Macro Segmentation and Content Analysis of TV Broadcast Stream

Abstract

This study addresses a non-supervised approach to extract TV programs via repetition based detection of the Inter-Programs (IPs) and audio based segmentation and classification algorithm to analyze the massive raw TV stream. Acoustic and visual information are both adopted for IPs detection so as to avoid missing true-positive. Novel audio fingerprints scheme and shot based indexing algorithm are introduced to guarantee the efficient and superior detection performance. After the TV programs are further segmented into clips, Gaussian Mixture Models (GMMs) are used to classify the clips into three types, namely, pure speech, non-pure speech, and non-speech. Experiments on a test dataset composed of more than 500 hours content-unknown TV streams show that the F-measure of the programs extraction and content analysis achieve 0.986 and 0.887 respectively. The experiments also demonstrate that the proposed algorithm for detecting repeated IPs outperforms the state-of-art approach.