Application of Radial Basis Function Network and Locality Preserving Projections for Face Recognition

Locality preserving projections (LPP) is a linear method that optimally preserves the local structure of the data set. However, because of the continuing existence of transformation difference, LPP subspace is failed to detect the important nonlinear variations of the face manifold. In order to improve the recognition performance, we propose to use radial basis function network (RBFN) to classify the features in the LPP subspace. The multi-quadrics function is taken as the activation function of the RBFN and the hidden layer of RBFN is trained via the gradient descent algorithm in our approach. The experimental results show that the LPP+RBFN method has achieved a better rate than Laplacianfaces, Fisherfaces and Eigenfaces.