网络结构

News

• 置顶: 介绍一款我们实验室开发的多功能且多平台的OCR软件,包含常用的各种OCR功能，例如PDF转word，PDF转excel，公式识别，表格识别以及自动去除水印功能，欢迎试用！
• 查看To-do lists，获取最新信息。

使用指南

环境配置

深度学习环境

• python >= 3.7
• pytorch >= 1.7.0
• torchvision >= 0.8.0
• lmdb >= 0.98
• pillow >= 7.1.2
• numpy
• six
• tqdm
• python-opencv
• easydict
• yaml

数据集

• 下载TextZoom数据集

相关权重文件

• 下载Aster model权重文件
• 下载Moran model权重文件
• 下载CRNN model权重文件

训练

1. 安装

git clone https://github.com/Lenubolim/TextDiff.git

2. 参数配置
   见config.yaml文件
   

3. 训练

python train.py

推理

python test.py

To-do lists

• [ ] 添加训练代码(To be released soon.)
• [ ] 添加推理代码(To be released soon.)
• [ ] 使用DPM_solver减少推理步长

效果图

感谢

• 如果你觉得TextDiff对你有帮助，请给个star，谢谢！
• 如果你有任何问题，欢迎提issue(issue通知与我邮箱绑定，看到后我会及时回复)。
• 如果你愿意将TextDiff作为你的项目的baseline，欢迎引用我们的论文。

References

• [1] Scene text telescope: Text-focused scene image super-resolution
• [2] Activating more pixels in image super-resolution transformer.
• [3] Srdiff: Single image super-resolution with diffusion probabilistic models.
• [4] DocDiff: Document Enhancement via Residual Diffusion Models
• [5] Improving Scene Text Image Super-Resolution via Dual Prior Modulation Network

:book: Citation

If you use (part of) my code or find my work helpful, please consider citing

@article{liu2023textdiff,
  title={TextDiff: Mask-Guided Residual Diffusion Models for Scene Text Image Super-Resolution},
  author={Liu, Baolin and Yang, Zongyuan and Wang, Pengfei and Zhou, Junjie and Liu, Ziqi and Song, Ziyi and Liu, Yan and Xiong, Yongping},
  journal={arXiv preprint arXiv:2308.06743},
  year={2023}
}

Acknowledgement

This code is developed relying on DocDiff and TATT. Thanks for these great projects. Among them, DocDiff is the main research content of my classmate, and I participated in part of the research.