Optimization of Micro-object Identification Based on the Mellin Transform and the Use of Parallel Computing

Abstract

Scientific and methodological foundations for the optimal identification of non-stationary objects based on the use of neural networks have been developed. Models and algorithms for detection, extraction of hidden relationships, useful properties and patterns in data, formation of a database and knowledge bases are proposed. Mechanisms have been developed for using the statistical, dynamic and specific characteristics of images, unique features of three, five-layer neural networks and combined models for setting variables with typical recognition and classification tools. Have been developed computational schemes for determining and adjusting the weights of neurons, choosing a suitable activation function, coefficients of synaptic and interneuronal connections, rational neural network architecture, the number of layers and neurons in the layers of the network, a set of functions of nonlinear dependencies "inputs-outputs". Data pre-processing algorithms are implemented that perform the functions of informative features selection, segmentation, object image contour extraction, search based on methods with annealing, prohibition, and stochastic search. Tested neural networks of Hopfield, Hamming, Hebb, Kohonen, bidirectional associative memory were tested. Schemes for two and three-dimensional image reconstruction based on the synthesis of tools for calculating Mellin transform functions, initial values of centroids, and the formation of a suboptimal set of variables are proposed. The identification software package in C++ was developed and implemented in the CUDA parallel computing environment.