Based on the low efficiency and high cost of conventional manual and electrical methods for detecting defects in PCB production, a PCB defect detection method based on YOLOv5 algorithm is proposed, which adds a prediction head for small object detection to form a four-dimensional detection, so as to improve the detection effect of small objects; ASFF (adaptive feature space fusion) is added to YOLOv5s original FPN + PANNET structure for feature fusion to ensure that each space can adaptively fuse different levels of feature information; GAM(global attention mechanism) is added to the original network, and attention operation is applied in all three dimensions , which strengthens the ability of model information extraction. The experimental results show that the improved defect detection method can accurately classify six kinds of defects, and the average accuracy can reach 98.8%. It has a certain reference value for the deep learning PCB defect detection method.