<![CDATA[In this paper, we propose a deep learning-based method for wafer flaw detection and classification in lab-label photolithography, known as a core step of the semiconductor manufacturing process. In photolithography, defects due to particles or process errors are critical to product yield and reliability. To detect these flaws, images were collected and efficientnet deep learning method was applied. Data augmentation and model lightweighting techniques were also applied to improve the limitations of the dataset. experimental results showed the relation between model complexity and the amount of training data. For EfficientNetB5, the massive architecture, caused the overfitting problem because of learning even noise in small datasets. But EfficientNetB0, the lightweight model, with batch normalization and early stopping techniques shows improvement of the reliability. In conclusion, this study provides practical guidelines for building and efficient flaw detection method in a data-limited research environment.]]>
Copyrights © 2026
