Skip to content

Latest commit

 

History

History
141 lines (101 loc) · 4.53 KB

README.md

File metadata and controls

141 lines (101 loc) · 4.53 KB
Raw

Enhancing Low-light Images Using Infrared Encoded Images

Shulin Tian*,  Yufei Wang*,  Renjie Wan,  Wenhan Yang,  Alex C. Kot,  Bihan Wen  
Nanyang Technological University, Hong Kong Baptist University, Peng Cheng Laboratory

Paper | Project Page | Dataset

The visibility of low-light images is enhanced by increasing the number of income photons (The right sides of (a) and (b) are amplified by a factor of 3.5 for better visualization).

Dataset

In this work, we are using a resized version - IR-RGB-resize [Google Drive] for our experiments. The file structure is constructed as follows:

data_root # The paths need to be specified in the training configs under folder `./code/confs/xx.yml`
└── train/
    ├── high/  
    └── low/
└── eval/
    ├── high/
    ├── low/
    └── low-rgb/

We also relased the original size of images for broadening research purposes IR-RGB [Google Drive], feel free to download and explore!

Results

Quantitative results

The evauluation results on IR-RGB dataset are as follows:

Method PSNR SSIM LPIPS
RetinexNet 11.14 0.628 0.586
LIME 11.31 0.639 0.560
Zero-DCE 11.40 0.592 0.443
KinD 14.73 0.714 0.357
EnlightenGAN 16.95 0.715 0.357
KinD++ 17.84 0.830 0.249
MIRNet 22.23 0.833 0.224
LLFlow 25.46 0.890 0.130
ELIEI (Ours) 26.23 0.899 0.116

Qualitative results

Comparison with other methods

Our method shows better performance in controlling color distortion and detail preservation.

Usage of IR-RGB dataset

(a), (c) are images captured from RGB and IR-RGB space separately under low-light conditions, (d), (f) are respective high-light outputs.

Usage of color alignment loss (CAL)

(c) is the result of our model trained directly w/o adding CAL, and (d) is the output from the same architecture but w/ CAL.

Get Started

Dependencies and Installation

  • Python 3.8
  • Pytorch 1.9
  1. Clone Repo
git clone https://github.com/shulin16/ELIEI.git
  1. Create Conda Environment
conda create --name ELIEI python=3.8
conda activate ELIEI

or you can just simply use conda env create -f environment.yml to install all the packages you need.

  1. Install Dependencies
cd ELIEI
pip install -r requirements.txt

Citation

If you find our work useful for your research, please cite our paper

@inproceedings{tian2023enhancing,
  title={Enhancing Low-Light Images Using Infrared Encoded Images},
  author={Tian, Shulin and Wang, Yufei and Wan, Renjie and Yang, Wenhan and Kot, Alex C and Wen, Bihan},
  booktitle={2023 IEEE International Conference on Image Processing (ICIP)},
  pages={465--469},
  year={2023},
  organization={IEEE}
}

Acknowledgment

We thank Yufei Wang for his work LLFlow. This work was done at Rapid-Rich Object Search (ROSE) Lab, Nanyang Technological University.