Attention Based CycleDehaze

Attention-based Single Image Dehazing Using Improved CycleGAN, IJCN IEEE WCCI 2022. Official Pytorch based implementation.

Model Architecture

Dependencies

• Python 3
• PyTorch >= 1.0
• NVIDIA GPU+CUDA

Dataset

Dataset used : RESIDE

File Structure

project
│   README.md
│   dataset.py
│   main.py
│   metrics.py
│   option.py
│   utility.py
└───inputs   
|
└───outputs   
|   
└───models   
|   |   dehaze.py
|   |   dicriminator.py
|   |   generator.py
|   └───DCNv2_latest   
|      
└───data
│   └───haze
│   |   |   *.png
│   |   
│   └───clear
│   |   |   *.png
│   |   
│   └───SOTS
│       └───indoor
│       |   └───haze
│       |   |   |   *.png
│       |   |   
│       |   └───clear
│       |       |   *.png
│       |       
│       └───indoor
│           └───haze
│           |   |   *.png
│           |   
│           └───clear
│               |   *.png
|
└───trained_models

Metrics update

Methods	Indoor(PSNR/SSIM)	Outdoor(PSNR/SSIM)
Paired Models	-	-
AOD-NET	19.06/0.8504	20.29/0.8765
DehazeNet	21.14/0.8472	22.46/0.8514
FFA-Net	36.39/0.9886	33.57/0.9840
Unpaired Models	-	-
DCP	16.62/0.8179	19.13/0.8148
Improved CycleGAN (with ssim loss)	20.05/0.8307	21.14/0.8919
Dehaze-GLCGAN	23.03/0.9165	26.51/0.9354
Ours	31.67/0.9612	36.17/0.9745

Usage

Train

Unzip DCNv2_latest.zip inside models and build the files. Train the model in ITS dataset.

python main.py

Test

Put your images in input.

python main.py --eval

the dehazed image will be saved at output

Samples

Acknowledgements

The code for DCN module implementation in PyTorch has been taken from DCNv2_latest.