Authors:
Júlia Rodrigues
and
Joel Carbonera
Affiliation:
Institute of Informatics, Federal University of Rio Grande do Sul, Porto Alegre, Brazil
Keyword(s):
Graph Neural Networks, Image Classification, Superpixels, Graph Convolutional Networks.
Abstract:
Graph Neural Networks (GNNs) is an approach that allows applying deep learning techniques to non-euclidean data such as graphs and manifolds. Over the past few years, graph convolutional networks (GCNs), a specific kind of GNN, have been applied to image classification problems. In order to apply this approach to image classification tasks, images should be represented as graphs. This process usually involves over-segmenting images in non-regular regions called superpixels. Thus, superpixels are mapped to graph nodes that are characterized by features representing the superpixel information and are connected to other nodes. However, there are many ways of transforming images into graphs. This paper focuses on the use of graph convolutional networks in image classification problems for images over-segmented into superpixels. We systematically evaluate the impact of different approaches for representing images as graphs in the performance achieved by a GCN model. Namely, we analyze the
degree of segmentation, the set of features chosen to represent each su-perpixel as a node, and the method for building the edges between nodes. We concluded that the performance is positively impacted when increasing the number of nodes, considering rich sets of features, and considering only connections between similar regions in the resulting graph.
(More)