关键词: 视频水印, 低维流形, 流形学习, 几何攻击, AVS标准

Abstract:

How to resist geometric attacks is currently the focus and difficulty in video watermarking fields. This paper presents a video dual watermarking method which could effectively resist geometric attacks. The method mines a geometric deformation invariant, namely the low-dimensional manifold of video, and theoretically derives and experimentally demonstrates geometric invariance. First of all, the method generates zero-watermark or dynamic semantics watermark online according to low-dimensional manifolds of different video shots. In order to balance the robustness and transparency, the method embeds watermark in the IF DCT coefficients of AVS predicted residuals with large energy. This paper studies deeply the three-dimensional space-time characteristics of video, effectively quantifies and characterizes video motion information by low-dimensional manifold of video,establishes the human dynamic masking model including movement features, brightness features and texture features, so as to adaptively control watermark embedding strength. Experiments and analysis show that this method can effectively resist center cutting, irregular cutting, row cutting, rotation, scaling, translation and other high intensity geometric attacks and combined attacks, such as rotation and corner cutting, center cutting and rotation, etc. It also has strong robustness to the conventional video signal processing and various video attacks.

Key words: video watermarking, low-dimensional manifold, manifold learning, geometric attack, audio video coding standard (AVS)