Although many systems exist for automatic classification of faces according to their emotional expression, these systems do not explicitly estimate the strength of given expressions. This paper describes and empirically evaluates an algorithm capable of estimating the degree to which a face expresses a given emotion. The system first aligns and normalizes an input face image, then applies a filter bank of Gabor wavelets and reduces the data's dimensionality via principal components analysis. Finally, an unsupervised Fuzzy-C-Mean clustering algorithm is employed recursively on the same set of data to find the best pair of principle components from the amount of alignment of the cluster centers on a straight line. The cluster memberships are then mapped to degrees of a facial expression (i.e. less Happy, moderately happy, and very happy). In a test on 54 previously unseen happy faces, we find an orderly mapping of faces to clusters as the subject's face moves from a neutral to very happy emotional display. Similar results are observed on 78 previously unseen surprised faces.
展开▼
