Papers › Anomaly Detection Using Normalizing Flow-Based Density Estimation and Synthetic Defect...
Anomaly Detection Using Normalizing Flow-Based Density Estimation and Synthetic Defect Classification
Seungmi Oh
We propose a novel deep learning-based anomaly detection (AD) system that combines a pixelwise classification network with conditional normalizing flow (CNF) networks by sharing feature extractors. We trained the pixelwise classification network using synthetic abnormal data to fine-tune a pretrained feature extractor of the CNF networks, thereby learning the discriminative features of the in-domain data. After that, we trained the CNF networks using normal data with the fine-tuned feature extractor to estimate the density of normal data. During inference, we detected anomalies by calculating the weighted average of the anomaly scores from the pixelwise classification and CNF networks. Because the proposed system not only has learned the properties of in-domain data but also aggregated the anomaly scores of the classification and CNF networks, it showed significantly improved performance compared to existing methods in experiments using the MvTecAD and BTAD datasets. Moreover, the proposed system does not increase computations intensively since the classification and the density estimation systems share feature extractors.
Code
Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.
Code Syntology ran Syntology
Not run by Syntology. Nothing on this page verifies that the listed code works.
Tasks
Results from the paper archive 2025-07-28
| Task | Dataset | Model | Metric | Value | Rank at snapshot | Leaderboard | Report |
|---|---|---|---|---|---|---|---|
| Anomaly Detection | BTAD | AD-CLSCNFs | Detection AUROC | 95.93 | #4 of 15 | Archive leaderboard | report |
| Anomaly Detection | BTAD | AD-CLSCNFs | Segmentation AP | 55.86 | #4 of 15 | Archive leaderboard | report |
| Anomaly Detection | BTAD | AD-CLSCNFs | Segmentation AUPRO | 72.77 | #4 of 15 | Archive leaderboard | report |
| Anomaly Detection | BTAD | AD-CLSCNFs | Segmentation AUROC | 97.13 | #4 of 15 | Archive leaderboard | report |
| Anomaly Detection | MVTec AD | AD-CLSCNFs | Detection AUROC | 98.85 | #49 of 148 | Archive leaderboard | report |
| Anomaly Detection | MVTec AD | AD-CLSCNFs | Segmentation AP | 74.32 | #49 of 148 | Archive leaderboard | report |
| Anomaly Detection | MVTec AD | AD-CLSCNFs | Segmentation AUPRO | 96.01 | #49 of 148 | Archive leaderboard | report |
| Anomaly Detection | MVTec AD | AD-CLSCNFs | Segmentation AUROC | 98.74 | #49 of 148 | Archive leaderboard | report |
Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.
Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections