|Title||:||Methods of Score-level fusion in Multi-biometric systems|
|Speaker||:||Renu Sharma (IITM)|
|Details||:||Tue, 2 Aug, 2016 4:00 PM @ BSB 361|
|Abstract:||:||Biometrics authentication involves automatic recognition of a person based on the unique physiological or behavioral traits. Depending on the biometric identifier (or cue) used, many challenges appear in real-world applications, such as user acceptability and cooperation, low performance in unconstrained environments, low recognition rate due to occlusion, pose, illumination, aging, expression or makeup etc. Initially, we had focused our work on specific issues of face recognition using a single training sample (STS) and degraded set for training, which restrict the efficiency of face recognition algorithms in most practical applications. We have proposed a unified framework for handling both these challenges simultaneously, by using a suitable data dictionary and sparse representation of face images.
Unimodal biometric systems, designed to work using a single source of information for identification, face issues of lack of uniqueness, non-universality, non-permanence, intra-class variations, matcher limitation, noisy data, low recognition rate, spoof attacks and low fault tolerance. These shortcomings can be alleviated to a large extent by the use of multimodal biometric systems for recognition. In multimodal biometric systems, human identification is performed by combining cues from different biometric sources. Fusion of information can be performed at different levels like sensor-level, feature-level, score-level, rank-level and decision-level. At score-level, there is heterogeneity in the scores obtained from different unimodal biometric systems. To aid the issue of heterogeneity, we have proposed a score normalization method based on extreme value theory (EVT) using a Generalized Extreme Value (GEV) distribution. Thereafter, we have also proposed a superior score-level fusion technique based on Dezert-Smarandache Theory (DSmT). A new multimodal biometric framework for information fusion at rank-level methods has also been devised. Extensive set of experimental performed using chimeric and real-world datasets show the efficiency of our proposed methods.