Given a real textured image, we wish to determine the parameters of a GMRF model which could be used to reconstruct the original texture through the samplers given in the previous section.

This section develops parallel algorithms for estimating the parameters of a GMRF texture. The methods of least squares (LSE) and of maximum likelihood (MLE), both described in [6], are used. We present efficient parallel algorithms to implement both methods. The MLE performs better than the LSE. This can be seen visually by comparing the textures synthesized from the LSE and MSE parameters, or by noting that the asymptotic variance of the MLE is lower than the LSE ([3], [25]).

David A. Bader

dbader@umiacs.umd.edu