Lightweight image super-resolution network based on graph-based deep learning

Abstract

Recently, convolution neural network-based models dominated the field of computer vision. These models are based on using the image as a grid, which has some limitations in modeling irregular and complex objects. Based on this limitation, we proposed a graph super-resolution network (GCSRN) for the task of single image super-resolution. The idea of the model is to represent the image as a graph to extract graph-level features. The GCSRN can use advanced graph neural network research to solve low-level image tasks such as image super-resolution (SR). This graph creation is done by dividing the image into patches viewed as nodes and connecting the nearest neighbors to build a graph. So, this GCSRN model is transforming and exchanging information among all the nodes. The GCSRN is built based on a graph layer, and multi-layer perception is used to create a graph convolution block (GCB) that looks like the Transformer block. Then, based on the GCB, a deep graph convolution block is designed to be used as a building block for the GCSRN network. The model achieved state-of-the-art SR benchmark results with good visual quality and reconstruction accuracy.