Tensor-based Orthogonal Locality Sensitive Discriminant Analysis and its Application on Face Recognition

In this paper, a novel appearance-based feature extraction method called Tensor-based Orthogonal Locality Sensitive Discriminant Analysis (Tensor-OLSDA) is presented for feature extraction problem in face recognition. Tensor-OLSDA preserves the intrinsic local manifold structure and the geometrical information as well as strengthens the discriminant power. And it also overcomes the Metric distortion due to the non-orthogonality, which distorts the local geometrical structure of the data sub-manifold, and reduces the difficulty for dimension estimation, therefore, improves the separability of face data and gives a better recognition result. With high-order tensor representation of the face data, the extraction is made along each order of the unfold data and the feature subspace is obtained by OLSDA with orthogonal constraints. At last, the original face data is projected onto this feature subspace for recognition. Experiments based on the AT&T and YaleB face database show the impressive classification capability of the proposed method. Experimental results show Tensor-OLSDA achieves the top average recognition rate in the several compared methods which also confirms that the locality preserving ability is enforced by computing the mutually orthogonal basis functions iteratively with tensor data representation.