A Novel Framework of Robust Video Watermarking Based on Statistical Model

Abstract

This paper is to investigate a novel framework of robust video watermarking based on the statistical model with robustness against multiple attacks. The main contribution is threefold. First, the Laplacian distribution is proposed to model each naive video frame, referring to as the original frame; meanwhile the noisy frame, referring to as the one with adding Gaussian-distributed noise, is modeled using the Gaussian distribution. Second, we propose a novel mechanism of embedding watermark by artificially adding noise or not, corresponding to watermark bit 1 or 0. Third, it is proposed to cast the problem of watermark extraction into the framework of hypothesis testing theory. In the ideal context, with knowing all the model parameters, the Likelihood Ratio Test (LRT) is smoothly established with verifying the feasibility of the designed watermark extraction based on the statistical models. In the case of estimating model parameters, we propose to design the Generalized Likelihood Ratio Test (GLRT) to deal with the practical problem of watermark extraction. Finally, compared with some prior arts, extensive experimental results show that our proposed novel framework of robust video watermarking can achieve the high video quality with robustness against various attacks such as re-scaling, crop**, and compression.