Performance analysis of over-determined noisy ICA: Bayesian approach versus signal transformation

This paper proposes a new analysis on two robust methods for solving the blind source separation problem of noisy linear over-determined mixtures using 2-Stage ICA and Bayesian approach. A new method has also been developed to determine the optimal SNR threshold as the selection index for choosing the better method under the varying influence of the noise levels. An experimental simulation has been analytically conducted to verify the proposed method. An in-depth analysis has been carried out between the two methods regarding to their different performances through out the noise level from -10 dB to 30 dB. It is further shown that the threshold selection can be generalized to more complex cases that have the same ratio between the number of observed signals and the number of sources.