A Novel Video Classification Method Based on Hybrid Generative/Discriminative Models

We consider the problem of automatically classifying videos into predefined categories based on the analysis of their audio contents. In detail, given a set of labeled videos (such as news, sitcoms, sports, etc.), our objective is to classify a new video into one of these categories. To solve this problem, a novel audio features based video classification method combining an unsupervised generative model named probabilistic Latent Semantic Analysis (pLSA) with a multi-class discriminative classifier is proposed. Since general audio signals usually show complicated distribution in the feature space, k-means clustering method is firstly used to group temporal signal segments with similar low-level features into natural clusters, which are adopted as “audio words”. Then, the audio stream of a video is decomposed into a bag of “audio words”. To classify those bags of “audio words” which extracted from videos, latent “topics” are discovered by pLSA, and subsequently, training a multi-class classifier on the “topic” distribution vector for each video. Encouraging classification results have been achieved in our experiments.