Handwriting recognition for various Indic scripts using deep learning(STN, CNN, LSTM)
For more info about the pipeline, please refer to IIIT-INDIC-HW-WORDS: A Dataset for Indic Handwritten Text Recognition
- Create LMDB files for train, validation and test splits.
python tools/create_dataset.py --root_dir <dataset_dir> --save <lmdb_dst_path>
The dataset folder should follow the same structure as IIIT-INDIC-HW-WORDS structure.
- Generate a file containing Unicode symbols/characters to be used for prediction. Move this file to alphabet/ folder. This repo already contains the sorted alphabet list for Indic scripts in the alphabet/ folder.
To train model(TPS-ResNet-BiLSTM-CTC) from scratch:
python lang_train.py --mode train --lang <lang_code> --trainRoot <train_lmdb_path> --valRoot <val_lmdb_path> --cuda
Refer to lang_train.py and config.py for default settings and additional parameter settings.
Language codes for Indic scripts:
| Bengali | Gujarati | Gurumukhi | Odia | Kannada | Malayalam | Tamil | Urdu |
|---|---|---|---|---|---|---|---|
| bn | gu | pn | od | kn | ma | ta | ur |
To generate predictions for a file, try:
python lang_train.py --lang <lang_code> --mode test --val_dir <test-lmdb-path> --output <save-predictions-path>
To evaluate the generated predictions, try the following:
python tools/score.py --preds <save-predictions-path>
or
python tools/oov_score.py --preds <save-predictions-path> --vocab <path-to-train-vocab>
to get WER and CER for OOV words only.
Please check the code for more config options.
Will be released soon.
We thank these repositories crnn.pytorch, deep-text-recognition-benchmark for releasing their codes. Code structure is inspired from aster.pytorch.
Please consider citing this work in your publications if it helps your research.
@InProceedings{10.1007/978-3-030-86337-1_30,
author="Gongidi, Santhoshini
and Jawahar, C. V.",
title="iiit-indic-hw-words: A Dataset for Indic Handwritten Text Recognition",
booktitle="Document Analysis and Recognition -- ICDAR 2021",
year="2021",
}