biography research projects publications courses students home

data compression lab:

multiple description
source codes


related papers


Multiple description source codes are source codes designed for systems where source reconstruction based on only a portion of the original data stream may be required. Multiple description codes are useful for applications like packet-based networks (this includes wireline systems like the internet as well as wireless systems including the majority of the wireless data systems), diversity communication systems (e.g., communications systems using antenna diversity), and distributed data storage systems. For example, in a packet-based communication system, where data is broken into small segments, each of which is separately sent through the channel, we would like to be able to reconstruct our signal under a full range of packet-loss scenarios. Similarly, in diversity channel systems, where a signal is sent from source to destination over a collection of channels, we would like to be able to reconstruct some version of our data string even when one or more of the channels in our collection fails.

Likewise in distributed data storage systems, where data storage is distributed across a number of hosts, we would like to be able to reconstruct the stored data even when some subset of the hosts is unavailable. In each of these applications, a multiple description code could be designed to allow reconstruction based on the received partial information. That is, we could design multiple description codes to allow source reconstruction under any packet-loss scenario in a packet-based network, under any channel failure condition in a diversity systems, and under any host availability situation in a distributed data storage system. (Multi-resolution source codes, which allow for reconstruction under a subset of these conditions, may be thought of as a special case of multiple description codes.)



In designing a multiple description code, we would like for the distortion achieved in a source reconstruction based on any possible subset of packets to be as close as possible to the lowest possible distortion at the received rate. Unfortunately, these best possible performances are not simultaneously achievable, as can be seen immediately from the multi-resolution source coding example. The goal in the theoretical research into the mutliple description problem is to completely describe the space of received rates and distortions achievable by a multiple description source code. Current results in this area include bounds on this space, but existing bounds are not yet tight.



Ongoing work in multiple description source coding concerns the design of practical multiple description source codes both within the optimal coding paradigm provided by vector quantizers and within other coding settings such as wavelet-based codes. New coding structures for general multiple description codes developed in the lab include a novel ternary code tree for multiple description source code design. Another recent advance in this area is the introduction of a new performance criterion for the design of multiple description codes. This criterion allows for the inclusion of a probability or priority function over the full range of packet-loss scenarios within a mutliple description coding system.