YOLO: An Improved High-Accuracy Method for PCB Defect Detection

Throughout the manufacturing process of printed circuit boards (PCBs), any flaws on the surface can negatively affect the quality of the product, leading to potential issues with equipment performance stability and reliability. Small defects are rare in PCBs, making their features weak and easily affected by background interference, which makes detection difficult. To tackle the issue, this study proposes the VSP-YOLO model. By introducing the VSConv feature extraction module, which combines RepVGG blocks and space-to-depth (SPD) technology, the model optimizes information extraction and retention processes. Inspired by RepVGG's multi-path structure and SPD's spatial segmentation, VSConv effectively captures and retains information related to small defects. Moreover, the model employs the Sequential Polarized Self-Attention (SPSA) mechanism in the detection head to enhance target recognition while removing the detection head used for large targets, thereby reducing the parameter count. Experimental results demonstrate that the VSP-YOLO model achieves an average precision [email protected] of 97.9% and a [email protected]:0.95 of 62.14% in detecting small targets and high-resolution images, respectively. Compared to the traditional YOLOv8 model, the VSP-YOLO model improves performance by 2.4% and 3.5%, respectively, while reducing parameters by 23% and achieving a frame rate of 136 frames per second (FPS).