Structural recognition with kernelized softassign

Abstract

In this paper we address the problem of graph matching and graph classification through a kernelized version of the classical Softassign method. Our previous experiments with random-generated graphs have suggested that weighting the Softassign quadratic cost function with distributional information coming from kernel computations on graphs yields a slower decay of matching performance with increasing graph corruption. Here, we test this approach in the context of automatically building graph prototypes and classifying graphs in terms of the distance to the closer prototype. In all cases we use unweighted graphs coming from real images and having sizes from 30 to 140 nodes. Our results show that this approach, consisting on applying graph kernel engineering to matching problems, has a practical use for image classification in terms of pure structural information. ©Springer-Verlag Berlin Heidelberg 2004.