Parameter estimation for two-dimensional Ising fields corrupted by noise (Q915328): Difference between revisions

The studied random field on a finite subset of \({\mathbb{Z}}^ 2\) is a possible model for describing images. The suggested estimation of its parameters can serve for recognition of images. The process arises as the sitewise product of a free-boundary Ising model (with values \(\pm 1)\) with inverse temperature \(\beta\) and of the Bernoulli field with the probability \(\epsilon\) for -1 being ordered to a site from a finite \(\Lambda \subset {\mathbb{Z}}^ 2\). Estimators for \(\beta\) and \(\epsilon\) are suggested which are consistent as \(\Lambda\) tends to \({\mathbb{Z}}^ 2.\) The proof of the consistency contains some inequalities for elliptic integrals occurring in the expressions for correlations of the limit Ising random field in \({\mathbb{Z}}^ 2\). Numerical results are introduced where the ``true'' Gibbs field was simulated by a numerical algorithm. The conditional probability and maximum likelihood method are used to estimate the value at each site of the ``true'' image, and the true and estimated values of the parameters \(\beta\) and \(\epsilon\) are compared.