Maksim Gabdeev (Pavol Jozef Šafárik University in Košice): Morphological classification of eclipsing binary stars using computer vision methods
→
Europe/Stockholm
90102
90102
Description
Abstract:
Developing robust methods for the automatic classification and parameters prediction of eclipsing binaries are crucial for leveraging the vast datasets produced by modern photometric surveys. This presentation focuses on our recent work applying computer vision (CV) techniques to the morphological classification of EB light curves. We investigated the effectiveness of fine-tuning pre-trained CV models for this astronomical task, specifically comparing the performance of two prominent architectures: ResNet50 a Convolutional Neural Network and Vision Transformers. A key challenge was adapting these models, designed for natural images, to work effectively with light curves. We addressed this by proposing and testing a novel light curve-to-image transformation method based on polar coordinates and hexbin density mapping. This transformation proved essential in overcoming overfitting issues encountered with standard representation of a light curve for binary classification distinguishing detached and overcontact systems. We trained models on the same image representation to identify starspots in the systems. Highlighting at the end the success achieved, particularly in binary classification on real observational data, and discussing the future challenges for fully automated analysis and parameters prediction.
