Answers are grounded in this paper and cite their sources.
Industrial anomaly detection commonly relies on normal images because representative samples of diverse defects are difficult to collect. In unified multi-class settings, a shared model must also accommodate product-specific normal appearances while identifying local defects. Feature reconstruction localizes anomalies through discrepancies between observed and reconstructed features, but may also reconstruct anomalous information and weaken defect responses. Scalar cosine distances can also obscure distinct channel-wise mismatch patterns. We propose NRC-Dinomaly, extending Dinomaly with neighborhood prediction and relation-guided residual calibration. Center-excluded neighborhood reconstruction predicts intermediate features from surrounding positions and decodes them through an independent branch, reducing direct reliance on center features. RGRC uses channel-wise differences and agreement between observed and reconstructed features to learn bounded adjustments to cosine residuals. References are learned from normal images; local synthetic anomalies supervise calibration. Compared with Dinomaly, NRC-Dinomaly increases pixel-level average precision (P-AP) from 69.29% to 70.78% on MVTec AD, 53.23% to 55.70% on VisA, 59.53% to 59.89% on MPDD, and 70.10% to 73.38% on BTAD, with improvements in 29 of 36 product categories. These results indicate that neighborhood-based reconstruction references and relation-informed residual scoring help shared models distinguish defect pixels from normal backgrounds.
No takes yet. Share an insight, caveat, or question.
Deng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: