Scene Text Image Generation

This software implements the scene text image generation network based on GAN. For details, please refer to our paper.

Scene Text Image Generation

• Python 3.x
• PyTorch 1.x
• TorchVision
• Pillow

Data Preparation

Please convert your own training dataset according to the proposed one in samples/. There should be three sub-folders: image/ that contains real scene text images, mask/ that contains corresponding masks and masktr/ that contains masks in which the text area are dilated.

For testing, simply put the masks of scene text images in a folder.

Train and Test

To train a new model, simply execute python train.py --data_path {train_data_path} --cuda. If you need to set other parameters, explore train.py for details.

To test a trained model, you need to explore and execute generate.py.

Citation

@article{gong2020generating,
  title={Generating Text Sequence Images for Recognition},
  author={Gong, Yanxiang and Deng, Linjie and Ma, Zheng and Xie, Mei},
  journal={Neural Processing Letters},
  pages={1--12},
  year={2020},
  publisher={Springer}
}