Research map / Detection and segmentation / Open-set and few-shot detection
A Comprehensive Augmentation Framework for 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 a comprehensive anomaly simulation framework (transparent, opaque, near-distribution and rotation anomalies applied selectively per class) plus a split training strategy for reconstruction-based anomaly detection, reporting state-of-the-art localization AP on MVTec AD.
Problem
Data augmentation for simulating anomalies in reconstruction-based surface anomaly detection often replicates biased real-world anomaly distributions and ignores that normality standards differ across classes; two-stage reconstruction/localization training also suffers from overfitting and a reconst
Approach
Categorize simulated anomalies into transparent, opaque, near-distribution (NDAA) and rotation types, select augmentation combinations per class based on human judgment of what counts as anomalous, and split the training data by index parity so the discriminative network sees reconstructions of the same quality as at inference.
Key results
- Anomaly localization on MVTec AD: 98.0% AUROC and 70.9% AP, surpassing DRAEM (97.3/68.4), NSA (95.0/58.8) and PatchCore (97.6/56.9) on average.
- Image-level detection on MVTec AD: 98.3% AUROC and 99.3% AP, slightly below PatchCore (98.6/99.5).
- Removing rotation augmentation without the split strategy causes overfitting; the split strategy restores downstream performance (ablation averages up to 98.0/70.9 with the full method).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the comprehensive anomaly simulation framework and split training strategy instead of DRAEM | Reconstructive anomaly detection · MVTec AD anomaly localization | AP: 68.4 -> 70.9 | Helped |
| Use the comprehensive anomaly simulation framework and split training strategy instead of PatchCore | Anomaly detection · MVTec AD image-level anomaly detection | AUROC: 98.6 -> 98.3 | Hurt |
| Use the comprehensive anomaly simulation framework and split training strategy instead of PatchCore | Anomaly detection · MVTec AD image-level anomaly detection | AP: 99.5 -> 99.3 | Hurt |
| Use the comprehensive anomaly simulation framework and split training strategy instead of PatchCore | Anomaly detection · MVTec AD anomaly localization | AP: 56.9 -> 70.9 | Helped |
| Use our anomaly simulation methods with the reconstructive network | Anomaly localization model · Simulated anomalies localization | AP: 81.4 -> 82.7 | Helped |
| Remove rotation augmentation from DRAEM (ablation) | DRAEM · MVTec AD anomaly localization | AUROC: 97.3 -> 94.8 | Hurt |
| Replace DRAEM simulation with our simulation framework in the DRAEM architecture (ablation) | DRAEM architecture · MVTec AD anomaly localization | AUROC: 97.3 -> 95.4 | Hurt |
| Replace our simulation framework with DRAEM simulation in our architecture (ablation) | Our architecture · MVTec AD anomaly localization | AP: 70.9 -> 66.9 | Hurt |
| Use split training with our simulation framework rather than the DRAEM simulation in our architecture (ablation) | Our architecture · MVTec AD anomaly localization | AP: 66.9 -> 70.9 | Helped |
Surprising findings
- Image-level detection is slightly below PatchCore on average, despite the localization gains.
- Directly combining the DRAEM architecture with the proposed simulation methods performs poorly, attributed to overfitting without rotation augmentation and split training.
Limitations
- Per-class augmentation selection relies on manual human judgment, which is not automated.
- Underperforms PatchCore and DRAEM on several classes such as cable and capsule at image level; attributed to limited anomaly categories and inaccurate labels.
- Simulated dataset evaluation cannot guarantee real-world generalization.
Experimental setup
Datasets: MVTec AD
Benchmarks: MVTec AD simulated anomaly dataset
Metrics: AUROC AP
Baselines: DRAEM NSA PatchCore
Builds on: DRAEM NSA Cutpaste anomalib
Closest papers by meaning
- RealNet: A Feature Selection Network with Realistic Synthetic Anomaly for Anomaly Detection (CVPR 2024 · similarity 0.87)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.86)
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.86)
- Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.86)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · 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.