Research map / Label-efficient and robust learning / OOD and anomaly detection
Generating and Reweighting Dense Contrastive Patterns for Unsupervised Anomaly Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GRAD, a prior-less unsupervised anomaly detector that generates local-structure contrastive patches with a receptive-field-limited diffusion model (PatchDiff) and trains a lightweight FCN patch detector with self-supervised density-based reweighting.
Problem
Unsupervised industrial anomaly detection often depends on ImageNet-pretrained feature extractors or ad-hoc anomaly simulation, which limits adaptability to diverse anomaly types such as logical anomalies in MVTec LOCO.
Approach
Generate contrastive images with a diffusion model whose receptive field is limited and attention removed so only local structures are kept, then train a lightweight patch-level detector whose unlabeled, long-tailed generated patches are reweighted by feature density.
Key results
- On MVTec LOCO, GRAD reaches 87.5% average image-level AU-ROC, versus 83.7% for DADF and 86.8% for SINBAD among the compared methods.
- On MVTec AD, GRAD achieves 98.7% image-level AU-ROC (best listed), while its pixel-level AU-ROC of 96.8% trails DRAEM's 97.3%.
- GRAD runs at 0.799 ms latency (1251.6 FPS) on a Tesla V100 on MVTec LOCO, reported as more than 16 times faster than GCAD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GRAD instead of DRAEM | MVTec AD anomaly detection · MVTec AD | image-level AU-ROC: 98.0 -> 98.7 | Helped |
| Use GRAD instead of DRAEM | MVTec AD anomaly localization · MVTec AD | pixel-level AU-ROC: 97.3 -> 96.8 | Hurt |
| Use GRAD instead of SINBAD | MVTec LOCO anomaly detection · MVTec LOCO | image-level AU-ROC: 86.8 -> 87.5 | Helped |
| Use GRAD instead of GCAD | MVTec LOCO inference · MVTec LOCO | latency (ms): 12.9 -> 0.799 | Helped |
| Add feature and gradient regularization (ablation) | GRAD Level-34 detector · MVTec LOCO | AUROC: +3.4 | Helped |
| Add feature and gradient regularization (ablation) | GRAD Level-68 detector · MVTec LOCO | AUROC: +3.1 | Helped |
| Add feature and gradient regularization (ablation) | GRAD Level-136 detector · MVTec LOCO | AUROC: +0.9 | Helped |
| Add noisy reweighting to the regularized detector (ablation) | GRAD Level-34 detector · MVTec LOCO | AUROC: +0.9 | Helped |
| Add noisy reweighting to the regularized detector (ablation) | GRAD Level-68 detector · MVTec LOCO | AUROC: +1.6 | Helped |
| Add noisy reweighting to the regularized detector (ablation) | GRAD Level-136 detector · MVTec LOCO | AUROC: +6.9 | Helped |
| Add long-tail reweighting to the detector with regularization and noisy reweighting (ablation) | GRAD Level-34 detector · MVTec LOCO | AUROC: +2.7 | Helped |
| Add long-tail reweighting to the detector with regularization and noisy reweighting (ablation) | GRAD Level-68 detector · MVTec LOCO | AUROC: +2.9 | Helped |
Surprising findings
- Generated patches from PatchDiff with a 17x17 receptive field expose logical anomalies such as a missing or extra pushpin, which boosts pushpins AU-ROC to 99.7%.
- The gradient-based and feature-based regularizers and reweighting each add accuracy, with the long-tail reweighting giving +2.7 to +3.0 points on LOCO levels.
Limitations
- Image-level gains on MVTec LOCO and AD depend on per-subdataset training of PatchDiff and detectors, requiring multiple levels of models per category.
- Pixel-level localization on MVTec AD is not the best (96.8 vs DRAEM 97.3).
- Level-136 detectors perform poorly on structural anomalies due to 64x64 input resolution.
Experimental setup
Datasets: MVTec AD MVTec LOCO
Benchmarks: MVTec AD MVTec LOCO
Metrics: image-level AU-ROC pixel-level AU-ROC inference latency (ms) FPS
Baselines: IGD PSVDD FCDD CutPaste NSA DRAEM DSR SPADE PaDiM S-T PatchCore GCAD
Builds on: DDPM denoising autoencoder CutPaste DRAEM
Closest papers by meaning
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.89)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.88)
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.88)
- Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.88)
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.88)
- Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.