A Lightweight PCB Defect Detection Algorithm Based on an Improved YOLOv11s

Le résumé fourni par la source

To address the limited accuracy of existing PCB surface defect detection methods, the challenge of balancing detection performance with lightweight design, and the need for real-time deployment in electronic manufacturing systems, we propose a lightweight PCB defect detection algorithm based on an improved YOLOv11s. First, a lightweight EfficientViT module is introduced into the backbone to replace the original backbone, reducing computational resource usage while enhancing inter-channel information exchange. Second, we integrate FasterNet partial convolution (PConv) with an efficient multi-scale attention mechanism (EMA), and design a C3k2_Faster_EMA convolution to replace standard convolutions in the neck module, further decreasing parameter count and strengthening fine-grained feature extraction. Finally, by combining Inner_IoU auxiliary bounding boxes with the dynamic focusing mechanism of Wise_IoU, we construct an Inner_Wise_IoU loss to replace the original CIoU loss, improving localization accuracy for small targets and bounding-box regression performance. Experiments show that the improved method achieves strong performance on the PKU-Market-PCB and DeepPCB datasets, attaining an mAP of 92.3%. Compared with the original YOLOv11s, parameters are reduced by 61.7%, computational complexity by 63.9%, and model size by 58.3%, down to 8.0MB, while reaching an inference speed of 227.54 FPS. To validate deployment feasibility on edge devices, we deploy the model on the LubanCat-5 platform. The results indicate that the quantized model size is 6.1MB, with an average frame rate of 23.094/s and an average inference time of 43.301ms; relative to the non-quantized version, the frame rate increases by 130% and inference time decreases by 56%. These results demonstrate that the proposed model achieves a favorable balance between lightweight design and detection accuracy, offering strong engineering value and practicality for modern electronic manufacturing systems.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.