A non-supervised approach for repeated sequence detection in TV broadcast streams

In this paper, a novel method for repeated sequence detection in an audio-visual TV broadcast is proposed. This method is required for TV broadcast macro-segmentation which is at the root of many novel services related to TV broadcast and in particular to the TV-on-Demand service. Repeated sequence detection allows inter-program detection (commercials, jingles, credits, ……), which allows the segmentation of the TV broadcast and the extraction of useful programs. Our method is completely non-supervised, that is, it does not require a manually created reference database. It relies on a micro-clustering technique that groups similar audio/visual feature vectors. Clusters are then analyzed and repeated sequences are detected. This method is able to continuously analyze the TV broadcast and to periodically return analysis results. The efficiency and effectiveness of the method have been shown on two real broadcasts of 12 h and 7 days.