---
license: apache-2.0
datasets:
- iapp/thai_handwriting_dataset
language:
- th
metrics:
- cer
base_model:
- microsoft/trocr-base-handwritten
tags:
- ocr
- handwriting
- thai
- trocr
- vision
- transformer
---

# Thai Handwritten OCR (TrOCR)

A Thai Handwritten OCR model fine-tuned from Microsoft TrOCR for recognizing Thai handwritten text.

## Model Details

### Model Description

This model is developed to convert Thai handwritten images into text using the TrOCR architecture, which combines Vision Transformer (ViT) for image processing and Transformer Decoder for text generation.

- **Developed by:** Warit Sirikosityanggoon
- **Model type:** Vision Encoder-Decoder (TrOCR)
- **Language(s):** Thai (th)
- **License:** Apache 2.0
- **Finetuned from:** [microsoft/trocr-base-handwritten](https://huggingface.co/microsoft/trocr-base-handwritten)

### Model Sources

- **Repository:** [waritkan/Thai-Hand-Written-TrOCR-Webapp](https://github.com/waritkan/Thai-Hand-Written-TrOCR-Webapp)

## Uses

### Direct Use

This model can be used directly for converting Thai handwritten images into text. Suitable for:
- Converting Thai handwritten documents
- Real-time handwriting recognition systems
- Digitizing handwritten notes

### Out-of-Scope Use

- Not suitable for languages other than Thai
- May not perform well on extremely difficult handwriting or low-quality images

## Training Details

### Training Data

Trained on [iapp/thai_handwriting_dataset](https://huggingface.co/datasets/iapp/thai_handwriting_dataset), which contains Thai handwritten images paired with their corresponding text labels.

### Tokenizer

Uses **SentencePiece with Unigram algorithm** instead of Dictionary-based Word Segmentation because:
- Handles Out-of-Vocabulary words effectively
- Supports misspelled or incomplete words from handwriting
- No pre-tokenization required

**Tokenizer Configuration:**
- Vocab Size: 30,000
- Character Coverage: 0.9995
- Algorithm: Unigram

### Training Hyperparameters

| Parameter | Value |
|-----------|-------|
| Epochs | 250 |
| Batch Size | 16 |
| Learning Rate | 1e-5 |
| Optimizer | AdamW |
| Training Regime | fp16 mixed precision |

### Training Infrastructure

- **Hardware:** NVIDIA GPU (HPC Cluster)
- **Framework:** PyTorch + Hugging Face Transformers

## Evaluation

### Metrics

| Metric | Value |
|--------|-------|
| **CER (Character Error Rate)** | **0.488%** |

### How to Evaluate

```python
import editdistance

def calculate_cer(pred, label):
    """Character Error Rate (lower is better)"""
    if len(label) == 0:
        return 1.0 if len(pred) > 0 else 0.0
    distance = editdistance.eval(pred, label)
    return distance / len(label)
```

## How to Get Started with the Model

### Installation

```bash
pip install transformers torch sentencepiece pillow
```

### Usage

```python
import torch
from PIL import Image
import sentencepiece as spm
from transformers import VisionEncoderDecoderModel, ViTImageProcessor

# Load model
model = VisionEncoderDecoderModel.from_pretrained('microsoft/trocr-base-handwritten')
image_processor = ViTImageProcessor.from_pretrained('microsoft/trocr-base-handwritten')

# Load Thai tokenizer
sp = spm.SentencePieceProcessor()
sp.Load('thai_sp_30000.model')

# Load trained weights
checkpoint = torch.load('best_model.pt', map_location='cpu')
model.decoder.resize_token_embeddings(sp.GetPieceSize())
model.load_state_dict(checkpoint['model_state_dict'], strict=False)
model.eval()

# Inference
image = Image.open('handwriting.jpg').convert('RGB')
pixel_values = image_processor(image, return_tensors='pt').pixel_values

with torch.no_grad():
    generated_ids = model.generate(
        pixel_values,
        max_length=128,
        num_beams=4,
    )

# Decode
ids = generated_ids[0].tolist()
text = sp.DecodeIds(ids)
print(text)
```

## Model Architecture

```
Input Image
    |
    v
Vision Transformer (ViT) Encoder
    |
    v
Cross-Attention
    |
    v
Transformer Decoder
    |
    v
SentencePiece Tokenizer (Unigram)
    |
    v
Thai Text Output
```

## Limitations

- Performance depends on image quality and handwriting clarity
- May not perform well on handwriting styles significantly different from training data
- Supports Thai language only

## Citation

```bibtex
@misc{thai-handwritten-trocr,
  author = {Warit Sirikosityanggoon},
  title = {Thai Handwritten OCR using TrOCR},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://github.com/waritkan/Thai-Hand-Written-TrOCR-Webapp}}
}
```

## Acknowledgements

- [Microsoft TrOCR](https://huggingface.co/microsoft/trocr-base-handwritten) for Pretrained Model
- [iApp Technology](https://huggingface.co/datasets/iapp/thai_handwriting_dataset) for Thai Handwriting Dataset
- [SentencePiece](https://github.com/google/sentencepiece) for Tokenizer

## Model Card Contact

- **Author:** Warit Sirikosityanggoon
- **GitHub:** [waritkan/Thai-Hand-Written-TrOCR-Webapp](https://github.com/waritkan/Thai-Hand-Written-TrOCR-Webapp)