Optimization of Identification of Micro-Objects with Setting the Values of Variables of the Image Model

Constructive approaches, principles, and methods for optimal identification of images of micro-objects with mechanisms for extracting and using statistical parameters, dynamic and specific characteristics of information, and unique properties of neural networks have been developed. A hybrid model for the synthesis of computational schemes for training three-layer neural networks, modified algorithms for the structural components of the network, mechanisms aimed at correcting the weights of neurons, synoptic, interneuronal connections, variable activation functions, appropriate architecture, input-output dependence, selection and formation of reference points of the image contour is proposed. Mechanisms for preliminary processing of information, extraction and segmentation of image contours, selection of an informative training set, selection of a suitable identification model, tools for optimizing the training of neural networks using probabilistic search for points based on random enumeration of all possible options, with annealing, prohibitions, and stochastic modeling have been implemented. A generalized algorithm for identifying images of micro-objects has been implemented, combining the capabilities of polynomial models, the Daubechies interpolation spline of 4 and 5 orders, and a cubic spline, the software modules of which have been tested in recognizing and classifying unicellular microorganisms in medical diagnostic systems and in the management of industrial and technological complexes.