Research map / Label-efficient and robust learning / OOD and anomaly detection
Exploring the Relationship Between Samples and Masks for Robust Defect Localization
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 one-stage GAN-based framework for texture defect localization that learns directly from unpaired synthetic defects and random masks, without modeling normality, using a scaled pattern loss, cycle-consistent auxiliary network, and dynamic masked correction.
Problem
Unsupervised defect localization in textures typically models normal sample distributions, which limits generalization under environmental changes such as illumination and hue shifts.
Approach
Trains a GAN generator to map defective inputs to masks using unpaired synthetic defects and random Perlin masks, stabilized by a scaled pattern loss (invisible pattern on normal images), a cycle-consistent auxiliary generator, and a dynamic correction that masks defect regions using the generator's own predictions; no ground truth or normality modeling is used.
Key results
- On MVTec AD texture classes, the method achieves 68.9% average F1-Score versus 66.0% for the best prior method (DRAEM), a 2.9-point gain.
- On DTD-Synthetic and Woven Fabric Textures, average F1 is 66.7% versus 64.0% for DRAEM (2.7-point gain).
- Under simulated illumination and hue changes, the method reports 59.7% F1 (general) and 49.7% F1 (hard) and claims 12.4% and 18.3% gains over the best baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed framework instead of the strongest baseline, DRAEM | Defect localization · MVTec AD texture classes | F1-Score: 2.9% | Helped |
| Use the proposed framework instead of the strongest baseline, DRAEM | Defect localization · DTD-Synthetic and Woven Fabric Textures | F1-Score: 2.7% | Helped |
| Use the proposed framework instead of the strongest baseline | Defect localization under environmental variation · MVTec AD texture classes, general setting | F1-Score: 12.4% | Helped |
| Use the proposed framework instead of the strongest baseline | Defect localization under environmental variation · MVTec AD texture classes, hard setting | F1-Score: 18.3% | Helped |
| Add scaled pattern loss to GAN (ablation) | GAN-only model · MVTec AD texture classes | F1-Score: 2.8 -> 37.2 | Helped |
| Add cycle-consistent auxiliary network to GAN with scaled pattern loss (ablation) | G.S · MVTec AD texture classes | F1-Score: 37.2 -> 35.4 | Hurt |
| Add dynamic mask to GAN with scaled pattern loss and cycle consistency (ablation) | G.S.C · MVTec AD texture classes | F1-Score: 35.4 -> 51.1 | Helped |
| Use progressive opacity instead of fixed opacity 0.1 (ablation) | G.S.C.M · MVTec AD texture classes | F1-Score: 15.9 -> 68.9 | Helped |
| Use progressive opacity instead of fixed opacity 0.5 (ablation) | G.S.C.M · MVTec AD texture classes | F1-Score: 40.4 -> 68.9 | Helped |
| Use progressive opacity instead of fixed opacity 0.9 (ablation) | G.S.C.M · MVTec AD texture classes | F1-Score: 30.9 -> 68.9 | Helped |
Surprising findings
- Adding the cycle-consistent auxiliary network alone (G.S.C) lowered performance versus G.S (35.4 vs 37.2 AVG), which the authors attribute to unrecovered defect regions.
Limitations
- Evaluated only on texture classes; object anomalies are explicitly excluded by design.
- Performance drops noticeably under unseen illumination and hue changes (e.g. 49.7% F1 in hard setting).
- Training is volatile; the cycle-consistent component alone degrades performance, requiring the dynamic correction mechanism.
Experimental setup
Datasets: MVTec AD texture classes DTD-Synthetic Woven Fabric Textures
Benchmarks: MVTec AD texture classes DTD-Synthetic Woven Fabric Textures
Metrics: F1-Score AUROC
Baselines: CFlow DRAEM PatchCore PyramidFlow PaDiM
Builds on: CycleGAN DRAEM
Closest papers by meaning
- Few-Shot Defect Image Generation via Defect-Aware Feature Manipulation (AAAI 2023 · similarity 0.88)
- DefectFill: Realistic Defect Generation with Inpainting Diffusion Model for Visual Inspection (CVPR 2025 · similarity 0.85)
- PatchGuard: Adversarially Robust Anomaly Detection and Localization through Vision Transformers and Pseudo Anomalies (CVPR 2025 · similarity 0.84)
- Wavelet and Prototype Augmented Query-based Transformer for Pixel-level Surface Defect Detection (CVPR 2025 · similarity 0.84)
- Boosting Fine-Grained Visual Anomaly Detection with Coarse-Knowledge-Aware Adversarial Learning (AAAI 2025 · similarity 0.84)
- MVREC: A General Few-shot Defect Classification Model Using Multi-View Region-Context (AAAI 2025 · similarity 0.83)
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.