Unsupervised Credit Detection in TV Broadcast Streams

This paper proposes an unsupervised method for detecting credits in TV streams. Identifying credits in TV streams allows to precisely determining boundaries of TV programs and hence, to extract specific and high valuable TV programs. The proposed detection solution is based on the temporal stability of opening and closing credits. Consequently, from a linear TV stream, we detect sequences that are broadcasted several times on a stable schedule with a clustering-based approach. These repeated sequences include opening and closing credits but also commercials, trailers etc. In order to select, among repeated sequences, those which are effectively credits, their temporal stability and their metadata consistency are analyzed. Since recurring programs are usually broadcasted at the same time(s) of the day, the temporal distribution of occurrences of each repeated sequence is studied. The Electronic Program Guide (EPG) is then used to validate the selected sequences and to distinguish between opening and closing credits. This method is entirely unsupervised and no prior information is required. Extensive experimental results validating our approach are presented.