論文ID: 21.20240054
Defect pattern detection of wafer bin maps (WBMs) is vital in wafer quality improvement owing to preventing further defects and resource waste. We proposed two Mixup approaches to train Vision Transformer under only single defect WBM samples for mixed-type defects recognition. We use UnionMixup and Token level Max-Min-Saliency Mixup to generate mixed-type defect WBMs to feed Vision Transformers. In the recognition of two-mixed defect types WBMs, our method improves 17.1 % compared to baseline (none mixup) and we have 1.7% accuracy gain compared with state-of-the-art mixup approaches. In the recognition mixed defect samples containing more than two-mixed defects (three-mixed and four-mixed), we gain at least 24.7% (compared with baseline) and 11.1% (compared with single SOTA mixup) respectively. The combination of Union Mixup and Token level Max-Min-Saliency Mixup become better than other SOTA mixup methods obviously in mixed patterns including more than three defects.
