Image super-resolution via channel attention and spatial graph convolutional network

Abstract

Recently, deep convolutional neural networks (CNNs) have been widely explored in single image super-resolution (SISR) and obtained remarkable performance. However, most of the existing CNN-based SR methods mainly focus on wider or deeper architecture design, neglecting to discover the latent relationship of features, hence limiting the representational ability of networks. To address this issue, we propose a channel attention and spatial graph convolutional network (CASGCN) for more powerful feature obtaining and feature correlations modeling. The CASGCN is formed by several channel attention and spatial graph (CASG) blocks that incorporate global spatial and channel inter-dependencies for rendering features of each pixel. Inside the CASG block, channel branch and spatial branch are first arranged in a paralleled way, and then are concatenated to effectively learn the representation of each image pixel. Specifically, we use attention mechanism to extract informative features in channel branch while the spatial-aware graph is used in spatial branch to model the global self-similar information. Furthermore, the adjacency matrix in spatial-aware graph is dynamically generated via the Gram matrix to model global correlations between pixels and is shared across the whole network without auxiliary parameters. Extensive experiments on SISR with different degradation models show the effectiveness of our CASGCN in terms of quantitative and visual results.