DehazeFlow

Code and trained models for reproducing results of:

DehazeFlow: Multi-scale Conditional Flow Network for Single Image Dehazing
Hongyu Li, Jia Li, Dong Zhao, Long Xu
ACM Conference on Multimedia (ACM MM), 2021.

[Paper]

News

The latest performance of our method after parameter adjustment is as follows:

	PSNR	SSIM	LPIPS
indoor	40.88	0.9897	0.0025
outdoor	34.50	0.9859	0.0051

Link of trained models: https://drive.google.com/drive/folders/1NtQyK5dVu47E2LBk5ETcivPO96YFomPN?usp=sharing

Note: Opencv of this version is different from before. You need to run:

conda uninstall opencv
pip install opencv-python==4.5.3.56

Environment

• python==3.8.0
• lpips==0.1.3
• pytorch==1.8.0
• scikit-image==0.18.1
• opencv==4.0.1

Note: Different versions of opencv may cause different data reading results.

Datasets

We use different parts of the RESIDE dataset for training and validation.

• ITS (indoor training set)
• OTS (outdoor training set)
• SOTS (testing set)
• RTTS (real world testing samples)

Use /code/img2path.py to read the image paths and generate path files.

Testing

Download the trained models via https://drive.google.com/drive/folders/1Vb9BNYrDqKykfLpbX2lhNaus5YQw-s6V?usp=sharing.

Modify DehazeFlow.yml to:

1. set dataroot_GT and dataroot_HZ to paths containing testing images and ground-truths.
2. set test_mode to 'indoor' or 'outdoor'.
3. set model_path to pth file path.
4. set heat (standard deviation) to an appropriate value.

Run:

python test.py

Training

Modify DehazeFlow.yml to:

1. set path_root to path files for training and validation.
2. set other parameters to appropriate values.

Run:

python -m torch.distributed.launch --nproc_per_node=2 train.py

Comparison

Citation

@inproceedings{10.1145/3474085.3475432,
  author = {Li, Hongyu and Li, Jia and Zhao, Dong and Xu, Long},
  title = {DehazeFlow: Multi-Scale Conditional Flow Network for Single Image Dehazing},
  year = {2021},
  isbn = {9781450386517},
  publisher = {Association for Computing Machinery},
  address = {New York, NY, USA},
  url = {https://doi.org/10.1145/3474085.3475432},
  doi = {10.1145/3474085.3475432},
  pages = {2577–2585},
  numpages = {9},
  keywords = {normalizing flow, single image dehazing, attention},
  location = {Virtual Event, China},
  series = {MM '21}
}

Acknowledgement

This repository is based on the implementation of SRFlow: Learning the Super-Resolution Space with Normalizing Flow.