Coarse-to-fine based matching for audio commercial recognition

In this paper, an automatic audio commercial recognition based on coarse-to-fine matching strategy is proposed. First, to index audio commercials in the database efficiently, the locality sensitive hash (LSH) is applied to accelerating the initial retrieval procedure. Second, with respect to the collision occurred in the initial LSH matching, a fine matching procedure based on Fine Granularity Successive Elimination (FGSE) is triggered out to eliminate rapidly those irrelevant candidates which have passed the initial coarse matching. At last, an elaborate post processing with correlation in time domain is presented for further decreasing false alarms following fine matching. The promising experiment results show that satisfactory precision and recall related to audio commercial recognition are obtained, achieving 99.4% and 97.5% respectively.