Design of defect detection system for process parts based on machine vision

This study addresses the problem of surface defects in parts produced on traditional production lines and designs a process defect detection system based on machine vision. Firstly, taking the detection of production defects in seat components as an illustrative example, we select appropriate imaging equipment and construct imaging platforms specifically designed for horizontal movement detection and horizontal rotation detection, respectively. Secondly, improvements are made on the basis of the YOLO v5 model. By introducing rotating rectangular frames, it avoids overlapping of real boxes when identifying target characteristics, prevents large real boxes from completely surrounding small real boxes, and reduces redundancy of target box information. Lastly, in terms of lightweight network model design, by adding the ShuffleNet-V2 network to the YOLO v5 model, the use of depthwise separable convolutions not only improves the model's performance but also significantly reduces its computational complexity, significantly reducing the required computing and storage resources for deployment on mobile terminals. The experimental results revealed that the enhanced YOLO v5 model achieved an average accuracy improvement of 3%, 11.5%, and 8.2%, respectively, over the classic YOLO v5s, SSD, and Faster RCNN models. Its detection speed also surpassed that of the SSD and Faster RCNN models. The on-site verification of the production line showed that the improved model has greatly improved in both the speed and accuracy of testing compared to the original model.