Lt tLatent Fi i tFi ngerprint Q litQ uality AtA...
Transcript of Lt tLatent Fi i tFi ngerprint Q litQ uality AtA...
L t t Fi i t Q lit A tLatent Fingerprint Quality Assessment
Anil K. Jain, Soweon Yoon, Eryun Liu and Kai CaoMichigan State University
Project # 12S 04W 12Project # 12S‐04W‐12
Image Quality• Image quality indicates “perceived” image degradation with/without relation to a referencedegradation with/without relation to a reference image
• Factors affecting image qualityFactors affecting image quality– Sharpness, contrast, noise, distortion, resolution, dynamic range,..
S d d D fi i i Hi h D fi iti• Quality assessment
– Qualitative (Good/bad/ugly)
Standard Definition High Definition
vs.Quantitative (SNR)
Fingerprint Image Quality• Prediction of AFIS performance for feature extraction and matching
– “Perceived” fingerprint image quality may not necessarily correlate with AFIS performance
Good quality fingerprint image (NFIQ* = 1) Poor quality fingerprint image (NFIQ = 5)
* NIST Fingerprint Image Quality; value is from 1 (highest quality) to 5 (lowest quality)
Rolled vs. Latent Quality Rolled Fingerprint
Assessment• Rolled quality assessmentq y
– Clean background; central part of the finger– Clarity of ridge and valley structures– Minutiae quality and number
• Latent quality assessment– Background noise, off‐center finger, skin
dLatent Fingerprint
distortion– Local ridge quality alone not adequate– No existing work to predict AFIS
performanceperformance• Latent ridge clarity assessment in [1] targets at correlating the automatic metric to human examiner’s assessment
[1] R. A. Hicklin et al., “Assessing the clarity of friction ridge impressions”, Forensic Science International, 2013
Latent Quality Assessment by Examinersy y
• ACE‐V methodologygy• Examiner determines latent value in analysis phase:
– Value for Individualization (VID)( )– Value for Exclusion Only (VEO)– No Value (NV)
• Only VID or VEO latents are searched via AFIS• Concern: Reliability and consensus of examiners
– Visual perception, expertise of examiners, workload, etc.
Goals of Our Studyy• Provide an objective measure of latent quality to avoid
i l di l i i bj ti lit ( l ) l timisleading conclusions in subjective quality (value) evaluation
• Identify latents which can be processed in “Lights‐out” mode
Good quality latent (VID) Poor quality latent (NV)
Both latents were identified at rank 1 by AFIS
Latent Value vs. Identification Rate
Value for Value for No ValueIndividualization Exclusion Only No Value
NIST SD27* (258 latents) 210 41 7
WVU (449 latents) 370 74 5
Rank‐1 ID Rate 491 (85%) 46 (40%) 1 (8%)
Rank‐100 ID Rate 525 (91%) 72 (63%) 7 (58%)
A significant number of VEO or NV latents can be successfully identified by AFIS
Identification rate is obtained by combining multiple AFIS; if the mate of a latent is retrieved within rank m by any of the AFIS, it is considered as a successful match within rank m
* Hicklin et al., “Latent Fingerprint Quality: A Survey of Examiners”, Journal of Forensic Identification, 61(4), 2011
Latent Quality DefinitionLatent Quality Definition
• Salient featuresSalient features– Local ridge clarity in presence of background noiseVicinity of good quality ridge areas– Vicinity of good quality ridge areas
– Position of markMi ti li bilit– Minutiae reliability
h h• Matcher‐independent vs. matcher‐dependent
Matcher‐Independent vs. Matcher‐DependentMatcher Independent vs. Matcher Dependent
• Matcher‐Independent Quality MeasureMatcher Independent Quality Measure– A latent is considered VID if any one of the AFIS can successfully retrieve its mate from a referencecan successfully retrieve its mate from a reference database within the candidate list
• Matcher‐Dependent Quality MeasureMatcher Dependent Quality Measure– A latent is considered VID if a specific AFIS can successfully retrieve its mate from a referencesuccessfully retrieve its mate from a reference database within the candidate list
AFIS Interoperabilityp y
Retrieval Rank AFIS 1 AFIS 2 AFIS 3
Proprietary Minutiae 32 561 222
Markup Minutiae 31,997 156 1
Local Ridge Qualityg Q y• Ridge Clarity
FFTx =
• Ridge Continuity
Gaussian Mask
Power Spectrum
Local Block
• Ridge ContinuityContinuous blocks Discontinuous blocks
Orientation Frequency Phase
Minutiae Reliability: Learningy g• Minutiae patch dictionary learning
Hi h Q lit
Dictionary Elements
High QualityMinutiae Patches (48x48)High Quality Fingerprints
…
…
Minutiae ReliabilityMinutiae ReliabilityLatent Fingerprint Minutia Patch Dictionary Elements
…
P D = {dm|m = 1, 2, …, M}
Reliability of a patch P (Qm) is defined as the Structural SIMilarity (SSIM) between P and its
{ m| , , , }
Structural SIMilarity (SSIM) between P and its closest dictionary element dm:
Qm = max {SSIM(P, D)}Qm max {SSIM(P, D)}
Wang et al., “Image Quality Assessment: From Error Visibility to Structural Similarity”, TIP, 13(4), 2004
Reference Point DetectionReference Point Detection
Orientation Field Curvature map Weight Map(Reconstructed from minutiae)
• Reference point is determined as the point where the curvature is maximum
A. Yoshida and M. Hara, “Fingerprint Image Quality Metrics That Guarantees Matching Accuracy”, Biometric Quality Workshop, 2006.
Latent Fingerprint Image Quality (LFIQ)
• For each minutiae triangle Ti,
3
1jMMRT ijijii
WQQQ
QMMij TQM thatsuchfor QMMij TQMijthatsuchfor
iR TQi
in quality ridge Average :
iM TjQij
in minutia th theof Reliablity :
• Quality score of a latent:
positionfingertheonbasedWeight :ijMW
• Quality score of a latent:
N
iTi
QLFIQ1
latentintrianglesof Number :N
LFIQ Distribution15
)
707 Latents
10
rcen
tage
(%)
5Per
0 20 40 60 80 100 120 140 160 180 2000
LFIQ
30Quality Index = 1 10050 80
LFIQ PerformanceLFIQ Performance• 707 latents from NIST SD27
d WVU L D b90
100
and WVU Latent Database• 31,997 exemplar fingerprints• Three COTS matchers; 70
80
90
Rat
e (%
)
matcher‐independent quality• Minutiae from examiners’
markup and AFIS40
50
60
Iden
tific
atio
n
• Oracle: Knows the true retrieval rank of the latent
10
20
30
Ran
k-1
Oracle (Q = True Retrieval Rank)Latent Quality Measure in [1]
≡
10 20 30 40 50 60 70 80 90 1000
Quality Index
Proposed LFIQ
[1] S. Yoon, E. Liu, and A. K. Jain, “On Latent Fingerprint Image Quality”, International Workshop on Computational Forensics, 2012
Successful Prediction• Quality Index 96 (LFIQ = 70); Mate retrieved at rank 1; examiner labeled it as NV latent
Latent Mated Rolled Print
Unsuccessful Prediction• High quality, but low matching performance
Latent Mated Rolled PrintLatent Mated Rolled Print
• Quality Index 93 (LFIQ = 64)• Value determination by examiner: VID• Retrieval rank of the mate: 600
Latent Examiners’ Value Determination vs. LFIQ
100
LFIQ is correlated with value determination by latent examiners
80
90
100
Mean LFIQ of VEO Latents
Mean LFIQ of VID Latents
50
60
70
ndex
Mean LFIQ of NV Latents
Mean LFIQ of VEO Latents
30
40
50
LFIQ
I
10
20
VID VEO NV0
Value Determination by Latent Examiners
LFIQ as Tenprint Quality MeasureLFIQ as Tenprint Quality MeasureLFIQ COTS Quality Measure
0.8
0.9
1
0.8
0.9
1
0 4
0.5
0.6
0.7
atch
Sco
re
0 4
0.5
0.6
0.7
atch
Sco
re0 1
0.2
0.3
0.4Ma
0 1
0.2
0.3
0.4MaOracle
(Q ≡Match Score)
0 20 40 60 80 1000
0.1
Quality Index
0 20 40 60 80 100
0
0.1
Quality Index
(Q ≡ Match Score)
• LFIQ scores are highly correlated with COTS match scores for NIST SD4 LFIQ can also be used as a tenprint quality measure
LFIQ vs. NFIQ for Rolled FingerprintsLFIQ vs. NFIQ for Rolled Fingerprints
LFIQ NFIQ
0.8
0.9
1
0.8
0.9
1
0 4
0.5
0.6
0.7
atch
Sco
re
0 4
0.5
0.6
0.7
atch
Sco
re
0 1
0.2
0.3
0.4Ma
Oracle(Q ≡Match Score) 0 1
0.2
0.3
0.4Ma
0 20 40 60 80 1000
0.1
Quality Index
(Q ≡ Match Score)
1 2 3 4 50
0.1
NFIQ
• Both LFIQ and NFIQ are correlated to genuine match scores of fingerprints in NIST SD4
Conclusions
• Latent image quality measure is crucial for properly g q y p p ydetermining latent value as forensic evidence
• Defining latent quality in terms of a single index is g q y gchallenging
• Proposed LFIQ is an objective measure of latent qualityProposed LFIQ is an objective measure of latent quality– Can distinguish latents that can be processed in “lights‐out” mode
– Complement latent examiners’ value determination
Future WorkFuture Work
• Incorporate features from ridge clarity that areIncorporate features from ridge clarity that are robust to noise; develop a mapping function from feature vector to quality scorefrom feature vector to quality score
• Normalize and quantize LFIQ score to provide better interpretationbetter interpretation
Publications & PresentationPublications & Presentation
• Publicationsub cat o s– S. Yoon, K. Cao, E. Liu, and A. K. Jain, "LFIQ: Latent Fingerprint Image Quality", BTAS, Washington, D.C., S t 29 O t 2 2013Sept. 29‐Oct. 2, 2013
– S. Yoon, E. Liu, and A. K. Jain, "On Latent Fingerprint Image Quality", ICPR IWCF Workshop, Tsukuba, Japan, g Q y , p, , p ,Nov. 11, 2012
• Presentation– “Latent Fingerprint Image Quality”, Fingerprint Image Quality Assessment NFIQ 2.0, Winchester, U.K., April 26 201326, 2013
Thank you