Selecting discriminant eigenfaces for face recognition

In realistic face recognition applications, such as surveillance photo identification, supervised learning algorithms usually fail when only one training sample per subject is available. The lack of training samples and the considerable image variations due to aging, illumination and pose variations, make recognition a challenging task. This letter proposes a development of the traditional eigenface solution by applying a feature selection process on the extracted eigenfaces. The proposal calls for the establishment of a feature subspace in which the intrasubject variation is minimized while the intersubject variation is maximized. Extensive experimentation following the FERET evaluation protocol suggests that in the scenario considered here, the proposed scheme improves significantly the recognition performance of the eigenface solution and outperforms other state-of-the-art methods.