The work presented here attempts at answering the question: how do we decide when a given adetection is a planet or just residual noise in exoplanet direct imaging data? To this end we present a method implemented within a Bayesian framework: (1) to unify 'source detection', and, 'source characterization' into one single rigorous mathematical framework; (2) to enable an adequate hypothesis testing given the S/N of the data; (3) to enhance the detectability of planets faint signal in the presence of instrumental and background noise and to optimize the characterization of the planet. As a proof of concept we implemented a routine named ${\tt PlanetEvidence}$ that integrates the nested sampling technique (Multinest) with a post-processing technique, the Karhunen-Loeve Image Processing (KLIP), algorithm. This is a first step to recast such post-processing method into a fully Bayesian perspective. We test our approach on real direct imaging data, specifically using GPI data of $\beta$ Pictoris b, and, on synthetic data. We find that for the former the method strongly favors the presence of a planet (as expected) and recovers the true parameter posterior distributions. While for the latter case our approach allows us to detect (true) dim sources invisible to the naked eye as real planets, rather than background noise, and set a new lower threshold for detection at the 2$\sigma$ level approximately. Further it allows us to quantify our confidence that a given detection is a real planet and not just residual noise (for example residual speckles). The next natural step is to extend this approach to construct a Bayesian-based algorithm for blind detection, that is, not requiring an initial guess as to the location of the planet. This is the subject of ongoing work.