Open AccessGeneralization of Convolutional Neural Networks for Searching for O-Star Spectra Using Generative Adversarial Networks
Author(s) -
Zipeng Zheng,
Bo Qiu
Publication year - 2020
Publication title -
journal of physics. conference series
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.21
H-Index - 85
eISSN - 1742-6596
pISSN - 1742-6588
DOI - 10.1088/1742-6596/1626/1/012017
Subject(s) - convolutional neural network , classifier (uml) , computer science , pattern recognition (psychology) , artificial intelligence , spectral line , generalization , data set , star (game theory) , stellar classification , artificial neural network , spectrum (functional analysis) , set (abstract data type) , stars , mathematics , astrophysics , physics , computer vision , astronomy , mathematical analysis , quantum mechanics , programming language
In the stellar spectral data released by LAMSOT, the O-star spectrum is very rare, and the total amount of O-star spectra that can be utilized is only 156. We recommend generating a simulated real spectrum to overcome the above limitations. Using the real O-star spectrum as the model spectral image, we propose a one-dimensional spectral generation confrontation network (1D SGAN) to create artificial spectra based on real data sets. We use a combination of real and artificial spectra to train a one-dimensional convolutional neural network (1D CNN) to create a classifier that classifies the stellar spectra into seven categories. We demonstrate that using the proposed balanced data set with 1D SGAN generated images improves the performance of the 1D CNN classifier compared to the same 1D CNN trained with only the original data set.