Parameter identification for noisy image via the EM algorithm (Q922420)

The main problem considered in this paper, is to restore a discrete monochromatic image degraded by observation noise. The original image is represented by a semicausal 2-D autoregressive (AR) model which can be written as a rector 1-D AR model the matrix coefficients of which have a special structure. Using the discrete sine transform (DST), the reactor AR model is approximately decoupled in a number of scalar 1-D AR models. Assuming white Gaussian observation noise, the parameters of the latter models are estimated by the estimation-maximization (EM) algorithm. The original parameters and signals are then determined by LSM and the inverse DST. The restored image is obtained as a by-product of the estimation step in the EM parameter estimation algorithm.