Application of Image Edge Detection Technology in Welding Environment Recognition
Foreword Welding production automation and intelligence has become an important direction for the development of welding technology. In the intelligent process of welding, the application of sensors is one of the key technologies. Among them, optical sensors have high tracking accuracy, fast dynamic response and abundant information, and are one of the most studied sensors. Image sensing has the characteristics of non-contact and is less affected by welding arcs and electromagnetic fields. It has been widely used in the field of welding and provides powerful means for the automation of welding operations. The small size and low price of visual sensors such as CCDs have also made their industrial applications possible. At present, visual sensing and image processing technology have been widely applied to weld identification, weld pool dynamic intelligent control, weld tracking, predicting welding organization, structure and performance, etc. [1-3]. 1. Characteristics of weld edge types In image processing theory [7], image edge points may correspond to different physical meanings. According to the corresponding physical meanings, the edges can be divided into the following 4 categories (as shown below): Image Image Edge Type (1) Type 1 edge. For the edge line labeled as 1 in Figure 1, the edge of type 1 is the intersection of two different curved surfaces or planes. At this point, the normal direction of the surface of the object is not continuous, and the gray value of the image is on both sides of the type 1 edge line. There are significant changes. 2. Common edge extraction algorithms and their applicability in welding environment identification As mentioned earlier, the edge extraction method can be roughly divided into gradient detection methods (such as Roberts operator, Prewitt operator and Sobel operator, etc.), second derivative zero crossing detection method, statistical method, wavelet multi-scale detection. , fuzzy mathematical methods, as well as mathematical morphology, neural networks, edge flow method and other detection methods. So many algorithms can only say which one is better for a specific application area. In 1986, Canny [8] summarized the results of past theories and practices and proposed Canny's three criteria of edge detection: good detection results, good positioning, and low repetition response to single edges, and their mathematical expressions are given. Based on the characteristics of the welding environment, the applicability of the existing edge extraction algorithm to the recognition of the welding environment is analyzed below. 2.1 Differential Operator A gradient is a vector whose magnitude and phase are: The algorithm needs to calculate every pixel position, so that the operation is very large. In practice, small templates are often used to approximate the convolution operation. Gx and Gy each use a template. The commonly used template is the Robert operator. The more complex common templates are Prewitt operator, Sobel operator, and Kirsch operator. The quality of these types of operators depends on the structure of the noise. If the noise is the same at each point, the Prewitt operator is better. If the noise near the edge is twice the edge, then Sobel operator is better. This gradient operator considers both the amplitude and direction parameters. In the identification of the welding environment, we hope to obtain the information of the weld in real time, and at the same time we hope to predict the direction of the progress of the weld below. Therefore, the gradient operator provides us with a feasible method. However, the gradient operator is sensitive to noise. This can identify many false edges for a complex welding environment. Therefore, the gradient operator alone is not ideal for the weld edge extraction. You can smooth the image first to improve the results, smooth out some edges that are close together, and affect the positioning of the edges. The edges obtained after convolving with these templates may span several points instead of one point. Therefore, two factors should be considered simultaneously. The edge pixels are not only larger than the threshold, but also the size of the gradient in the direction of the gradient is larger than that of the edge. The former and its latter, this method is called non-extreme suppression. 2.1.2 Laplacian Operator Where s is a set of neighbors centered on f(m,n), which may be 4 or 8 neighbors. Can be directly Gradient operators and Laplacian operators are all sensitive to noise. This can be done before the edge extraction with the neighborhood averaging method for smoothing, while the Gaussian two-dimensional low-pass filter can be used to filter the image, and then for Laplacian edge extraction. This is the commonly used Laplacian-Gauss operator. There are many noise points on the welding workpiece, and the differential operation has a “diffusion†effect on those isolated noise points, especially the Laplacian operator, which not only spreads but also significantly increases the intensity. Therefore, it is better to remove the noise before the differential operator detects the edges. For butt welds, the edge feature is shown as a thin line when the gap is small, and it becomes wider after the differential operation. Because the gradient operator can detect the direction of the image edge, it is more suitable for the recognition of the welding environment. 2.2 wavelet multi-scale edge detection method [9] [10] For the welding environment, this method has a good adaptability. It can search the target from a large scale on the workpiece or the weld pool and then extract the interesting details. 2.3 Mathematical morphology method [11] 2.4 Sub-pixel edge detection algorithm The edge detection algorithms described above are performed at the pixel level. Sub-pixel edge detection refers to decomposing pixels near the edge to accurately determine the edge. The sub-pixel edge detection maps the image data to a Hilbert space with 9 parameters to determine the edge parameters. Ghosal and Mehrotal first proposed the use of Zernike moments (Zernike Moments ZMs) to detect subpixel edges. In their algorithm, they established an ideal step grayscale model for the edges. By calculating the three different order ZMs of the image, The four parameters of the ideal step grayscale model are mapped into three ZMs, and then through these three ZMs, the parameters of the straight line where the edge is located are calculated to determine the sub-pixel level coordinates of the edge. Li Jinquan [12] conducted a more in-depth study of the ZMs algorithm, and pointed out its inadequacies and proposed a corresponding improved algorithm, which was applied to the weld seam recognition, the edge detection has high precision, self-refining edges and Strong anti-interference and other advantages. 3. Conclusion The direction of most welds will not change too sharply and are continuous straight lines or curves. In a small local area can be seen as two parallel lines. Therefore, when the welding environment is identified, the weld seam can be detected by searching for a straight line. In these existing algorithms, the gradient operator can detect the edge of the weld and can also predict the direction, so that real-time image processing can also predict the direction of the weld advancement, and is more suitable for the identification of the welding environment. However, the differential operator has poor anti-interference performance. For complex welding environments, it cannot be used directly, but it should be improved and combined with other algorithms. Wavelet multi-scale, morphological edge detection algorithms are one of the research hotspots in this field. Its characteristics are suitable for complex welding environment identification and should be further studied. Some sub-pixel detection algorithms can obtain more accurate detection results, which is one of the efforts to improve image processing accuracy and welding results. The number of edge detection methods is numerous and has its specific scope of application. When selecting or developing new algorithms, we must consider the characteristics of welding itself. Ferrite Magnet Ring,Multipole Ferrite Magnet,Ferrite Ring Magnet,Sintered Ferrite Ring Magnet Sunny Fore Magnet Company Limited (Jiangmen Xinlike Magnet CO.,LTD) , https://www.magnet-xl.com
To visually identify welds or extract weld pool characteristics for weld tracking, penetration control, etc., one of the important steps is to extract the characteristic parameters of the weld or weld pool. For example, on a workpiece image, the gray level of the weld seam and the workpiece that forms the weld seam is discontinuous. Therefore, the weld seam appears as an edge on the workpiece image. For the bath image, the difference in the gray level between the bath and the arc is also reflected on the edge of the image, reflecting the shape characteristics of the bath. Therefore, the edge is an important feature. Poggio et al. [4] defined edge detection as "mainly metrics, detection, and localization of gray changes." The edge is related to the boundary of the object in the image but different. To use optical sensing and image processing technology for welding intelligence, extraction of the weld or weld pool edge is a necessary process.
In 1959, the earliest mention in the literature was edge detection [5]. In 1965 LG Roberts first began systematically studying edge detection [6]. Every year there will be many articles on edge detection. Most of the important articles are published in IEEE Trans. On Pattern Analysis and Machine Intelligence, CVGIP: Image Processing, IEEE Trans. On Image Processing, Journal of the ACM, etc. Although there are thousands of edge detection methods, none of them have versatility and wide adaptability. They can be used directly in a particular application, so the characteristics of existing algorithms are explored to find suitable welding characteristics. The algorithm and the development of new algorithms for the welding process have certain practical significance. 
(2) Type 2 edges. Type 2 edge lines result from different material types or color differences. The above figure is composed of two different materials. Due to their different reflection coefficients of light, the two sides of the two edge lines have significant changes in the gray level.
(3) 3 types of edges. Class 3 edge lines are the boundary between the object and the background. As shown in the figure above, there are two types of edge lines on the cylinder, which are generally called outer contour lines. In the three types of edge points, the normal direction of the surface of the three-dimensional object is continuous. When the edge point is seen, the three types of boundary points are the boundary between the object and the background. Since the object and background differ greatly in light conditions and reflectance of the material, the gray levels of the image also significantly change on both sides of the three types of edges. The edge marked 3' in the figure is the boundary between the object and the background, and it is also the discontinuity of the surface normal on the object, but the cause of the gray transition on both sides is the former.
(4) 4 types of edges. 4 is the shadow caused edge. As a part of the surface of the object is blocked by another object, it is not illuminated by the light source, which causes a large difference in the gray values ​​on both sides of the edge point.
For joints before welding, either butt welds, beaded welds, or overlapped welds exhibit two distinct edges, which can be seen as overlapping of edges for overlap. For the weld pool, it shows a sharp edge profile corresponding to the shape of the bath. At the same time, the welding environment is more complicated, such as the inconsistency of the surface of the weldment (scratches, oxidation color, marks, oil, etc.), and the interference of arc. For strongly reflective workpieces such as aluminum, there are also reflections of light, reflections of torches, and the like, making it difficult to identify. Analyzing these characteristics of welds and weld pools, it can be seen that the edges of the welding environment include the above-mentioned edges 2, 3, and 4. The influence of these factors must be considered when selecting and proposing new edge extraction algorithms.
2.1.1 Gradient Operators The gradient corresponds to the first derivative, and the corresponding gradient operator corresponds to the first derivative operator. For a continuous function f(x,y), its gradient at (x,y) is defined as follows: 
The Laplacian operator is a second-order differential operator whose general representation in digital images is: 
As edge pixel grayscale, it can also be The pixels act as edges.
The wavelet transform can be understood as the following: compare the original signal with the left end of the wavelet function, find the similarity coefficient of the two functions, and then shift the wavelet function to the right distance of a wavelet function, perform comparison and calculation, and complete the entire process. Signal operations; this gives a scale of wavelet coefficients. The wavelet function is expanded and the above process is repeated to obtain a series of scale wavelet coefficients. The catastrophic point in the image is a key feature in analyzing the image, usually the edge feature of interest. Edge detection is the sudden change from the direction of the gradient of the wavelet coefficients. In order to detect the edge and detail features of the large target structure in the image, the researchers proposed the concept of multi-scale edge detection, ie detecting the large edge of the target on a large scale and detecting the target details on a small scale. Related theories can be found in [9][10]. This method is one of the hot spots in current image processing and has a good development prospect. The existing literature has applied it to the processing of weld pool images [10].
Mathematical morphology is the theory of studying the morphological structure of digital images and the method of rapid parallel processing. It is the purpose of structural analysis and feature extraction through morphological transformation of target images. Mathematical morphology takes the morphological characteristics of images as its research object. Its main content is to design a set of concepts, transformations, and algorithms to describe the basic features and basic structure of an image, that is, to describe the elements and elements in the image, and between the parts and the parts. Relationship. The object and image features in the image directly depend on the shape. The purpose of mathematical morphology is to study the shape in the time domain, so the morphology is suitable for image processing. Corrosion, expansion, opening, and closing in morphological operations are based on sets of operations. The structural elements play a key role in these operations, adjusting the geometry of the image feature transformation. With the aid of morphological operations, image edge detection operators can be introduced. The dilation and erosion operations in mathematical morphology have a very intuitive geometric background. They can make the processed image thicker or thinner in a certain direction. The difference between the original image and the two operations can be used as a full-scale edge. Detection, or you can detect the edge of the image. In addition, the morphological method can also modify the edge of the acquired image through an adaptive method, and gradually adjust the size of the structural element window to achieve the purpose of effectively enhancing the fuzzy edge and appropriately eliminating the influence of noise.