A practical Approach Based on Gaussianization for Post-Nonlinear Underdetermined BSS

This work deals with the blind source separation (BSS) problem in presence of more sources than sensors and post-non linear (PNL) mixing. The interest on the subject is increased by the fact that very few related contributions have appeared in the literature so far. The proposed method is made of three separate steps: compensation of nonlinearity (based on the Gaussianization concept), mixing matrix recovery and final unknown source estimation. The first one has been already considered for nonlinear complete BSS, but not in the over-complete case, whereas the latter two represent the typical two-step approach in underdetermined BSS. Performed computer simulations have shown the effectiveness of the idea, even in presence of strong nonlinearities and synthetic mixture of real world data (like speech signals)