Most common quantization schemes assume quantizer design based on training sets and/or source models which can represent the statistics of the entire input. The performance of quantizers using such a priori knowledge of the input is largely affected by the choice of the training set or the input model, resulting in a loss of performance if there is a mismatch between the actual input statistics and the design assumption. In practical situations of quantizing complex data, it may be hard to have a good training set or sufficient knowledge on the input model. Thus there exists a motivation for adaptive quantization schemes which do not require any (or as little as possible) a priori information on the signal of interest.

We can categorize adaptive quantization schemes into two broad classes: forward adaptation and backward adaptation. In forward adaptive quantization, the encoder makes a decision on how to update the quantizer by probing current and future inputs. Since the encoder's decision is based on information unavailable to the decoder, side information has to be sent to the decoder to specify the changes. As an example of forward adaptive quantization, JPEG encoders can selectively use a specific quantization table for each image. Optimized tables for a given image must be stored in the header of compressed JPEG file for correct dequantization while reconstructing the image.

In backward adaptation, quantizers are updated based only on the previously quantized data which are available to both encoder and the decoder. Although backward adaptation has the drawback of requiring enocder and decoder to have similar complexities, it has the advantage of avoiding the need for overhead information transmission to the decoding end. When discussing adaptive techniques, we are particularly interested in this "overhead-free" backward adaptive quantization.

Related Publications

Youngjun Yoo and Antonio Ortega,"Adaptive Quantization without Side Information Using SVQ and TCQ," in Proc. of the 29th Asilomar Conf. on Signals, Sys., and Computers, Pacific Grove, CA, Oct 30 - Nov 1, 1995