Research map / Label-efficient and robust learning / OOD and anomaly detection
AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios
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 AnomalyNCD, a self-supervised novel-class-discovery network that classifies multi-class industrial anomalies using anomaly-centered sub-images (MEBin), mask-guided ViT attention, and area-weighted region merging.
Problem
Multi-class anomaly classification in industrial scenarios, where anomalies are non-prominent and weakly semantic, so clustering-based methods lack anomaly-prior knowledge.
Approach
Reframes anomaly classification as novel class discovery: MEBin extracts stable main anomaly regions from any anomaly detector's map to crop anomaly-centered sub-images; mask-guided attention in the last ViT layers directs the [CLS] token to anomalous regions; pseudo labels are corrected with anomaly scores; and area-weighted merging assigns image-level classes.
Key results
- With zero-shot MuSc as the detector, AnomalyNCD improves over Anomaly Clustering on MVTec AD by 8.8% NMI, 9.5% ARI and 10.8% F1 (NMI 0.613, ARI 0.526, F1 0.712).
- On MTD, the method reaches NMI 0.268, ARI 0.228, F1 0.509 in the unsupervised setting, above Anomaly Clustering's 0.179/0.120/0.346.
- Mask-guided attention on the [CLS] token outperforms applying masks to all or patch tokens (NMI 0.613 vs 0.507 and 0.563) on MVTec AD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| combine MuSc with AnomalyNCD instead of using AC (Unsup.) | MVTec AD anomaly classification · MVTec AD | F1: 0.604 -> 0.712 | Helped |
| combine MuSc with AnomalyNCD instead of using AC (Unsup.) | MTD anomaly classification · MTD | F1: 0.346 -> 0.509 | Helped |
| combine CPR with AnomalyNCD instead of using AC (Semi-sup.) | MVTec AD anomaly classification · MVTec AD | F1: 0.652 -> 0.805 | Helped |
| combine PatchCore with AnomalyNCD instead of using AC (Semi-sup.) | MTD anomaly classification · MTD | F1: 0.490 -> 0.617 | Helped |
| use MEBin instead of fixed-threshold or Otsu binarization (ablation) | AnomalyNCD · MVTec AD | F1: 0.499 -> 0.712 | Helped |
| add mask-guided attention on the class token (ablation) | AnomalyNCD · MVTec AD | F1: 0.698 -> 0.712 | Helped |
| apply mask-guided attention to all tokens rather than only the class token (ablation) | AnomalyNCD · MVTec AD | F1: 0.712 -> 0.600 | Hurt |
| apply mask-guided attention to patch tokens rather than only the class token (ablation) | AnomalyNCD · MVTec AD | F1: 0.712 -> 0.686 | Hurt |
| use mask-guided attention in the last 9 layers instead of 1 layer (ablation) | AnomalyNCD · MVTec AD | ARI: 0.508 -> 0.526 | Helped |
| use mask-guided attention in the last 9 layers instead of 1 layer (ablation) | AnomalyNCD · MTD | F1: 0.420 -> 0.509 | Helped |
| use area-weighted region merging rather than average prediction (ablation) | AnomalyNCD · MVTec AD | F1: 0.709 -> 0.712 | Helped |
| use area-weighted region merging rather than average prediction (ablation) | AnomalyNCD · MTD | F1: 0.501 -> 0.509 | Helped |
Surprising findings
- Masking patch tokens or all tokens hurts performance relative to no mask attention (all tokens NMI 0.507 vs 0.598 without MGA), so the mask is applied only to the class token.
Limitations
- Performance is bounded by the quality of the upstream anomaly detection method.
- Evaluated only on two industrial datasets (MVTec AD and MTD) with combined anomaly classes removed.
Experimental setup
Datasets: MVTec AD MTD AeBAD-S
Benchmarks: MVTec AD MTD
Metrics: F1 NMI ARI AUPRO FPR FNR
Baselines: Anomaly Clustering UniFormaly IIC GATCluster SCAN UNO GCD SimGCD AMEND
Builds on: DINO GCD SimGCD Anomaly Clustering Otsu thresholding
Closest papers by meaning
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.89)
- ADPretrain: Advancing Industrial Anomaly Detection via Anomaly Representation Pretraining (NeurIPS 2025 · similarity 0.89)
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.88)
- Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.88)
- Normal-Abnormal Guided Generalist Anomaly Detection (NeurIPS 2025 · similarity 0.88)
- CNC: Cross-modal Normality Constraint for Unsupervised Multi-class Anomaly Detection (AAAI 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.