A Galaxy Morphology Classification Model Based on Momentum Contrastive Learning

Abstract

The taxonomy of galaxy morphology plays an important role in astrophysics and provides great help for the study of galaxy evolution. To integrate the advantages of unsupervised learning without labels and supervised learning with high classification accuracy, this paper proposes a galaxy morphology classification model based on a momentum contrastive learning algorithm named Momentum Contrastive Learning Galaxy (MCL-Galaxy), which mainly includes two parts (i) pre-training of the model, where the ResNet_50 backbone network acts as an encoder to learn the galaxy morphology image features, which are stored in the queue and their consistency is ensured by using the momentum contrastive learning algorithm; and (ii) transfer learning, where Mahalanobis distance can assist in improving classification accuracy in downstream tasks where both encoder and queue are transferred. To evaluate the performance of MCL-Galaxy, we use the data set of the Galaxy Zoo challenge project on Kaggle for comparative testing. The experimental results show that the classification accuracy of MCL-Galaxy can reach 90.12%, which is 8.12% higher than the unsupervised state-of-the-art results. Although it is 3.1% lower than the advanced supervised method, it has the advantage of no label and can achieve a higher accuracy rate at the first epoch of classification iteration. This suggests that the gap between unsupervised and supervised representation learning in the field of Galaxy Morphologies classification tasks is well bridged.