【作者】 张维纬；

【导师】 钮心忻；

【作者基本信息】 北京邮电大学 ， 信息安全， 2013， 博士

【摘要】 近年来,随着高清数字电视、视频点播等技术的迅速发展及多媒体共享网站的盛行,视频媒体正逐渐成为人们生活的重要组成部分。如何有效地对数字视频进行版权保护和内容认证,已成为当前多媒体安全领域研究的热点。视频数据量大,通常在存储和传输过程中需要压缩编码,H.264/AVC是当前市场最受欢迎的视频编解码标准,因此,研究基于H.264/AVC编解码标准的视频水印技术的需求极为迫切。本文针对当前基于H.264/AVC的视频水印算法的一些不足,结合H.264的压缩编码标准特性及人类视觉特性HVS(Human Visual System),重点研究了用于版权保护的视频鲁棒水印和用于内容认证的视频脆弱水印和半脆弱水印。具体的研究内容包括四个部分：基于纹理特征的差分能量视频水印算法、基于人眼可察觉变化步长-JND(Just noticeable difference)模型的鲁棒视频水印算法、基于内容特征的H.264实时可逆脆弱水印算法以及基于误差补偿策略的半脆弱水印算法,其中主要的工作和创新点如下：1.针对现有的一些基于H.264/AVC的视频水印算法只在宏块的1个系数上嵌入1比特水印,导致水印鲁棒性较差的问题,结合H.264压缩编码标准的特性,提出了一种新的基于纹理特征的差分能量视频水印算法。算法先对当前帧4×4块进行整数离散余弦变换,判断其是否是纹理块,再采用差分能量的方式自适应选择系数嵌入水印。所提的算法提高了水印的鲁棒性,能有效抵抗高斯噪声、低通滤波、重编码等常见的视频水印攻击。并且对视频质量和码率的影响较小。2.现有的基于8×8浮点DCT变换的JND模型不能直接应用于采用4×4整数变换的H.264/AVC视频水印算法中。首先分析了现有的基于4x4DCT变换的JND模型的不足,提出了一种新的更加精确的JND模型。由于并不是所有JND值大于量化值的整数DCT系数都适合嵌入水印,本文对水印的嵌入位置进行了研究,提出了在H.264/AVC编码过程中量化后的残差块DCT系数的直流和低频分量不适合嵌入水印的结论,并通过理论分析和实验说明了结论的合理性。为了增加水印的鲁棒性,根据不同频带的DCT系数,采用不同的水印嵌入和提取算法。所提算法的水印不仅满足不可见性和鲁棒性,嵌入容量也较高。3.已提出的基于H.264/AVC的视频脆弱水印算法大多数是不可逆的。分析了不可逆的原因,提出了一种基于内容特征的实时可逆脆弱水印算法。算法先计算当前宏块预测残差块量化的DCT系数的哈希值以生成认证码,再把认证码作为水印信息以可逆的方式嵌在下一个相邻宏块活性最大的4×4块的最后一个非零量化DCT系数中。在解码端,通过比较认证码和提取的水印信息进行视频数据完整性的认证和原始数据的还原。算法对视频质量和码率的影响较小,并且能对认证通过的视频数据进行还原和对认证失败的视频Ⅰ帧遭篡改区域进行有效的定位。4.为进一步提高半脆弱视频水印特征码的鲁棒性,通过消除特征码之间的相关性,提出了一种新的特征码生成算法。针对半脆弱水印要均匀嵌在视频图像帧中以实现准确定位篡改区域的要求,引入了一个最小失真选择函数,通过这个函数在残差4×4块中选择一个对视频质量和码率影响最小的DCT系数嵌入水印。和已有的算法相比,采用最小失真选择函数可以较显著地降低嵌入水印对视频质量和码率的影响,并且半脆弱水印对于可接受的“内容保持”类修改具有鲁棒性,而对“内容篡改”类修改具有敏感性,并且对恶意篡改能有效、准确地定位。更多还原

【Abstract】 In recent years, with the rapid development of HD digital TV and VOD (video on demand) technology accompanied by the popularity of multimedia sharing site, video multimedia has gradually become a significant part in people’s daily life. Thus, the efficient methods for copyright protection and content authentication of digital video have been research hotspots in academic field of multimedia security.Because of the large amount of data, video need to be compressed in the process of storing and transferring and H.264/AVC is the most popular encoding and decoding standard on market nowadays, which makes the demand for video watermarking based on H.264/AVC more and more urgent.Analysizing on the shortcomings of the existing video watermarking algorithms based on H.264/AVC and considering the properties of H.264/AVC and HVS (Human Visual System), this articlefocuses on video robust watermarking used for copyright protection and fragile watermarking and semi-fragile watermarking used for content authentication.The comprehend content of this research is consisted by energy-differential video watermarking algorithm based on textural features, video robust watermarking based on JND model, H.264real-time reversible fragile watermarking algorithm based on content features and semi-fragile watermarking algorithm based on error compensation strategy. And the main innovation points of this article could be concluded as following:Firstly, to solve the problem of robustness insufficiency of video watermarking based on H.264/AVC caused by only embedding lbit watermark in one coefficient of macro-block, this article put forward a new energy-differential video watermarking algorithm based on textural features by considering the properties of H.264compressing and encoding standard. The proposed algorithm firstly performs the4x4integer Discrete Cosine Transform on the current frame in purpose of identifying whether it is the texture block. Then the watermarking is embedded by adopting the way of differential energy for selecting the appropriate coefficients adaptively. The experimental result shows that the proposed algorithm has little impact on the video quality and bit-rate, moreover, its robustness can effectively resist the common attacks such as the gauss noise, low pass filter and re-coded performed on the video watermarking.Secondly, the existing JND models based on8x8floating point DCT transform can’t be directly applied to video watermarking algorithms based on H.264/AVC because of which adopts the4x4integer transform. After analyzing on the shortcomings of the existing JND models based on the4×4integer transform, the article proposes a new and more accurate JND model. Since not all of the JND values of integer DCT coefficients which arelarger than the quantization are suitable for embedding watermarks, after researching the embedding position for watermarks, we come to a conclusion that the DC and low-frequency component of the DCT coefficients of the quantized residual blocksare not suitable for embedding watermarks, and the theoretical analysis and experimentsshow that our conclusion is reasonable. In order to improve the robustness of watermarking, we propose different watermarking embedding and extraction algorithmsaccording to the different frequency ranges of DCT coefficients. The proposed watermarking algorithm not only meetsthe imperceptibility and robustness requirement, but also has highembedding capacity.Thirdly, the majority of existing fragile video watermarking algorithms based on H.264/AVC are irreversible. We analyze the reasons and propose a real-time reversible fragile video watermarking based on content features. The proposed algorithm firstly performs the hash value of quantized DCT coefficients of predictive residue blocks in current macro-block, and then creates the authentication code. And after that, embed the code into the last nonzero quantized DCT coefficient in the most active4×4block of the next adjacent macro-block as watermark information in the reversible manner. The integrality authentication is carried out and the original data isrestoredint thethe decoding end by the comparison of the authentication code and the extracted watermark information. The experimental result shows that the proposed algorithm has little impact on the video quality and bit-rate, moreover, the original video can be recovered if the authentication is successful and the tampered area of I frame can be effectively located if the authentication fail.Fourthly, to further improve the robustness of the feature codes in semi-fragile video watermarking algorithm, we propose a new generation algorithm for feature codes by eliminating the correlation of them. In order to meet the requirement that the semi-fragile watermarks should be uniformly embedded into the video frames to locate the tampered region accurately, this paper introduces a minimal distortion selection function to select a DCT coefficientin the4×4block residualto embed watermark which has minimal impact on the video quality and bit rates.compared with existing algorithms, the proposed algorithm can significantly reduce impact on video quality and bit rate because of embedding the watermarking by adopting the minimum distortion selection function. Experimental results show that the scheme has robustness against the content-preserving manipulations and it is sensitive to malicious manipulations. What’s more, the retrieved feature code and watermarking can detect and locate malicious manipulations.更多还原