A New Improved YOLO Based Network for PCB Surface Defect Detection

PCB (Printed Circuit Boards) surface defect detection plays a vital role in upgrading product quality. However, the defect detection of small targets still faces great challenges. Considering that, this paper proposes a new multi-scale feature enhancement and feature fusion module(MSFE-FFM) to improve processing multi-scale features in both the backbone and the neck based on YOLO. Moreover, a multi-scale convolutional attention (MSCA) module is added to improve the feature extraction for small targets with low resolution. Experiments are conducted to compare with other art-state models on two public datasets DeepPCB and HRIPCB. Ablation experiments are also conducted to show the effects of our proposed modules. Evaluation results show that mAP increased to 97.5% on DeepPCB and 99.1% on HRIPCB using our proposed method. It has higher mAP valules than the art-state methods.