Neural Network Architectures for Recognizing Military Objects on Satellite Images

Abstract

This paper is devoted to the research and comparison of neural network architectures for recognizing military objects on satellite images. The paper analyzes the current state of the problem of object recognition, in particular, military objects on satellite images. In addition, the authors researched existing technologies and tools for recognizing military objects on satellite images, and also implemented models of neural networks with different architectures for detecting military objects. The influence of learning indicators of neural network models on the recognition and detection of objects on satellite and aerial photographs was investigated. As a result of the work, one multilayer perceptron, three models of convolutional neural networks, and neural networks with VGG and XCeption architectures were implemented and investigated, their main advantages and disadvantages were determined, and software with corresponding neural network architectures was developed using the Google Colab cloud service. The best results were shown by sixth model with a convolutional neural network architecture, the main difference of which was a gradual reduction in the number of filters and the size of the kernel on the convolutional layers.