Improving the efficiency of recognition of micro-objects based on the use of redundant information structures of images

Abstract

Scientific and methodological foundations have been developed for solving problems of identification, recognition, classification of micro-objects - pollen grains, unicellular organisms on the basis of mechanisms for determining their variety, belonging to any class, optimization using information redundant structures in the form of geometric shapes, morphology, dynamic, specific characteristics, features of neural networks. The systems of management of production and technological complexes, palynology, monitoring of the environment and ecology, medical diagnostics have been investigated. Mechanisms for optimization of identification and adaptive learning of neural networks in conditions of a priori insufficiency, uncertainty of parameters, and low accuracy of data have been developed. Optimization methods based on adjusting the parameter values to the properties of changing the points of the image contours, adjusting the network structure, variable activation function, filtering noise, random emissions, bursts in the image are proposed. Estimates of the root mean square error are obtained for various dynamic models of identification, filtration, approximation, interpolation, and extrapolation of images. A software package for identification, recognition, and classification of micro- objects has been implemented, which has been tested according to real data from a regional clinic in the diagnosis of tuberculosis patients based on a three-layer, recurrent, loosely coupled neural network, Kohonen network according to network learning algorithms with and without a teacher, which is synthesized with cubic, biquadratic, interpolation spline functions.