Fusion of eigenface and fisherface for face verification

Abstract

Principle Component Analysis (PCA) and Linear Discriminant Analysis (LDA) are two well-known methods that are used for face verification and recognition systems. Both of these methods perform dimensionality reduction. In similarity based face verification approach, a function known as similarity function assigns a matching score to a new given face image and its claimed identity, by measuring the distance between the face image and its stored template. In this work new data are first projected into both Eigenspace and Fisher space and matching-score is obtained for each space separately. The information from these two score is combined together with three different fusion rules, i.e., sum, max, and mean rule. Two different face databases, ORL and YALE, are used to evaluate the verification accuracy based on Equal Error Rate (EER). The EER for ORL database by using fusion approach is 7.12%-7.22% compared with the PCA of 7.6% and the LDA of 7.87%. The same experiment for Yale database shows by fusing the classifiers the error is 14.22% to 15.20% while by only using PCA the error is 15.27% and the error for LDA is 15.5%. Thus, fusing Eigenface and Fisherface system outperforms using them individually.