A Feature Pyramid Residual Network with Attention Mechanism for Wafer Map Mixed Defect Detection

In the semiconductor industry, wafer defect detection is facing unprecedented challenges as a key application of object detection. Lots of complex steps in the wafer manufacturing process can lead to defects and identifying these defect patterns accurately is critical to preventing defects and improving process reliability and product yields. With the complexity of the manufacturing process, the number of wafer map mixed defect has increased, which puts forward higher requirements for recognition accuracy. The existing deep learning-based defect recognition methods do not take into account the existence of multi-scale defects with different sizes when dealing with wafer map mixed defect, and the mixed diversity of wafer defect also limits the accuracy of the existing recognition methods. To solve the above problems, this paper proposes a feature pyramid residual network (FPRN) with convolutional block attention module (CBAM) named FPRN-CBAM. The FPRN-CBAM model not only realizes multi-scale feature fusion for wafer map mixed defect detection (WMMDD), but also enhances the defect feature representation by adaptively assigning channels and spatial weights. To verify the effectiveness and superiority of the proposed model, this paper uses the MixedWM38 dataset for experiments. The results show that the proposed model is effective in detecting multi-scale features of wafer maps, which leads to better performance in recognizing mixed defect patterns.