DW-GAN: A discrete wavelet transform GAN for NonHomogenous Image Dehazing - NTIRE 2021

This is the official PyTorch implementation of DW-GAN.
1st place solution of NTIRE 2021 NonHomogeneous Dehazing Challenge (CVPR Workshop 2021).

Email contact: [email protected]

See more details in [report] , [paper], [certificates]

Environment:

• Ubuntu: 18.04

• CUDA Version: 11.0

• Python 3.8

Dependencies:

• torch==1.6.0
• torchvision==0.7.0
• NVIDIA GPU and CUDA

Pretrained Weights & Dataset

1. Download ImageNet pretrained weights and Dehaze weights and place into the folder ./weights.
2. Download the NH-HAZE and NH-HAZE2 dataset.

Test

For inference, run following commands. Please check the test hazy image path (test.py line 12) and the output path (test.py line 13) .

python test.py

Qualitative Results

Results on NTIRE 2021 NonHomogeneous Dehazing Challenge validation images:

Results on NTIRE 2021 NonHomogeneous Dehazing Challenge testing images:

Acknowledgement

We thank the authors of Res2Net, MWCNN, and KTDN. Part of our code is built upon their modules.

Citation

If our work helps your research, please consider to cite our paper:

@InProceedings{Fu_2021_CVPR,
    author    = {Fu, Minghan and Liu, Huan and Yu, Yankun and Chen, Jun and Wang, Keyan},
    title     = {DW-GAN: A Discrete Wavelet Transform GAN for NonHomogeneous Dehazing},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
    month     = {June},
    year      = {2021},
    pages     = {203-212}
}