Synergistic application of advanced machine learning and computer vision techniques for the detection of exoplanet and star: leveraging contrastive learning in astronomical imagery

Abstract

The detection of exoplanets through direct imaging has become increasingly feasible with the advent of modern telescopes like the James Webb Space Telescope (JWST) and its MIRI coronagraph. These instruments have significantly expanded the availability of high-quality direct imaging datasets, offering new opportunities for astronomical research. However, despite these advances, traditional machine learning methods, such as Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs), face limitations when applied to exoplanet detection. These limitations include difficulties in detecting exoplanets that are faint and often obscured by the bright light of nearby stars, as well as challenges posed by small and imbalanced datasets. To address these challenges, this thesis introduces a novel machine learning approach that leverages contrastive learning and few-shot learning. Contrastive learning is used to improve the distinguishing ability of the model focusing on separating different celestial objects such as exoplanets and stars. It learns representations that distinguish between these objects despite the high noise, low contrast and complex nature of astronomical images. In addition, to address the problem of few training images, the approach called few-shot learning is used to achieve good results even when only a few pixel-labeled images of exoplanets are available. The components of the proposed framework include integration of contrastive learning with ResNet18 and Siamese Networks followed by experimenting with two YOLO models. The decision to increase the model’s complexity going to provide higher accuracy over the simple method in processing noise contexts or in the limited availability of exoplanet instances. Through the sophisticated learning methodologies presented in this work, our model achieved an accuracy of 92.16%, significantly enhancing detection capabilities. By utilizing YOLOv5, our approach also achieved high performance in Dice Coefficient (1.0 for star and 0.9 for exoplanet) and Intersection over Union (IoU) metrics (1.0 for star and 0.8 for exoplanet), underscoring the model’s effectiveness in direct imaging exoplanet detection. This advancement not only contributes to current astronomy efforts but also enables more powerful discoveries in the future.