Robust Approach for Blind Source Separation in Non-gaussian Noise Environments

In this contribution, we address the issue of Blind Source Separation (BSS) in non-Gaussian noise. We propose a twostep approach by combining the fractional lower order statistics (FLOS) for the mixing matrix estimation and minimum entropy criterion for noise-free source components estimation with the gradient-based BSS algorithms in an elegant way. First, we extend the existing gradient algorithm in order to reduce the bias in the demixing matrix caused by the non-Gaussian noise. In the noise cancellation step, we derive a new kind of nonlinear function that depends on the noise distribution and we discuss the optimal choice of this nonlinearity assuming a generalized Gaussian noise model. The optimal choice, in the minimum entropy sense, is robust against the influence of Gaussian and non-Gaussian noise including heavy-tailed model. The effectiveness and the robustness of the proposed separating algorithm are shown on numerical examples.