GraDeT-HTR: A Resource-Efficient Bengali Handwritten Text Recognition System utilizing Grapheme-based Tokenizer and Decoder-only Transformer

📖 Introduction

This is the official GitHub repository for GraDeT-HTR: A Resource-Efficient Bengali Handwritten Text Recognition System utilizing Grapheme-based Tokenizer and Decoder-only Transformer.

📢 News: This work has been accepted at the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP) System Demonstrations.

🔗 Resources:

• 📄 Paper
• 🌐 Live Demo
• 🛠️ System Walkthrough
• 🎥 Video Presentation

📑 Table of Contents

• 📖 Introduction
• ⚙️ Installation
• 🎓 Training
• 🚀 Inference
• 📧 Request Access to Weights/Synthetic Datasets
• 📝 Citation

⚙️ Installation

Follow these steps to set up the project on your local machine:

1. Clone the Repository

git clone https://github.com/mahmudulyeamim/GraDeT-HTR.git
cd GraDeT-HTR

2. Create a Virtual Environment (Recommended)

It is recommended to create and activate a virtual environment (e.g., venv or Conda) before installing dependencies.

3. Install Dependencies

Install the required packages using pip:

pip install -r requirements.txt

4. Set Up YOLOv5 in BN_DRISHTI

Navigate to the BN_DRISHTI directory and clone the YOLOv5 repository:

cd BN_DRISHTI
git clone https://github.com/ultralytics/yolov5

After cloning, replace the default detect.py file with the custom version provided in this repository:

cp detect.py yolov5/detect.py

You can now safely delete the detect.py file from the BN_DRISHTI directory:

rm detect.py

5. Download Detection Model Weights

Download the required pre-trained detection model weights into the model_weights directory:

wget -P model_weights https://huggingface.co/crusnic/BN-DRISHTI/resolve/main/models/line_model_best.pt
wget -P model_weights https://huggingface.co/crusnic/BN-DRISHTI/resolve/main/models/word_model_best.pt

Detection model weights provided by: BN-DRISHTI

Note: If you don't have wget installed, you can install it using your package manager (e.g., apt-get install wget on Ubuntu, brew install wget on macOS). Alternatively, you can download the files manually from the links above and place them in the BN_DRISHTI/model_weights directory.

You're all set! The installation is complete.

🎓 Training

Training the Text Recognition Model

To train the text recognition model, navigate to the GraDeT_HTR directory and run the train.py file. A sample training dataset is provided to help you test whether the model runs correctly.

Quick Test with Sample Dataset:

cd GraDeT_HTR
python train.py

This will train the model using the provided sample_train dataset, allowing you to verify that everything is set up properly.

🚀 Inference

Preparing Your Input

Before running inference, create an input_pages directory in the root of the project.

mkdir input_pages

Then copy your input images (.jpg, .png, etc.) or PDF files into this directory. Sample test files are available in the sample_test directory for quick testing.

Running Inference

To perform inference, run the inference.py with the path to your pre-trained text recognition model weights.

For image inputs:

python inference.py --weights path/to/your/text_recognition_model_weights.pth

For PDF inputs: Add the --pdf true flag

python inference.py --weights path/to/your/text_recognition_model_weights.pth --pdf true

Output

The inference results will be saved in the output_texts directory (created automatically if it doesn't exist).

📧 Request Access to Weights/Synthetic Datasets

To request access to our pre-trained text recognition model weights or synthetic datasets, please email mahmudulyeamim@gmail.com with the following information:

• Your name and institutional affiliation
• Intended purpose and use case
• Brief project description
• Confirmation that you will properly acknowledge and cite this work

Access will be granted upon review of your request.

Important: Pre-trained weights and synthetic datasets are provided strictly for academic and research purposes.

📝 Citation

If you use GraDeT-HTR in your research, please cite our paper:

@inproceedings{hasan-etal-2025-gradet,
  title     = {GraDeT-HTR: A Resource-Efficient Bengali Handwritten Text Recognition System Utilizing Grapheme-based Tokenizer and Decoder-only Transformer},
  author    = {Hasan, Md. Mahmudul and Choudhury, Ahmed Nesar Tahsin and Hasan, Mahmudul and Khan, Md Mosaddek},
  booktitle = {Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations},
  year      = {2025},
  pages     = {696--706},
  publisher = {Association for Computational Linguistics},
  url       = {https://aclanthology.org/2025.emnlp-demos.52/},
  doi       = {10.18653/v1/2025.emnlp-demos.52}
}

---

Contact: For questions or issues, please open an issue on this repository or contact mahmudulyeamim@gmail.com