Detection and classification of objects in Metis solar corona images using deep learning

Detection of star-like objects in the background of solar corona images is a challenging task, as faint signals are often embedded within complex coronal structures and instrumental noise. Apart from the real stars, these objects may include sun-grazing comets, asteroids, or debris, which are of interest for further research. Traditional analysis methods, however, often rely on cataloging and manual inspection of their corresponding images, an approach that is not scalable for large datasets. As a possible solution, we present a deep-learning framework for automatic object detection using data from the Metis coronagraph onboard the Solar Orbiter ESA Mission. Starting from raw Level 0 (L0) visible-light (VL) images, our pipeline extracts candidate images cutouts associated with potential objects. Stars and irrelevant features, such as noisy samples, bright pixels, or broken images, are systematically discarded. The refined dataset is then pre-labeled using clustering methods, enabling the construction of a new catalog. Then a curated subset of this catalog is used to train a customized YOLO-based model optimized for detecting small-scale objects in full-resolution (2048 × 2048) images. This work highlights the potential of combining systematic cataloging with deep learning to automate the detection of objects in the Metis dataset, thereby supporting further scientific investigation and expanding its detection capabilities.