GD$^2$ASR: Graph Based Dual Domain Attention for Image Super-Resolution
In this paper, we proposed a graph-based dual-domain attention network for the super-resolution task (GD$^2$ASR), including the graph domain and the CNN domain. The graph domain attention is related to a recent class of neural networks operation on graphs. Instead, we design a graph representation to leverage the natural connections among pixels, which are crucial for understanding the structural relationships. Inspired by the advantages of Graph Convolution Networks (GCNs) on learning the global contextual information, we present a deep network structure to combine the attention features from the CNN domain and GCN domain. To implement the projection of the features between the graph domain and CNN domain, a GCNs-based module is designed to extract features in the graph domain, and to convert the feature vectors onto the coordinate space. Furthermore, a dual attention fusion scheme is to further combine the graph domain attention and the CNN domain attention. Experimental results verify the superiority of the proposed GD$^2$ASR model, performing favorably against some state-of-the-art methods in terms of both PSNR and SSIM.