STRUCTURAL PECULARITIES OF NEUROLIKE OBJECT CLASSIFIER

Abstract

Classification of the objects of various designation is the most involved procedure in the sphere of image recognition. Classification procedure is especially efficient in medical diagnostics where input features are biomedical symptoms and the output data is the disease diagnosis. In case, when the statistical methods of objects description are used, discriminant analysis has performed well, in particular, on the base of discriminant functions. On the other hand, methods of classification, applying neural technologies are of great interest.

The given study contains the analysis of structural peculiarities of neurolike object classifier, used in the process of discriminant functions classification. Kohonen map SOFM was taken as a basic model, it has 2D organization and determines metric and topologic dependences of the input signals. The study also considers the alternative approach to the quantitative measure of proximity as classification criterion. The approach is used where the formation of linear discriminant functions and their pairwise comparison is not performed, this enables «not to grow» linear discriminant functions but process on the level of their addends with gradual resetting to the moment, when one non-zero linear discriminant function is left. In this case there exists the possibility to form object occurrence ranks to the determined classes.

Two dimensional structure of neurolike classifier is suggested, basic unit of which is matrix calculator (maximizer). It is realized in the form of two maps – 2D computational map and 1D map of features. Neurosimilarity of the structure of the suggested classifier is stipulated by the fact that for the formation of the computation map three basic self-organization processes are used, namely, competition, cooperation and synaptic adaptation. The given paper contains the table with the comparative characteristic of Kohonen map and the suggested matrix calculator as a part of neurolike classifier.