Examination of Firearms Review: 2007 to 2010 - Interpol

spurious minutiae around the crease, within a distance of 5 pixels. The combined use
of minutiae and creases diminishes error rates in matching with respect to matching
based exclusively on minutiae, but only for the database of elderly people. On a
general population database this improvement is only very small.
Level II detail
The best way of analysing data in the form of match scores from a fingerprint
matcher has been the object of two articles (143, 144). ROC (receiver operating
characteristic) curves are compared to likelihood ratios (LR). Here, a LR above one
yields the classification result that two impressions belong to the same person and a
LR below one that they belong to different persons. For obtaining LRs, the within-and
between-finger variabilities are modelled using gamma- and normal distributions. The
likelihood ratio method outperforms ROC curves, but the performance of the two
methods is similar when large numbers of minutiae are available (143). However, an
analysis of four different matching systems shows that there is no definitive
underlying distribution function for match and non-match scores. Nonparametric
analyses on the discrete distribution functions of the four sets of scores (issued from
the four different systems) are presented, and take the form of ROC curves (144).
Note that while Srihari and Srinivasan (143) privilege LRs over ROC curves for
performance reasons, Wu and Wilson (144) use ROC curves in the nonparametric
approach. This second approach is, according to these authors, the proper way of
analysing such data, due to differences in the distributions obtained when different
matchers are used on one hand and the fact that the data is discrete (or can be
transformed to discrete data) on the other hand.
A study on the variability of minutiae in a Spanish population is presented by
Gutierrez and al. (127). The authors studied in particular the total count of the
minutiae, the count of each different type (where combined minutia types were used),
both in the whole fingerprint and in the centre (defined as a circle with a radius of 18
ridges) and periphery separately. The frequencies of the different minutia types are
given. Also, the number of minutiae on the fingerprint is compared between male and
female subjects. Differences are observed for the whole fingerprint, when the general
pattern is whorl or loop, and differences are significant for all general patterns in the
peripheral region, males having more minutiae than females. In the central area,
none of the differences observed between male and female subjects are significant.
Zhu and co-workers (123) use minutiae location and orientation in a generative
model. This work and related ones has been reviewed also by Dass et al. (23). Here,
a joint distribution is proposed, where minutiae locations are modelled by a bivariate
normal mixture, and orientations by a Von-Mises distribution. A different approach,
also using the model of Zhu et al. (123) as a basis, integrates image quality by
adding a model for errors in minutiae detection and localisation (124). An internal
report gives an overview of models for fingerprint individuality (145), proposing a
classification of models into grid, fixed probability, ridge-based, relative
measurement, and generative models. Generative models for birthdays, height and
fingerprints are described (117). The fingerprint generative model includes minutiae
locations and directions, as well as ridge lengths and the location and direction of
ridge points, where one ridge point is selected in medium ridges and two ridge points
for long ridges (none for short ridges). The parameters of the model are then
estimated on the basis of data (100 fingerprints and 8 impressions from each
234