Pattern recognition in interrelated data: the problem, fundamental assumptions, recognition algorithms

As an adjunct to the classical pattern recognition theory dealing with single objects, a new approach to supervised pattern recognition is proposed for a variety of practical problems in which the class-memberships of several interrelated objects making an entire data array are to be estimated jointly. It is assumed, first, that the known structure of the array has the form of an undirected graph of immediate pair-wise adjacency of objects represented by their feature vectors, and, second, that the priori knowledge on expected combinations of classes is expressed as a hidden Markov random field on that graph. The presence of pronounced a priori information on interdependence of class-memberships of immediately adjacent objects allows for drawing much more reliable decisions from relatively unreliable features than in the classical case when the classes of single object are a priori considered as independent.