Research map / Label-efficient and robust learning / OOD and anomaly detection
Distribution Prototype Diffusion Learning for Open-set Supervised 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 DPDL, which learns multiple Gaussian prototypes with a Schrödinger bridge to map normal features compactly while pushing anomalies away, plus a hyperspherical dispersion loss, for open-set supervised anomaly detection.
Problem
Open-set supervised anomaly detection (OSAD) with few observed anomalies: existing methods rely on pseudo-anomaly generation and ignore the distributional priors of diverse normal samples, yielding weak decision boundaries and poor generalization to unseen anomalies.
Approach
Learns multi-Gaussian prototypes as a latent space for normal samples, couples them with a Schrödinger bridge that diffusively transports normal features toward the prototypes while pushing anomalies away, and adds a von Mises-Fisher-based dispersion loss on hyperspherical features to increase inter-sample separation.
Key results
- Under the general setting with one training anomaly, DPDL improves over AHL by more than 8.3% AUC on AITEX, ELPV, and Mastcam.
- Under the hard setting, DPDL beats AHL by 0.5% to 8.9% AUC across datasets.
- With ten training anomalies, DPDL reports the best AUC on all nine general-setting datasets, e.g., 0.977 on MVTec AD and 0.937 on ELPV.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add distribution prototype learning (DPL) (ablation) | DPDL · Industrial anomaly datasets, general and hard settings | AUC | Helped |
| Add dispersion feature learning (DFL) (ablation) | DPDL · Open-set anomaly detection, general and hard settings | AUC | Helped |
| Remove the residual anomaly classifier Mr (ablation) | DPDL · AITEX, ELPV, and Mastcam, general and hard settings | AUC | Hurt |
| Remove the anomaly classifier Ma (ablation) | DPDL · AITEX, ELPV, and Mastcam, general and hard settings | AUC | Hurt |
| Remove the normal classifier Mn (ablation) | DPDL · AITEX, ELPV, and Mastcam, general and hard settings | AUC | Hurt |
| Increase prototype count C excessively (ablation) | DPDL · General setting, datasets with more categories | AUC | Hurt |
| Increase diffusion parameter ϵ (ablation) | DPDL · AITEX, Mastcam, and MVTec AD | AUC | Hurt |
| Increase dispersion concentration κ above 10 (ablation) | DPDL · General setting | AUC | Mixed |
| Increase dispersion loss weight λ (ablation) | DPDL · General setting | AUC | Hurt |
Limitations
- Performance declines as the number of prototypes or the loss weight lambda grows, suggesting sensitivity to hyperparameters.
- Medical datasets BrainMRI and HeadCT are small and single-class, so results there are only competitive rather than best.
Experimental setup
Datasets: MVTec AD Optical SDD AITEX ELPV Mastcam Hyper-Kvasir Brain-MRI HeadCT
Benchmarks: MVTec AD Optical SDD AITEX ELPV Mastcam Hyper-Kvasir Brain-MRI HeadCT Carpet Metal nut
Metrics: AUC
Baselines: DevNet FLOS SAOE MLEP DRA AHL
Builds on: DevNet DRA AHL LightSB I2-SB
Closest papers by meaning
- Anomaly Heterogeneity Learning for Open-set Supervised Anomaly Detection (CVPR 2024 · similarity 0.88)
- On Diffusion Modeling for Anomaly Detection (ICLR 2024 · similarity 0.87)
- A Diffusion-Based Framework for Multi-Class Anomaly Detection (AAAI 2024 · similarity 0.86)
- Out-of-Distribution Detection with a Single Unconditional Diffusion Model (NeurIPS 2024 · similarity 0.86)
- Correcting Deviations from Normality: A Reformulated Diffusion Model for Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.86)
- Vague Prototype-Oriented Diffusion Model for Multi-Class Anomaly Detection (ICML 2024 · similarity 0.85)
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.