Glow in the Dark: Low-Light Image Enhancement with External Memory (TMM 2023)

Official Pytorch implementation of "Glow in the Dark: Low-Light Image Enhancement with External Memory" published in IEEE Transactions on Multimedia (TMM).

[Paper-official]

Dongjie Ye, Zhangkai Ni,Wenhan Yang, Hanli Wang, Shiqi Wang, Sam Kwong

Requirements and Installation

pytorch=1.8.1

(Please refer EMNet_env.yml for our operational environment.

Be aware that EMNet_env.yml includes redundant libraries. Kindly install the relevant libraries according to your needs.)

Testing (Running pretrained models)

• Checkpoint

Checkpoints trained on LOL-v1 and LOL-v2 dataset can be found from Google Drive or Baidu Netdisk (提取码：zhwy).

• Inference

1. Unzip the checkpoint file and place all the files in the ./pre_trained_logs/enhancer/ and ./pre_trained_logs/memory/ directory.
2. Edit the test_script.sh file to modify your python path and the [--input_dir] by specifying the path to your test datasets.
3. Excute the test script below:

   sh test_script.sh
   

4. The results are displayed in the './log_eval/lolv1' and './log_eval/lolv2' directories. In the '[ori]' folder, you can find the original outputs from the image enhancer, while in the '[mem]' folder, you can find the outputs after utilizing external memory.
5. You may also replace our image enhancer with other existing methods if desired.

Training (Training from scratch)

• Image Enhancer

Beforer training the external memory, you need to train the image enhancer first.

1. Edit the train_enhancer_script.sh file to modify your python path and the [--train_dir], [--val_dir] by specifying the path to your training datasets.
2. Excute the training script for image enhancer below:

   sh train_enhancer_script.sh
   

3. Find the trained image enhancer weight in the ./log/[env]/models/ folder.

• External Memory

Training the external memory requires a pre-trained image enhancer.

1. Edit the train_memory_script.sh file to modify your python path, the [--pretrain_weights] by specifying the path to your pretrained_weights of image enhancer, and the [--train_dir], [--val_dir] by specifying the path to your training datasets.
2. Excute the training script for external memory below:

   sh train_memory_script.sh
   

3. Find the trained memory weight in the ./log/[env]/models/ folder.

• Test

1. Edit the test_script.sh file to modify your python path, [--weights] by specifying the path to your image enhancer, [--mem_weights] by specifying the path to your external memory, and the [--input_dir] by specifying the path to your test datasets.
2. Excute the test script below:

   sh test_script.sh
   

3. The [input_dir] directory structure will be arranged as:

[your input dir]
    |- high
        |- 695.png (or alternative img formats)
        |- ...
    |- low
        |- 695.png (or alternative img files)
        |- ...

Citation

If this code is useful for your research, please cite our paper:

@article{emnet,
  author={Ye, Dongjie and Ni, Zhangkai and Yang, Wenhan and Wang, Hanli and Wang, Shiqi and Kwong, Sam},
  journal={IEEE Transactions on Multimedia}, 
  title={Glow in the Dark: Low-Light Image Enhancement with External Memory}, 
  year={2023},
  volume={},
  number={},
  pages={1-16},
  doi={10.1109/TMM.2023.3293736}}

Contact

Thanks for your attention! If you have any suggestion or question, feel free to leave a message here or contact Dongjie Ye ([email protected]).