Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.

1, February 2021

RESEARCH ON NOISE REDUCTION AND

ENHANCEMENT ALGORITHM OF GIRTH
WELD IMAGE
Xiang-Song Zhang1 Wei-Xin Gao1 and Shi-Ling Zhu2
1
College of Electronic Engineering, Xi'an Shiyou University, Xi’an, China
2
Communication and Information Engineering, Xi’an University of Post and
Telecommunications, Xi’an, China

ABSTRACT
In order to eliminate the salt pepper and Gaussian mixed noise in X-ray weld image, the extreme value
characteristics of salt and pepper noise are used to separate the mixed noise, and the non local mean
filtering algorithm is used to denoise it. Because the smoothness of the exponential weighted kernel
function is too large, it is easy to cause the image details fuzzy, so the cosine coefficient based on the
function is adopted. An improved non local mean image denoising algorithm is designed by using weighted
Gaussian kernel function. The experimental results show that the new algorithm reduces the noise and
retains the details of the original image, and the peak signal-to-noise ratio is increased by 1.5 dB. An
adaptive salt and pepper noise elimination algorithm is proposed, which can automatically adjust the
filtering window to identify the noise probability. Firstly, the median filter is applied to the image, and the
filtering results are compared with the pre filtering results to get the noise points. Then the weighted
average of the middle three groups of data under each filtering window is used to estimate the image noise
probability. Before filtering, the obvious noise points are removed by threshold method, and then the
central pixel is estimated by the reciprocal square of the distance from the center pixel of the window.
Finally, according to Takagi Sugeno (T-S) fuzzy rules, the output estimates of different models are fused by
using noise probability. Experimental results show that the algorithm has the ability of automatic noise
estimation and adaptive window adjustment. After filtering, the standard mean square deviation can be
reduced by more than 20%, and the speed can be increased more than twice. In the enhancement part, a
nonlinear image enhancement method is proposed, which can adjust the parameters adaptively and
enhance the weld area automatically instead of the background area. The enhancement effect achieves the
best personal visual effect. Compared with the traditional method, the enhancement effect is better and
more in line with the needs of industrial field.

KEYWORDS
X-ray image, Mixed noise, Noise separation, noise reduction, image enhancement

1. INTRODUCTION
Due to the influence of equipment and acquisition environment in X-ray welding image
acquisition, welding images are limited by low contrast and large noise, which will interfere with
the segmentation and recognition of welding image defects. With the development of image
processing technology, image denoising and contrast enhancement methods have emerged in
large numbers, and some achievements have been achieved. The weld images obtained by X-ray
generally are characteristic of fuzzy defect edges and various image noise [1]. To detect internal
defects more accurately, researchers should reduce noise and enhance the weld images. There are
many research findings in these fields. For instance, an improved Gabor filter image contrast
denoising algorithm was proposed, which can reduce the noise and retain the image details and

DOI: 10.5121/sipij.2021.12102 9
 Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.1, February 2021
defect areas [2]. A local image enhancement method was proposed by determining the local pixel
nonuniformity factor, and the combination of histogram equalization and contrast limitation can
adaptively improve the defect detection accuracy [3]. However, due to the large noise of weld
images, the contrast is low and the difference between the image background and the target gray
is not obvious. If the nonuniformity factor is not set accurately, the background and the target
will be enhanced, which interferes with the realization of the enhancement effect. A new
anisotropic diffusion imaging enhancement method can smooth, maintain edges and sharpen at
the same time [4]. We compared various denoising methods based on anisotropic diffusion, and
evaluated their performances with mean square error (MSE) and peak signal-to-noise ratio
(PSNR), which underlay further image segmentation and feature extraction experiments on noisy
weld X-ray images [5]. Firstly, wavelet denoising was used to remove the random high-frequency
noise generated in the detection of scattering radiation, then Shape from Shading(SFS) method
was used to distinguish the moire optical characteristics of low- and high-gradient defect areas,
and finally, Harris corner detection(HCD) method was adopted to detect and highlight the types
of residual corners related to different welding defects [6]. Because the effect of wavelet de-
noising on Gaussian noise is obvious and the noise in a weld image is Gaussian and salt-pepper
mixed noise, the effect of wavelet de-noising alone is not ideal. A gas tungsten arc weld (GTAW)
deep learning enhancement method based on multi-source remote sensing image was proposed
[7]. Two integrated methods including generative antagonism network and classic convolution
neural network were designed, which combined multiple neural networks to improve the
modeling performance for the unobserved data in different experimental environments. An
enhancement method based on the scale variable stochastic resonance model and a stochastic
resonance image enhancement method based on the genetic optimization algorithm were
proposed, which made full use of the parallel optimization of the genetic algorithm and the weak
signal enhancement of stochastic resonance and overcame the shortcomings of traditional
methods [8]. A series of improvement measures based on the similarity between local blocks of
the image were put forward [9]. Local binary pattern (LBP) texture feature and edge structure
tensor were used to improve the comparison of similar blocks and the effect of noise reduction. In
a two-step noise suppression based method for low-light-level image perception enhancement,
firstly the noise level function was used for contrast enhancement of noise perception, and then
the just-noticeable-distortion (JND) model was used to reduce the noise while keeping the details,
estimate noise visibility according to the intensity change of brightness and extract the details
through contrast masking and visual regularity [10]. With an improved nonlocal mean filtering
model, the rotation angle between image blocks was calculated by using the main direction of
corner points, and the similarity between blocks was computed more accurately [11]. However,
the noise reduction result was not ideal for weld images, because of the overall low contrast of
weld images and the indistinguishable similarity between image blocks [11]. When histogram
equalization algorithm was combined with high-frequency lifting filter, the image information
entropy and contrast after processing were higher and the processing time was shorter, which can
meet the real-time requirements of image processing [12]. In an adaptive noise reduction method
of image sensor, the image was decomposed by curve transform and the variational technique
was used to reduce noise and improve PSNR [13]. In a laser 3D image enhancement system
based on virtual reality, point cloud data were used to build a 3D image model that conformed to
normal distribution [14]. According to the distribution of its point image elements, the image was
enhanced effectively by the 3D image transformation enhancement system. A low-illuminance
nonlinear laser image edge adaptive enhancement device based on radial hill was adopted [15].
Moreover, an nsmission estimation method based on the improved dual region filter and the
guide filter was proposed to reduce the complexity of image defogging algorithms and solve the
white halo of image edges, and was verified experimentally to be effective and efficient [16].In
reference [17], an adaptive dynamic weighted median filtering algorithm is proposed, which
achieves good filtering effect.In reference [18], a block median filtering algorithm based on
10
 Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.1, February 2021
probability decision-making is proposed, which improves the traditional median filtering, and the
effect is remarkable.[19] An iterative nonlocal mean filtering algorithm is proposed to remove
impulse noise.[20]In this paper, a convolutional neural network (CNN) model algorithm based on
deep learning is proposed. The algorithm is superior to sparse representation and block based
method in accuracy and robustness.

Based on the above analyses and in view of the Gaussian and salt-and-pepper mixed noise in
weld images, firstly noise separation was carried out, then an improved nonlocal mean filter was
proposed to reduce the Gaussian noise and finally, an adaptive Takagi-Sugeno (T-S) fuzzy fusion
filter algorithm which can identify the probability of salt and pepper noise was proposed in
combination with the median filter. To improve the image contrast, we proposed an adaptive
parameter adjustment enhancement algorithm without human intervention. The enhancement
effect of the algorithm can reach the best vision of human eyes. Finally, experiments show that
the proposed algorithm has obvious advantages in noise reduction and enhancement.

2. IMAGE NOISE REDUCTION

The noise of an X-ray weld image is mainly the photoelectron noise caused by the photoelectron
conversion of the image sensor during shooting, and the electronic noise caused by the random
thermal movement of electrons. Based on mathematical analysis, the noise can be mainly divided
into Gaussian noise and salt-and- pepper noise. The noise in an actual image is a mixture of the
two. The model of the mixed noise can be expressed as follows:

f (i, j )  f y  z g  zs
1

z
Where f (i, j ) and f y represent the image with mixed noise and the original image respectively; g
and z s represent Gaussian noise and salt-and-pepper noise respectively.

2.1. Noise Model

As for Gaussian noise, its probability density is described by Gaussian normal distribution. A
larger mean value of Gaussian noise means a whiter picture; the larger variance indicates the
image is at lower resolution and is more fuzzy. Such noise in a digital image is mainly caused by
uneven illumination or high temperature.

1
p( z )  e ( z   ) 2 2
2

2 (2)

where p ( z ) represents the probability of the gray value z ,  and  are the mean and variance.
2

Salt and pepper noise refers to two kinds of noise: salt noise and pepper noise. The white noise is
salt noise and has a high gray value (255), and the black noise is pepper noise and has a low gray
value (0). a and b represent the limit gray values in the image.


 pa za
p( z )   pb z b

0 else


2.2. Mixed Noise Separation

The gray values of salt and pepper noise points are characteristic of extreme values, and are
generally the positive (negative) maxima or minima of gray value in the region where they are
11
 Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.1, February 2021
located. However, these positive (negative) maxima or minima may not be noise pixels, because
there are also positive (negative) maxima or minima in the undisturbed image gray smooth area
or the area with obvious edge. Therefore, it is necessary to further judge whether the extreme
point is a salt and pepper noise point. Let an m n window with x(i, j) be the center. If
x(i, j )  f (i, j )max , count the pixels x(i, j)  f (i, j )max in the window, where there are n1 in all; if
x(i, j )  f (i, j )min , count the pixels x(i, j )  f (i, j)min in the window, where there are n2 in all. When n1  t
or n2  t ( t is the threshold value of salt and pepper noise and generally varies between 12 and
16), the pixel point is judged as a salt and pepper noise point; when n1  t or n2  t , it is not judged
as a salt and pepper noise point.

2.3. Gauss Noise Filtering

2.3.1. Nonlocal Mean Filtering

The basic idea of original nonlocal mean (onlm) filtering is to use a large amount of redundant
information in the image to calculate the weighted average value of gray values of all pixels in
the same neighborhood according to the weighted coefficient. The core problem is to determine
the weighted kernel function by the weighted Euclidean distance between this function and the
image. The exponential kernel function (EKF) used for weighting makes the image details too
smooth and fuzzy. Therefore, on the basis of previous works, we discuss the establishment of the
weighted kernel function and propose a new cosine Gaussian kernel function (CGKF) combining
the Gaussian kernel function (GKF) and the cosine kernel function (CKF). The noise signal is
assumed to be image-independent additive Gaussian white noise, and can be modelled as follows:

Z (i)  X (i)  N (i)



where X (i) is the original image without noise pollution; N (i) is the Gaussian white noise with
mean value of 0 and variance of  ; Z (i) is the noise-contaminated image. Let z  {z(i) | i  I } be
2

a pair of discrete noisy images, where I is the image domain. For any pixel i in the image, the
onlm filter uses the weighted average of all pixel gray values in the image to estimate the gray
value of the point:

NL[ z ](i)   w(i, j ) z (i)

jI


where the weight z (i ) depends on the similarity between pixels i and j , and satisfies 0  z (i)  1
and
 j
z (i, j )  1
. The similarity between pixels i and j is determined by the similarity between
H
gray value matrices H i and j ,where H i represents a known-size square neighborhood centered
on pixel i .The similarity between the gray value matrices of each neighborhood is measured by
their Gaussian weighted Euclidean distance d (i, j ) :

2
d (i, j )  Hi  H j
2,


12
 Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.1, February 2021

where  is the standard deviation weighted by Gauss and   0 ; Function 2 is L norm, so a

. 2

higher neighborhood similarity indicates a smaller distance.Weight w(i, j ) is defined as:

1
w(i, j )  f k (d (i, j ))
C (i)
(7)

C (i )   f k (d (i, j ))
where j
is the normalized parameter. The core problem of nonlocal mean (nlm)
filtering given in Eq. (6) is the kernel function f k (.) in Eq. (8) and it plays an important role in the
denoising performance of the algorithm. The nlm method adopts EKF, which is defined as:

d (i, j )
f k (d (i, j ))  exp( )
h2 

where h is the attenuation factor, which controls the attenuation speed of the exponential
function and affects the filtering degree and denoising performance of the algorithm.

2.3.2. Improved Nlm Filtering

The core problem of nlm filtering is to determine the weighted kernel function. The core idea of
weighting is to give more weight to the neighborhood with higher similarity, and less weight to
the neighborhood with lower similarity. This is because the neighborhood with low similarity or
dissimilarity will increase the computation of the algorithm and affect its denoising effect. Under
certain conditions, the ideal kernel function should output larger weights when the distance
between pixels is small, and the output will decrease rapidly to 0 as the distance is enlarged. The
weighted kernel function plays an important role in denoising performance, including the cosine
type and the Gauss type. Specifically, CKF is defined as:


f k (d (i, j ))  cos( d (i, j ) 2h) d (i, j )  h
0 d (i, j )  h
(9)

and GKF is defined as:

d 2 (i, j )
f k (d (i, j ))  exp( ) d (i, j )  h
h2 (10)

When the noise intensity is weak, the denoising performance is better than that of nlm, but the
over weighting of CKF and the insufficient weighting of GKF lead to the decrease of denoising
performance when the signal intensity increases. Based on the comparative analysis of EKF, CKF
and GKF, a new cosine Gaussian kernel function (CGKF) was proposed:

 d 2 (i, j )
exp( ) cos( d (i, j ) 2h 2 ) d (i, j )  h2
f k (d (i, j ))   h12
 0 d (i, j )  h2
(11)

where h1 and h2 are filter parameters. CGKF adds a cosine coefficient on basis of GKF, so the
improved algorithm can perform denoising better at different noise levels. In comparison of Eq.
(8) to Eq. (11), the response curve is shown in Figure 1.
13
 Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.1, February 2021
1.0
EKF
0.9
CGKF
0.8 CKF
GKF
0.7

Weighted value
0.6

0.5

0.4

0.3

0.2

0.1

0.0
0 20 40 60 80 100 120
Euclidean distance

Figure 1. Response curve of weighted kernel function

Compared with the EKF of onlm, the response curve of GKF is flatter when the distance between
pixels is large, and drops rapidly as the distance increases, and the response curve of CKF is
relatively flat in the whole region (Fig. 1). The improved CGKF synthesizes their advantages and
has greater weighting when the distance is small, and the weight decreases at the enlarged
distance, avoiding over-weighting or under-weighting. Hence, we can make full use of the high-
similarity neighborhood to filter out noise, and effectively reduce the interference of low-
similarity and dissimilar neighborhood. Experimental results show that the denoising
performance is better than that of the onlm algorithm.

2.4. Salt and Pepper Noise Filtering

For salt and pepper noise, the median filter is obviously effective on it, but has a fixed window.
Firstly, a small window such as a 3  3 template can effectively retain image details, but its noise
reduction effect is not good when the noise density increases. Secondly, a large window such as a
5  5 template can effectively reduce salt and pepper noise, but loses the image details. Moreover,
median filtering cannot estimate the probability of salt and pepper noise. Therefore, we propose
an adaptive salt and pepper noise elimination algorithm that can automatically adjust the filtering
window to identify the noise probability. This algorithm is described in detail below.

2.4.1. Noise Probability Estimation

The salt and pepper noise probability is estimated as follows:

Step1 After the image is partitioned, select any point in each area, and take S1 , S2 ...Sn , n modules
centered on this point;

Step2 Use the adaptive median filtering window of 3  3 , 5  5 , 7  7 to filter n regions, and 3n
3  3 rectangular regions, and obtain S '11 , S '12 ...S '1n , S '21 , S '22 ...S '2 n S '31 , S '32 ...S '3n ;

Step3 In different filtering windows, subtract each point in the filtered area S '11 , S '12 ...S '1n ,
S '21 , S '22 ...S '2 n S '31 , S '32 ...S '3n
, from the gray level of each point in the previous area S1 , S2 ...Sn .
After the absolute value of the subtracted result is taken, count the points over a certain positive
number u11 , u12 ...u1n ; u21 , u22 ...u2n ; u31 , u32 ...u3n ;
14
 Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.1, February 2021
u (i  1, 2...n, j  1, 2...n)
Step4 Regroup the number of pixels ij in any small area, and take the
average value of the middle pixels in the small area, i.e. 、a2、a3 .The weighted average of
m a1

a1、a2、a3 is 9m higher than the noise probability.

2.4.2. T-S fuzzy Model Fusion Filtering

To facilitate the following discussion, let the image size be M  N , any pixel point in the image
be f (i, j ) , and the filtered pixel be f '(i, j ) , where i  1, 2,3...M , j  1,2,3...N . The defect of the
median filter is that it can rank all pixels without the ability of noise point recognition and noise
rate estimation. Herein, the variable scale sliding window is used to flexibly adopt windows of
different sizes for different noise probabilities. It can reduce the computation amount and the blur
caused by relatively large windows when the noise probability is low, and can improve the
filtering accuracy when the noise probability is high. Figure 2 shows four window modes, which
will be combined into four filtering modes, where (a) to (d) are Model1 , Model2 , Model3 and Model4 ,
respectively and "Star" is the pixel participating in filtering.

(a) (b) (c) (d)

Figure 2. maximum window model

FM1
(1) (abbreviation of Filter Model ), window, Model is sliding filter window, and window
1 1 Model2

as the backup is the largest window;

(2) FM 2 : FM 2 is the main window and FM 3 is the backup;
(3) FM 3 : FM 3 is the main window and FM 4 is the backup;
(4) FM 4 ： FM 4 is the main window.

After determination of noise probability, the fusion filtering results of different noise probability
values and fuzzy correspondence under T-S model of filtering model are given in Figure. 3.
Fusion degree

FM1 FM2 FM3 FM4

1

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Noise probability

Figure 3. Corresponding fuzzy relation between noise probability and filtering mode

2.4.3. Filtering Algorithm

Based on the above analysis, the steps of the adaptive salt and pepper noise elimination algorithm
proposed here are as follows:

Step1 Noise probability calculation;

15
 Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.1, February 2021
Step2 Model fusion: according to the noise probability, the filtering algorithm under the T-S
model is used. If the probability of salt and pepper noise is estimated to be 0.4, FM 3 is used; if the
probability is 0.5, FM 3 and FM 4 are filtered at the same time. Finally, the results of FM 3 and FM 4
are fused and outputted according to the T-S fuzzy relationship.

Filtering under the known model: firstly, the pixel v(i, j ) is detected. If the pixel is not noise-
polluted, the result according to Eq. (12) is outputted and then the next pixel v(i, j  1) is filtered;
otherwise, the pixels of the filtering window in the filtering mode are detected one by one, and
the number of non-noise-polluted pixels nn is counted. If nn  0 , the output of noise pixel is
estimated as per Eq. (13); if nn  0 , the detection range of the pixel is extended to the range of
the standby window and Eq. (14) is adopted:

v '(i, j )  v(i, j ) min  v(i, j )  max (12)

R R R R
v '(i, j )   nm v(m, n)  mn
v(i, j )  min 或v(i, j )  max 且nn  0
m n m n (13)

v '(i, j )  med (Wij ) v(i, j )  min 或v(i, j )  max 且nn  0

(14)

W
where v(i, j ) is the center pixel of the current window ij ( R  R , R is odd); v '(i, j ) is the
estimated value after filtering of the current operating pixel; min and max are the noise

thresholds; Med is the median value of the pixel, mn is the weighting of the undisturbed pixel,
i.e. 1 ( x  y ) ; x and y are the horizontal and vertical pixel spacing respectively from any pixel
4 4

point to the center pixel point.

3. IMAGE ENHANCEMENT
3.1. Sin Function Transformation

The gray-scale transformation curve of a sin function is characterized by gentle transformation of

upper and lower wave heads, and large changes in the middle (Fig. 4). Since the gray level of the
weld image with poor contrast concentrates in a certain range and the gray level of the range is
stretched by the sin function, the gray level of the background area larger than the average gray
value increases and that of the weld area smaller than the average gray value decreases. This
feature is very effective in distinguishing the gray values of the background area and the weld
area. The transformation method of the sin function is (15):

    f ( x, y)   (a  b)  
g ( x, y)  127 1  sin   
  b-a 2  (b  a)  
(15)

where f ( x, y ) and g ( x, y ) are the gray values before and after transformation respectively; a and
b are the lowest and highest gray values before transformation respectively. The gray ranges of
the sin function before and after transformation are the same.

3.2. Logarithmic Transformation

The expression of logarithmic transformation is:

16
 Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.1, February 2021
ln  f ( x, y )  1
g ( x, y )  a 
b  ln c (16)

Where a , b and c are parameters for adjusting the position and shape of the curve. Logarithmic
transformation can stretch the low-gray area of the image and compress the high-gray area. The
transformation curve is shown in Fig. 5.

g(x, y) g(x, y)
255
255
output grayscale

output grayscale
f (x, y) f (x, y)
0 0
a enter grayscale b enter grayscale

Figure 4. Sin transformation curve Figure 5. Logarithmic transformation curve

3.3. XE Function Transformation

At present, the image of X-ray pipeline girth weld is fuzzy and low-quality. Neither manual nor
automatic detection is conducive to identify the defects in the weld. The existing image
enhancement algorithms mostly use histogram dynamic range stretching, which can stretch both
the target area and the background area. However, there is still a lack of X-ray image
enhancement technology for girth weld region of interest without parameter selection. Herein, a
method of industrial X-ray girth weld image enhancement was used. The gray value of a weld
area was determined by retrieving the gray value corresponding to the second peak of gray
histogram. The gray value of the weld area was enhanced by introducing a specific continuous
function, and that of background was suppressed. It can automatically enhance the weld area
instead of the background area, without manual intervention in parameter selection, and has
strong robustness. To effectively enhance the X-ray image of girth weld to the best value of
personal visual perception, we conducted a number of experiments. The success rate of defect
detection is improved from about 95% without enhancement to about 98% after the enhancement
of xe function. The gray-scale enhancement is expressed as follows, 
f gray  i, j  
,
g  gray  i, j  
is the gray value before and after enhancement.The xe enhancement function is
shown in Figure 6

g ( gray (i, j ))

f ( gray (i, j ))

Figure 6. Schematic diagram of xe function enhancement

17
 Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.1, February 2021
Because the collected girth weld images are placed horizontally, the xe function enhancement
method is designed, as shown in the figure above. It can be seen from the figure that the closer
the xe function image is to 0, the more obvious the change of gray level is.

4. EXPERIMENTAL ANALYSIS
4.1. Peak Signal-to-Noise Ratio (PSNR) Comparison After Filtering

Totally 200 weld images were randomly selected from the image library and were filtered by the
nlm filter with different kernel functions. After that, 10 filtered images were randomly selected
and numbered as 1 to 10. The PSNR of these 10 images were compared with that of mean filter
(meaf), median filter (medf) and algorithm in Ref. 10. To write conveniently, we used different
kernel functions to represent these images. For different filters, the PSNRs of the filtered images
are shown in Table 1.

Table 1. PSNR of filtered image

Image CGKF MEAF GKF EKF MEDF CKF Document

No. psnr psnr psnr psnr psnr psnr 10 psnr
1 53.4475 52.8513 50.4475 48.8328 47.8411 51.6415 48.4135
2 54.7413 53.2919 50.7413 48.0659 47.5142 51.5501 48.4492
3 53.5043 52.9523 50.5043 49.3936 48.3503 51.4403 49.3903
4 53.9554 52.2902 51.9554 49.7863 48.7751 51.5855 48.3516
5 53.6042 52.3925 50.6042 48.8776 48.0851 50.6551 48.5509
6 54.2037 53.9670 49.2037 47.8695 47.1965 51.4965 48.5565
7 53.4404 52.3465 50.4014 48.5321 47.6547 51.5547 49.5145
8 54.6302 53.4509 49.6302 47.2283 46.7074 51.6475 48.6075
9 53.6641 52.4761 50.6641 48.5963 47.7474 52.5364 49.5614
10 52.2199 53.7228 51.2199 50.1466 48.8726 51.4827 48.5425

55
CGKF
54 EAF
MEAF
53 MEDF
Document 10
52 GKF
CKF
PSNR

51

50

49

48

47

46
1 2 3 4 5 6 7 8 9 10
Image number

Figure 7. The psnr curves

18
 Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.1, February 2021
4.2. Image Quality Comparison After Filtering

4.2.1. Filtered Images and Corresponding 3D Histogram

The nlm filter with different kernel functions was used to denoise the weld images, and compare
the filtering effect of various kernel functions (EKF, GKF, CKF and GCKF) as well as the
algorithm in the literature.

Figure 8. The filtered images

100
100 100

80
80 80
Grayvalue

60
Grayvalue

Grayvalue
60 60

40 40 40

20 20 20

0 0 0
150 150 150
50 100 50 50
100 40 100
40 40
30 30 30
50 20 50 20 50 20
10 10 10
Pixelpoint 0 0 Pixelpoint 0 0 0 0
Pixelpoint Pixelpoint Pixelpoint Pixelpoint

(a) (b) (c)

100 100 100

80 80 80
Grayvalue

60
Grayvalue

Grayvalue

60 60

40 40 40

20 20 20

0 0 0
150 150 150
50 100 50 50
100 100
40 40 40
30 30 30
50 20 50 20 50 20
10 10 10
0 0 0 0 Pixelpoint 0 0
Pixelpoint Pixelpoint Pixelpoint Pixelpoint Pixelpoint

(d) (e) (f)

Figure 9. Three-dimensional histograms of filtered images

Figures 9(a) to (f) show the original image, filtered image (through cgkf NLM filter, mean filter,
GKf NLM filter, EKF NLM filter, median filter, CKF NLM filter) filtered image and
corresponding three-dimensional histogram. The simulation results show that the histogram of
3D image is smoother and the gray distribution is more uniform after cgkf NLM filtering, which
indicates that cgkf NLM filter is most suitable for weld image filtering.

4.3. Enhanced Image and Corresponding Histogram

XE function, logarithmic function, histogram equalization and sin function were used to enhance
the same randomly-selected weld images, and compare their enhancement effects with the
algorithms in the literature.

19
 Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.1, February 2021

(a) (b) (c) (d) (e) (f)

Figure 10. Images enhanced by various methods

140 140 160

120 120 140

120
100 100
100
80 80
80
60 60
60
40 40
40

20 20 20

0 0 0
Graylevel
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
Graylevel Graylevel Graylevel

(a) (b) (c)

140

120
150 100

100
80
80
100
60
60
40
40 50

20 20

0 0 0

0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
Graylevel Graylevel Graylevel

(d) (e) (f)

Figure 11. Histograms of enhanced images

Figure 10 (a)-(f) shows the images enhanced by xe function, logarithmic function, histogram
equalization, sin function, and algorithms in Refs. 3 and 6. Figure 10 shows the gray histograms
corresponding to Figure 11.

5. CONCLUSIONS
Through theoretical analysis and calculation, the mixed noise was separated in the image filtering
part. For Gaussian noise, a new Gaussian cosine kernel nonlocal mean filtering algorithm was
proposed by comparing the filtering effects of different kernel functions. Practical verification
showed the new algorithm had obvious advantages over the traditional algorithm. Moreover,
noise estimation was proposed for salt and pepper noise. This algorithm solved the problem of
fixed median filter window and the inability to estimate noise intensity. Experiments showed the
effect of noise reduction was obvious. In the image enhancement part, the ex function
enhancement method proposed here can automatically and adaptively enhance the weld area
instead of the background area. The best effect of human vision can be achieved by adjusting
parameters without human intervention. Finally, the feasibility and advantages of the proposed
method were verified by comparing with different algorithms.

ACKNOWLEDGEMENTS
First of all, I would like to thank my tutor, Mr. Gao Weixin, for giving me a lot of guidance and
suggestions in the process of writing the thesis; thirdly, I would like to thank Zhu Shiling for
giving me a lot of opinions on the writing and grammar of English papers; finally, I would like to
thank all the reviewers for their sincere opinions on the thesis, which benefited me a lot.
20
 Signal & Image Processing: An International Journal (SIPIJ) Vol.12, No.1, February 2021
REFERENCES

[1] W. H.Wang, Image denoising algorithm based on noise detection and dynamic window,Journal of
Graphics,40(1)(2019)112-115.
[2] Yahaghi E, Movafeghi A. Contrast enhancement of industrial radiography images by gabor filtering
with automatic noise thresholding, Russian Journal of Nondestructive Testing, 55(1)(2019)73-79.
[3] Lin Z, Yingjie Z, Bochao D, Welding defect detection based on local image enhancement, Iet Image
Processing,13(13)( 2019)2647-2658.
[4] Gharsallah M B, Mhammed I B, Braiek E B, Improved geometric anisotropic diffusion filter for
radiography Image enhancement, Intelligent Automation and Soft Computing,24(2)(2018)231-240.
[5] Muthukumaran M, Prabaharan L, Sivapathi A,A comparative analysis of an anisitropic diffussion
image denoising methods on weld x-radiography images,Far East Journal of Electronics and
Communications,17(2)(2017)267-281.
[6] Yahaghi E, Hosseiniashrafi M. Enhanced defect detection in radiography images of welded
objects,Nondestructive Testing and Evaluation,34(1)(2019)13-22.
[7] Feng.Y, Chen.Z, Wang.D, Chen.J,et al. DeepWelding: a deep learning enhanced approach to gtaw
using multi-source sensing images,IEEE Transactions on Industrial Informatics,1(16)(2020)465-474.
[8] W.L.Mu,G.J.Liu,Research on stochastic resonance enhancement of x-ray images based on a genetic
algorithm,Non-Destructive Testing and Condition Monitoring,58(5)(2016)246-250.
[9] H.Liu.Research of non-local mean Image denoising algorithm based on structural
similarity,master,Wuhan University of Technology,Wuhan,China,2017.
[10] Su H, Jung C. Enhancement of Low light images based on perceptual two-step noise
suppression,IEEE Access,6(2018)7005–7018.
[11] N.H.Chang, The research of non-local means filtering to remove the image gaussian noise ,master,Xi
an,University of Electronic Science and Technology of China,Chengdu,China,2015.
[12] Y.N.Liu,L.R.Wang,Research on fuzzy image enhancement based on laser sensor,Laser
journal,41(3)(2020)123-125.
[13] H.Q.Zeng,Study on adaptive noise reduction of image sensor,Laser journal,12(67)(2019)67-71.
[14] C.Y.Zhu,X.Zhang,Research on laser 3D image enhancement system based on virtual reality,Laser
journal,38(3)(2017)114-117.
[15] J.J.Liu,Design of edge adaptive enhancement device for low illumination nonlinear laser,Laser
journal,38(3)(2017)122-125.
[16] D.D Chen,L. Chen.A new single image dehazing method based on improved double-area filter and
guided filter,Journal of Computers,29(4)(2017),230-240.
[17] Khan S, Lee D H. An Adaptive Dynamically Weighted Median Filter for Impulse Noise Removal[J].
Eurasip Journal on Advances in Signal Processing, 2017, 2017(1):67.
[18] Balasubramanian G, Chilambuchelvan A,Vijayan S, etal. Probabilistic Decision Based Filter to
Remove Impulse Noise Using Patch else Trimmed Median[J]. AEU-International Journal of
Electronics and Communications, 2016, 70(4):471-481.
[19] Wang X T, Shen S S, Shi G M, et al. Iterative Non-Local Means Filter for Salt and Pepper Noise
Removal[J]. Journal of Visual Communi- cation and Image Representation, 2016, 38(2016):440-450.
[20] Mohammad Tariqul Islam, SM Mahbubur Rahman, M Omair Ahmad, and MNS Swamy, “Mixed
gaussian-impulse noise reduction from images using convolutional neural network,” Signal
Processing: Image Communication, vol. 68, pp. 26–41, 2018.

AUTHORS

Zhang Xiangsong (1986 -), master degree, research direction: image

processing.

Gao Weixin (1973 -), Professor, doctor, postgraduate supervisor, research direction: image and
signal processing.

Zhu Shiling (1990 -), master degree, research direction: Internet of things technology and
application.
21