Face analysis based on reference samples

Humans are not sensitive to variations in facial appearance and are capable of performing face analysis tasks reliably under realistic conditions when compared with current computer-based face analysis techniques. This can be partly explained by the ability of humans to make effective use of previously encountered known faces for both internal representation and processing. This thesis focuses on establishing computational models to account for the cognitive findings related to internal face representation and two fundamental perception processes (distinctiveness and familiarity), and developing novel methods based on the models for face analysis. Specifically, a set of reference samples that may or may not contain any labeling information and any instance of the person whose face is under consideration are proposed to model previously encountered faces. The non-negative matrix factorization which affords part-based representation is extended to learn reusable local facial patterns for representation from the reference set. Computational models are developed for locating distinctive areas and measuring familiarity of faces with respect to the reference set. By employing the proposed face representation, distinctiveness and familiarity models, novel schemes are developed to recognize faces from single sample per person and estimate ages and head poses of faces.