Angle-Based Non-local Recurrent Network for Single Image Deraining

Abstract

Single image deraining is always an important problem in the field of image processing. With the development of deep learning over the years, image rain removal has also made substantial progress. However, the existing deep learning networks tend to regard rain streaks as a neat pattern, without considering the direction and physical characteristics of rain streaks, and they can not completely remove long rain streaks. To solve these problems, We proposed an Angle-based Nonlocal Recurrent Network(ANRN), which consists of Angle information extraction block and non-local recurrent convolution block. Angle information extraction block can effectively model rain streaks, and generates abstract feature maps and weights based on the amount of rain streaks, making full use of the direction and physical characteristics of rain streaks. In order to get rid of long rain streaks, we have used a non-local recurrent convolution structure, which can better connect long distance contextual information and reduce the size of the network. We did a large number of experiments in both synthetic and real data sets, and the results had shown that our network can remove rain streaks in various directions and densities, perform better on long rain streaks, and is superior to current state-of-the-art methods.