Customized Orthogonal Locality Preserving Projections With Soft-Margin Maximization for Face Recognition

Face recognition still is a challenging task since face images may be affected by changes in the scene, such as in head pose, face expression, or illumination. In addition, face pattern representation often requires several dimensions, which poses additional challenges for face recognition. We propose a novel face recognition method based on projections of high-dimensional face image representations into lower dimensionality and highly discriminative spaces. This is achieved by a modified orthogonal locality preserving projection (OLPP) method that uses a customized locality definition scheme to preserve the face class structure in the lower dimensionality face feature space. The proposed method can work with sparse and dense face image representations (i.e., it can use subsets or all face image pixels) and tends to be robust to data outliers and noise. Besides, we introduce a sparse representation using interpolated landmarks, designed to preserve important details in high-resolution color face images (e.g., eyes), and compensate for uncertainties in landmark positioning during face image feature extraction. The face images are classified in this lower dimensionality feature space using a trained soft-margin support vector machine, so it performs better than the nearest neighbor rule used in the typical OLPP method. A set of experiments was designed to evaluate the proposed scheme under various conditions found in practice (such as changes in head pose, face expression, illumination, and in the presence of occlusion artifacts). The experimental results were obtained using five challenging public face databases (namely, Poznan University of Technology, Fundação Educacional Inaciana, Facial Recognition Technology, Yale, and Our Database of Faces). These experiments suggest that our sparse representation for high-resolution face color images, integrated to the proposed lower dimensionality feature space and classification scheme, tends to obtain higher accuracy values than those obtained using typical sparse and dense representations for the same face images in grayscale. To evaluate the generality of our lower dimensionality feature space and classification scheme, additional tests using full low-resolution grayscale face images were performed, as often used in face recognition (e.g., typical OLPP method). Our experiments suggest that the proposed approach can also provide higher accuracy values than comparable state-of-the-art methods available in the literature when using full low-resolution grayscale face images (i.e., dense representations).