Automatic Latent Value Determination finalbiometrics.cse.msu.edu/Presentations/TarangChugh_AutomaticLatent... · ICB 2016 Automatic Latent Value Determination Kai Cao†, Tarang Chugh†,

ICB2016

AutomaticLatentValueDetermination

KaiCao†,TarangChugh †,JiayuZhou †,Elham Tabassi ‡ andAnilK.Jain †

†MichiganStateUniversity‡NationalInstituteofStandardsandTechnology

June14,2016

ICB2016

ICB2016

WhatareLatentFingerprints?

Rolled Plain Latent

2

ICB2016

ChallengesinLatentMatching

PoorRidgeClarity PartialRidgeArea ComplexBackground

• AFISPerformance(Rank-1accuracy)– Plain:98.5%– Latent:67.2%(70.2%withimage+markup)

C.Watson, G.Fiumara,E.Tabassi,S.L.Cheng,P.Flanagan,W.Salamon.Fingerprint VendorTechnologyEvaluation,NISTIR,8034, 2012.* M.Indovina,V.Dvornychenko,R.Hicklin, andG.Kiebuzinski.ELFT-EFSEvaluationofLatentFingerprint Technologies:Extended FeatureSets,NISTIR,2012. 3

ICB2016

IndividualizationExclusion

Inconclusive

LatentMatching:ACE-V Protocol

Fingerprint examiner

ValueDetermination

FeatureMarkup

AFIS

CandidateList

C1

C2 Ck

C3..

..

Decision

ValueforIndividualization(VID)ValueforExclusionOnly(VEO)

NoValue(NV)

Retain/Discard

EvaluationVerificationbysecondexaminer

OfValue

4

ComparisonAnalysis

ICB2016

LimitationsofExaminerValueDetermination

• Highly subjective• repeatability(intra-examinervariability):84.6%• reproducibility(inter-examinersimilarity): 75.2%

• Dependsuponexaminer’sskillandexperience

• Time-consuming

Ulery etal.,“Repeatability andreproducibility ofdecisionsbylatent fingerprint examiners,” PloSone,7(3):e32800,2012.Ulery etal.,“Accuracyandreliability offorensiclatentfingerprint decisions,” PNAS,108(19):7733–7738, 2011.

Needforautomaticvaluedetermination

5

ICB2016

Minutiaefeatures

Ridgefeatures

Singularityfeatures

Orientationfeatures

AutomaticFeatureExtraction

3 4 5 6 7 8 9 14 15 16 17 18 19

FeatureVector

PredictedValue TrainedSVM

LatentwithmarkedROI

𝑽𝑰𝑫

𝑽𝑰𝑫

ProposedAutomaticMethodforValueDetermination

𝑽𝑰𝑫

𝑽𝑰𝑫

6

ICB2016

FeaturesforValueAssessmentFeatureNo. Description

1 Numberofminutiae

2- 8 Sumofminutiae reliabilitywithreliability≥t,t=0,0.1,...,0.6

9 AverageareaofminutiaeDelaunaytriangulation

10 Areaoftheconvexhullofminutiaeset

11- 17 Sumofridgequalityblockswith qualityvalue ≥t,t=0,0.1,...,0.6

18 Numberofsingularpoints(coreanddelta)

19 Standarddeviationoftheridgeflow intheforeground

7

ICB2016

Feature Extraction

InputLatent withROI RidgeFlowEstimation Normalization

8

K.CaoandA.K.Jain,LatentOrientationFieldEstimationviaConvolutionalNeuralNetwork, ICB,2015

ICB2016

Image Normalization

LocalridgeorientationOriented

window

X

X-signature

maxX

minX

OrientedwindowandX-signature

Inputlatentimage

Normalizedimage 9

ICB2016

Feature Extraction

RidgeEnhancement

InputLatent withROI RidgeFlowEstimation Normalization

10

ICB2016

Ridge Enhancement

• Dictionaryconstruction• Ridgeenhancementusingdictionary

Inputlatent EnhancedbydictionaryEnhancedbyGaborfiltering

11

ICB2016

Ridge Enhancement• Dictionaryconstruction

+=

Dictionaryelement

Valleyimage

Ridgeimage

Asubsetofdictionaryelementsusedforlatentridgeenhancement

Notethatridgeandvalleywidthscanbedifferent

12

ICB2016

RidgeQualityEstimation

Feature Extraction

RidgeEnhancement

InputLatent withROI RidgeFlowEstimation Normalization

13

ICB2016

Ridge Quality Estimation

(a) Normalizationofinput latent

(b)NormalizationofEnhancedlatent

(c)Orientationcoherenceof(a) (d)Orientationcoherenceof(b) (f)FusedQualitymapof(c),(d)&(e)

(e)Similaritybetween(a)and(b)

14

ICB2016

RidgeQualityEstimationMinutiaeExtraction

Feature Extraction

RidgeEnhancement

InputLatent withROI RidgeFlowEstimation Normalization

15

ICB2016

MinutiaeandSingularPoints Extraction

LatentID:G007 LatentID:B106 LatentID:U228

ManuallymarkedminutiaeAutomaticextractedminutiaeAutomaticdetectedcorepoint

16

N.K.Ratha,S.ChenandA.K.Jain,“Adaptivefloworientation-basedfeatureextractioninfingerprintimages”,PatternRecognition, 1995

ICB2016

Experiments• Latent Databases (Total of 707 latents)• NIST SD27: 258 latents• WVU Latent DB: 449 latents

• Latent value determination posed as two class(VID and not-VID) classification

• Two sources of ground truth• Value Determination by Examiners• Value Determination by latent AFIS*

• Latentmates retrieved at Rank-1 determined to be VID

• 10-fold cross validation protocol

*OneofthebestperforminglatentAFIS inELFTevaluation 17

ICB2016

GroundTruthConsistency

Consistency:77.8%

Agreement

AFIS:VIDExaminer:VID

AFIS:VIDExaminer:VID

18

ICB2016

GroundTruthConsistency

Consistency:77.8%

Disagreement

AFIS:VIDExaminer:VID

AFIS:VIDExaminer:VID

19

ICB2016

CorrectlyClassified

Examiner:VIDOurMethod:VID

Examiner:VIDOurMethod:VID

LatentID:G020 LatentID:U204

LatentValueGroundTruthbyExaminers

ClassificationAccuracy:85.6± 2.4% 20

ICB2016

IncorrectlyClassified

Examiner:VIDOurMethod:VID

Examiner:VIDOurMethod:VID

LatentID:B120 LatentID:G057

LatentValueGroundTruthbyExaminers

ClassificationAccuracy:85.6± 2.4% 21

ICB2016

LatentValuegroundtruthbyAFISCorrectlyClassified

AFIS:VIDOurMethod:VID

AFIS:VIDOurMethod:VID

LatentID:G004 LatentID:U213

ClassificationAccuracy:79.5± 7.2% 22

ICB2016

IncorrectlyClassified

AFIS:VIDOurMethod:VID

AFIS:VIDOurMethod:VID

LatentID:U260 LatentID:B148

LatentValuegroundtruthbyAFIS

ClassificationAccuracy:79.5± 7.2% 23

ICB2016

Conclusions

• Automatic and objective value determination of aquery latent

• No manual featuremarkup is required

• Experimental results on NIST SD27 and WVU latentdatabases demonstrated the efficiency of proposedapproach

24

ICB2016

ThankYou

AnyQuestions?