YOLOv11-DMSFD: A PCB Defect Detection Algorithm with Multi-Scale Feature Enhancement

Abstract

To address weak micro-defect features, large-scale variations, and edge information loss during feature fusion in PCB defect detection, this paper proposes a multi-scale feature enhanced algorithm named YOLOv11-DMSFD. First, a C3K2DD module is designed by introducing the DWRD feature extraction unit, which combines the multi-scale residual feature extraction strategy of DWR with the large-receptive-field reparameterization mechanism of DRB to enhance local and global feature representation. Second, ADown and DySample are introduced to optimize downsampling and upsampling, respectively, improving micro-defect feature preservation and boundary recovery. Finally, a Multi-Scale Attention Module and a Full-path Aggregation and Distribution mechanism are incorporated into the neck network to enhance multi-scale feature interaction and information transmission. Experimental results show that YOLOv11-DMSFD improves mAP50 by 2.12 percentage points and mAP50-95 by 11.54 percentage points compared with YOLOv11n, while reducing parameters by 0.54M and model size by 0.8 MB. The proposed method achieves improved PCB defect detection performance while maintaining lightweight characteristics.