Video Coding Using Multihypothesis Motion Compensation in the Redundant Wavelet Domain

在视频编码中，多假设运动补偿（multi - hypothesis motion compensation， MHMC）是由多个运动补偿的假设预测组成的预测，以对抗运动估计过程中固有的不确定性。MHMC技术——例如那些选择在空间上多样化的预测（例如，分数像素预测），以及那些利用时间多样性（例如，双向预测）的技术——已经形成了现代视频编码器的一个组成部分。本研究探讨了一种新的MHMC形式——冗余小波多假设（RWMH）——它在过完备或冗余小波变换域中执行运动补偿，以产生变换阶段不同的运动补偿预测。变换的冗余产生了不同的阶段，从不同的角度看待运动，每个阶段都对运动的真实性质做出了自己的假设，导致率失真性能明显优于传统的单相系统。本研究的重点是将RWMH纳入现代视频编码系统，以证明RWMH技术与其他先进的视频编码技术相辅相成。具体考虑的是将RWMH纳入由运动补偿预测、变换和量化组成的传统混合系统，即最先进的H.264参考模型。此外，RMHW被部署到采用运动补偿时间滤波（MCTF）的现代3D编码器中，以提供完整的时间、分辨率和保真度可扩展性。在这两个系统中，RWMH的优点是可以用更多的假设来预测低分辨率的信息，这对应于在低空间分辨率下估计运动的更大困难。这项工作的广泛影响源于这样一个事实，即所有生成的源代码都在开源许可下作为大型库的一部分提供给视频编码社区。

ABSTRACT0310864James E. FowlerMississippi State UIn video coding, multihypothesis motion compensation (MHMC) forms a prediction composed of multiple motion-compensated hypothesis predictions in an effort to combat the uncertainty inherent in the motion-estimation process. MHMC techniques---such as those that choose predictions that are diverse spatially (e.g., fractional-pixel prediction), as well as those exploiting temporal diversity (e.g., bidirectional prediction)---have formed an integral part of modern video coders. This research investigates a new form of MHMC---redundant wavelet multihypothesis (RWMH)---which performs motion compensation in the domain of an overcomplete, or redundant, wavelet transform in order to produce motion-compensated predictions that are diverse in transform phase. The redundancy of the transform yields distinct phases that view motion from different perspectives with each phase contributing its own hypothesis as to the true nature of the motion, resulting in rate-distortion performance significantly superior to traditional single-phase systems.The present research focuses on the incorporation of RWMH into modern video-coding systems in order to demonstrate that the RWMH technique functions complementary to other advanced video-coding techniques. Specifically considered is the incorporation of RWMH into a traditional hybrid system consisting of motion-compensating prediction, transform, and quantization, namely the state-of-the-art H.264 reference model. Additionally, RMHW is deployed into modern 3D coders employing motion-compensated temporal filtering (MCTF) in an effort to provide full temporal, resolution, and fidelity scalability. In both systems, RWMH has the benefit that lower-resolution information is predicted with a greater number of hypotheses, corresponding to the greater difficulty inherent in estimating motion at low spatial resolution. Broader impacts of the work stem from the fact that all resulting source code is provided to the video-coding community as part of a large library under an open-source license.