Application Of Deep Learning Algorithm In Celestial Object Image Classification

With the continuous development of sky survey projects in recent years,the amount of astronomical data has increased exponentially.The massive observation target sources obtained by telescopes contain a variety of different types of celestial bodies.This thesis mainly studies the classification of galaxies,stars,and quasars in celestial types.Deep learning algorithms are able to automatically learn their features from large amounts of data and enable classification or parameter measurement tasks.Therefore,people began to apply deep learning algorithms to solve the problem of accurate classification of astronomical images.This thesis first obtains the data set in the Sloan Digital Sky Survey project,and uses the deep learning algorithm to solve the classification problem of galaxies,stars and quasars,and the algorithm with better classification performance is improved.The main research contents of this thesis are as follows :(1)Acquisition and preprocessing of astronomical image data.In this thesis,we first select the high quality and large number of sources in the Sloan Digital Sky Survey(SDSS).Secondly,because the astronomical image is stored in the form of a single-band FITS file,the g,r and i bands are selected for image synthesis.After that,the target features of galaxies,stars and quasars in the composite image are cut.In order to verify the influence of image size on the experimental results,this thesis cuts into four different size images,and finally preprocesses the data,including scaling,vertical flipping,and horizontal flipping.(2)Application of classical and advanced convolutional neural network algorithms in automatic classification of celestial targets.Firstly,three classical convolutional neural algorithms(AlexNet,GoogLeNet-V1,VGGNet)and four advanced lightweight algorithms(MoblieNet,ConvNet,ShuffleNet,DenseNet)were used to classify celestial targets automatically.Then,through the comparison of experimental results,it is verified that the GoogLeNet-V1 algorithm has superiority in the classification of galaxies,stars and quasars in astronomical images.Finally,we apply six different GoogLeNet series algorithms(GoogLeNet-V2,GoogLeNet-V3,GoogLeNet-V4,Inception-ResNet-V1,Inception-ResNet-V2,Xception)to the classification of celestial target sources.The experimental results show that the Xception algorithm has the best effect among the six GoogLeNet series algorithms.(3)An improved celestial object automatic classification algorithm based on Xception model structure-Xception-AS is proposed.Based on the excellent Xception model structure,the algorithm is improved by adding different attention mechanisms and updating the selection activation function to further improve the efficiency,so that the automatic classification of celestial target sources can be realized efficiently.Through the comparison of experimental results,this thesis shows that the improved algorithm-Xception-AS proposed in this thesis has good classification accuracy in solving the classification problems of galaxies,stars and quasars.