Jim Wayman（英國首席專家）：人臉和指紋：協調不同的生物特征圖像質量概念.pdf

《Jim Wayman（英國首席專家）：人臉和指紋：協調不同的生物特征圖像質量概念.pdf》由會員分享，可在線閱讀，更多相關《Jim Wayman（英國首席專家）：人臉和指紋：協調不同的生物特征圖像質量概念.pdf（15頁珍藏版）》請在三個皮匠報告上搜索。

1、Reconciling Differing Concepts of Biometric Image QualityJames L.Wayman,Ph.D.,LFIEEE104/01/2025The Importance of DefinitionsNo quantity can be measured with more accuracy or precision than that in which it is defined Definitional accuracy and precision limits measurement accuracy and precisionDiffer

2、ent definitions will lead to different values and different applications for a 204/01/2025Complexities in Defining“Biometric Quality”ISO/IEC 29794-1“Information technology Biometric sample quality Part 1:Framework”presents 3 interrelated conceptualizations of“quality”Character:set of attributes asso

3、ciated with a biometric characteristic that cannot be controlled during the biometric acquisition process Biometric fidelity degree to which a biometric sample is representative of its source biometric characteristic Biometric utility degree to which a biometric sample supports biometric recognition

4、 304/01/2025“Character”“Biometric Recognition”is a human/machine technology “Character”,mentioned explicitly in all Parts of 29794,refers to human characteristics that“cannot be controlled”(by collector)but sometimes can be tested Implies Platos idealized concept of“form”Inherent demographic bias As

5、sumed but not tested Faces have 2 aligned eye sockets Time invariant Explicitly tested:Fingers have coherent friction ridges Face not occluded beards Iris boundaries concentric,iris/pupil/sclera gray scale difference 404/01/2025“Fidelity”Primarily technical characteristics of“representative”collecti

6、on Collection device characteristics defined in Annex D of relevant parts of ISO/IEC 39794“Extensible Biometric Interchange Formats”Limitations:Part 4(Contact FP),Part 5(Machine Readable Travel Docs)Concepts of“fidelity”differ across modes Illumination specified for Part 5(Face)and Part 6(Iris)but n

7、ot Part 4(Contact FP)No light reflection from finger,so skin albedo is not a factor Other collection device specs:Modulation transfer function Spatial Sampling Optical distortion Pixel aspect ratio Signal to Noise 504/01/2025“Utility Supports Recognition Performance”Two inter-dependent pillars of er

8、ror rate performance Stability:supports false non-match rate Distinctiveness:supports false match rate Influence of quality on false match rate not addressed in Parts 5 and 6 DoD“Cotton Ball”problem established theoretical dependence of distinctiveness on 604/01/2025Performance and ConformanceBOLD a

9、dded ISO/IEC 29794-1:2024“Part 1:Framework”7.1.3 Quality measure(quality score or q”uality component)or errorQuality scores enable discrimination between distinct levels of performance.A quality score shall predict performance metrics such as false match and false non-match rates when comparisons ar

10、e made to references developed under stated collection policies.ISO/IEC 29794-5:2025“Part 5:Face image data”1.Scope “This document establishes requirements on implementations that quantify how a face images properties conform to those of canonical face images,for example those specified in ISO/IEC 3

11、9794-5:2019,Clause D.1.”04/01/7The Complexity of“Quality”Character,Fidelity and Utility are not separate and not independent Utility=performance=f(error rates,throughput)Error rates=f(stability,distinctiveness,mode/algorithm)Distinctiveness=f(stability,character,fidelity,algorithm)Stability=f(collec

12、tion conformance,character consistency,mode/algorithm)So “Quality”as focused on“Utility”is a complicated function of character,fidelity,collection conformance,mode/804/01/2025ISO/IEC 29794-1:2024 Examples of Use“real-time quality feedback(in)biometric capture process”“data fusion”“hardening systems

13、against presentation attacks using or targeting low quality biometric samples”“correlating quality measures to other system metrics can be used to diagnose problems and highlight potential areas of performance improvement.”904/01/2025Points of Application in Biometric SystemISO/IEC SC27 Standing Doc

14、ument 1004/01/2025Points of Application in Biometric SystemISO/IEC SC27 Standing Document 11Front EndBack End04/01/2025Enhancing Performance on Front End Front-end application Data capture subsystem“Failure to Capture”Signal processing subsystem “Failure to Acquire”Reference creation Enhancement thr

15、ough“Discard Rates”Only a concept in Parts 1 and 5 No mention in Parts 4 and 1204/01/2025Enhancing Performance on Back End“Pairwise”comparison of quality components“Quality comes in pairs”P.Jonathon Phillips Part 1:10.Pairwise quality“In some applications,there are no assumptions regarding the confo

16、rmance of either the probe or the reference to any collection best practices or requirements”.Part 6:6.4 Iris image quality metrics computed from two images Part 5:Annex B.1“Q=F(X1),predicts(Score)because it implicitly assumes the comparison V(X1,XPORTRAIT).Quality scores are evaluated as predictors

17、 of mated comparison scores;”Use of quality score in comparison algorithm choice Quality support for data fusion Nandakumar,et al(2006)Poh,Kittler,and Bourlai(2010)1304/01/2025PARTCONFORMANCMETRICCHARACTER(BIOLOGY)FALSE NON-MATCH RATEFALSE MATCH RATEDISCARD(FRONTEND)PAIRWISE(BACKEND)1404/01/2025Reco

18、mmendationsAll parts of ISO/IEC 29794 should be harmonized to:1.Acknowledge both“performance”and“conformance”as reasonable approaches to quality2.Include“pairwise”mutual quality comparison for backend applications3.Acknowledge impact of quality on false match rate4.Inform use of quality metrics in data 1504/01/2025