Residual Contextual Hourglass Network for Single-Image Deraining

Abstract

Abstract Rain streaks could blur and distort images, significantly impacting further image processing. Single-image deraining is a hotspot and has practical application value, while most existing methods still have problems such as residual rain streaks and inadequate recovery of detail textures. To address these issues, we propose a Residual Contextual Hourglass Network (RCHNet) for single-image deraining, which could adapt to remove rain streaks in complex environments. Firstly, a contextual distillation block is presented to obtain local and global features across different scales. Further, residual downsampling block and residual upsampling block are used to maintain the residual nature of our architecture and better restore the details of the image. Finally, a dual attention mechanism is introduced to compensate for the spatial information lost by the downsampling. Extensive experiments on five synthetic datasets and a real-world dataset demonstrate that our proposed RCHNet outperforms existing state-of-the-art deraining approaches. On average across all synthetic datasets, the PSNR score of RCHNet is as high as 33.31 dB.