Abstract

Exoplanets are planets located outside the solar system. One method to detect these distant worlds is through the direct detection of their thermal emission. The so-called direct imaging technique is suitable for observing young planets far from their star.

Due to the star emissions, these are very low signal-to-noise-ratio measurements. Moreover, the limited and highly unbalanced ground truth hinders the use of supervised learning approaches to automatically detect planets signals in the images.

In this talk, we show how to bypass the scarcity of real data by training a Generative Adversarial Network. The synthetic images produced by the generative model can be assumed to not contain any planet and are augmented by artificially injecting planets signals. The data obtained are not just labeled but, for the positive samples, the exact position of the object to detect is known. CNN detectors trained on this synthetic dataset exhibit good predictive performance and, on real data, the models can re-confirm bright sources detection. In this sense, the above technique shows the potential to go beyond the current state of the art in exoplanet discovery via direct imaging.