An alternative natural gradient approach for multichannel blind deconvolution

This paper presents an alternative natural gradient based solution to the multichannel blind deconvolution (MBD) problem. The derived learning rule comes from the definition of a new Riemannian metric in the linear system space, rather than the one relative to the algorithm already existing in the literature. Moreover, it is proved that this novel approach satisfies the equivariance property if the MBD problem is formulated in a certain way. Experimental results have shown that the two gradients lead to the the same deconvolution performances.