Improved Bayesian Approach for Face Recognition

In subspace face recognition, PCA, LDA and Bayesian are the most commonly used methods. Each of them has their own advantages and disadvantages in recognizing human faces. Their recognition rates depend much on the methodologies used in selecting/transforming the eigenvectors using the eigenvalues obtained from the face subspaces. In this paper we compare all these three methods and propose a new methodology for selecting the eigenvalues in the face subspace which can be used for measuring the residual reconstruction error in the partial Karhunen-Loeve Transformation (KLT) for Bayesian face recognition. We compare the recognition performances of all these methods on FERET image database on various image sizes. Experimental results using a large set of faces - 2388 images drawn from 1194 subjects separated into training, gallery and probe datasets show that our proposed method consistently improves the performance over the Bayesian, LDA and PCA approaches.