Over-complete Signal Decomposition

CCR-9903368 Zakhor高效压缩图像和运动视频在各种新兴应用中正变得越来越重要。这些应用可以根据它们所需的比特率进行分类。广播质量的视频点播系统可能需要每秒几兆比特，而商业视频会议系统可能只需要每秒几百千比特。在规模的极低端，可视电话和无线通信系统通常被限制在每秒20千比特以下的速率。混合运动补偿离散余弦变换(DCT)结构是几乎所有现有视频编码标准的一部分，例如H.263和MPEG1，2，4，以及商业上可用的视频压缩系统。在将DCT应用于极低比特率视频编码时，块效应变得相当明显。该研究致力于在当今的视频编码系统中采用非块定向的分解技术来取代DCT。加州大学伯克利分校的VIP小组最近开发了一种这样的方案，该方案涉及基于Gabor字典的更大、更灵活的过完备基集上的信号分解。它使用统计文献中的贪婪、匹配追踪算法来寻找扩展系数。与基于DCT的方法相比，这种方法可以显著提高PSNR和视觉质量。此外，由于在其上执行信号分解的基函数的过完备集合上不存在约束，因此集合的元素可以被选择为没有块伪影，从而产生无伪影的重建帧，而不需要计算昂贵的后处理。这项研究的主要目标是研究过完备字典上的信号分解及其在视频编码中的应用。具体地，关于基于过完全信号扩展的视频编码系统要解决的问题包括：性能的基本界限、最佳幅度和位置编码算法、匹配基本集合的非分块性质的非基于块的运动补偿技术、词典设计、分解和搜索策略、以及扩展到：(A)时空可伸缩性，(B)静止图像编码，(C)多描述编码和(D)合成视频/图像/图形。工作水平声明：在建议的支持水平下，PI将尽一切努力满足项目的原始范围和水平。

CCR-9903368ZakhorEfficient compression of image and moving video is becoming increasingly more important in a wide range of emerging applications. These applications can be classified by the bit rates they require. A broadcast quality video-on-demand system might require several megabits per second, while a commercial video conferencing system might use only a few hundred kilobits per second. At the extreme low end of the scale, video telephone and wireless communication systems are often restricted to rates below 20 kilobits per second. The hybrid motion-compensated Discrete Cosine Transform (DCT) structure is a part of nearly all existing video coding standards such as H.263 and MPEG-1,2,4, and commercially available video compression systems. In applying DCT to very low bit rate video coding, blocking artifacts become quite pronounced. This research is focused on non-block oriented decomposition techniques to replace DCT in today's video coding systems.One such scheme recently developed at the VIP group at U.C. Berkeley involves signal decomposition over a larger, more flexible over-complete basis set based on Gabor dictionary. It uses the greedy, matching pursuit algorithm from the statistics literature in order to find the expansion coefficients. This approach is found to result in significant PSNR and visual quality improvement over DCT based schemes. In addition, since there are no constraints on the overcomplete set of basis functions over which the signal decomposition is carried out, the elements of the set can be chosen to be free of blocking artifacts, resulting in artifact free reconstructed frames, with no need for compute expensive post processing. The main goal of this research is to investigate signal decomposition along overcomplete dictionaries with applications to video coding. In particular the issues to be addressed regarding video coding systems based on over complete signal expansion include: fundamental bounds on performance, optimum amplitude and position coding algorithms, non-block based motion compensation techniques to match the non-blocky nature of the basis set, dictionary design, decomposition and search strategies, and extension to: (a) spatio-temporal scalability, (b) still image coding, (c) multiple description coding and (d) synthetic video/imagery/graphics. Level of Effort Statement: At the recommended level of support, the PI will make every attempt to meet the original scope and level of the project.