Yu Cao
To address the issues of strong background texture interference, small defect scales, and the tendency for feature loss in the detection of defects on Printed Circuit Boards (PCBs) within complex industrial environments—which lead to high miss rates and false positive rates in existing detection algorithms—this paper proposes a PCB defect detection method based on multi-scale feature enhancement and background suppression. An end-to-end detection network is introduced. First, a multi-scale contextual feature enhancement module is designed. By constructing a parallel atrous convolutional pyramid with different dilation rates, it captures global context rich in semantic information. This is combined with shallow, high-resolution feature maps to fully preserve the edge and texture details of defects, addressing the information loss caused by down-sampling in deep networks. Second, a background decoupling and suppression attention mechanism is introduced. This mechanism separates foreground responses from background noise in the feature maps, generating a discriminative attention weight map. It adaptively enhances the response values in defect regions while suppressing interfering responses from complex circuit traces and copper foil backgrounds. Finally, an improved localization loss function is incorporated to optimize the regression accuracy of the predicted bounding boxes. Experimental results on the public PCB defect dataset PKU-Market-PCB demonstrate that the proposed method achieves a mean Average Precision (mAP50) of 96.8%, which is a 4.3 percentage point improvement over the mainstream algorithm YOLOv8n (92.5%). Particularly for defects such as nicks and pinholes, the recall rate is improved by more than 5.8%. Furthermore, the model achieves a detection speed of 85 frames per second, and its parameter count is reduced by 22% compared to the baseline, meeting the real-time requirements of industrial production lines. The method proposed in this paper effectively suppresses complex background interference and enhances the feature representation for multi-scale defects, significantly improving detection accuracy while maintaining detection speed, demonstrating good robustness and industrial application value.