Split Attention Mechanism of Faster RCNN for PCB Defect Detection

With the development of the electronics industry, the structure of printed circuit boards (PCBs) has become increasingly complex, making defect detection a crucial part of the production process. Existing PCB detection technologies face challenges such as difficulty in detecting minor defects and positional deviation. To address these issues, this paper proposes the SP-Faster RCNN network for real-time and accurate identification and localization of small defects on PCBs. Based on Faster RCNN, this network employs Resnet50 with a split attention mechanism combined with FPN as the feature extraction network, enhancing the message passing ability for features in different scales. Additionally, K-means clustering and GIOU are utilized to make the network more suitable for the identification and localization of small defects. Experimental results show that the SP-Faster RCNN network achieves an accuracy of 98.1 % on the dataset, with a detection speed of 1.34s per iteration, meeting the practical detection requirements of PCB industrial production.