Online Handwritten Text Recognition with PyTorch

A clean PyTorch implementation of "Fast multi-language LSTM-based online handwriting recognition" by Carbune et al. (2020) 🚀⚡🔥.

Contributions are always welcome!

📌 Introduction

In this repository I provide a clean PyTorch model implementation of the paper "Fast multi-language LSTM-based online handwriting recognition" by Carbune et al. (2020) from Google; see this paper.

This model lets you predict handwritten text as shown in the project demo below.

The present work is part of my attempt to build a handwriting recognition system for Xournal++, a fabulous open-source handwriting notetaking software. Most of this aforementioned attempt is captured in the Xournal++ HTR repository where I publish working solutions based on models like the one implemented in this repository.

📺 Project Demo

(Click on GIF or here to get to video.)

🚀 Quickstart

The following explanation sets you up to use the scripts src/draw_and_predict_sample.py and src/draw_and_store_sample.py to predict your own handwritten text as shown in the 📺 project demo and to store the written sample, respectively:

1. Install the project according to the installation section in this README and activate the corresponding environment.
2. Download the model weights here and place it in models/dataIAMOnDB_featuresLinInterpol20DxDyDtN_decoderGreedy/ after unpacking it.
3. Invoke the following command from the root of this repository: python src/draw_and_predict_sample.py.

If you want to store your own handwriting sample in a CSV file, then execute python src/draw_and_store_sample.py.

🏋️ Training from scratch

1. Follow installation procedure provided in Installation.
2. Set up the training data as described in the training data section.
3. Start training with bash scripts/train_dataIAMOnDB_featuresLinInterpol20DxDyDtN_decoderGreedy.sh from repository root directory after you activated the conda environment that you use.

Installation

This repository uses a conda environment in which packages are installed using pip.

Follow these steps from the root of the repository to install this package:

1. conda create --prefix <path> python=3.10.11
2. conda activate <path>
3. pip3 install torch torchvision torchaudio
4. pip install -r requirements.txt
5. pip install -e . (do not forget the dot, .)
6. make test to confirm that installation was successful. Alternatively, you can run make test-full to run all tests including the slow ones.

Training data

IAM On-Line Handwriting Database is used as training and validation data. Register on their website to download the dataset for free. Afterwards, place the following folders and files from their dataset in this repository's subfolder data/datasets/IAM-OnDB:

1. Download the following files that are listed on the above stated dataset website: data/original-xml-part.tar.gz, data/writers.xml, data/lineStrokes-all.tar.gz, data/lineImages-all.tar.gz, data/original-xml-all.tar.gz, data/forms.txt & ascii-all.tar.gz.
2. Extract the content of each of those files into the data/datasets/IAM-OnDB/<file_base_name> folder where <file_base_name> denote the basenames of each of the downloaded files.
3. This is how it should look like:

├── data/datasets/IAM-OnDB
│   ├── ascii-all/
│   ├── forms.txt
│   ├── lineImages-all
│   ├── lineStrokes-all
│   ├── original-xml-all
│   ├── original-xml-part
│   └── writers.xml

4. Confirm with make test-training-data that you placed the training data correctly.

Available models & their model cards

• dataIAMOnDB_featuresLinInterpol20DxDyDtN_decoderGreedy
  • How it was trained: Using the IAM-OnDB dataset, this model is obtained by training on raw stroke data with channels (dx, dy, dt, n) where (dx, dy) are coordinate differences, dt is time difference and n is 1 if a point is the start of a new stroke with 0 otherwise. Prior to computing the differences and n, the raw stroke data is linearly interpolated to feature 20 points per unit length.
  • Download the model weights here.
  • Train this model yourself executing bash scripts/train_dataIAMOnDB_featuresLinInterpol20DxDyDtN_decoderGreedy.sh after you activated the conda environment that you use.

⌛ Open tasks

All contributions are welcome! :-)

• Allow inference on own handwriting.
• Implement CTC beam decoding with language model.
• Implement Bezier curve fitting algorithm as data preprocessor.
• Publish trained models on 🤗 Hugging Face for easy access.

👩‍💻 Contributing

I would love you to contribute! Let's make it a great project that people can benefit from :-).

🙏🏼 Acknowledgments

Thanks Leonard Salewski and Jonathan Prexl for super useful discussions on training PyTorch models! Thanks Harald Scheidl for providing both great content and code around handwritten text recognition.

I thank the department where I do my PhD, the School of Physics and Astronomy of The University of Edinburgh, for providing computational resources to train the models.

The scaffold of this code is based on the awesome lightning-hydra-template by ashleve - thank you for providing this!

License

This repository is licensed under the MIT License.