Signal Processing

Manuscript Draft

Manuscript Number: SIGPRO-D-19-01568

Title: An effective weighted vector median filter for impulse noise

reduction based on minimizing the degree of aggregation

Article Type: Research Paper

Keywords: impulse noise;

noise removal;
aggregated distance;
degree of aggregation

Abstract: Impulse noise is regarded as an outlier in the local window of

an image. To detect noise, many methods based on noise-like features have
been proposed, including the aggregated distance, spatially weighted
aggregated distance, n nearest neighbor distance, local density, and
angle-weighted quaternion aggregated distance methods. However, the
noise-like features calculated by these methods ignore the weight of each
pixel. This paper introduces the concept of degree of aggregation and
proposes a weighting method that minimizes the degree of aggregation to
obtain the weight vector of the pixels. The obtained weight vector has a
larger component on the signal pixel than on the noise pixel. This weight
vector is then fused with the aggregated distance to form a weighted
aggregated distance that can reasonably characterize the noise and
signal. The weighted aggregated distance, along with an adaptive
segmentation method with a double-ended threshold, can effectively detect
the noise. An adaptive selection strategy is also used to replace the
noise pixel with the weighted channel combination optimization estimate
value. The experimental results exhibit the validity of the proposed
method by showing better performance in terms of both objective criteria
and visual effects.
*Manuscript
Click here to view linked References

An effective weighted vector median filter for impulse noise

reduction based on minimizing the degree of aggregation
Xiangxi Menga , Tongwei Lub,∗ , Feng Minb and Tao Lub
a School of Computer Science and Engineering,Wuhan Institute of Technology ,430074 Wuhan,China
b Hubei Key Laboratory of Intelligent Robot(Wuhan Institute of Technology) ,430074 Wuhan,China

1 ARTICLE INFO ABSTRACT

2 Keywords: Impulse noise is regarded as an outlier in the local window of an image. To detect noise, many
3 impulse noise methods based on noise-like features have been proposed, including the aggregated distance,
4 noise removal spatially weighted aggregated distance, n nearest neighbor distance, local density, and angle-
5 aggregated distance weighted quaternion aggregated distance methods. However, the noise-like features calculated
6 degree of aggregation by these methods ignore the weight of each pixel. This paper introduces the concept of degree
of aggregation and proposes a weighting method that minimizes the degree of aggregation to
7
obtain the weight vector of the pixels. The obtained weight vector has a larger component on
8 the signal pixel than on the noise pixel. This weight vector is then fused with the aggregated
9 distance to form a weighted aggregated distance that can reasonably characterize the noise and
10 signal. The weighted aggregated distance, along with an adaptive segmentation method with a
11 double-ended threshold, can effectively detect the noise. An adaptive selection strategy is also
used to replace the noise pixel with the weighted channel combination optimization estimate
12
value. The experimental results exhibit the validity of the proposed method by showing better
13 performance in terms of both objective criteria and visual effects.
14
15
16
17
18 1. Introduction
19 Digital images have become integral to people’s production and daily lives. However, due to sensor errors during
20 imaging, hardware failure during image storage, and signal interference during image transmission, images are often
21 contaminated by noise. Impulse noise is a very common type of noise that can be found in images[1, 2]. When
22 an image contains impulse noise, the contaminated pixels change to a random value, while the other pixels remain
23 unchanged. Moreover, the intensity (single-channel image) or intensity vector (three-channel image) of impulse noise
24 has a relatively large difference from the spatial neighborhood.
25 Generally speaking, color image filtering techniques can be categorized as component-wise methods, vector based
26
methods, and other methods such as wavelet domain denoising methods[3, 4], convolutional neural network methods[5–
27
8], partial differential equation(PDE) methods[9], low-rank methods[10–14]. Component-wise methods separately
28
29 operate on individual color channels, so it can make full use of many existing filter technologies for grayscale images.
30 However these methods neglect the inherent correlation between different channels, which may result in pixel output
31 values that are different from the input values with possible chromaticity shifts[15]. In contrast, vector based filter
32 treats the color pixel as a vector, so it is more effective for the color image impulse noise removal[16]. Many popular
33 filtering approaches, such as median filters[17], weighted median filters[18], are based on the theory of robust order
34 statistics. The middle-positioned samples in the ordered sequence represent the robust estimates. However, these
35 classical median type of vector filters introduce too much smoothing regardless of whether image pixels are corrupted
36 by impulse noise or not[15]. So switching filters, also known as switching vector median filter (SVMF)[19–22] have
37 been developed. Many other vector based methods have been published such as the vector median filter(VMF)[23],
38 robust vector median filter[24], vector directional filter(BVDF)[25], directional distance Filter [26], adaptive vector
39 median filter[27], peer group filter [28], center weighted vector median filter[29], adaptive center-weighted vector di-
40 rectional filter[30], vector sigma filter[31],fuzzy logic vector filter[32, 33], hybrid vector filters[15],adaptive weighted
41 quaternion distance filter (AWQD)[34].
42 However, such as the rank-ordered absolute differences (ROAD)[35] and fast averaging peer group filter (FAPGF)[28],
43 have used the local density of local pixel groups in the color space as a noise-like feature. The ROAD value of a pixel
44 is computed as the sum of the Euclidean distances between the pixel and the four closest (color field) pixels in the local
45
46 ORCID (s):
47
48
Xiangxi Meng et al.: Preprint submitted to Elsevier Page 1 of 14
49
 An effective weighted vector median filter

window. Meanwhile, FAPGF first defines a threshold distance of a color space and then calculates the number of other
pixels in the window that contains this same threshold range. This number is used as an independent variable to define
a function that is regarded as a noise-like feature. Although the local density information can reflect the characteristics
of the pixel (local density is higher, more signal-like; local density is lower, more noise-like), such information greatly
depends on parameters and therefore cannot easily adapt to different noise levels.
The adaptive vector median filter (AVMF)[27] uses the Euclidean distance between the mean vector of several
sample vectors, which are close to the vector median and the current color vector, to determine the noise, and then, the
noise is replaced by the output of the standard VMF. Meanwhile, the fast peer group filter for VMF (FPGFVMF)[19],
regards the current center pixel in the local window as a signal pixel only if the number of surrounding pixels, which are
1 close enough to the current center pixel (in color space), is sufficiently large. These two types of switching mechanisms
2 use fixed thresholds and parameters, and they heavily depend on parameters too.
3 Analyzing the above noise removal filters, aggregated distance[36] has been widely used as a noise-like feature
4 in several methods. The aggregated distance of a pixel is computed as the sum of distances from the pixel to its
5 surrounding pixels. However, the distance to the noise pixels’ value does not have much meaning compared with
6 the distance to the signal pixels’ value because impulse noise usually does not have valid information for the original
7 image. To reasonably characterize a pixel, we first propose the concept of a degree of aggregation, based on which
8 we propose a pixel weighting method that minimizes the degree of aggregation (where the weight vector’s component
9 obtained is smaller on the noise pixel than on the signal pixel, which can reduce the effect of noise on the noise-
10 like feature calculation) and then adds these appropriate weights to form a weighted aggregated distance. Then an
11 adaptive data segmentation method for weighted aggregated distance sequence is designed to adaptively detect the
12 noise pixels. In the image recovery phase,we uses the values optimized by channel combination to replace the noise
13 pixels. Experimental results show that the proposed method exhibits competitive results over other state-of-the-art
14 color image denoising methods.
15
16 The rest of this paper is organized as follows. Section 2 discusses the noise analysis and detection method. Section 3
17 describes the noise replacement. Section 4 presents the experiment results. Finally, the conclusion is drawn in Section
18 5.
19
20
21 2. Noise analysis and detection
22
2.1. Degree of aggregation and minimum aggregation weighting method
23
In the partial window (n*n) (for the convenience of description, the window size is set to 3*3 hereinafter), the pixel
24
25 group in the window is mapped to the nodes in the color space. These nodes are connected by undirected edges, where
26 the value of each edge is computed as the Euclidean distance of the two nodes associated with the edge. In this case,
27 these nodes and edges can be used to form a complete graph. The adjacency matrix D of the graph is built as follows
28 (where 𝐷𝑖,𝑗 represents the color distance between the i’th pixel and the j’th pixel):
29
30
⎛ 0 𝐷0,1 𝐷0,2 𝐷0,3 𝐷0,4 𝐷0,5 𝐷0,6 𝐷0,7 𝐷0,8 ⎞
31 ⎜𝐷 ⎟
0 ... ...
32 ⎜ 1,0 ⎟
33 ⎜𝐷2,0 ... 0 ⎟
34 ⎜𝐷3,0 0 ⎟
⎜𝐷4,0 0 ⎟. (1)
35 ⎜ ⎟
36 ⎜𝐷5,0 0 ⎟
37 ⎜𝐷6,0 0 ⎟
38 ⎜𝐷7,0 0 ⎟
⎜𝐷 0 ⎟⎠
39 ⎝ 8,0
40
41 The aggregated distance is defined as
42
43
44
45 ∑
𝑗=𝑛−1

46 𝐺𝑖 = 𝐷𝑖,𝑗 . (2)
𝑗=0
47
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 2 of 14
49
 An effective weighted vector median filter

The aggregated distance can be perceived as the sum of the distance of the pixels from the neighboring pixels. In
the matrix D the aggregated distance of the ith pixel is computed as the sum of elements of the ith row or column.
However, the distance (color space) of the pixels from the noise pixels in the window is not as important as the distance
(color space) from the signal pixels in the window. We want the aggregated distance to only include the distance to
the signal pixel and exclude the distance to the noise pixel. Otherwise, one cannot learn in advance which pixels are
noise and which are signals. To reduce the influence of noise as much as possible, we propose the concept of degree
of aggregation.

Given that the Euclidean distance from the node to itself is 0, the diagonal elements of the adjacency matrix D are
all 0. The adjacency matrix can be reasonably modified in the two ways:
1
2 1. Modify the diagonal element to the sum of the elements of the row or column.
3
4
5 ∑
𝑗=𝑛−1

6 𝐷𝑖,𝑖 = 𝐷𝑖,𝑗 (3)

7 𝑗=0
8
2. Modify the diagonal element to be slightly larger than the negative of the minimum eigenvalue 𝑀𝑣 of 𝐷.
9
10
11
𝐷𝑖,𝑖 = −1 ∗ 𝑀𝑣 + Δ (4)
12
13
14 The modified positive definite adjacency matrix 𝐷1 is obtained as follows:
15
16 ⎛𝐷0,0 𝐷0,1 𝐷0,2 𝐷0,3 𝐷0,4 𝐷0,5 𝐷0,6 𝐷0,7 𝐷0,8 ⎞
17 ⎜𝐷 ⎟
𝐷1,1 ... ...
18 ⎜ 1,0 ⎟
19 ⎜𝐷2,0 ... 𝐷2,2 ⎟
20 ⎜𝐷3,0 𝐷3,3 ⎟
⎜𝐷4,0 𝐷 4,4
⎟ (5)
21 ⎜ ⎟
22 𝐷
⎜ 5,0 𝐷 5,5 ⎟
23 ⎜𝐷6,0 𝐷6,6 ⎟
24 ⎜𝐷7,0 𝐷7,7 ⎟
⎜𝐷 𝐷8,8 ⎟⎠
25 ⎝ 8,0
26
27 Define the weight vector of the node 𝐖T :
28 ( )
29 𝑊0 𝑊1 𝑊2 𝑊3 𝑊4 𝑊5 𝑊6 𝑊7 𝑊8 ,
30
31 and its constraint:
32
33
34 ∑
𝑖=𝑛−1
𝑊𝑖 = 1
35
𝑖=0
36
37 𝑊𝑖 > 0 𝑖 >= 0 𝑎𝑛𝑑 𝑖 < 𝑛.
38
The initial value of the weight vector is 1/n.
39
The degree of aggregation is defined as
40
41
∑
𝑖,𝑗=𝑛−1 ∑
𝑖=𝑛−1
42 𝑋 = 𝑊 𝑇 ∗ 𝐷1 ∗ 𝑊 = 𝐷𝑖𝑗1
∗ 𝑊𝑖 ∗ 𝑊𝑗 (𝑖! = 𝑗) + 𝐷𝑖𝑖1 ∗ 𝑊𝑖 ∗ 𝑊𝑖 (6)
43 𝑖,𝑗=0 𝑖=0
44
45 The right part of formula (6) contains two terms: the first term is the weighted sum of the off-diagonal elements of
46 the adjacency matrix 𝐷1 , and the second term is the weighted sum of the diagonal elements. The first term enlarges
47
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 3 of 14
49
 An effective weighted vector median filter

the gap between the weight components, while the second term minimizes the weight component. In other words, the
penalty factor can also control the sensitivity of the change in the weight vector. If there is no noise in the local area,
its degree of aggregation is small, and the presence of noise increases the local window’s degree of aggregation. Then,
the degree of aggregation X is minimized, and the optimal weight vector obtained can weight more on the signal pixels
than on the noise pixels. Finally, the weight vector is merged into the aggregated distance, which is computed as the
weighted aggregated distance.
∑
𝑗=𝑛−1
1
𝐺𝑖𝑤 = 𝐷𝑖,𝑗 ∗ 𝑊𝑗 (7)
𝑗=0
1
The following example illustrates the process of the weighting method. Consider the nine pixels in the 3*3 window
2
3 that take the following values:
4
⎛ 100 101 101 101 99 100 100 100 100 ⎞
5 ⎜ 100 100 99 ⎟.
100 100 100 99 100 100
6 ⎜ ⎟
7 ⎝ 100 100 100 100 100 100 100 99 101 ⎠
8 Suppose that all the nine pixels are signal pixels. The initial weight vector is
9
( )
10 1∕9 1∕9 1∕9 1∕9 1∕9 1∕9 1∕9 1∕9 1∕9 .
11
12 According to formula(7), the initial weighted aggregated distance is
13 ( )
14 𝑖𝑛𝑖𝑡𝑖𝑎𝑙 ∶ 0.89 1.12 1.12 1.12 1.34 1.12 1.34 1.28 1.28 .
15 According to formula(6), the degrees of aggregation of different modifications (formulas(3) and (4)) are 2.36 and
16 1.645. Afterwards, we minimize the degree of aggregation to obtain the optimal weight vector:
17 ( )
18 𝑝𝑟𝑜𝑝𝑜𝑠𝑒𝑑1 ∶ 0.181 0.123 0.123 0.123 0.074751 0.123 0.074751 0.088 0.088
19 𝑝𝑟𝑜𝑝𝑜𝑠𝑒𝑑2 ∶ 0.184 0.185 0.185 0.185 0.000111 0.185 0.000111 0.036 0.036
20
and obtain that the new minimum aggregation agrees with 2.28 and 1.58. It can be found that the newly minimized
21
22 degree of aggregation is only slightly smaller than the original degree of aggregation. After all, the optimal weighted
23 aggregated distance is
24 ( )
𝑝𝑟𝑜𝑝𝑜𝑠𝑒𝑑1 ∶ 0.82 1.03 1.03 1.03 1.38 1.03 1.38 1.27 1.27
25 .
26 𝑝𝑟𝑜𝑝𝑜𝑠𝑒𝑑2 ∶ 0.82 0.81 0.81 0.81 1.56 0.81 1.56 1.31 1.31
27 In sum, when there is no noise, the weighted aggregated distances are similar and their value is small.
28
29 Consider the 9 pixels in the 3*3 window; the pixels value are
30
31 ⎛ 100 101 101 101 100 100 25 101 151 ⎞
32 ⎜ 100 99 99 98 102 99 99 180 50 ⎟ .
33 ⎜ ⎟
⎝ 100 100 99 100 100 100 200 20 100 ⎠
34
Among them, 𝑃0 to 𝑃5 represent the value of the signal pixels, while the others represent the value of the noise
35
36 pixels. The initial weight vector is
37 ( )
38 1∕9 1∕9 1∕9 1∕9 1∕9 1∕9 1∕9 1∕9 1∕9 .
39 According to formula(7), the initial weighted aggregated distance is
40
( )
41 𝑖𝑛𝑖𝑡𝑖𝑎𝑙 ∶ 35.33 35.23 35.38 35.52 36.13 35.27 125.81 116.94 83.68 .
42
43 According to formula(6), the degrees of aggregation are 119.851 and 95.142. After minimizing the degree of aggre-
44 gation, the following optimal weight vector is obtained.
45 ( )
𝑝𝑟𝑜𝑝𝑜𝑠𝑒𝑑1 ∶ 0.167 0.168 0.167 0.166 0.160 0.168 0.000111 0.000111 0.000111
46 ,
𝑝𝑟𝑜𝑝𝑜𝑠𝑒𝑑2 ∶ 0.167 0.167 0.167 0.166 0.163 0.167 0.000111 0.000111 0.000111
47
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 4 of 14
49
 An effective weighted vector median filter

and we obtain that the newly minimum aggregation agrees with 54.870 and 54.664, which are both smaller than the
initial degree of the aggregation. Moreover, the optimal weighted aggregated distance is
( )
𝑝𝑟𝑜𝑝𝑜𝑠𝑒𝑑1 ∶ 1.42 1.28 1.50 1.72 2.65 1.32 125.44 113.39 70.72
.
𝑝𝑟𝑜𝑝𝑜𝑠𝑒𝑑2 ∶ 1.43 1.29 1.51 1.72 2.64 1.33 125.44 113.39 70.73
It can be seen that the weighted aggregated distance used as a noise-like feature is more distinct between noise and
signal than the unweighted aggregated distance.

What should be mentioned is that the classic VMF can be regarded as a reduced version of the minimum of the
1 degree of aggregation weighting method. Considering the first setting of the diagonal elements (set to the aggregated
2 distance), if the adjacency matrix 𝐷1 ignores the off-diagonal elements (set to 0), then only the diagonal element 𝐷𝑖,𝑖
3 (the unweighted aggregated distance of node i) remains, as shown by matrix 𝐷2 below:
4
5 ⎧𝐷 0 ... ... ... ... ... ... ... ⎫
6 ⎪ 0,0 ⎪
7 ⎪ 0 𝐷1,1 ... ... ⎪
⎪ 0 𝐷2,2 ⎪
8 ⎪ 0 𝐷3,3 ⎪
9 ⎪ ⎪
10 ⎨ 0 𝐷4,4 ⎬
11 ⎪ 0 𝐷5,5 ⎪
⎪ 0 𝐷6,6 ⎪
12 ⎪ ⎪
13 ⎪ 0 𝐷7,7 ⎪
14 ⎪ 0 𝐷8,8 ⎪
⎩ ⎭
15
16 The same constraint minimum aggregation optimization is performed for matrix 𝐷2 as a reduced version of the
17 minimum aggregation optimization. By using the Lagrange multiplier method, the optimal solution vector 𝑊 1 is
18 obtained as
19
1
20 𝐷𝑖,𝑖
21 𝑊𝑖1 =∑ (8)
𝑖=𝑛−1 1
22 𝑖=0 𝐷𝑖,𝑖
23
24 In formula(8), the pixel with the minimum unweighted aggregated distance is selected by the VMF and has the largest
25 component on the weight vector 𝑊 1 . In this case, the VMF only considers the aggregated distance (diagonal elements)
26 and ignores the other nondiagonal elements of the adjacency matrix.
27
28 2.2. Noise detection based on adaptive segmentation of noise-like features
29 When pixels contaminated by impulse noise are present in a window, the complete graph becomes confusing. The
30 aggregated distance of the noise nodes greatly increases along with that of remaining signal pixels, but the amplitude of
31 the increase in the signals is much lower than that of the increase in the noise (shown in the example list in the Section
32
2.1). A segmentation method that can adaptively divide the aggregated distance sequence into signal and noise parts
33
can be applied to solve this problem. The weighted aggregated distance mentioned in section 2.1 is used as the data
34
35 sequence segmentation to find the noisy pixels. However, the segmentation of this sequence is not straightforward
36 because the former portion of the sequence that represents the signals’ noise-like features value is usually small and
37 convergent, and the following portion of the sequence representing the noises’ noise-like features is usually large and
38 divergent. In addition, the weighted aggregated distance of the signal nodes has a natural cluster, while that of the
39 noisy nodes has a natural divergence (which is a nature of noise). To address this problem, a segmentation method
40 based on the sigmoid function may be effective. The sigmoid function formula can be expressed as follows:
41
42
1
43 𝑓 (𝑥) = (9)
44 1 + 𝑒−(𝑥−𝑃 )
45 The curve of the sigmoid function is shown in Fig. 1, where the red squares in the curve represent the noises, the
46 green squares represent the signals, and the blue square represent the segmentation point P.
47
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 5 of 14
49
 An effective weighted vector median filter

1
2
3
4
5
6
7 Figure 1: Noise detection based on sigmoid segmentation
8
9
10
We want the signal part of the weighted aggregated distance sequence, which is small and convergent, to fit the
11
high slope area (which is near the center point and on the negative half-axis of the sigmoid function) and the remaining
12
13 noise part in the sequence, which is larger and divergent to fit the gentle slope area (which is far from the center point
14 and on the positive half-axis of the sigmoid function). The point P is considered as the segmentation point to determine
15 the noise and signal.
16 Because the standard sigmoid function has most of the range of slopes in the domain 0 and 4, which is sufficient for
17 model constrained optimization, a scale transformation is performed on the weighted aggregated distance sequence.The
18 adaptive segmentation algorithm consists of the following four steps:
19 1. The weighted aggregated distance is sorted in an ascending order.
20
21 2. The weighted aggregated distance is transformed between 0 and 4.
22 3. The scaled weighted aggregated distance is mapped to the sigmoid function’s value, 𝑓 (𝑥) in formula (9).
23 4. Optimal point P is obtained by using the maximum interclass variance method for the mapped values.
24 However, the double-ended threshold strategy is used to supplement the adaptive segmentation method. If the
25
weighted aggregated distance of the pixel is no more than a threshold 𝑇1 , then the pixel is classified as a signal pixel.
26
If the distance exceeds the threshold 𝑇2 , then the pixel is classified as a noise pixel.
27
28
29 3. Noise replacement
30
31 3.1. Adaptive selection strategy
32 In the image recovery phase, the recovery strategy can greatly influence the recovery effect. For the recovery strategy,
33 whether or not the recovered value is used should be discussed. If the restored part of the window is used as the original
34 value, then the image cannot be easily recovered from heavy noise. Meanwhile, if the recovery value in the restored
35 part of the window is used instead of the original value, then the details and edges of the image may be misinterpreted
36 as noisy due to the uneven noise distribution. To balance these problems, [37] proposed a recovery strategy named
37 course-to-fine detection. The course-to-fine detection strategy was used to detect the obvious noise and recover them in
38 the course state, reducing the noise density and leaving the less obvious noise, which would be detected and recovered
39 in fine state. However, the problem lies in assuming that coarse detection is accurate, as the quality of recovery cannot
40 be easily guaranteed when the noise level is high. Using a less obvious noise pixel as a replacement will negatively
41 affect the fine detection.
42 Consider the following processing window,
43
44 ⎛ 𝑉𝑠 ∕𝑉𝑑 𝑉𝑠 ∕𝑉𝑑 𝑉𝑠 ∕𝑉𝑑 ⎞
45 ⎜ 𝑉𝑠 ∕𝑉𝑑 𝑉𝑐 𝑉𝑠 ⎟,
46 ⎜ ⎟
47 ⎝ 𝑉𝑠 𝑉𝑠 𝑉𝑠 ⎠
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 6 of 14
49
 An effective weighted vector median filter

where 𝑉𝑠 is the original value, 𝑉𝑑 is the recovered value, and 𝑉𝑐 is the center point being processed.

In the processing window, the recovered value usually has better properties. The probability for the recovered pixel
to be a noisy pixel may be low. In a sense, it is equivalent to artificially reduce the noise density if the recovered value
is used in the local window, which is extremely important in high noise. Consequently, we propose the following
adaptive selection mechanism to adaptively replace some original values with recovered values.

⎧ −1∗ 𝑆
−𝑒−1
𝑤𝑖𝑡ℎ 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑚𝑎𝑥(0, 𝑒
𝐻
⎪ 𝑉 )
𝑉 =⎨ 𝑠 1−𝑒−1
−1∗ 𝑆
(10)
1 −1
⎪ 𝑉𝑑 𝑤𝑖𝑡ℎ 𝑝𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑚𝑖𝑛(1, 1 − 𝑒 𝐻 −𝑒
)
2 ⎩ 1−𝑒−1
3
4 where 𝑆 is the Euclidean distance between the original value and the recovered value at the same image position,
5 and 𝐻 is a positive threshold.
6
7
8 3.2. Replacement with value of channel combination optimization
9 Most methods choose an "optimal" pixel as the output of the filter, but this solution has two drawbacks:
10
11 1. A possible situation is that the pixel does not have a neighbor that is exactly equal to it on the original uncorrupted
12 image.
13 2. The values on some channels of the pixels may be good, while those on the other channels may be quite different
14 from the correct values.
15 In sum, the pros and cons of a pixel cannot be easily evaluated as a whole. To combine the excellent channels of pixels,
16
we propose a strategy for combinatorial optimization as follow:
17
18 ⎛ 𝑅0 𝑅1 𝑅2 𝑅3 𝑅4 𝑅5 𝑅6 𝑅7 𝑅8 ⎞
19 ⎜ 𝐺0 𝐺1 𝐺2 𝐺3 𝐺4 𝐺5 𝐺6 𝐺7 𝐺8 ⎟
20 ⎜ ⎟
21 ⎝ 𝐵0 𝐵1 𝐵2 𝐵3 𝐵4 𝐵5 𝐵6 𝐵7 𝐵8 ⎠
22 where r, g, and b are the serial numbers of the R, G, and B channels, respectively.
23
24 𝑟, 𝑔, 𝑏 ∈ {0, 1, 2, 3, 4, 5, 6, 7, 8}
25
26 𝑃𝑡 = (𝑅𝑟 , 𝐺𝑔 , 𝐵𝑏 )
27
28 𝑃𝑖 = (𝑅𝑖 , 𝐺𝑖 , 𝐵𝑖 )
29 ∑
𝑖=8
30 𝑎𝑟𝑔𝑚𝑖𝑛(𝑟, 𝑔, 𝑏) ‖𝑃𝑖 , 𝑃𝑡 ‖ ∗ 𝑊 (𝑖)
31 𝑖=0
32
The combined scheme (r, g, b) and color vector (𝑅𝑟 ,𝐺𝑔 ,𝐵𝑏 ) with the smallest weighted aggregated distance (value of
33
34 channel combination optimization) is considered as the ideal filter output.
35
36 4. Experiments
37
38 4.1. Noise model
39 To evaluate the restoration effects of different methods, an artificial impulse noise model capable of simulating natural
40 impulse noise is needed. A widely used noise model is
41 {
42 𝑁𝑖,𝑗 with probability r
𝑋𝑖,𝑗 = ,
43 𝑂𝑖,𝑗 with probability 1-r
44
45 where r represents the noise probability, 𝑂𝑖,𝑗 represents the uncontaminated pixel value at position (i,j) of the image,
46 and 𝑁𝑖,𝑗 represents the noise value contaminated at image (i,j). The simulation of the noise value can be divided into
47
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 7 of 14
49
 An effective weighted vector median filter

1
2
3
4
5
6
7
8
9
10
11
12
13 Figure 2: 15 natural images
14
15
16 two steps. First, the three color channels independently accept the corrosion of the noise value with probability r, and
17 then, for each contaminated pixel, a correlation factor, such as 𝜌 = 0.5, is used to simulate channel correlation. That is,
18 if a pixel is contaminated on at least one channel, each of the other uncontaminated channels receives impact corrosion
19 with a probability of 0.5. Moreover, if the correlation factor 𝜌 = 0 is taken, then the corrosion of the three color
20 channels is irrelevant, and each channel independently receives the corrosion of impact noise with the probability r.
21
22 We use four widely used objective criteria for evaluating restoration effects, namely, PSNR (peak signal to noise
23 ratio), SSIM (structural similarity)[38], NCD (color similarity), and FSIMc (feature similarity)[39]. The formula for
24 these three objective criteria (PSNR, SSIM, NCD) is as follows:
25
26 2552
27 𝑃 𝑆𝑁𝑅 = 10 ∗ lg ∑𝑖=𝐻 ∑𝑗=𝑊
1
+ 𝑖=1 𝑗=1
‖𝑦(𝑖, 𝑗) − 𝑜(𝑖, 𝑗)‖22
28 3∗𝐻∗𝑊
29
(2𝜇𝑜 𝜇𝑦 + 𝐶1 ) ∗ (2𝛿𝑜𝑦 + 𝐶2 )
30 𝑆𝑆𝐼𝑀 = .
31 (𝜇𝑜2 + 𝜇𝑦2 + 𝐶1 )(𝛿𝑜2 + 𝛿𝑦2 + 𝐶2 )
32 ∑𝑖=𝐻 ∑𝑗=𝑊
33 𝑖=1 𝑗=1
‖𝑦𝐿𝐴𝐵 (𝑖, 𝑗) − 𝑜𝐿𝐴𝐵 (𝑖, 𝑗)‖2
34 𝑁𝐶𝐷 = ∑𝑖=𝐻 ∑𝑗=𝑊 𝐿𝐴𝐵
35 𝑖=1 𝑗=1
‖𝑜 (𝑖, 𝑗)‖2
36
37 4.2. Performance comparison
38 The proposed algorithm is compared with other effective image restoration algorithms, namely, AWQD[34], FAPGF[28],
39 and L0TV[40], BVDF[25] on 15 natural images, shown in Fig. 2, taken from the CSIQ database (http://vision.okstate.edu/csiq)
40 and the classic Lena image.
41 A random impulse noise with a density range from 0.1 to 0.5 is added to each image, and the recovery effect is
42 compared by using four objective criteria (PSNR, SSIM, NCD, FSIMc). Since the Lena image is generally considered
43 as the most common test data in image denoising, the performance of these methods on the Lena image is examined
44 in detail. Tables 1 to 4 show the performance of the proposed method and the other methods (AWQD, FGAPF and
45 L0TV) on the Lena image under different objective criteria (PSNR, SSIM, NCD, FSIMc) and different noise levels
46 (0.1 to 0.5).
47
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 8 of 14
49
 An effective weighted vector median filter

Table 1
PSNR performance on the Lena image under a noise level ranging from 0.1 to 0.5
PSNR 0.1 0.2 0.3 0.4 0.5
Noisy 15.98 13.27 11.83 10.89 10.24
VMF 29.73 26.80 23.32 19.49 16.28
BVDF 28.09 24.54 20.50 16.02 12.89
AWQD 34.59 30.48 26.55 22.36 18.19
FAPGF 32.47 28.85 26.29 23.73 20.72
L0TV 31.37 28.61 26.43 24.68 22.97
proposed1 34.92 32.05 29.49 26.67 23.34
1 proposed2 34.37 31.82 29.39 26.85 24.05
2
3
4 Table 2
5 SSIM performance on the Lena image under a noise level ranging from 0.1 to 0.5
6 SSIM 0.1 0.2 0.3 0.4 0.5
7 Noisy 17.98 9.41 6.16 4.47 3.32
8 VMF 80.78 73.96 64.00 52.10 46.26
9 BVDF 77.04 69.12 56.33 37.59 26.45
10 AWQD 93.59 86.01 74.89 58.01 40.73
11 FAPGF 91.56 82.38 73.14 62.46 50.80
12 L0TV 90.56 83.07 75.52 67.82 60.39
13 proposed1 93.29 87.03 79.51 69.05 55.78
14 proposed2 93.08 86.92 79.74 70.63 59.13
15
16
Table 3
17 NCD performance on the Lena image under a noise level ranging from 0.1 to 0.5
18
19 NCD 0.1 0.2 0.3 0.4 0.5
20 Noisy 0.0985 0.1814 0.2506 0.3066 0.3506
21 VMF 0.0275 0.0389 0.0608 0.1088 0.1781
22 BVDF 0.0311 0.0443 0.0726 0.1447 0.2368
23 AWQD 0.0088 0.0185 0.0334 0.0631 0.1293
FAPGF 0.0103 0.0204 0.0322 0.0489 0.0748
24
L0TV 0.0117 0.0213 0.0314 0.0423 0.0553
25
proposed1 0.0085 0.0161 0.0251 0.0392 0.0665
26
proposed2 0.0089 0.0163 0.0249 0.0368 0.0567
27
28
29
Tables 1 to 4 show that in the PSNR indicator, the proposed method is clearly superior to other methods except for
30
L0TV at a noise level of 0.5. However, when the noise level ranges from 0.1 to 0.4, proposed method obtains a better
31
32 score than the L0TV. In terms of PSNR, the proposed method outperforms the other algorithms at each noise level.
33 When the noise density is as high as 0.5, although L0TV shows a higher score in terms of SSIM and FSIMc, L0TV
34 shows a poor performance in texture protection compared with the proposed method and produces fake color blocks.
35 Therefore, the proposed method outperforms L0TV in terms of PSNR as shown in Table1.
36 To reflect the robustness and universality of the proposed algorithm, the following tables present the comprehensive
37 performance of the aforementioned algorithms on the four objective criteria of 15 different images at a noise level
38 ranging from 0.3 to 0.4.
39
40
41
42 Tables 5 to 12 demonstrate that the proposed method outperforms two other methods under medium-density noise
43 levels. Fig. 3 visually compares the performance of the aforementioned methods under a noise level ranging from 0.1 to
44 0.5. The proposed1 method shows better detail protection under low noise, while the proposed2 method shows higher
45 robustness under high noise. AWQD demonstrates a good recovery performance at the 0.1 noise level and has well
46 preserved both the color and image structure; however, its recovery ability is greatly reduced when the noise gradually
47
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 9 of 14
49
 An effective weighted vector median filter

Table 4
FSIMc performance on the Lena image under a noise level ranging from 0.1 to 0.5
FSIMc 0.1 0.2 0.3 0.4 0.5
Noisy 0.78 0.69 0.63 0.58 0.55
VMF 0.96 0.91 0.83 0.71 0.61
BVDF 0.95 0.89 0.81 0.64 0.54
AWQD 0.99 0.98 0.93 0.83 0.67
FAPGF 0.98 0.97 0.94 0.89 0.81
L0TV 0.99 0.97 0.95 0.92 0.89
proposed1 0.99 0.99 0.97 0.93 0.86
1 proposed2 0.99 0.98 0.97 0.93 0.87
2
3
4 Table 5
5 PSNR performance on the 15 images under a 0.3 noise level
6 PSNR 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Noisy 10.67 11.62 9.58 11.95 10.81 11.56 12.00 11.87 12.17 10.76 10.71 9.61 10.71 11.52 12.39
7 VMF 18.03 19.54 18.50 19.11 18.00 19.32 22.50 18.20 22.58 16.61 18.80 17.85 18.35 21.44 19.40
8 BVDF 15.48 18.69 11.48 17.32 16.76 17.31 21.85 17.24 22.04 14.45 16.97 11.57 15.87 19.91 17.21
AWQD 21.36 22.88 23.59 22.46 22.31 23.10 27.25 21.53 26.33 19.88 22.63 22.53 23.60 26.52 21.72
9 FAPGF 21.60 21.69 24.37 21.93 23.28 23.09 26.58 21.07 25.88 20.34 22.64 23.44 24.79 26.58 21.17
10 L0TV 21.08 21.28 24.68 20.24 22.54 22.21 27.13 20.05 25.35 19.22 21.74 22.90 24.18 26.69 20.45
11 proposed1 23.47 23.65 27.20 23.56 25.37 25.27 29.65 22.79 28.64 21.62 24.66 25.71 27.21 29.65 22.87
proposed2 23.20 23.45 26.85 23.15 25.31 24.99 29.60 22.69 28.44 21.62 24.46 25.62 27.59 29.97 22.62
12
13
14 Table 6
15 SSIM performance on the 15 images under a 0.3 noise level
16 SSIM 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Noisy 9.57 9.79 4.66 13.43 11.66 11.25 5.37 17.47 6.07 14.99 8.59 7.74 8.03 5.31 14.23
17 VMF 43.00 47.84 39.15 56.51 48.48 45.83 64.48 35.07 79.19 29.48 42.97 33.37 47.59 68.21 37.99
18 BVDF 39.56 43.54 37.89 50.50 46.50 38.57 59.76 34.39 77.55 26.18 42.09 25.22 40.86 61.93 32.33
AWQD 64.23 68.80 67.82 76.88 70.69 70.93 79.20 66.89 87.69 56.78 66.28 62.43 71.79 80.34 63.78
19 FAPGF 60.86 62.25 74.16 69.32 75.01 69.87 73.37 62.39 77.19 57.02 65.63 69.89 76.59 76.92 57.52
20 L0TV 63.10 65.15 76.44 66.62 75.04 69.58 77.78 58.50 79.57 53.01 66.43 70.61 77.41 79.33 57.18
proposed1 71.75 71.01 80.70 77.62 82.15 79.24 81.05 73.68 84.05 65.94 74.94 77.75 83.09 83.31 68.89
21 proposed2 71.51 70.85 81.75 76.87 82.74 79.19 81.34 73.36 84.11 65.74 74.88 78.62 84.28 83.93 67.95
22
23
24 Table 7
25 NCD performance on the 15 images under a 0.3 noise level
26 NCD 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Noisy 0.2697 0.2484 0.3929 0.2182 0.2457 0.2735 0.1880 0.2085 0.1763 0.2989 0.2857 0.3723 0.3228 0.2631 0.2222
27 VMF 0.1164 0.0877 0.1294 0.0824 0.1066 0.1082 0.0482 0.1041 0.0352 0.1520 0.1091 0.1377 0.1216 0.0662 0.0949
28 BVDF 0.1269 0.0846 0.2204 0.0912 0.1046 0.1147 0.0443 0.1045 0.0320 0.1632 0.1002 0.2303 0.1352 0.0652 0.1128
AWQD 0.0594 0.0463 0.0459 0.0446 0.0504 0.0545 0.0219 0.0570 0.0169 0.0822 0.0513 0.0568 0.0513 0.0293 0.0584
29 FAPGF 0.0532 0.0504 0.0385 0.0477 0.0429 0.0525 0.0235 0.0592 0.0210 0.0756 0.0477 0.0478 0.0429 0.0292 0.0610
30 L0TV 0.0580 0.0569 0.0395 0.0631 0.0493 0.0629 0.0225 0.0768 0.0224 0.0991 0.0562 0.0558 0.0463 0.0299 0.0733
proposed1 0.0440 0.0421 0.0279 0.0399 0.0357 0.0422 0.0178 0.0496 0.0152 0.0680 0.0391 0.0377 0.0340 0.0217 0.0511
31 proposed2 0.0434 0.0420 0.0276 0.0407 0.0351 0.0421 0.0175 0.0500 0.0152 0.0669 0.0385 0.0370 0.0322 0.0210 0.0518
32
33
34 increases because the detection threshold set by this algorithm is difficult to control and the coarse-to-fine detection
35 and recovery may be difficult to achieve a good denoising effect with high noise density. In the coarse detection stage
36 of high-density noise, the image has a high noise density, and the recovery value of the coarse detection stage is very
37 likely to be a less obvious noise. To make the original noise pixels that remain in the fine stage become less obvious,
38 the original signal pixel correspondingly becomes more anomalous, thereby leading to errors in fine detection and
39 producing many structures where the original image does not exist.
40
41 L0TV is less sensitive to noise density because it is based on data fidelity regularization and is robust at high noise
42 levels. However, in the case of low to medium noise density, the image is excessively smoothed, and L0TV lacks a
43 sensitive detail protection mechanism, as seen in Fig. 4 (the left image represents the L0TV result, while the right
44 image represents the proposed method’s result). Meanwhile, when the local noise similarity is high, L0TV produces
45
many fake color spots, so the L0TV cannot effectively preserve the chromaticity as shown in Fig. 5.
46
47
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 10 of 14
49
 An effective weighted vector median filter

Table 8
FSIMc performance on the 15 images under a 0.3 noise level
FSIMc 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Noisy 0.6724 0.6590 0.5560 0.7422 0.7287 0.7306 0.6277 0.7312 0.5712 0.7145 0.6533 0.6258 0.6974 0.5837 0.7432
VMF 0.6960 0.7338 0.7192 0.6771 0.7490 0.7416 0.7849 0.7040 0.8476 0.6832 0.7254 0.6697 0.7363 0.8325 0.7016
BVDF 0.6781 0.7525 0.6294 0.7429 0.7593 0.7473 0.8036 0.7477 0.8663 0.6945 0.7261 0.5791 0.7159 0.8061 0.7160
AWQD 0.8601 0.8975 0.8776 0.8941 0.9076 0.9123 0.9239 0.8989 0.9422 0.8641 0.8821 0.8557 0.9133 0.9373 0.8782
FAPGF 0.8779 0.8890 0.9235 0.9056 0.9355 0.9257 0.9331 0.9130 0.9254 0.8891 0.9006 0.9052 0.9424 0.9410 0.8862
L0TV 0.8639 0.8732 0.9365 0.8445 0.9352 0.9191 0.9503 0.8880 0.9351 0.8504 0.8796 0.9016 0.9500 0.9539 0.8529
proposed1 0.9062 0.9124 0.9532 0.9167 0.9484 0.9477 0.9587 0.9334 0.9633 0.9071 0.9278 0.9282 0.9594 0.9700 0.9084
proposed2 0.9009 0.9095 0.9548 0.9094 0.9489 0.9459 0.9605 0.9316 0.9626 0.9066 0.9251 0.9321 0.9637 0.9729 0.9029

1 Table 9
2 PSNR performance on the 15 images under a 0.4 noise level
3 PSNR 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
4 Noisy 9.76 10.72 8.67 11.03 9.90 10.63 11.07 10.95 11.23 9.81 9.80 8.70 9.80 10.60 11.46
VMF 15.17 16.99 13.97 17.30 14.69 16.47 18.79 16.04 19.49 14.07 15.32 13.79 14.96 17.72 17.75
5 BVDF 12.67 16.71 9.63 14.63 13.87 14.12 18.13 15.06 19.46 12.13 14.18 9.68 12.68 16.05 15.04
6 AWQD 17.89 20.25 17.38 19.58 17.98 19.59 22.84 18.75 22.41 16.58 18.44 16.99 18.81 21.74 19.71
7 FAPGF 19.72 20.33 22.15 19.88 20.83 20.99 24.27 19.27 23.32 18.55 20.68 21.18 22.09 23.96 19.68
L0TV 20.18 20.33 23.17 19.16 20.81 20.77 25.24 18.95 23.68 18.38 20.84 21.48 22.34 24.84 19.66
8 proposed1 21.34 22.10 24.39 21.39 22.16 22.59 26.66 20.72 25.37 19.32 22.53 22.87 23.45 26.24 21.42
9 proposed2 21.56 22.18 24.87 21.06 22.41 22.79 26.85 20.95 25.76 20.03 22.98 23.28 24.54 27.07 21.40

10
11 Table 10
12 SSIM performance on the 15 images under a 0.4 noise level
13 SSIM 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
14 Noisy 6.88 7.08 3.32 9.71 8.36 7.95 3.82 12.39 4.45 10.63 6.17 5.52 5.68 3.83 9.93
VMF 32.73 37.93 27.26 46.36 32.65 32.17 56.85 22.72 70.38 19.18 33.22 17.77 32.75 57.24 29.23
15 BVDF 29.94 35.33 21.85 31.59 31.36 25.77 48.07 22.36 68.15 16.29 31.03 15.49 24.84 44.83 23.23
16 AWQD 41.72 52.28 33.63 60.51 46.67 50.21 65.39 45.08 75.16 34.91 43.87 28.78 48.64 64.72 46.24
FAPGF 48.44 50.95 63.81 57.19 63.76 58.01 64.01 48.46 67.72 43.53 53.37 57.90 65.91 67.53 44.98
17 L0TV 54.85 56.42 69.12 57.58 65.96 59.85 70.19 48.08 72.06 44.34 58.55 62.20 68.90 72.01 48.26
18 proposed1 57.63 60.12 64.85 65.23 68.17 66.16 71.99 58.37 76.14 48.73 61.76 60.82 69.34 74.35 56.75
proposed2 59.53 60.98 71.57 65.13 70.50 67.53 72.23 59.92 75.70 52.31 63.89 66.22 72.88 75.53 56.55
19
20
21 Table 11
22 NCD performance on the 15 images under a 0.4 noise level
23 NCD 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
24 Noisy 0.3283 0.3014 0.4769 0.2656 0.2996 0.3337 0.2300 0.2549 0.2160 0.3643 0.3472 0.4518 0.3923 0.3199 0.2711
VMF 0.1846 0.1315 0.2561 0.1105 0.1718 0.1624 0.0792 0.1394 0.0603 0.2143 0.1781 0.2561 0.2057 0.1147 0.1209
25 BVDF 0.1879 0.1124 0.3195 0.1428 0.1582 0.1715 0.0689 0.1386 0.0476 0.2268 0.1527 0.3181 0.2201 0.1162 0.1502
26 AWQD 0.1105 0.0747 0.1432 0.0725 0.1009 0.0963 0.0419 0.0908 0.0333 0.1428 0.1048 0.1495 0.1105 0.0600 0.0845
FAPGF 0.0749 0.0685 0.0596 0.0686 0.0644 0.0757 0.0353 0.0814 0.0333 0.1035 0.0687 0.0713 0.0664 0.0456 0.0811
27 L0TV 0.0709 0.0706 0.0534 0.0780 0.0664 0.0813 0.0306 0.0936 0.0308 0.1167 0.0682 0.0720 0.0644 0.0414 0.0869
28 proposed1 0.0663 0.0577 0.0478 0.0582 0.0605 0.0645 0.0277 0.0707 0.0250 0.1012 0.0579 0.0611 0.0602 0.0346 0.0679
proposed2 0.0603 0.0551 0.0410 0.0578 0.0563 0.0611 0.0268 0.0685 0.0244 0.0905 0.0529 0.0549 0.0523 0.0322 0.0671
29
30
31
32 5. Conclusion
33 In this paper, the degree of aggregation is proposed, based on which we proposed a weighting method to obtain an
34 optimal weight vector by minimizing the degree of aggregation. This weight vector, which has a larger component
35 on the signal pixel than on the noise pixel, is integrated with the aggregated distance to form a weighted aggregated
36 distance that can effectively distinguish signals from noises. Since the weighted aggregated distance has been calcu-
37 lated, an adaptive segmentation method, which is based on the sigmoid function, is integrated with the double-ended
38 threshold to detect the noises and signals. To reduce the local window noise density and avoid edge distortion, the
39 adaptive selection strategy and isotropic recovery framework are also proposed. Finally, the detected noise is replaced
40
by the weighted maximum likelihood estimation replacement value of the channel combination optimization. Experi-
41
ments show that compared with other methods, the proposed algorithm can effectively protect the image details while
42
43 ensuring an effective noise recovery.
44
45
46
47
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 11 of 14
49
 An effective weighted vector median filter

Table 12
FSIMc performance on the 15 images under a 0.4 noise level
FSIMc 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Noisy 0.6362 0.6189 0.5264 0.6908 0.6899 0.6899 0.5876 0.6936 0.5268 0.6742 0.6213 0.5894 0.6643 0.5442 0.7003
VMF 0.5784 0.6305 0.6188 0.5459 0.6039 0.6073 0.7045 0.5690 0.7444 0.5540 0.6166 0.5413 0.5981 0.7193 0.5955
BVDF 0.5913 0.6792 0.5297 0.6442 0.6632 0.6522 0.6956 0.6573 0.7823 0.6215 0.6393 0.5290 0.6165 0.6714 0.6345
AWQD 0.7350 0.8061 0.6752 0.7753 0.7867 0.8017 0.8248 0.7847 0.8560 0.7340 0.7454 0.6626 0.7871 0.8443 0.7811
FAPGF 0.8198 0.8413 0.8655 0.8188 0.8905 0.8819 0.8853 0.8664 0.8631 0.8379 0.8461 0.8484 0.8980 0.8877 0.8408
L0TV 0.8256 0.8342 0.9025 0.7898 0.8993 0.8777 0.9202 0.8417 0.8946 0.8072 0.8459 0.8562 0.9137 0.9233 0.8142
proposed1 0.8326 0.8626 0.8892 0.8406 0.8910 0.8870 0.9127 0.8601 0.9208 0.8193 0.8642 0.8454 0.9002 0.9320 0.8521
proposed2 0.8335 0.8624 0.9099 0.8295 0.8927 0.8963 0.9204 0.8721 0.9221 0.8434 0.8720 0.8694 0.9160 0.9412 0.8505

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27 Figure 3: Lena results under noise levels ranging from 0.1 to 0.5 (from left to right: noise image, VMF results,
28 BVDF results, AWQD results, FAPGF results, L0TV results, proposed1 results and proposed2 results).
29
30
31 6. Acknowledgment
32
33 This work is supported by the Department of Education of Hubei Province of China under Grant No. D20181504.
34
35
References
36
37 [1] Eduardo Abreu, Michael Lightstone, Sanjit K Mitra, and Kaoru Arakawa. A new efficient approach for the removal of impulse noise from
highly corrupted images. IEEE transactions on image processing, 5(6):1012–1025, 1996.
38
[2] Zhou Wang and David Zhang. Progressive switching median filter for the removal of impulse noise from highly corrupted images. IEEE
39 Transactions on Circuits and Systems II: Analog and Digital Signal Processing, 46(1):78–80, 1999.
40 [3] Javier Portilla, Vasily Strela, Martin J Wainwright, and Eero P Simoncelli. Image denoising using scale mixtures of gaussians in the wavelet
41 domain. IEEE Trans Image Processing, 12(11), 2003.
42 [4] Anna Brook. Three-dimensional wavelets-based denoising of hyperspectral imagery. Journal of Electronic Imaging, 24(1):013034, 2015.
43 [5] Hamid Reza Shahdoosti and Zahra Rahemi. Edge-preserving image denoising using a deep convolutional neural network. Signal Processing,
44 159:20–32, 2019.
[6] Yi Chang, Luxin Yan, Houzhang Fang, Sheng Zhong, and Wenshan Liao. Hsi-denet: Hyperspectral image restoration via convolutional neural
45
network. IEEE Transactions on Geoscience and Remote Sensing, 57(2):667–682, 2018.
46
47
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 12 of 14
49
 An effective weighted vector median filter

1
2
3
4
5
6
7
8
9 Figure 4: Performance on the Lady image under a 0.2 noise level
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29 Figure 5: Performance on the Lady image under a 0.5 noise level
30
31
32 [7] Wenhua Zhang, Lianghai Jin, Enmin Song, and Xiangyang Xu. Removal of impulse noise in color images based on convolutional neural
33 network. Applied Soft Computing, page 105558, 2019.
34 [8] Lianghai Jin, Wenhua Zhang, Guangzhi Ma, and Enmin Song. Learning deep cnns for impulse noise removal in images. Journal of Visual
35 Communication and Image Representation, 2019.
36 [9] Seongjai Kim. Pde-based image restoration: A hybrid model and color image denoising. IEEE Transactions on Image Processing, 15(5):1163–
37 1170, 2006.
[10] Lixia Chen, Pingfang Zhu, and Xuewen Wang. Low-rank constraint with sparse representation for image restoration under multiplicative
38
noise. Signal, Image and Video Processing, 13(1):179–187, 2019.
39 [11] Zhengwei Shen and Huitong Sun. Iterative adaptive nonconvex low-rank tensor approximation to image restoration based on admm. Journal
40 of Mathematical Imaging and Vision, 61(5):627–642, 2019.
41 [12] Qiang Wang, Zhaojun Wu, Jing Jin, Tiancheng Wang, and Yi Shen. Low rank constraint and spatial spectral total variation for hyperspectral
42 image mixed denoising. Signal Processing, 142:11–26, 2018.
43 [13] Kaixuan Wei and Ying Fu. Low-rank bayesian tensor factorization for hyperspectral image denoising. Neurocomputing, 331:412–423, 2019.
44 [14] Linwei Fan, Ran Meng, Qiang Guo, Miaowen Shi, and Caiming Zhang. Image denoising by low-rank approximation with estimation of noise
energy distribution in svd domain. IET Image Processing, 13(4):680–691, 2019.
45
[15] Jihong Pei, Hongguang Fan, Xuan Yang, and Xiaofeng Feng. A hybrid filter with neighborhood analysis for impulsive noise removal in color
46
47
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 13 of 14
49
 An effective weighted vector median filter

images. Signal Processing, 153:197–209, 2018.

[16] Lianghai Jin, Zhiliang Zhu, Xiangyang Xu, and Xiang Li. Two-stage quaternion switching vector filter for color impulse noise removal. Signal
Processing, 128:171–185, 2016.
[17] E Ataman, V Aatre, and K Wong. A fast method for real-time median filtering. IEEE Transactions on Acoustics, Speech, and Signal Processing,
28(4):415–421, 1980.
[18] David RK Brownrigg. The weighted median filter. Communications of the ACM, 27(8):807–818, 1984.
[19] Bogdan Smolka and Andrzej Chydzinski. Fast detection and impulsive noise removal in color images. Real-Time Imaging, 11(5-6):389–402,
2005.
[20] M Emre Celebi, Gerald Schaefer, and Huiyu Zhou. A new family of order-statistics based switching vector filters. In 2010 IEEE International
Conference on Image Processing, pages 97–100. IEEE, 2010.
[21] Jyohei Matsuoka, Takanori Koga, Noriaki Suetake, and Eiji Uchino. Switching non-local vector median filter. Optical Review, 23(2):195–207,
1 2016.
2 [22] Wei Wang and Peizhong Lu. An efficient switching median filter based on local outlier factor. IEEE Signal Processing Letters, 18(10):551–554,
2011.
3
[23] Jaakko Astola, Petri Haavisto, and Yrjo Neuvo. Vector median filters. Proceedings of the IEEE, 78(4):678–689, 1990.
4 [24] Samuel Morillas and Valentín Gregori. Robustifying vector median filter. Sensors, 11(8):8115–8126, 2011.
5 [25] Panos E Trahanias, Damianos Karakos, and Anastasios N Venetsanopoulos. Directional processing of color images: theory and experimental
6 results. IEEE Transactions on Image Processing, 5(6):868–880, 1996.
7 [26] Rastislav Lukac. Optimised directional distance filter. Machine Graphics & Vision International Journal, 11(2/3):311–326, 2002.
8 [27] Rastislav Lukac. Adaptive vector median filtering. Pattern Recognition Letters, 24(12):1889–1899, 2003.
9 [28] Lukasz Malinski and Bogdan Smolka. Fast averaging peer group filter for the impulsive noise removal in color images. Journal of Real-Time
Image Processing, 11(3):427–444, 2016.
10
[29] Justin Varghese, Mohamed Ghouse, Saudia Subash, Madappa Siddappa, Mohamed Samiulla Khan, and Omer Bin Hussain. Efficient adaptive
11 fuzzy-based switching weighted average filter for the restoration of impulse corrupted digital images. IET Image Processing, 8(4):199–206,
12 2014.
13 [30] Rastislav Lukac. Adaptive color image filtering based on center-weighted vector directional filters. Multidimensional Systems and Signal
14 Processing, 15(2):169–196, 2004.
15 [31] Rastislav Lukac, Bogdan Smolka, Konstantinos N Plataniotis, and Anastasios N Venetsanopoulos. Vector sigma filters for noise detection and
16 removal in color images. Journal of Visual Communication and Image Representation, 17(1):1–26, 2006.
17 [32] Gaihua Wang, Hu Zhu, and Yunyan Wang. Fuzzy decision filter for color images denoising. Optik, 126(20):2428–2432, 2015.
[33] Amarjit Roy, Joyeeta Singha, Salam Shuleenda Devi, and Rabul Hussain Laskar. Impulse noise removal using svm classification based fuzzy
18 filter from gray scale images. Signal Processing, 128:262–273, 2016.
19 [34] Lianghai Jin, Zhiliang Zhu, Enmin Song, and Xiangyang Xu. An effective vector filter for impulse noise reduction based on adaptive quaternion
20 color distance mechanism. Signal Processing, 155:334–345, 2019.
21 [35] Roman Garnett, Timothy Huegerich, Charles Chui, and Wenjie He. A universal noise removal algorithm with an impulse detector. IEEE
22 Transactions on image processing, 14(11):1747–1754, 2005.
23 [36] P Roji Chanu and Kh Manglem Singh. A two-stage switching vector median filter based on quaternion for removing impulse noise in color
24 images. Multimedia Tools and Applications, 78(11):15375–15401, 2019.
[37] Yali Amit, Donald Geman, and Xiaodong Fan. A coarse-to-fine strategy for multiclass shape detection. IEEE Transactions on Pattern Analysis
25
& Machine Intelligence, (12):1606–1621, 2004.
26 [38] Wang Zhou. Image quality assessment: from error measurement to structural similarity. IEEE transactions on image processing, 13:600–613,
27 2004.
28 [39] Lin Zhang, Lei Zhang, Xuanqin Mou, and David Zhang. Fsim: A feature similarity index for image quality assessment. IEEE transactions
29 on Image Processing, 20(8):2378–2386, 2011.
30 [40] Ganzhao Yuan and Bernard Ghanem. l0tv: A new method for image restoration in the presence of impulse noise. In Proceedings of the IEEE
31 Conference on Computer Vision and Pattern Recognition, pages 5369–5377, 2015.
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48 Xiangxi Meng et al.: Preprint submitted to Elsevier Page 14 of 14
49