Novel Ica Algorithm With Nonparametric Estimation Based On Ggd Kernel
Fasong Wang,Hongwei Li,Yujie Zhang,Rui Li
2006-01-01
Abstract:A basic element, in most independent component analysis (ICA) algorithm, is how to choose a model of the score functions of the unknown, source signals. This is usually based on the parametric approximations, but the parametric approach will not be truly blind to the underlying distribution of the sources. In this paper, we describe an efficient, scheme for achieving the probability density function (PDF) of sources by the generalized Gaussian density (GGD) kernel estimation, based nonparametric method. The proposed, ICA algorithm. is able to separate the hybrid mixtures of source signals which include sub-Gaussian and super-Gaussian sources, symmetric and asymmetric sources using only a flexible score function model. Moreover, the GGD kernel can, adaptively adjust its shaping and scaling parameters according to the high order statistics of the source signals. It paves the way to wider applications of ICA methods to real world signal processing. Simulations show that our proposed ICA algorithm can, separate a wider range of sources than most of the state-of-the-art parametric method.