Unsupervised Galaxy Morphological Visual Representation with Deep Contrastive Learning

Abstract Galaxy morphology reflects structural properties that contribute to the understanding of formation and evolution galaxies. Deep convolutional networks have proven be very successful in learning hidden features allow for unprecedented performance morphological classification Such mostly follow supervised paradigm, which requires sufficient labeled data training. However, labeling a million galaxies is an expensive complicated process, particularly forthcoming survey projects. In this paper, we present approach, based on contrastive learning, with aim galaxy visual representation using only unlabeled data. Considering low semantic information contour dominated images, feature extraction layer proposed method incorporates vision transformers network provide rich via fusion multi-hierarchy features. We train test our three classifications sets from Zoo 2 SDSS-DR17, four DECaLS. The testing accuracy achieves 94.7%, 96.5% 89.9%, respectively. experiment cross validation demonstrates model possesses transfer generalization ability when applied new sets. code reveals pretrained models are publicly available can easily adapted surveys. 6 https://github.com/kustcn/galaxy_contrastive