Complexity scalable and robust motion estimation for video compression

Unrestricted Thanks to the fast development of network technology, transmission of high quality multimedia data becomes essential. However, the growth of data transferring capability is not always matched with the growth of bandwidth, specifically for wireless mobile environment. This is why the multimedia compression is so important for information technology development. Due to this importance, multimedia compression standards such as JPEG and JPEG2000 for still image compression and ISO/IEC MPEG-1, MPEG-2, MPEG-4, ITU-T H.261, H.263, H.263+, and JVT H.264/MPEG-4 AVC for video compression have been developed since the early 90's.; Video compression is achieved through computationally complex encoding operations. Among these, the motion estimation / compensation is most complex. The complexity and the memoryrequirement is further increased in the long-term memory motion compensation (LTMC) that significantly improves coding efficiency by utilizing multiple reference frames. Therefore, in this dissertation, we propose low complexity motion estimation algorithms which use information of previously encoded and multiresolution frames to speed up the search. The main novelty of our proposed work comes from defining search and complexity reduction techniques that are optimized for LTMC. Also, we propose an efficient memory management control technique toreduce the decoder memory requirement for LTMC. For this we design a novel greedy search algorithm which searches for a subset of reference frames that results in minimal performance degradation rather than checking all the combination of reference frames as the optimal solution does.; Also, we propose a novel system-level error tolerance scheme specifically targeted for multimedia compression algorithms. While current manufacturing process classifies fabricated systems intotwo classes, namely, perfect and imperfect, our proposed scheme employs categories which are based on acceptable / unacceptable performance degradation. By enabling the use of systems that would otherwise have been discarded we seek to increase the overall yield rate in the system fabrication process. To achieve this, we propose hardware testing algorithms that aim at determining if faults in a given chip produce acceptable performance degradation, and a technique that can cancel the effect of those among the acceptable faults that can be compensated.

得益于网络技术的飞速发展，高质量多媒体数据的传输已成为信息技术领域的迫切需求。然而，数据传输能力的增长并非总能与带宽扩容的速度相匹配，在无线移动环境中这一矛盾尤为凸显。这也正是多媒体压缩技术对信息技术发展至关重要的核心原因。自20世纪90年代初起，各类多媒体压缩标准陆续推出：静态图像压缩方向涵盖JPEG、JPEG2000；视频压缩方向则包括ISO/IEC MPEG-1、MPEG-2、MPEG-4、ITU-T H.261、H.263、H.263+以及JVT H.264/MPEG-4 AVC。

视频压缩依赖计算复杂度极高的编码操作实现，其中运动估计/补偿（motion estimation / compensation）环节的复杂度最为显著。而长期记忆运动补偿（long-term memory motion compensation, LTMC）通过引入多参考帧机制大幅提升了编码效率，但其自身的计算复杂度与内存占用需求进一步提升。为此，本论文提出一类低复杂度运动估计算法，通过利用先前编码帧与多分辨率帧的信息加速搜索流程。本研究的核心创新在于，定义了针对LTMC场景优化的搜索策略与复杂度降低技术；此外，我们还提出一种高效的内存管理控制技术，以降低解码器在LTMC模式下的内存需求。为此，我们设计了一种新型贪心搜索算法：该算法仅搜索能够带来最小性能损失的参考帧子集，而非如最优解法一般遍历所有参考帧组合。

与此同时，我们还提出一种专为多媒体压缩算法打造的新型系统级容错方案。当前的半导体制造工艺将已流片完成的系统划分为两类：完美良品与不良品。而我们提出的方案则采用基于“可接受性能损耗”与“不可接受性能损耗”的分类标准。通过复用本应被废弃的不良系统，我们旨在提升系统制造流程的整体良率。为达成这一目标，我们提出两类硬件测试算法：其一用于判断给定芯片中的故障是否仅会引发可接受范围内的性能损失；其二则可抵消可接受故障中可被补偿的部分所带来的影响。