The "DARK: Denoising, Amplification Restoration Kit" project introduces an innovative computational framework for enhancing images captured in low-light conditions. The project leverages the principles of Retinex theory combined with advanced image restoration techniques facilitated by convolutional neural networks. By incorporating streamlined architectural elements inspired by MIRNet-v2 and the Retinexformer, the model focuses on efficient, context-sensitive image processing, significantly improving image clarity and color fidelity while maintaining a minimal computational footprint.
Provide step-by-step series of examples and explanations about how to get a development env running.
This repository is built in PyTorch 1.11 and tested on Ubuntu 16.04 environment (Python3.7, CUDA10.2, cuDNN7.6). Follow these intructions
git clone git@github.com:hollinsStuart/dark.git
cd dark
conda create -n dark python=3.7
conda activate dark
pip install -r requirements.txt
python setup.py develop
We use the following datasets:
Lol_train https://drive.google.com/file/d/1K29vsPfMUsAkYvmNLcaUgiOEYGMxFydd/view?usp=sharing
Lol_test https://drive.google.com/file/d/1jUGpsih3T-1H7t3gqpEdj7ZD5GcU_v0m/view?usp=sharing
Please modify the parameters in Enhancement/Options/dark_train_config.yml
.
python3 basicsr/train.py -opt Enhancement/Options/dark_train_config.yml
python3 basicsr/inference.py
The image outcome is at results/Enhancement_test folder
This project is licensed under the MIT License
Zhuoheng Li zhlii@umich.edu
Yuheng Pan extomato@umich.edu
Houchen Yu hollinsy@umich.edu
Zhiheng Zhang alexzh@umich.edu
Inspiration from MIRnet_v2 https://github.com/swz30/MIRNetv2, Retinexformer https://github.com/caiyuanhao1998/Retinexformer and basicSR https://github.com/XPixelGroup/BasicSR.