Robust face recognition with class dependent factor analysis

Abstract

A general framework for face recognition under different variations such as illumination and facial expressions is proposed. The model utilizes the class information in a supervised manner to define separate manifolds for each class. Manifold embeddings are achieved by a nonlinear manifold learning technique. Inside each manifold, a mixture of Gaussians is designated to introduce a generative model. By this way, a novel connection between the manifold learning and probabilistic generative models is achieved. The proposed model learns system parameters in a probabilistic framework, allowing a Bayesian decision model. Experimental evaluations with face recognition under illumination changes and facial expressions were performed to realize the ability of the proposed model to handle different types of variations. Our recognition performances were comparable to state-of art results.