A Contrast for Ica Based on the Knowledge of Source Kurtosis Signs

We propose a new contrast criterion for independent component analysis (ICA) based on the prior knowledge of the source kurtosis signs. After prewhitening, the contrast can be optimized by a pairwise processing approach in which plane rotations are found at low computational cost at each iteration. It is proved that the indeterminacy associated with this contrast is a scaled permutation matrix composed of two blocks, each corresponding to a source kurtosis sign. Hence, if the source of interest has a kurtosis sign different from that of the others, it can be extracted without separating the whole mixture. Experimental results show that the source estimation performance improves with the kurtosis gap.