Texture surface defect detection of plastic relays with an enhanced feature pyramid network

Deep learning has seen its promising applications in manufacturing processes. In this study, a deep network named Cascade Tri-DFPN based on the two-stage target detection algorithm is proposed for detecting the defects on the texture surface of plastic relays. The network adopts a derivative Resnet-101d as the backbone to obtain a loss-reduced feature extraction. Meanwhile, an enhanced feature pyramid module is put forward to enhance the feature representation of the network by adding a dense connection of feature layers through the self-attentive block. Moreover, the defect region proposals are optimized by introducing a cascade module to obtain high-quality defective proposal boxes. Experimental results on the augmented data set of relays' surface defect reveal an average accuracy of 88.57% and an average recall rate of 94.58%, much higher than those of traditional RCNN or FPN detectors, demonstrating the remarkable improvement of the proposed network. Robustness of the method is also verified by performing tests with deteriorative image processing, which indicates an eligible defect detection under relatively complex scenarios such as image blurring. The proposed deep network could be used in surface defect detection of plastic relays and other potentially related industrial defect detection fields.