AttSwin-YOLO: beyond efficient YOLOv5 for PCB defect detection

As the most fundamental and crucial component in electronic devices, Printed Circuit Board (PCB) has a significant impact on the stability and reliability of electronic products. However, owing to the intricate manufacturing processes involved in PCB production, various defects frequently arise, resulting in missed detections and false alarms. Against the backdrop of continuous advancements in deep learning, research on object detection has made remarkable progress. Applying object detection algorithms to the field of PCB defect detection can efficiently and accurately detect defects, thereby reducing missed detections and false alarm rates and consequently lowering detection costs. Therefore, this paper proposes AttSwin- YOLO, which takes the lightweight YOLOv5s as a baseline and adjusts the backbone of the network based on the Swin Transformer to better enhance the network's feature extraction and generalization capabilities. In addition, Coordinate Attention is introduced to provide better feature correlation, further improving the detection accuracy. Extensive ablation experiments are conducted on printed circuit board dataset from Peking University. Compared to baseline network, AttSwin-YOLO achieves an mAP of 95.4%, which is a 2.4% improvement over the baseline. Meanwhile, the mAP for multiple types of defects is also improved. Furthermore, the model achieves an accuracy of 96.7%. These experimental results fully demonstrate the effectiveness of AttSwin-YOLO.