MSBDN-DFF

The source code of CVPR 2020 paper "Multi-Scale Boosted Dehazing Network with Dense Feature Fusion" by Hang Dong, Jinshan Pan, Zhe Hu, Xiang Lei, Xinyi Zhang, Fei Wang, Ming-Hsuan Yang

Updates

(2020.12.28) Releasing the training scripts and the improved model.

Dependencies

• Python 3.6
• PyTorch >= 1.1.0
• torchvision
• numpy
• skimage
• h5py
• MATLAB

Test

1. Download the Pretrained model on RESIDE and Test set to MSBDN-DFF/models and MSBDN-DFF/folder, respectively.

2. Run the MSBDN-DFF/test.py with cuda on command line:

   MSBDN-DFF/$python test.py --checkpoint path_to_pretrained_model

3. The dehazed images will be saved in the directory of the test set.

Train

We find the choices of training images play an important role during the training stage, so we offer the training set of HDF5 format: Baidu Yun (Password:v8ku)

1. Download the HDF5 files to path_to_dataset.

2. Run the MSBDN-DFF/train.py with cuda on command line:

   MSBDN-DFF/$python train.py --dataset path_to_dataset/RESIDE_HDF5_all/ --lr 1e-4 --batchSize 16 --model MSBDN-DFF-v1-1 --name MSBDN-DFF

3.(Optional) We also provide a more advanced model (MSBDN-RDFF) by adopting the Relaxtion Dense Feature Fusion (RDFF) module.

MSBDN-DFF/$python train.py --dataset path_to_dataset/RESIDE_HDF5_all/ --lr 1e-4 --batchSize 16 --model MSBDN-RDFF --name MSBDN-RDFF

By repalcing the DFF module with RDFF module, the MSBDN-RDFF outperforms the original MSBDN-DFF by a margin of 0.87 dB with less parameters on the SOTS dataset. More details will be released soon.

Citation

If you use these models in your research, please cite:

@conference{MSBDN-DFF,
    author = {Hang, Dong and Jinshan, Pan and Zhe, Hu and Xiang, Lei and Xinyi, Zhang and Fei, Wang and Ming-Hsuan, Yang},
    title = {Multi-Scale Boosted Dehazing Network with Dense Feature Fusion},
    booktitle = {CVPR},
    year = {2020}
}