FDDNet: A Fine-grained Detection Network for PCB Defects

Abstract

Defect detection is extremely important to improve the quality of PCB production. Although defect detection using traditional methods has achieved good results, a large number of false detections and missed detections cannot be avoided. In response to solve this problems, we propose a fine-grained defect detection network (FDDNet) model to improve the detection performance of PCB defects. This model increases the dimension of spatial context features in PCB defect detection to fuse multi-scale features, which helps the model to deal with more complex scenes. To facilitate the efficiency of feature fusion, we propose an improved channel attention module to enhance the learning efficiency of the network for effective features. To cooperate with the multiplexing of multi-scale feature maps in the backbone network, we propose a module capable of enhancing image recognition to extract pure shallow information. Finally, the experimental results on the PCB defect dataset show that the proposed method can achieve a mAP50 index of 97.32%.