Core Concepts
A novel approach combining convolutional neural networks and Bayesian networks to effectively capture the hierarchical structure of galaxy morphology and improve classification accuracy.
Abstract
This work proposes a novel method called Bayesian and Convolutional Neural Networks (BCNN) for hierarchical morphological classification of galaxies. The method consists of two main components:
A convolutional neural network (CNN) is trained on images of different galaxy classes to output the probability for each class in the hierarchy.
The CNN's outputs are then fed into a Bayesian network that represents the hierarchical structure of the galaxy classes. The Bayesian network helps improve the prediction accuracy by combining the CNN's outputs while maintaining the hierarchical constraint, where an instance associated with a node must also be associated with all its ancestors.
The authors also address several challenges in galaxy classification, including class imbalance, image quality variations, and the need for models that can generalize well across different surveys and observing conditions. To tackle these issues, the authors employ data augmentation techniques and fine-tune the CNN model.
Experiments on a dataset of galaxy images show that the proposed BCNN method outperforms several standalone CNN models, achieving 67% exact match, 78% accuracy, and 83% hierarchical F-measure. The incorporation of the Bayesian network, data augmentation, and fine-tuning all contribute to the improved performance.
The authors conclude that the BCNN approach effectively captures the hierarchical structure of galaxy morphology and provides a robust solution for this challenging classification task.
Stats
The dataset contains 1931 galaxy images, with the number of images per class shown in Table 2.
Quotes
"Galaxy morphological classification is essential for understanding galaxies evolution and studying stellar populations and their physical properties."
"The ambiguity in the diverse and complex morphology that some galaxies may present makes it difficult to define clear boundaries between the classes, allowing overlap between them, which hinders accurate classification."
"As data sets grow in size and complexity, accuracy and computational efficiency become significant challenges."