Automatic Search For Large Samples Of LSB Galaxies For The Sloan Digital Sky Survey

Low surface brightness galaxies(LSBG)are a class of galaxies with a faint surface brightness of at least one magnitude fainter than the night sky.Low surface brightness galaxies are an important part of the Universe,and the study of their properties is important for exploring the formation of galaxies and the evolution of the Universe.However,due to the difficulty of detecting low surface brightness galaxies,the number of known samples of low surface brightness galaxies is much lower than the estimated value.Traditional methods for detecting low surface brightness galaxies require multiple steps such as sky light reduction and surface source photometry,and require manual setting of model parameters,which is not suitable for handling large amounts of astronomical data.In this paper,we propose a deep learning-based method for detecting low surface brightness galaxies,which aims to locate and classify low surface brightness galaxies from astronomical photometric images more efficiently.Instead of annotating a bounding box,the method uses a smooth response region created by a Gaussian function to represent the range of the target’s centroid position.Using this target detection model,this paper has detected about 930,000 astronomical images,resulting in the largest Sloan Digital Sky Survey catalog of low surface brightness galaxies.The research in this paper focuses on:(1)Construction of a sample dataset of low surface brightness galaxies for model training.The Sloan Digital Sky Survey Science Archive Server was used to screen and acquire a sample of 765 certified low surface brightness galaxies for model training and validation.(2)A target detection model for identifying low surface brightness galaxies from Sloan Digital Sky Survey photometric images is developed.To address the problems of tedious annotation process and centroid coordinate shift in the rectangular candidate frame based detection method,a method is proposed to predict only the target centroid.The experimental results show that compared with the bounding box-based method,the centroid-based target detection method proposed in this paper shows better performance in terms of reducing the predicted centroid offset as well as the accuracy and recall rate.(3)A deep learning regression model for predicting the photometric parameters of low surface brightness galaxies from Sloan Digital Sky Survey images is developed.The fully connected layers of the VGG16,ResNet50 and DenseNet 121 networks are modified to predict the photometric parameters such as surface brightness,magnitude and effective radius of low surface brightness galaxies in the images.The experimental results show that the DenseNet121 network-based model outperforms other models in predicting galaxy photometric parameters and is more efficient than traditional methods.(4)The Sloan Digital Sky Survey(SDSS)images are detected using the established low surface brightness galaxy detection model.The process is as follows:firstly,930,000 g,r and i-band synthetic images of the Sloan Digital Sky Survey DR16 are acquired;secondly,the trained model is used to detect low surface brightness galaxies in these images;then,duplicate sources are removed from the candidates and the anomaly is removed using the One-class Deep Support Vector Data Description)anomaly detection method is used to remove the anomalous data.The established parametric regression model was used to predict the parameters of the low surface brightness galaxy candidates,resulting in a catalog of 37,563 low surface brightness galaxy candidates.This study is the first to use a target detection algorithm without bounding boxes to detect low surface brightness galaxies from astronomical images,and to search and build the largest Sloan Digital Sky Survey low surface brightness galaxy catalog so far,providing a richer sample data for the study of low surface brightness galaxies.