Skip to Main Content
The appearance of a face is severely altered by illumination conditions that makes automatic face recognition a challenging task. In , we introduced an illumination- invariant face identification method based on Gaussian Mixture Models (GMM) and the phase spectra of the Fourier Transform of images. In this paper we explore the application of this identification scheme on the Yale database that contains images with a greater degree of illumination variations. The novelty of our approach is that the model is able to capture the illumination variations so aptly that it yields satisfactory results without an illumination normalization unlike most existing methods. Identification based on a MAP estimate achieves misclassitication error rate of 3.5% and a low verification rate of 0.4% on this database with 10 people and 64 different illumination conditions. Both these sets of results are significantly better than those obtained from traditional PCA and LDA classifiers. We next show that upon illumination normalization, our method succeeds in attaining near-perfect results using the reconstructed images. A rigorous comparison with existing state-of-the-art approaches demonstrates that our proposed technique outperforms all of those. Furthermore, some statistical analyses pertaining to Bayesian model selection and large-scale performance evaluation based on random effects model are included.