Image Space Reconstruction Based Overcomplete ICA Algorithm

In this paper, based on image space reconstruction algorithm and regression ICA, a new algorithm for overcomplete ICA is developed. In this algorithm, it first reconstructs part of the observations and then predicates the missing observations from their probability distributions and the available observations. After these steps, the standard ICA algorithm is employed to estimate all the independent components. In the paper, the convergence of the proposed algorithm is also analyzed. Experiment demonstrates that the algorithm can obtain good fidelity for blind signal separation. In the simulation, three speech signals are successfully separated while only two mixtures of the three signals are given.