| # PatchCore Defect Detection — Metal Nut |
|
|
| ## Model Description |
|
|
| This is an anomaly detection model trained |
| on the MVTec AD metal nut category. It detects |
| surface defects in manufactured components including |
| scratches, bends, colour contamination, and |
| orientation errors — without ever being trained on |
| defective examples. |
|
|
| ## Performance |
|
|
| | Metric | Score | |
| |--------|-------| |
| | Image AUROC | 0.9976 | |
| | Pixel AUROC | 0.9868 | |
|
|
| ## Training Data |
|
|
| Trained on 220 defect-free images from the |
| [MVTec AD dataset](https://www.mvtec.com/company/research/datasets/mvtec-ad) |
| metal nut category. |
|
|
| ## Algorithm |
|
|
| PatchCore (Roth et al., CVPR 2022) using |
| WideResNet50 backbone with: |
| - Feature layers: layer2, layer3 |
| - Coreset sampling ratio: 0.1 |
| - Number of neighbours: 9 |
| - Calibrated threshold: 0.577 |
|
|
| ## Threshold Calibration |
|
|
| Threshold calibrated at 99th percentile of good |
| part scores: |
| - Good part pass rate: 95% (21/22) |
| - Defect detection rate: 96% (89/93) |
|
|
| ## Usage |
|
|
| This model is deployed as a Streamlit web |
| application. See the associated Space for the |
| live demo. |
|
|
| ## Citation |
|
|
| Roth, K., Pemula, L., Zepeda, J., Scholkopf, B., |
| Brox, T., & Gehler, P. (2022). Towards Total Recall |
| in Industrial Anomaly Detection. CVPR 2022. |
