Three-Stage Progressive Residual Network For Single Image Deraining

The information of image obtained in the rainy day is incomplete and even seriously affects subsequent application and processing. To solve this kind of problem, we proposed a three-stage progressive residual image deraining network. The network consists of Long Short-Term Memory depth-wise convolution network (LSTMDNet), dilated dense depth residual network (DDDRNet), multi-scale detail reconstruction network (MSDRNet) three modules. The first module utilizes Long Short-Term Memory (LSTM) and depth-wise convolution residual structure to extract global and local features. The second module uses depth-wise convolution structure to further extract image space features and reduce network computation, combining dilated convolution structure to enlarge the receptive field and improve adaptability to rain streaks, in addition, network introduces dense connection to aggregate the features of different depth. The third module designs a multi-scale parallel convolution structure to interact and fuse the context information of the recovered images to improve and optimize the details. Experiments show our network has good and efficient effect.