A High-Precision Wafer Defect Inspection System Integrating DMD-Based Parallel Confocal Imaging, Bright-Field Inspection, and MutexMatch Semi-Supervised Learning1

Wafer defect inspection is a critical process in semiconductor manufacturing, directly impacting product yield and overall production quality. This study presents a high-precision inspection system that integrates bright-field imaging with digital micromirror device (DMD)-based parallel confocal laser scanning. The system was physically implemented with an optimized optical and mechanical design to achieve fast and accurate defect detection. To simulate real industrial scenarios, a simulated wafer defect image dataset was constructed using the proposed prototype system via bright-field inspection of optical standard test boards. The dataset approximates actual wafer defect inspection conditions. To further enhance classification performance under limited labeled data scenarios, we have incorporated and improved a semi-supervised learning approach, MutexMatch, which introduces a dynamic, adaptive and class-wise high-confidence threshold mechanism and achieves excellent classification performance even with extremely scarce annotations-reaching 84.12% accuracy with only one labeled sample per class. It significantly improves pseudo-label utilization and reduces reliance on manual labeling. Experimental results show that MutexMatch outperforms multiple baseline methods in classification accuracy, demonstrating strong robustness and effectiveness. Overall, the proposed system provides a stable and efficient solution for wafer defect inspection, offering substantial potential for enhancing quality control and yield in advanced semiconductor manufacturing.