Interaction-Guided-Two-Branch-Image-Dehazing-Network

Huichun Liu , Xiaosong Li^*, Tianshu Tan
IEEE/CVF Asian Conference on Computer Vision (ACCV2024)
Arxiv Link: https://arxiv.org/abs/2410.10121

---

Network Architecture

Datasets

Dataset	NH-HAZE	Dense-HAZE	6k
Baidu Cloud	Download (801y)	Download (ixen)	Download (gtuw)

data

Please download the corresponding datasets and put them in the folder ./data/<datase_name>. (The train and valid split files are provided in ./data/<datase_name>. )

The data directory structure will be arranged as: (Note: please check it carefully)

data
   |-Dense-Haze
      |- train_dense
         |- haze
            |- 01_hazy.png 
            |- 02_hazy.png
         |- clear_images
            |- 01_GT.png 
            |- 02_GT.png
         |- trainlist.txt
      |- valid_dense
         |- input 
            |- 51_hazy.png 
            |- 52_hazy.png
         |- gt
            |- 51_GT.png 
            |- 52_GT.png
         |- val_list.txt
   |-NH-Haze
      |- train_NH
         |- haze
            |- 01_hazy.png 
            |- 02_hazy.png
         |- clear_images
            |- 01_GT.png 
            |- 02_GT.png
         |- trainlist.txt
      |- valid_NH
         |- input 
            |- 51_hazy.png 
            |- 52_hazy.png
         |- gt
            |- 51_GT.png 
            |- 52_GT.png
         |- val_list.txt

Pre-trained Models

Dataset	NH-HAZE	Dense-HAZE	6k
Baidu Cloud	Download (yzcs)	Download (2fzx)	Download (xtds)

Then, place the models to ckpts/<dataset_name> directory, separately.

The directory structure will be arranged as:

ckpts
   |- Dense
      |- DENSE-16.42ssim0.5235.pt  
   |- NH
      |- NH-20.10ssim6716.pt.pt
   |- 6k
      |- 6K-30.20ssim0.9643.pt

Training

See python src/train.py --h for list of optional arguments, or train.sh for examples.

An example of NH-HAZE dataset.

CUDA_VISIBLE_DEVICES=0,1 python src/train.py \
  --dataset-name NH \
  --train-dir ./data/train_NH/ \
  --valid-dir ./data/valid_NH/ \
  --ckpt-save-path ../ckpts \
  --ckpt-overwrite \
  --nb-epochs 5000 \
  --batch-size 2\
  --train-size 800 1200  \
  --valid-size 800 1200 \
  --loss l1 \
  --plot-stats \
  --cuda   

Or you can just run it

bash train.sh

Testing

See python test_PSNR.py --h for list of optional arguments, or test.sh for an example.

An example of NH-HAZE dataset.

CUDA_VISIBLE_DEVICES=1 python src/test_PSNR.py \
  --dataset-name NH   

Or you can just run it

bash test.sh

Full-size Evaluation Results

Dataset	PSNR	SSIM	Entriopy	LPIPS
NH-HAZE	20.10	0.6716	7.5319	0.3210
Dense-HAZE	16.42	0.5235	6.9600	0.6696
6K	30.20	0.9643	7.4325	0.1749

Citation

If you find this repo useful, please give us a star and consider citing our papers:

@inproceedings{liu2024interaction,
  title={Interaction-Guided Two-Branch Image Dehazing Network},
  author={Liu, Huichun and Li, Xiaosong and Tan, Tianshu},
  booktitle={Proceedings of the Asian Conference on Computer Vision},
  pages={4069--4084},
  year={2024}
}

@article{liu2024interaction,
  title={Interaction-Guided Two-Branch Image Dehazing Network},
  author={Liu, Huichun and Li, Xiaosong and Tan, Tianshu},
  journal={arXiv preprint arXiv:2410.10121},
  year={2024}
}

Contact

If you have any question, please feel free to contact us via [email protected].

Acknowledgments

This code is based on Dehamer, DeHazeFormer.