【作者】 魏伟；

【导师】 侯正信；

【作者基本信息】 天津大学 ， 信号与信息处理， 2006， 硕士

【摘要】 本文主要讨论基于可变形块匹配(DBMA)的运动估计与补偿,用以减小传统方块匹配算法(BMA)的预测误差,提高帧间运动估计的准确性,实现对非平移运动的预测。提出了基于节点搜索的可变形块匹配算法(NS-DBMA),并在此基础上提出几个优化方法,进一步降低运算量和提高预测质量,弥补了BMA的不足。本文首先介绍视频压缩编码的系统结构,分析运动估计与补偿的基本原理与重要作用;介绍BMA的发展和几种有代表性的快速算法,分析快速BMA的共同特点并对加快匹配速度的方法进行总结。然后分析视频信号的实际属性,从根本上解释了BMA所采用的平移模型的缺陷,提出了适合实际运动特性的基于节点位移的可变形块运动模型。为了实现此模型的运动估计,首先介绍基于梯度的可变形块匹配算法(GB-DBMA),并指出GB-DBMA的预测精度和运算量都达不到实际应用的要求;进而提出NS-DBMA,它能有效提高预测精度,并降低运算量,有效地预测帧间非平移运动。在NS-DBMA的基础上,进一步提出了计算过程的优化方法、结合BMA和DBMA的双模式混合估计方法、在预测质量足够好时结束搜索的适可而止估计方法、使用简化点阵的交叉搜索算法和提高预测质量的分数像素精度预测算法,使DBMA更加适合实际应用的要求。测试结果表明,NS-DBMA的运算量仅为GB-DBMA的一半,而预测质量平均提高0.71dB,能得到更好的主客观预测质量,且更易于硬件实现;综合各种改进方法的DBMA的运算量只相当于全搜索法BMA的42%,而预测质量平均提高1.53dB,可以替代BMA应用在实际视频压缩系统中。更多还原

【Abstract】 Motion estimation and compensation using deformable block matching algorithm (DBMA) is discussed. DBMA is developed to reduce the prediction error of traditional block matching algorithm (BMA), that is, to increase the accuracy of motion estimation and to estimate the nontranslational motion. Here we propose a Nodal-search based DBMA. Some ameliorative methods are further proposed, which can alleviate the computation and improve the quality more effectively.This thesis first describes the video compression system to indicate the importance of motion estimation and compensation. Then the famous BMA and some typical fast algorithms are introduced. Finally some fundamental characteristics of fast BMAs are summarized. By analyzing the property of real video sequence, the translational model adopted by BMA is demonstrated to be inaccurate in the interframe prediction, and a nodal-displacement-based deformation model is proposed. To estimate the nodal displacements, we developed a nodal search-based algorithm (NS-DBMA), which has lower computation and better performance than original gradient-based algorithm (GB-DBMA). Some improved methods are further proposed, such as the two-mode hybrid method which combines BMA and DBMA, the auto-terminate method which stops searching when the prediction is good enough, the cross-search method which uses simple search pattern, the sub-pixel accuracy prediction which denotes motion more accurately and computing optimization method.Experimental results show that the basic NS-DBMA has only half computation of GB-DBMA but gets better subjective and objective quality, and is much easier to implement by VLSI. The most effective method, which synthesizes all ameliorative methods, has only 42% computation of the Full Search while outperforms 1.53dB. It could be used as a substitute for BMA in real coding system.更多还原