Multi-Scale Dense Selective Kernel Spatial Attention Network for Single Image De-raining

Abstract

Different rain streaks attached to a single image will affect the visibility of the image. According to earlier researches, it is a challenging problem for single image rain streaks removal due to the various orientations and scales of rain streaks. We propose a multi-scale dense selective kernel spatial attention network for single image de-raining. To extract the information of the features from rain streaks effectively and accurately, we adopt selective kernel units which allow each neuron to adaptively adjust its receptive field with the multi-scale manner on the channel dimension. To better guide the feature extraction, we utilize spatial attention blocks which focus on the informative parts of spatial map via the multi-scale mechanism on the spatial dimension. Additionally, we further exploit densely connected network architecture to strengthen the feature reuse and ensure the useful information propagate persistently. We conduct extensive experiments on several datasets to demonstrate the considerable performance improvement of our proposed method over other state-of-the-art methods. Simultaneously, we present an ablation study to indicate the performance improvement obtained by different modules of our proposed method.