Efficient Training for Automatic Defect Classification by Image Augmentation

Abstract

At semiconductor wafer production sites, an automatic defect classification (ADC) system is used to analyze defects. The ADC system automatically classifies defect images into user-defined defect categories. Every manufacturing process requires the ADC system to be set up by acquiring training samples and teaching defect categories in advance since different defects occur for each process. Since the setup is time-consuming, there is a need for more efficient training from a small number of training samples. This paper describes an ADC system that can be efficiently trained by image augmentation to improve the efficiency of the defect analysis. The image augmentation of the proposed ADC system includes rotation/flip and distortion processing. The rotation/flip processing restricts rotation angles and flip axes to generate new images as if the images were acquired by actual tools. The distortion processing generates new images of defects that are likely to exist by considering distortion amount and continuity. These processes increase effective appearance variations for the training while maintaining the consistency of defect appearances in each category. Experimental results demonstrate that the proposed ADC system maintains the classification accuracy even with half the number of actual samples for the training.