Logistic regression to boost exoplanet detection performances
Direct imaging of exoplanets requires to separate the background noise from the exoplanet signals. Statistical methods have been recently proposed to avoid subtracting any signal of interest as opposed to initial self-subtracting methods based on Angular Differential Imaging (ADI). However, unless conservative thresholds are chosen to claim for a detection, such approaches tend to produce a list of candidates that include many false positives. Choosing high, conservative, thresholds leads to miss the faintest planets. We extend a statistical framework with a logistic regression to filter the list of candidates. Features with physical/optical meaning (in two wavelengths) are used, leading to a very fast and pragmatic approach. The overall method requires a simple edge detection (image processing) and clustering algorithm to work with sub-images. To estimate its efficiency, we apply our approach to targets observed with the ESO/SPHERE high contrast imager, that were previously used as tests for blind surveys. Experimental results with injected signals show that either the number of false detections is considerably reduced or faint exoplanets that would otherwise not be detected can be sometimes found. Typically, on the blind tests performed, we are now able to detect around 50% more of the injected planets with an SNR below 5, and with a very low number of additional candidates.