ref paper3

8/7/2019 ref paper3

1/13

NIST Fingerprint Image QualityNIST Fingerprint Image Quality

Elham Tabassi

Biometric Consortium Conference

September 20, 2005

8/7/2019 ref paper3

2/13

BCC 2005 [email protected]

backgroundbackgroundo quality is important

o previous research mostly defined as a measure of the extractability of the

features used for recognition such as minutiae.

local orientation information (Bolle et al, Shen et al, Hong etal., )

global features (Hong et al, Lim & Yao, Nill & Bouzas, )

o almost all used subjective quality assessment toevaluate their quality algorithm

size of fingerprint, pressure, humidity, amount of dirt,

8/7/2019 ref paper3

3/13


quality as prediction of performancequality as prediction of performance

we define fingerprint image quality as a prediction of a matcherperformance, e.g. a samples quality score reflects the predictive

positive or negative contribution of an individual sample to the overallperformance of a fingerprint matching system.

FAR

TAR

excellent quality

samples result inhigh performance

poor qualitysamples result inlow performance

8/7/2019 ref paper3

4/13


use of quality to improve performanceuse of quality to improve performance

o recapture samples of insufficient quality

pruning the poorest quality samples (1.65% of dataset) reduced EERfrom .0047 t0 .0024 (sdkI - dos - ri)

o process samples differently based on their qualities

o collect relevant statistics

compare capture devices and/or environments

correlation among fingers

p(nfiq(ri)=5) = 0.011 p(nfiq(li)=5) = 0.016

p(nfiq(li)=5| nfiq(ri)=5) = 0.22

o cause higher quality sample dominate fusion

8/7/2019 ref paper3

5/13


qualitynumberNFIQ


=1=5

NFIQs 5 levels of quality are intended to be predictive of therelative performance of a minutia based fingerprint matchingsystem.

NFIQ=1 indicates high quality samples, so lower FMR and/orFNMR is expected.

NFIQ=5 indicates poor quality samples, so higher FMR and/orFNMR is expected.

8/7/2019 ref paper3

6/13


performance targetperformance targetdegree of separation between a sampledegree of separation between a samples genuine and imposter distributionss genuine and imposter distributions

4 3 6 6 8 1 0 0 1 3 2 1 6 4 1 9 6 2 2 8

0 .00

0 .05

0 .10

0 .15 m a tch sco res

non m a t ch sc o res

s d 2 9 - v tb m a t c h a n d n o n m a t c h s c o re s h i s to g ra m

nomatch

match

excellent

poor

quality of a biometric sample xi prediction of the bin its normalized

match score falls

8/7/2019 ref paper3

7/13


pairpair--wise qualitywise quality

fingerprint

matching algorithmsimilarity

score

Q1

Q2

pairwise qualityH(Q1,Q2) QQ1212

when the enrollment sample

is of good quality and better

than that of the use phase(search) sample, the search

samples quality is sufficient

to predict performance.

8/7/2019 ref paper3

8/13


qualitynumberNFIQNFIQ


o feature extraction: computes appropriate signal or image fidelitycharacteristics and results in an 11-dimensional feature vector.

o neural network:classifies feature vectors into five classes of qualitybased on various quantiles of the normalized match scoredistribution.

o quality number: an integer value between 1(highest) and 5(poorest).

feature neuralextraction network

8/7/2019 ref paper3

9/13


NFIQ effectivenessNFIQ effectivenesso evaluation criterion is rank ROC as a function of image quality

o used various fingerprint matching algorithms and variousdatasets to evaluate NFIQ

15 different COTS fingerprint matching algorithms

22 different datasets of different fingers captured by various devicesand at different operational settings

each test dataset has 2 fingerprint images of 6000 person

o compared (TAR,FAR) of levels of quality at a fixed threshold

as quality degrades, true accept rate decreases for all the matchers,FAR increase for some.

o levels 2,3,4, and 5 are statistically separable.

o It takes about one third of a second to compute quality of a flatfingerprint image.

8/7/2019 ref paper3

10/13

6000 subjects - Right index

Threshold @ (far,tar)=(0.012,0.99)

quality1

excellent2veryGood

3good

4fair

5poor

FAR 0.0037 0.0083 0.0131 0.0216 0.0477

TAR 0.997 0.994 0.993 0.9496 0.926

8/7/2019 ref paper3

11/13


separable levels of qualityseparable levels of quality

For each quality level,we calculated 95%confidence intervals of

TARs @ FAR=0.1% forsix matchers and sixteendatasets.

nfiq levels 2,3,4, and 5are statistically separate.

8/7/2019 ref paper3

12/13


conclusionconclusion

o a novel definition of fingerprint image qualityo NFIQ works as a rank statistic for performance for all 330

combinations of COTS fingerprint matchers andoperational datasets tested

o NFIQ levels 2,3,4, and 5 are statistically independento NFIQ can be used for real-time quality assessment

o NFIQ is publicly available but subject to US export control

laws (fingerprint.nist.gov)

8/7/2019 ref paper3

13/13

thanksthanks

[email protected]@nist.gov

301 975 5292301 975 5292
mailto:[email protected]:[email protected]:[email protected]

ref paper3

Documents

Transcript of ref paper3