Abstract
Biometric recognition protocols involving single sources of information for human authentication which are commonly termed as unimodal systems though show satisfying performance, still suffered from problems relating to non-universality, permanence, collectability, convenience and susceptibility to circumvention. This paper emphasizes the priority of biometric information fusion by analyzing two kinds of fusion: Fusion of multiple representations of single biometric trait and Fusion of multiple biometric traits. As biometric traits possess large variance between persons and small variance between samples of the same person, it is important to capture this information using multiple representations at both global-level and local-level and perform fusion at feature-level. As a feature set is a straightforward representation of raw biometric data, it is theoretically presumed to incorporate richer information. Hence, we propose to use a fusion method that maximally correlates information captured from both the features and eliminates the redundant information giving a more compact representation. Fusion of multiple biometric traits is realized using fingerprint, palmprint and iris modalities. We explore this kind of fusion using two architectures: Parallel architecture and Hierarchical-cascade architecture. Multi-biometric recognition systems designed with hierarchical architecture not only are robust, fast and highly secure but also mitigate problems like missing and noisy data associated with parallel and serial architectures respectively, not to be forgotten that parallel architectures are preferred in high security-demanding defense/military applications as they evidently provide more precision for the reason that they combine more modalities and evidences about the user for recognition. Parallel framework proposed in this work takes advantage of score-level fusion. Score-level fusion is widely put to use as it offers best trade-off between ease and efficiency. We propose two score-level fusion techniques which rely on Equal Error Rates of individual modalities. Since error rate is a percentage of misclassified samples, we attempt to minimize the overlapped area between genuine and imposter curves by choosing to maximize the stability of the modality with superior performance. The proposed rule addresses the fusion problem from error rate minimization point of view so as to increase the decisive efficiency of the fusion system. To take the advantage of feature-level fusion, serial/cascade architecture and hierarchical architectures, we also propose a two-stage cascading frame-work based on fusion of fingerprint and palmprint feature sets in the first stage and iris features to eliminate the ambiguity of false matches in the next stage. The proposed frame work takes advantage of both unimodal and multimodal architectures. Proportionate experimental results reported on both real and virtual databases in this work demonstrate the superior performance of a multimodal recognition system over a unimodal system but however infers that the design of a multimodal biometric system predominantly depends on the application criteria and so is difficult to anticipate the best fusion strategy. The review of biometric based recognition systems indicate that a number of factors including the accuracy, cost, and speed of the system may play vital role in assessing its performance.
But today with the cost of biometric sensors constantly diminishing and high speed processors and parallel programming techniques widely available to affordable research, accuracy performance has become predominant focus of biometric system design. The main aim of the present work is to improve the accuracy of a multimodal biometric recognition system by reducing the error rates.