One-class learning project for anomaly detection using real industrial dataset
- The model is based on original CS-Flow model that has been modified.
- We tested our model against CS-Flow and Fastflow (current SOTAs) on our own benchmark datasets that are much larger than MVtech-AD (Camera lens, TCP boards)
- We were able to increase the inference speed for more than >2x and outperforming CS-Flow by more than >0.5% in AUROC and significant improvement in False Postitive Rate.
Dataset
Camera Lens | TCP board | |
---|---|---|
Train | 422 | 1432 |
Test | 802 | 2897 |
Results
Camera Lens (AUROC/FPR) | TCP board (AUROC/FPR) | Inf. speed (ms) | |
---|---|---|---|
Proposed | 99.5/7.5% | 99.9/26.9% | 36.5 |
CS-Flow | 98.7/55.3% | 99.7/68.0% | 92.8 |
ROC Curve
SMT | Camera Lens |
---|---|
This code is heavily borrowed from the CS-Flow implementation (https://github.com/marco-rudolph/cs-flow)