---
license: cc-by-nc-3.0
task_categories:
- image-segmentation
tags:
- vision
- document-processing
- binarization
- ocr-preprocessing
size_categories:
- 10K<n<100K
---

# Tzefa Binarization Dataset

Training dataset for the Tzefa document binarization model.

## Structure
```
images/    # 40,085 input RGB document images (.jpg)
masks/     # 40,085 binary ground-truth masks (.png)
```

Each image has a corresponding mask where:
- **Black (0)** = ink / text
- **White (255)** = paper / background

## Statistics
| Split | Images | Masks |
|-------|--------|-------|
| Unified | 40,085 | 40,085 |

Images are collected from multiple batches (Batch 1-14) and unified into a single training set.
Tiles are cropped at 640x640 during training with ImageNet normalization.

## Usage
```python
from datasets import load_dataset
dataset = load_dataset("WARAJA/Tzefa-Binarization-Dataset")
```

Or load images directly:
```python
from PIL import Image
img = Image.open("images/BATCH_10_copy_000003.jpg")
mask = Image.open("masks/BATCH_10_copy_000003.png")
```

## Related
- [Binarization Model (mit_b5)](https://huggingface.co/WARAJA/b5_model)
- [Binarization Demo](https://huggingface.co/spaces/WARAJA/Tzefa-Binarization)
- [Full Tzefa Pipeline](https://huggingface.co/spaces/WARAJA/Tzefa)