An end-to-end wafer map defect recognition model

Wafer map defect recognition is one of the hot research topics in the semiconductor production field. This paper adopts a CNN-Transformer architecture to address the unified single and multiple defect wafer map problems, achieving accurate identification of different types of defects while maintaining high accuracy and efficiency. The model utilizes a U-shaped network and adaptable convolutional operations to construct the CNN backbone, generating high-resolution feature maps. Transformer is employed to capture global contextual information from the feature maps, producing a set of target query vectors for accurate prediction of the category and location of one or multiple defect patterns in the wafer map. Experimental results demonstrate significant performance improvement of the proposed model in wafer map defect recognition tasks.