| There are some rare celestial objects with important research value in the universe.Intelligent and efficient identification of rare stars from the massive astronomical observation data can better assist astronomers to study the universe.Due to the lack of rare celestial objects data and few identifiable features,the traditional rare celestial objects recognition method requires manual selection of features,which cannot realize intelligent recognition.This thesis studies the data enhancement and recognition model for rare celestial objects from the following three aspects.Firstly,this thesis introduces the preprocessing of spectral data and image data and proposes a rare celestial objects spectral data augmentation model LCGAN based on locally connected convolution and generative adversarial network.Experiments prove that compared with the existing spectral data enhancement methods,LCGAN model can learn the features of different bands and generate rare celestial objects data with authenticity and diversity.Secondly,by analyzing the characteristics of spectral data,this thesis proposes a rare celestial objects spectral recognition model FACR based on deep learning.In this model,spectral features are extracted by convolutional neural network,and important features are focused by Multi-Head Self-Attention.Moreover,indistinguishable samples are focused by Focal Loss,and residual structure is designed to simplify model training.Experimens in identifying easily confused stars prove that the FACR model has some advantages over the existing methods.Finally,for celestial images containing a small amount of information,this thesis introduces spectral data as an auxiliary,and proposes a multimodal rare celestial objects recognition model AOR-CR.The model uses the multimodal coordinated learning method to convert the image features into the spectral feature space,and finally combines the image features and the converted spectral features to jointly complete the rare star recognition task.Experiments on datasets of QSO and Star show that the model can learn the prior knowledge of spectra during training and can better complete the task of rare celestial objects recognition than using image data alone. |
