Modified Recognition Algorithms Based on the Construction of Models of Elementary Transformations

Abstract

This paper deals with the issues of building a model of recognition algorithms targeted at classifying objects in conditions of high dimensionality of the feature space. A model of recognition algorithms based on the estimate calculation has been considered as the source one. The distinctive features of the approach under consideration are as follows: 1) forming subsets of correlated features; 2) selecting a set of representative features when developing recognition algorithms; 3) building models of elementary transformations in the subspace of representative features. The main advantage of the proposed algorithms is the selection of the preferred models of elementary threshold rules with the subsequent calculation of the object membership estimate and ensuring a significant reduction in the number of computational operations when recognizing unknown objects. This characteristic is very important for real-time recognition systems. To test the operability of the proposed model, experimental studies have been carried out in solving a number of problems; in particular, model problems have been generated in the space of correlated features; as well as facial recognition tasks. These experimental studies considered: a model of recognition algorithms based on the estimate calculation; a model of recognition algorithms based on the potential function; the proposed model of recognition algorithms. A comparative analysis of the above models of recognition algorithms shows that the proposed model allows increasing the recognition accuracy by an average of 7-10%. The developed recognition algorithms can be used in medical and technical diagnostic, geological forecasting, biometric identification and other areas, in which solving problems of classification of objects specified in the high dimension feature space.