BLIND SOURCE SEPARATION FROM MULTI-CHANNEL OBSERVATIONS WITH CHANNEL-VARIANT SPATIAL RESOLUTIONS

We propose a Bayesian method for separation and reconstruction of multiple source images from multi-channel observations with different resolutions and sizes. We reconstruct the sources by exploiting each observation channel at its exact resolution and size. The source maps are estimated by sampling the posteriors through a Monte Carlo scheme driven by an adaptive Langevin sampler. We use the t-distribution as prior image model. All the parameters of the posterior distribution are estimated iteratively along the algorithm. We experimented the proposed technique with the simulated astrophysical observations. These data are normally characterized by their channel-variant spatial resolution. Unlike most of the spatial-domain separation methods proposed so far, our strategy allows us to exploit each channel map at its exact resolution and size.