Space-variant Neural Network Approach to Blind Image Deconvolution

A new space-variant neural network based blind image deconvolution method is proposed for restoring degraded images and the space-variant blur simultaneously. The neural network model is based on an autoregressive moving average (ARMA) process in which the AR part of the neural network defines the image model and its coefficients, while the MA part defines the degradation process. Since the blur affects textured region and smooth regions differently the image is divided into blocks, which have been categorized into four classes according to their activity. Furthermore it is assumed that blur is space-invariant in each block. This classification enhances the texture of relatively textured regions while suppressing the noise in relatively smoother backgrounds. A new cost function motivated by the human visual perception system is also proposed for the neural network. This comprises of two new terms for matching the second order local statistics of each block in order to improve the visual quality of the restored image. Improvement in signal to noise ratio (ISNR) and normalized mean square error (NMSE) of the estimated blur have been used as figures of merit. Our proposed algorithm has given better results in terms of ISNR and NMSE compared to some of the results available in the literature.