Using convolutional neural networks to automatically select small artificial space objects on optical images of a starry sky

Abstract

This article discusses the use of convolutional neural networks to solve the problem of automatically selecting moving objects on a moving starry background when their images exhibit speed blur. The article gives the results of testing several networks that have substantially less structural complexity than does the prototype. The estimates obtained for the accuracy and selection rate of several of the networks studied here are evidence that it is promising to use such networks to detect, classify, and estimate the location of two types of objects in the instrument’s coordinate system when resources are severely limited.