Automatic Classification Of Galaxy Morphology Based On The RegNet Algorithm

The morphological classification of galaxies is a key tool for understanding galaxy formation and evolution and plays a crucial role in modern astrophysics.With the explosion of data growth,deep learning has been applied to research the classification of galaxy morphology.However,it remains an important problem how to design deep convolutional neural networks(CNNs)so that they can efficiently classify galaxy morphologies.In the above background,several good deep learning algorithms from the Image Net dataset are applied to astronomical images.By comparing the experimental results,the algorithm with the best classification performance is selected as the basic model,and then the algorithm is improved.The main research contents of this thesis are as follows:(1)The morphology of galaxies is determined to be divided into seven categories:lenticular,barred spiral,spiral,completely round smooth,in-between smooth,cigarshaped smooth and irregular galaxy.The clean samples of pre-class galaxies are selected from the Extraction de Formes Id’ealis’ees de Galaxies en Imagerie(EFIGI)and Galaxy Zoo 2(GZ2)databases to form a new data set by using EFIGI Hubble sequence classification and GZ2 decision tree.Based on the features of the image,the pre-processing operations are performed,including centre cropping,down sampling,denoising,data augmentation,normalization,and standardization.(2)Together with Alex Net,VGG,Goog LeNet,ResNet-34,Dense Net,Mobile Net,Shuffle Net,Efficient Net and Efficient Net V2,the Reg Net algorithm is applied to the galaxy image dataset for training,validation and testing.Experimental results show that the Reg Net model performs better than other neural networks and can be effectively applied to the task of automatic classification of galaxy morphologies.(3)Based on RegNet X-4.0GF network,this thesis studies 10 improvement strategies from the perspective of attention mechanism and regularization.The experimental results are analyzed in the same setting,and the optimal model is identified as Reg Net XCBAM3,where only the Convolutional Block Attention Module is added to the block structure with stride one.The accuracy of the Reg Net X-CBAM3 model on the test set reached 92.02%,and the average precision,recall,F1-score and AUC value reached92.14%,92.13%,92.10% and 0.9827,respectively.Moreover,considering the effect of classification bias on the morphological classification of galaxies,this thesis establishes a method to calculate probability confidence based on Softmax output values,and explores the relationship between the performance of the Reg Net X-CBAM3 model and the redshift in the EFIGI data set.Finally,the working mechanism of CNNs has been studied using feature map visualization,and some valuable findings have been obtained.