Research and implementation of an improved YOLOv11n-based method for PCB defect detection

Abstract

This paper presented YOLO-MUBL, an enhanced YOLOv11n-based approach designed specifically for PCB defect identification tasks. The original C3k2 building block was substituted with a novel C3k2-MambaOut-UniRepLK composite module, which combined gated convolutional feature interaction with large-kernel spatial modeling to strengthen the feature representation of minute defects. Furthermore, we embedded a Bi-Level Routing Attention (BRA) module into the backbone framework, which served to preserve high-resolution spatial details and support dynamic cross-scale feature communication. We also focused on improving edge deployment efficiency. To this end, we devised a lightweight shared convolution detection head (LSCD-Head). This design achieved two major reductions: It shrank the model’s parameter budget and also lessened the computational burden during inference. Experimental findings demonstrated that YOLO-MUBL attained 94.1% mAP@ 0.5 and 50.7% mAP@ 0.5:0.95 on the PCB defect dataset, representing improvements of 2.0% and 1.3%, respectively, over the baseline model. The parameter count decreased from 2.583 to 2.278 M, and GFLOPs dropped from 6.3 to 6.2. We proceeded with deployment on the Fibocom SC171V2 edge platform. Performance evaluation showed that each frame required only 30.8 ms for inference, while the system consistently operated at 25 frames per second. These results affirmed that our approach is well suited for industrial use scenarios.