Pre-trained CNN-based TransUNet Model for Mixed-Type Defects in Wafer Maps

• DOI: 10.37394/23209.2023.20.27
• Publication Date: 2023-07-19T10:53:08Z
• Authors: Youngjae Kim; Jee-Hyong Lee; Jongpil Jeong
• Keywords: Feature (linguistics)

Classifying the patterns of defects in semiconductors is critical to finding root cause production defects. Especially as concentration density and design complexity semiconductor wafers increase, so do size severity The increased likelihood mixed makes them more complex than traditional wafer defect detection methods. Manually inspecting for costly, creating a need automated, artificial intelligence (AI)-based computer vision approaches. Previous research on analysis has several limitations, including low accuracy. To analyze mixed-type defects, existing requires separate model be trained each type, which not scalable. In this paper, we propose segmenting by applying pre-trained CNN-based TransUNet using N-pair contrastive loss. proposed method allows you extract an enhanced feature repressing extraneous features concentrating attention want discover. We evaluated Mixed-WM38 dataset with 38,015 images. results our experiments indicate that suggested performs better previous works accuracy 0.995 F1-Score 0.995.