Stochastic models of video structure for program genre detection

Abstract

In this paper we introduce stochastic models that characterize the structure of typical television program genres. We show how video sequences can be represented using discrete-symbol sequences derived from shot features. We then use these sequences to build HMM and hybrid HMM-SCFG models which are used to automatically classify the sequences into genres. In contrast to previous methods for using SCGFs for video processing, we use unsupervised training without an a priori grammar. © Springer-Verlag Berlin Heidelberg 2003.