High Performance Multiple Description Code Design

在多重描述（MD）问题中，目标是生成表示信号的给定数量的比特流（即描述），例如K，这样每个比特流可以被解码到可接受的质量，同时更多的比特流相互改进以改进信号重建。MD编码在有包损耗（有线或无线）网络上的多媒体信号传输中得到了应用，因为它可以确保在由于严格的延迟限制而无法重传时从丢失中优雅地恢复。其他应用还包括分集通信系统、分布式数据存储和多速率组播。大多数实际的MD代码处理两种描述的情况，而一般K的实际设计很少，而且它们的性能与理论最优值有一定距离。另一方面，虽然传统的MD码被设计用来对抗描述丢失，但它们对抗比特错误的能力已经成为一个副产品。这个特性在无线通信中特别有用。然而，利用这一特性设计MD码却很少受到重视。提出的研究目的是开发高性能MD码，提高从损失和误码中恢复的能力。这将通过高度灵活的MD设计来实现，它可以控制接收描述的不同子集的重建质量之间的权衡，并通过构建具有良好汉明距离属性的索引分配来实现。这项研究的一个重要组成部分将是训练高度合格的人员。数名研究生将参与本研究，并将接受多媒体编码与通信、信息理论与优化方面的培训；为他们在信息技术领域的工程或学术生涯做准备的培训。拟议的研究将促进MD代码设计领域的知识和理解，所开发的技术和见解也将适用于其他源网络编码问题（例如，在健壮的分布式源编码中）。

In the multiple description (MD) problem, the goal is to generate a given number, say K, of bitstreams (i.e., descriptions) representing a signal, such that each bitstream can be decoded to an acceptable quality, while more bitstreams refine each other improving the signal reconstruction. MD coding finds applications in transmission of multimedia signals over packet lossy (wired or wireless) networks as it ensures graceful recovery from losses when retransmission is not an option due to stringent delay constraints. Other applications are in diversity communications systems, distributed data storage and multirate multicast. Most practical MD codes address the case of two descriptions, while practical designs for general K are scarce and their performance is some distance from the theoretical optimum. On the other hand, while traditionally MD codes were designed to combat description loss, their ability to combat bit errors has emerged as a side product. This feature is especially useful in wireless communications. However, the design of MD codes to take advantage of this property has received little attention. The purpose of the proposed research is to develop high performance MD codes with increased capability to recover from losses and bit errors. This will be achieved through MD designs of high flexibility in controlling the trade-off between the quality of reconstruction for different subsets of received descriptions and through construction of index assignments with good Hamming distance properties. An important component of this research will be the training of highly qualified personnel. Several graduate students will be engaged in this research and will be trained in multimedia coding and communications, information theory and optimization; training that will prepare them for an engineering or academic career in the information technology sector. The proposed research will advance the knowledge and understanding in the area of MD code design and the techniques and insights developed will be also applicable to other source network coding problems (for instance, in robust distributed source coding).