Commercial Recognition in TV Streams Using Coarse-to-Fine Matching Strategy

In this paper, a coarse-to-fine matching strategy and corresponding approaches are proposed to address the two key research issues in commercial recognition, which are robust signature generation and efficient indexing structure. Specially, aiming at resisting the visual perception distortion, some novel and robust video content signatures are extracted by exploiting the global and local photometric and spatial properties. Then, to introduce a coarse-to-fine indexing structure, Locality Sensitive Hash (LSH) is applied to accelerate the initial coarse retrieval procedure. And the Fine Granularity Successive Elimination (FGSE) is triggered to eliminate rapidly those irrelevant candidates which have passed the coarse matching process. In addition, a heuristic temporal consistency maintenance approach is presented for further decreasing false alarms after fine matching process in regard to the temporal correlation. The promising experimental results show the effectiveness of the proposed strategy with respect to large video data collections.