Abstract:

In this work, an approach that can unambiguously classify objects and patterns based on identification of distinctive features in labeled training sets is described. By considering a pattern as a representation of extracts of information regarding various features of an object, most established recognition methods tend to achieve classification by identifying the resemblances amongst the class members. This paper looks at the recognition act differently, through negative recognition. It argues that the basic functioning of the established methods also implies that the members of distinct classes must exhibit different characteristics resulting in different values for some or all of the features that describe the objects under consideration. That is, the categorization can also be based on recognition of differences between objects that belong to different classes. Such characteristics, when identified, will form the distinctive features of patterns and objects, in our proposed approach. In other words, using training sets, distinctive features for all or at least some of the classes are determined. The distinctive features are then used to classify all objects, even for complex systems.