An accurate detection method for randomly distributed welding slags using an improved Yolo v3 network

Abstract

Irregular welding slag spatters during welding of the car body are considered as extremely unavoidable defects. Detecting these tiny defects exhibiting properties in position and number is a challenging task. In this paper, a novel computational framework by modifying the structure of the Yolo v3 network is developed to accurately detect the randomly distributed welding slags. It has been verified that the improved Yolo v3 algorithm has improved accuracy by 7.4%, recall rate Rr by 6.8%, and real-time detection frame rate of 31.25 fps compared with Yolo v3, the detection performance is excellent in comparison with Faster RCNN and Yolo v3. Therefore, the improved algorithm can improve the accuracy of welding slag detection, which is beneficial to real-time welding slag detection and promote automated production and processing in industry.