Research On Automatic Search Of Edge-on Low Surface Brightness Galaxies In SDSS

Low surface brightness galaxies(LSBG)are galaxies whose surface brightness is generally fainter than the night sky and occupy a significant proportion of the universe.Edgeon LSBGs are LSBGs with their sides facing toward the earth,compared with face-on LSBGs,edge-on LSBGs can show edge-on structures due to their unique angles.Therefore,the studies of them can help to understand the formation and evolution of LSBGs.Collecting and expanding the sample of edge-on LSBGs is the primary task for further related studies.However,the relevant parameters of edge-on LSBGs are difficult to extract automatically,making it impossible to conduct a large-scale search for edge-on LSBGs.Currently,the Sloan Digital Sky Survey(SDSS),as one of the largest sky survey projects in the world,has released a large number of images.In order to automate the search process for edge-on LSBGs in SDSS,this thesis applies deep learning techniques of object detection and anomaly detection to photometric images,and aims to investigate new methods to search for a sample of edge-on LSBGs directly from photometric images,and make the search and selection process of the samples more efficient.The main study contents of this thesis are as follows:(1)A sample of 281 edge-on LSBGs and 281 corresponding SDSS photometric images are obtained,the dataset is labeled and divided to create a training set and a test set for the object detection algorithm.The object detection algorithm YOLOv5 is used to build the detection model,and the data enhancement method,network structure and loss function with the best training effect are selected according to the characteristics of the training data.After completing the process of building model,the detection effect of the model is evaluated by testing the test set using different confidence thresholds to determine a reasonable confidence threshold for subsequent large-scale search.(2)All SDSS DR16 gri-band composite images are obtained,and the built object detection search model is used to conduct a large-scale,automated search for edge-on LSBGs directly among 938,046 composite images,and the search results are processed subsequently.The anomaly detection algorithm Deep-SVDD is used to build the anomaly detection model,and the anomaly scores of all the candidates are obtained,and the visual inspection method is combined to remove the other objects mixed in the candidates,so as to achieve further purification of the candidates,obtain the catalog of selected candidates,and provide data for subsequent analysis.(3)The obtained catalog of candidates is cross-matched with all galaxies in SDSS DR16,and the unmatched candidates are visually inspected to remove a very small number of objects which are not edge-on LSBGs.The catalog of 49,972 cross-matched candidates and 252 unmatched candidates were obtained,and the properties of the candidate samples are analyzed.The user interface(UI)is designed and implemented to display the images of the candidate samples,and the users are provided with instructions on how to use the detection method in this thesis,so that the users can fully understand how to use the object detection model and the anomaly detection model to conduct the subsequent study of the search for edge-on LSBGs.This study applies deep learning algorithms to search edge-on LSBGs for the first time,searches for the largest sample set of edge-on LSBG candidates to date,and provides the catalog of all the samples.The object detection model built in this study is able to locate edge-on LSBG candidates directly in photometric images by predicting coordinates with known coordinates,free from the dependence on photometric parameters.In addition,the anomaly detection model built in this study can effectively identify anomalous samples,achieving further purification of the detected candidates.This study not only significantly expands the existing samples of edgeon LSBG candidates in terms of quantity,but also searches for edge-on LSBG candidates that are not detected by the SDSS DR16 photometric pipeline.The detection method of this thesis can be used to search for edge-on LSBGs in other sky surveys in the future.