Recently, deep learning methods are often used in wafer map failure pattern classification. CNN requires less feature engineering but still needs preprocessing, e.g., denoising and resizing. Denoising is used to improve the quality of the input data, and resizing is used to transform the input into an identical size when the input data sizes are various. However, denoising and resizing may distort the original data information. Nevertheless, CNN-based applications are focusing on studying different feature map architectures and the input data manipulation is less attractive. In this study, we proposed an image hash layer triggered CNN framework for wafer map failure pattern retrieval and classification. The motivation and novelty are to design a CNN layer that can play as a resizing, information retrieval-preservation method in one step. The experiments proved that the proposed hash layer can retrieve the failure pattern information while maintaining the classification performance even though the input data size is decreased significantly. In the meantime, it can prevent overfitting, false negatives, and false positives, and save computing costs to a certain extent.

Recent research applying deep learning to the field of defect pattern recognition in wafer maps has greatly accelerated the process of defect detection. However, when different defects are mixed on the same wafer, the mixed type is very complex, and it is still difficult to recognize the defect pattern. In this article, we propose a new framework to segment different defect patterns on the wafer map by using a semantic segmentation approach. This method works well on single and known and unknown mixed types. First, we extract the defects from the single-defect wafer map of the MixedWM38 dataset to generate single-defect pixel-level labels. Then, a mixed defect pattern dataset suitable for semantic segmentation is generated using single-defect wafer maps and single-defect pixel-level labels. The average accuracy of the test set on our synthetic dataset reaches over 97%, and using the trained model for testing on MixedWM38, we can get an average accuracy of 95.8%.