Fulin Peng, Yunmeng Liu, Lei Ding
Galaxy morphology classification is a fundamental task for studying galaxy formation and evolution. While deep learning has significantly improved automated classification, most existing approaches rely mainly on visual features and show limited capability in fine-grained morphology recognition, particularly when structural differences are subtle. This work aims to improve fine-grained galaxy morphology classification by combining learned image representations with structural descriptors and by enhancing robustness to arbitrary galaxy orientations. We propose DBMR-GalaxyNet, a galaxy morphology classification framework that combines learned image representations with structural descriptors. A ConvNeXt backbone is used to extract hierarchical representations, followed by a multi-scale rotation-equivariant adapter to improve orientation robustness. A feature pyramid network aggregates the features, after which two branches are constructed: a visual branch that encodes morphology cues and a structural-descriptor branch trained with auxiliary regression targets computed from the input images. The two representations are fused through a learnable gating mechanism and optimised using a joint classification and regression objective. Our method was evaluated on two datasets derived from Galaxy Zoo: a seven-class Galaxy7 dataset and the Galaxy10 DECaLS dataset for ten-class classification. The proposed model achieves five-run mean accuracies of (95.21 ± 0.25%) on Galaxy7 and (90.45 ± 0.18%) on Galaxy10 DECaLS, indicating stable and competitive performance for fine-grained galaxy morphology classification. Ablation studies show that both the multi-scale rotation-equivariant adapter and the dual-branch design contribute to the performance improvement. The proposed dual-branch framework demonstrates strong performance and provides a practical approach for incorporating structural descriptors into deep learning models for large-scale galaxy surveys.