Easter2.0: IMPROVING CONVOLUTIONAL MODELS FOR HANDWRITTEN TEXT RECOGNITION

This repo provides the model and code for our paper: Easter2.0: IMPROVING CONVOLUTIONAL MODELS FOR HANDWRITTEN TEXT RECOGNITION

[[Arxiv-PDF-Link]]

Overview

In this paper, we proposed a convolutional architecture for the task of handwritten text recognition that utilizes only 1D convolutions, dense residual connections and a SE module. We also proposed a simple and effective data augmentation technique-’Tiling and Corruption (TACo)’ useful for OCR/HTR tasks. We have presented experimental study on components of Easter2.0 architecture including dense residual connections, normalization choices, SE module, TACo variations and few-shot training. Our work achieves SOTA results on IAM-Test set when training data is limited, also Easter2.0 has very small number of trainable parameters compared to other solutions. The proposed architecture can be used in search of smaller, faster and efficient OCR/HTR solutions when available annotated data is limited.

How to use?

The following steps can help setting up Easter2 fast:

Download checkpoint from release, and put it inside /weigths directory. The download link is here
Download IAM dataset, and update data path in /src/config.py (sample notebook - /notebooks/iam_dataset_download.ipynb)
install requirements as per the file requirements.txt
Modify /src/config.py as per your needs
run the train() function from /src/easter_model.py
sample training and testing notebooks are given in /notebooks directory

How to predict for single line image?

Go to the src folder.
Run the python predict_line.py --path img_path

Contributing

This is a basic keras implementation of Easter2.0 model as per the paper (paper link above). If there is an issue or feature request, feel free to open an issue. Additionally, a PR is always welcome.

Citation

If you find our work helpful, please cite the following:

@article{chaudhary2022easter2,
  title={Easter2. 0: Improving convolutional models for handwritten text recognition},
  author={Chaudhary, Kartik and Bali, Raghav},
  journal={arXiv preprint arXiv:2205.14879},
  year={2022}
}