Single-image de-raining with a connected multi-stream neural network

Abstract

Single-image de-raining is extremely challenging, because rainy images may contain rain streaks with various shapes, and at differing scales and densities. In this paper, we propose a new connected multi-stream neural network for removing rain streaks. In order to better extract rain streaks under different conditions, we use three dense networks with different kernel sizes that can efficiently capture the rain information at different densities. We show that providing useful additional information helps the network to effectively learn about the rain streaks. To guide the removal of rain streaks, we utilize a high pass filter to generate a rain region feature map, which focuses on the structure of the rain streaks and ignores the background in the image. Experiments illustrate that the proposed method significantly improves the removal of rain streaks in both synthetic images and real-world images.