.png){kind=icon}
Return
Galaxy morphological classification using Dynamic Multiscale Attention Network
.png){kind=icon}
OA
AI
DOI：10.1093/mnras/staf1037.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The implementation of large-scale digital sky surveys has collected enormous photometric images, making galaxy morphology classification a current research hotspot. This paper presents a galaxy morphology classification network based on dynamic large convolution kernels and attention feature fusion, named as the Dynamic Multiscale Attention Galaxy Network (DMAGNet). In this study, 15 266 galaxy images from the Galaxy Zoo DECaLS (Dark Energy Camera Legacy Survey) project are used with six categories of galaxy morphologies. The six categories are edge-on, cigar, in-between, round, spiral, and merger. On the test set, the model is demonstrated with superior performance, achieving 97.1 per cent accuracy, 96.8 per cent recall, and a 96.8 per cent F1 score. Furthermore, the proposed network is subjected to extensive evaluation and ablation experiments, and the t-distributed stochastic neighbour embedding (t-SNE) method is used to visualize the morphological features extracted from the test set. The accuracy of Maxvit is 0.969, while the accuracy of DMAGNet reaches 0.971. The experimental results show that this model outperforms known classification models in the galaxy morphology classification task. Finally, this method was applied to 14 174 190 galaxy images from the BASS + MzLS data set of the DESI Legacy Surveys, constructing a galaxy catalogue covering six morphological categories: Edge-On (192 577), In-Between (2353 937), Cigar (299 725), Round (353 358), Spiral (16 067), and Merger (1266 961). Additionally, an Error column was introduced, containing a total of 9691 564 samples, to label images that may be damaged or have quality issues by incorporating information from other catalogues.
Keywords:
methods: data analysis
techniques: image processing
Galaxies: structure
Journal
IF:
4.8
Papers:
7.0W
Citations:
25.0W
Organization
Cited Papers
.png){kind=icon}
ShareSaveShareSave
View more