【作者】 马行；

【导师】 陈贺新；

【作者基本信息】 吉林大学 ， 通信与信息系统， 2009， 博士

【摘要】 随着立体视频的广泛应用,大量的视频数据需要存储和传输,如何对立体视频数据进行高效的压缩编码是当前研究的热点问题。本文以经过校正的立体视频/图像为研究对象,提出以下几种立体视频压缩方法。论文首先研究了马尔科夫随机场建立视差图模型求立体匹配的问题,针对置信度传播信息时具有时变自适应支持域的特点,提出一种基于时域相关性的置信度传播立体视频压缩算法。然后针对立体匹配容易在对象边界部分出现较大误差,从而影响恢复图像主观效果的问题,提出一种依据纵向亮度连续性约束准则,将图像每列划分为互不交叠的亮度段,接下来以对象边界的视差值为参考,确定每条亮度段自适应搜索范围的立体视频压缩算法。最后提出一种将立体视频构建在四维矩阵中,选择代表帧进行运动估计,然后利用四维矩阵DCT变换进一步去除时域、空域以及双目之间冗余的立体视频压缩算法。根据四维矩阵划分方式不同,分为规则型四维矩阵立体视频压缩算法和基于内容的四维矩阵立体视频压缩算法两种。实验结果表明,对于真实的自然场景,视图加深度的立体视频压缩方式由于缺乏遮挡/暴露区的残差补偿,图像的恢复质量将受到明显影响,基于内容的四维矩阵立体视频压缩算法具有更高的压缩效果。更多还原

【Abstract】 2D video is incomparable with stereo video, which could provide viewers depth sense according to stereo disparity theory. By prevailing of stereo video, enormous of video data need to be stored and transmitted, especially in real-time system, high resolution and multi points-points communications. Recently, 2D encoding theory and technology were already maturity, but not in stereo video and multi points. How to compressing stereo video data became a hotspot in communication and it has also important for application of stereo video. Nowadays, stereo video compression has been paid attention by many researchers.DISTIMA project was an integrated stereo video communication system, which was based on MPEG-2 coding standard and developed by Germany, France and some other European nations. ATTEST proposed a novel stereo video compression that was based on a more flexible joint transmission of monoscopic color video and associated per-pixel depth information. From this 3D data representation format, one or more“virtual”views of a 3D scene can then be synthesized in real-time at the receiver side by means of so-called depth-image-based rendering (DIBR) techniques. International standard organization MPEG built a stereo video coding special team, which mainly discussed stereo video requirements and techniques in apply, and for purpused of a uniform stereo video standard. There were two stereo video schemes in all, which was depth based image rendering, and another one was block-based estimation with residual compensation.This paper mainly researched on stereo video coding in 3DTV system, and a pair of rectified images with small baseline were as input. Three stereo video compression algorithms were proposed and organized as follows.1. More exactly disparity estimation could get smaller compression stream and finer quality of reconstructed image by depth based image rendering. Loop belief propagation could get exactly matching by minimizing global energy, but it would take heavy computational cost. This paper proposed a temporal correlation based belief propagation (BP) stereo video coding, which was based on BP’s message passing provides a time-varying adaptive support region for stereo matching to deal with textureless regions and depth discontinuities elegantly. In textureless regions, for example, the influence of a message can be passed far away. On the other hand, the influence in discontinuous regions will fall off quickly. Stereo video were organized in group. Disparity of every I stereo pair was calculated by standard loop BP, and disparity of every P stereo pair was computed selectively by difference of temporal frames. Finally, reference video and disparity sequences were encoded.2. If algorithms of disparity estimation were not referred objects, there will be large residual around object edges. And then this would be effected subjective quality of reconstructed image. So this paper proposed a disparity estimation algorithm based on objects.Firstly, mainly edges of object were extracted by Canny, then feature points were detected by Harris, and original matching relation was constructed, finally outlier disparities were deleted by fundamental matrix estimation processing. And a sparse disparity map was constructed, and then disparity in object needed to be calculated.It is a difficult problem for selecting unit for calculating disparity of object. It will be too dispersed to be calculated by pixel-based and it is also difficult for selecting size of block by block-based. Because of disparity discontinuous always occurs with large changes in intensity, this paper proposed every column of image was adaptable divided into non-overlapping segments by continuous constraint of intensity in vertical, and every segment was considered as a unit for matching. And then size of searching windows was referenced by disparity of left and right adjacent object edges. Finally, reference and disparity sequences were encoded.3. Due to view points, intensity source, and noise effect, some information could not be reconstructed by depth base image rendering, but they could be remedied by residual compensation. This paper proposed a content-based four dimensional (4D) matrix stereo video algorithm. Differ from typical block-based method, this motion estimation was calculated by reference frames, which would reduce almost estimation computation. And then redundant of temporal, spatial and disparity correlation was reduced by 4D matrix DCT. Coefficients after transforming were assembled in low frequency and temporal, and then correlation between coefficients could be enhanced by 4D matrix Z scan, which was fitable for variable length coding.Several stereo video and image were tested, and experimental results were compared between these methods.In contrast of standard loop BP algorithm, temporal correlation based BP stereo video coding algorithm could reduce 83.8% computation time under the same reconstructed image quality. So it was an efficient disparity sequence algorithm for stereo video transmitting and free view point rendering. Experiment results showed that this algorithm could preserve details well, but some errors occured in edge of object. This might affect stereo sense because it could not be remedied in further by lacking residual compensation process.Object and continuous constraint of intensity based stereo video coding could reserve well in edges and details, but not in discontinuous intensity. Object edges could be reconstructed more integrity by this method than temporal correlation based BP stereo video coding.Stereo video could be reconstructed well by 4D matrix based stereo video coding. With accretion of bit rate, the quality of reconstructed image could get higher. By compared with the other two algorithm proposed in this paper, 4D matrix based stereo video coding may be blockness in textureless region at lower bit rate, but it has higher fidelity, which reserved original image information well and may be helpful for reconstructed stereo sense.Experiment results showed that it would be finer quality of reconstructed image by depth based image rendering for simple content stereo video or image. But for complexity content stereo video, residual compensation could resolve errors by occlusion and exposure well.更多还原