README.md

---
license: apache-2.0
tags:
- image-classification
- index-cards
- library-archives
- timm
datasets:
- custom
metrics:
- accuracy
- f1
library_name: timm
pipeline_tag: image-classification
---

# Index Card Classifier

A fine-tuned image classifier for detecting library index cards vs other image types
(verso/back of cards, covers, blank pages).

## Model Details

- **Base Model**: efficientnet_b0
- **Task**: Binary classification (index_card vs other)
- **Training Data**: 70 images
- **Validation Data**: 30 images
- **Framework**: PyTorch + timm

## Training Data Distribution

```
{
  "cover": 1,
  "blank": 1,
  "needs_review": 1,
  "verso": 49,
  "index_card": 46,
  "other": 2
}
```

## Performance

| Metric | Value |
|--------|-------|
| Validation Accuracy | 96.7% |
| Validation F1 (index_card) | 0.966 |
| Validation F1 (other) | 0.968 |

### Classification Report

```
              precision    recall  f1-score   support

       other      1.000     0.938     0.968        16
  index_card      0.933     1.000     0.966        14

    accuracy                          0.967        30
   macro avg      0.967     0.969     0.967        30
weighted avg      0.969     0.967     0.967        30

```

### Confusion Matrix

```
[[15  1]
 [ 0 14]]
```

## Usage

```python
import timm
import torch
from huggingface_hub import hf_hub_download
from PIL import Image
from safetensors.torch import load_file
from torchvision import transforms

# Download and load model from Hub
weights_path = hf_hub_download(
    repo_id="davanstrien/nls-index-card-classifier",
    filename="classifier.safetensors"
)
model = timm.create_model('efficientnet_b0', pretrained=False, num_classes=2)
model.load_state_dict(load_file(weights_path))
model.eval()

# Preprocess
transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

# Inference
image = Image.open('card.jpg').convert('RGB')
input_tensor = transform(image).unsqueeze(0)

with torch.no_grad():
    output = model(input_tensor)
    probs = torch.softmax(output, dim=1)
    pred = output.argmax(1).item()
    confidence = probs[0, pred].item()

classes = ['other', 'index_card']
print(f"Prediction: {classes[pred]} ({confidence:.1%})")
```

## Training

Trained using frozen backbone with only classifier head fine-tuned.

```bash
python train_classifier.py --model efficientnet_b0 --epochs 20 --val-split 0.3
```

## Context

This model was developed for the National Library of Scotland to help process
digitized manuscript index cards from the Advocate's Library collection.

## License

Apache 2.0