This is a simple text image super-resolution package.

This is a simple baseline (ESRGAN) trained using synthetic data from our CVPR paper MARCONet. This model is trained on Chinese and English Characters. When the degradation is not severe, it may also perform well in other languages, like Japanese.

This package can post-process the text region with a simple command, i.e.,

textbsr -i [LR_TEXT_PATH] -b [BACKGROUND_SR_PATH]

• [LR_TEXT_PATH] is the LR image path.
• [BACKGROUND_SR_PATH] stores the results from any blind image super-resolution methods.
• If the text image is degraded severely, this method may still fail to obtain a plausible result.

Dependencies and Installation

• numpy
• torch>=1.8.1
• torchvision>=0.9
• cnstd==1.2

# Install with pip
pip install textbsr

Basic Usage

# On the terminal command
textbsr -i [LR_TEXT_PATH]

or

# On the python environment
from textbsr import textbsr
textbsr.bsr(input_path='./testsets/LQs')

Parameter details:

(1) Text Region Restoration

# On the terminal command
textbsr -i [LR_TEXT_PATH]

or

# On the python environment
from textbsr import textbsr
textbsr.bsr(input_path='./testsets/LQs', save_text=True)

(2) Post-process the Text Region from Any Blind Image Super-resolution (BSR) Methods

# On the terminal command
textbsr -i [LR_TEXT_PATH] -b [AnyBSR_Results_PATH] -s

or

# On the python environment
from textbsr import textbsr
textbsr.bsr(input_path='./testsets/LQs', bg_path='./testsets/AnyBSRResults', save_text=True)

When [AnyBSR_Results_PATH] is None, we only restore the text region and paste it back to the LR input, with the background region unchanged.

Real-world LR Text Image	AnyBSR Method	Post-process using our textbsr

(3) Example for restoring the aligned text region

# On the terminal command
textbsr -i [LR_TEXT_PATH] -a

or

# On the python environment
from textbsr import textbsr
textbsr.bsr(input_path='./testsets/LQs', aligned=True)

Aligned LR Text Image	TextBSR

Acknowledgement

This project is built based on BSRGAN. We use cnstd for Chinese and English text detection.

:bookmark_tabs: Citation

If you find this package helpful, please kindly consider citing our CVPR23 paper MARCONet:

@InProceedings{li2023marconet,
author = {Li, Xiaoming and Zuo, Wangmeng and Loy, Chen Change},
title = {Learning Generative Structure Prior for Blind Text Image Super-resolution},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
year = {2023}
}

:scroll: License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.