Research On Automatic Detection And Segmentation Of Welding Seam Targets In Dust Environment

In recent years,under the background of the deep integration of advanced manufacturing and information technology,the transformation and upgrading of digital intelligence of enterprises has been gradually accelerated,and the level of domestic industrialization and intelligent manufacturing have reached a new level.At present,the grinding and polishing industry is gradually abandoning the traditional manual grinding method,with robot based on visual technology positioning to automate operations,but the actual workshop of the harsh drilling,milling and grinding environment is easy to interfere with the computer vision system or industrial application equipment,such as the metal dust suspended and accumulated in the air and uneven illumination lead to the reduction of weld imaging quality,and the weld target contour and its boundary with the background base metal are not clear,as a result,detection accuracy and speed are still tough to meet industrial needs.In order to help the weld grinding and polishing industry transform from automation to intelligence,optimizing the neural network model structure according to the characteristics of actual working conditions has become a research hotspot,so as to improve the accuracy,speed and robustness of the weld target detection model,while improving the ability of its migration,application and deployment.Aiming at the task of automatic detection and segmentation of weld targets in the industrial dust environment,this paper proposes to construct a high-precision integrated recognition,positioning and segmentation algorithm system of weld targets.The main research contents of this paper are summarized as follows:1.The image of degraded weld is taken as the research object,and based on the physical model of image degradation and the traditional dark priori theory,the image restoration principle is explored;Optimize the image restoration algorithm to improve the problem of dark,distorted and blurred outline of the restored image,and to obtain clear weld images with strong visibility.2.The fundamental theories and logical connection of dominant image segmentation algorithms based on deep learning are depicted,and analyzes the realization principle and network structure of two instance segmentation algorithms,Mask R-CNN and YOLACT;The effectiveness of weld target detection is verified by two segmentation algorithms and their superiorities and weaknesses are compared and elaborated.3.The expansion of weld image samples comes from the traditional image data enhancement method and the application of generative countermeasure network.Based on the results of K-means clustering analysis,the anchor frame generation mechanism and RPN layer of Mask R-CNN algorithm are improved;The network detection frame selection module was improved by Soft-NMS,and multiple groups of control experiments showed that the weld identification accuracy reached 93.8%,while the detection rate was increased to 0.341 pieces /s,indicating good robustness of the model.4.Design and develop a graphical user interface based on Py Qt5 in the Windows environment,packaged into an executable program that can be independent of the deep learning environment,the software can realize the preliminary application of the improved algorithm for the system camera,local video or pictures;Self-built a welding seam detection platform,and experimentally verify each functional module and detection effect of the software,weld inspection visualization results can be quickly and easily obtained by inspection personnel.