基于深度残差网络的星系形态分类

Galaxy Morphology Classification Using Deep Residual Networks

星系形态与星系的形成和演化有着密切的联系,因此星系形态分类(galaxy morphology classification)成为研究不同星系物理特征的重要过程之一。斯隆数字巡天(Sloan Digital Sky Survey, SDSS)等大型巡天计划产生的海量星系图像数据对星系形态的准确、实时分类提出了新的挑战,而深度学习(deep learning)算法能有效应对这类海量星系图片的自动分类考验。面向星系形态分类问题提出了一种改进的深度残差网络(residual network, ResNet),即ResNet-26模型。该模型对残差单元进行改进,减少了网络深度,并增加了网络宽度,实现了对星系形态特征的自动提取、识别和分类。实验结果表明,与Dieleman和ResNet-50等其他流行的卷积神经网络(convolution neural network, CNN)模型相比,ResNet-26模型具有更优的分类性能,可应用于未来大型巡天计划的大规模星系形态分类系统。

Galaxy morphology is closely related to the galaxy formation and evolution,so galaxy morphology classification is one of the most important processes in the study of the physical characteristics of different galaxies.The massive galaxy images data produced by the large scale surveys,such as Sloan Digital Sky Survey(SDSS),poses a new challenge to classifying galaxy images accurately and real time,and deep learning algorithm can effectively and automatically deal with the kind of very large collections of galaxy images.In this paper,a modified residual network(ResNet),namely ResNet-26,is proposed for galaxy morphology classification.This model improves the residual unit,while reduces the depth of the network and widens the width of the network,and realizes the automatic extraction of galaxy morphological features to identification and classification.The experimental results show that ResNet-26has better classification performance compared with other popular convolution neural network models such as Dieleman and ResNet-50,and can be applied to large-scale galaxy classification in forthcoming surveys.