Feature extraction approach to blind source separation

Local independent component analysis is formulated as a task involving the extraction of local geometric structure in the joint distribution. Because the geometrical structure of statistical independence is not well captured by statistical descriptions such as moments and cumulants, we use feature detection tools from image analysis to locate the local independent component coordinate system. The resulting approach to source separation can be implemented in real time using conventional image analysis hardware. The generality of this approach is demonstrated by blind source separation of multi-modal sources, and the pseudo-separation of three sources from two mixtures.