Abstract

Introduction Previous work on Blind Source Deconvolution has focused mainly on the problem of deconvolving sound samples. It is assumed that the observed sound samples are temporally convolved versions of the true source samples. Blind Deconvolution algorithms have fallen into two types, those where the inverse of the convolution lter is learnt [1],[3] and those where the aim is to learn the lter itself [1]. When applying these ideas to the problem of deconvolving images two problems become apparent. Firstly in many real data sets (for instance the images generated by telescopes observing the sky or the power spectrum from a Nuclear Magnetic Resonance (NMR) spectrometer) the pixel values correspond to intensities. So the pixel values must be positive. The standard blind separation approaches of assuming that the sources are distributed as 1 cosh [3] or mixtures of Gaussians [2] lose this positivity of the source images. Deconvolution without a positivity con

Citations