Variational blind source separation toolbox and its application to hyperspectral image data

The task of blind source separation (BSS) is to decompose sources that are observed only via their linear combination with unknown weights. The separation is possible when additional assumptions on the initial sources are given. Different assumptions yield different separation algorithms. Since we are primarily concerned with noisy observations, we follow the Variational Bayes approach and define noise properties and assumptions on the sources by prior probability distributions. Due to properties of the Variational Bayes algorithm, the resulting inference algorithm is very similar for many different source assumptions. This allows us to build a modular toolbox, where it is easy to code different assumptions as different modules. By using different modules, we obtain different BSS algorithms. The potential of this open-source toolbox is demonstrated on separation of hyperspectral image data. The MATLAB implementation of the toolbox is available for download.