Point estimation for multi-spectral distributed random matrices (Q1017617)

The authors propose a method for better recognition of noised images. The image vector is assumed to be normal \((M,\Sigma)\) and presents a sum of independent normal signal vectors with unknown mean and covariance matrix and a noise vector with zero mean and known covariance matrix. It is assumed that the problem of signals from noise separation may be solved by one of methods cited. Theorems 1-5 present maximum likelihood estimators for \(M\) and \(\Sigma\) and of signal distribution parameters. A numerical experiment of ten digits recognition from noised patterns shows the advantage of the maximum likelihood prognoses over human eye recognition.