A fast procedure for the computation of similarities between Gaussian HMMS

An appropriate definition and efficient computation of similarity (or distance) measures between stochastic models are of theoretical and practical interest. In this work a similarity measure for Gaussian hidden Markov models is introduced based on the generalized probability product kernel. An efficient scheme for computing the similarity measure is presented. The out of precision problem, which is a significant implementation issue, is considered and a scaling procedure is provided. The effectiveness of the proposed method has been evaluated on texture classification and preliminary experimental results are presented.