Due to increasing amount of data, galaxy image classification is being automated using machine learning, with the most common approach being convolutional neural networks (CNNs). Using domain-specific observations about galaxy morphology, we propose two CNN layers – Focal and Wave. The Focal layer highlights information in the center, while the Wave layer is designed to highlight spiral arms. A comprehensive evaluation of layers is performed using datasets of 4 and 11 (not yet deeply studied) galaxy types, with and without augmentations. With no augmentations, our models outperform best results of models used in other studies with Wilcoxon signed-rank test's p-value of pF1 = 0.002 for both 4 and 11 classes in terms of F1 score (used to better evaluate performance due to inherent class imbalance). Moreover, our layers help reduce overfitting and reliance on augmentation combinations, as our models, despite benefiting less from augmentations, still achieve results comparable to other augmented models. Additionally, our models tuned for 4-class data perform equally well on 11-class data, and vice versa. Finally, an ensemble of our models trained without augmentations achieves comparable results to our augmented models.
This work is licensed under a Creative Commons Attribution 4.0 International License.