Improved fast LBG training algorithm in Hadamard domain

Abstract

To reduce computational complexity of the Linde-Buzo-Gray (LBG) algorithm, an improved fast training method in the Hadamard domain is proposed. This method uses an optimally-ordered Hadamard transform kernel based on the statistical energy distribution of transformed training vectors in k-dimensional space and a three-step elimination criterion to more efficiently reject impossible codevectors. Experimental results demonstrate the effectiveness of the proposed method in terms of arithmetical operations and search space.