Academic Journal of Science and Technology 
ISSN: 2771-3032 | Vol. 12, No. 2, 2024 

 

125 

Research	on	Weld	Seam	Marking	Based	on	Machine	
Vision	Technology	
Shunyang Hu, Chenlei Zhao, Qian Fu 

Geely University of China, Chengdu 641402, China 

 

Abstract: The rapid development of machine vision technology has made it widely used in various industrial fields. The 
detection of welding quality can also be achieved through this technology. The core of welding quality inspection is to realize 
the marking of weld features. To solve the problem of weld feature marking, the GUI weld marking interface is built based on 
MATLAB. After the process of image acquisition, image preprocessing, edge detection and feature extraction, the weld marking 
experiment of is finally completed. Based on data analysis of the experiment, it can be concluded that Canny edge detection 
operator has the best marking effect and the most stable detection for the research object used. 

Keywords: Machine Vision Technology; Image Processing; Weld Marking. 
 

1. Introduction 
Machine vision technology is used to measure and identify 

the target shape and size, color and other information of the 
workpiece, and with this advantage, it has slowly replaced 
manual work in the industrial production line and greatly 
enhanced the workpiece quality and its sensitivity and 
accuracy. Relying on machine vision weld technology, China 
has achievindustrial automation with much lower costs in 
labor, time costs money costs, etc., which made the technolog 
a rigid demand for manufacturing. The automation and 
intelligent development of the industry will only become 
more and more rapid, providing a solid foundation for 
welding technology. The weld images obtained by machine 
vision technology can be directly provided after computer 
processing along with the feature data of various images for 
optimizing the welding method and parameters, so that the 
welding quality is guaranteed. 

2. Detection and Principle of Machine 
Vision Weld Marks 

 
Figure 1. Image processing 

 
Machine vision is to replace the human eyes with a 

machine to measure and judge the work. The system converts 
the captured object into an image signal and transmits it to a 
special image processing system. The image can be analyzed 
according to the relationship between pixel distribution and 
brightness, color and other information. Combined with 
related algorithms, the system can obtain the required 
information for recognition, and then realize the recognition 
and judgment of the target weld. Then the data is processed 
by the computer and output on the display. The complete 
image processing includes image acquisition, image 
preprocessing, detection of target edge and feature 
information extraction and so on. It can simulate human 
visual process by computer and control it automatically. 

2.1. Image Preprocessing 
2.1.1. Grayscale Conversion 

The gray-scale transformation is mainly based on 

independent pixels, which enhances the quality of the picture 
by changing the gray-scale range of the original image data, 
making the image clearer and easier to recognize. In practice, 
it is also necessary to pay attention to the problems that may 
occur at various stages for the machine vision system to better 
play a role. Usually, the dynamic range of pixels is enlarged 
by gray-scale transformation, and the image contrast is 
expanded. For different types of gray transformation, 
different effects can be obtained. If it is assumed that the pixel 
value k in the source image is located at a certain image 
position and transformed by the algorithm formula F (k), the 
pixel value at the position becomes H. Since there is a 
mapping relationship between gray images of function F (k), 
the formula after gray transformation is the expansion of its 
mapping formula. 

2.1.2. Binary Image 
Usually a normal view is that black and white image refers 

to gray scale image, in practice, in addition to gray scale 
image, there is a class of binary image. Strictly speaking, 
binary image is a real black and white image, because the gray 
scale of the binary image can only reach 0 and 255, the color 
of 0 represents black, 255 white. In this paper, a new 
histogram equalization method based on gray space is 
introduced, which can adjust the brightness and color of the 
image according to the actual needs. When the gray value of 
an image pixel exceeds the set threshold, the pixel value in 
the figure is 255, otherwise, it is 0. 

2.2. Image Denoising 
The existing digital images are all noisy images, which 

means that the image data contains some useless clutter 
information and interference information, which is usually in 
the form of pixels. Pixel blocks and other visual effects are 
strong, so it is very necessary to carry out image processing 
after image noise reduction, which not only improves the 
quality of the image, but also avoids its useless information  
interfering with people's detection of the data. Image filtering 
can suppress the noise in the target image while ensuring that 
the image features are preserved. Usually, there are two kinds 
of filtering, linear and nonlinear, while the mean filtering is 
linear, averaging all pixel values in a certain area and giving 
the average value of pixels at the center of the image. Non-
linear is more difficult than linear, and the maximum or 



 

126 

minimum value of pixels in a certain range, which is nonlinear, 
needs to be calculated.  

2.2.1. Linear Filtering 
Commonly used kinds of linear filtering include mean 

value filtering, Gaussian filtering, etc. Mean filtering 
processes the average value of the pixel in the target area, and 
gives the average value of the pixel in the center position. 
Gaussian filtering is a method which averages all pixels in the 
image and then calculating the coordinates of the center 
points. Gaussian filtering refers to the filtering of the 
unknown pixel in the center after the Gaussian weighting of 
the pixel values in the target range. This method is suitable 
for images with little change in gray level or strong local 
contrast. The mapping function of linear filtering is linear, so 
the convolution form is expressed as Y=h*X, and the 
convolution kernel is h [9], so formulas 1 are as follows: 

, ∑ 	 ∑ 	 , ,   (1) 

2.2.2. Nonlinear Filtering 
The nonlinear filtering usually used is the median filtering, 

bilateral filtering, etc., each nonlinear filtering has no specific 
manifestation, so the general nonlinear filtering is to filter the 
pixel value in a certain region after the maximum or minimum 
value, and the pixel value in the center position will be 
replaced. The Gaussian filter method is suitable for images 
with little change in gray level or strong local contrast. 
Median filtering is to arrange the pixel values within the 
region in an orderly manner and then filter, replacing the 
center position of the pixel values in the region with the 
middle value. 

Formula 2 of nonlinear filtering is as follows: (X represents 
the input image, i represents the abscissa, j represents the 
ordinate) 

, — , — ,… , , , … , ,   (2) 

2.3. Object Detection Algorithm 
The feature extraction of weld marks generally includes 

image features of texture, color, shape, spatial relationship, 
etc. After image reading, image features are extracted, and 
image feature data is collected and uploaded to the classifier. 
In the process of image feature data classifying, the basic 
principle is to distinguish the distinctive feature data of the 
image from other images. Due to the diversity of the image 
itself, the extracted feature vectors are also different, which 
makes it impossible for the traditional method to achieve the 
expected effect. 

2.3.1. Textural Features 
Texture is an attribute that shows the relationship between 

the structure arrangement of the surface of things and the 
environment, and also reflects the homogenous visual 
features in the image. Usually, the image features that are 
repeatedly displayed by the structure arrangement and local 
pattern are texture features. Texture features have the 
advantages of resisting noise to a certain extent, and can also 
be used for statistical calculation of areas containing several 
pixels. In the face of rotating photos, data can also be 
extracted correctly. However, there will be big errors in 
obtained features data under the influence of blurred images 
or light irradiation. Texture feature extraction usually adopts 
statistical method, geometric method, model method and 
structure analysis method. 

2.3.2. Colour Signature 
The color feature in an image is one of the important 

features, which generally refers to the properties of the 

corresponding surface in the image area. Texture feature 
extraction methods are mainly divided into two algorithms 
based on mathematical morphology and neural network. The 
advantage of color features is that color is not sensitive 
enough to the shape, rotation and other changes of the image, 
and the connection between color and the object contained in 
the image is closer. Color feature extraction is usually carried 
out by the method of color histogram, color aggregation 
amount, color set and so on. 

2.3.3. Shape Feature 
The shape of the general image is composed of points, lines 

and surfaces, and the shape features have limitations, and 
some are even one-sided, usually used on the edge features of 
the image, but when the picture is rotated, blurred, under the 
action of deformation and other factors, the data extracted 
from the shape features will lose reliability and cannot be used 
in the test. In order to overcome the above shortcomings, an 
improved algorithm is proposed to extract the feature 
information with better stability and accuracy, so as to obtain 
more complete and accurate shape feature data. There are two 
types of shape features in general. One is the feature of the 
whole area of the image, and the other is the contour feature 
of the edge area of the local image. 

2.3.4. Spatial Relationship Feature 
The so-called spatial relationship refers to the three types 

of adjacency and cross, overlap or inclusion relations between 
the multiple objects segmented in the image. In order to 
overcome the above shortcomings, an improved algorithm is 
proposed to extract the feature information with better 
stability and accuracy, so as to obtain more complete and 
accurate shape feature data. Generally, spatial position 
information can be divided into relative spatial information 
and absolute spatial position information. 

3. Weld Marking and Inspection 
Feature extraction of welds is to use machine vision 

technology to mark welds in the image region. Since welds 
are linear features, experiments and analysis can be carried 
out by using edge extraction algorithms. At present, the most 
widely used edge detection methods include Sobel, Prewitt, 
Roberts, Canny and other operators. In order to correctly 
extract useful feature data, the above four operators will be 
compared and analyzed combined with weld features. 

3.1. Edge Detection Operator 
Edge detection is actually a filtering algorithm, and the 

difference between it and other filtering algorithms lies in the 
selection of filters. The characteristic of ordinary filters is that 
Gaussian filters use the normalized kernel formed by 
discretized Gaussian functions, while the filter algorithm of 
edge detection generates grayscale matrix by calculating the 
target image through Gaussian kernel. Edge detection 
algorithm is a is edge detection technology widely used in 
computer vision which can display the small structure and 
details on the surface clearly. 

(1) Robert operator 
Among these classical operators, edge detection operator is 

the simplest. Robert operator uses some regions of the image 
to search edges by difference, and locates edge positions with 
greater advantages than other operators. However, for photos 
with large noise, the effect is poor, and for target images, the 
edge lines are inaccurate. This algorithm is often used as an 
image with steep low noise. The basic image is f (x,y), and 



 

127 

the edge detection image of Robert operator is g (x,y). The 
algorithm 3 can be expressed as: 

 g x,y ={[f(x,y)-f(x+1,y+1)]²+[f(x+1,y)-f(x,y+1)]²}0.5   (3) 
(2) Prewitt operator 
Compared with Robert operator, Prewitt operator has 

stronger resistance to noise and is less affected by waves. In 
essence, it is a differential operator, that is, the gray-scale of 
pixels in the image in the region is used for differentiation. 
Therefore, Prewitt operator is more suitable for images with 
high noise waves. Suppose the basic image representation 
function is f (x,y), and the pixel at a certain point in the image 
is represented by (i,j). The horizontal difference ( ) and 
vertical difference (j): 

 
G(i)=|[f(i-1,j-1)+f(i-1,j)+f(i-1, j+1)]-[f(i+1,j-1)+f(i+1, 

j)+f(i+1, j+1)]|                (4) 
G(j)=|[f(i-1,j+1)+f(i,j+1)+f(i+1, j+1)]-[f(i-1,j-1)+f(i,j-

1)+f(i+1, j-1)]|              (5) 

By formula 6: P(i,j)=max[G(i),G(j)]           (6) 

Or formula 7: P(i,j)=G(i)+G(j)             (7) 

The gradient amplitude P(i,j) of pixel point (i,j) is 
calculated. By setting threshold T0, when P(i,j) is less than T0, 
it can be judged as contour edge, and the contour feature is 
marked. 

(3) Sobel operator 
Sobel operator belongs to a class of discrete differentiation, 

it is a combination of Gaussian smoothing and differential 
derivation, usually used to approximate the calculation of 
image brightness, according to the target image edge 
processing of the region, the region of a brightness greater 
than a certain threshold value of the special point marked. The 
edge detection technology is widely used in computer vision, 
which can clearly show the tiny structures and details existing 
on the surface of objects. When detecting edges, Sobel 
operator uses the gradient ᐁf to represent the probability and 
direction of the edge of an image f at the position of pixel 
(x,y) : 

ᐁ grad
g
g 															         (8) 

(4)Canny operator 
Canny operator is a multi-level detection algorithm. 

Compared with other basic edge detection operators, it is one 
of the most advanced edge detection operators at present. The 
main optimal edge is defined as the number of marks that the 
algorithm can mark the edge of the target image more times, 
and the marked edge can be closer to the edge in the actual 
image. While using Canny operator for edge detection, firstly, 
the noise interference in the target image should be eliminated 
by Gaussian filtering, and then the target image should be 
processed gradiently, and then the non-maximum value of the 
gradient image should be suppressed. Due to the edge 
interference in the gradient image, the maximum value of the 
pixel in the target area should be searched by non-maximum 
suppression method. Finally, to achieve the goal, the method 
of double threshold edge connection is adopted, and the 
existence of false edge is further tested. That is, the single-
channel gray image is filtered and denoised, and then the 
amplitude of the image gradient (x,y and direction Angle θ) is 
calculated, and the image edge is finally determined by the 

double threshold. 

(x,y=                         (9) 

θ=arctan ()                    (10) 

In the formula, the gradient amplitudes in the horizontal 
and vertical directions of the image are respectively 
represented. 

3.2. Comparison and Analysis of Experimental 
Results 

3.2.1. Experimental Raw Data 
As shown in the following figure 2, the research objects of 

weld marks are: 

 
Figure 2. The research objects of weld marks 

3.2.2. Experimental Analysis 
For edge processing in 3.2 (a), Canny operator has the 

clearest detection, while Prewitt and Roberts operators are 
affected by image noise. Although Prewitt operator is 
interfered with, it obviously has some resistance, and Roberts 
operator is the most interfered with, which seriously affects 
weld marks. However, among these operators, Sobel operator 
is the most accurate in detecting regional edge markers and 
the least affected by interference factors. 

 
Figure 3. The detection effect of operators 

 

 
Figure 4. Operator detection is affected by interference 

 
In the original figure 4, the detection of Roberts, Prewitt 

and Canny operators were all affected by interference, but the 
detection results of Roberts and Canny operators were the 



 

128 

clearest, so the detection results of Roberts and Canny 
operators were better. Although Sobel operator marked the 
contour, However, Sobel operator leads to unsatisfactory 
image result because of the large pixel distance in the image 
area. 

3.2.3. Analysis of Experimental Results 
In terms of the overall data of the experiment, the detection 

of Canny operator is the most stable, and the picture is very 
clear. With the interference of various image factors 
considered, the operator itself has a high resistance. The 
detection of Roberts operator can achieve an ideal level in the 
processing of low-noise images. When the Prewitt operator 
labels high-noise images, the significant resistance effect is 
more significant. However, when comparing the results of 
Roberts and Canny operator, the clarity of the images detected 
by the Prewitt operator is significantly lower than that of the 
two operators. Through the analysis and comparison of 
several commonly used algorithms, it is found that for 
different types of objects, the corresponding image processing 
technology can get more satisfactory processing results, and 
these methods have certain applicability. When the Sobel 
operator is marked as the weld contour, the image marking 
effect is much more obvious than that of the rest operator. 
However, in the face of objects with not dense pixels in the 
image area, the result is not clear enough and cannot be 
accurately marked. 

4. Conclusion 
In the weld seam marking, interference by different factors 

is unavoidable, and the high-noise image should be pre-
processed by noise reduction to achieve the effect of 
optimizing the image quality and reducing the interference 
degree. Based on the edge detection algorithm of Sobel, 
Roberts, Prewitt and Canny operator, the experimental 
verification is carried out for the analysis of image weld 

marks. The experimental results show that the detection effect 
of Canny operator is the most stable. 

References 
[1] Kiddee, Prasarn, Zaojun Fang, and Min Tan. "An automated 

weld seam tracking system for thick plate using cross mark 
structured light." The International Journal of Advanced 
Manufacturing Technology 87 (2016): 3589-3603. 

[2] Sreedharan, J., and A. K. Jeevanantham. "Optimization of 
injection molding process to minimize weld-line and sink-mark 
defects using Taguchi based grey relational analysis." 
Materials Today: Proceedings5.5 (2018): 12615-12622. 

[3] Wei, Guoqian, et al. "Experiment and simulation investigation 
of multiple cracks evolution at the weld toe." International 
Journal of Fatigue 144 (2021): 106037. 

[4] Shi, Wen, et al. "Wire melted mark metallographic image 
recognition and classification based on semantic segmentation. 
"Expert Systems with Applications 238 (2024): 122146. 

[5] Wang, Weifeng, et al. "Study on the microstructure 
characteristics of melting marks in copper-clad aluminum 
wires under overcurrent faults." Journal of Thermal Analysis 
and Calorimetry149.1 (2024): 345-357. 

[6] Wang, Ren**, and Yong Lei. "Application of Computer 
Virtual Reality Technology in Welding Simulation Training of 
Complex Structures Based on Unity3D." Journal of Electrical 
Systems 20.2 (2024): 641-652. 

[7] Wiegand, Michael, et al. "High-strength dissimilar welds 
between a NiTi shape memory alloy and titanium obtained by 
intermixing niobium using pulsed laser beam welding." 
Metallurgical and Materials Transactions A55.1 (2024): 278-
290. 

[8] He, Jiadong, Haibo Wang, and Yajun Zhang. "Weld quality 
measurement of T-joints based on three-dimensional 
scanner." The International Journal of Advanced 
Manufacturing Technology (2024): 1-12.