Statistical-Analysis-Based Imaging-Free Method for Efficient and Fast Wafer Defect Detection

This article proposes an imaging-free method based on single-pixel imaging and statistical analysis for efficient and fast wafer defect detection. First, the chi-square test, a commonly used statistical analysis approach, is used to determine the correlation between projected patterns and the presence or absence of defects in wafers. The selection of the most efficient patterns is informed by the top k chi-square values. Moreover, a straightforward feature calculation method is used instead of deep learning algorithms to reduce the computing resource requirement. Results show that the average recognition accuracy is 97.16% at a sampling rate of 30.5% and still reaches 84.69% at a low sampling rate of 1.5%. A comparison of the proposed method with the traditional feature extraction strategy based on convolutional neural network (CNN) models demonstrates that the former decreases the model size by 4–5 orders while maintaining comparable accuracy. The statistical-analysis-based imaging-free method is efficient and resource-friendly, and it could be a potential solution for wafer defect detection.