基于最短描述长度的序列图像运动分割

MDL BASED MOTION ESTIMATION AND SEGMENTATION FOR LOW BIT RATE VIDEO CODING

提出了一种分两个阶段的运动估计和分割方法．首先采用小块对图像做传统全局搜索的块匹配运动估计，得到每块相应搜索范围内的误差图．根据误差图判决相邻块运动的一致性，将一致性好的块合并成一个区域并同时得到区域的运动矢量．进一步的区域合并在最短描述长度准则指导下进行，每对区域合并的依据是合并后新区域的描述长度（编码码字）小于合并前两区域描述长度之和，如此迭代合并直至所有相邻区域均不满足条件．由于迭代过程总是趋向于使编码码字减少，从而能得到更有利于编码的分割结果．初步实验证明了该方法的可行性．

In this paper, motion estimation and segmentation based on the region merging and the minimum description length (MDL) principle is studied. The proposed technique is an unsupervised bottom up merging method which could be separated into two parts: initial segmentation and further segmentation. In initial segmentation, a small block is used to fulfill the full searching block matching motion estimation to obtain an error map and the initial motion vector of the current block. Two neighbor regions with good motion coherence are merged according to their error maps, and the motion vector for the newly merged region is obtained via the new error map generated by the two old ones. The MDL principle is employed in the further region fusion. The MDL based criteria incorporates bit rate minimization into the merging criterion, which will generally give more satisfying results. In every step of iterative operation, the two neighbor regions will be merged if the sum of their description length is larger than that of the fused region. Further fusion will stop if the up mentioned condition is in no circumstance satisfied. The preliminary experiments show that promising results can be reached in the viewpoint of the video coding.