Institut national de recherche scientifique français Univerité Pierre et Marie Curie Université Paris Diderot - Paris 7

NA-SODINN : a deep learning algorithm for exoplanet image detection based on residual noise regimes

mardi 23 mai 2023, par Carles Cantero, Université de Liège

Jeudi 1er juin 2023 à 16h00 , Lieu : Salle de rĂ©union du 1er Ă©tage du bâtiment 16 et visioconfĂ©rence

As of today, there exists a plethora of post-processing algorithms for exoplanet imaging. Their performance has been assessed using different data sets and metrics, which caused confusion in the HCI community when comparing their detection ability. In order to homogenize the comparison of these algorithms, the Exoplanet Imaging Data Challenge (EIDC) was born. With twenty eight algorithm submissions, the first EIDC phase (exclusively dedicated to exoplanet detection) provided two interesting conclusions :

  1. detection algorithms that exploit the local behavior of image noise obtained the highest detection score ;
  2. supervised machine learning algorithms, such as the SODINN deep neural network developed at ULiège, produced a relatively high number of false positives.

With the aim of improving the robustness of SODINN against false alarms, we built a more advanced version, referred to as Noise-Adaptive SODINN, which relies on two new strategies that help the training to capture stronger local image noise correlations.

First, unlike its predecessor, NA-SODINN trains an independent classification model per image noise regime in the processed frame.

Second, its network is fed with S/N curves, local discriminators that contain additional physical-motivated features and help the trained model to better disentangle an exoplanet signature from speckle noise.

NA-SODINN is evaluated against SODINN through a Receiver Operating Characteristics (ROC) analysis, in which we observe a clear improvement in both sensitivity and specificity. Then, it is submitted to EIDC, where we observe that it is ranked at the top (first or second position) of the challenge leaderboard for all considered evaluation metrics.

Lien pour assister au séminaire :