Reliable face recognition using feature selection and image rejection based on probabilistic face model

Though many studies on robust face recognition have shown good performances at their experiments, they still suffer from diverse environmental variations in real world. In addition, most face recognition methods focus only on classifying subjects using well-aligned facial images, and thus their reliabilities are dependent upon the performance of pre-processors such as face detector. The purpose of this study is to develop a reliable face recognition system that can deal with two common errors caused by automatic face detectors: incorrect localization and false detection of non-facial images. Based on the previous framework using probabilistic face model of local features, we add a feature selection module that can deal with localization error as well as a rejection module that can effectively treat detection error. Through computational experiments using benchmark data and real world images including translation variations and/or detection errors, we confirm significant improvement in the classification performance and reliability.