J. Vis. Commun. Image R.

36 (2016) 110

Contents lists available at ScienceDirect

J. Vis. Commun. Image R.

journal homepage: www.elsevier.com/locate/jvci

A neutrosophic filter for high-density Salt and Pepper noise based

on pixel-wise adaptive smoothing parameter q
Xianying Qi a,b, Boqiang Liu a,, Jianwei Xu c
a
School of Control Science and Engineering, Shandong University, Jinan, Shandong 250010, China
b
Department of Radiology, Taishan Medical University, Taian, Shandong 271000, China
c
Medical Imaging Department, Taian Tumor Prevention and Treatment Hospital, Taian, Shandong 271000, China

a r t i c l e i n f o a b s t r a c t

Article history: Image indeterminacy has been neglected in most traditional filtering algorithms. This paper proposes a
Received 2 January 2015 pixel-wise adaptive neutrosophic filter based on neutrosophic indeterminacy feature to remove
Accepted 6 January 2016 high-level Salt-and-Pepper noise. In the proposed algorithm, the indeterminacy of a pixel is quantified
Available online 11 January 2016
by a Neutrosophic Set and innovatively exploited as an efficient characteristic of measuring the similarity
of pixels. In order to adjust the smoothing parameter of the weight function pixel-wise adaptively, the
Keywords: uncertainty of a pixel is utilized as an indicator of image contents. Extensive experiments on numerous
Image denoising
images demonstrate that with a 3
3 window, our method outperforms many existing denoising
Salt and Pepper noise
Adaptive neutrosophic weight function
methods in terms of noise suppression and detail preservation.
Pixel-wise adaptive smoothing parameter 2016 Elsevier Inc. All rights reserved.
Indicator of image contents
Similarity measurement
Neutrosophic Set
Indeterminacy

1. Introduction median filter (AMF) has been put forward in [4,5], but at high noise
density (HND), the biggest template size has reached 39
39, and
Noise is an unwanted signal that corrupts the original image in the computation is exceedingly heavy. Another drawback of MF is
various processes such as image acquisition, transmission and stor- that it executes identically on all pixels in the image, yet an ideal fil-
age. The aim of image denoising is to remove noise while retaining ter should be applied only to noisy pixels while leaving noise-free
useful details as much as possible. One of the most destructive pixels intact. Hence, a switching median filter (SMF) [6] was intro-
noises is Salt-and-Pepper noise (SPN) that replaces the original pix- duced to avoid the injuring of uncontaminated pixels. In SMF, noisy
els with the maximum or minimum gray level of the image. Qual- pixels are firstly identified by certain strategies and then removed
ity of the original image is deteriorated significantly, so it is by specialized regularizations. There are some well-known meth-
imperative to eliminate noise before subsequent image processing, ods with the switching scheme, such as decision-based algorithm
such as object recognition and image segmentation. (DBA) [7], switching-based adaptive weighted median (SAWM)
Numerous methods have been proposed to remove SPN. The [8], modified decision based unsymmetrical trimmed median filter
most popular nonlinear filter is the median filter (MF) [1], but (MDBUTM) [9], fuzzy-based decision algorithm (FBDA) [10], and
strong, undesired contouring effect is produced by MF. It performs decision-based trimmed median filter (DBTMF) [11]. In DBA, if a
well only in low-density noise and under high-density (>70%) envi- pixel has a gray value of 0 or 255 in 8-bit image, it is considered
ronment [2], a larger window size can improve the performance of to be a noise candidate. In noise reduction stage, the noise candi-
noise removal; however, it makes a looser correlation between the date is replaced by the median value as long as at least one noise-
median value and the corrupted pixel, which leads to the blurring of free pixel exists in the 3
3 window, otherwise it will be replaced
image details [2,3]. Furthermore, the most appropriate window size by its neighborhood pixels. Nevertheless, this kind of repetitive
varies with the density of noise, so it is rather difficult to choose the replacement results in annoying artifacts in HND. To overcome this
optimal one. To automatically adjust the window size, an adaptive problem, when all pixels are contaminated, MDBUTMF substitutes
the mean value of all noisy elements in the filtering template for
q
the center pixel, which generates artificial spots in the restored
This paper has been recommended for acceptance by Yehoshua Zeevi.
Corresponding author. image. It is clear that the drawback common to DBA and MDBUTMF
E-mail addresses: [email protected] (X. Qi), [email protected] (B. Liu). is that they all ignore the correlation between pixels. In FBDA, the

http://dx.doi.org/10.1016/j.jvcir.2016.01.005
1047-3203/ 2016 Elsevier Inc. All rights reserved.
2 X. Qi et al. / J. Vis. Commun. Image R. 36 (2016) 110

maximum and minimum of window gray value are regarded as A new concept hNeut-Ai is introduced to express the case of neither
noisy points, and the pixels similar to these polluted ones are all hAi nor hAnti-Ai and used to describe the indeterminacy of an event
eliminated by a fuzzy mechanism. Then the median of the remain- [21]. For example, if hAi = black, then hAnti-Ai = white, hNeut-
ing elements is exploited to refresh the center pixel. The strategy of Ai = red, green, purple, cyan, blue, yellow, etc. (any color except
DBTMF is similar to that of FBDA and the only difference between white and black). In the contaminated image, sometimes it is diffi-
them is that the median value is computed after eliminating pixels cult to distinguish whether a pixel is noisy or noise-free due to the
with the value of 0 or 255. The major disadvantage of these existing of textures or boundaries. In this paper, we denote the
extreme-compression methods is that only the high reliability of background of image as hAi, the edge or texture as hNeut-Ai, and
the median is considered while the local information of pixels has noise as hAnti-Ai. Three neutrosophic components denoted by T, I
been neglected. Consequently, details and edges cannot be recov- and F are applied to estimate the degrees of truth, indeterminacy
ered satisfactorily. In order to overcome this shortcoming, the the- and false, respectively. Let T, I and F be non-standard or standard
ory of image inpainting has been introduced to preserve edges. In real subsets of ]0, 1+[ with sup T t sup; inf T t inf;
[12,13], a three-stage filter (TSF) and an adaptive iterative convolu- sup I i sup; inf I i inf; sup F f sup; inf F f inf, and n sup
tions filter (ACIF) have been proposed, respectively. They can not t sup i sup f sup;n inf t inf i inf f inf [29]. Where
only suppress noise, but also preserve edge information efficiently. x_sup and x_inf are the superior and inferior limits of subset x. There
However, on account of iterative inpainting, the phenomenon of are no restriction on n_sup and n_inf, so  0 6 n sup 6 3 and
undue blurring or over-smoothing is obvious in the restored image 
0 6 n inf 6 3 . T, I and F can be any real sub-unitary subsets and
at HND. In non-local means (NLM) [14,15] filter, a weighting are not necessarily intervals. Besides, the three sets may overlap
scheme is applied, where the weight is determined by the similarity or be converted from one to the other [30]. An element A (t, i, f)
of local patches. NLM is powerful against Gaussian noise, but per- belongs to the set in the following way: it is t true (t 2 T), i indeter-
forms poor in SPN. Besides, in the weight functions of most NLM fil- minate (i 2 I), and f false (f 2 F). In Neutrosophy, if an event hAi is t %
ters, the value of the smoothing parameter h is usually set true, it does not necessarily mean it is (1  t) % false, but can be f %
manually. Methods for adaptive h have been studied [15,16], but false and i % indeterminate simultaneously. However, in traditional
on account of having not taken image contents into consideration, logic, if an event hAi is t % true, it must be (1  t) % false.
most algorithms still use a globally constant value of h.
Due to the complexity of noise sources and the imperfection of 2.2. Neutrosophic image
certain definitions, such factors as ambiguity, vagueness and
imprecision are widespread in image processing. Nevertheless, In neutrosophic domain, a neutrosophic image PNS is repre-
algorithms above-mentioned have not taken them into account. sented by three sets I, T and F [19], and a neutrosophic pixel is
Unlike conventional algorithms, in [17], the indeterminacy infor- denoted as Pt; i; f . The expression of neutrosophic pixel signifies
mation is introduced to image denoising and a Neutrosophic the point is t % true (background), i % indeterminate (texture or
entropy filter (NEF) has been presented. In NEF, the contaminated edge) and f % false (noise), where i varies in I, t varies in T, and f var-
image is transformed into Neutrosophic Set (NS) domain at first,
ies in F, respectively. A pixel Pi; j in traditional image can be
then an iterative c-median filtering is applied to reduce the inde-
transformed into neutrosophic domain in the following way:
termination degree evaluated by neutrosophic entropy. It can
PNS i; j fTi; j; Ii; j; Fi; jg. In order to make full use of the
remove different kinds of noises effectively, but the iterative oper-
uncertain information to restore the corrupted image, an efficient
ation of c-median generates undue blurring of images. At present,
estimation of I is critical. It is well known that the median value
Neutrosophy has been used not only in image denoising but also in
has stronger immunity to SPN, so it can be used to construct the
image segmentation [18,19]. Consequently, how to further maxi-
function of I. Besides, the maximum absolute luminance difference
mize the potential of indeterminacy in image restoration is a
defined as Eqs. (1) and (2) can reflect the distinction between the
meaningful task.
center pixel and its neighboring pixels, hence it can be utilized to
In this paper, by exploiting the indeterminacy information
evaluate the degree of current pixel being noise-free.
based on NS, a new powerful neutrosophic filter is presented for
the reduction of high density SPN. On one hand, the uncertainty Dif i; j maxfdif i k; j lg; 1
of a pixel is regarded as a feature of measuring pixel similarity;
on the other hand, by using the indeterminate information to dis- dif i k; j l jpi k; j l  pi; jj with i k; j
tinguish the regional type to which a pixel belongs, a pixel-wise
li; j; 2
adaptive smoothing function is put forward to minimize the nega-
tive influence resulting from a globally fixed h. Experimental Motivated by Ref. [17], Ii; j; Ti; j and Fi; j are formulated as
results on a series of images have demonstrated that the proposed
di; j  dmin
adaptive neutrosophic weighted filter (ANWF) outperforms the Ii; j ; 3
counterparts in terms of visual quality and objective performance. dmax  dmin
The outline of the paper is organized as follows. NS and the
quantification of uncertainty are briefly introduced in Section 2. di; j jPi; j  mi; jj; 4
Section 3 describes ANWF in detail. Results of the proposed algo- 2 3
rithm are described and discussed in Section 4. Section 5 concludes pi  1; j  1 pi  1; j pi  1; j 1
6 7
the paper. mi; j median4 pi; j  1 pi; j pi; j 1 5; 5
pi 1; j  1 pi 1; j pi 1; j 1
2. Quantification of indeterminacy
Difmax  Dif i; j
2.1. Neutrosophic Set Ti; j ; 6
Difmax  Difmin

Proposed by Florentine Smarandache, Neutrosophy is the foun- Fi; j 1  Ti; j: 7

dation of neutrosophic logic, neutrosophic statistics, Neutrosophic
Set and neutrosophic probability [20,21]. In neutrosophic logic, where mi; j is the local median value of the w
w window
hAi is a theory, entity or event, and hAnti-Ai is the opposite of hAi. w 2n 1 n P 1 centered at i; j, and in a 3
3 window, it
 X. Qi et al. / J. Vis. Commun. Image R. 36 (2016) 110 3

Fig. 1. Images of indeterminacy in NS. (a) Original image. (b) Indeterminacy image of original Babala. (c) Indeterminacy of noisy image (r = 1%). (d) Indeterminacy of noisy
image (r = 10%).

Fig. 2. Standard images used for searching optimal K. (a) Lena. (b) Peppers. (c) Tree-hill. (d) Cat. (e) Buildings.

can be presented as Eq. (5). di; j denotes the absolute difference Ti; j and Fi; j are just to illustrate the process of transforming
between Pi; j and mi; j and Dif i; j is the maximum absolute an image from traditional field to NS domain, but they have not
luminance difference in working window. For an image with a size been used in noise reduction.
of M
N, dmax, dmin, Difmax and Difmin are described as
2.3. Properties of neutrosophic indeterminacy
dmax maxfds; tg; dmin minfds; tgs 2 M; t 2 N: 8
From Eqs. (35), we can deduce that if Pi; j has a similar value
Difmax maxfDif s; tg; Difmin minfDif s; tgs 2 M; t 2 N: 9 with the median of a squared window, its Ii; j will be lower;
therefore pixels in smooth area will have a lower indeterminacy.
From above formulas, we can conclude that if di; j is larger, While in edge zones, due to the stronger gray fluctuation, pixels
Pi; j will has a lower reliability, and vice versa. So normalized will have a slightly larger indeterminacy. Nonetheless, compared
by the factor of dmax  dmin , di; j can estimate the indeterminacy with the vibration intensity of SPN, the gray variation on edge is
degree of Pi; j. Likewise, a lager Dif i; j signifies Pi; j has a looser comparatively moderate. This is because there are certain correla-
relation with surrounding pixels; hence, after normalized by tions between pixels in edge zones and their intensity variation are
Difmax  Difmin , it has the capability to calculate the truth degree relatively continuous. Yet SPN has no relevancy with their neigh-
of Pi; j being signal pixel. In this paper, the introductions of borhoods and they only take the maximum or the minimum value

Fig. 3. The relation between K and the PSNR at different noise levels. (a) Lena image. (b) Cat image.
4 X. Qi et al. / J. Vis. Commun. Image R. 36 (2016) 110

Fig. 4. The relation between K and the PSNR at certain noise level. (a) r = 10%. (b) r = 60%. (c) r = 70%. (d) r = 90%.

Table 1
The optimal values of K at every peak for different images with various noise density.

Images Optimal values of K

10% 20% 30% 40% 50% 60% 70% 80% 90%
Building 0.0653 0.0718 0.0790 0.0790 0.0790 0.0956 0.0956 0.0956 0.1399
Cat 0.0593 0.0653 0.0718 0.0790 0.0790 0.0869 0.0956 0.0956 0.1156
Tree-hill 0.0593 0.0653 0.0718 0.0790 0.0790 0.0869 0.0956 0.0956 0.1156
Pepper 0.0593 0.0653 0.0718 0.0790 0.0869 0.1051 0.1051 0.1051 0.1399
Lena 0.0653 0.0718 0.0718 0.0790 0.0869 0.1051 0.1051 0.1051 0.1399

abruptly. Accordingly, noisy point will has a much higher indeter- 3. Adaptive neutrosophic weight function
minacy in comparison with points in boundary areas.
Taking the Babala image as an example, we observe the valu- The powerful Gaussian kernel function described by Eq. (10) has
able properties of neutrosophic indeterminacy. With a 3
3 filter- been widely exploited in spatial filtering [25]. In the proposed fil-
ing widow, the original and contaminated images are transformed ter, the weight of every pixel in the filtering window is modeled by
into NS domain by using Eqs. (37), and their indeterminacy this function, too.
images are shown in Fig. 1. In weak texture or edge area such as  
x2
the ellipse region in Fig. 1(b), pixels uncertainty are slightly stron- Fx exp  2 : 10
ger, while in flat region such as the rectangle area, they are 2h
decreased even down to zero. However, as shown in Fig. 1 In Eq. (10), h controls the degree of decay and affects the denois-
(c) and (d), the indeterminacy of noisy pixels in images corrupted ing performances, and it is related not only to the density of noise
with 1% and 10% noise density has been significantly enhanced. but also to the contents of images. The bigger the value of the
Consequently, from above-descriptions, a conclusion can be parameter is, the smoother the restored image is, and vice versa.
drawn that indeterminacy information has the capability to distin- In [31], B. Smolka has proposed that image structure can be used
guish the type of the area that the current pixel locates in. to acquire the optimal h. Since neutrosophic indeterminacy can
 X. Qi et al. / J. Vis. Commun. Image R. 36 (2016) 110 5

P
y2Dx wy
gy
gx0 P : 11
y2Dx wy

where gx0 is the denoising result of gx, wy denotes the weight

of the pixel at the location y, and D(x) is the filtering window cen-
tered at x.

3.1. Similarity measurement

The weight function w() is determined by the similarity of pix-

els in D(x) and plays a crucial role in noise removal. The fundamen-
tal difference of various filters is the strategy of designing weight
function, in which a suitable and efficient feature used to measure
pixel similarity is pretty critical. Spatial distance has been used in
Gaussian filters and intensity distance has been taken into account
in Yaroslavsky neighborhood filter [22]. While Bilateral Filter con-
siders both Euclidean distance and gray-level of the neighboring
Fig. 5. Local zones in tested images. (a) Baboon. (b) Lena.
pixels without smoothing edges [23]. In addition, gradient infor-
mation has been adopted in [24]. In this paper, the feature of neu-
trosophic indeterminacy is innovatively regarded as a criterion of
accurately discern the property of areas, it can be exploited to measuring similarity. Just as mentioned above, if a pixel has a
design the function for an adaptive h, and this will be discussed lower indeterminacy, it signifies the current pixel has a reasonably
in Section 3.2. higher similarity and closer correlation with surrounding pixels.
The basic idea of image filtering is to update the detected noise Therefore, it should be allocated a larger weight in denoising. By
by using its neighborhood information, and the restored value of contrast, if the uncertainty value is much stronger, this implies
the current pixel can be estimated as the current pixel has a much lower homogeneity and similarity

Fig. 6. Results of different filters for Baboon image. (a) Noisy image with 60% Salt and Pepper noise. (b) MF. (c) AMF. (d) DBA. (e) NEF. (f) MDBUTM. (g) FBDA. (h) DBTMF. (i)
AICF. (j) TSF. (k) ANWF. (l) Original image.
6 X. Qi et al. / J. Vis. Commun. Image R. 36 (2016) 110

Fig. 7. Results of different filters for Lena image. (a) Noisy image with 90% Salt and Pepper noise. (b) MF. (c) AMF. (d) DBA. (e) NEF. (f) MDBUTM. (g) FBDA. (h) DBTMF. (i) AICF.
(j) TSF. (k) ANWF. (l) Original image.

with neighboring pixels, so it should be distributed a weaker the standard deviation and the recommended value of K is
weight. In conclusion, indeterminacy can be selected as a robust between 10 and l5. In [7,16,26,27], methods of adaptive h have
characteristic of measuring similarity. been studied deeply, but its value is maintained the same for all
Modeled by Gaussian kernel function, an adaptive neutrosophic pixels in the whole image. However, in different parts of the image,
weight function based on indeterminacy can be considered as it is rather difficult to find a globally fixed optimal parameter that
below, can simultaneously ensure perfect denoising result and effective
! protection of details [28].
1 Iy2 In detailed area, a smaller h is conducive to preserve the edges
wy exp 2
; 12
C1 2h and textures, while in a flat region, a larger one can make all neigh-
boring pixels almost have the equal weight and can enhance the
!
X Iy2 smoothness of the restored area. That is to say, the value of h
C1 exp 2
: 13 should be modified in different type of regions. Since pixels inde-
y2Dx 2h
terminacy has the faulty to distinguish the type of a region, it can
be used to adjust the value of h, so a pixel-wise adaptive decay
where I(y) is the indeterminacy of the pixel locating at y and calcu-
function on base of indeterminacy can be expressed as
lated by Eqs. (3)(5). h is the smoothing parameter of Gaussian ker-
nel function and determines the approximate threshold above hy exp2Iy=Imax =K: 14
which to penalize high indeterminacy value. C1 is the normalization
where h(y) is the smoothing parameter of the pixel locating at y,
constant described as Eq. (13). Since I(y) varies from point to point,
and I(y) is the pixels indeterminacy. Imax is the maximum in uncer-
the value of w(y) can be adaptively regulated with the variation of
tain map of the corrupted image, and the coefficient 2 before I(y) is
pixels indeterminacy.
used to reinforce the role of indeterminacy on h(y).
Combining formulas (1214), the weight of every pixel in the
3.2. Pixel-wise adaptive smoothing parameter
filtering window can be obtained as
In most denoising methods, h is selected by trial and error and 1  
kept constant in the whole image. For example, in [14], Buades wy exp Iy2 =2exp 2Iy=Imax =K 2 ; 15
C2
et al. suggested h = K b for Gaussian noise reduction, where b is
 X. Qi et al. / J. Vis. Commun. Image R. 36 (2016) 110 7

Fig. 8. Results of different filters for Man image. (a) Noisy image with 90% Salt and Pepper noise. (b) MF. (c) AMF. (d) DBA. (e) NEF. (f) MDBUTM. (g) FBDA. (h) DBTMF. (i) AICF.
(j) TSF. (k) ANWF. (l) Original image.

Fig. 9. Partial enlarged details of DBTMF and ANWF. (a) Partial enlarged detail of Lena image with DBTMF. (b) Partial enlarged detail of Lena image with ANWF. (c) Partial
enlarged detail of Man image with DBTMF. (d) Partial enlarged detail of Man image with ANWF.

X  
C2 exp Iy2 =2exp2Iy=Imax =K2 : 16 2552
PSNR 10 log ; 17
y2Dx MSE
As an automatic way of adjusting h(y) pixel by pixel, our object PM1 PN1
i0 j0 jYi; j  Zi; jj
is to let K play the dominant role while I(y) acts as an auxiliary fac- MAE ; 18
tor. In other words, K keeps constant while h(y) varies in accor- MN
dance with I(y). To achieve this objective, it is necessary to where MSE is expressed as
research the laws between the optimal K and a perfect denoising PM1 PN1 2
performance. The performance criterions applied in this paper i0 j0 Yi; j  Zi; j
MSE : 19
are PSNR (Peak Signal to Noise Ratio) and MAE (Mean Absolute MN
Error) defined by
8 X. Qi et al. / J. Vis. Commun. Image R. 36 (2016) 110

Table 2
Comparisons of different filters for Baboon image at different noise levels.

Noise density (%) Criterion PSNR and MAE of different filters for restored images (Baboon)
MF [1] AMF [4] DBA [7] NEF [17] MDBU TMF [9] FBDA [10] DBTMF [11] AICF [13] TSF [12] ANWF Ours
10 PSNR 23.06 23.09 31.64 25.41 32.15 26.57 24.81 32.35 32.78 31.69
MAE 10.73 10.23 1.39 8.33 1.31 3.85 4.95 1.36 1.35 1.38
20 PSNR 21.05 22.32 28.12 24.85 28.79 26.05 24.31 29.06 29.95 28.35
MAE 14.83 11.38 2.93 9.01 2.71 4.49 5.92 2.83 2.51 2.85
30 PSNR 20.80 21.50 25.93 24.02 26.75 25.38 23.72 27.29 27.87 26.46
MAE 15.21 12.72 4.62 10.59 4.21 5.38 7.05 4.27 3.85 4.34
40 PSNR 20.27 20.81 24.32 24.88 25.30 24.61 23.21 24.83 25.09 25.10
MAE 16.71 14.04 6.47 9.90 5.75 6.50 8.17 6.92 6.78 5.85
50 PSNR 19.94 20.16 22.81 24.11 23.82 23.65 22.58 22.94 23.87 23.84
MAE 17.29 15.38 8.63 10.33 7.55 7.93 9.46 9.84 8.50 7.59
60 PSNR 19.57 19.49 21.44 23.18 22.49 22.69 21.96 22.10 22.30 22.94
MAE 18.31 16.89 11.11 11.36 9.65 9.63 10.87 11.39 10.27 9.04
70 PSNR 18.65 18.60 20.05 21.34 20.77 21.20 21.17 21.10 21.64 21.90
MAE 19.96 18.86 14.14 13.89 12.65 11.18 12.62 13.69 12.72 11.44
80 PSNR 14.79 16.09 18.77 15.77 18.44 20.28 20.24 19.42 20.04 20.43
MAE 28.70 24.15 17.88 33.77 17.84 14.65 14.90 19.18 15.34 14.32
90 PSNR 9.47 11.28 17.15 8.44 15.67 19.11 18.91 18.62 19.32 19.53
MAE 61.55 45.15 23.24 74.58 27.84 18.02 18.48 22.50 17.93 17.24

Table 3
Comparisons of different filters for Lena image at different noise levels.

Noise density (%) Criterion PSNR and MAE of different filters for restored images (Lena)
MF [1] AMF [4] DBA [7] NEF [17] MDBU TMF [9] FBDA [10] DBTMF [11] AICF [13] TSF [12] ANWF Ours
10 PSNR 34.47 35.02 42.23 32.84 43.83 37.51 39.95 42.04 43.91 43.23
MAE 2.20 1.88 0.34 3.42 0.29 0.89 0.41 0.38 0.18 0.31
20 PSNR 30.63 32.80 37.97 31.92 39.87 36.20 36.97 39.06 40.21 38.98
MAE 3.92 2.49 0.76 3.73 0.63 1.15 0.83 0.68 0.41 0.67
30 PSNR 29.76 31.01 35.06 30.47 37.37 35.29 34.80 36.89 37.32 36.97
MAE 4.28 3.16 1.28 4.83 1.02 1.40 1.33 1.16 0.86 1.03
40 PSNR 27.79 29.44 32.32 28.38 34.95 34.15 33.11 34.02 34.89 34.75
MAE 5.44 3.91 1.96 6.61 1.52 1.77 1.89 1.91 1.22 1.51
50 PSNR 26.87 28.28 30.42 27.90 32.34 33.02 31.75 32.24 32.86 33.36
MAE 5.90 4.62 2.72 6.50 2.09 2.17 2.47 2.70 1.79 1.97
60 PSNR 25.48 26.94 28.13 26.22 28.65 31.02 30.50 30.98 31.24 31.66
MAE 7.04 5.44 3.79 7.32 3.06 3.12 3.13 3.39 2.41 2.58
70 PSNR 22.53 23.81 25.57 24.32 24.04 29.93 28.92 28.53 29.21 29.97
MAE 8.84 7.02 5.04 8.79 5.43 3.48 4.03 4.97 3.26 3.41
80 PSNR 15.82 18.29 23.12 19.35 19.36 27.75 27.13 26.43 27.40 28.19
MAE 17.79 12.08 7.08 17.79 11.21 4.61 5.23 6.91 4.73 4.50
90 PSNR 9.42 11.52 19.58 11.38 15.06 24.52 24.12 23.89 24.96 25.43
MAE 53.79 35.32 13.07 41.74 25.34 7.05 7.74 11.08 6.93 6.64

where Y denotes the original image with a resolution of M
N and Z To further verify the persuasiveness and universality of above
is the reconstructed one. conclusions, extensive experiments on a large number of images
Tested images used in this paper are listed in Fig. 2. Taking Lena have been implemented repeatedly. Fig. 4(a)(d) are the typical
and Cat images as examples, Fig. 3 indicates the relation between K results of the contaminated images with the noise density of
and PSNR performance at various noise level (r = 1080%). It is can 10%, 60%, 70% and 90%, respectively. From these results, it is can
be noted that When K <= 0.002, PSNR values are almost stable, and be seen that the law between PSNR and K almost remains the same
around K = 0.07, a distinct peak of PSNR appears. But with the as the rule shown in Fig. 3. To diverse images, the optimum values
increasing of K, PSNR values descent sharply. When the value of of K under various noise level are shown in Table 1.
K is near to 1.4, the results remain stable again. When r >= 60%, As can be seen from Table 1, to a same image with a varying
two peaks appear in these curves. The first peak is around noise density from 10% to 90%, the values of K are not precisely
K = 0.002 and the other is near to 0.07, but the corresponding PSNR the same. Likewise, to diverse images with the same noise density,
of the second peak is much higher than that of the first one. the values of K are not fully identical, either. In a sense, these data
 X. Qi et al. / J. Vis. Commun. Image R. 36 (2016) 110 9

Table 4
Comparisons of different filters for Man image at different noise levels.

Noise density (%) Criterion PSNR and MAE of different filters for restored images (Man)
MF [1] AMF [4] DBA [7] NEF [17] MDBU TMF [9] FBDA [10] DBTMF [11] AICF [13] TSF [12] ANWF Ours
10 PSNR 28.77 29.28 36.69 26.65 36.17 31.20 31.77 37.06 38.12 36.23
MAE 4.84 4.37 0.73 8.69 0.80 2.10 1.32 0.71 0.64 0.84
20 PSNR 25.78 27.83 33.24 23.60 32.90 30.22 30.45 34.03 34.53 33.03
MAE 7.29 5.19 1.48 8.34 1.53 2.58 2.08 1.43 1.39 1.56
30 PSNR 25.32 26.51 30.66 25.76 30.27 29.29 29.28 29.31 32.45 31.25
MAE 7.65 6.08 2.36 9.09 2.40 3.17 2.86 3.39 2.06 2.22
40 PSNR 23.85 25.28 28.46 25.74 28.07 28.23 28.12 28.91 30.86 29.47
MAE 9.28 7.11 3.46 7.83 3.42 3.94 3.79 3.91 2.82 3.15
50 PSNR 23.32 24.27 26.59 25.09 25.92 27.05 27.15 27.62 29.32 28.15
MAE 9.80 8.12 4.75 8.05 4.67 4.86 4.72 5.06 3.66 4.04
60 PSNR 22.27 23.12 24.84 24.03 23.21 26.02 26.20 26.00 27.08 26.89
MAE 11.22 9.33 6.37 7.90 6.59 5.90 5.76 6.78 5.28 5.09
70 PSNR 20.38 21.31 22.87 22.00 20.05 24.90 25.07 24.59 25.49 25.70
MAE 12.99 11.05 8.56 11.53 10.16 7.25 7.08 8.65 6.70 6.32
80 PSNR 14.99 16.90 20.68 17.71 16.53 23.22 23.52 23.02 23.59 24.07
MAE 21.83 16.47 11.90 21.55 17.71 9.26 8.85 11.15 9.10 8.12
90 PSNR 8.82 10.61 17.89 10.40 12.84 20.84 21.17 20.88 21.90 22.04
MAE 56.73 39.88 18.26 46.40 34.88 13.28 12.30 15.43 11.14 11.22

can explain the reason why a globally constant smoothing param- 4. Experimental results
eter is inadvisable. In addition, it is clear that when r >= 60%, the
values of K are relatively larger than those of r < 60%. In order to In ANWF, if a pixel has a gray value of 0 or 255 in 8-bit images, it
simplify the assigning of K, when r < 60%, as the median of all val- is labeled as noise candidate and its location is recorded in the
ues, 0.0718 is selected as the most optimal value of K, and 0.0956 is noise matrix BNM. To a noisy pixel, the corresponding value in
regarded as the optimum in other cases. On the whole, the con- BNM is 0, otherwise, it is 1. In Eq. (20), r is used to determine
stant K can be assigned as follows the value of K, but from Figs. 3 and 4 and Table 1, it is can be
deduced that in a certain smaller range, noise density has small
 impact on K. Accordingly, a rough estimation of r can meet our
K 0:0718; r < 60%
20 requirements. Hence it can be roughly estimated as follows
K 0:0956; others
After replacing K with 0.0718 or 0.0956 in Eq. (14), the value of number of 00 s in BNM
r : 21
h(y) will be different from point to point on account of a pixel-wise total number of pixels in the image
I(y). For example, in Fig. 5, Baboon and Lena are contaminated by
In filtering stage, noise candidates are disposed while noise-free
the noise with density of 90%, and two local zones in these two
points are kept intact. What is more, only the uncontaminated pix-
images are selected, respectively. In a 3
3 filleting window, the
els are exploited to replace the corrupted one.
smoothing parameter matrixes hB1 and hB2 of Baboon and hL1 and
The proposed ANWF is compared with some state-of-the-art
hL2 of Lena are indicated below.
methods, such as MF [1], AMF [4], DBA [7], NEF [17], MDBUTMF
2 3 [9], FDBA [10], DBTMF [11], TSF [12] and AICF [13]. In order to com-
40:3627 1:8849 4:0974 pare fairly, MF, NEF, AMF and TSF are operated with their most suit-
6 7 able window sizes changing from 3
3 to 9
9, while ANWF and
hB1 4 13:9276 13:9276 13:9276 5
3:1879 13:9276 8:2349 the left algorithms are implemented with a fixed 3
3 window.
2 3 Extensive experiments have been performed on various images cor-
12:9783 13:3920 5:3495
6 7 rupted by SPN with varying noise density from 10% to 90% with
hB2 4 6:1125 5:9704 10:5842 5 increment of 10%. All experiments are executed in MATLAB
5:3495 5:2251 5:1438 R2013a and run on a PC with Intel Core i7-4700MQ 2.4 GHz CPU.

2 3 4.1. Visual quality evaluation

1:8849 1:8849 1:8849
6 7
hL1 4 3:7587 3:5025 13:9726 5
Visual comparisons of the recovered images such as Lena,
3:7587 13:9276 13:9726 Baboon and Man are shown in Figs. 69, respectively. In Fig. 6,
2 3
5:3077 13:8188 5:0242 when r = 60%, all filters can eliminate noise efficiently, but image
6 7 details are blurred slightly by MF and NEF. Moreover, it may be
hL2 4 5:3916 4:9460 5:2663 5
the iterative operation of c-median that makes the blurring of
1:8849 5:4341 4:9460
image is much heavier in Fig. 6(e) than in Fig. 6(b). When noise
It is obvious that in hB1, hB2, hL1 and hL2, the value of every pixels level soars to 90%, the performances of MF, AMF, NEF, DBA and
smoothing parameter is adaptively regulated from point to point. MDBUTMF are all influenced badly, and as shown in Figs. 7 and
So far, the object that the decay parameter can be pixel-wise 8, noise blocks increase considerably in their restored images. In
adjusted according to the specific contents of images has been the images recovered by FBDA, noisy points have not been elimi-
achieved. nate completely and a lot of stains are still remained. In addition,
10 X. Qi et al. / J. Vis. Commun. Image R. 36 (2016) 110

images restored by ACIF look fairly hazy. The visual qualities of [4] R.H. Chan, C.W. Ho, M. Nikolova, Salt-and-pepper noise removal by median-
type noise detectors and detail-preserving regularization, IEEE Trans. Image
DBTMF and TSF are as excellent as ANWF; however, as it seen from
Process. 14 (10) (2005) 14791485.
the partially enlarged details in Fig. 9(a)(d), the jagged edges of [5] H. Ibrahim, N.S.P. Kong, T.F. Ng, Simple adaptive median filter for the removal
DBTMF are more evident than that of ANWF. Although edges are of impulse noise from highly corrupted images, IEEE Trans. Consum. Electron.
preserved satisfactorily by TSF, unfortunately, the over- 54 (2008) 19201927.
[6] S.J. Horng, L.Y. Hsu, T. Li, Using sorted switching median filter to remove high-
smoothing phenomenon is pretty apparent, which leads to a bad density impulse noises, J. Vis. Commun. Image Represent. 24 (2013) 956967.
visual effect. For example, the textures of Lena hat and the mans [7] K.S. Srinivasan, D. Ebenezer, A new fast and efficient decision-based algorithm
hair cannot be clearly identified in Figs. 7(j) and 8(j). In a word, for removal of high-density impulse noises, IEEE Signal Process. Lett. 14 (3)
(2007) 189193.
ANWF performs significantly better than many other existing tech- [8] X. Zhang, Y. Xiong, Impulse noise removal using directional difference based
niques in term of visual quality. noise detector and adaptive weighted mean filter, IEEE Signal Process. Lett. 16
(4) (2009) 295298.
[9] S. Esakkirajan, T. Veerakumar, Adabala N. Subramanyam, et al., Removal of
4.2. Quantitative evaluation high density salt and pepper noise through modified decision based
unsymmetric trimmed median filter, IEEE Signal Process. Lett. 18 (5) (2011)
To further demonstrate the effectiveness of ANWF, the compar- 287290.
[10] Madhu S. Nair, G. Raju, A new fuzzy-based decision algorithm for high-density
ative results of PSNR and MAE of the ten methods are shown in impulse noise removal, Signal Image Video Process. 6 (2012) 579595.
Tables 24. When noise level is less than 40%, the most competitive [11] V.R. Vijaykumar, G. Santhanamari, New decision-based trimmed median filter
PSNR and MAE performances are produced by TSF. In Table 2, when for high-density salt-and-pepper noise removal in images, J. Electron. Imag. 23
(3) (2014) 033011033014.
noise level approximately rise to 5060%, the PSNR value of NEF
[12] G. Gao, Y. Liu, An efficient three-stage approach for removing salt and pepper
excels ANWF, but from Fig. 6(e) and (j), it is can be detected that noise from digital images, Int. J. Light Electron Opt. 126 (2015) 467471.
ANWF has more prominent visual effect than the former. Espe- [13] X.Q. Zhang, F. Ding, Z.H.J. Tang, Salt and pepper noise removal with image
cially when noise level amounts to 7090%, the proposed algo- inpainting, Int. J. Electron. Commun. 69 (2015) 307313.
[14] A. Buades, B. Coll, J.M. Morel, A review of image denoising algorithms, with a
rithm shows its superior performances in PSNR and MAE. new one, Multiscale Model. Simul. 4 (2005) 490530.
[15] D. Van De Ville, M. Kocher, SURE-based non-local means, IEEE Signal Process.
Lett. 16 (2009) 973976.
5. Conclusion [16] T. Tasdizen, Principal neighborhood dictionaries for nonlocal means image
denoising, IEEE Trans. Image Process. 18 (2009) 26492660.
In this paper, an adaptive neutrosophic filter based on pixels [17] Y.H. Guo, H.D. Cheng, Y. Zhang, A new neutrosophic approach to image
denoising, New Math. Natural Comput. 5 (2009) 653662.
indeterminate information has been presented to remove high-
[18] M. Zhang, L. Zhang, H.D. Cheng, A Neutrosophic approach to image
density SPN. In order to accurately measure pixel similarity, a segmentation based on watershed method, Signal Process. 90 (2010) 1510
novel and effective criterion based on Neutrosophic indeterminacy 1517.
[19] Y.H. Guo, H.D. Cheng, New Neutrosophic approach to image segmentation,
is utilized in the weight function. Meanwhile, for the purpose of
Pattern Recogn. 42 (2009) 587595.
adjusting the smooth parameter according to the contents of [20] P.A. Loeb, M. Wolff, Nonstandard Analysis for the Working Mathematician,
image, the Neutrosophic indeterminacy is exploited to distinguish Kluwer, Dordrecht, Netherlands, 2000.
the area type to which a pixel belongs, and as a regulatory factor, it [21] Y.H. Guo, A. Sengur, A novel color image segmentation approach based on
neutrosophic set and modified fuzzy c-means, Circ. Syst. Signal Process. 32
controls the variation of smoothing parameter in cooperation with (2013) 16991723.
a constant K. With a fixed 3
3 window, the proposed method can [22] B. Zhang, J.P. Allebach, Adaptive bilateral filter for sharpness enhancement and
obtain remarkable performances in terms of visual quality and noise removal, IEEE Trans. Image Process. 17 (2008) 664678.
[23] A. Senjian, B. Farid, B. Mohammed, S. Ferdous, Quantitative error analysis of
quantitative evaluation especially for high-level SPN. bilateral filtering, IEEE Signal Process. Lett. 22 (2015) 202206.
[24] P. Perona, Malik, Scale-space and edge detection using anisotropic diffusion,
Acknowledgments IEEE Trans. Pattern Anal. Mach. Intell. 12 (1990) 629639.
[25] J. Portilla, V. Strela, M.J. Wainwright, E.P. Simoncelli, Image denoising using
scale mixtures of Gaussians in the wavelet domain, IEEE Trans. Image Process.
This work was supported by the National Natural Science 12 (2003) 13381351.
Foundation of China (Grant No. 61203330), the Department of [26] G. Xu, J. Tan, A universal impulse noise filter with an impulse detector and
nonlocal means, Circ. Syst. Signal Process. 33 (2014) 421435.
Science and Technology of Shandong Province (Grant No.
[27] M. Nasri, S. Saryazdi, H. Nezamabadi-pour, A fast adaptive salt and pepper
2015ZDXX0801A01) and the Fundamental Research Funds of noise reduction method in images, Circ. Syst. Signal Process. 32 (2013) 1839
Shandong University (Grant No. 2015QY001). 1857.
[28] T. Brox, D. Cremers, Iterated nonlocal means for texture restoration, in:
Proceedings of the 1st International Conference on Scale Space and Variational
References Methods in Computer Vision, vol. 4485, 2007, pp. 1324.
[29] A.Q. Ansari, Ranjit Biswasb, Swati Aggarwalc, Neutrosophic classifier: an
[1] J.W. Tukey, Nonlinear methods for smoothing data, in: Proceedings of extension of fuzzy classifier, Appl. Soft Comput. 13 (2013) 563573.
Electronics and Aerospace Systems Conference, 1974, pp. 673681. [30] Abdulkadir Sengur, Yanhui Guo, Color texture image segmentation based on
[2] M. Nasri, S. Saryazdi, H. Nezamabadi-pour, SNLM: a switching non-local means neutrosophic set and wavelet transformation, Comput. Vis. Image Understand.
filter for removal of high density salt and pepper noise, Scientia Iranica 20 115 (2011) 11341144.
(2013) 760764. [31] B. Smolka, R. Lukac, A. Chydzinski, Fast adaptive similarity based impulsive
[3] K.S. Srinivasan, D. Ebenezer, A new fast and efficient decision-based algorithm noise reduction filter, Real-Time Imag. 9 (2003) 261276.
for removal of high-density impulse noises, IEEE Signal Process. Lett. 14 (2007)
189192.