Efficient Object Detection and Recognition of Body Welding Studs Based on Improved YOLOv7

Abstract

The welding stud is a widely used part in automobile manufacturing, and its welding quality plays a crucial role in component assembly efficiency and vehicle quality. In welded stud target inspection, the complex body environment and different lighting conditions will have a certain impact on the inspection accuracy, and most of the existing methods have limited efficiency. In this paper, in order to solve the problems of low accuracy and slow speed in the stud target inspection process, we propose an innovative welding stud target inspection method based on YOLOv7. First, the EfficientFormerV2 backbone network is adopted to utilize the new partial convolution, which can extract spatial features more efficiently, reduce redundant computation, and improve the detection speed. Secondly, the bounding box loss function is changed to NWD, which reduces the loss value, accelerates the convergence speed of the network model, and better improves the detection of studs. After the test, the improved YOLOv7 network model is better than the traditional network in both speed and accuracy of welded stud target detection. (1) The mAP0.5 increased from 94.6% to 95.2%, and the mAP0.5:0.95 increased from 63.7% to 65.4%. (2) The detection speed increased from 96.1 f/s to 147.1 f/s. The results of the study can provide technical support for the subsequent tasks of automatic detection and position estimation of body welding studs.