Graph-Based Contextual Attention Network for Single Image Deraining

Abstract

Rain streaks degrade the images and badly affect the outdoor vison tasks, and deep learning based single image deraining approach has witnessed the continuously growing and achieved great success. However, traditional convolution operation which uses a slide window can only extract the local feature patch, most of them ignore the correlation among the features for image deraining. We propose a graph-based contextual attention network for single image deraining. Firstly, we project the input feature map to a latent global context representation, and then the global information flow to spatial graph attention and channel graph attention, respectively. After the two graph attention, a grouped linear layer is adopted to mine the correlation among the information and enhance the global contextual information. The experiments on several datasets demonstrate that our methods achieve better results than the previous state-of-art methods.