biography research projects publications courses students home

data compression lab:

universal multi-resolution
source coding


related papers


Like traditional (single-resolution) source codes, multi-resolution source codes are data dependent. The optimal multi-resolution source code for a particular source guarantees good coding performance on that source, but may achieve poor performance on other sources. In the interest of designing source-independent multi-resolution source codes to achieve good performance across a broad class of possible sources, we have recently introduced the new field of universal multi-resolution source coding.



The goal in universal multi-resolution source coding is to design a single code that will -- in the long run -- do as well on each source in some class of sources as if it were designed for the source in operation, and to do so without prior knowledge of the source to be compressed. At first glance, universal multi-resolution source codes, like universal (single-resolution) source codes seem too good to be true. Yet early work in this nascent field has included not only proofs of the existence of these codes but also given bounds on their performance.



While universal multi-resolution source coding theory proves the existence of universal multi-resolution source codes, the proofs are not constructive. That is, the proofs demonstrate the existence of good codes without describing how to find those codes in practice. Recent work in practical universal multi-resolution source coding treats the problem of optimal universal multi-resolution source code design. This work takes an approach similar to that used in the weighted universal vector quantization algorithm, but generalizes that approach by building a collection of multi-resolution source codes rather than a collection of single-resolution source codes and by allowing the first-stage source description itself to be multi-resolution in nature.