Pair of projections based on sparse consistence with applications to efficient face recognition

Dimension reduction based feature extraction and classification method show significant performance on the high-dimensional face images. The traditional dimension reduction methods learn a projection based on the Fisher criterion or local structure of the face images. This work aims at learning a pair of projection based on sparse consistence which is measured by sparse constraint and label information for efficient face recognition. The first projection maps the original high-dimensional face images into a low-dimensional space where each face is sparse, and the second one which can also be treated as a classifier guides the sparse low-dimensional face images to the right label. The pair of projections is optimized together using alternative update rules efficiently. Due to the discriminant power of sparse face images and the supervised classifier, the proposed algorithm integrates the supervised and unsupervised information and is more efficient than them for face recognition on both learning and classifying. Experimental results on the challenging Extended Yale B, AR, and LFW face image databases demonstrate the proposed algorithm on both accuracy and efficiency.