Wafer Map Defect Recognition (WMDR) is an essential stage in the semiconductor manufacturing process. It is of great significance to detect and recognize wafer map defects precisely, so as to trace back and locate problems in the manufacturing process and solve them for improving the reliability and productivity of the semiconductor manufacturing process. The current intelligent methods for WMDR are limited in their recognition performance due to their complex structure and lack of effective solutions to the problem of feature weakness of the defects. Therefore, this paper proposes a WMDR model: cosine-weighted interactive enhancement network (CIENet), which is plugged into a cosine-weighted interactive enhancement module (CIEM). CIEM achieves feature enhancement for weak defects by performing an interactive cosine similarity calculation between feature maps and weighing them. Meanwhile, a pretrain-finetune strategy is proposed to train CIENet, which decouples the traditional training process to differentially and purposively optimize CIENet. To verify the effectiveness and superiority of the proposed method, comparative and ablation experiments are conducted on real-world semiconductor wafer datasets. The results show that the proposed model has higher recognition performance than other models, and the proposed pretrain-finetune strategy further improves the recognition performance of the model.

wafer map,feature enhancement,pretrain-finetune strategy,defect recognition