Recursive-Multi-Scale-Image-Deraining-With-Sub-Pixel-Convolution-Based-Feature-Fusion-and-Context Aggregation

In this paper, we propose a new single image deraining architecture with competitive deraining performance. Specifically we use a recursive technique of deraining with two modules cascaded each other. The first module -called front-end module- is dedicated to remove the rain at coarse level and is based on dense fusion of lower label features followed by sub-pixel convolutions. The second module -called refinement module- is dedicated to further remove the remnant of rain streaks and is based on context aggregation networks. The overall architecture is trained end to end and is capable to produce high-quality results -on both real-world and synthetic datasets- that are superior to several famous techniques in the literature.

Download pdf

Requirements:

-Ubuntu >=16.04
-Python3.6
-Pytorch >=0.4
-opencv-python, tensorboardX

Training:

-Download the dataset Rain100H, Rain100L, Rain12, DDN dataset and put them in subsequent folder inside Datasets/train/ and /Datasets/test/.
-Set the data and model saving directories in the script "train.py". -CD to the master folder and run the command for training as given in commands.txt file (python3.6 train.py).

Testing:

• Put the test dataset in /Datasets/test/ directory.
• Run the command "python3.6 test.py".

Results:

Results comparison on various state-of-the-art techniques on several synthetic(top) and real-world(bottom) datasets.

Acknowledgement:

--Training and data processing scripts are taken from pytorch implementation of PreNet. We are very much thankful for the authors for sharing their codes.