Automatic classification of galaxy morphology based on the RegNetX-CBAM3 algorithm

Abstract

ABSTRACT This paper focuses on the automatic classification of galaxy morphology based on deep learning. Through applying a variety of improvement strategies and comparing the results of a large number of experiments, an algorithm named RegNetX-CBAM3 with good performance is proposed to implement the task of automatic classification of galaxy morphology. The RegNetX-CBAM3 algorithm is applied along with many other popular neural networks in the data set consisting of the Extraction de Formes Idéalisées de Galaxies en Imagerie (EFIGI) catalogue and Galaxy Zoo 2 (GZ2), and there are the following seven types of the galaxy morphology in this data set: lenticular, barred spiral, spiral, completely round smooth, in-between smooth, cigar-shaped smooth, and irregular, respectively. Experimental results show that the RegNetX-CBAM3 algorithm achieves the state-of-the-art results over many other excellent algorithms, with the accuracy of 0.9202, purity of 0.9214, completeness of 0.9213, F1-score of 0.9210, and AUC value of 0.9827 on the test set. Moreover, we establish a method of probability confidence calculation considering the classification bias. The confidence degree of galaxies calculated by this method is basically consistent with that of GZ2 and EFIGI, which demonstrates the rationality of this method and also proves that the RegNetX-CBAM3 algorithm can effectively classify galaxies. Therefore, the RegNetX-CBAM3 algorithm can be applied to effectively solve the problem of automatic classification of galaxy morphology. On EFIGI data, the performance of the RegNetX-CBAM3 algorithm does not change substantially with the redshift range. In addition, it should be noted that the use of deep neural networks, manual annotation, and data enhancement may cause classification bias in galaxy images.