Master Thesis - Department of Computer Science

[105] in combining face and hand-geometry features have met with only limited
success.
Since the features contain richer information about the input biometric data than the
matching scores obtained at decision level, integration at feature level is believed to
be more effective than decision level fusion. But feature fusion is often not possible
for the unknown or incompatible relationship between different feature spaces, and
often simple concatenation of features may cause “curse of dimensionality problem”.
2.3.3.2 Fusion After Matching
Information fusion after matching can be divided into two main categories: dynamic
classifier selection and decision fusion.
A dynamic classifier selection technique chooses the result of that classifier which
is most likely to give correct decision for the specific input pattern [132]. This is
well-known as winner-take-all approach and the device that performs the selection is
known as associative switch [24].
Information fusion at decision level can be at abstract, rank and measurement
level based on the type of matcher’s output.
• Abstract level: The biometric matcher individually decides on the best match
based on the input presented. Methods like majority voting [71], behavior
knowledge space [70], weighted voting based on the Dempster-Shafer theory of
evidence [135], AND and OR rules [31], etc. are use to obtain the final decision.
• Rank level: The output of each biometric matcher is a subset of possible
matches sorted in decreasing order of confidence. Ho et al. [46] described three
methods to combine the ranks assigned by different matchers. In the highest
rank method, each possible match is assigned the highest (minimum) rank as
computed by different matchers. Final decision is made based on the combined
ranks and ties are broken randomly. The Borda count method uses the sum of
the ranks assigned by the individual matchers to calculate the combined ranks.
The logistic regression method is a generalization of the Borda count method
where weighted sum of the individual ranks is calculated and the weights are
38