Rain-Density Squeeze-and-Excitation Residual Network for Single Image Rain-removal

Abstract

The removal of rain streaks in a single image is an extremely challenging task due to the uneven rainfall density in the image. Methods based on deep learning have boosted the performance of rain removal significantly in recent years. However, most of these methods have a certain demand for different density of rain marks in the training data, which prevent them to further improve the performance in some outdoor scenarios. In this paper, we present a novel Rain-Density Squeeze-and-Excitation residual network (RDSER-NET), which adopts the squeeze-and-excitation blocks into the ResNet framework. The proposed network remove rain streaks based on single density of rain marks in the training data, reducing the limitation of multi-density proposals and achieving better results. Extensive experiments on synthetic and real datasets demonstrate that the proposed network outperform the recent state-of-the-art methods greatly.