Facial Expression Recognition by Supervised ICA with Selective Prior

摘要

Feature selection is required when using the Independent Component Analysis (ICA) in feature extraction for pattern classification. Selection during ICA might provide a better candidate set of features. We propose a supervised ICA with a selective prior for the de-mixing coefficients so that those features with higher significance in discrimination could emerge easier during the learning. We formulate the learning rule for the supervised ICA in a form of the natural gradient approach and develop the algorithm of supervised ICA in facial expression analysis. The efficiency of the proposed algorithm has been investigated by numerical experiments.