Documente Academic
Documente Profesional
Documente Cultură
Future Directions
Salil Prabhakar
Digital Persona Inc.
Fingerprint Applications
Scale
1010
105
Usability
Reasons for Accuracy Challenges
■ Information Limitation
– Due to individuality, poor presentation, and inconsistent acquisition
■ Representation Limitation
– Design and choice of representation (features) and quality of feature
extraction algorithms (especially for poor quality fingerprints)
■ Invariance Limitation
– Incorrect modeling of invariant relationships among features
Fingerprint Individuality Estimation
Accuracy; Information Limitation
Probability that one of one input minutiae matches any of the m template minutiae:
mC
A
Probability that two of two input minutiae matches any of the m template minutiae:
2x
mC A − mC
x
A A −C
Probability of a False Correspondence
Accuracy; Information Limitation; Fingerprint Individuality Estimation
Let l be such that P(min(|θ ’i-θ j|,360-|θ ’i- θ j|)≤ θ 0) =l. Then,
m M − m
min( m ,n ) ρ
n − ρ ρ q ρ −q A
p (M , m , n , q ) = ∑ × (l ) (1 − l ) , where M ≡
ρ =q
M
q C
n
Finally, since minutiae can lie only on ridges, i.e., along a curve of length A/w,
where w is the ridge-period, M is modified as:
A /w
M≡ where 2r0 is the length tolerance in minutia location.
2r0
Upper Bound on Fingerprint Accuracy
Accuracy; Information Limitation; Fingerprint Individuality Estimation
Theoretical
M,m,n,q P(Correspondence)
248, 46, 46, 46 1.33 x 10-77
248, 46, 46, 12 5.86 x 10-7
70, 12, 12, 12 1.22 x 10-20
Empirical
Database m,n,q P(Correspondence)
MSU_DBI 46, 46, 12 5.8 x 10-2
The probabilities of false correspondences for various values of q are computed from
our theoretical model based on the parameters estimated from a Ground Truth database
and the MSU_DBI databases and compared with the empirical probability of false
correspondence obtained from the MSU_DBI database using an automatic fingerprint
matcher.
The entry (70, 12, 12, 12) corresponds to the 12-point guideline.
Lower Bound on Fingerprint Accuracy
Accuracy; Information Limitation; Fingerprint Individuality Estimation
Minutiae-based representation
may not be most suitable Fingerprints from two different fingers
■ Minutiae-based
– Sequential design based on the following modules:
Segmentation, local ridge orientation estimation (singularity and
more detection), local ridge frequency estimation, fingerprint
enhancement, minutiae detection, and minutiae filtering and
post-processing.
■ Ridge Feature-based
– Size and shape of fingerprint, number, type, and position of
singularities (cores and deltas), spatial relationship and
geometrical attributes of the ridge lines, shape features, global
and local texture information, sweat pores, fractal features.
Representations: Future Directions
Accuracy; Representation Limitation
(( ) (
P = x1P , y1P ,θ1P , , xMP , y MP ,θ MP ))
Q = (( xQ
1 , y1Q ,θ1Q ), , ( x Q
N , y NQ ,θ NQ ))
■ where x, y, and q are the x-coordinate, y-coordinate, and
minutiae direction.
a) b) c)
Scale: Future Directions
■ Continuous classification
■ Feature-based indexing (search and retrieval) schemes
(e.g., minutiae triplets)
■ Fast matchers
■ Classifier combination
70
minutiae triplets
60
ration
Penet
(%)
50
orientation image
40
30 Combination
20 FingerCode
10
0
0 5 10 15 20 25 30 35 40
Error (%)
Multiple Biometrics; Fusion
■ A decision (and lower) level fusion of multiple biometrics can
improve performance