This repository contains the official implementation of the ACMMM 2021 paper ReLLIE: Deep Reinforcement Learning for Customized Low-Light Image Enhancement.
To tackle the low-light image enhancement (LLIE) problem, we propose a novel deep reinforcement learning based method, dubbed ReLLIE, for customized low-light enhancement. Specifically, ReLLIE models LLIE as a markov decision process, i.e., estimating the pixel-wise image-specific curves sequentially and recurrently. Given the reward computed from a set of carefully crafted non-reference loss functions, a lightweight network is proposed to estimate the curves for enlightening of a low-light image input. For more details, please refer to our orginal paper.
You can install the required libraries by the command pip install -r requirements.txt
. We checked this code on cuda-10.0 and cudnn-7.3.1.
If you want to train the model
git clone git@github.com:GuoLanqing/ReLLIE.git
python train.py
If you want to test the pretrained model on noisy low-light images (enhancement with denoising)
python test.py
or on high-quality low-light images (enhancement without denoising)
python test_el.py
We used the publicly avaliable pretrained models of FFDNet as the denoiser module.
We obtained the LOL and LIME dataset from
Our implementation is based on PixelRL. We would like to thank them.
Preprint available here.
In case of use, please cite our publication:
R. Zhang, L. Guo, S. Huang and B. Wen, "ReLLIE: Deep Reinforcement Learning for Customized Low-Light Image Enhancement," ACM MM 2021.
Bibtex:
@article{zhang2021rellie,
title={ReLLIE: Deep Reinforcement Learning for Customized Low-Light Image Enhancement},
author={Zhang, Rongkai and Guo, Lanqing and Huang, Siyu and Wen, Bihan},
journal={arXiv preprint arXiv:2107.05830},
year={2021}
}
If you have any questions, please contact rongkai002@e.ntu.edu.sg or lanqing001@e.ntu.edu.sg