Confidence interval of feature number selection for face recognition

Abstract

We propose a multistep statistical procedure to determine the confidence interval of the number of features that should be retained in appearance-based face recognition, which is based on the eigen decomposition of covariance matrices. In practice, due to sampling variation, the empirical eigenpairs differ from their underlying population counterparts. The empirical distribution is difficult to derive, and it deviates from the asymptotic approximation when the sample size is limited, which hinders effective feature selection. Hence, we propose a new technique, MIZM (modified indifference zone method), to estimate the confidence interval of the number of features. MIZM overcomes the singularity problem in face recognition and extends the indifference zone selection from PCA to LDA. The simulation results on the ORL, UMIST, and FERET databases show that the overall recognition performance based on MIZM is improved from that using all available features or heuristically selected features. The relatively small number of features also indicates the efficiency of the proposed feature selection method. MIZM is motivated by feature selection for face recognition, but it extends the indifference zone method from PCA to LDA and can be applied in general LDA tasks.