【作者】 邹晓春；

【导师】 冯燕；

【作者基本信息】 西北工业大学 ， 信号与信息处理， 2005， 硕士

【摘要】 序列图象的运动估计作为视频处理和计算机视觉的关键技术,已在社会生产生活的各个方面显示出越来越不可忽视的地位和作用,尤其在工业、商业、国防、医学、网络技术等领域,具有广泛的应用价值。近年来,块匹配(block matching)技术由于计算复杂度低及很高的压缩效率,被许多视频编码标准(如ISOMPEG-1/2以及ITU-T H.261/263) 所采用,在理论研究和实践应用中得到了不断的发展和完善,成为数字视频技术的一个研究热点。本文在深入分析,全面总结当前块匹配技术研究的基础上,提出了一种基于运动矢量分布特性的块匹配算法,并将其应用于基于小波变换的分层块匹配算法中。试验结果充分表明,在提高精度、降低计算复杂度等方面,新方法有着明显的优越性。论文首先充分回顾了几种经典的运动估计方法,其中基于块的匹配算法,方法简单,且估计效果较好,具有满足系统实时计算的潜力。在理解块匹配问题本质的基础上,系统地归纳总结了目前国内外现有的块匹配算法。根据二维对数搜索法、交叉搜索法、基于块的梯度下降搜索法和菱形搜索法等算法所采用的不同搜索模板与搜索策略,分析了它们的适用范围及优缺点。在基于运动矢量分布特性的运动估计方面,本文首先介绍了序列图象运动矢量的分布特性,即中心倾向分布特性、运动相关性和SAD值的方向性,这是本文研究的基础。然后对现有起点预测方法、匹配准则一一作了分析,并在此基础上提出了改进方案--基于运动相关性的起点预测方法和基于视觉特征的运动估计匹配准则。针对现有搜索模板与搜索策略对小运动块进行估计时造成的运算冗余,本文又提出了菱形、正方形相结合的DSM模板及沿SAD值下降方向进行搜索的策略。综合利用上述准则、方法、搜索模板与搜索策略的优点,设计了基于运动矢量分布特性的块匹配算法--快速块匹配LCB算法,并在算法计算复杂度、解的精确度方面进行了比较,实验结果验证了新方法的正确性和有效性。提出了一种改进的基于小波变换的分层块匹配算法。首先,介绍了小波变换和多尺度间的关系,并对二维小波变换的分解与重构算法和二维图象的多尺度表示进行了简要介绍。分层运动估计理论是多尺度理论和小波理论在序列图象分析中的应用,它利用小波变换良好的多尺度特性,得到不同分辨率的子图象,并在不同尺度上进行块匹配。改进方法是先在低分辨率下利用第三章所介绍的基于运动矢量分布特性的运动估计算法进行块匹配,然后在高分辨率下进行更精确的估计。实验表明,这种方法既减少了迭代时落入局部最小点的可能性,又提高了运算速度,因而也能获得比在单一尺度上更好的处理效果。更多还原

【Abstract】 As the key technique of computer vision and digital video processing, motion estimation of sequence images shows more and more important position and role not to be ignored in the various directions of production and life in society and has wide application value in industry, commerce, country defense, medicine, internet technology, etc. Recently, due to its low computational complexity and high compression efficiency, block matching technique has been adopted by many video coding standards such as ISO MPEG-1/2 and ITU-T H.261/263, and gets great improvements and rapid developments in methodology and application and becomes a new research hot point of digital video technique. On the base of comprehensive summarization and deep analysis for current block matching technique research work, a block matching algorithm based on motion vector distribution has been proposed in this thesis and has been applied to hierarchical block-matching algorithm based on the wavelet transform. Experimental results show the new algorithm has obvious advantage in respect of enhencing accuracy and reducing the computational complexity, etc.At the beginning, a fully survey about motion estimation algorithms is given. Matching algorithm based on block becomes to be considered, owing to its simplistic method, good estimation efficiency and potential to be implemented for the real-time application using the available hardware. On the base of understanding for block matching problem in nature, a systematic summarize is made for all current block matching algorithm in the world. From the different search algorithm and search criteria adopted by two-dimension logarithmic search algorithm, cross search algorithm, gradient descend search algorithm based on block, and diamond search algorithm, etc, we have analysed their applicability, and compared them in terms of computational complexity.In the aspect of motion estimation based on motion vector distribution characteristics, we firstly introduce the motion vector distribution characteristics of sequence images, i.e., center biased distribution characteristics of motion vector, the correlation among the motions of adjacent blocks and the direction of SAD value. This is basic of this thesis. Then the analysis is made to the current start search point prediction method, matching criterion. Taking good advantaging of these ideas, a developing algorithm, start search point prediction method based on motion correspondence and motion estimation matching criterion based on vision feature, is proposed. Aiming to the computational redundancy which is made by the current search mask and search strategy using in micro-block motion estimation, a new DSM mask which combines the diamond and square mask and a new search strategy which is along the descend of the SAD value is proposed . Taking good advantaging of these criteria, method, search mask and search strategy, a new block matching algorithm, fast block matching LCB algorithm, which is based on motion vector distribution characteristics, is exploited. At last, the computational complexity and accuracy of the solution is compared with existing algorithms. Experiment result shows the good quality performance, efficient and robust of the new algorithm.A improved method of hierarchical block matching motion estimation based on the wavelet transform is presented. Firstly, the relation of the wavelet transform and the multiscale is mentioned and a concision introduction of 2D wavelet decompose and reconstruction algorithm and multiscale discription of 2D image are made. Hierarchicalmotion estimation theory is the application of multiscale theory and the wavelet theory in the sequence image analysis. It used multiscale feature of the wavelet to get sub-images which own different resolution. Then block matching in the different scale is done. The developed method firstly uses the motion estimation algorithm based on motion vector distribution characteristic which is mentioned in the chapter 3 to do block matching in the low resolution, then does more precise estimat更多还原