Robust Part-Based Face Recognition Using Boosting and Transduction

The challenge for biometrics is to withstand image variability and defend against impostors seeking to breach security. The impostors attempt to hide and/or alter the information needed for their identification. While faces can be partially occluded and/or disguised some of their parts remain unchanged and can still be properly detected and authenticated. Towards that end this paper advocates robust part-based face recognition using boosting and transduction. The face representation used spans a multi-resolution (golden ratio) grid that captures partial information at different scales in order to accommodate different surveillance scenarios including human identification from distance. The face components are defined across the eyes, nose, mouth, eye and nose, nose and mouth, and the like, and encode both facial parts and their second order relationships. The parts, clusters of local patches described using similar SIFT features, are modeled using an exemplar based representation. The model free and non-parametric weak learners found by transduction, which correspond to parts and their relationships, compete to build up a strong boosting classifier. The feasibility of the novel approach, using FRGC (UND) database, shows robustness to uncontrolled lighting condition, different facial expressions, and occlusion.