Answers are grounded in this paper and cite their sources.
Symmetry and repetitive patterns are common in industrial products, where small defects often disrupt local structural regularity. Fixed-resolution reconstruction can weaken this evidence, while high-resolution cropping changes visual context and anomaly-score distributions. We propose HiFRA-AD, a high-frequency residual-aware sparse refinement framework that couples local feature sensitivity with calibrated evidence aggregation. A fixed-coefficient token-grid residual operator enhances local spatial variation before reconstruction using a frozen DINOv2 encoder. At inference, the global anomaly map selects four peak-suppressed regions for re-evaluation by the shared model. Robust median/quantile alignment and a soft evidence gate convert the local responses into nonnegative residual updates. HiFRA-AD-S achieves P-AP/AUPRO scores of 71.6/95.5, 55.1/96.1, and 46.8/59.8 on MVTec AD, VisA, and MVTec AD 2, respectively. The MVTec AD 2 evaluation uses the official 2045-image TESTpriv partition and AU-PRO@0.30. In backbone-matched MVTec AD comparisons, P-AP improves from 67.5 to 71.6 with ViT-S and from 69.3 to 72.8 with ViT-B. Five paired ViT-S training runs yield a 4.10-point P-AP gain (95% CI [3.95, 4.25]) and a 1.10-point AUPRO gain (95% CI [0.98, 1.22]). With four local regions, ViT-S inference takes 20.8 ms per image on an RTX 4090, compared with 10.0 ms for the global baseline. A defect-subtype analysis records larger paired per-image AP gains for structural disruptions (+4.99 points) than for other anomalies (+2.68 points). These results support selective local refinement for fine-grained industrial anomaly localization.
No takes yet. Share an insight, caveat, or question.
Yangu Chen (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: