A NEW OPTIMIZATION METHOD FOR REFERENCE-BASED QUADRATIC CONTRAST FUNCTIONS IN A DEFLATION SCENARIO

摘要

This paper deals with the problem of blind source separation of convolutive MIMO mixtures by a deflation procedure. Contrast functions showing a quadratic dependence with respect to the searched parameters have recently been proposed. Combined with a fast SVD-based optimization technique, they proved to be very efficient for the extraction of one source signal. In this contribution, we examine how these contrast functions behave in a deflation scenario. We show that the SVD-based optimization method requires a good knowledge of the filter orders due to its sensitivity on a rank estimation. To overcome the difficulty, we propose an optimal step size gradient algorithm.