Multi-scale Attentive Residual Network for Single Image Deraining

Removing rain streaks from a single image is extremely challenging since the appearance of rain streaks in shapes, scales and densities is ever changing. Therefore, we propose a novel end-to-end two-stage multi-scale attentive residual network that is both location-aware and density-aware, in order to preferably remove various rain streaks. Specifically, in the first stage, a multi-scale progressive attention sub-network is designed to automatically locate the distribution of diverse rain streaks and further to guide the following deraining. Then the second stage with the guidance of the attention map generated in the former stage aims to efficiently remove various rain streaks. To aggregate the characteristics of rain streaks with different scales and densities, we construct a multi-scale residual sub-network in which dilated convolution and residual learning are used to combine these features. As a result, these two sub-networks make up the whole network, and accomplish the process of joint detection and removal of diverse rain streaks fairly well. Extensive experiments on both synthetic and real-world rainy images demonstrate that our proposed method significantly outperforms several recent state-of-the-art approaches.