An Iterative Blind Source Separation Method for Convolutive Mixtures of Images

The paper deals with blind source separation of images. The model which is adopted here is a convolutive multi-dimensional one. Recent results about polynomial matrices in several indeterminates are used to prove the invertibility of the mixing process. We then extend an iterative blind source separation method to the multi-dimensional case and show that it still applies if the source spectra vanish on an interval. Based on experimental observations we then discuss problems arising when we want to separate natural images: the sources are non i.i.d. and have a band limited spectrum; a scalar filtering indeterminacy thus remains after separation.

• Type: Conference
• Year: 2004
• Where: ICA
• Authors: Marc Castella, Jean-Christophe Pesquet
• Keywords: Blind Source Separation, ICA 2004, Source Separation Method, Source Spectra Vanish