基于EfficientNet的星系形态分类研究

星系的结构和形态能够反映星系自身的物理性质,其形态的分类是后续分析研究的一个重要环节.EfficientNet模型使用复合系数对深度网络模型的深度、宽度、输入图像分辨率进行更加结构化的统一缩放,是一种新的深度网络优化扩展方法.将该模型应用于星系数据形态的分类研究中,结果表明基于EfficientNetB5模型的平均准确率、精确率、召回率以及F1分数(精确率与召回率的调和平均数)都在96.6%以上,与残差网络(Residual network, ResNet)中ResNet-26模型的分类结果相比有较大的提升.实验结果证明EfficientNet的深度网络优化扩展方法可行且有效,可应用于星系的形态分类.

Research on Galaxy Morphology Classification Based on EfficientNet

The structure and morphology of galaxies can reflect the physical properties of the galaxy itself, and the classification of its morphology is an important part of subsequent analysis and research. The EfficientNet model uses composite coefficients to unify the depth, width, and input image resolution of the deep network model in a more structured manner. This is a new deep network optimization and extension method. This study applies the model to the classification of galaxy data morphology, and the results show that the average accuracy, precision, recall and F1 score (Harmonic mean of precision and recall) based on the EfficientNet-B5 model are all large than 96.6%, which is a significant improvement compared with the classification results of the ResNet-26 model. The experimental results prove that the deep network optimization extension method of EfficientNet is feasible and effective, and can be applied to the morphological classification of galaxies.