Semester

Fall

Date of Graduation

2005

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Natalia A Schmid

Abstract

Biometric systems utilize physiological or behavioral traits to automatically identify individuals. A unimodal biometric system utilizes only one source of biometric information and suffers from a variety of problems such as noisy data, intra-class variations, restricted degrees of freedom, non-universality, spoof attacks and unacceptable error rates. Multimodal biometrics refers to a system which utilizes multiple biometric information sources and can overcome some of the limitation of unimodal system. Biometric information can be combined at 4 different levels: (i) Raw data level; (ii) Feature level; (iii) Match-score level; and (iv) Decision level. Match score fusion and decision fusion have received significant attention due to convenient information representation and raw data fusion is extremely challenging due to large diversity of representation. Feature level fusion provides a good trade-off between fusion complexity and loss of information due to subsequent processing. This work presents generic feature information fusion techniques for fusion of most of the commonly used feature representation schemes. A novel concept of Local Distance Kernels is introduced to transform the available information into an arbitrary common distance space where they can be easily fused together. Also, a new dynamic learnable noise removal scheme based on thresholding is used to remove shot noise in the distance vectors. Finally we propose the use of AdaBoost and Support Vector Machines for learning the fusion rules to obtain highly reliable final matching scores from the transformed local distance vectors. The integration of the proposed methods leads to large performance improvement over match-score or decision level fusion.

Share

COinS