Fast Rauch-Tung-Striebel smoother-based image restoration for noncausal images

We describe a technique for restoration of blurred images corrupted with additive noise. Our algorithm uses a practical implementation of the Rauch-Tung-Striebel (RTS) smoother-based on noncausal prediction that models the blurred image as a finite-lattice Gauss-Markov random process (GMRP). The one-sided regressors of the GMRP converge at a geometric rate to shift-invariant values along the rows of the image. This leads to a steady-state solution for the RTS filter. Experimental results illustrate the superiority of our RTS-based algorithm over Wiener filter, deterministic filters, and filters that use the one-sided causal state model.