Referenced Contrast: A New Tool for Blind Separation of Gaussian Sources

This work deals with the problem of source separation in the case when the observations result from MIMO instantaneous mixing system. In a blind framework, high order contrast functions constitue an efficient tool for extracting sources. However, all of them are limited to the case of at most one source is gaussian. In this paper, a reformulation of Referenced Contrast (RC) is presented. This novel method applies to separate all sources having any distributions. Our proposition allows to assess the capabilities of RC to break the obviousness that High-Order Statistics based methods are restricted to the mixtures allowing at most one Gaussian source. Our design constrains the reference signals choice to avoid the cancelation of the cross-cumulant involved in the proposed contrast. Simulation studies are presented to support the potential of the approach in terms of source separation.