Ride the Storm: Navigating Through Unstable Periods / Katerina Rudko (Belka G...
Ijcb2011 final
1. IJCB 2011
A Bayesian Approach to Fingerprint Minutia
Localization and Quality Assessment using
Adaptable Templates
Nathan Short, A. Lynn Abbott, Michael S. Hsiao,
Edward A. Fox
Virginia Tech
October 11th, 2011
2. Authors
Bradley Department of Electrical and
Computer Engineering
Nathan Short
Lynn Abbott
Michael Hsiao
Department of Computer Science
Edward Fox
10/24/2011
3. Motivation
Rolled/Plain Fingerprints Latent Fingerprints
Large sample of good Few good quality
quality features features for matching
Supervised acquisition of Low quality
sample fingerprint Low fingerprint surface
area
10/24/2011
6. Motivation (cont.)
Automated Fingerprint Identification System
(AFIS)
Minutia based
Aimed towards Plain/Rolled fingerprint matching
Large sample size
Automated Latent fingerprint match
performance suffers
Increase quantity of features
Minutiae
Extended features
Improve quality of features
Refine minutia descriptors
10/24/2011
Accidental friction ridge skin impression left on a surface (crime scene)Typically not visible, made visible by chemicals like powders ninhydrin then photographed or lifted with adhesive
X_t state Y_t observationAllows for use of prior state estimations when current observation has a high level of noise in low quality regionCondensation algorithm, object tracking and localization
- Particle filter
MultivariateNormal distribution centered atprevious directionT is an ideal minutia templateS is ROI around m_t^IGives a quality of least squares fitting to original image region
d_i,j found by distance transform of skeleton imagedistance to nearest non-zero pixelRange again is [-1,1]Adjust to local mean and variance
Ugly group had best improvementTAR 69.3 to 74.6 FAR 30.6 to 25.37
T – is number of iterations- Values associated with levels are empirically based
TAR increase 95.9 to 97 for goodFAR increase 4.1 to 3.0 for good
Test to see if improved locations are not a result of random chanceNon- parametric (independent of distribution of data)Paired difference testT is test statisticAlpha is critical level (95% and 99% confidence interval)-wilcoxon is more definitive
4.55x10-6 for avg. min or ~ 1/350,0003.0 x10-10 for all min pairs or ~ 1/333mil