Evaluation of fusion score for face verification system

Abstract

Given an individual face image and a claimed ID, the face verification problem is to determine whether or not he is the person he claims to be. Although this task seems to be easy for a human, this problem is one of the most challenging problems in the area of computer vision. Eigenface and fisherface are two well-known and successful face verification approaches. Despite an assumption that face verification systems based on fisherface is thought to be more accurate than eigenface system, recent studies reveal that the idea is not always true. In this research, in order to leverage on the strength of both eigenface and fisherface techniques, a fusion of these two techniques by using different fusion method is examined. Four fusion methods, namely, sum-rule, Artificial Neural Network (ANN), Linear Support Vector Machines (Linear SVM), and Gaussian Support Vector Machines (Gaussian SVM) are considered. ORL database is used to evaluate and compare different approaches. The experiments show that the Total Error Rate for individual eigenface and fisherface systems are 12.5% and 9.4% respectively. This error for the fusion based systems that use sum-rule, ANN, Linear SVM, and Gaussian SVM, as fusion techniques are 9.9%, 5.9%, 6.7%, and 6.3% respectively. The results demonstrate that fusion-based face verification system outperforms both eigenface and fisherface systems when used individually