Embedding Non-Local Mean in Squeeze-and-Excitation Network for Single Image Deraining

Abstract

Images captured in rainy outdoor usually have poor visual quality due to the appearance of raindrops blur or rain streaks in the image. For many practical vision systems, such as autonomous driving and video surveillance, this problem is urgently required to be solved. In this work, a novel network for single image de-raining has been proposed. The proposed network consists of three stages, encoder stage, Dense Non-Local Residual Block (DNLRB) stage, and decoder stage. To better capture spatial contextual information, which has been analyzed to be meaningful for image de-raining, we adopt squeeze-and-excitation enhancing on feature maps in each convolution layer. We also embed non-local mean operations in DNLRB, which effectively leverages spatial contextual information for extracting rain components. Quantitative and qualitative experimental results demonstrate the superiority of the proposed method compared with the state-of-the-art de-raining methods.