DM2F-Net-improved

By Yuqi Zhang

This repo is an improvement of "Deep Multi-Model Fusion for Single-Image Dehazing" (ICCV 2019), written by Zijun Deng at the South China University of Technology. The original repo can be found at here.

Results

Dataset	O-HAZE				HazeRD
Method	PSNR	SSIM	MSE	CIEDE	PSNR	SSIM	MSE	CIEDE
DM2F-Net	25.113	0.7742	0.0032	5.2000	14.212	0.8145	0.0724	16.8331
DM2F-Net-improved	25.602	0.7752	0.0030	4.9714	15.774	0.8291	0.0589	15.5905

The checkpoint and dehazing results can be found at Baidu Drive.

Installation & Preparation

Make sure you have Python>=3.7 installed on your machine.

Environment setup:

1. Create conda environment

   conda create -n dm2f
   conda activate dm2f
   

2. Install dependencies (test with PyTorch 2.3.0):

   1. Install pytorch==2.3.0 torchvision==0.18.0

   2. Install other dependencies

      pip install -r requirements.txt
      

• Prepare the dataset

  • Download the RESIDE dataset from the official webpage.

  • Download the O-Haze dataset from the official webpage.

  • Make a directory ./data and create a symbolic link for uncompressed data, e.g., ./data/RESIDE.

Training

1. Set the path of datasets in tools/config.py
2. Run by python train.py

Use pretrained ResNeXt (resnext101_32x8d) from torchvision.

Hyper-parameters of training were set at the top of train.py, and you can conveniently change them as you need.

Training a model on a single NVIDIA A100 GPU takes about 4 hours.

Testing

1. Set the path of five benchmark datasets in tools/config.py.
2. Put the trained model in ./ckpt/.
3. Run by python test.py

Settings of testing were set at the top of test.py, and you can conveniently change them as you need.

License

DM2F-Net-improved is released under the MIT license.

Citation

If you find the code helpful to your research, please give a star to me!