DOI: 10.28979/jarnas.1957308 ISSN: 2757-5195

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

İsmail Hakkı Kınalıoğlu
Show PDF Cite
Automated visual quality inspection often operates with few normal referenceimages but still requires an explicit policy for asymmetric operational errors. This paperpresents Risk-Calibrated Few-Shot Industrial Anomaly Detection Plus (RC-FS-IAD+), afew-normal-support, labeled-calibration-assisted framework that maps anomaly scores toautomatic pass, manual review, and automatic rejection. The few-shot designation refers tothe normal support set used to construct the anomaly representation; the decision stage issemi-supervised and uses a separate labeled calibration set. To avoid score-fitting/thresholddata reuse, the main Visual Anomaly (VisA) evaluation uses stratified nested calibration withdisjoint score-fusion and threshold subsets. At k=4, RC-FS-IAD+ achieves Area Under theReceiver Operating Characteristic Curve (AUROC) 0.848, Area Under the Precision-RecallCurve (AUPR) 0.579, F1 0.584, mean missed-defect rate 4.17%, and manual-inspection rate45.34%; 65.0% of category-seed trials meet the empirical 5% missed-defect target. Acrossk ∈ {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 slightlyhigher 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 baselinesin aggregate ranking. Additional analyses quantify category-level target violations, practicalinspection workload, normalized operating cost, controlled distribution shifts, human-reviewerror, and component ablations. The reported risk quantities are held-out empirical operatingcharacteristics rather than finite-sample conformal guarantees.

More from our Archive