EfficientDerain

we propose EfficientDerain for high-efficiency single-image deraining

<img align="center" src="./results/structure.png" swidth="750">

Requirements

• python 3.6
• pytorch 1.6.0
• opencv-python 4.4.0.44
• scikit-image 0.17.2
• torchvision 0.9.1
• pytorch-msssim 0.2.1

Datasets

• Rain100L-old_version https://github.com/nnUyi/DerainZoo/blob/master/DerainDatasets.md
• Rain100H-old_version https://github.com/nnUyi/DerainZoo/blob/master/DerainDatasets.md
• Rain1400 https://xueyangfu.github.io/projects/cvpr2017.html
• SPA https://stevewongv.github.io/derain-project.html

Pretrained models

Here is the urls of pretrained models (includes v3_rain100H, v3_rain1400, v3_SPA, v4_rain100H, v4_rain1400, v4_SPA) :

direct download: http://www.xujuefei.com/models_effderain.zip

google drive: https://drive.google.com/file/d/1OBAIG4su6vIPEimTX7PNuQTxZDjtCUD8/view?usp=sharing

baiduyun: https://pan.baidu.com/s/1kFWP-b3tD8Ms7VCBj9f1kw (pwd: vr3g)

Train

• The code shown corresponds to version v3, for v4 change the value of argument "rainaug" in file "./train_*.sh" to the "true" (train_*.sh means it's the training script of dataset *)
• Unzip the "Streaks_Garg06.zip" in the "./rainmix"
• Change the value of argument "baseroot" in file "./train.sh" to the path of training data
• Edit the function "get_files" in file "./utils" according to the format of the training data
• Execute

sh train.sh

Test

• The code shown corresponds to version v3
• Change the value of argument "load_name" in file "./test.sh" to the path of pretained model
• Change the value of argument "baseroot" in file "./test.sh" to the path of testing data
• Edit the function "get_files" in file "./utils" according to the format of the testing data
• Execute

sh test.sh

Results

The specific results can be found in “./results/data/DERAIN.xlsx”

<img align="center" src="./results/psnr_ssim-time.png" swidth="750"> <img align="center" src="./results/table-ssim_psnr.png" swidth="750"> <table> <tr> <td ><center><img src="./results/gt_vs_rcdnet_0.gif" > <p align="center">GT vs RCDNet</p> </center></td> <td ><center><img src="./results/gt_vs_efderain_0.gif" > <p align="center">GT vs EfDeRain</p> </center></td> <td ><center><img src="./results/input_vs_gt_0.gif" > <p align="center">Input vs GT</p> </center></td> </tr> <tr> <td ><center><img src="./results/gt_vs_rcdnet_1.gif" > <p align="center">GT vs RCDNet</p> </center></td> <td ><center><img src="./results/gt_vs_efderain_1.gif" > <p align="center">GT vs EfDeRain</p> </center></td> <td ><center><img src="./results/input_vs_gt_1.gif" > <p align="center">Input vs GT</p> </center></td> </tr> </table> <table> <tr> <td ><center><img src="./results/gt_vs_v1_0.gif" > <p align="center">GT vs v1</p> </center></td> <td ><center><img src="./results/gt_vs_v2_0.gif" > <p align="center">GT vs v2</p> </center></td> <td ><center><img src="./results/gt_vs_v3_0.gif" > <p align="center">GT vs v3</p> </center></td> <td ><center><img src="./results/gt_vs_v4_0.gif" > <p align="center">GT vs v4</p> </center></td> </tr> <tr> <td ><center><img src="./results/gt_vs_v1_1.gif" > <p align="center">GT vs v1</p> </center></td> <td ><center><img src="./results/gt_vs_v2_1.gif" > <p align="center">GT vs v2</p> </center></td> <td ><center><img src="./results/gt_vs_v3_1.gif" > <p align="center">GT vs v3</p> </center></td> <td ><center><img src="./results/gt_vs_v4_1.gif" > <p align="center">GT vs v4</p> </center></td> </tr> </table>

Bibtex

@inproceedings{guo2020efficientderain,
      title={EfficientDeRain: Learning Pixel-wise Dilation Filtering for High-Efficiency Single-Image Deraining}, 
      author={Qing Guo and Jingyang Sun and Felix Juefei-Xu and Lei Ma and Xiaofei Xie and Wei Feng and Yang Liu},
      year={2021},
      booktitle={AAAI}
}