An analysis of competitive and re-initialization learning for adaptive vector quantization

This paper describes an analysis of competitive and reinitialization learning (CRL) for adaptive vector quantization (AVQ) which is a version of vector quantization (VQ) in order for digital coding of signals to adapt to changing statistics of the signal sources. The CRL has been designed for achieving equi-distortion or asymptotically optimal quantization to overcome the under-utilization problem or the local minimum problem of vector quantization networks, while its performance in adaptation speed and obtained distortion level is shown to be higher than the conventional AVQ algorithms such as the optimal adaptive k-means algorithm (OPTM) and diversity oriented competitive learning II (DOCL-II). In this paper, after reviewing the CRL algorithm, we examine how the CRL algorithm works for various source signals such as nonstationary 2D vectors and high dimensional images. Furthermore, we compare the performance of the CRL with the OPTM and the DOCL-II