Awesome
Low-light Image Enhancement via Breaking Down the Darkness
by Xiaojie Guo, Qiming Hu.
:boom: Update Online Replicate Demo:
▶️ Online Colab Demo: 
1. Dependencies
- Python3
- PyTorch>=1.0
- OpenCV-Python, TensorboardX
- NVIDIA GPU+CUDA
2. Network Architecture
3. Data Preparation
3.1. Training dataset
- 485 low/high-light image pairs from our485 of LOL dataset, each low image of which is augmented by our exposure_augment.py to generate 8 images under different exposures. (Download Link for Augmented LOL)
- To train the MECAN (if it is desired), 559 randomly-selected multi-exposure sequences from SICE are adopted (Download Link for a resized version).
3.2. Tesing dataset
The images for testing can be downloaded in this link.
<!-- * 15 low/high-light image pairs from eval15 of [LOL dataset](https://daooshee.github.io/BMVC2018website/). * 44 low-light images from DICM. * 8 low-light images from NPE. * 24 low-light images from VV. -->4. Usage
4.1. Training
- Multi-exposure data synthesis:
python exposure_augment.py - Train IAN:
python train_IAN.py -m IAN --comment IAN_train --batch_size 1 --val_interval 1 --num_epochs 500 --lr 0.001 --no_sche - Train ANSN:
python train_ANSN.py -m1 IAN -m2 ANSN --comment ANSN_train --batch_size 1 --val_interval 1 --num_epochs 500 --lr 0.001 --no_sche -m1w ./checkpoints/IAN_335.pth - Train CAN:
python train_CAN.py -m1 IAN -m3 FuseNet --comment CAN_train --batch_size 1 --val_interval 1 --num_epochs 500 --lr 0.001 --no_sche -m1w ./checkpoints/IAN_335.pth - Train MECAN on SICE:
python train_MECAN.py -m FuseNet --comment MECAN_train --batch_size 1 --val_interval 1 --num_epochs 500 --lr 0.001 --no_sche - Finetune MECAN on SICE and LOL datasets:
python train_MECAN_finetune.py -m FuseNet --comment MECAN_finetune --batch_size 1 --val_interval 1 --num_epochs 500 --lr 1e-4 --no_sche -mw ./checkpoints/FuseNet_MECAN_for_Finetuning_404.pth
4.2. Testing
- [Tips]: Using gamma correction for evaluation with parameter --gc; Show extra intermediate outputs with parameter --save_extra
- Evaluation:
python eval_Bread.py -m1 IAN -m2 ANSN -m3 FuseNet -m4 FuseNet --mef --comment Bread+NFM+ME[eval] --batch_size 1 -m1w ./checkpoints/IAN_335.pth -m2w ./checkpoints/ANSN_422.pth -m3w ./checkpoints/FuseNet_MECAN_251.pth -m4w ./checkpoints/FuseNet_NFM_297.pth - Testing:
python test_Bread.py -m1 IAN -m2 ANSN -m3 FuseNet -m4 FuseNet --mef --comment Bread+NFM+ME[test] --batch_size 1 -m1w ./checkpoints/IAN_335.pth -m2w ./checkpoints/ANSN_422.pth -m3w ./checkpoints/FuseNet_MECAN_251.pth -m4w ./checkpoints/FuseNet_NFM_297.pth - Remove NFM:
python test_Bread_NoNFM.py -m1 IAN -m2 ANSN -m3 FuseNet --mef -a 0.10 --comment Bread+ME[test] --batch_size 1 -m1w ./checkpoints/IAN_335.pth -m2w ./checkpoints/ANSN_422.pth -m3w ./checkpoints/FuseNet_MECAN_251.pth
4.3. Trained weights
Please refer to our release.
5. Quantitative comparison on eval15
6. Visual comparison on eval15
7. Visual comparison on DICM
8. Visual comparison on VV and MEF-DS
