Classification of Images with Linear Objects in All-Sky Astronomical Survey Data using Convolutional Neural Networks

This paper focuses on the classification of astronomical data, specifically images with or without linear features, which can be of different origin like meteors, asteroids, or satellites. Such models might be helpful in fast recognition of contaminated images from astronomical surveys, where satellites or other phenomena might decrease scientific value of survey for astrophysicists. It is quite large problem in astronomy today, due to increasing number of new satellites. On the other hand, such identification might also bring information on objects that pass through cameras field of view. Therefore, any fast way to recognize such cases is helpful. To achieve our goal of classification, we propose two main methods. Firstly, we focus on annotation project of images to create a dataset divided into two classes based on the occurrence of linear objects. Secondly, we suggest three different experiments utilizing convolutional neural networks. The results of each experiment are crucial as we create a sequence of experiments, with the best result serving as input for the subsequent experiment. To determine the success of our experiments, we define a threshold value of 98% for accuracy, which we were able to achieve in the final of the three experiments.