Feature Extraction And Automatic Search Of Galaxies In Wide Field Surveys

Astronomy is an observational subject with a long history.Accompanied with the progress of science and technology as well as the continuous upgrading of observation equipment,our understanding of the universe continues to improve.With the advent of the era of wide-field survey,our have obtained a large amount of galaxy observation data,which has improved the data base for data-driven research methods.Researchers can better study the characteristics and properties of galaxies in astronomy,meanwhile,machine learning,deep learning and other methods have been widely applied to this kind of work.This thesis aims to research the feature extraction and automatic search of galaxies in wide-field surveys,which includes the three parts.The first part is feature extraction and automatic search of gravitational lens system based on deep learning.Gravitational lens system is an important reference index to study the distribution of galaxies and dark matter,so studying gravitational lens system is of great significance to the excavation and exploration of galaxies.In this part,supervised learning method of deep learning is applied to feature extraction and automatic search research of gravitational lens system.Feature extraction of images is carried out by convolutional neural network,and then dimension reduction of high-dimensional images is carried out by pooling layer,and finally sent to the activation layer for classification.This method avoids the manpower consumption in human eye recognition and the complex feature engineering in traditional modeling methods.Experimental results show that this method is superior to traditional machine learning methods in classification.The second part is extraction and automated classification of galaxy morphological features based on few shot learning.Although deep learning can help us complete the classification task well,traditional supervised deep learning has strong dependence on the size and authenticity of data sets(as model dependence problem).In order to avoid the model dependence problem,we try to use few shot learning method to classify galaxy images,and use small and real data sets to realize feature extraction and automatic classification of galaxies.This study proposed the Siamese Convolutional Network(SC-Net)model.The experiment found that when the data set was 10000,the model reaches the accuracy of the traditional supervised method when the data set was 28793.In addition,when the data set was of the same size,SC-Net model is better than the traditional supervised method.The final part is optimization and improvement of SC-Net model based on few shot learning.In the previous work,we verified the feasibility of few shot learning in galaxy morphological feature extraction and automatic search,and proposed the SC-Net model.In this part,SC-Net model is optimized and improved from two dimensions of similarity measurement method and network structure.The experimental results show that the improved method can improve the model effect by2% when the training parameters are reduced by 58%.