Professional enrichments on independent component analysis for blind source separation [articol]

Abstract

Independent Component Analysis (ICA) and its variants were proposed for Blind Source Separation (BSS) wherein most of the algorithms assumed that the sources are independent, non-negative and well-grounded. Some variants of ICA made the independence assumption unnecessary by utilizing information from theoretically based metrics; however, these methods are very slow in the process. In this paper an efficient Clustering of Mutual Information based least dependent Component Analysis (CMILCA) is proposed to cluster based least dependent components in a computationally efficient way by utilizing a squared-loss variant of mutual information. However, CMILCA provides better results just for a less number of sources and observations. In order to overcome this, Clustering in Conjugate Gradients of Mutual Information based least dependent Component Analysis in Riemannian manifold (CCGMILCA) is proposed to achieve convergence faster. The Riemannian directional derivative is used for local minimum efficiency, which results in the estimation of a weight matrix by giving independent components corresponding to various sources of mixture signals. The results of our experiments show that proposed algorithms take less time and high signal to noise ratio than the existing Blind Source Separation techniques.