Demystifying Galaxy Classification: An Elegant and Powerful Hybrid Approach

The process of identifying a galaxy's structure from telescopic images and then classifying it into its appropriate group has long been a topic of extensive research. The importance of this classification task stems from the fact that the structure as well as the physical properties such as color and texture hold important clues to the contents of the galaxy, its age, and the course of its evolution over time. In fact, the nature of its evolution process provides insights towards prediction of its future state and the time required to reach it. Consequently, such classification studies of galaxies shed light on the understanding of the large-scale structure of the universe as well as its course of evolution. This article proposes a novel, simple yet powerful approach for galaxy classification. One of the major contributions of the proposed approach includes a well-designed effective pre-processing stage to extract only the galaxy portion from the noisy telescopic images, which is itself an important but challenging task. This is especially necessary since noise and artifacts in telescopic images adversely affect the training of the classification network and interfere with the learning of the useful galaxy features. Another significant contribution is the introduction of CoAtNet0-R deep architecture, which is an augmentation of the existing CoAtNet-o network by introducing Relative Self Attention for boosting its performance. The incorporation of two additional structural features for fine-tuning the classifier performance is yet another feather in the cap. The recently annotated dataset Galaxy10 DECals is used for experimentation. The performance of the proposed algorithm for classification into 4 and 10 classes of galaxy structures establishes, in no uncertain terms, the power of the proposed strategy. Hence, this work effectively gives rise to an elegant, hybrid scheme that achieves classification of galaxies at a level hitherto unattained by state-of-the-art networks.