EfficientNetV2-S-Triplet7: 一种改进的星系newline 形态学分类算法

星系的形态与星系的形成和演化息息相关, 其形态学分类是星系天文学后续研究的重要一环. 当前海量天文观测数据的出现使得天文数据自动分析方法越来越得到重视, 针对此问题, 利用先进的深度学习骨干网络EfficientNetV2, 分析不同的注意力机制类型和使用节点对网络性能的影响, 构建了一种命名为EfficientNetV2-S-Triplet7 (即在EfficientNetV2-S stage7的$1\times1$卷积层后加入Triplet模块)的改进算法模型来实现星系形态学的自动分类. 使用第二期星系动物园(Galaxy Zoo 2, GZ2)中超过24万张的测光图像作为初始数据进行实验测试. 在对数据进行预处理时采取了尺寸抖动、翻转、色彩畸变等图像增强手段来解决图像数量的不平衡问题. 在同一系列经典和前沿的深度学习算法模型AlexNet、ResNet-34、MobileNetV2、RegNet进行对比实验后, 得出EfficientNetV2-S-Triplet7算法在分类准确率、查全率和F1分数等指标上具有最好的测试结果. 在9375张测试图像中的3项指标值分别可达到89.03%、90.21%、89.93%, 查准率达到89.69%, 在其他模型中排在第3位. 该结果表明将EfficientNetV2-S-Triplet7算法应用于大规模星系数据的形态学分类任务中有很好的效果.

EfficientNetV2-S-Triplet7: An Improved Algorithm for Morphological Galaxy Classification

The morphology of galaxies is closely related to the formation and evolution of galaxies, and its morphological classification is a significant part of the follow-up research of galaxy astronomy. With the emergence of massive astronomical observation data, the automatic analysis of astronomical data has attracted more and more attention. To solve this problem, the advanced deep learning backbone network EfficientNetV2 is utilized to analyze the effects of different attention mechanism types and usage nodes on network performance, and an improved algorithm model named EfficientNetV2-S-Triplet7 is constructed to realize automatic classification of galaxy morphology. More than 240 thousand photometric images from Galaxy Zoo 2 are used as initial data for experimental tests. In the process of data preprocessing, image enhancement methods such as size jittering, flipping and color distortion are adopted to solve the problem of image number imbalance. After conducting comparative experiments on the same series of classic and cutting-edge deep learning algorithms AlexNet, RegNet, MobileNetV2 and ResNet-34, it is concluded that the EfficientNetV2-S-Triplet7 algorithm has the best test results in classification accuracy, recall and F1-score. In 9375 test images, the three index values can reach 89.03%, 90.21% and 89.93%, respectively, and the precision can reach 89.69%, ranking the third among other models. The results show that EfficientNetV2-S-Triplet7 algorithm can be effectively applied to the morphological classification of large-scale galaxy data.