Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
tags:
|
| 3 |
+
- anomaly-detection
|
| 4 |
+
- computer-vision
|
| 5 |
+
- patchcore
|
| 6 |
+
- defect-detection
|
| 7 |
+
- industrial
|
| 8 |
+
datasets:
|
| 9 |
+
- mvtec-ad
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# DefectVision — PatchCore Anomaly Detection (bottle)
|
| 13 |
+
|
| 14 |
+
PatchCore implemented from scratch with PyTorch on MVTec AD.
|
| 15 |
+
|
| 16 |
+
## Why PatchCore over YOLOv8?
|
| 17 |
+
| Model | Approach | Key Metric |
|
| 18 |
+
|-------|----------|------------|
|
| 19 |
+
| YOLOv8n | Supervised | mAP50 = 0.09 |
|
| 20 |
+
| PatchCore | Unsupervised | AUROC = 0.9976 |
|
| 21 |
+
|
| 22 |
+
YOLOv8 struggles with MVTec AD because defect images are scarce.
|
| 23 |
+
PatchCore trains on normal images only and generalizes to unseen defects.
|
| 24 |
+
|
| 25 |
+
## Architecture
|
| 26 |
+
- **Backbone**: WideResNet50 (ImageNet pretrained)
|
| 27 |
+
- **Layers**: layer2 + layer3 → 1536-dim patch features
|
| 28 |
+
- **Memory Bank**: coreset 10% of train patches
|
| 29 |
+
- **Scoring**: max nearest-neighbor distance (k=9)
|
| 30 |
+
|
| 31 |
+
## Metrics — bottle
|
| 32 |
+
- **AUROC** : 0.9976
|
| 33 |
+
- **Best F1**: 0.9920
|
| 34 |
+
|
| 35 |
+
## Usage
|
| 36 |
+
```python
|
| 37 |
+
import torch
|
| 38 |
+
from huggingface_hub import hf_hub_download
|
| 39 |
+
|
| 40 |
+
path = hf_hub_download(repo_id='Chasston/defect-vision-patchcore-bottle', filename='memory_bank.pt')
|
| 41 |
+
data = torch.load(path)
|
| 42 |
+
memory_bank = data['memory_bank']
|
| 43 |
+
```
|
| 44 |
+
|
| 45 |
+
## Author
|
| 46 |
+
[Chasston](https://huggingface.co/Chasston)
|