Subspace Linear Discriminant Analysis for Face Recognition
W Zhao, R Chellappa, P. Jonathon Phillips
In this correspondence, we describe a holistic face recognition method based on subspace Linear Discriminant Analysis (LDA). Like existing methods, this method consists of two steps: first, the face image is projected into a face subspace via Principal Component Analysis (PCA) where the subspace dimension is carefully chosen, and then the PCA projection vectors are projected into the LDA to construct a linear classifier in the subspace. Experiments show that the dimension of the face subspace is fixed regardless of the image size as long as the image size exceeds the subspace dimension. The property of relative invariance of the subspace dimension enables the system to work with smaller face images without sacrificing performance. The choice of such a fixed subspace dimension is mainly based on the characteristics of the eigenvectors instead of the commonly used eigenvalues. Such a choice of the subspace dimension enables the system to generate class-separable features via LDA from the full subspace representation. Hence the generalization/overfitting problem can be addressed. In addition, a weighted distance metric guided by the LDA eigenvalues is employed to improve the performance of the subspace LDA method. Experimental results using FERET datasets and the MPEG-7 content set are presented.
IEEE Transactions on Image Processing
face recognition, FERET, linear discriminant analysis