Home

Awesome

Learning Enhancement From Degradation: A Diffusion Model For Fundus Image Enhancement

The official implementation of the paper Learning Enhancement From Degradation: A Diffusion Model For Fundus Image Enhancement

<p align="center"> <img src="https://img.shields.io/github/license/QtacierP/LED" /> <img src="https://img.shields.io/github/issues/QtacierP/LED" /> <img src="https://img.shields.io/github/forks/QtacierP/LED" /> <img src="https://img.shields.io/github/stars/QtacierP/LED" /> </p> <img src="./docs/led.gif" width="300" height="300">

Highlights

image info

image info

image info

image info

image info

image info

Start LED with few lines

from led.pipelines.led_pipeline import LEDPipeline
led = LEDPipeline()
led.cuda()
led_enhancement = led('./doc/example.jpeg')[0]

Furthermore, you can combine LED with any existing SOTA methods as external backend. Current supported backends include:

Try I-SECRET backend with only one line

led = LEDPipeline(backend='I-SECRET', num_cond_steps=200)

For more details, please read example.ipynb. Please feel free to pull your proposed fundus enhancement methods as backend.

Catalog

Train

For training your own LED, you need to update few lines in configs/train_led.yaml

    train_good_image_dir: # update to training hq images directrory
    train_bad_image_dir: # update to training lq images directrory
    train_degraded_image_dir: # update to training degraded images directrory
    val_good_image_dir:  # update to validation hq images directrory
    val_bad_image_dir: # update to validation lq images directrory

Please note that train_degraded_image_dir should contain degraded high-qualty images by any data-driven methods. We will inculde related codes in our future workspace. However, you can consider using some existing repos instead, like CUT or CycleGAN.

To train LED, simply run

accelerate launch  --mixed_precision fp16 --gpu_ids 0 --num_processes 1 script/train.py

More GPUs will take significant performance improvement.

Acknowledgement

Thanks for PCENet, ArcNet and SCRNet for sharing their powerful pre-trained weights! Thansk for diffusers for sharing codes.

Citation

If this work is helpful for your research, please consider citing the following BibTeX entry.

@article{cheng2023learning,
  title={Learning Enhancement From Degradation: A Diffusion Model For Fundus Image Enhancement},
  author={Cheng, Pujin and Lin, Li and Huang, Yijin and He, Huaqing and Luo, Wenhan and Tang, Xiaoying},
  journal={arXiv preprint arXiv:2303.04603},
  year={2023}
}

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.