Three-Stage Channel Split Dense Fusion Network for Single Image Deraining

Single image deraining is a considerable challenge in the domain of image processing. The enhancement of image quality under adverse weather conditions heavily relies on the development of efficient single image rain streaks removal algorithms. However, prevailing rain streaks removal algorithms often suffer from an abundance of parameters. Furthermore, as the number of network layers increases, the extraction of image features tends to become redundant, detrimentally impacting both the efficacy of rain streaks removal and the preservation of background details. Therefore, a three-stage channel split dense fusion network (TCSDFNet) is proposed to obtain better rain removal performance. Specifically, multiple channel split blocks (CSB) are used in the network for dense feature fusion to gather the high and low level features of the rain streaks. CSB uses channel split operation to split the rainy image into multiple channels, and applies different rain streaks removal methods according to different levels of features to reduce redundant features and network parameters, and improve the performance ability and computational efficiency of the model. At the same time, the residual dual attention block (RDAB) is proposed, which is applied to the channel split to ensure that the background image will not be destroyed or blurred when the rain streaks are removed. In addition, due to the strong correlation between rain streaks, the network adopts a three-stage learning strategy in order to better capture the distribution of rain streaks in the image. TCSDFNet greatly reduces the number of parameters in the rain removal network, extracts the features of different levels more comprehensively, and reduces the loss of background image detail to a greater extent. A large number of experiments on synthetic datasets and real datasets show that the proposed network achieves good rain removal effect.