Face recognition using KFD-Isomap

Facial images with high dimension often belong to a manifold of intrinsically low dimension. Subspace methods utilize different algorithms to extract and analyze the underlying manifold for face recognition. Isomap is a recently proposed algorithm for manifold learning and nonlinear dimensionality reduction. However, since the Isomap is developed based on minimizing the reconstruction error with multi-dimensional scaling, it may not be optimal from classification viewpoint. In this paper, an improved version of Isomap, namely KFD-Isomap, is proposed using kernel Fisher discriminant (KFD) method for face recognition task. In KFD-Isomap, the matrix of geodesic distances between all pairs of points as feature vectors is applied to the kernel Fisher discriminant for finding an optimal projection direction. In face recognition experiments, KFD-Isomap is used as a feature extraction process compared with Isomap, Ext-Isomap, and two other baseline subspace algorithms, eigenfaces and Fisherfaces, combined with a nearest neighbor classifier. Experimental results show that KFD-Isomap excels the other methods.