Vial PatchCore — Industrial Anomaly Detection
A PatchCore-based industrial anomaly detection model trained for the Vial category of the MVTec AD 2 dataset.
The model is designed to distinguish anomalous vial images from normal vial images and provides an anomaly map for visualizing regions that contribute to the detected anomaly.
🚀 Live Demo
Try the deployed model through the interactive Gradio application:
Industrial Vial Defect Detection — Hugging Face Space
Upload a vial image and receive:
- Normal / Defective prediction
- Anomaly score
- Anomaly heatmap
- Visual localization of suspicious regions
🧠 Model Overview
Architecture
PatchCore
PatchCore is an industrial anomaly detection approach that represents image patches using deep visual features and compares them against a memory bank representing normal samples.
The model is trained to learn the visual characteristics of normal vial images. During inference, deviations from the learned normal representation produce higher anomaly scores.
Model Details
| Property | Value |
|---|---|
| Architecture | PatchCore |
| Framework | PyTorch |
| Library | Anomalib |
| Dataset | MVTec AD 2 |
| Category | Vial |
| Input Resolution | 256 × 256 |
| Task | Industrial Anomaly Detection |
| Model File | model.ckpt |
📊 Evaluation
The trained Vial model was evaluated on the MVTec AD 2 public test set.
| Metric | Score |
|---|---|
| Image AUROC | 0.7578 |
| Image F1 Score | 0.5315 |
| Pixel AUROC | 0.9142 |
| Pixel F1 Score | 0.1197 |
Interpretation
Image AUROC — 0.7578
The model provides useful separation between normal and anomalous vial images at the image level.
Image F1 Score — 0.5315
The F1 score reflects the balance between precision and recall under the evaluation threshold used during the original evaluation.
Pixel AUROC — 0.9142
The high pixel-level AUROC indicates that the model can effectively rank anomalous regions relative to normal regions.
Pixel F1 Score — 0.1197
The lower pixel-level F1 indicates that precise defect segmentation remains challenging, even though the anomaly map provides useful localization information.
🎯 Deployment Threshold
During development, the default image-level threshold resulted in a relatively high number of false negatives for the Vial model.
An additional threshold analysis was performed using the MVTec AD 2 public test set.
The experimental deployment threshold selected for the interactive demo is:
0.09