Least-commitment graph matching with genetic algorithms

Abstract

This paper concerns the correspondence matching of ambiguous feature sets extracted from images. The first contribution made in this paper is to extend Wilson and Hancock's Bayesian matching framework (Wilson and Hancock, IEEE Trans. Pattern Anal. Mach. Intell. 19 (1997) 634–648) by considering the case where the feature measurements are ambiguous. The second contribution is the development of a multimodal evolutionary optimisation framework which is capable of simultaneously producing several good alternative solutions. Previous multimodal genetic algorithms have required additional parameters to be added to a method which is already over-parameterised. The algorithm presented in this paper requires no extra parameters: solution yields are maximised by removing bias in the selection step, while optimisation performance is maintained by a local search step. This framework is in principle applicable to any multimodal optimisation problem where local search performs well. An experimental study demonstrates the effectiveness of the new approach on synthetic and real data.