A novel signature for fast video retrieval

Abstract

Many video signatures have been used in video copy detection and retrieval, but most of them are extracted from frames of video. So the time cost would become intolerable when the video database become large enough. To implement the fast retrieval or copy detection on large scale video database, we propose a novel video signature based on data count fluctuation independent upon video frame content just relevant about frame data count for compress domain MPEG video, which first calculate the data count for each frame on compressed domain as signature, then slide the window in the step of GOP(Group of Picture) length after the alignment of I frame, and then compare the difference of data count fluctuation between the query clip and the target video, at last the similarity result was given. This signature could be used both in video database retrieval and online video matching. The experiments show that the novel signature based method is much faster than current retrieval method, and the recall can achieve a high level.