Latent semantic indexing for video content modeling and analysis

An extension of LSI work to index and compare video shots in a large video database and the distributions of LSI features among semantic classes are estimated to detect concepts present in video shots.

Abstract

In this paper, we propose to adapt latent semantic index-ing (LSI) to model video contents. This well-known tech-nic used to describe text documents provides a rich and efficient representation of the content. We include in the original method a way to efficiently include various fea-tures extracted from the video content. In particular we focus on color and texture features. The distributions of LSI features among semantic classes is then estimated to detect concepts present in video shots. K-Nearest Neigh-bors and Gaussian Mixture Model classifiers are evalu-ated and compared. Finally, performances obtained on LSI features are compared to a direct approach based on raw features that are color histograms and Gabor’s filter energies.