NIST Fingerprint Image Quality - NIST Visual Image Processing Group

NIST FingerprintImage Qualityand relation to latent printsElham TabassiImage Group – NISTApril 6 2006elham.tabassi@nist.gov

quality as prediction of performancewe define fingerprint image quality as a prediction of amatcher performance, e.g. a sample’s quality scorereflects the predictive positive or negative contribution ofan individual sample to the overall performance of afingerprint matching system.TARexcellent qualitysamples result inhigh performanceApril 6 2006poor qualitysamples result inlow performanceFARelham.tabassi@nist.gov

Statement of performancepoornomatchmatchexcellentthe quality measure should be indicative of the degree to whichthe distribution of match scores is separated from thedistribution of non-match scores.April 6 2006elham.tabassi@nist.gov

… but it is not easyQ 1Q 2pairwise fingerprint qualitysimilarityH(Q 1 ,Q 2 ) Qscore12matching algorithmbecause matching involves two samples, butquality is defined and measured per sample.April 6 2006elham.tabassi@nist.gov

pair-wise qualityApril 6 2006when the enrollment sample is of good quality andbetter than that of the authentication sample, theauthentication sample’s quality is sufficient topredict performance.elham.tabassi@nist.gov

NIST Fingerprint Image Quality (NFIQ)NFIQqualitynumber=1 =5excellent poor quality qualitysamples result inhigh low performanceNFIQ=1TARNFIQ=5FARApril 6 2006elham.tabassi@nist.gov

NIST Fingerprint Image QualityNFIQfeature neuralextraction networkquality number{1,2,3,4,5}feature extraction: computes appropriate signal or imagefidelity characteristics and results in an 11-dimensional featurevector.neural network:classifies feature vectors into five classes ofquality based on various quantiles of the normalized matchscore distribution.quality number: an integer value between 1(highest) and 5(poorest).April 6 2006elham.tabassi@nist.gov

poor quality samplesdistorted source e.g.scars on a fingertipdistortion in one or moresteps of the process e.g.capture or compressionnfiq=5nfiq=5April 6 2006elham.tabassi@nist.gov

NFIQ is not about human perceptionnfiq=5nfiq=2low character source the sample maysubjectively be assessed as “good”quality, but a matcher may not be ableto match it to its mate.April 6 2006elham.tabassi@nist.gov

NFIQ effectivenessfalse non-match rate1 - better quality samplesi.e. higher NFIQ numbersgive lower false rejectionrates2- five levels of arestatistically separate5 4 3 2 1NFIQ

public releaseOCTOBER192004technical report:ftp://sequoyah.nist.gov/pub/nist_internal_reports/ir_7151/ir_7151.pdf

open source????2006

Uses of NFIQ to improve performance• Conditional reacquisition– pruning the poorest quality samples (1.65% of dataset)reduced EER from .0047 t0 .0024 (sdkI - dos - ri)• Initiate invocation of special processing or matchingalgorithms• collect relevant statistics– compare capture devices and/or environments– correlation among fingersp(nfiq(ri)=5) = 0.011 p(nfiq(li)=5) = 0.016p(nfiq(li)=5 | nfiq(ri)=5) = 0.22• Multi-biometic fusionApril 6 2006elham.tabassi@nist.gov

Q - can we apply NFIQ approach tolatent quality?• perhaps not today!– Because NFIQ predicts performance of an automatedfingerprint matcher, but “currently all comparisonsof latents with candidate matches are performed byhuman examiners” i.e. no automated matchingalgorithm exists for latents.• unknowns: feature set, ground truth quality• unavailable: large enough dataset to train and test a qualityapparatusApril 6 2006elham.tabassi@nist.gov

Q -what are uses of latent quality?• NFIQ - Conditional reacquisition– Recapture of latent is extremely unlikely if not impossible• NFIQ - Initiate invocation of special processing or matching algorithms– No latent algorithm currently exist• NFIQ - collect relevant statistics– compare capture devices and/or environments• This might be the only use of latent quality to compare differentmethods and/or chemical to lift latent prints– correlation among fingers• we do not care about it in latent.• NFIQ - Multi-biometic fusion• quite unlikely• What else?April 6 2006elham.tabassi@nist.gov

conclusion• latent quality is difficult• research is lacking on automated latent matching• limited data is available• many open questions:– should latent quality be defined as prediction ofperformance?– should latent quality be a vector or just a scalar? Ifscalar is a binary (good/bad) enough?– what are the uses of latent quality?– ….April 6 2006elham.tabassi@nist.gov

thankselham.tabassi@nist.gov301 975 5292April 6 2006elham.tabassi@nist.gov

extraApril 6 2006elham.tabassi@nist.gov

Vendor F – VISIT_POE – Right indexthreshold=350 (far,tar)=(0.012,0.99)12345quality excellent veryGood good fair poorFAR 0.0037 0.0083 0.0131 0.0216 0.0477TAR 0.997 0.994 0.993 0.9496 0.926