Image Super-Resolution Based on Gated Residual and Gated Convolution Networks

Single image super-resolution based on deep neural network has been a hot topic in recent years. In this paper, we propose a gated residual and gated convolution network (GRGCN) to deal with super-resolution reconstruction. Gated residual (GR) and gated convolution (GC) are the products of introducing attention mechanism (AM) into deep neural network. Specifically, GR enables the network to focus on more important residual channel information and GC allows the network to pay attention on more important feature extraction regions. Meanwhile, we propose multi-level residual connection, which includes long connection, secondary connection and short connection. Three different levels of residual connection correspond to three scales of feature extraction module, namely residual cluster, residual group and residual block. By fusing the feature information on different scales, the convolution layer can efficiently extract the feature information. Experiments on benchmark datasets demonstrate that our proposed GRGCN has better performance and visual effects than state-of-the-art methods.