摘要：

数字技术的进步使得用户更便捷地在互联网上上传和下载数据，但是由此引发了盗取伪造、非法滥用等问题。数字水印技术可以有效遏制滥用数据的行为。针对现有视频水印算法在面对视频帧攻击时抵抗能力差的问题，提出一种结合张量分解与视频场景分割的鲁棒视频水印算法，提高嵌入水印图像后载体视频的不可见性。将载体视频分割为若干个视频场景，并将视频场景从RGB颜色空间转换到YCbCr颜色空间，根据视频场景中每帧的Y分量构造高阶张量，得到相应的张量特征图并进行不重叠分块处理。根据特征图子块联合熵确定嵌入水印的子块，利用奇异值分解对选定块分解得到U矩阵，并将水印信息嵌入到U矩阵中，通过对水印图像进行Arnold置乱处理，增加水印图像的安全性。实验结果表明，在载体视频“suzie.avi”的MPSNR值为55 dB以上时，该算法对不同攻击参数的高斯噪声、椒盐噪声、泊松噪声、模糊、帧丢失、帧平均以及帧交换等攻击都具有较强的鲁棒性，归一化互相关系数均在0.938之上。

关键词: 视频水印, 张量分解, 场景分割, 联合熵, 奇异值分解

Abstract:

The advancement of digital technology has made it easy for users to upload and download data on the Internet, leading to prevalent issues of theft, forgery, and illegal abuse.To effectively address data abuse, digital watermarking technology has emerged. This paper proposes a robust video watermarking method that combines tensor decomposition and video scene segmentation, aiming to improve the resistance to video frame attacks and enhance the invisibility of watermarking images embedded in a carrier video.The proposed approach involves dividing the carrier video into multiple video scenes.Each video scene is then converted from the RGB color space to the YCbCr color space.High-order tensors are constructed based on the Y component of each frame within the video scene, generating the corresponding tensor feature map.Non-overlapping block processing is applied, and the sub-blocks for watermarking embedding are determined based on the joint entropy size of the feature map sub-blocks. Singular Value Decomposition(SVD) is used to decompose the selected block to obtain the U matrix, and the watermarking information is embedded into the U matrix.Additionally, Arnold scrambling is applied on the watermarking image to enhance its security.The experimental results demonstrate the strong robustness of the proposed algorithm against various attacks such as Gaussian noise, salt and pepper noise, Poisson noise, blurring, frame loss, frame averaging, and frame swapping when the MPSNR value of the carrier video "suzie.avi" exceeds 55 dB. The NCC(Normalized Cross-Correlation Coefficient) values for all attacks are consistently above 0.938.

Key words: video watermarking, tensor decomposition, scene segmentation, joint entropy, Singular Value Decomposition(SVD)