Research On Automatic Classification Algorithm Of Astronomical Objects Based On Convolutional Neural Network

With the advancement of science and technology,astronomical observation has developed to a new stage.Large-scale sky survey projects are in full swing around the world.With the development of large-scale sky survey projects,the amount of astronomical data increases exponentially.In the face of huge astronomical data,conventional artificial classification methods of astronomical objects have not been able to meet the actual demand.How to classify astronomical objects efficiently and quickly has become a difficult problem for astronomical researchers to solve.As deep learning has taken off in the field of computer vision in recent years,astronomical researchers have also begun to experiment with applying deep learning algorithms to astronomical object classification tasks.Based on the convolutional neural network in deep learning and the data obtained from the Sloan Digital Sky Survey(SDSS),this thesis studies the classification of galaxies,quasars and stars by mainstream deep learning algorithms,and makes a series of improvements to the algorithm.The main contents are as follows:(1)The preprocessing of astronomical image data.The image data used in this thesis are from the fourth release of the fourth phase of the Sloan Digital Sky Survey.According to the characteristics of the image data released by the Sloan Digital Sky Survey,a series of data preprocessing including image synthesis,object cropping,image denoising and data augmentation are carried out in this thesis.Then the processed data is divided into training set,verification set and test set for the subsequent experiment.(2)Research on automatic classification algorithm of astronomical objects based on convolutional neural network.Nine mainstream neural network algorithms are used to classify galaxies,quasars and stars,and the experimental results are compared.The experimental results show that ResNet50 has a good classification performance for galaxies,quasars and stars.(3)Based on the comparative analysis of nine deep learning algorithms,an improved deep convolutional neural network ResNet50-GQS is proposed in this thesis.Based on the deep residual network,the network aims to efficiently classify galaxies,quasars and stars by adding attention mechanism,modifying the convolution kernel size,improving the residual unit structure,reducing the activation function and the use of normalization.By comparing the experimental results,the superiority of Resnet 50-GQS in classifying galaxies,quasars and stars is verified in this thesis.