A Discriminant Subspace Learning Based Face Recognition Method

In the system of face recognition, the tradition method of data dimension reduction method is used to rearrange the face image vectors, resulting in the structural characteristics of the data itself lost and the recognition accuracy not high. In this paper, we develop a data dimension reduction method based on tensor-multilinear discriminant subspace projection. The algorithm directly describes the face with tensor and projects the tensor data into the vector discriminant subspace through a new projection mode tensor to vector projection (TVP). This method finds a set of orthogonal projection vector sets to maximize the dispersion between the data classes and minimize the intra-class dispersion in the discriminant subspace. Then, the high-dimensional tensor data is mapped to low-level vector data by TVP. These vector features after dimension reduction will be the most representative feature data in the whole face data under the appropriate constraint condition. Finally, these feature data are classified by the K-nearest neighbor classifier. Experiments on classical face database Olivetti Research Laboratory and face recognition technology verify the effectiveness of this method.