Abstract
Tasks of image recognition become important components for multimodal man-machine interface. For developing feasible components, problems of huge dimensionality and non-linearity must be resolved. We have been applying Self-Organizing Map (SOM) to feature representation stage of lip-reading and face recognition, and have appealed the advantages of SOM for dimensionality reduction and nonlinear feature representation. However, tasks of trial and error to decide the appropriate dimensionality of SOM can be a difficulty for developing image recognition. For this problem, we propose a dimensionality estimation method for SOM by using spectral clustering (SC). SC also has a characteristic of non-linear topographic mapping, and its fruitage suggests the dimensionality of feature space. In the section of experimental results, we will show relations between estimated dimensionalities of SOM and total recognition accuracies. The results emphasize feasibility of this proposed method.