Risk-Calibrated Few-Shot Industrial Anomaly Detection with Human-in-the-Loop Inspection

Automated visual quality inspection often operates with few normal reference images but still requires an explicit policy for asymmetric operational errors. This paper presents Risk-Calibrated Few-Shot Industrial Anomaly Detection Plus (RC-FS-IAD+), a few-normal-support, labeled-calibration-assisted framework that maps anomaly scores to automatic pass, manual review, and automatic rejection. The few-shot designation refers to the normal support set used to construct the anomaly representation; the decision stage is semi-supervised and uses a separate labeled calibration set. To avoid score-fitting/threshold data reuse, the main Visual Anomaly (VisA) evaluation uses stratified nested calibration with disjoint score-fusion and threshold subsets. At k=4, RC-FS-IAD+ achieves Area Under the Receiver Operating Characteristic Curve (AUROC) 0.848, Area Under the Precision-Recall Curve (AUPR) 0.579, F1 0.584, mean missed-defect rate 4.17%, and manual-inspection rate 45.34%; 65.0% of category-seed trials meet the empirical 5% missed-defect target. Across k ∈ {1, 2, 4, 8, 16} on VisA, AUROC increases from 0.828 to 0.872; manual-inspection rate (MIR) is 46.7% at k = 1 and 39.3% at k = 16, with variation across intermediate support sizes. Under the same held-out partition and threshold-calibration policy, WinCLIP yields slightly higher average ranking metrics but a higher missed-defect rate (5.45%), whereas RC-FS-IAD+ outperforms the DINOv2 patch-memory control and the controlled feature-memory baselines in aggregate ranking. Additional analyses quantify category-level target violations, practical inspection workload, normalized operating cost, controlled distribution shifts, human-review error, and component ablations. The reported risk quantities are held-out empirical operating characteristics rather than finite-sample conformal guarantees.