MULTIMODAL PERSONAL AUTHENTICATION USING GAUSSIAN MIXTURE MODEL AND SUPPORT VECTOR MACHINE

Abstract

ABSTRACT In this work, different approaches like Gaussian Mixture Models (GMM) and Support Vector Machine (SVM) have been investigated for multimodal biometric authentication. There are various schemes for multimodal fusion like normalization techniques, classifier based approaches and density evaluation approaches. Finite mixture models like GMM have been used for multimodal biometric systems and which produces significantly good results on multimodal databases. Multimodal databases can be constituted of multiple instances of same biometric trait or can be obtained from various matching algorithms which determine the scores of genuine and imposter classes. The experiments have been performed on palmprint, knuckleprint and iris databases. Different feature extraction techniques like Gabor-Scale Invariant Feature Transform (SIFT), Gabor-Harris and Gabor-HOG has been used for feature extraction in palmprint and knuckleprint databases. These techniques work significantly well for both constrained(IITD) and unconstrained(PolyU) databases which are used in the work and produce efficient results for performance measures as compared to the prior techniques like line based approaches, texture based and appearance based approaches. GMM has been used for classification from the scores generated by these feature extraction techniques. The parameters evaluation for the GMM technique has been evaluated using Expectation-Maximization (EM) algorithm. Maximum Likelihood Estimation (MLE) further improvises the results of GMM. SVM which is active learning methods also results in good results for palm and knuckle database and gives optimal performance. The results show that these techniques are quite robust for multimodal authentication schemes.