---
license: cc-by-nc-3.0
pipeline_tag: image-to-image
tags:
- vision
- document-processing
- binarization
- segmentation
---

# Tzefa Binarization Model (mit_b5 HighResMAnet)

Custom-trained document binarization model for the Tzefa OCR pipeline.

## Architecture
- **Encoder:** MiT-B5 (Mix Transformer)
- **Decoder:** MAnet with custom High-Resolution Stem + Fusion Head
- **Framework:** segmentation-models-pytorch
- **Input:** RGB image tiles (640x640)
- **Output:** Binary mask (ink=0, paper=255)

## Usage
```python
from huggingface_hub import hf_hub_download
import torch

# Download weights
ckpt_path = hf_hub_download("WARAJA/b5_model", "b5_model.pth")

# Load model (see Tzefa Binarization Space for full architecture code)
checkpoint = torch.load(ckpt_path, map_location="cpu")
```

## Related
- [Binarization Demo](https://huggingface.co/spaces/WARAJA/Tzefa-Binarization)
- [Full Tzefa Pipeline](https://huggingface.co/spaces/WARAJA/Tzefa)
- [Binarization Dataset](https://huggingface.co/datasets/WARAJA/Tzefa-Binarization-Dataset)