Composite orthonormal basis for person-independent facial expression recognition

We propose a novel framework, involving a composite orthornormal basis (COB), for person-independent facial expression recognition from images. In combination with localized features, the new approach decouples expression from identity information as much as possible. Experimental results show that such an approach, when applied to wide-ranging databases, significantly improves expression recognition accuracy in comparison with the results in the literature. It can also be successfully generalized to cross-database expression recognition which is believed to be the first of its kind.