A Fast Detection Algorithm for Surface Defects of Bare PCB Based on YOLOv8

Bare PCB (Printed Circuit Board) surface defect detection is an important step in the automated production of PCB. To meet the requirements of production efficiency, the detection must have high real-time performance, therefore we propose the MOP-Block module to replace the C2f module and also improve the YOLOv8 network structure to reduce the inference time. Then, considering that most of the surface defects on the bare PCB are small in size compared with the overall product, we propose the SiLU-SE attention module to enhance the feature map for improving the detection accuracy at a lower computational cost. The experimental results show that compared with the original YOLOv8n, our algorithms improves the FPS (batch=1, without TensorRT acceleration) on the Tesla T4 GPU by about 70.17%, the size of model is reduced by about 77.01 %, the parameters of the model are reduced by about 84%, and the FLOPs of the model are reduced about 53.08%, but map@0.5 and map@0.5:0.95 are only reduced by about 2.6(%) and 0.9(%) respectively.