Performance Improvement of Second-Order-Statistics-based Noisy BSS

A. Tanaka and M. Miyakoshi (Japan)


blind source separation, second-order-statistics, joint diagonalization, stationary noise


The aim of blind source separation (BSS) is to recover mutually independent unknown source signals from obser vations obtained through an unknown linear mixture system. In many existing methods, it is assumed that the observations are not contaminated by an observation noise. Although methods for a noisy observation model are proposed, assumptions for the noise, such as Gaussianity, spatially or temporally whiteness, may limit the application area of these methods. In our previous work, we proposed a novel noisy BSS method based on a joint diagonaliza tion of differences of correlation matrices in which only stationarity is imposed on the noise. However, this method was unstable in practical situations. In this paper, we improve our previous method by applying the idea of a stable second-order-statistics-based BSS algorithm to the noisy setting. We also verify the efficacy of the proposed method by computer simulations.

Important Links:

Go Back