A fixed-point ICA algorithm with initialization constraint

Abstract

We propose a novel approach, the fixed-point algorithm (FastICA) with initialization constraint, for performing independent component analysis (ICA). In order to achieve a reliable convergence during estimating the blind source components, both the third- and fourth-order statistics are taken into account when diagonalizing the cumulant tensors. By combining these high-order statistics, an initialization constraint is introduced into the decomposition procedure of the independent components by FastICA. The experimental results demonstrate that the improved algorithm can achieve a better performance than the original FastICA without increasing the computation cost dramatically. The simulations involving source signals with different distributions show that our algorithm can adapt most source signals, including some complicated and asymmetric distributions.