Adaptive objective function of ICA by Gaussian approximation in second-order polynomial feature space
Yoshitatsu Matsuda, Kazunori Yamaguchi · 2016
In this paper, we propose an objective function of ICA with adaptive estimation of the kurtoses of sources. It is derived by applying the Gaussian approximation to the distribution of sources in the second-order polynomial feature space. This objective function (called the adaptive ICA function (AIF)) is a simple form given as a summation of weighted 4th-order statistics, where the weights are determined by adaptively estimated kurtoses. It is proved by the convergence analysis that the ICA solution is a stable maximum of the objective function. We also propose a natural gradient algorithm optimizing the function. Experimental results show that the proposed method is effective in blind image separation problems.