This repository contains the implementation of the following paper:

Learning Lightweight Low-Light Enhancement Network using Pseudo Well-Exposed Images
Seonggwan Ko*, Jinsun Park*, Byungjoo Chae and Donghyeon Cho
IEEE Signal Processing Letters
[pdf] [dataset] [StudentNetwork] [TeacherNetwork(Coming Soon)]

The following packages must be installed to perform the proposed model:

• PyTorch 1.7.1
• torchvision 0.8.2
• Pillow 8.2.0
• TensorBoardX 2.2
• tqdm

Test datasets should be arranged as the following folder dataset/test.

dataset
│   ├── test
│   │   ├── LIME
│   │   ├── LOL
│   │   ├── DICM
│   │   └── ...
└── ...

If you set up the folder, you can make it run.

python test.py

To train the proposed model, the following options are required:

python train.py --lowlight_images_path 'your_dataset_path' --gt_images_path 'your_GT_dataset_path' --pretrain_dir  'your_pretrain_path'

lowlight_images_path is the path of your low-light image

gt_images_path is the path of your ground-truth image

pretrain_dir is the path of your pretrained teacher model path

We provide 9,983 training pairs and 366 test images.

Please click here to download Low-light CNU(LLCNU) dataset.

• We collected 25,967 low-light images from BDD100k(4,830 images) and Dark Zurich(5,336 images), LoLi-Phone(now LLIV)(6,442 images), ExDark(7,263 images), SICE(1,611 images), LOL(485 images).
• Then, we generate pseudo well-exposed images using the pretrained EnlightenGAN, and additionally reduce noise using DnCNN.

@article{ko2021learning,
  title={Learning Lightweight Low-Light Enhancement Network using Pseudo Well-Exposed Images},
  author={Ko, Seonggwan and Park, Jinsun and Chae, Byungjoo and Cho, Donghyeon},
  journal={IEEE Signal Processing Letters},
  year={2021},
  publisher={IEEE}
}

The code framework is mainly modified from Zero-DCE, AdaBelief and SPKD. Please refer to the original repo for more usage and documents. Thanks to authors for sharing the codes!