Content Based Robust Video Fingerprinting

With the rapid development of the multimedia technology, computer technologyand network technology, the data volume of the digital video information in the world isexploding. Digital video information provide greatly convenient for processing, copyand modification, while significantly increase trasnsmission rate and expressionefficiency. However, because of the openness of the internet and digital videoprocessing software, the copyright protection and the management of video content is anew challenge. Content based video fingerprting which is a new and potentialtechnology is absorbing widespread attention by the academic and business.In general, the video fingerprints need to satisfy some properties like robustness,pairwise independence and search efficiency. The main work of this paper is listed asfollows:First, a robust video fingerprinting based on shape feature is proposed. Thepyramid histogram of gradient orientation extracted in the downsampled video frames isused as video fingerprint. The method solves the underutilized deficiency of the shapeinformation in the existing methods. The experimental results show that the proposedscheme is robust against scale transformation, adding noise and global attacks. Theperformance of pairwise independence and search efficiency are also better thantraditional methods.Second, a robust video fingerprinting based on spatio-temporal features isproposed. In the input video clips the spatio-temporal interest points are detected. Foreach local region around the interest points, weighted contrast histogram is used tocalculate the intensity differences, and the unit vector obtained by normalizing thehistogram is used as the local fingerprints. The experimental results show that theproposed scheme haves obvious advantage in robustness and pairwise independencecompared with existing methods.Third, an improved algorithm combining global features and local features isproposed to enhance the robustness. The shape features and spatio-temporal featuresextracted and normalized at the same time are weighted combined as the fingerprint. Inthe experiments, the improved algorithm performs better than the two previousproposed schemes.