LAUSR

In this paper, we propose a novel deep learning the based deraining method. The proposed method is motivated by the idea that an effective deraining algorithm should have the ability to remove various remaining rain streaks, which have been processed by the deraining method, in a repeated way. So, we design the deraining network in a coarse-to-fine manner that is multi-stage processing procedure and the parameters are shared in each stage. As the spatial contextual information is important for single image deraining, a densely connected dilation convolution block is proposed to deal with rain streaks with different sizes. Moreover, outer dense connections are used to guide the subsequent deraining procedures by fusing all the previous estimated rain-free images. The quantitative and qualitative experimental results demonstrate the superiority of the proposed method compared with recent state-of-the-art deraining methods on Rain100H, Rain1200, and Rain1400 datasets, while the number of parameters of our proposed method is greatly reduced due to the shared parameters strategy.