Abstract:
Gibbs random fields with multiple pairwise pixel interactionshave good
potentialities in modeling natural image texturesbecause allow for learning both the
structure and strengths ofpixel interactions from a given training sample. The
learningscheme is based on the maximum likelihood estimate (MLE) ofGibbs
potentials that specify the interaction strenghts. Thisscheme is amplified here by
deducing an explicit, to scalingfactors, analytic form of the potentials from an
additionalfeasible top rank principle. It suggests that the training samplemay
possess a feasible top rank in its total Gibbs energy withinthe parent population.
Under this condition, only the scaling factorshave to be learnt using their MLE. As
a result, the introducedconditional MLE of the potentials extends capabilities of the
Gibbsimage models under consideration.