Research map / Label-efficient and robust learning / OOD and anomaly detection
Boosting Fine-Grained Visual Anomaly Detection with Coarse-Knowledge-Aware Adversarial Learning
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 CKAAD, a coarse-knowledge-aware adversarial feature-reconstruction method that uses a small set of coarsely labeled anomalous images plus a patch-level energy-based discriminator to suppress reconstruction of anomalies for fine-grained anomaly detection and localization.
Problem
Reconstruction-based unsupervised anomaly detection fails on fine-grained anomalies because auto-encoders also reconstruct anomalies well; existing weakly supervised methods use anomaly images poorly and ignore localization.
Approach
An energy-based discriminator trained with a small set of image-level anomaly labels aligns reconstructed feature distributions with normal ones; a patch-level version treats all patches of anomalous images as anomalous and is proven to still yield alignment despite noisy patch labels.
Key results
- On medical datasets with 5% labeled anomalies, CKAAD reaches 92.66 AUC on ISIC2018 and 99.79 AUC on Br35H, outperforming weakly supervised baselines like DevNet and DRA.
- On MVTec AD, CKAAD achieves 99.5 I-AUC, 98.4 P-AUC and 95.2 PRO, competitive with state-of-the-art methods like DMAD and RD++.
- Ablation on medical datasets shows patch-level alignment (CKAPatch) improves AUC over image-level CKAImg (e.g., 92.6 vs 91.4 on ISIC2018).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add 1% labeled anomalies and coarse-knowledge-aware adversarial learning | Recon · ISIC2018 | AUC: 87.51 -> 91.02 | Helped |
| Add 1% labeled anomalies and coarse-knowledge-aware adversarial learning | Recon · Chest X-ray | AUC: 85.59 -> 93.22 | Helped |
| Add 1% labeled anomalies and coarse-knowledge-aware adversarial learning | Recon · Br35H | AUC: 97.37 -> 98.89 | Helped |
| Add 1% labeled anomalies and coarse-knowledge-aware adversarial learning | Recon · OCT | AUC: 99.33 -> 99.88 | Helped |
| Increase labeled anomaly ratio from 1% to 5% | CKAAD · ISIC2018 | AUC: 91.02 -> 92.66 | Helped |
| Increase labeled anomaly ratio from 1% to 5% | CKAAD · Chest X-ray | AUC: 93.22 -> 96.57 | Helped |
| Replace normal-only reconstruction with maximizing reconstruction error on anomalous samples (ablation) | Recon · MVTec AD | PRO: 94.4 -> 86.2 | Hurt |
| Replace vanilla GAN with a coarse-knowledge-aware discriminator (ablation) | GAN · ISIC2018 | AUC: 91.0 -> 91.4 | Helped |
| Add patch-level alignment to image-level coarse-knowledge-aware alignment (ablation) | CKAImg · ISIC2018 | AUC: 91.4 -> 92.6 | Helped |
| Add patch-level alignment to image-level coarse-knowledge-aware alignment (ablation) | CKAImg · Chest X-ray | AUC: 96.0 -> 96.6 | Helped |
| Add patch-level alignment to image-level coarse-knowledge-aware alignment (ablation) | CKAImg · MVTec AD | P-AUC: 98.3 -> 98.4 | Helped |
| Add patch-level alignment to image-level coarse-knowledge-aware alignment (ablation) | CKAImg · MVTec AD | PRO: 95.1 -> 95.2 | Helped |
Surprising findings
- Naively using anomalies to enlarge reconstruction error (ReconSub) substantially hurts MVTec pixel-level localization (P-AUC 95.0 vs 98.2 for Recon).
- Pure GAN alignment with anomalies gives only marginal gains over the reconstruction baseline on medical datasets.
Limitations
- Industrial experiments rely on synthetic anomalies from elastic transformation rather than real anomalous images.
- Patch-level alignment assumes all patches of anomalous images are anomalous, which is noisy, and the theory assumes disjoint normal/anomaly distributions.
Experimental setup
Datasets: ISIC2018 Chest X-ray Br35H OCT MVTec AD Visa
Benchmarks: ISIC2018 Chest X-ray Br35H OCT MVTec AD Visa CIFAR10
Metrics: AUC best F1 ACC I-AUC P-AUC PRO
Baselines: GANomaly DifferNet Fastflow AE-FLOW ReContrast PReNet Deep SAD AA-BiGAN DRA DevNet PaDiM DRAEM
Builds on: GAN Deep SAD DevNet DRA PReNet
Closest papers by meaning
- PatchGuard: Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies (CVPR 2025 · similarity 0.89)
- Contamination-Resilient Anomaly Detection via Adversarial Learning on Partially-Observed Normal and Anomalous Data (ICML 2024 · similarity 0.86)
- RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection (CVPR 2024 · similarity 0.85)
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.85)
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.85)
- ADPretrain: Advancing Industrial Anomaly Detection via Anomaly Representation Pretraining (NeurIPS 2025 · similarity 0.84)
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.