A unified backward approach to the code book design problem

Abstract

This paper presents a fast and robust approach for the code book design problem. Unlike the classical forward approach (CFA), which carries out the design process in two independent stages and in a forward direction; quantization then index assignment, the proposed unified backward approach (UBA) employs a statistics based mapping to generate the initial code vectors from the indices. This association faces a challenge due to the cardinally inequality, and this leads to utilizing the discrete cosine transform for compaction and dimension scaling. This mapping generates code vectors with two qualities; implicit tendency towards the constellation centroids and indices distance relationships analogy. Tests have shown that the UBA outperforms the CFA in more than 90% in terms of source distortion, while consistently having significant reduction in the channel distortion. Moreover, the improved initialization has caused some reduction in run time for higher bit rates designs.