Combination Framework
Four types of combination algorithms are distinguished based
on the number of input parameters they consider and the number of
trainable combination functions: low , medium I, medium II and high
complexity types. Low and medium I type combinations use only matching
scores related to one class to calculate the combined score for this
class, while medium II and high complexity combinations might use
matching scores for other classes as well. Low and medium II
combinations have only one trainable combination function; these
combinations are class-generic. On the other hand, medium I and high
complexity combinations use different combination functions for
different classes.
The diagrams below represent these four types. Entries in the
lattices are matching scores; columns correspond to scores assigned to
one class, and rows correspond to classifiers (thus we consider N
classes and M classifiers).
Main results:
- We argued that medium II combinations are most suitable type of
combinations for problems with large number of classes (such as
biometric identification problems)
- We proved that low and medium II complexity types are indeed
different, with medium II combinations having potentially better
performance
- Two combination algorithms of meduim II complexity type were developed using the concept of identification model showing superior performance in both verification and identification operational modes