| Title |
Diversity-Aware Diffusion Augmentation for Long-Tailed Wafer Bin Map Defect Classification |
| Authors |
유석호(SeokHo You) ; 김현진(HyunJin Kim) |
| DOI |
https://doi.org/10.5573/ieie.2026.63.9.99 |
| Keywords |
Wafer bin map; Data augmentation; Diffusion model |
| Abstract |
Wafer bin map (WBM) defect classification is essential for semiconductor yield management, but real manufacturing data often exhibit a long-tailed distribution with few samples for rare defect classes. We fine-tune Stable Diffusion with DreamBooth per defect class, use real defect-free WBM maps for domain-aligned prior preservation, and select valid, diverse synthetic images using robust CLIP similarity thresholding and farthest-point sampling in CLIP space. To separate augmentation method from data quantity, we compare real-only training, geometric augmentation, size-matched generative augmentation, and their combination. Experiments on five WM811K defect classes show that generative augmentation improves macro-F1 and balanced accuracy over no augmentation and geometric augmentation under both 30/80 sample per class and imbalanced long-tail settings. |