Multimedia Systems Компьютерные сети и коммуникации
Process-prior-constrained segmentation for interpretable quality assessment of wafer dicing marks
Аннотация
Wafer dicing marks are typically characterized by slender morphology, low contrast, and strong directional continuity, and their morphological quality directly affects chip yield and packaging reliability. Existing wafer defect segmentation methods mainly rely on general-purpose pixel-level modeling, which limits their ability to exploit the structural prior of dicing streets. They also lack an effective mechanism for translating segmentation results into interpretable quality decisions. To address these issues, this paper proposes a process-prior-constrained segmentation method for interpretable quality assessment of wafer dicing marks. First, a street-aligned coordinate system guides learnable region-of-interest (ROI) extraction to focus on the dicing street and suppress background interference. A spatial–frequency-aware dual-branch network is then designed, with the spatial branch preserving boundary continuity and the frequency-aware branch enhancing the high-frequency saliency of low-contrast dicing marks. Meanwhile, an ROI-constrained structural-consistency semi-supervised optimization strategy is introduced, which integrates Teacher–Student consistency, dual-branch consistency, and street topology constraints within dicing-mark-related regions to improve structural preservation under limited-sample conditions. Finally, process-related indicators, including area, width, length, and branch complexity, are extracted from the segmentation masks to achieve traceable OK/NG quality assessment. Across 10 repeated stratified image-level splits of the WaferDM dataset, the proposed method achieved an mIoU of 0.9694 ± 0.0064 and an mDice of 0.9843 ± 0.0033. NG recall was 1.0000 in all evaluated test splits.