Defect Pattern Classification of Wafer Graph based on Dual Channel Convolutional Neural Network

For the issue of wafer defect classification in semiconductor manufacturing, we propose a lightweight method for nonhomologous two-channel feature extraction method based on a convolutional neural network (CNN), employing CNNA and CNNB channels. In both channels, we replace the traditional fully connected layer with a global average pooling layer, thereby reducing parameter computation. Initially, addressing the imbalance in wafer defect categories using oversampling and standardize image size through nearest neighbor interpolation. Subsequently, to comprehensively capture wafer image features, CNNA employs larger filter sizes while CNNB utilizes smaller ones, supplemented with zero padding. Finally, we fuse the feature vectors from both channels for classification. The method is validated on the WM-811K dataset, achieving a classification accuracy of 75% without denoising. Comparison with other models highlights the superiority of the proposed model.