Extracting Astrophysical Sources from Channel-Dependent Convolutional Mixtures by Correlated Component Analysis in the Frequency Domain

A second-order statistical technique (FD-CCA) for semi-blind source separation from multiple-sensor data is presented. It works in the Fourier domain and allows us to both learn the unknown mixing operator and estimate the source cross-spectra before applying the proper source separation step. If applied to small sky patches, our algorithm can be used to extract diffuse astrophysical sources from the mixed maps obtained by radioastronomical surveys, even though their resolution depends on the measurement channel. Unlike the independent component analysis approach, FD-CCA does not need mutual independence between sources, but exploits their spatial autocorrelations. We describe our algorithm, derived from a previous pixel-domain strategy, and present some results from simulated data.