ASWMF outperforms the other filters 90% impulse noises are made in FIGURE 11 and 12, respec-
at almost all the noise densities, except that ASWMF lags tively. For a visually clear comparison, only a zoomed-in local
behind the DAMF at very high noise density, showing good part of each recovery image is shown.
capability in keeping the variation of image intensity and the FIGURE 11 shows the improvements in visual appearance
image textures. given by ASWMF against the state-of-the-art filters for image
These can be attributed to that the ASWMF is designed in Barbara; two numbers in the parentheses below each recov-
such a way that it discriminates the noisy pixels accurately ery image are the corresponding PSNR and SSIM, respec-
from the noise free ones and estimates the intensity of noisy tively. By observing FIGURE 11 carefully, three important
pixel effectively. conclusions can be inferred. (i) The ASWMF gives a more
visually pleasant recovery image, and provides a signifi-
E. VISUAL PERCEPTION OF RECOVERY RESULTS cant contribution towards preserving the structural infor-
Here, the images Barbara and Jetplane in SET12 are selected mation. (ii) In the recovery images of the other filters,
as experimental images. Because the performance of a filter either residual noises or obvious blur effects can be seen.

VOLUME 7, 2019 158553

 J. Chen et al.: Adaptive Sequentially Weighted Median Filter for Image Highly Corrupted by Impulse Noise

FIGURE 10. Average EPI and entropy of the state-of-the-art filters and ASWMF on BSD68 dataset. (a)EPI; (b)Entropy.

FIGURE 11. Recovery results of image Barbara with 80% impulse noises.(a) Zoomed-in local part of Barbara with impulse noise;
Results of (b) DAMF; (c) NDBINF; (d) PVGF; (e) ADWMF; (f) RBFI; (g) SVMFF; (h) NAISM; (i) INLM; (j) ASWMF.

(iii)Again, the PSNR and SSIM values below the recovery level with better capability of local features preservation. And
images confirm the superiority of ASWMF over the other fil- the corresponding PSNR and EPI are also consistent with the
ters. These mean that the ASWMF possesses better capability conclusion derived by visual perception. The ASWMF still
of noise removal and structural information preservation than works fairly well, even at a very high noise density; this is
the existing filters, and the image produced by the ASWMF due to the same reason as explained above.
is visually sharper and more distinctive than those obtained Here, these inferences by visual perception are consistent
by the other filters. with the comments given above.
The recovery images of all filters on image Jetplane with
90% impulse noises, are shown in FIGURE 12; two numbers F. COMPUTATIONAL TIME
in the parentheses below each recovery image are the corre- Denoted by n the pixel number of one image and Cx
sponding PSNR and EPI, respectively. On careful observation a constant. Consider the complexity of each subprocess
from FIGURE 12, the visual result of ASWMF is percep- in the proposed ASWMF: noise detection by extreme
tually better than those of the other considered filters; the intensity— O(n); detection by 3σ principle or local
recovery results confirm that our result is far superior to those intensity statistics— O(C1 n), here C1 = 49 or C1 = 25;
of the other filters, indicating that the ASWMF still produces search noise free pixels in neighborhood of adaptive
consistently higher quality image at such a high corruption size— O(C2 n); weighted processing for each noisy

158554 VOLUME 7, 2019

J. Chen et al.: Adaptive Sequentially Weighted Median Filter for Image Highly Corrupted by Impulse Noise

FIGURE 12. Recovery results of image Jetplane with 90% impulse noises. (a) Zoomed-in local part of Jetplane with impulse noise;
Results of (b) DAMF; (c) NDBINF; (d) PVGF; (e) ADWMF; (f) RBFI; (g) SVMFF; (h) NAISM; (i) INLM; (j) ASWMF.

TABLE 6. Average computational time (seconds) of each filter on six V. CONCLUSION

selected images.
In this paper, we proposed a new method ASWMF for image
restoration from impulse noise, consisting of a simple and
effective noise detector, and a noise removal technique capa-
ble of removing impulse noise thoroughly and preserving the
structure and edge information very well. The noise detector
in ASWMF takes full advantage of the 3σ principle of normal
distribution and the local intensity statistics; and the noise
removal technique in ASWMF gets support from the adaptive
sequentially weighted median processing. The 3σ principle
of normal distribution and local intensity statistics employed
in noise detection and the adaptive sequentially weighted
operator employed in noise removal are the remarkable con-
tributions of the proposed ASWMF; they jointly make the
denoising performance advanced considerably. With exten-
sive experimental results on various images with various
density of impulse noise, it is observed that quantitatively and
pixel— O(C3 n); replace each noisy pixel with median— qualitatively, the proposed ASWMF performs superiorly to
O(C4 n). Therefore, summing them together, the total com- the state-of-the-art filters in the presence of impulse noise.
plexity of ASWMF is O((1+C1 +C2 +C3 )n). Hence, it can In addition, no significant superiority in computational time
be concluded that the computational complexity of the pro- is shown by the proposed ASWMF; we will advance it in
posed ASWMF is acceptable. Further, we end this section the further work, making it applicable for real-time image
by considering the computational time of our ASWMF and denoising.
the state-of-the-art filters; they are performed on image
Lenna, Mandril, Man, Boat, Barbara, and Jetplane, and the ACKNOWLEDGMENT
average computational time of each filter is tabulated as The authors would like to thank the editors and anonymous
TABLE 6. reviewers for their constructive suggestions to the
A careful look at TABLE 6 and reference to the above improvements of this article.
performance analyses reveal that some state-of-the-art filers
often provide such a good performance only at the cost of REFERENCES
computational time, such as INLM and RBFI. Among all [1] K. H. Jin and J. C. Ye, ‘‘Sparse and low-rank decomposition of a Hankel
filters, no significant superiority in computational time is structured matrix for impulse noise removal,’’ IEEE Trans. Image Process.,
shown by the proposed ASWMF; however, its running speed vol. 27, no. 3, pp. 1448–1461, Mar. 2018.
[2] L. Liu, C. P. Chen, Y. Zhou, and X. You, ‘‘A new weighted mean filter
is comparable to most of the state-of-the-art filters, and sur- with a two-phase detector for removing impulse noise,’’ Inf. Sci., vol. 315,
passes some other filters, such as ADWMF, INLM and RBFI. pp. 1–16, Sep. 2015.

VOLUME 7, 2019 158555

 J. Chen et al.: Adaptive Sequentially Weighted Median Filter for Image Highly Corrupted by Impulse Noise

[3] W. Ye, ‘‘Optimality of the median filtering operator,’’ Circuits, Syst. Signal [27] S. Arora, M. Hanmandlu, and G. Gupta, ‘‘Filtering impulse noise in
Process., vol. 30, no. 6, pp. 1329–1340, 2011. medical images using information sets,’’ Pattern Recognit. Lett., to be
[4] N. Singh and U. Oorkavalan, ‘‘Triple threshold statistical detection fil- published, doi: 10.1016/j.patrec.2018.06.002.
ter for removing high density random-valued impulse noise in images,’’ [28] M. González-Hidalgo, S. Massanet, D. Ruiz-Aguilera, and A. Mir,
EURASIP J. Image Video Process., vol. 2018, no. 1, p. 22, 2018. ‘‘Improving salt and pepper noise removal using a fuzzy mathemati-
[5] M. S. Nair and P. M. A. Mol, ‘‘Direction based adaptive weighted switching cal morphology-based filter,’’ Appl. Soft Comput., vol. 63, pp. 167–180,
median filter for removing high density impulse noise,’’ Comput. Elect. Feb. 2018.
Eng., vol. 39, no. 2, pp. 663–689, 2013. [29] P. Nair and K. N. Chaudhury, ‘‘Fast high-dimensional bilateral and non-
[6] C. Yuan and Y. Li, ‘‘Switching median and morphological filter for local means filtering,’’ IEEE Trans. Image Process., vol. 28, no. 3,
impulse noise removal from digital images,’’ Optik, vol. 126, no. 18, pp. 1470–1481, Mar. 2019.
[30] X. Wang, S. Shen, Y. Xu, P. Zhang, and G. Shi, ‘‘Iterative non-local
pp. 1598–1601, 2015.
means filter for salt and pepper noise removal,’’ J. Vis. Commun. Image
[7] A. S. Awad, ‘‘Standard deviation for obtaining the optimal direction in
Represent., vol. 38, pp. 440–450, Jul. 2016.
the removal of impulse noise,’’ IEEE Signal Process. Lett., vol. 18, no. 7, [31] H. C. Burger, C. J. Schuler, and S. Harmeling, ‘‘Image denoising with
pp. 407–410, Jul. 2011. multi-layer perceptrons, part 1: Comparison with existing algorithms and
[8] S. K. Meher and B. Singhawat, ‘‘An improved recursive and adaptive with bounds,’’ Comput. Sci., Nov. 2012, arXiv:1211.1544. [Online]. Avail-
median filter for high density impulse noise,’’ Int. J. Electron. Commun., able: https://arxiv.org/abs/1211.1544
vol. 68, no. 12, pp. 1173–1179, 2014. [32] S. Harmeling, C. J. Schuler, and H. C. Burger, ‘‘Image denoising: Can plain
[9] J. Y. Lee, S. Y. Jung, and P. W. Kim, ‘‘Adaptive switching filter for neural networks compete with BM3D?’’ in Proc. IEEE Conf. Comput. Vis.
impulse noise removal in digital content,’’ Soft Comput., vol. 22, no. 5, Pattern Recognit., Jun. 2012, pp. 2392–2399.
pp. 1445–1455, 2017. [33] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, ‘‘Image quality
[10] Y. Wang, J. Fu, H. Dihn, and R. Adhami, ‘‘A novel learning-based switch- assessment: From error visibility to structural similarity,’’ IEEE Trans.
ing median filter for suppression of impulse noise in highly corrupted Image Process., vol. 13, no. 4, pp. 600–612, Apr. 2004.
colour images,’’ Imag. Sci. J., vol. 64, no. 1, pp. 15–25, 2016. [34] V. Soni, A. K. Bhandari, A. Kumar, and G. K. Singh, ‘‘Improved sub-band
[11] U. Erkan, L. Gökrem, and S. Enginoǧlu, ‘‘Different applied median filter adaptive thresholding function for denoising of satellite image based on
in salt and pepper noise,’’ Comput. Electr. Eng., vol. 70, pp. 789–798, evolutionary algorithms,’’ IET Signal Process., vol. 7, no. 8, pp. 720–730,
Aug. 2018. 2013.
[12] S. Esakkirajan, T. Veerakumar, A. N. Subramanyam, and [35] S. Zhang, Y. Liu, and X. Li, ‘‘Fast entropy minimization based autofocus-
C. H. PremChand, ‘‘Removal of high density salt and pepper noise ing technique for ISAR imaging,’’ IEEE Trans. Signal Process., vol. 63,
through modified decision based unsymmetric trimmed median filter,’’ no. 13, pp. 3425–3434, Jul. 2015.
IEEE Signal Process. Lett., vol. 18, no. 5, pp. 287–290, May 2011. [36] R. C. Gonzalez and R. E. Woods, Digital Image Processing, 3rd ed.
[13] A. K. Samantaray, P. Kanungo, and B. Mohanty, ‘‘Neighborhood deci- Beijing, China: Electronic Industry Press, 2017.
sion based impulse noise filter,’’ IET Image Process., vol. 12, no. 7,
pp. 1222–1227, 2018.
[14] Z. Li, G. Liu, Y. Cheng, and Y. Xu, ‘‘Modified directional weighted filter
for removal of salt & pepper noise,’’ Pattern Recognit. Lett., vol. 40, JIAYI CHEN was born in 1983. He received the
pp. 113–120, Apr. 2014. M.S. degree in computer science and technology
[15] Q.-Q. Chen, M.-H. Hung, and F. M. Zou, ‘‘Effective and adaptive algorithm from the Guangdong University of Technology,
for pepper-and-salt noise removal,’’ IET Image Process., vol. 11, no. 9, Guangzhou, China, in 2010. He is an Associate
pp. 709–716, 2017. Professor with Guangdong Medical University,
[16] V. R. Vijaykumar, G. S. Mari, and D. Ebenezer, ‘‘Fast switching based Zhanjiang, China. He has published more than
median–mean filter for high density salt and pepper noise removal,’’ AEU- 30 articles in his research areas. His research inter-
Int. J. Electron. Commun., vol. 68, no. 12, pp. 1145–1155, 2014. ests include wavelet analysis, signal processing,
[17] G. Gao and Y. Liu, ‘‘An efficient three-stage approach for removing salt and image processing.
& pepper noise from digital images,’’ Optik, vol. 126, no. 4, pp. 467–471,
2015.
[18] C.-T. Lu, M.-Y. Chen, L.-L. Wang, C.-C. Hsu, and J.-H. Shen, ‘‘Removal
of salt-and-pepper noise for X-ray bio-images using pixel-variation gain
factors,’’ Comput. Elect. Eng., vol. 71, pp. 862–876, Oct. 2018. YINWEI ZHAN (M’12) was born in 1966.
[19] M. S. Nair and P. M. A. Mol, ‘‘An efficient adaptive weighted switching He received the M.S. degree in applied mathemat-
median filter for removing high density impulse noise,’’ J. Inst. Eng. India, ics from Jilin University, Jilin, China, in 1989, and
B, vol. 95, no. 3, pp. 255–278, 2014. the Ph.D. degree in applied mathematics from the
[20] H. Hosseini, F. Hessar, and F. Marvasti, ‘‘Real-time impulse noise sup- Dalian University of Technology, Dalian, China,
pression from images using an efficient weighted-average filtering,’’ IEEE in 1992. He is a Professor and a M.S. Super-
Signal Process. Lett., vol. 22, no. 8, pp. 1050–1054, Aug. 2015. visor with the Guangdong University of Tech-
[21] V. Kishorebabu, G. Packyanathan, V. Shankar, and H. Kamatham,
nology, Guangzhou, China. He has published
‘‘An adaptive decision based interpolation scheme for the removal of
more than 100 articles in his research areas.
high density salt and pepper noise in images,’’ EURASIP J. Image Video
Process., vol. 2017, no. 1, 2017, Art. no. 67.
His research interests include image processing,
[22] S. Khan and D.-H. Lee, ‘‘An adaptive dynamically weighted median filter
wavelet analysis, pattern recognition, and computer vision.
for impulse noise removal,’’ EURASIP J. Adv. Signal Process., vol. 2017,
Sep. 2017, Art. no. 67.
[23] Y. Wang, J. Wang, S. Xiao, and L. Han, ‘‘An efficient adaptive fuzzy
HUIYING CAO was born in 1975. She received
switching weighted mean filter for salt-and-pepper noise removal,’’ IEEE
Signal Process. Lett., vol. 23, no. 11, pp. 1582–1586, Nov. 2016.
the M.S. degree in applied electronics from
[24] F. Taherkhani and M. Jamzad, ‘‘Restoring highly corrupted images by Guangxi Normal University, Guangxi, China,
impulse noise using radial basis functions interpolation,’’ IET Image Pro- in 2009, and the Ph.D. degree in physics from
cess., vol. 12, no. 1, pp. 20–30, 2018. Nanjing University, Nanjing, China, in 2012. She
[25] A. Roy, J. Singha, S. S. Devi, and R. H. Laskar, ‘‘Impulse noise removal is an Associate Professor with Guangdong Med-
using SVM classification based fuzzy filter from gray scale images,’’ ical University, Zhanjiang, China. She has pub-
Signal Process., vol. 128, pp. 262–273, Apr. 2016. lished more than 20 articles in her research
[26] M. Habib, A. Hussain, M. Ali, and S. Rasheed, ‘‘Adaptive fuzzy inference areas. Her research interests include physical nano
system based directional median filter for impulse noise removal,’’ AEU- materials and image processing.
Int. J. Electron. Commun., vol. 70, no. 5, pp. 689–697, 2016.

158556 VOLUME 7, 2019