mlsanityData to 14 Oct 2026

What Remains Normal? Clean Images Miss Useful Near-Defect Normal Patches for Anomaly Detection

arXiv 2608.23299 ↗First posted 24 Aug 2026cs.CVcited by 0

Not on the map. Why?

Abstract (arXiv)

Normal-only industrial anomaly detectors use patches from clean training images as normal references or reconstruction targets. This assumes that clean patches are sufficient for the normal regions encountered at test time. We test that assumption directly. On MVTec AD, admitting ground-truth-normal patches from real defect images to a DINOv2 memory candidate pool raises pixel average precision (P-AP) from 73.34 to 76.95 while keeping the encoder, test-time score, and number of stored references fixed. Patches within two patch cells of the annotated defect recover 94.70% of this gain. We then ask whether useful patches of this kind can be exposed using clean training images alone. BoundarySupport inserts a procedural synthetic defect to alter surrounding context, excludes every token intersecting the nominal insertion or a detected RGB change, and learns only from pixel-preserved neighboring patches. Across three paired seeds, the same principle improves P-AP in all six memory and reconstruction settings across MVTec, VisA, and Real-IAD. Matched controls identify the altered-context feature itself as the useful normal evidence: with synthetic input or selected positions fixed, altered-context features outperform their clean-view counterparts as both reconstruction targets and memory references. On MVTec memory, the final score change is also spatially selective, with larger reductions on normal patches next to defects than on mid-distance or far-normal patches in all 15 categories. Code is publicly available at https://github.com/jw-chae/boundary_support.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 20 of its 23 references that are in the map. First posted on arXiv 24 Aug 2026.

Builds on

20 of 23 references.

References · 20

  1. 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,974
  2. 2023Vision Transformers Need Registerscited by 871
  3. 2021Towards Total Recall in Industrial Anomaly Detectioncited by 776
  4. 2021CutPaste: Self-Supervised Learning for Anomaly Detection and Localizationcited by 531
  5. 2021DRAEM -- A discriminatively trained reconstruction embedding for surface anomaly detectioncited by 514
  6. 2021Intriguing Properties of Vision Transformerscited by 446
  7. 2022Anomaly Detection via Reverse Distillation from One-Class Embeddingcited by 421
  8. 2022SPot-the-Difference Self-Supervised Pre-training for Anomaly Detection and Segmentationcited by 400
  9. 2022DeSTSeg: Segmentation Guided Denoising Student-Teacher for Anomaly Detectioncited by 154
  10. 2024Real-IAD: A Real-World Multi-View Dataset for Benchmarking Versatile Industrial Anomaly Detectioncited by 139
  11. 2021Natural Synthetic Anomalies for Self-Supervised Anomaly Detection and Localizationcited by 116
  12. 2024Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detectioncited by 93
8 more
  1. 2024A Unified Anomaly Synthesis Strategy with Gradient Ascent for Industrial Anomaly Detection and Localizationcited by 74
  2. 2024SoftPatch: Unsupervised Anomaly Detection with Noisy Datacited by 55
  3. 2024MuSc: Zero-Shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Imagescited by 52
  4. 2025Exploring Intrinsic Normal Prototypes within a Single Image for Universal Anomaly Detectioncited by 48
  5. 2024Progressive Boundary Guided Anomaly Synthesis for Industrial Anomaly Detectioncited by 17
  6. 2024FUN-AD: Fully Unsupervised Learning for Anomaly Detection with Noisy Training Datacited by 8
  7. 2025INP-Former++: Advancing Universal Anomaly Detection via Intrinsic Normal Prototypes and Residual Learningcited by 7
  8. 2026ProCon: Projection-Consistency Memory for Training-Free Anomaly Detectioncited by 1

Built on by

Not cited yet.