Collaborative representation-based discriminant neighborhood projections for face recognition

Manifold learning as an efficient dimensionality reduction method has been extensively used. However, manifold learning suffers from the problem of manual selection of parameters, which seriously affects the algorithm performance. Recently, applications of collaborative representation have received concern in some fields such as image processing and pattern recognition research. Based on manifold learning and collaborative representation, this paper develops a new algorithm for feature extraction, which is called collaborative representation-based discriminant neighborhood projections (CRDNP). In CRDNP, we first construct intra-class and inter-class neighborhood graphs of the input data as well as a weight matrix based on collaborative representation model and class label information. Then, a projection to a reduced subspace is obtained by margin maximization between the between-class neighborhood scatter and within-class neighborhood scatter. CRDNP not only characters the inherent geometry relationship of the dataset using L2-graph, but also enhances the between-class submanifold separability. In addition, the discriminating capability of CRDNP is further improved by obtaining the orthogonal projection vectors. Experiment results on public face datasets prove that CRDNP can achieve more accurate results compared with the existing related algorithms.