Advances in Graph Matching for Image Interpretation

Abstract

In structural pattern recognition, graphs are a powerful and flexible data structure, allowing for the description of complex relationships between data elements. This flexibility comes at a cost, as the unconstrained nature of graphs results in a high computational complexity for graph matching algorithms. Various algorithms have been proposed to mitigate this complexity and make graph matching tractable. Additionally, in domains where the number of graph nodes is low, or where the data provides additional constraints, such as node and edge labels, graph matching has been effectively applied. Such applications include chemical structure matching, protein-protein interaction networks, and network analysis. In the domain of computer vision, graphs have been successfully applied to a number a problems including image segmentation, partitioning, and matching. However, for practical reasons, many of the image matching techniques that utilise graphs do not match graph topology directly. Instead, image graphs are used only to constrain feature locations, or spectral embedding is used to transform image graphs into vectors, which are then matched.