New Signal Processing Techniques for Next Generation Video Compression

视频压缩（VC）是实现无线网络上视频内容的流媒体、下载和广播的关键技术之一。VC的重要目的是尽量减少捕获内容所需的数据量，同时不影响视觉质量。考虑到与大多数视频内容相关的大量数据，如果不首先应用非常有效的压缩来减少所需的网络带宽，即使是最快的无线链路，也几乎不可能发送高质量的实时视频。随着无线网络容量的不断增长，诸如8K高清视频、高动态范围（HDR）和宽色域（WCG）视频等新兴应用的原始比特率也将继续增长。思科系统公司（Cisco Systems Inc.）估计，到2021年，超过80%的互联网流量将是视频。许多业内人士认为，如果不从根本上改变当前VC技术所基于的传统运动补偿变换编码（MCTC）框架，就无法实现视频压缩效率的实质性提高。这项研究计划将从根本上研究新的颠覆性VC技术，这些技术可以帮助超越MCTC的性能限制。所提出的研究在两个主要方面偏离了既定的实践，具有实现视频压缩效率数量级改进的潜力。首先，它的目的是放弃将一系列处理操作以启发式方式组合在一起的传统方法，以实现复杂的功能，如运动补偿（MC）和率失真优化（RDO）。相反，重点将放在整体方法上，这些功能将从真实视频数据中“机器学习”。近年来，基于深度卷积神经网络（DCNN）的机器学习已经解决了计算机视觉中的许多难题。鉴于视频压缩也可以被视为计算机视觉问题，本研究将追求将DCNNs用于MC和RDO的新思路。提出的研究的第二个主要区别是使用生物启发视觉模型进行性能优化。例如用于量化原始视频和编码视频之间失真的视网膜模型和用于人眼信息编码的模型。这些模型可以用于大量压缩视频数据，而对人类观看者所感知的质量影响可以忽略不计。这与目前基于像素级差异压缩视频的做法形成对比，后者会在压缩视频中产生干扰的可见伪影。预计本研究将为MC和RDO引入新的框架，有助于开发新的视频压缩算法，实现非常高的带宽效率。本研究以研究生为主体，对通信工程专业高素质人才的培养也将起到重要作用。

Video compression (VC) is one of the key technologies that enables streaming, downloading, and broadcasting of video content over wireless networks. The important purpose of VC is to minimize the amount of data required to capture the content, without compromising the visual quality. Given the sheer amounts of data associated with most video content, it is practically impossible to send high quality live video over even the fastest wireless link, without first applying very efficient compression to reduce the required network bandwidth. While the capacity of wireless networks will continue to grow, so will the raw bitrates of emerging applications such as 8K HD video, high dynamic range (HDR) and wide color gamut (WCG) video etc. Cisco Systems Inc. estimates that, by 2021, over 80% of the Internet traffic would be video. Many in the industry are of the opinion that substantial improvements in video compression efficiency can no longer be achieved without fundamental changes to the conventional motion compensated transform coding (MCTC) framework on which the current VC technology is built. This research program will investigate fundamentally new disruptive VC technologies which can help go beyond the performance limits of the MCTC. The proposed research departs from the established practice in two major ways with the potential to achieve order-of-magnitude improvements in video compression efficiency. First,  it aims to forgo the conventional approach of putting together a series of processing operations in a heuristic manner to implement complex functions such as motion compensation (MC) and rate-distortion optimization (RDO).  Instead, the focus will be on a holistic approach where such functions are "machine learned" from real video data. Recently, a  number of difficult problems in computer vision  have been solved with machine learning based on deep convolution neural-networks (DCNN).  Given that video compression can also be viewed as a computer vision problem,  this research will pursue the novel idea of using DCNNs for MC and RDO . The second major difference of proposed research is the use of bio-inspired vision models for performance optimization. Examples are retinal model for quantifying the distortion between original and coded video and models for information coding in human eye. These models can be used to heavily compress video data with a negligible effect on the quality as perceived by a human viewer. This is in contrast to the current practice of  compressing video based on  pixel-level differences which results in disturbing visible artifacts in compressed video. It is anticipated that this research will introduce novel frameworks for MC and RDO  which can help develop new video compression algorithms achieving a very high bandwidth efficiency. Being mainly carried out by graduate students, this research will also play an important role in the training of highly qualified personnel (HQP) in communication engineering.