Research on Quartz Wafer Defect Detection Method Based on Deep Learning Model Pruning

With the vigorous development of semiconductor wafer, the optical inspection of surface defects of quartz wafer has attracted increasing attention. However, current algorithms for surface defects detection are difficult to meet the following industrial requirements: sufficient generalization ability, robustness and fast processing speed. To meet those requirements, this paper proposes a quartz wafer defect detection method based on deep learning model pruning. First, we propose a light-weight target detection network, named Wafer-net, which is suitable for wafer defect detection. Compared with the classic light-weight YOLOv4-tiny network, the mAP and processing speed of Wafer-net on our quartz wafer dataset increases. Then, the Wafer-net network is further structurally pruned to obtain a Prune-net network. Finally, when tested on the server CPU side, the Prune-net network can reach 118 FPS, which is almost 9 times faster than the YOLOv4-tiny network.