Abstract: To address the limitations of existing defect detection methods, namely weak edge feature extraction for small targets, inadequate multi-scale perception under complex backgrounds and irregular morphologies, and limited localization accuracy, an improved YOLO11n-based PCB surface defect detection model, termed EMD-YOLO, is proposed. An Edge Enhancement (EE) module and a Multi-Scale Edge Information Enhancement (MSEI) module are introduced to strengthen the representation of edge features for minute and structurally complex defects. Additionally, a lightweight multi-scale pooling scheme is employed to construct a feature pyramid, by which the network’s multi-scale perception capability is enhanced. Furthermore, the Distance-IoU (DIoU) loss function is adopted during training to optimize bounding box regression, by which localization accuracy is further improved. Experimental results on the PKU-Market PCB dataset demonstrate that a 3.8 percentage point improvement in mean average precision (mAP@50) is achieved by EMD-YOLO compared to the baseline YOLO11n model, while nearly unchanged computational complexity is maintained, thereby significantly enhancing the model’s practicality and robustness.

Key words: PCB, defect detection, computer vision, loss function, multi-scale feature fusion