SRGNet: A GRU Based Feature Fusion Network for Image Denoising

Abstract

Image denoising is an essential pretreatment for most of image processing pipelines, which has been extensively studied for several decades. Recently, convolutional neural networks with skip connections show promising performances on image denoising due to their discriminative feature modeling and utilizing of features from former layers. However, they only apply coarse feature fusion strategies like pixel-wise addition or concatenation which are insufficient in image denoising. In this work, we propose a novel network architecture to exploit finer feature fusion. Specifically, a module based on gate recurrent unit is introduced into the architecture, fusing features from different layers and adopting finer feature selection at the same time. Experiments on multiple challenging datasets show the effectiveness of the proposed network.