Classification Of Space Debris And Stars In Astronomic Images Based On Deep Learning

With the continuous development of space exploration and spacecraft are continuously sent to space,the collision between spacecraft directly leading an explosive increase in the number of space debris.The increase in the number of space debris poses a serious security threat to space exploration in turn,which led to the study of relevant issues in space debris.At present,radar detection and telescope observations are the main ways to detect space debris.This article is about telescope observation.Currently,the widely used software for identifying space debris is SExtractor.Its principle is to process the astronomical images observed by the telescope and classify the stars and space debris by the difference in the movement speeds of the stars and space debris.Deep learning originated in neural networks in machine learning,and it has as one of the most important explorations into artificial intelligence because its structure simulates the human brain.Among them,the convolutional neural network in deep learning as an excellent classifier is also one of the important methods of image processing.This paper discusses the feasibility and correctness of the recognition of space debris by the convolutional neural network through the observation images of the telescope.The rate of this article is as follows:(1)In this paper,a convolutional neural network is constructed under the framework of caffe for space debris recognition experiments and compared with the accuracy of SExtractor.Among them,regarding the learning rate parameter of convolutional neural network,this paper proposes a method of screening learning rate by Pearson test.(2)Due to the particularity of astronomical images,this paper evaluates the noise tolerance of convolutional neural networks.After verifying the robustness of convolutional neural network through experiments,a simple image processing method is proposed to improve image signal-to-noise ratio for astronomical images and experimental verification is performed.