MCAD-Net: Multi-scale Coordinate Attention Dense Network for Single Image Deraining

Pengpeng Li, Jiyu Jin, Guiyue Jin, Jiaqi Shi, Lei Fan

Abstract: Single image rain removal is an urgent and challenging task. Images acquired under natural conditions are often affected by rain, which leads to a serious decline in the visual quality of images and hinders some practical applications. Therefore, the research of image rain removal has attracted much attention. However, both the model based method and the deep learning based method can not adapt to the spatial and channel changes of rain feature information. In order to solve these problems, this paper proposes an end-to-end Multi-scale Coordinate Attention Dense Network (MCAD-Net) for single image deraining. MCAD-Net can accurately identify and characterize rain streaks and remove them, while preserving image details. To better solve the problem, the Multi-scale Coordinate Attention Block (MCAB) is introduced into the MCAD-Net to improve the ability of feature extraction and representation of rain streaks. MCAB first uses different convolution kernels to extract and fuse multi-scale rain streaks features, and then uses the coordination attention module with adaptation module to recognize rain streaks in different spaces and channels. A large number of experiments have been carried out on several commonly used synthetic datasets and real datasets. The quantitative and qualitative results show that the proposed method is superior to the recent state-of-the-art methods in improving the performance of image rain removal and preserving image details.