Underwater Image Enhancement Toolbox

This is an open source underwater image enhancement toolbox based on PyTorch. Meanwhile, this is the code of the implementation of the UIEC^2-Net. If you use our code, please consider citing our paper. Thanks.

Documentation: https://uwenhancement.readthedocs.io/

Introduction

This is a code that summarizes deep learning-based underwater image enhancement. We converted relevant Tensorflow code to PyTorch, and transcoded the Tensorflow checkpoint provided by the author so that it can be read by PyTorch. We confirmed the reliability of transcoding, and uploaded the PyTorch checkpoint models(.pth file) to Google Drive and Baidu Cloud. (You can download at Model Zoo)

This is convenient for everyone to use the same environment to run experiments. We currently support UIEC^2-Net, UWCNN, WaterNet, and we will continue to update. (More details in Model Zoo)

Highlights

• TTA x8 (Test Time Argument x8 images)

• fp16

• Visualize Network Structure

• Training process monitoring (Use Visdom package)

• Training log records, and loss visualize(Use TensorboardX package)

• Tensorflow to PyTorch.

• ...

You can get more details at useful_tools.

TODO:

• We will support GAN soon.
• In the future we will support PIxel-Level related tasks, such as Dehazing, Salient Object Detection(SOD) etc.

Installation

Please refer getting_start.md for installation.

Prepare Data

Using write_txt.py prepare train.txt and test.txt. All datasets should put in DATA folders, the directory structure is:

.
├── ...
├── DATA
│   ├── Train
│   │   ├── train
│   │   ├── gt
│   │   └── train.txt
│   ├── Test
│   │   ├── test
│   │   ├── gt(if need)
│   │   └── test.txt
└── ...

If you want to run WaterNet, you should prepare GC, WB, HE images and put in DATA/Train/WaterNet and DATA/Train/WaterNet, the directory structure is:

.
├── WaterNet
│    ├── ce
│    ├── gc
│    └── wb

Prepare Training and Testing

1. Clone the repo.

   git clone https://github.com/BIGWangYuDong/UWEnhancement.git UW

2. put the checkpoint(*.pth) in checkpoints/
3. change args.config and args.load_from root path.
4. change config/XXX.py data root path, and some other important options.

If you want to use our toolbox to add some New Models, Data Pipelines, loss functions or some other operation, We provide Tutorials.

Testing

python test.py
# or 
python test_UWModels/test_XXX.py

Training

python -m visdom.server         # open this in a single terminal
python train.py --gpus XX       # XX means use numbers of gpus

Contact Us

If you have any questions, please contact us (yudongwang@tju.edu.cn or yudongwang1226@gmail.com).

Bibliography entry for citation:

@article{wang2021uiec,
title = {{UIEC\^{}2-Net}: CNN-based underwater image enhancement using two color space},
journal = {Signal Processing: Image Communication},
volume = {96},
pages = {116250},
year = {2021},
issn = {0923-5965},
doi = {https://doi.org/10.1016/j.image.2021.116250},
url = {https://www.sciencedirect.com/science/article/pii/S0923596521001004},
author = {Yudong Wang and Jichang Guo and Huan Gao and Huihui Yue},
}