An enhanced network model for PCB defect detection: CDS-YOLO

This study introduces CDS-YOLO, an enhanced YOLOv5s network model, to boost accuracy and real-time performance in detecting defects in Printed Circuit Boards (PCBs). Initially, the model incorporates a Convolutional Block Attention Module (CBAM) into its architecture to heighten focus on diverse positions within the input image. Subsequently, the model employs Distribution Shifts Convolution (DSConv) to replace traditional convolution modules, reducing parameter count and boosting computational efficiency. Furthermore, the substitution of the conventional Complete Intersection over Union (CIoU) loss function with the SCYLLA Intersection over Union (SIoU) loss function optimizes bounding box prediction accuracy. Lastly, a high-resolution detection head (160 \(\times\) $\times$ 160 \(\times\) $\times$ 255) was incorporated to improve the model’s capability in identifying small-sized defects. Evaluations on the publicly available PKU-Market-PCB dataset showed that the CDS-YOLO model not only achieved an impressive mAP50 of 97.4% and a recall rate of 92.9%, but also operated with a low computational load of 13.5 GFLOPs and a fast detection speed of 117.5 FPS. The findings demonstrate that the CDS-YOLO model proficiently satisfies the stringent defect detection standards required for industrial-grade PCBs.