Master Thesis - Department of Computer Science

combine the discriminative informations obtained from null space and range space
for constructing a dual space. Then forward or backward selection technique is used
to select optimal feature set from dual space. On the other hand, decision fusion
based method constructs two different classifiers on the null space and range space
separately and then combines them using decision fusion strategies. Results of all the
three methods have been shown over three standard databases.
There are severe drawbacks and risks of using a unifeature biometric recognition
system, specifically recognition based on fingerprint or face alone. A machine which
has the ability to only analyze data, but has no power to determine if the data has
been fed into by an authorized person or not, is vulnerable to such acts of impostors
who are always on the lookout to create innovative means of breaking into a system.
Multimodal biometry which uses multiple sources of information for decision-making
is a first choice solution for the above defined problem.
In this thesis, a new approach for combining evidences from face and fingerprint
classifiers at the decision level has been proposed. In this approach, each of the face
and fingerprint classifier is separately exploited on the basis of availability of class-
specific information to enhance combination performance. Results using face and
fingerprint databases, show that the proposed methodology of using class-specific
information at classifier’s response outperforms the state-of-art fusion techniques.
Keywords: Face, dual space, subband face, decision fusion, feature fusion,
eigenspace, multimodal, biometry.
xix