Face recognition using data driven local appearance features

A novel data driven face descriptor based on point-to-subspace metric is proposed for subspace classifiers. Unlike conventional feature descriptors which are carefully designed by hand, the newly proposed method uses supervised learning to derive more robust and more discriminative descriptors. During the feature extraction process, the point-to-subspace distance is used as the inner mechanism to train parameters including filters and weights of different pixels. Experimental results on FERET and Extended Yale B database show that when using subspace classifiers, the proposed feature descriptor is more discriminative and yields higher recognition rate over other features.