Research map / Detection and segmentation / Open-set and few-shot detection
Few Shot Part Segmentation Reveals Compositional Logic for Industrial 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 PSAD, which uses a few-shot part segmentation model (histogram matching and entropy losses on unlabeled images) to build three memory banks for logical and structural anomaly detection on MVTec LOCO AD.
Problem
Detecting logical anomalies in industrial images (wrong quantity, arrangement, or composition of components) requires accurately segmenting visually similar components, but pixel-level annotation is costly and existing few-shot and unsupervised part segmentation methods fail on such images.
Approach
Train a part segmentation model with few labeled images plus unlabeled images sharing logical constraints, using a histogram matching loss and entropy loss, then build class-histogram, component-composition and patch-level memory banks whose anomaly scores are adaptively rescaled using training data and summed.
Key results
- PSAD reaches 98.1% average AUROC for logical anomaly detection on MVTec LOCO AD versus 89.6% for the best competing method (SLSG).
- PSAD achieves 91.6% average AUROC for structural anomaly detection, the best among compared methods (SLSG 91.4%).
- Adaptive scaling raises LA AUROC from 96.8% to 98.1% when combining the three memory banks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PSAD instead of SLSG | MVTec LOCO AD · Average (LA) | AUROC: 89.6 -> 98.1 | Helped |
| Use PSAD instead of SLSG | MVTec LOCO AD · Average (SA) | AUROC: 91.4 -> 91.6 | Helped |
| Use PSAD instead of SLSG | MVTec LOCO AD · Average | AUROC: 90.3 -> 94.9 | Helped |
| Add entropy loss (L_H) to supervised losses (ablation) | PSAD segmentation · MVTec LOCO AD, Average (LA) | AUROC: 95.9 -> 96.3 | Helped |
| Add entropy loss (L_H) to supervised losses (ablation) | PSAD segmentation · MVTec LOCO AD, Average (SA) | AUROC: 89.6 -> 90.0 | Helped |
| Add histogram matching loss (L_hist) to supervised and entropy losses (ablation) | PSAD segmentation · MVTec LOCO AD, Average (LA) | AUROC: 96.3 -> 98.1 | Helped |
| Add histogram matching loss (L_hist) to supervised and entropy losses (ablation) | PSAD segmentation · MVTec LOCO AD, Average (SA) | AUROC: 90.0 -> 91.6 | Helped |
| Add component composition and patch representation banks to the class histogram bank (ablation) | PSAD · MVTec LOCO AD, Average (LA) | AUROC: 94.2 -> 96.8 | Helped |
| Add component composition and patch representation banks to the class histogram bank (ablation) | PSAD · MVTec LOCO AD, Average (SA) | AUROC: 71.1 -> 87.6 | Helped |
| Apply adaptive scaling when combining all three memory banks (ablation) | PSAD · MVTec LOCO AD, Average (LA) | AUROC: 96.8 -> 98.1 | Helped |
| Apply adaptive scaling when combining all three memory banks (ablation) | PSAD · MVTec LOCO AD, Average (SA) | AUROC: 87.6 -> 91.6 | Helped |
| Reduce normal training images from 100% to 50% (ablation) | PSAD · MVTec LOCO AD, LA detection | AUROC: 97.4 -> 97.1 | Hurt |
Surprising findings
- Patch-level memory bank alone scores only 73.9% LA AUROC, and naively summing unscaled scores lowers SA performance.
Limitations
- Few-shot segmentation models such as VAT, RePRI and SegGPT perform poorly on components with similar textures, and the method depends on segmentation quality.
- Evaluation is limited to the single MVTec LOCO AD benchmark with five categories.
Experimental setup
Datasets: MVTec LOCO AD
Benchmarks: MVTec LOCO AD
Metrics: AUROC
Baselines: PatchCore RD4AD DRAEM ST AST GCAD SINBAD ComAD SLSG SCOPS Part-Assembly SegGPT
Builds on: PatchCore RePRI SCOPS Wide ResNet
Closest papers by meaning
- MuSc: Zero-Shot Industrial Anomaly Classification and Segmentation with Mutual Scoring of the Unlabeled Images (ICLR 2024 · similarity 0.87)
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.87)
- 3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised Anomaly Detection (AAAI 2025 · similarity 0.86)
- LogicAD: Explainable Anomaly Detection via VLM-based Text Feature Extraction (AAAI 2025 · similarity 0.86)
- Towards Training-free Anomaly Detection with Vision and Language Foundation Models (CVPR 2025 · similarity 0.86)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.86)
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.