Research map / Detection and segmentation / Open-set and few-shot detection
Supervised Anomaly Detection for Complex Industrial Images
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.
Introduces the Valeo Anomaly Dataset (VAD) of real industrial images with defects and proposes SegAD, which feeds segment-wise statistics of anomaly maps (plus an optional supervised classifier score) into a Boosted Random Forest for supervised anomaly detection.
Problem
Anomaly detection in industrial visual inspection is limited by public datasets that lack real defect images, and existing supervised anomaly detectors overfit to seen defects or need pixel-level labels, so they perform poorly on complex real-world objects.
Approach
SegAD computes per-segment statistics (99.5% quantile, skewness, kurtosis, mean) from anomaly maps of one-class detectors over a segmentation map, and trains a Boosted Random Forest on these features plus an optional supervised classifier score, requiring only image-level labels.
Key results
- SegAD (All AD + Ours) reaches 90.4 Cl. AUROC on VAD one-class benchmark vs 89.1 for EfficientAD alone.
- AllAD+WRN+Ours reaches 96.5 Cl. AUROC and 16.4 FPR@95TPR on VAD high-shot supervised benchmark, vs 95.0 for WRN alone.
- On VisA supervised benchmark, EfficientAD + Ours reaches 98.3 Cl. AUROC vs 98.0 for EfficientAD alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Combine PatchCore, FastFlow, and RD4AD anomaly maps with SegAD | SegAD · VAD one-class benchmark | Cl. AUROC: 89.1±0.8 -> 90.4±0.5 | Helped |
| Combine all anomaly detectors and WRN with SegAD | SegAD · VAD high-shot supervised benchmark | Cl. AUROC: 95.0±0.6 -> 96.5±0.3 | Helped |
| Combine all anomaly detectors with SegAD | SegAD · VAD low-shot supervised benchmark | Cl. AUROC: 78.7±1.8 -> 92.7±0.5 | Helped |
| Apply SegAD to EfficientAD anomaly maps | EfficientAD · VisA supervised benchmark | Cl. AUROC: 98.0 -> 98.3 | Helped |
| Use all anomaly detector maps with SegAD | SegAD · VisA supervised benchmark | Cl. AUROC: 98.0 -> 98.4 | Helped |
| Use proposed features and BRF with one segment (ablation) | SegAD · VAD high-shot supervised benchmark | Cl. AUROC: 87.3 -> 88.6 | Helped |
| Compute maximum anomaly-map value per segment (ablation) | SegAD · VAD high-shot supervised benchmark | Cl. AUROC: 87.3 -> 89.4 | Helped |
| Replace BRF with Boosted Tree (ablation) | SegAD · VAD high-shot supervised benchmark | Cl. AUROC: 90.1 -> 88.4 | Hurt |
| Replace BRF with Random Forest (ablation) | SegAD · VAD high-shot supervised benchmark | Cl. AUROC: 90.1 -> 89.8 | Hurt |
| Use SegAD with EfficientAD anomaly maps and generated defects | EfficientAD · VAD one-class benchmark | Cl. AUROC: 89.1±0.8 -> 85.6±1.6 | Hurt |
Surprising findings
- Adding SegAD to EfficientAD on the VAD one-class benchmark lowers the score to 85.6 AUROC with FPR 100, because synthetic defects differ from real defect anomaly maps.
- Supervised classifiers like WRN and DevNet/DRA perform poorly on unseen defect types, so supervised scores alone cannot detect them.
Limitations
- Requires segmentation maps; static maps work only for aligned objects like VAD, and unaligned objects need SAM or a trained segmenter.
- Gains are smaller for less complex objects (VisA) than for VAD.
- Anomaly map discrepancies between training and test data may be hard to mitigate.
Experimental setup
Datasets: VAD VisA MVTec AD
Benchmarks: VAD VisA
Metrics: Classification AUROC FPR@95TPR
Baselines: PatchCore FastFlow RD4AD EfficientAD DevNet DRA Wide ResNet50 classifier
Builds on: PatchCore FastFlow RD4AD EfficientAD XGBoost
Closest papers by meaning
- Real-IAD: A Real-World Multi-View Dataset for Benchmarking Versatile Industrial Anomaly Detection (CVPR 2024 · similarity 0.90)
- 3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised Anomaly Detection (AAAI 2025 · similarity 0.88)
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.88)
- Towards Training-free Anomaly Detection with Vision and Language Foundation Models (CVPR 2025 · similarity 0.87)
- CableInspect-AD: An Expert-Annotated Anomaly Detection Dataset (NeurIPS 2024 · similarity 0.87)
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (CVPR 2025 · similarity 0.87)
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.