Source Clustering for Codebook Compression

Clustering algorithms can be applied to the design of N codebooks to be shared by M sources, 1< M < N. We previously introduced a constrained storage vector quantizations algorithm for this design problem. In this work, we extend the algorithm to additionally design simple parametric expandor functions to enhance codebook sharing efficiency. We apply the particular case of scaling expandor functions to the compression of tree structured vector quantization codebooks. By allowing different levels of the tree codebook to share a library of feature (residual) codebooks, we were able to achieve in our experiment a 4 : 1 reduction of storage without compromising rate-distortion performance. For very deep trees, an earlier design method which effects sharing only within each level of the tree is more effective.