Single image deraining has received considerable progress based on deep convolutional neural network. Most existing deep deraining methods follow residual learning in image denoising to learn rain streak layer, and perform limited in restoring background image layer. In this work, we propose bilateral recurrent network (BRN) to allow the interplay between rain streak and background image layers. In particular, two recurrent networks are coupled to simultaneously exploit these two layers. Instead of naive combination, we propose bilateral LSTMs, which not only can respectively propagate deep features across stages, but also bring the interplay between these two SRNs, which is essential in separating two layers from rainy observation. The experimental results demonstrate that our BRN notably outperforms state-of-the-art deep deraining networks on synthetic datasets quantitatively and qualitatively. The proposed method also performs more favorably in terms of generalization performance on real-world rainy dataset.
- Python 3.6, PyTorch >= 0.4.0
- Requirements: opencv-python, tensorboardX
- Platforms: Ubuntu 16.04, cuda-10.0 & cuDNN v-7.5
- MATLAB for computing evaluation metrics
SRN and BRN are evaluated on seven datasets*: Rain100H [1], Rain100L [1], RainHeavy*[5], RainLight*[5], Rain12 [2], Rain1400 [3] and SPA-data [4]. Please download the testing datasets from BaiduYun or OneDrive, download the RainHeavy*[5] and RainLight*[5] from here, and download the testing generalization dataset SPA-data [4] from GoogleDrive. And then place the unzipped folders into './datasets/'. Make sure that the path of the extracted file is consistent with '--data_path'.
*We note that:
(i) The datasets of Rain100H and Rain100L have been updated by the authors. We notate them as RainHeavy* and RainLight*, that can be downloaded from here.
(ii) We upload the old datasets of Rain100H and Rain100L to BaiduYun or OneDrive. For Rain100H, we strictly exclude 546 rainy images that have the same background contents with testing images.
We have placed our pre-trained models into ./logs/.
Run shell scripts to test the models:
bash test_RainHeavy.sh # test models on RainHeavy
bash test_RainLight.sh # test models on RainLight
bash test_Rain100H.sh # test models on Rain100H
bash test_Rain100L.sh # test models on Rain100L
bash test_Rain12.sh # test models on Rain12
bash test_Rain1400.sh # test models on Rain1400
bash test_real.sh # test models on SPA-data(i) On RainHeavy* [5] and RainLight* [5], we re-train all the competing methods. We have uploaded all the trained models to ./logs/RainHeavy/ and ./logs/RainLight/. You can use their source codes to reproduce the results in the paper.
(ii) All the results in the paper are also available at GoogleDrive. You can place the downloaded results into ./results/, and directly compute all the evaluation metrics in this paper.
We also provide the MATLAB scripts to compute the average PSNR and SSIM values reported in the paper.
cd ./statistic
run statistic_RainHeavy.m
run statistic_RainLight.m
run statistic_Rain100H.m
run statistic_Rain100L.m
run statistic_Rain12.m
run statistic_Rain1400.m
run statistic_real.m
python train.py --save_path path_to_save_trained_models --data_path path_to_training_dataset*If you use the new dataset by yourself, please make sure to define new function for preprocessing training patches in DerainDataset.py.
[1] Yang W, Tan R, Feng J, Liu J, Guo Z, and 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, and Brown M. Rain streak removal using layer priors. In IEEE CVPR 2016.
[3] Fu X, Huang J, Zeng D, Huang Y, Ding X, and Paisley J. Removing rain from single images via a deep detail network. In IEEE CVPR 2017.
[4] Wang T, Yang X, Xu K, Chen S, Zhang Q, and Lau R. Spatial attentive single-image deraining with a high quality real rain dataset. In IEEE CVPR 2019.
[5] Yang W, Tan R, Feng J, Liu J, Yan S, and Guo Z. Joint rain detection and removal from a single image with contextualized deep networks. IEEE T-PAMI 2019.