Multivariate polynomial identification for blind image separation

This paper presents a new approach for optimizing the non-normalized kurtosis under unit power constraint. It is inspired from the famous FastICA algorithm which uses a fixed-point algorithm and computes fourth-order statistics at each step of the optimization, which is very demanding in terms of both computation time and memory space, especially when the number of samples is high. Our method avoids this by using a polynomial identification at the beginning of the algorithm, which lets us perform the optimization in a more efficient computation space. Our so-called O-FICA algorithm is particularly interesting in blind image separation because of the present size increase of light sensors.