12.11.2012 Views

Examination of Firearms Review: 2007 to 2010 - Interpol

Examination of Firearms Review: 2007 to 2010 - Interpol

Examination of Firearms Review: 2007 to 2010 - Interpol

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

fingerprint, FVC2002 DB1), and the model is used <strong>to</strong> compute a probability <strong>of</strong><br />

random correspondence given the number <strong>of</strong> minutiae and ridge points in the<br />

reference and query images, as well as the number or correspondences between the<br />

two. This model is further described by Su and Srihari (118) : a mixture <strong>of</strong> bivariate<br />

Gaussian distributions is used for minutiae location, von Mises distributions are used<br />

for minutiae orientation and the ridge lengths are considered <strong>to</strong> be distributed<br />

uniformly. Ridge points are described by the distance from the minutia, the direction<br />

between the minutia and the ridge point as well as the orientation <strong>of</strong> the ridge point.<br />

The distance <strong>of</strong> ridge points <strong>to</strong> minutiae is modelled using a one-dimensional<br />

Gaussian distribution. The model, with a computation <strong>of</strong> specific random<br />

correspondence probabilities, is the subject <strong>of</strong> (119).<br />

Chen and Moon carry out an investigation <strong>of</strong> discriminative power using minutiae<br />

data (location, direction). Minutiae location is modelled using a uniform (spatial)<br />

distribution, and a Von Mises distribution is used for the differences between<br />

orientations <strong>of</strong> two compared minutiae. First, the model for locations is compared <strong>to</strong><br />

observed values (obtained on different available fingerprint databases), and second<br />

the model for both locations and directions is compared <strong>to</strong> known nonmatchcomparisons.<br />

The ‘score’ is constituted <strong>of</strong> matching pairs <strong>of</strong> minutiae between the<br />

(nonmatching) impressions. Good correspondence between the model and<br />

observations are shown (121). However, in a following paper, the authors show that<br />

the location <strong>of</strong> minutiae is not distributed according <strong>to</strong> complete spatial randomness,<br />

invalidating this first model (122), and the authors propose <strong>to</strong> model the minutiae<br />

locations using a quantitative s<strong>to</strong>chastic model. The results obtained are again<br />

compared <strong>to</strong> observed data as well as <strong>to</strong> the first model, and an improvement (better<br />

correspondence <strong>to</strong> empirical data) is observed.<br />

The neighbourhood <strong>of</strong> minutiae is analysed by Hsu and Martin. The locations <strong>of</strong><br />

minutiae are shown <strong>to</strong> be non-random (a preference <strong>of</strong> four inter-ridge distances is<br />

observed). The relative minutiae orientations and positions are distributed nonisotropically.<br />

Probabilities for various numbers <strong>of</strong> minutiae and numbers <strong>of</strong> nearest<br />

neighbours are given. For example, the probability for 12 minutiae is inferior <strong>to</strong> 10 -30 .<br />

The authors also show that manually annotated minutiae yield results that differ from<br />

those obtained on au<strong>to</strong>matically extracted minutiae (120).<br />

Ridge points are described in (125), and their use for improving fingerprint matching<br />

is assessed using a modified version <strong>of</strong> the Bozorth and k-minutiae matchers.<br />

Minutiae types and orientations, as well as ridge types (defined using the ridge points<br />

described in (125)) are used in order <strong>to</strong> compute a PRC (probability <strong>of</strong> random<br />

correspondence) for fingerprints by Fang and co-workers (126).<br />

Two studies using AFIS (Au<strong>to</strong>mated Fingerprint Identification System) in the context<br />

<strong>of</strong> fingerprint identification have been carried out. One makes use <strong>of</strong> AFIS scores <strong>to</strong><br />

determine the validity <strong>of</strong> fingerprint identification (115). Comparing mated inked<br />

prints, and using a logistic regression classifier, a threshold is used <strong>to</strong> declare a<br />

match. When this threshold is set <strong>to</strong> a probability <strong>of</strong> 0.5, the classification error is <strong>of</strong><br />

2.4 ± 0.1 %. When marks acquired by research assistants were used in a second<br />

experiment, this classification error was <strong>of</strong> 4.5 ± 0.4%. In the third experiment, where<br />

marks from a NIST database were used (NIST Special Database 27), a classification<br />

error <strong>of</strong> 5.4 ± 1.0% was obtained (115).<br />

235

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!