High-contrast, direct imaging of exoplanets is a key objective for the future of astronomy. Underpinning both the development and successful deployment of future direct imaging missions is the ability to analyze the resulting data. Therefore, in order to inform noise budget requirements and fully leverage high-resolution direct images, it is critical to develop post-processing algorithms that are versatile, efficient, and reliable. This work leverages recent advancements in starshade simulation technology and presents a Convolutional Neural Network (CNN) for detecting exoplanet signals from direct images. The proposed architecture modifies U-Net, a semantic segmentation architecture, to directly regress the coordinate location of exoplanet signals from a synthetic star-shade imaging dataset. Results from a limited test set of 144 images indicate that the CNN achieves an average precision and recall of 0.925 ± 0.012 and 0.923 ± 0.018, respectively. Additionally, the CNN does not require processing the data through conventional background estimation algorithms prior to detection and performs inference on single images, showcasing the potential of supervised machine learning as a versatile approach for exoplanet detection.
|