In this paper, we propose a novel dual attention residual group networks (DARGNet) for better deraining performance. Specifically, the proposed framework of dual attention includes spatial attention and channel attention. The spatial attention extracts the multiscale feature. Meanwhile, channel attention separates channel domain and spatial attention which can extract multiple attributes from different channels and guide the selection of the most important features. Furthermore, to simplify the structure, we integrate the dual attention module and convolution layers to the residual groups. The residual information can enhance the information flow. Extensive experiments on synthesized and real-world datasets verify the superiority of the proposed network over the state-of-the-art image deraining.

• Python 3.6, PyTorch >= 0.4.0
• Requirements: opencv-python, tensorboardX
• Platforms: Ubuntu 16.04, cuda-8.0 & cuDNN v-5.1 (higher versions also work well)

our model are evaluated on three datasets: Rain100H , Rain100L , Rain12 .
Please download the testing datasets from BaiduYun Access Code ：xbts
and place the unzipped folders into ./datasets/test/.

To train the models, please download training datasets:
RainTrainH , RainTrainL from BaiduYun Access Code ：xbts
and place the unzipped folders into ./datasets/train/

We have placed our pre-trained models into ./logs/. Run scripts to test the models:

python test_Rain100H.py   # test models on Rain100H
python test_Rain100L.py   # test models on Rain100L
python test_Rain12.py     # test models on Rain12
python test_real.py       # test models on real rainy images

All the results in the paper are also available at BaiduYun Access Code：bj1j

Run scripts to train the models:

python train_rain100H.py   # test models on Rain100H
python train_rain100H.py   # test models on Rain100L