Reverse engineering vector quantizers by training set synthesis

In this paper we present a technique for reverse engineering vector quantizers by synthesizing a training set that has similar statistics to the original training set used in designing the vector quantizer. Most VQ codebooks are designed using the LBG or generalized Lloyd algorithm which is similar to the construction of nonuniform bin histograms. Thus the VQ codebook and the number of training set vectors allocated to each of its codebook vectors is an approximation of the underlying pdf of the training set. This observation is used to synthesize a training set that has a histogram similar to the original training set. This synthesized training set can be used to construct VQs to describe subspaces of the original vector space or spaces transformed by a linear transformation.