A Quadratic Lagrange Multipliers Based Approach to Feature Extraction in Face Recognition

Abstract

Fisher Linear Discriminant Analysis (LDA) has been widely used for feature extraction in face recognition. However, it cannot be used when each object has only one training sample because, within-class scatters cannot be statistically measured in this case. In addition, the respective axes of the projection matrix are not necessarily orthogonal in the strict sense. In this paper, a new method is proposed to solve those problems by quadratic Lagrange multipliers fisher linear discriminant analysis, which could not only eliminate the singular problem, but also obtain the optimal orthogonal projection matrix. The proposed approach is compared to the 2D-LDA method on the well-known ORL, CMU and YALE B+Extend YALE B face databases. It shows that the proposed method achieves better recognition accuracy and faster computational speed than 2D-LDA method does, especially in solving the matrix singular problem with only one training sample.