Optimization of recognition of micro-objects based on reducing excessive information structures of images

Abstract

Scientific and methodological foundations of identification, recognition, classification of micro-objects using redundant information structures - morphometric, histological, fractal characteristics of images have been developed. Mechanisms for extracting statistical, dynamic, specific characteristics and rarefaction of redundant information structures are proposed. Dynamic models have been developed, combined with the capabilities of neural networks. Computational schemes for preliminary processing of images, texture, segmentation, filtering, approximation, regulation of variable values and optimization have been developed. Methods of recognition and classification of micro-objects with tools for obtaining images from a photo, video camera, digital microscope, interactive measurement, counting, structure determination, statistical analysis, isolation and segmentation, and the formation of informative (reference) points of image contours have been investigated. The traditional computational schemes of a multilayer neural network, a Kohonen neural network, and a radial-basic network are investigated. Modified algorithms for training a network with mechanisms for adjusting the values of variables, monitoring errors along the boundaries of permissible values, accounting for stationary, quasi-stationary and non-stationary behavior of image points during the formation of training samples have been developed. Generalized algorithms for the identification of images of pollen grains are proposed. The efficiency of the algorithms was investigated according to the criteria of the root mean square error and the speed of information processing. A software package for visualization, recognition, classification of images of pollen grains has been developed, the implementations of which have been tested taking into account the conditions of a priori insufficiency, uncertainty and nonstationarity of processes.