FCD-YOLO: Improved YOLOv5 Based on Decoupled Head and Attention Mechanism for Defect Detection on Printed Circuit Board

Defect detection technology is an indispensable quality control technology in PCB manufacturing. However, the effect of the existing PCB defect detection algorithm is not good. Therefore, we propose an improved PCB defect detection algorithm FCD-YOLO based on YOLOv5. Firstly, we add a set of anchors suitable for detecting small objects, and at the same time add a shallow prediction layer. Secondly, we modify the neck network and integrate more shallow features to improve the detection effect of small objects and texture features. Thirdly, we integrate a CBAM attention module into the neck network to improve the ability of the model to extract the features of the region of interest on complex PCB background. Finally, we integrate a decoupled head mechanism into the head network to help the model converge quickly and improve the detection effect of the model. The experiment adopts the public PCB defect image released by the laboratory of Peking University. The experiment proves that the precision of this method is increased by 1.4 %, the recall is increased by 1.6 %, the mAP@0.5 is increased by 0.6 %, and the mAP@0.5:.95 is increased by 1.7% compared with YOLOv5s, which is more suitable for detecting PCB defects than YOLOv5s.