The object of the research is a neural network represented by a perceptron, designed for classifying observation objects based on their coordinate features. The subject of the research is the mathematical apparatus of the neural network, constructed relative to the perceptron, to which angular and linear coordinates of the observation object are fed using a monocular technical vision system (TVS). The perceptron contains four neurons in the input layer, three neurons in the output layer, and an intermediate layer. The goal of the research is to determine the optimal parameters of the neural network for predicting classifications of observation objects with minimal errors, calculated using software tools. The mathematical apparatus of the neural network is derived according to the graph of forward and backward propagation. A feature of the mathematical apparatus is its sequential structuring with the derivation of the formula for the cross-entropy loss function. Traditional methods necessary for training the neural network are used to construct the mathematical apparatus: forward propagation, backward error propagation, and gradient descent. The software implementation was carried out using the Python programming language in the PyCharm environment. The methodology of the research is based on the development and modeling of neural network software according to its mathematical apparatus. The work aims to improve the accuracy of classification predictions of objects using a neural network in a video surveillance channel equipped with a monocular TVS. The presented neural network can be used in combined systems of controlled gyro-stabilizers and technical vision. In this case, the video surveillance channel with the neural network will classify the object and determine its coordinates, while the gyro-stabilizer will aim the optical line of sight at the target. The results of the simulation are presented in the form of graphs of the loss function against the number of software iterations during the training of the neural network. The software training time and the number of satisfactory predictions from the test sample for various activation functions have been determined. The integration of the neural network into the video surveillance channel of combined optoelectronic systems is a new and promising solution to the problem of object identification based on coordinate features and subsequent targeting.
Anton Dmitrievich Mihed (Thu,) studied this question.