鲁棒视频水印算法及相关技术研究-智能与分布计算实验室

鲁棒视频水印算法及相关技术研究

姓名	邹复好
论文答辩日期	2006.05.09
论文提交日期	2008.03.24
论文级别	博士
中文题名	鲁棒视频水印算法及相关技术研究
英文题名	Research of Robust Video Watermarking Algorithms and Related Techniques
导师1	卢正鼎
导师2
中文关键词	数字水印;视频水印;实时性;均值参考水印;差分能量水印;水印容量;信道编码;Turbo码
英文关键词	Digital Watermark;Video Watermarking;Real-Time;Mean Value Reference Watermarking;Differential Energy Watermarking;Watermarking Capacity;Channel Coding;Turbo Code 论文总页码 111
中文文摘	目前，数字水印作为多媒体数据的版权保护、内容认证、拷贝控制、操作跟踪和广播监控等应用的一项重要技术，正受到国内外学者的广泛关注。在现存的水印研究文献中，图像水印远多于视频水印。然而，在实际应用中，视频水印更加有用，因而应获得更多的关注。尽管，现存的一些图像水印算法可以直接扩展到视频应用中，但视频水印并不是图像水印简单延伸，视频水印应满足一些特殊要求，如实时性、随机检测性、盲水印等。通常，一个好的视频水印方案应该至少同时满足实时性和鲁棒性要求。在现存的视频水印文献中，尽管大多数视频水印算法有很好的实时性，但这些算法为了满足实时性要求，通常忽略了鲁棒性。为此提出一个能同时满足实时性和鲁棒性的算法----均值参考视频水印(MRW)算法。MRW算法设计时使用了有助于提高水印鲁棒性的三种策略：水印嵌在人类视觉系统感知最重要的直流系数部分、使用Watson视觉模型控制水印嵌入达到最大强度以及选择有助于产生稳定均值点的扩展M序作为水印模式。同时，MRW算法直接运行在VLC域，可以避免一些复杂性计算，如逆DCT变换、DCT变换、运动补偿等，因而MRW算法的复杂性较低。然而，MRW算法中，存在某些水印子区域的能量与均值偏差大，水印嵌入时，如果强行把该水印子区域的能量调制到预定的值，势必会产生视觉失真。针对此问题，提出了优化的均值参考视频水印(OMRW)算法，OMRW算法给出两种解决方法：随机置乱和牺牲检测响应值。与MRW算法相比，OMRW算法在不降低水印系统的整体性能的前提下，进一步提高了嵌入水印的视频的视觉质量。然而，MRW和OMRW算法的水印模式限定为扩展M序，该水印模式只能对用户进行唯一表示，并不表示任何具体的含义。但是，视频水印有些应用要求水印信号是有意义的信息。因此又提出一个对水印模式没有任何限制的水印算法??低频系数域差分能量水印(L-DEW)算法。L-DEW算法借鉴差分能量水印算法能量差思想，在DCT块的低频系数域嵌入水印信息。由于要在DCT块的低频系数域产生一个能量差，必须回答两个问题：一是能否在低频系数域构建能量近似的水印子区域；二是如何在低频系数域构建能量近似的水印子区域。针对问题一，专门对量化的DCT系数进行统计建模，建模的结果表明低频系数空间存在大量“碎片”，可以在低频系数域构造能量近似的水印子区域。针对问题二，多维均衡水印子空间构造研究给出了解决此问题的两种算法：遗传算法和ILPT算法，实验表明ILPT算法更适合视频水印。在这两个问题成功解决的基础之上，L-DEW算法使用ILPT算法在DCT块的低频系数域构造能量近似的水印子区域，并在Watson视觉模型控制下，对低频系数细微修改以在两个水印子区域间产生一个能量差。通过理论分析和实验检验，表明在视觉质量、水印容量、鲁棒性和水印提取用时等性能指标上，L-DEW算法均优于差分能量水印算法。对L-DEW算法的信道分析基础上，把该水印算法信道建模为二进制对称信道，并给出求解该信道容量的方法。然后，在L-DEW算法中引入RS码和Turbo码，进一步提高水印系统的可靠性。实验结果表明，Turbo码纠错性能更好，但其计算复杂性较高。
英文文摘	In recent years, digital watermark is regarded as one of key techniques for some applications, such as copyright protection, content authentication, copy protection, fingerprinting and broadcast monitoring, and this makes digital watermark receive a plenty of attention from a number of researchers. Among the existing watermarking literature, image watermarking algorithms are far more than video’s. However, video watermarking algorithms are more useful than image’s, and should be received more attentions. Although the existing image watermarking concept can be extended to video watermarking, the video watermarking can not be considered as the simple extension of image watermarking. Video watermarking should be met with some special requirements, such as real-time, random detection and blind extraction. Usually, a perfect video scheme should be at least satisfied with real-time and robustness property simultaneously. In the existing video watermarking literatures, the robustness of video watermarking algorithms usually be neglected in order to meet with real-time property. So, a video watermarking algorithm (so-called mean value reference watermarking, MRW) with real-time and robustness property is proposed. To satisfy the robust requirement, the MRW algorithm adopts three strategies: watermark is embedded into the perceptual significant component, the Watson model is used to control the watermarking strength to maximum and the extended m-sequences is selected as watermark pattern to generate stable mean reference point. In addition, the MRW algorithm is directly performed in the VLC domain and can avoid some complex computations, such as inverse DCT, DCT and motion compensating. Then the MRW algorithm’s complexity is very low. However, there still exist some sub-region’s energy is far away from the mean value, it will lead to visual quality distortion when the watermark embedding function mandatorily modulates those sub-region’s energy to a predefined value. To solve this problem, we propose an optimized mean value watermarking algorithm (OMRW). The OMRW algorithm presents two methods, which are shuffle and decrease detection response value, for this problem specially. Compared with the MRW algorithm, the OMRW algorithm has a better visual quality. However, the MRW and OMRW algorithms’ watermark pattern is restricted as the extended m-sequences, which is only identified the user and dose not represent any detailed meaning. Some applications of the video watermarking require watermark pattern to represent some meaningful information. So we propose a differential energy watermarking in the low frequency domain (L-DEW), whose watermarking pattern has not any restrictions. The L-DEW algorithm borrows the energy difference concept from the DEW algorithm and embeds watermark in the low frequency coefficients. Due to enforce an energy difference in the low frequency domain, it must confront with two problems: one is whether it can construct a set of sub-regions with approximate energy; another is how to construct a set of sub-regions with approximate energy. To answer the first problem, we perform a study of statistical model of quantized DCT coefficients. The study results show that there exist a great deal of fragments (i.e. small value coefficient) in the low frequency domain, and this represents it can construct approximate energy sub-regions with low frequency coefficients. To answer the second problem, we perform a research of constructing multi-dimension approximate energy sub-regions. In this research, we present genetic algorithm and improved LPT algorithm. The experiments show that the improved LPT algorithm more suitable for video watermarking. On basis of resolution to the above-mentioned two problems, the L-DEW algorithm first constructs a set of approximate energy sub-regions using the improved LPT algorithm, and then enforces an energy difference between every two sub-regions to embed watermarking bits under the control of the Watson visual model. From the theory analysis and experiments, it is obviously that the L-DEW is indeed better than DEW at the aspect of visual quality, data payload, robustness, and time cost of watermarking embedding. At last, this thesis first models the channel of L-DEW as binary symmetrical channel and then estimates the capacity toward this channel. Then, the RS code and Turbo code, which are suitable for high signal-noise ratio and low signal-noise ratio cases separately, are introduced into the watermarking system for improving the reliability. In the case of L-DEW scheme, Turbo code is better.