DUAL RECURSIVE NETWORK FOR FAST IMAGE DERAINING

Introduction

In this paper, we propose a dual recursive network (DRN) for fast image deraining as well as comparable or superior deraining performance compared with state-of-the-art approaches. Specifically, our DRN utilizes a residual network (ResNet) with only 2 residual blocks (ResBlock), which is recursively unfolded to remove rain streaks in multiple stages. Meanwhile, the 2 ResBlocks can be recursively computed in one stage, forming the dual recursive network. Experimental results show that DRN is very computationally efficient and can achieve favorable deraining results on both synthetic and real rainy images.

Prerequisites

• 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)
• MATLAB for computing evaluation metrics

Datasets

DRN is evaluated on three datasets*: Rain100H [1], Rain100L [1] and Rain12 [2]. Please download the testing datasets from BaiduYun, and place the unzipped folders into ./test/.

To train the models, please download training datasets: RainTrainH [1] and RainTrainL [1] from BaiduYun, and place the unzipped folders into ./train/.

*We note that: (i) The datasets in the website of [1] seem to be modified. But the models and results in recent papers are all based on the previous version, and thus we upload the original training and testing datasets to BaiduYun. (ii) For RainTrainH, we strictly exclude 546 rainy images that have the same background contents with testing images. Our DRN is trained on remaining 1,254 training samples.

Getting Started

1) Testing

We have placed our pre-trained models into ./logs/.

test on Rain100H :

python test_Rain100H.py 

test on Rain100L :

python test_Rain100L.py 

test on Rain12 :

python test_Rain12.py 

We have placed four real rainy images into ./test/real, you can test on real rainy images :

python test_real.py 

2) Training

Read the configuration guide for more information on model configuration.

train the model for Rain100H:

python train_Rain100H.py

train the model for Rain100L:

python train_Rain100L.py

3) Evaluation metrics

We also provide the MATLAB scripts to compute the average PSNR and SSIM values reported in the paper.

 cd ./statistics
 run statistic_Rain100H.m
 run statistic_Rain100L.m
 run statistic_Rain12.m

Model Configuration

The following tables provide the configurations of options.

Training Mode Configurations

Option	Default	Description
batchSize	16	Training batch size
intra_iter	7	Number of intra iteration
inter_iter	7	Number of inter iteration
epochs	100	Number of training epochs
milestone	[30,50,80]	When to decay learning rate
lr	1e-3	Initial learning rate
save_freq	1	save intermediate model
use_GPU	True	use GPU or not
gpu_id	0	GPU id

Testing Mode Configurations

Option	Default	Description
use_GPU	True	use GPU or not
gpu_id	0	GPU id
inter_iter	7	Number of unfolding stages
intra_iter	7	Number of recursive ResBlock

References

[1] Yang W, Tan RT, Feng J, Liu J, Guo Z, Yan S. Deep joint rain detection and removal from a single image. In IEEE CVPR 2017.

[2] Li Y, Tan RT, Guo X, Lu J, Brown MS. Rain streak removal using layer priors. In IEEE CVPR 2016.