Graph Spectral Image Processing

10Graph Neural Networks for Image Processing

Giulia FRACASTORO and Diego VALSESIA

Politecnico di Torino, Turin, Italy

10.1. Introduction

Processing visual data can be a daunting task due to the complexity and nuances of their representations. For this reason, decades of research have made great strides in defining more and more sophisticated models. Traditionally, such methods have been focused on images, which are still a major focus of the research in the field. However, new visual data types, such as 3D point clouds, are becoming increasingly relevant and extending image processing techniques to effectively process them provides a new set of challenges.

Recently, graph signal processing (GSP) has provided new powerful tools that are particularly suitable for visual data. On the one hand, traditional signal types, such as images, can benefit from richer representations induced by the graph structure (Liu et al. 2014; Kheradmand and Milanfar 2014; Bai et al. 2018; Pang and Cheung 2017). On the other hand, new data types with an irregular domain, such as point clouds, can now be effectively processed (Zeng et al. 2019; Chen et al. 2018; Thanou et al. 2016).

Concurrent to the emergence of GSP, data-driven solutions, based on neural networks have shown impressive performances in a variety of tasks, including low-level tasks, such as image restoration. The workhorse of data-driven methods is the convolutional neural network (CNN), which has been shown to capture highly complex ...