Graph neural networks for image processing

Graph signal processing (GSP) has provided new powerful tools that are particularly suitable for visual data. 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. This chapter presents a few applications of graph convolutional neural networks (GCNNs) to visual data. It provides an overview of the most relevant supervised methods based on graph neural networks. GCNNs have provided an elegant and effective way to overcome these limitations and have been successfully applied in point cloud processing. The chapter discusses image denoising, which is a long-standing problem in image processing. It presents two graph-convolutional generative models that can capture complex representations of the data without requiring supervisory signals. The chapter also presents an approach to the shape completion problem, that is, reconstruction of the missing parts of a 3D shape as a result of partial scans.