A low cost and reliable Polak-Ribiere conjugate gradient deflation ICA algorithm for real signals

In this paper a new more robust version of a classical deflation independent component analysis algorithm (namely DelLR) which was introduced by Delfosse and Loubaton[1] for real context, is implemented. The new version uses one of the nonlinear conjugate gradient optimization methods, Polak-Ribiere. We compared the new algorithm with the classical one and the FastICA technique in order to propose a performance comparison of different deflation based ICA algorithms in terms of estimation accuracy and numerical complexity on different signals, audio and images. We show that the new algorithm is faster and has higher quality of separation and a lower computational cost than others.