PageNet: Towards End-to-End Weakly Supervised Page-Level Handwritten Chinese Text Recognition

The official implementation of PageNet: Towards End-to-End Weakly Supervised Page-Level Handwritten Chinese Text Recognition (IJCV 2022).

Environment

We recommend using Anaconda to manage environments.

conda create -n pagenet python=3.7 -y 
conda activate pagenet
git clone https://github.com/shannanyinxiang/PageNet
cd PageNet
pip install -r requirements.txt

Dataset

• ICDAR2013 Competition Dataset: BaiduNetDisk or Google Drive
• SCUT-HCCDoc: Please apply for this dataset at SCUT-HCCDoc_Dataset_Release.
• MTHv2: BaiduNetDisk or Google Drive

Download the datasets and put them into the datasets folder following the file structure below.

datasets
├─IC13Comp
├─MTHv2_test
└─raw
   └─SCUT-HCCDoc
      │  hccdoc_test.json
      │  hccdoc_train.json
      └─image

Then run the following command to generate the SCUT-HCCDoc dataset in lmdb format.

python tools/convert_hccdoc_to_lmdb.py \
  --image_root datasets/raw/SCUT-HCCDoc/image/ \
  --annotation_file datasets/raw/SCUT-HCCDoc/hccdoc_test.json \
  --dict_path dicts/scut-hccdoc.txt \
  --lmdb_root datasets/SCUT-HCCDoc_test

Inference

ICDAR2013 Competition Dataset

1. Download the pretrained weights from BaiduNetDisk or Google Drive and put it into the outputs/casia-hwdb/checkpoints folder.

2. Run the following command:

python main.py --config configs/casia-hwdb.yaml

The results will be saved at outputs/casia-hwdb/val_log.txt.

SCUT-HCCDoc

1. Download the pretrained weights from BaiduNetDisk or Google Drive and put it into the outputs/scut-hccdoc/checkpoints folder.

2. Run the following command:

python main.py --config configs/scut-hccdoc.yaml

The results will be saved at outputs/scut-hccdoc/val_log.txt.

MTHv2

1. Download the pretrained weights from BaiduNetDisk or Google Drive and put it into the outputs/mthv2/checkpoints folder.

2. Run the following command:

python main.py --config configs/mthv2.yaml

The results will be saved at outputs/mthv2/val_log.txt.

Model Performance

The performance of the provided models on these datasets should be:

Dataset	$AR^*$	$CR^*$
ICDAR2013 Competition Dataset	92.87	93.34
SCUT-HCCDoc	78.70	84.29
MTHv2	93.76	96.03

Training

Currently the training codes are not available. For questions about model training, please contact Prof. Lianwen Jin (eelwjin@scut.edu.cn) and Mr. Dezhi Peng (eedzpeng@mail.scut.edu.cn).

Note: In the spatial matching of the weakly supervised learning, we found it better to simply delete the matching pairs whose IoUs are equal to zero.

Citation

@article{peng2022pagenet,
  title={PageNet: Towards End-to-End Weakly Supervised Page-Level Handwritten Chinese Text Recognition},
  author={Peng, Dezhi and Jin, Lianwen and Liu, Yuliang and Luo, Canjie and Lai, Songxuan},
  journal={International Journal of Computer Vision},
  pages={2623--2645},
  year={2022},
  volume={130},
  number={11},
  doi={10.1007/s11263-022-01654-0},
}

License

This repository should be used and distributed under Creative Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) License for non-commercial research purposes.

Copyright

This repository can only be used for non-commercial research purpose.

For commercial use, please contact Prof. Lianwen Jin (eelwjin@scut.edu.cn).

Copyright 2022, Deep Learning and Vision Computing Lab, South China University of Technology.