Research On Automatic Search Method Of Low Surface Brightness Galaxies Based On Deep Learning

Low surface brightness galaxies(LSBG)is an important part of the galaxy field.However,its darker surface brightness limits people's detection and search for it,resulting in the fact that the number of low surface brightness galaxies currently found is less.The contribution of surface brightness galaxies to the universe is seriously underestimated.Traditional methods of searching for low surface brightness galaxies include source selection,skylight reduction,and brightness calculation.Some of these steps are prone to failure and require manual intervention,making it difficult to achieve automated batch search.In order to efficiently search for low surface brightness galaxies from massive astronomical images and expand the sample of low surface brightness galaxies,this paper builds a new model for automatically searching for low surface brightness galaxies in Sloan Digital Sky Survey(SDSS)images-LSBG Auto Detect(LSBG-AD).LSBG-AD is a deep learning object detection framework based on convolutional neural networks and borrowed from the idea of YOLO.It first divides the SDSS image into a certain number of small cells,then uses the Residual Network(ResNet)to extract the features of all small cells at the same time,and then fits the extracted feature map to each small cell using a convolution network with a convolution kernel size of 1*1 to the probability of contain the target and the coordinate location of the target.Finally,the output of LSBG-AD summarizes the results of all the small cells.The training sample of LSBG-AD contains 1129 low-surface brightness galaxies.In this paper,we first obtain 1120 SDSS images based on the coordinates of these 1129 galaxies,and then process the data with band alignment,pixel compression and labeling according to the characteristics of astronomical images to adapt to the deep learning model.At the same time,in order to improve the recall rate of LSBG-AD,this paper puts forward a strategy of multiple searches according to the data characteristics,which effectively improves the recall rate of LSBG-AD.In this paper,1120 SDSS images were searched using trained LSBG-AD to obtain 1197 low surface brightness galaxies,of which 1081 were marked in training samples and 116 were newly discovered candidates.After verification,96.46%of the candidates are non-edge-on galaxies;92.04%of the candidates are disk-dominated galaxies and the characteristics of these candidates were consistent with those of the training samples.Since low surface brightness galaxies are generally determined by the brightness of the B-band center facet?0(B),this paper uses parameters from the SDSS database to calculate the center-facet brightness of candidates,and finds that the average value of the candidate is 22.6 mag arcsec-2,and 97.35%of the candidates are darker than 22.5 mag arcsec-2,which meets the conditions of low surface brightness galaxies.These conclusions indicate that most of the candidates are low surface brightness galaxies.In this paper,YOLOv5 and Faster RCNN are used to do comparative experiments,the results show that LSBG-AD has a higher recall and accuracy in low surface brightness galaxy search,while comparing traditional astronomical methods:in 93 SDSS images the traditional astronomical methods searched for 47 low surface brightness galaxies darker than 22.5 mag arcsec-2,46 of which were identified by LSBG-AD,and LSBG-AD also searched for 32 candidates,which shows that LSBG-AD has strong searching ability for low surface brightness galaxies darker than 22.5 mag arcsec-2.