Skip to content

Canjie-Luo/Text-Image-Augmentation

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

18 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Text Image Augmentation

Build Status

A general geometric augmentation tool for text images in the CVPR 2020 paper "Learn to Augment: Joint Data Augmentation and Network Optimization for Text Recognition". We provide the tool to avoid overfitting and gain robustness of text recognizers.

Note that this is a general toolkit. Please customize for your specific task. If the repo benefits your work, please cite the papers.

News

  • 2020-02 The paper "Learn to Augment: Joint Data Augmentation and Network Optimization for Text Recognition" was accepted to CVPR 2020. It is a preliminary attempt for smart augmentation.

  • 2019-11 The paper "Decoupled Attention Network for Text Recognition" (Paper Code) was accepted to AAAI 2020. This augmentation tool was used in the experiments of handwritten text recognition.

  • 2019-04 We applied this tool in the ReCTS competition of ICDAR 2019. Our ensemble model won the championship.

  • 2019-01 The similarity transformation was specifically customized for geomeric augmentation of text images.

Requirements

We recommend Anaconda to manage the version of your dependencies. For example:

     conda install boost=1.67.0

Installation

Build library:

    mkdir build
    cd build
    cmake -D CUDA_USE_STATIC_CUDA_RUNTIME=OFF ..
    make

Copy the Augment.so to the target folder and follow demo.py to use the tool.

    cp Augment.so ..
    cd ..
    python demo.py

Demo

  • Distortion

  • Stretch

  • Perspective

Speed

To transform an image with size (H:64, W:200), it takes less than 3ms using a 2.0GHz CPU. It is possible to accelerate the process by calling multi-process batch samplers in an on-the-fly manner, such as setting "num_workers" in PyTorch.

Improvement for Recognition

We compare the accuracies of CRNN trained using only the corresponding small training set.

Dataset IIIT5K IC13 IC15
Without Data Augmentation 40.8% 6.8% 8.7%
With Data Augmentation 53.4% 9.6% 24.9%

Citation

@inproceedings{luo2020learn,
  author = {Canjie Luo and Yuanzhi Zhu and Lianwen Jin and Yongpan Wang},
  title = {Learn to Augment: Joint Data Augmentation and Network Optimization for Text Recognition},
  booktitle = {CVPR},
  year = {2020}
}

@inproceedings{wang2020decoupled,
  author = {Tianwei Wang and Yuanzhi Zhu and Lianwen Jin and Canjie Luo and Xiaoxue Chen and Yaqiang Wu and Qianying Wang and Mingxiang Cai}, 
  title = {Decoupled attention network for text recognition}, 
  booktitle ={AAAI}, 
  year = {2020}
}

@article{schaefer2006image,
  title={Image deformation using moving least squares},
  author={Schaefer, Scott and McPhail, Travis and Warren, Joe},
  journal={ACM Transactions on Graphics (TOG)},
  volume={25},
  number={3},
  pages={533--540},
  year={2006},
  publisher={ACM New York, NY, USA}
}

Acknowledgment

Thanks for the contribution of the following developers.

@keeofkoo

@cxcxcxcx

@Yati Sagade

Attention

The tool is only free for academic research purposes.