Research On Spectral Noise Reduction Based On Generative Adversarial Network

Astronomy is an ancient discipline,and astronomical research is closely related to human activities,which profoundly affects human's views on nature,the universe,human daily production and life,and even human social views.The cross study of the of astronomy and artificial intelligence will definitely bring new collisions,promote each other,and bring better development opportunities for each other.At present,certain progress has been made in the use of artificial intelligence technology for astronomical research,such as the use of machine learning and deep learning technologies for spectral clustering,feature selection,and automatic classification.However,there are many problems that need to be solved in these studies,such as the spectral denoising problem when processing spectra.Problems such as high temperature cause low spectral quality.These spectra show very different spectral line characteristics from normal spectra.They have problems such as large noise,low local signal-to-noise ratio,and abnormal continuous spectrum.The above problem have brought inconvenience to our astronomical research work,which needs to be considered in our study.In order to solve the above problems,this article makes the following researches based on the application of deep learning and machine learning in astronomy:(1)The system summarizes the current noise reduction algorithms used for spectral image noise reduction and their performance characteristics and advantages and disadvantages in image noise reduction.Innovative idea of using image consistency generated by Cyclic Consistency Generative Adversarial Network(CycleGAN)to select the low-mass spectrum and unpaired high-quality spectrum in SDSS for counter training.The mass spectrum is transformed into a high-quality spectrum,which realizing the noise reduction processing of astronomical spectrum data.(2)The performance of the PCA reconstructed spectral method and the algorithm used in this paper in denoising spectral data are compared,and the advantage of the algorithm selected in this paper is highlighted.(3)we have selected a pair of existing original astronomical spectra to perform a noise reduction comparison test,that is,compare the generated noise-reduced astronomical spectrum with a pair of high quality astronomical spectra;We have selected multiple types of classification algorithms,such as ensemble learning(GBDT,XGBoost),a dedicated binary classification algorithm SVM,and compare the classification accuracy and accuracy of the astronomical spectrum before and after the noise reduction Rate and F1 indicators.The results of a large number of experiment show that the cycle-consistent generative adversarial network CycleGAN has achieved good results in noise reduction of spectral data.The astronomical noise reduction model CycleGAN used in this paper can effectively remove high noise and false lines in astronomical spectral data,which completed the reuse of low-quality astronomical spectrum.This can greatly expand the sample size of astronomical spectrum data and solve the problem of data insufficiency data.It is of great significance for astronomical spectrum data parameter measurement and galaxy research.