Dual attention residual group networks for single image deraining

Abstract

Single image deraining is one of challenges in image processing. An efficient algorithm for single image deraining can significantly improve the image quality in severe weather conditions. Existing deraining algorithms only pay attention to spatial characteristics or channel information, which leads to poor performance of the network. In this paper, we propose a novel dual attention residual group network (DARGNet) to get better deraining performance. Specifically, the framework of dual attention includes spatial attention and channel attention. The spatial attention can extract the multi-scale feature to adapt to different shapes and size of the rain streaks. Meanwhile, channel attention has established the dependence relationship among different channels. In addition, in order to simplify the structure, we integrate the dual attention module and convolution layers into the residual groups, which also improves information transmission. Extensive experiments on synthesized and real-world datasets demonstrate that the proposed network achieves a good effect of deraining tasks. The source code is available at https://github.com/zhanghai404.