Entropy-constrained reflected residual vector quantization: A realization of large block vector quantization

Abstract

Vector quantization (VQ) techniques have been used for a number of years for developing successful data compression and classification algorithms. Vector quantization is inherently an unsupervised data clustering technique that smartly adopts to the given data statistics and is thus used in large variety of tasks. The gains associated with a VQ system are largely linked to the size of the block used. However, the complexity increases exponentially with an increase in block length. In conventional methods, a dimensional reduction pre-processing step is invoked before the VQ system in order to reduce the block length. For some data sources this may be unacceptable approach such as with hyper-spectral data. Therefore, a large block VQ is needed to fully exploit the linear and non-linear correlation in a data source. We propose an entropy-constrained reflected residual VQ (EC-RRVQ) as one alternative for large block vector quantization. The reflected residual VQ, a type of multistage VQ, has constrained structure with multiple stages having small-size code-books. The use of multiple code-books and a reflection constraint makes the computational complexity linearly depend on the number of stages involved. The linear increase in complexity prompted us to pose EC-RRVQ as a contender in the list of options of large block vector quantization implementation algorithms. Experimental results indicate that good image reproduction quality can be accomplished at relatively low bit rates. The performance of a 64-stage 16×16 EC-RRVQ at 0.175 bits per pixel is 23.75dB with 96 vector distortion calculations per source vector, while the number for previously proposed large-dimensional entropy-constrained residual VQ (EC-RVQ) designed under the same specifications is 21.17dB with 1212 vector distortion calculations per source vector.