Wafer map defect recognition based on multi-scale feature fusion and attention spatial pyramid pooling

Wafers are products in semiconductor manufacturing and serve as the foundation for producing semiconductor chips. During the wafer testing stage, functional and electrical parameters are examined to identify defects in chip design and fabrication. The wafer map is the result of the wafer testing process. Analyzing and classifying defective information on the wafer map aids in defect source identification and optimization of the wafer production process. Deep learning has been employed for defect detection on wafer maps because of its superior image processing capabilities. Nevertheless, as semiconductor chip design integration and wafer size increase, more complex types of wafer defects tend to emerge in the production process, and the size, shape, and distribution of wafer defects can affect the final classification outcome. Accordingly, this study proposes a deep learning model, called ESPP-Net (Attention Spatial Pyramid Pooling Network), that combines a deep convolutional neural network and attention space pyramid pooling to recognize and classify single and mixed-type defect wafer maps. We evaluated our model on both the mixed-type dataset Mixed38WM and the single-type dataset WM-811K and compared it with state-of-the-art deep learning models. Experimental results show that our proposed model outperformed the preexisting models and demonstrated superior classification performance.