Method Research On L Dwarf Search Based On Deep Learning

In the past two decades,long-band and large-field survey projects,such as Sloan Digital Sky Survey(SDSS)and Two Micron All Sky Survey(2MASS),have been put into operation,and more and more astronomical photometric data have been accumulated.These huge amounts of astronomical photometric data have raised higher demands on astronomical research methods.In order to efficiently find more L dwarfs from the large amount of photometric images of SDSS and expand the sample of L dwarfs,this thesis uses deep learning to conduct related research.Currently,spectra need to be taken for L-dwarf identification,but the speed of taking spectra is much slower than the speed of taking photometric images.Therefore,it is necessary to use photometric images for initial source selection and further capture of spectra.The current photometric image source selection method does not use the photometric image directly,but extracts the parameters from the photometric image now and then performs the L dwarf selection,which may not only lose some image features in the photometric parameter extraction,but also may attract errors in the process of photometric parameter extraction.In order to solve the above problems,the following study is conducted in this thesis:(1)We identify a sample of 521 L dwarfs,and process the SDSS photometric images of these L dwarfs with pixel alignment and JPEG image synthesis,and annotate the L dwarfs in the images to produce a sample set of L dwarf images that can be used for deep learning.(2)Aiming at the characteristics of small size and large scale variation of L dwarf stars,this thesis designs L dwarf detection models based on Faster R-CNN,combining the ideas of Resnet and feature pyramid.And the noise elimination module is designed for the noise that may exist in the L dwarf training samples.Using 421 L dwarf images for training,we obtain the able L dwarf detection model LDAD.The effectiveness and correctness of LDAD in searching L dwarf stars are proved by the results of the model in several sample sets.(3)The LDAD model was used to search for L dwarfs from 941 SDSS images(covering a sky area of about 32.8238 square degrees),and the astronomical properties of the searched L dwarf candidates were analyzed,resulting in 85 new L dwarf candidates.(4)An L-dwarf detection system based on the LDAD model is designed and implemented.The system integrates the functions of downloading SDSS astronomical images,aligning and synthesizing multi-band images,and detecting L dwarfs using LDAD model,and adds the functions of image management,image preview and download,and export of results,so that users can easily detect L dwarfs from images of specified areas through web pages and facilitate their follow-up studies.The LDAD model studied in this thesis can effectively search for L dwarfs from astronomical photometric images.Compared with the traditional L dwarf search algorithm,the LDAD model directly uses the photometric images to fully exploit the information in the photometric images without relying on the photometric parameters,which can avoid the errors that may be introduced in the extraction of photometric parameters.This model will detect all the images of SDSS DR16 to perform a large-scale search for L dwarfs and expand the study sample of L dwarfs.