Source Coding Research at HBBCL

HBBCL members have considered quantizer design for both theoretical reasons and also for various applications. In particular, special attention was paid to rate scalable (successively refinable) and multiple description quantizers. Other areas of activity in HBBCL includes source coding using redundant expansions with low resolution quantization.

Successively Refinable Quantizers

Successively refinable (rate-scalable) quantizers have received considerable attention owing to the increasing importance of applications which require progressive transmission of data. An embedded output bit stream is provided by successively refinable source coders. i.e., by selecting different sub-streams of the output, various levels of encoding rate and distortion can be achieved. HBBCL members have invented the first successively refinable trellis coded quantizers. The structure and performance of our scalable trellis coded quantiizers and are reported in publications.

Multiple Description Coding



Multiple description coding is concerned with describing a data sequence using a number of coarse representations, each of which describes the data within accuracy which is acceptable for an application at hand, but at the same time these individual descriptions refine each other so that the desciription accuracy improves as more of them are received. Coding schemes of this kind are particularly useful for communication in packet loss scenarios.

The design of a multiple description coder, requires some form of joint optimization of rate distortion characteristics of individual and combined descriptions, and the resulting coders are often computationally very intensive. We approach the problem of multiple description coding using the techniques of trellis-coded quantization and propose a computationally simple coding scheme which provides remarkable rate-distortion performance.

Source Coding Using Redundant Expansions

 


The source-coding paradigm exploited in modern techniques for analog-to-digital (A/D) conversion is to attain high conversion accuracy by applying low-resolution quantization to redundant signal expansions. This approach to commercial A/D conversion, as illustrated in the above figure, is motivated by the need to avoid the high costs of implementing high-resolution quantization in the current technology. In simple oversampled analog-to-digital conversion, samples of bandlimited signals taken above the Nyquist rate are discretized in amplitude using uniform scalar quantization, while sigma-delta modulation employs sophisticated quantization schemes which better exploit the correlations between signal samples to achieve high-resolution amplitude discretization using only single-bit quantization. Relatively accurate signal representations in sensory systems of mammals, despite noise, coarse quantization, and system imperfections, are also attained by means of redundancy. Lately, we notice an increasing interest in quantized redundant expansions for multiple description coding for transmission over packet-loss networks, designed to ensure graceful degradation of signal quality as the probability of packet loss increases. Although the idea of gaining some form of robustness or accuracy by introducing redundancy is quite intuitive, precise quantitative characterization of the involved trade-offs for different degradation scenarios has proven to be difficult. Some members of HBBCL have been studying the topic extensively prior to joining Harvard University, and continued pursuing that direction of research here. Results of their work in this area have been presented in a number of publications.