Open-Set Semiconductor Defect Classification in SEM Images via Uncertainty-Aware Local Prototype Alignment

Automated classification of scanning electron microscope (SEM) defects is complicated by acquisition-domain shift, long-tailed classes, and target-only defect types that invalidate the closed-set assumption. We propose an offline open-set adaptation framework that combines a three-criterion reliable-known selector with global and defect-region-guided local prototype alignment. Confidence, entropy, and prototype distance jointly determine which unlabeled target samples may be aligned to known source prototypes; samples that fail this gate are excluded from known-class alignment and optimized by an unknown-rejection objective. Unlike global-only adaptation, foreground-weighted local prototypes emphasize compact defect morphology rather than domain-specific background texture. Averaged over four bidirectional SEM adaptation tasks and three random seeds per task, the method achieves 80.8% balanced accuracy, 78.1% Macro-F1, 88.7% unknown AUROC, 74.2% H-score, and 8.4% ECE. Compared with the strongest baseline for each individual metric, the proposed method improves balanced accuracy by 4.4 percentage points and Macro-F1 by 4.5 percentage points over M2DD, improves unknown AUROC by 6.9 percentage points and reduces ECE by 3.1 percentage points relative to the energy-based baseline, and improves H-score by 5.9 percentage points over CMU. Component removal and sensitivity analyses confirm complementary contributions from reliable-target selection, prototype alignment, unknown rejection, and localized alignment. The framework is directly applicable to offline SEM inspection when labeled historical data and unlabeled new-condition images are jointly available; its selective-alignment principle may also transfer to other localized visual-inspection tasks after domain-specific validation and calibration.