Recognizing Facial Expressions in the Orthogonal Complement of Principal Subspace
Swapna Agarwal, Sanjoy Mazumder, Dipti Prasad Mukherjee · 2014
We present a novel framework for recognition of facial expressions from a given face image. The framework is based on the assumption that expression information lies in the subspace orthogonal to the subspace representing expression-neutral faces. For deriving the principal subspace of the face images showing no expression, PCA is used as a tool. Then we derive a method to find the orthogonal complement (OC) of the subspace defined by the principal components. It is shown using different tools such as dendrogram and Davies-Bouldin cluster index that the OC of the principal subspace better represents the expressions as compared to the principal subspace in PCA analysis. We have done extensive experiments to validate the recognition capability of the proposed OC space. Two well known publicly available facial expression databases are used for the experiments. We also compare the expression discrimination capability of the OC subspace with some well known features for expression representation. The proposed framework exhibits higher (9.66% on average) recognition capability as compared to the present state-of-the-art works.