| Characteristics of Low Surface Brightness Galaxies(LSBGs)are very important for understanding the global characteristics of galaxies,and it is important to search and expand LSBGs using modern machine learning,especially deep learning algorithms.LSBGs are difficult to be identified automatically and accurately by traditional methods because of their not obvious features,but deep learning has the advantage of automatically finding out complex and effective features.To solve this problem,based on the deep learning algorithm,the LSBGs automatic search problem is studied by using the g,r,i three-band composite images in the large sample sky survey project,and the algorithm is improved.The main work contents are as follows:(1)Data acquisition and image preprocessing.Firstly,the data set used in this experiment is from Low Surface Brightness Galaxies in 40% of the sky area of the Arecibo Legacy Fast Arecibo L-band Feed Array Survey(ALFALFA).According to the characteristics of astronomical images,image preprocessing methods such as cropping,Mosaic data enhancement,random scaling and random distortion image width and height are introduced.Due to the small number of samples in this dataset,this thesis also uses the Deep Convolutional Generative Adversarial Networks(DCGAN)model to expand the experimental data,which is the second dataset used in this thesis The labeled data set is used as the training set and test set of the experiment.(2)LSBGs automatic search algorithm based on deep learning.Five classical target detection algorithms Faster R-CNN,Center Net,YOLOv4,YOLOv5 and YOLOX are applied to the above processed data set,and the experimental results of these algorithms are compared and analyzed,and the optimal algorithm--YOLOX is obtained.Then,different attention mechanisms and optimizers are combined,and pre-training weights are used to speed up training and save training costs.The experimental results of the two data sets are compared and analyzed,and the effectiveness of DCGAN model to improve algorithm performance and the superiority of YOLOX-CS model are verified.The recall rate and Average Precision(AP)value of YOLOX-CS model reach 98.10% and 98.46% in the unexpanded data set of 1106 samples.In the expanded data set of DCGAN model,the recall rate of experiments under the same test set reaches 98.37%,and the AP value reaches98.64%.Finally,the algorithm is put into other data sets for application verification,and the results verify the feasibility and superiority of the algorithm.(3)LSBGs candidate star list.The algorithm proposed in this thesis is applied to some sky areas in Sloan Digital Sky Survey(SDSS)for searching experiments.The astronomical image is clipped overlapped first,and then the clipped image is fed into the algorithm model to get the upper left and lower right coordinates of the target.Then,the program is used to convert the image coordinates into right ascension and declination,and Obj ID and all the important parameters of targets are obtained from the star list of astronomical website according to the declination of right ascension.Finally,a star list containing 765 candidate galaxies with low surface brightness is obtained. |
