Method for training codebook e.g. for vector quantizers, involves evaluating data training records recursively in stages

摘要

A method for training a codebook (C) with a number (M) of vectors for vector quantization of data with a training record (data set) (T) involves initially factorizing the codebook into N-part codebooks (CI) training the part codebooks (CI) with the starting training data set (TS), diving the data training set (TS) with the part codebook into further training sets (TI-TK) and evaluating the training data set recursively in stages (j). The number of code vectors in each of the codebooks is determined on the basis of an evaluation function of the respective cells of the codebook of stage (j). With the set of training sequences (TS), the codebooks for the next stage (j+1) are trained until the total number (M) of the code vectors is trained. An Independent claim is given for a computer program product.