Pattern classification using relative constraints

An approach to pattern classification based on relative constraints in a discrete relaxation framework is described. Classical pattern classification techniques partition feature spaces into disjoint decision regions where thresholds are absolute, i.e. fixed numerical quantities. The approach defines pattern classes relative to one another and so results in decision boundaries that depend on the data being classified. Such a formulation leads to a classification scheme based on finding unambiguous labelings using a discrete relaxation-labeling algorithm. Classes are defined exclusively in relative terms, using fairly weak constraints. As a result, there are not many locally incompatible hypotheses to eliminate by Waltz filtering. A ranking scheme is developed which orders hypotheses so that unambiguous labelings can be quickly found through depth-first search. When an unambiguous labeling does not exist, classes can be assigned by picking the most compatible hypotheses. Results of work in progress in classifying Landsat multispectral imagery are presented