The Innovations Approach to Single Frame Multichannel Blind Image Deconvolution

Abstract

The linear mixture model (LMM) has recently been used for multi-channel representation of a blurred image. This enables use of multivariate data analysis methods such as independent component analysis (ICA) to solve blind image deconvolution as an instantaneous blind source separation (BSS) requiring no a priori knowledge about the size and origin of the blurring kernel. However, there remains a serious weakness of this approach: statistical dependence between hidden variables in the LMM. The contribution of this paper is an application of the ICA algorithms to the innovations of the LMM to learn the unknown basis matrix. The hidden source image is recovered by applying pseudo-inverse of the learnt basis matrix to the original LMM. The success of this approach is due to the property of the innovations of being more independent and more non-Gaussian than original processes. Our good, consistent simulation and experimental results demonstrate viability of the proposed concept.