Xception-AS:一种基于Xception算法结构的天体目标自动分类算法  

Xception-AS:An Automatic Object Classification Algorithm Based on the Structure of Xception Model

作　　者：李馨 屠良平[1] 李娟 高翔 冯雪琦 仲峥迪 Li Xin;Tu Liangping;Li Juan;Gao Xiang;Feng Xueqi;Zhong Zhengdi(College of Science,University of Science and Technology Liaoning,Anshan 114000,China)

机构地区：[1]辽宁科技大学理学院,辽宁鞍山114000

出　　处：《天文研究与技术》2023年第3期267-274,共8页Astronomical Research & Technology

基　　金：国家自然科学基金(U1731128)资助.

摘　　要：提出了一种基于Xception结构的天体目标自动分类算法,该算法可以有效应用于星系、恒星和类星体的自动分类。算法以Xception为基础框架,通过选择最优激活函数,加入注意力机制等方式进行改进。随机选择SDSS-DR16测光图像数据中的11543个星系、10490个类星体和11967个恒星共34000个观测源g,r和i共3个波段的图像作为实验数据,并设计多组实验进行算法验证和测试,综合分析所有实验结果得出本文算法在准确率、精确率、召回率和F 1分数等关键指标分别达到了90.26%,90.01%,89.86%和89.85%。相同数据集与其他13种经典和流行的卷积神经网络(Convolutional Neural Network,CNN)算法的实验结果对比表明,本文提出的Xception-AS算法具有更加优异的分类性能,证明本文算法解决天体目标自动分类问题的优越性。In this paper,an algorithm based on Xception is proposed,which can be used to solve the problem of automatic classification of galaxies,stars and quasars.Based on Xception,the algorithm is improved by selecting the optimal activation function and adding attention mechanism.In this paper,11543 galaxies,10490 quasars and 11967 stars in SDSS-DR16 photometric image data are randomly selected as experimental data from 34000 observation sources in g,r and i bands,and multiple experiments are designed to verify and test the algorithm.A comprehensive analysis of all experimental results shows that the algorithm in this paper achieves 90.26%,90.01%,89.86%and 89.85%respectively in the key indicators such as precision rate,accuracy rate,recall rate and F 1 score.Compared with other 12 classical and popular convolutional neural network algorithms on the same data set,the proposed Xception-AS algorithm has better classification performance,which proves that the proposed algorithm has advantages in solving the problem of automatic classification of celestial objects.

关 键 词：天文图像分类 机器学习 Xception 卷积神经网络 

分 类 号：TN911.73[电子电信—通信与信息系统] P111[电子电信—信息与通信工程]