一个优化运动矢量码率分配的快速算法

从信息论的角度看,运动补偿的最终目的是极小化信源熵率。因此,对具有四叉树结构的运动矢量场,当使用叶部运动矢量替代相应根部运动矢量作运动补偿预测时,若编码运动矢量所增加的熵大于相应补偿帧差所减少的熵,则删除相应的叶部运动矢量,反之,则应使用叶部运动矢量替代根部的运动矢量。依据这一准则,该文通过对正向搜索的运动矢量场进行自下而上反向优化截断,得到了一个全局优化运动矢量场码率分配的快速算法。通过假设补偿帧差系数的选验概率分布,该文首先给出了一个闭式补偿帧差系数熵理论计算公式,考虑到视频压缩的有损特性,然后又对这个闭式理论公式作了带权修正,为了进一步提高运动矢量的编码效率,文中同时也给出了一种运动矢量融合方案。最后,该文在运动补偿的三维小波视频编码系统上对这个算法进行了验证和比较。模拟实验证实该文所提出的运动矢量码率分配算法不仅速度快而且收敛性态稳定,只要提供合适的搜索深度,总可以获得最佳的分配结果。由于优化算法避开了补偿帧差的具体后处理,因此该文的算法也可以应用到传统的DCT视频编码器。

A Fast Algorithm for Optimizing Bit Rate Allocation of Motion Vectors

Recently, many variable size region wise motion estimation techniques have been proposed to replace the conventional constant size block motion estimation, improve the performance of video codec, particularly at very low bit rates. In these region wise motion estimation approaches, Lagrangian multiplier method was commonly adopted to optimize the bit allocation of motion vector field under the constraint of the target bit rate. However the high computation complexity prevents them from being applied in practices. On analyzing the nature of motion compensation prediction, this paper proposes a fast algorithm of optimizing bit allocation for the quadtree structured motion vectors (MV) field based on the minimum entropy criterion, according to which the quadtree structured leaf MVs will be pruned if and only if the increased entropy of MVs is greater than the associated decreased entropy of compensation frame difference (CFD),otherwise, will replace the corresponding root MV. By adaptively pruning along the backward search path of the quadtree, the proposed fast algorithm reaches a globally optimal solution to this problem. Through assuming a prior distribution, a closed format of the theoretical entropy of CFD is presented. Considering the lossy nature of video coding, a modified scheme to the theoretical entropy of CFD is also described, in which theoretical entropy of CFD for each sub block is multiplied by a suitable weighting factor according to its position in the quadtree. To capture the high correlation among MVs, a merging approach of MVs is simultaneously described, in which the MV with minimum norm will be lifted to its parent node position and coded with higher priority. Simulation experiments on a motion compensated three dimensional wavelet video codec show that the proposed optimally pruning algorithm is not only fast but also robust. It will yield an optimal bit allocation across MVs and CFD if a suitable searching depth is provided. Avoiding the concrete post processing of CFD, the proposed algorithm can also be applied to the conventional DCT based hybrid video codec.