Galaxy Morphology Classification Using Multi-Scale Convolution Capsule Network

Abstract

The classification of galaxy morphology is a hot issue in astronomical research. Although significant progress has been made in the last decade in classifying galaxy morphology using deep learning technology, there are still some deficiencies in spatial feature representation and classification accuracy. In this study, we present a multi-scale convolutional capsule network (MSCCN) model for the classification of galaxy morphology. First, this model improves the convolutional layers through using a multi-branch structure to extract multi-scale hidden features of galaxy images. In order to further explore the hidden information in the features, the multi-scale features are encapsulated and fed into the capsule layer. Second, we use a sigmoid function to replace the softmax function in dynamic routing, which can enhance the robustness of MSCCN. Finally, the classification model achieving 97% accuracy, 96% precision, 98% recall, and 97% F1-score under macroscopic averaging. In addition, a more comprehensive model evaluation were accomplished in this study. We visualized the morphological features for the part of sample set, which using the t-distributed stochastic neighbor embedding (t-SNE) algorithm. The results shows that the model has the better generalization ability and robustness, it can be effectively used in the galaxy morphological classification.