Vector quantization based on Gaussian mixture models

Switched Conditional PDF-Based Split VQ Using Gaussian Mixture Model

“…We need to solve for the optimum bit allocation, , such that is minimized. While there has been an attempt [7] to analyze the distortion incurred for a GM source using an expectation-maximization (EM) type of iterative algorithm, the above linearized approach provides a simple closed-form solution using the optimum inter-cluster bit allocation. Since the source PDF of each Voronoi region is modeled as a multivariate Gaussian with known parameters, we can use the high-resolution quantization distortion of CSVQ [16] for an overall distortion expression of the GM density and then optimize for .…”

“…While a direct R/D performance analysis of a GMM source is not easy to carry out [7], we consider a linearized approach [8] to the overall quantization distortion of the source, expressed as (4) where is the incurred distortion of the quantizer for the th component Gaussian of GMM and is the allocated bits/vector to the quantizer. We need to solve for the optimum bit allocation, , such that is minimized.…”

“…We also address the R/D performance optimality of the proposed SCSVQ method in a GMM-based parametric framework where the source PDF of each Voronoi region is modeled as a multivariate Gaussian. In this context, we would like to indicate that currently, there is a growing interest to develop parametric PDF-based quantization methods using the GMM [7], [8], [11]. For the parametric SCSVQ method, we derive the optimum bit allocation criteria among the Voronoi regions using the variable bitrate scheme [8].…”

Robust Underdetermined Algorithm Using Heuristic-Based Gaussian Mixture Model for Blind Source Separation

“…Θ is the value of the parameter set Θ at iteration k (Hedelin & Skoglund, 2000), (Nikseresht & Gelgon, 2008). A poor initialization of set Θ would have great effect upon final performance; however, some elements are hard to give suitable initial values by experience of user.…”

A Statistical Approach to Packet Loss Concealment for Video

“…The EM algorithm discussed in [5] has been widely used on statistical problems throughout many disciplines. In the field of communication, it has been applied for fitting a Gaussian basis in order to improve prediction [6], quantizing [7] and classification [8].…”