Detecting Movie Segments Using Gaussian Mixture Models for VOD Lectures with Japanese Subtitles

Abstract

A search system for VOD (video on demand) lectures is useful if it can be applied beyond searches of only text. To facilitate better searching for movie segments to be used in VOD lectures with Japanese subtitles, we propose a method using subtitles and a solving maximum likelihood estimation from a mixture of Gaussian distributions. The detection was performed by a statistical method by using the expectation-maximization algorithm. This allows for the detecting of movie segments and determining their number. In addition to improving evaluation of movie segments, we provide movie segment rankings. Movie segment rankings are computed for each movie segment using a method that removes one Gaussian distribution from a mixture of Gaussian distributions.