Graph Kernels for Object Classification

Amal Mahboubi and Luc Brun


Graph kernels, object classification


Graphs constitute one of the few basic models used to encode structural properties of objects within the pattern recognition field. Definite positive graph kernels encode a similarity measure based on an implicit mapping from the set of graphs to some usually unknown Hilbert space. Graph kernels provide thus an implicit graph embedding hence allowing to combine structural and statistical pattern recognition fields. We propose in this paper, a graph kernel devoted to image indexation and based on oriented neighborhoods. Several heuristics proposed in this paper allow to adapt this kernel to different types of databases. The relevance of the proposed kernel together with the one of these heuristics is validated through several experiments on four public datasets.

Important Links:

Go Back