Multichannel blind deconvolution and source separation using the natural gradient
Шун-ичи Амари, S.C. Douglas, Andrzej S Cichocki · Institutional Repositories DataBase (IRDB) · 1998
Multichannel deconvolution is an important task for numerous applications in communications, signal processing, and control. In this paper, we extend the efficient natural gradient search method to derive a set of on-line algorithms for combined multichannel blind source separation and time-domain deconvolution of additive, convolved signal mixtures. The algorithms are derived from a maximum differential entropy cost formulation. We prove that the doubly infinite multichannel filtering system possesses the equivariance property such that its asymptotic adaptive behavior depends only on the normalized stochastic distribution of the source signals and not on the mixing characteristics of the unknown channel. Both approximate time- and frequency-domain finite-impulse-response (FIR) implementations of the methods are described. Extensive simulations indicate the ability of the proposed methods to perform efficient simultaneous multichannel deconvolution and source separation.!