Face Recognition Based On Bayesian Classifier

Face Recognition not only has significant theoretic values but also is believed having a greatdeal of potential applications in public security, law enforcement, information security, andfinancial security. As one of the most important face recognition methods, Moghaddam'sBayesian classifier has been well understood and extended in many ways, such as from singleGaussian Model to Gaussian Mixture Model (GMM), from image intensity feature to otherfeatures (e.g. Gabor). Based on these extensions, we propose our improved scheme both inclassifier designing and probability estimating, the main work of the thesis includes:(1) In classifier designing, we proposes an approach combining multiple Feature Block-basedBayesian Classifiers (FBBCs). Each FBBC is designed as a Maximal Likelihood (ML)classifier by using GMM of Gabor features computed from some image block uniformlypartioning the whole image. These FBBCs are finally combined together by computingaverage similarity to make the final decision. We also investigate the use of weighed averagesimilarity to further improve performance. Our experiments on FERET face database haveshown the effectiveness of the proposed method.(2) Another extension is that we use nonparametric method to estimate the posterior probabilityinstead of Gaussian model. We apply this method to the Bayesian Classifier and FBBC weproposed. Our experiments on FERET Database have shown that the nonparametric methodoutperforms the Bayesian Classifier using Gaussian model, and also the nonparametricmethod is easy to train.The main contribution of this thesis extends the application of Bayesian method in FaceRecognition, and achieves better performance than traditional Bayesian Classifier.