Temporal Video Registration for Watermark Detection

Video frame alignment with the original source signal is required in a variety of applications, for instance non-blind watermark detection. Robustness to temporal cropping, scene removal, frame-rate changes and other non-linear temporal distortions are required. This paper presents an algorithm for locating a video segment within a long video document with frame-accurate alignment. The video sequences are first summarized into one-dimensional "temporal profile" signals. Then frame-to-frame correspondence is obtained by a dynamic programming approach. An automatic assessment of the matching result is proposed. Extensive test results achieved 77% success rate.