Biometric Systems BIOM 426
Popular in Course
Popular in Information technology
This 163 page Class Notes was uploaded by Karli Wunsch on Saturday September 12, 2015. The Class Notes belongs to BIOM 426 at West Virginia University taught by Natalia Schmid in Fall. Since its upload, it has received 13 views. For similar materials see /class/202794/biom-426-west-virginia-university in Information technology at West Virginia University.
Reviews for Biometric Systems
Report this Material
What is Karma?
Karma is the currency of StudySoup.
You can buy or earn more Karma at anytime and redeem it for class notes, study guides, flashcards, and more!
Date Created: 09/12/15
Performance Evaluation BIOM 426 Instructor Natalia Schmid October 4 2005 WVU Statistical Measures for Biometrics I False Accept Rate FAR FAR is the probability that a user making a false claim about hisher identity will be verified as that false identity Major reasons threshold biometrics of claimed identity is close to biometrics of the user making claim FAR characterizes the strength of the matching algorithm I False Reject Rate FRR FRR is the probability that a user making a true claim about hisher identity will be rejected as himherself Major reasons threshold the presented biometric features are not close enough to the biometrics template in the database FRR characterizes the robustness of the algorithm I False to Enroll FTE FTE is the probability that a user attempting to biometrically enroll will be unable to Rule of 3 attempts Major reasons the biometric hardware an algorithm that is not tuned properly a user who is slow in learning how to submit biometrics I Equal Error Rate EER Will be defined later October 4 2005 WVU 2 Basic System Errors I Before designing any system biometric error rates have to be analyzed I Two important facts 39 Any biometric system will make a mistake 39 The true value of error rates cannot be theoretically derived 39 Since biometrics measurements 2D representation or templates are available for verification of individual a metric SIT a measure of closeness or difference between two templates has to be introduced 39 Sources of noise preprocessing mapping thermal noise discretization quantization acquisition error etc October 4 2005 WVU Hypothesis Testing Problem The problem of verification can be stated as a hypothesis testing problem Two hypothesis are possible H0 Two samples match null HA Two samples do not match alternative Another statement A and A are two biometrics H0 A A null HA A 7E A alternative To decide between two hypothesis the score SIT is computed H0 SIT gt 7 null HA SIT lt 7 alternative October 4 2005 WVU Score Distributions I The reliability of the score is in uenced by 39 Variations in the live biometrics in time 39 Variations from sensor to sensor 39 Variability in the sampling process 39 Because of these reasons SIT will never be one for templates from the same live biometrics and will never give zero for two templates from two different individuals 39 False Match decide that A A because SIT gt y when in fact A 75 A I False Reject decide A 75 A because SIT lt y when in fact A A October 4 2005 WVU Estimation of FAR and FRR I We need a set of M genuine pairs of biometrics samples M genuine scores and N imposter pairs of biometrics samples N imposter scores I A genuine score is formed as a similarity or difference metric between two templates belonging to the same class I An imposter score is formed as a similarity or difference metric between two templates belonging to two different classes I Fix a threshold 7 The best guess about FAR is A l FAR WZKSM Ji gt y I The best guess about FRR is A 1 FRR lSITlt MEN 1 y October 4 2005 WVU Histogram Histogram is a frequency plot of data samples versus discretized measurement range AONEInOEIEH 60 80 100 120 140 160 180 200 0 BINS 4000 3500 3000 2500 2000 1500 1000 Dataset of U of Bologna October 4 2005 WVU ROC curve 0 FAR and FRR can be estimated for each 7 0lt ylt1 for example 0 Define a function FARO2 y in ymin ymaX gt01 0 Also FRRW can be plotted vs FARW parameterized by 7 FR i I 5 i I 39 71 72 73 7 Y1 Y2 Y3 Y Y3 Y2 Y1 1 FAR FAR and FRR as functions of y and the resulting ROC curve October 4 2005 WVU 8 Operating points 0 Operating point can be speci ed by specifying y FAR and FRR can be estimated for a given 7 0 When comparing two matchers the operating point is speci ed by FAR andFRR 0 Most biometrics cannot guarantee that both rates are low Secure applications require low FAR Convenient application require low FRR 0 There is a tradeoff between FAR and FRR security High FRR convenient Low FRR I Low FAR 1 FAR High FAR October 4 2005 WVU Variatlons of ROC curves 0 Plot of correct Reject Rate l FRR vs FAR 0 FRR vs FAR with one of axes being on logarithmic scale 0 both axes are logarithmic Be careful when reading ROC curves October 4 2005 WVU 10 Using the ROC Curve 0 The ROC curve provides a complete specification of a single biometric matcher performance 0 ROC shows a trade off between FAR and FRR over a broad range of thresholds 0 Useful when two or more matchers have to be compared 0 Given two matchers A and B which one is more accurate 0 The performance of matcher depends on its operational point October 4 2005 WVU ll Using ROC Curve Example 1 FRR 0 Easy case 0 Matcher A is always A better than Matcher B FRR gtFRR B A 9 0 It s a rare case FARBFARA October 4 2005 WVU 12 Using ROC Curve Example 2 FRR At point C Matchers are equally accurate Matcher B is more accurate than Matcher A if the I operational point is to the right of C FRRB gt FRRA FRRB lt FRRA 39 Matcher A is more f accurate than Matcher B if I FAR the operational point is to FARE FAR A the left of C October 4 2005 WVU 13 The Equal Error Rate EER I When comparing two matchers a single number is desired I Single point performance measures 0 EER d prime 0 overall expected error I The EER is the operating point on ROC such that FAR FRR I In Example 1 A has lower EER than B A is better than B but over a narrow range of error rates I dprime is the measure of separation of two histograms Genuine and Imposter October 4 2005 WVU Dp m is the mean of Genuine pdf is the mean of Imposter pdf m G m1 d 0 012 039 is the variance of Genuine pdf 02 is the variance of Imposter pdf Aouanbag Aouenbeg m Genuine I Imposter Genuine Imposter I matchlng score matching score Example of distributions with equal dprime values October 4 2005 WVU Overall Expected Error 137 7r FARO 1 7 FRRW Where 71 is a prior probability to pick an Imposter sample often set to 05 October 4 2005 WVU Practical Case Example Generated data GS genuine score 15 score values IS imposter score 105 score values GS 51 90 74 100 75 73 95 78 61 86 95 51 86 70 95 IS 31x6 32x3 33x4 34x2 35Xl 36x2 37Xl 38x4 39x2 40x4 42x4 43x3 45x3 47x3 48x3 49x2 50x3 5lX3 52x8 53Xl 54Xl 57x2 58Xl 59x2 60x2 6lX4 62Xl 63x5 64Xl 65x5 66Xl 68Xl 69x3 7lX2 72x4 73x3 74Xl 75Xl 77x2 78Xl October 4 2005 WVU l7 SOOZ 17 1910130 HAM anaA woos anaA 31003 SI Frequency 0 Number of users with particular score 9V0 value Number of users with particular score value x x Jaisoduu augnuas enEA moss Frequency Z39O D o o 01 m 90390 amen 31003 augnuas Example 02 Genuine and lmposter False Reject and False Accept K0119quotij 1 0 False Accept T1 lt0 v 0 O Mulqeqmd N o 0 w m 20 40 l 60 80 100 0 20 40 60 80 Sec value T Threshold T1 T2 T3 False Reject and False Accept ROC curve w lt0 v 0 O O Mulqeqmd N o 4 False Reject False Accept 2 o o w 4 1 O jdeocv aspd N o October 4 2005 0 j m m 06 08 False Reject 20 40 60 80 100 0 02 04 Threshold WVU Example Displaying ROC curves 08 06 HEH 04 A A 02 NOIlOEI EIH lOEIHHOO FAR 2 08 g l o LIJ a E a U I o LIJ n a o o 02 08 06 06 02 O 02 04 06 08 1 0 02 04 06 08 1 CORRECT ACCEPTANCE CORRECT ACCEPTANCE October 4 2005 WVU Example dprime d meanl meanZ meanl 7866 mean2 5189 Std12 Std22 Stdl Std2 Expected Error E7 751 39FARW 752 39FRRO is a threshold 7 lengthUS 7r lengthGS 1 lengthISlengthGS 2 lengthGSlengthIS 71 2 0875 72 2 0125 October 4 2005 WVU 21 Example E lengthISlengthISlengthGSFAR lengthGSlengthGSlengthISFRR 4 SCALED FAR 4 SCALED FRR 4 EXPECTED ERROR HOHHEI GEIlOEIdXEI 0 1 0 20 3O 4O 50 60 7O 80 90 1 00 TH RESHO LD 1 October 4 2005 WVU 22 Example 2 Hand Geometry Observed data 5 users 2 independent hand measurements per user User 11 71 63 70 61 74 56 56 52 281 362 268 278 243 136 User 12 67 61 74 62 77 61 54 55 274 373 257 285 247 147 User 21 64 65 69 65 45 40 41 42 240 300 250 261 240 154 User 22 65 66 71 66 46 41 42 42 240 301 251 262 240 155 User 31 75 69 72 67 76 60 62 57 290 310 275 290 253 153 User 32 74 66 72 68 74 62 64 57 291 310 275 291 252 157 User 41 57 55 63 54 65 49 47 45 250 320 269 275 240 147 User 42 57 55 65 54 65 47 48 42 251 323 268 277 243 144 User 51 55 56 63 53 60 47 48 47 249 303 258 268 241 152 User 52 50 51 60 50 52 43 42 45 240 295 269 270 249 157 Form two training sets Genuine and Imposter Total number of measurements 10 Total number of imposter scores 40 Number of distinct airs P C 10W10 10 this number includes 4 5 genuine and imposter K 2 j 28 scores October 4 2005 WVU 23 Example 2 Hand Geometry Imposter Set User 11User 21 User 11User 31 User 11User 41 User 11User 51 User 12User 21 User 12User 31 User 12User 41 User 12User 51 User 21 User 31 User 21 User 41 User 21 User 51 User 22 User 31 User 22 User 41 User 22 User 51 October 4 2005 User 11User 22 User 11User 32 User 11User 42 User 11User 52 User 12User 22 User 12User 32 User 12User 42 User 12User 52 User 21User 32 User 21User 42 User 21User 52 User 22User 32 User 22User 42 User 22User 52 WVU User 31 User 41 User 31 User 42 User 31 User 51 User 31 User 52 User 32 User 41 User 32 User 42 User 32 User 51 User 32 User 52 User 41 User 51 User 41 User 52 User 42 User 51 User 42 User 52 Genuine Set User 11User 12 User 21User 22 User 31User 32 User 41User 42 User 51User 52 24 Example 2 Hand Geometry Results EER 0024 dprime 216 October 4 2005 SEIWOOan IO HEIGWnN SEIWOOan IO HEIEWI39IN GENUINE SCORE ll39 2 0 N I SCORE GENUINE AND IMPOSTER SCORES n u GENUINE IMPOSTER Q I V N 0 20 4O 6O 80 100 88d 0 20 4O 60 80 100 SCORE WVU IMPOSTER SCORE O 20 40 60 80 100 R00 CURVE 25 39 4 I r lt e A A V x 39 i 7 39 a a T Stamina aquot quotquot x v 39A 39 O I W M l unumu m Images are from Handbaak of F ingerprimRecognilian by D Maltoni et 211 August 30 2005 B10M426 1 Introduction Applications law enforcement access to computer network bank machine car home security applications US Visit National ID card Evinnon riseapansinmauui ed Slams August 30 2005 BIOM 426 liea ons about 1 sec 9 hmeasi nggnumb em f netwarks and ltezm eb transactions T le need I l I m linpul mm component of reliable security well accepted by public August 30 2005 BIOM 426 3 00 to 6000 BC by ancient bricks anon 1quot Babylon around 1792 1750 BC Figure 18 39 m 39 annn 30 Lee and Gaensslen 2001 c a Chinese Clay seal 300 BC Lee and Gaenssien 2001 cl an impression on a Palestinian lamp 400 AD Moensssns 1971 Although impres siuns on the Neolithic carvings and the Goal island standing slones might not be used to india caie iripniirv L 39 39 L39 J 39 F 39 on the Palestinian lamp were used in indicate the identily of the providers Figures caurlesy of A Moenssens R Gaenssien and J Berry u August 30 2005 BIQM 426 4 s 11 Francis Galton introduced minutiae features for matching in 1888 Ed wfard Henry 1897 established Henry sy em of fi ngerprint classi cation In th mid180039s two facts were established no two fingerprints have the same ridge pattern and ii fingerprint The nine pattems illustrated in Purkinje s thesis pattern have good permanence By the early 20th century the formations of fingerprian Were Well understood August 30 2005 BIOM 426 5 39on was accepted as a personal tentfingerprint acquisition e loped setup in 1924 with a database of 393 Computer proee began in 1960s quotntrodueition of computer hardware Basis 1 for automatic identification Since 19808 fingerprints are used in non crir39ninal applications due to personal computers and optical scanners Personal use due to introduction of inexpensive capture devises and reliable matching algorithms August 30 2005 BIOM 426 6 Size 1 X 1 i small Resolution 500 dpi Ultrasonic scanning high quality images August 30 2005 BIOM 426 mmmlcunvmer Regpmgihle fmommglniga ng with external devices Secure identification system requirements protectionenmyption secure identification system discard fake fingerprints Additional Issues storage for large AFIS compression methods August SO 2005 BIOM 426 7 compliant scanners correlation techniques August 30 2005 i Minimum resolution for FBI BIOM 426 9 Figure 23 Examples of ngerprint images acquired with an Optical scanner a good quality ngerprint 11 a u Ingm c a 39 a d an inlrin cally had ngerprint August 30 2005 BIOM 43926 10 9 4 ETE TET I g Fingerprint sensors can be embedded in a variety of devices for user recognition purposes The three families of sensors optical solid state and ultrasound August 30 2005 BIOM 426 CCD ur CMOS FUR based gerprint sensing ridges and mum WW rimming m pobmr phumdiede anaycinbedded in glas Electrooptical fingerprint sensor August 30 2005 AFTIR 7 Distortions are compensated using software hased calibration techniques or using pre molded plastic lenses Generally good image quality and large sensing area 1 Cannot be miniaturized like other optical devices Electrooptical The ist layer is a polymer that emits light when polarized with a proper voltage The second layer is an array of photodiodes BIOM 426 12 a a livescan FTIR based optical scanner b a livescan capacitive scanner a livescan piezoelectric scanner a livescan thermal scanner offline inked impression alatent ngerprint August 30 2005 BIOM 426 13 The lines that ow in various patterns across ngerprints are called ridges and the space between ridges are Valleys Width varies from 100 pm to 300 um Fingerprint features associated with some matching algorithms fl dg e pattern global pattern matching a Emdm d e minutiae ridge ending and ridge bifurcation minutiae matching WSW 1 fwd2 1399 ndin a si attributes type Xy 3 Is Human location orientation Point or island Spur August 30 2005 BIOM 426 14 on the nest level of s 60 pm10250 um 7 mm maca 39 August30 2005 BIOM 426 15 Database 1 Nan H 39 I r Fingerprint Quality L J Image 1 Feature Scanner 39 Check Enhancement i Extraction Class icatlon Database 2 39 OOO 39 A I 39 Databases ND Onilineproeessing Fingerprint Quality Image 5 Feature Minutia Scanner Check Enhancement Extraction Clasg catmn Matching I Matching Score August 30 2005 BIOM 426 16 gray ss a1 re o lutlon 1by inch Size August 30 2005 39 5quot lavas BIOM 426 7 J ET k i EQC cl ii Contextual ltering that provides a low pass effect along the ridge direction and performs a bandpass effect in a direction orthogonal to the ridges Operations include calculation of ridge frequency and orientation eld Estimation of orientation field gradient method slit sums etc Segmentation Separation of foreground from background Binarized Image August30 2005 BIOM 426 18 adaptive thIesholding 3 graythresh 39duces ridge Width to a single pixel nectiv ity and minimizes the number of erronems bifurcations Image processing is time consuming However the results of all subsequent operations depend on the quality of image as captured and processed at this stage August 30 2005 BIOM 426 e methods orientation is found for August 30 2005 BIOM 426 20 mg 2on mega 6m imam mwutmEmEm m5 mEEmEoo Ammo utmE 956 m5 wwmzsz Emm m5 amE 35me mc conmmtou m5 ucm mmmE caumEmto lt m vm Ezmi 4 i a 13 E 4 E a m inn 3 w Bifurcation w a spur two endings on a short line is line due to noise two endings closely opposing is a broken ridge endings at the boundary is due to projection August 30 2005 BIOM 426 22 L 111611 Q2521 951 3 it s Direction 8 bits Then 100 features require 2700 hits August 30 2005 310M426 A Method 2 Align ngerprints using landmarks core and delta Core and delta can be found using Poincare index or using estimated orientation ow August 30 2005 BIOM 426 24 compare ordered vectors eg length and curvature A ridge pattern Correlation late one image over another Find the sum Sum is the highest Method 67 Perform correlation matching in frequency domain Perform 2D FF T multiply two transfermed images sum multiplied values Correlation matching is less tolerable to noise and nonlinear transformation Problems translational rotational freedom depend on landmarks August 30 2005 BIOM 426 25 August 30 2005 BIOM 426 v at a curve 2 6 b Figure 23 Examples of ngerprint images acquired with an opiicai scanner a a Quad quality ngerprini b a ngerprint left by a dry nger c a ngerprint left by a wet nger d an intrinsi caliy bad ngerprint August 30 2005 B IOM 426 27 Q f gqp m from two different A produce a high Matching marmidual fnay pmduce a low Matching 33016 2111 error 1 r Two f ig p j ts frOIll mesame mumer In August 30 2005 310M426 28 e are two types of error 39 R ratio of number of instances of s of different ngerprints found to I e n e ously match to total number Genuine ERR ratio of number of instances of of sa m ngerprint are found not to mate to total number of match attempts Marching score 1 1 August 30 2005 BIOM 426 AFIS dam kg recognition performance I 100 FAR log August 30 2005 310M 426 30 defSVilpdbaseshtn 02002 httpwwwaa stshofescomldatahtml August 30 2005 BIOM 426 on Springer 2005 anon in Networked Society g Fingerprints Proceedings of the 199713p Wan and A J Fingerprint Image Enhancement Algorithm and Performance EValuation IEEE Tans on PAM V01 20 n0 8 1998 pp 777789 6 K Karu and AK Jain Fingerprint Classification Pattern Recognition V01 29 N0 3 pp 389404 1996 7 AK Jain L Hong and R B011e Online Fingerprint Verification IEEE Trans on PAM V01 19 N0 4 pp 302314 1997 August 30 2005 BIOM 426 32 Performance Evaluation Identification Systems Identification System Recognition System TlT2 TM Templates of M Different users Query Template 1 U Igt Matcher Igt Output A list of K matching users September 20 2007 BIOM 426 Winnowing Could be Select a set of most likely candidates Human supervision 0 If dataset is too large or Additional data age demographic data etc 0 Another biometric system Biometric is too weak Database Matcher gt Matcher 2 M Users Select Klikely H E Query Template 1 September 20 2007 BIOM 426 Three Approaches Three criteria used to select a subset l Threshold based a score is computed for each pair STkI kl M Scores are compared to a threshold 7 2 Rank based returns a sorted list of K scores Kl corresponds to the best match 3 Hybrid K highest scores is returned Each score is compared to a threshold 7 eX 1 2 3 4 5 6 Rank Score 141 99 42 1 September 20 2007 BIOM 426 Open and Closed World 0 Closed world assumes that users using the system are enrolled No imposters are assumed A threshold based system may return nonmatch Which is a mistake Open world assumes that an unknown user may attempt to access the system A rank based system Will always make a mistake Involve a hybrid system Where an anti user null hypothesis contain features of all imposters September 20 2007 BIOM 426 Evaluation Criteria 39 A Cumulative Match Characteristic CMC curve plots the probability of identification against the returned 12M candidate list size See BCC 2005 a talk by Higgins Probability Rank httpWWW staticccgatechedusunresearchcrncppt September 20 2007 BIOM 426 Confusion Matrix System Decision Hand 1 Hand 2 Hand 3 Hand 4 Hand 1 16 2 25 5 2 Q 8 a Hand 2 O U at 51gt Hand 3 Hand 4 Assume 25 templates is collected from each user Total number is 100 September 20 2007 BIOM 426 Comparison of Threshold Based Verification and Identification Problems 0 Let I and T be the query and claimed biometric templates vectors of extracted features For example I and T can be vectors of hand measurements I and T can be vectors of minutiae points With all their attributes included Xy location and orientation angle 0 Let SIT be a the matching score value then Acceptance Hypothesis S T I gt 7 Rejection Hypothesis S T I lt 7 September 20 2007 BIOM 426 ThresholdBased Identification System M Users are enrolled are the templates During the matching mode for each user a score is calculated The score is compared to a threshold If the individual matches operate independently that TTTM STmI is a decision of one matcher Acceptance or Rejection as User 2 does not in uence decision of another one then we can perform the following analysis of error probability September 20 2007 BIOM 426 Acceptance or Rejection as User 1 Acceptance or Rejection as User M mlM gt sltT1I 7 lt7 gt SltT2I 7 lt 7 gt7 ST 1 M Ly Identification Correct Reject Let us assume that the individual with the query template I is not enrolled The CORRECT REJECT will occur if S T 1 lt 7 5T1lt7 Mathematically it can be written as M PrCorrect Reject H PrS I Tm lt 7 I is imposter m1 The last probability is the conditional probability that I is imposter September 20 2007 BIOM 426 Derivation of FARM and FRRM Let FARM and FARM are the conditional error rates for identification problem with M users Since matchers are independent and often similarly distributed identically distributed at the extreme we have M PrCorrect Reject H 1 FARm 1 FARm M ml 0 Here F ARm is the False Accept Rate of the mth matcher If FARm is small then we can apply the Taylor Series expansion known from calculus FARM 1 PrCorrect Reject l FAR1M z 1 l mFARl Thus FARM for identification problem is a scaled version if the FAR1 False Accept Rate of a single matcher verification FARM z M FAR1 September 20 2007 BIOM 426 Derivation of FARM and FRRM 0 A correct identification occurs when the proper candidate score is matched regardless of What happens With the others PrCorrect Identification 1 F RR 1 0 The probability of Failed Identification FRRM 1 PrCorrect Identification FRRl September 20 2007 BIOM 426 11 Biometrics Personal Identi cation BIOM 426 Biometrics Systems Instructor Natalia Schmid August 25 2005 BIOM 426 Outline I Introduction I Applications I Identification methods I Requirements to biometrics I Biometrics technology I Automatic Identification design representation feature extraction matching evaluation I Privacy Issues August 25 2005 BIOM 426 Introduction password PIN August 25 2005 BIOM 426 Identi cation associating identity With an individual Two types of identification problems 0 verification confimiing or denying person39s identity AmI who I claim I am 0 identification or recognition establishing identity Who am I Introduction A August 25 2005 Facts BIOM 426 Master Card estimated fraud at 450 million per year 1 billion dollars worth of calls are made by cellular bandwidth thieves ATM related fraud 3 billion annually 3000 illegal immigrants crossing the Mexican border each day Identi cation methods Person s identity is everything What person represents and believes Engineering approach reduce the problem to i some possession quotsomething What he hasquot or ii some knowledge quotsomething What he knowsquot Another approach reduce it to a problem of authentication based on physical characteristics physiological or behavioral Definition Biometrics are person39s identification based on hisher physiological or behavioral characteristics quotsomething that you arequot August 25 2005 BIOM 426 5 New De nition Biometrics are automated methods of recognizing a person based on a physiological or behavioral characteristic BCC2003 August 25 2005 BIOM 426 V t Requirements to biometrics universality everyone should have it uniqueness small probability that two persons are the same in terms of this characteristic permanence invariance with the time collectability can be measured quantitatively performance high identification accuracy acceptability acceptance by people circumvention how easy to fool the system by fraudulent technique August 25 2005 BIOM 426 Accepted Biometrics Accepted and studied biometrics fingerprint voice hand geometry gait ear face iris retina infrared facial and hand vein thermograms key stroke signature DNA DNA signature and fingerprint are recognized in court of law August 25 2005 BIOM 426 8 Biometric Technology Overview Fingerprints are graphical flowlike ridges Their formation depends on embryonic development Factors i genetic ii environmental Fingerprint acquisition i scanning inked impression ii lifescan Major representations image ridges minutia features derived from ridges or pores Basic approaches to identi cation i correlation based ii global ridge patterns classes iii ridge patterns iv fingerprint minutiae ridge endings and bifurcations V pore pattern August 25 2005 BIOM 426 9 Biometric Technology Overview Face one of the most acceptable biometrics Two identi cation approaches i transform eigenvalues analysis of covariance matrix orthononnal basis vectors ii attIibutebased approach geometric features Factors that in uence recognition i facial disguise ii facial expressions iii lighting conditions iv pose variation BIOM 426 August 25 2005 Biometric Technology Overview Iris is one of the most reliable biometrics Frontal images are obtained using near infrared camera 320 X 480 pixels at distance lt 1 meter Iris images are i segmented and ii encoded Twins have different iris patterns August 25 2005 BIOM 426 1 1 Biometric Technology Overview Voice print is highly accepted biometrics Used for identification over the telephone Easy to fool the system Augunzsaoos Voice is a behavioral characteristic and is not sufficiently unique large database Processing signal subdivided into a few frequency bands The most commonly used feature is cepstral feature log of PT in each band Matching stranegies hidden Markoff vector quantization etc Types of Verificatio textedependent texteindependent languageeindependent BlOM 426 12 Biometric Technology Overview Infrared Facial and Hand Vein Thermogram Human bodies radiate heat Infrared sensors acquire an image of heat distribution along the body Images thermograms Imaging methods similar to visible spectrum photographs Processing raw images are normalized with respect to heat radiating from landmark features In uncontrolled environment other sources of heat could be disturbance August 25 2005 BIOM 426 13 Biometric Technology Overview Gait is the speci c way one walks Complex spatiogtemporal behavioral characteristic Gait is not unique and does not stay invariant over time It is in uenced by distribution ofbody weight injuries involving joints or brain aging Gait features are derived from a video sequence and consists of characterization of several movements computer vision problem August 25 2005 BIOM 426 14 Biometric Technology Overview Retinal Scan Retinal vasculature is rich in structure Unique characteristic of each individual and each eye Not easy to change or replicate Image capture requires person to peep into an eyepiece and focus on a specific spot A predetermined part of retinal vasculature is imaged Requires cooperation Not accepted by public Can reveal some medical conditions as hypertension August 25 2005 BIOM 426 15 Biometric Technology Overview Signature the way person signs hisher name l i ifiiff Highly acceptable behavioral biometrics Evolves over time and depends on physical and mental conditions Easily forged Modeling the invariance and automating signature recognition process is Challenging Two approaches to signature veri cation i static geometric features strokes ii dynamic strokes and acceleration velocity trajectory August 25 2005 BIOM 426 16 Biometric Technology Overview Hand and finger geometry is used for access control 50 of market System captures frontal and side Views of palm Measurements length and width of fingers various distances The representation requirements are only 9 bytes Hand geometry is not unique but highly acceptable August 25 2005 BIOM 426 17 Comparison of Biometrics Technologies From Biometrics Personal Identification in Networked Society p 16 August 25 2005 BIOM 426 18 Automatic Identification History 0 Prehistoric Chinese used thumb print for identification 0 Alphonse Bertillon s System of Anthropometric Identification 1882 is based on bodily measuments physical description and photographs 0 Henry s fingeprint classification system 1880 classifies in gt 100 classes Sets of rules are developed for i matching of biometrics ii searching databases Automatic identification is due to inexpensive computer resources advances in computer vision pattern recognition and image understanding August 25 2005 BIOM 426 19 Applications I Civil applications 0 Banking electronic funds transfer ATM security Internet commerce credit card transactions Physical access control airport 0 Information system security access to databases Via login 0 Customs and immigration identification based on hand geometry 0 Voterdriver registration 0 Telecommunications cellular bandwidth access control August 25 2005 BIOM 426 20 Example a typical human subject may wr this picture to belong to Al Gore and Presidenl Bill Clinton on closer inspection one could recognize that both the laces in the picture identically show Bill Clinton39s lacial features an the crown of hair on one of the laces has been digitally manipulated to appear similar to that at Al Gore August 25 2005 BIOM 426 Automatic Identi cation Design Enrollment Biometric Reader gt Feature Extractor C Feature Extractor Biometric Reader Feature Matcher 39 V Identification system operates in two modes i enrollment mode and ii identification August 25 2005 BIOM 426 22 Automatic Identification Enrollment mode biometric measurement is captured information from raw data extracted feature person information is stored ID is issued for verification Identification mode biometric is sensed livescan features are extracted from the raw data match is performed search of the database In verification mode person presents ID Then system performs match only against one template in the database August 25 2005 BIOM 426 23 Recognition System Object 11 D Acquisition I Feature Extraction Training Testing 1 Object Outcome Acquisition Feature Extraction Matching Architecture of a typical pattern recognition system see A K Jain et 211 p 22 August 25 2005 BIOM 426 24 Design Issues Given the speed accuracy and cost specifications 1 How to collect the input data 3D 2D multiple Views high or low resolution 2 Internal representation features for automatic feature extraction 3 How to extract features Algorithms etc 4 How to select the quotmatchingquot metric Measurements are made in specific space 5 How to implement it 6 Organization of database 7 Effective methods for searching a template in the database binning etc August 25 2005 BIOM 426 25 Acquisition Quality of collected data determines perfonnance of the entire system Associated tasks i quality assessment ii segmentation separation of the data into foreground and background Research efforts i richer data 3D color etc ii metrics for assessment quality of measurements iii realistic models Solutions enhancement August 25 2005 BIOM 426 26 Representation Which machinereadable representation captures the invariant and discriminatory information in the data Determine features st invariant for the same individual intraclass variation maximally distinct for different individuals interclass More distincive features offer more reliable identification Representation has to be storage space ef cient smart card 2 Kbytes Representation depends on biometrics August 25 2005 BIOM 426 27 Feature extraction Given raw data automatically extracting the given representation is dif cult problem Example manual fingerprint system uses about a dozen of features For automatic system many of them are not easy to reliably detect Feature extraction procedures are typically designed in ad hoc manner inefficient when measurements are noisy 0 Determining effective models for features will help to reliably extract them esp in noisy situations August 25 2005 BIOM 426 28 Matching Similarity metric should be robust against noisc structural and statistical variations aging and artifacts of fcaturc extraction module Examplc signaturc hard to dcfinc thc ground truth Pcrformancc is dctcrmincd by i rcprcscntation and ii similarity mctric Tradeoff bcttcr cnginccring dcsign vs more compch matchcr Examplc ngcrprint variations in fcaturcs and rigid matchcr vs Flexible matchcr August 25 2005 BIOM 426 29 Matching Fingerprint Sources of distortion and noise i inconsistent contact 3Dto2D ii nonuniform contact due to dryness of skin sweat dirt humidity in the air etc iii irreproducible contact injuries to the finger iV feature extraction artifacts measurement error V sensing itself adds noise August 25 2005 BIOM 426 30 Evaluation August 25 2005 An enduser questions i Does the system makes an accurate identification ii Is the system sufficiently fast iii What is the cost of the system Because of noise distortions and limited information no metric is adequate for reliable identification Decisions genuine individual imp osteri BIOM 426 31 Evaluation Imam dum buuon gimme Thrath It August 25 2005 Matching score m I Four types of outcomes 1 Genuine individual is accepted true 2 Genuine individual is rejected error 3 Imposter is rejected true 4 Imposter is accepted error FAR false acceptance rate FRR false rejection rate EER equal error rate Given a database performance is a RV and only can be estimated BIOM 426 32 Evaluation August 25 2005 Measure of performance ROC receiver operating curve Con dence Intervals Forensic Appljcmjnns an Applications 3 V a 2 High Securil Access Application False NoneMatch Rate FNMR BIOM 426 33 Useful Links httpWWWbi0metrics0rg publications and periodicals C O N S O R T I U M research and databases meetings and events httpwwwitlnistg0Vdiv895bi0metrics m MICHIGAN STATE 0 httpbiometricscsemsuedu u N I v E R s T v httpWWWtechpurdueeduitresourcesbiometrics August 25 2005 BIOM 426 34 Hand Geometry BIOM 426 Instructor Natalia A Schmid Outline 0 Motivation 0 Acquisition systems 0 Enrollment 39 Verification quot Feature Extraction 3 WVU 2 References Not much open literature is available Much information is in the form of 0 Patents for example Miller 7l Sidlauskas 88 Applicationoriented descriptions see IEEE Spectrum no 2 1994 A prototype system described by Jain et al 4 in 2003 394 Web pages of Recognition Systems and Biometch for example BFC be found in recent 20052007 Proc of ICASSP BCC etc WVU Motivation Attractive points Almost all of the working population has hand Exception processing can be easily engineered Measurements are easily collectable Nonintrusive compared to iris or retinal scan Simple method of sensing putations are easy gt system is easy to build integrate with other biometrics as fingerprint WVU 4 Evolution First devices electromechanica Identimation measures length of 4 used in nuclear weapon industry was retired in 1987 In the mid1980 s Sidlauskas developed electronic 3D pro le identi cation apparatus capacity 20000 users processing time is 12 sec 1994 weight is 45 kg 1994 9 byte representation Acquisition systems Features finger length Width thiclm curvatures and relative location of features Scanners use CCD camera infrared LEDs mirrors and re ectors No surface details no color no fingerprint lines are recorded Top and side views WVU 6 Acquisition systems Scanners use Camera 9 Optical path approx IT quot I between camera and platen Camera Dimensions 812 inches square by 10 inches in height 1139 28cm 539 125m l Platen a quot Helen I Z x v Scanner takes 96 measurements Direct Optical Path Folded Optical Path Microprocessor converts mer optics 9byte template Enrollment Dun39ng enrollment pins pegs help user to position hisher hand user places hisher hand 3 5 times scanner averages measurements and stores in the database Quality of enrollment affects FRR Factors platen heights training for example landing an airplane scenario Template averaging updating template after user is veii ed WVU 8 Veri cation User types PIN key pad Places hand on the platen Scanner takes measurements extracts features compares previous template with the input template ex Total absolute distance Euclidean distance etc generates a similarity score can be compared with a threshold for veri cation I The FRIUFAR was 01 manufacture s threshold 199 1 I A single trial reject rate is Characterized by 02 quot39 100000 events are tested WVU 10 Feature Extraction Typical image blackaand W Features nger length Width thickness curvatures and relative location of features Feature Extraction An example feature set for hand geomeuy 4 WVU 12 Metrics Euclidean distance d dE 2611 ri2 lt8E V121 Absolute distance d dA ZIQiri lt8A i1 Example User 2 71 63 70 61 74 56 56 52 281 362 268 278 243 136 er 2 169 63 74 62 73 57 57 55 276 366 259 282 245 141 15 55 56 63 53 60 47 48 47 249 303 258 268 241 152 Usfer2 141421 dA User2User2 42 758222 dAUser2 User15 203 WVU 13 Market Access Control Used to access Health clubs Day care centers Laboratories Prisons etc Time amp Attendance Application ranges from coal mines to clean rooms Personal Identification and Toronto airports Food Services systems at the University of quot wwwrecogsyscorn WVU 14 Applications 70000 HandReaders are installed throughout the world The 1996 Olympic Games used HandReaders to protect access to Oly r n 65000 people were enrolled 1 million transactions were handled over 28 Since 1991 at San Francisco Airport HandReaders produced more than 100 million verifications 180 doors and 18000 employees In the United Kingdom Her majesty s Prisons rely on the HandReaders for prisoner and Visitor tracing 1 Colleges eX University of Georgia use HandReaders for oncampus meal programs eguard access to dormitories and protect their computer centers WVU 15 1 Privacy Issues Hand geometry is used to verify identity Templates cannot be reverse engineered to identify users 2 Operation by Disabled People Hand scanners can be used for scanning left hand palm up gnaw be enabled for blind persons to use WVU 16 Positives and Negatives 855 Hand pmr cums c There seems ID exist a certain amount of public acceptance for the band bio metric because it is already used at Disney World INSPASS The United States Immigration and NamIaJization Service Passenger Accelerated Service Sys tem 90 and at vau39otts universities for veri cation in eg the University of Georgia 0 It is claimed that band geometry measuringis an easy Limityourselfquot npera lion There exists at least one scenario evaluation of hand geometry as a biomet ric 131 which shows that hand is a good biometric for veri cation But little is known about en39or rates for hand geometry recognition a Hand geometry being a relatively weak biometric can be used for veri ca misuse 9f the data foridenti cation As with ngerprint hand geometry is measured when a subject presses the L39 39 39 L L L J quot Such contact may be cause for some public hygiene concerns Hand 5 m the ne l Therefore apler 7 39 for pure identi I IIUL my minimquot cation Mareoven there exists only one scenario evaluation of blind geomen y as a 0 metric There is debate as to whether hand geometry is uuly a biometric ur Again some people do not have hands or measurable ngers for various rea sons Available Databases 1 University of Bologna database httpbiascsruniboitresearcl biolabbiotreehtml 2 MSU hand geometry database project at WVU multimodal biometrics WVU 18 References 1 Biometrics Personal Identification in Networked Society A Jame L V 2 Hand give me five by D Sidlauskas in Vital signs of identity IEEE Specrtam February 1994 pp24 25 3 D P Sidlauskas 3D hand profile identi cation apparatus US Patent No 4736203 1988 4 A K Jain A Ross and Sh Pankanti A Prototype Hand Geometrybased Verification System Proc of 2nd Int l Conf on Audio and Videobased Biometric Person Authentication Washington DC pp 166171 March 2224 7 Bolle et a1 Guide to Biometrics Springer New York 2004 pp 45 47 wrecogsyscom 39 39 quot Mbiometch two finger veri cation WVU 19 Preprocessing IMAGE 100 200 300 400 500 C 100 200 300 400 500 600 ENARZEDIMAGETQ5 100 200 300 400 500 100 200 300 400 500 600 VV JII HMAGEI STIXBRAM 15000 10000 5000 0 0 50 100 700 2390 250 100 200 300 400 500 x 100 200 300 400 500 600 20 Feature Extraction MASK FOR COMPUTING HAND WIDTH 100 100 Note shown is the reversed mask 200 200 300 300 4oo 400 500 500 100 200 300 400 500 600 100 200 300 400 500 600 SUPERPOSITION OF MASK and BINARY IMAGE 1 00 200 300 400 100 200 300 400 500 600 WVU 21 Feature Extraction Matlab Code gtgt IM imread filename tiff read tifffile gtgt BW im2bWIMO75 binarization gtgt sizeIM provides info about image size gtgt mask zeros512640 creates image filled with zeros gt gt mask 26019OI450 1 fills line with ones Feature 1BWmask extracts feature ength ndFeature gt 0 finds feature length in pixels WVU 22 BIOM 426 Instructor Natalia A Schmid November 13 2007 WVU Procedure I Illumination Compensation I Align images I Convert into vector form vectorreshapematrixNA2l I Calculate the mean image I Form scatter matrix I Find eigenvalues and eigenvectors I Eigenvectors are the axes of the transformed space I Project images onto the new space I Use Euclidean distance between weights to find the distance between two images November 13 2007 WVU Equations Training images 11 12 IM M 15 k1 The mean Image 7 i M M 7 7 Mk 7 mac 7 If k1 The scatter matrix c Use eigenvalue decomposition eig in matlab Project each new image in the space of eigenvectors M 7 m z w u 1 where wi u 1m7 1 11 Distance between two images dwkwl 241WZUC 7wlZ 11 November 13 2007 WVU Iris Images 64x64 IRIS 1 IRIS 2 AVERAGE IRIS November 13 2007 WVU Eigenlrises Reconstruction Eigenvalues EigVa 10e008 474025426761636 174354598828577 104998230910908 101508593967614 080637230813956 070820383258604 043921571631076 035745729185653 000000000000001 November 13 2007 WVU Comparison REszTRucTED IRIS 1 ORIOGINAL IRIS 1 v I l November 13 2007 WVU Hand Geometry BIOM 426 Instructor Natalia A Schmid Outline Motivation Acquisition systems Enrollment 39 Verification Feature Extraction Applications 2 References Not much open literature is available Much information is in the form of Patents for example Miller 7 l Sidlauskas 88 Applicationoriented descriptions see IEEE Spectrum no 2 1994 39 Exclusion prototype system described by Jain et al 4 Web pages of Recognition Systems and Biometch s for example BFC WVU 3 Motivation Attractive points Almost all of the working population has hand Exception processing can be easily engineered 0 Measurements are easily collectable Nonintrusive compared to iris 0r retinal scan Simple method of sensing iomputations are easy gt system is easy to build WVU 4 Evolution Y i7 geometry systems es of the hand L i WVU First devices electromechanical Identimation measures length of 4 used in nuclear weapon indus was retired in 1987 In the mid1980 s Sidlauskas developed electronic 3D pro le identi cation apparatus capacity 20000 users processing time is 12 sec 1994 weight is 45 kg 1994 9 byte representation Acquisition systems Features finger length Width curvatures and relative 10cati0 n 9 features Scanners use CCD camera infrared LEDs mirrors and re ectors No surface details no color no fingerprint lines are recorded Top and side Views Acquisition systems El Camera r 1139 28cm I L Direct Optical Path Folded Optical Path negropticsr Scanners use Optical path approx 1T 39 between camera and platen Dimensions 812 inches square by 10 inches height Scanner takes 96 measurements Microprocessor converts 9byte template Enrollment Dun39ng enrollment pins pegs help user to position hisher hand user places hisher hand 3 5 times scanner averages measurements and stores in the database Quality of enrollment affects FRR Factors platen heights training for example landing an airplane scenario Template averaging updating template after user is veii ed WVU 8 Veri cation User types PIN key pad Places hand on the platen Scanner takes measurements extracts features compares previous template with the input template ex Total absolute distance Euclidean distance etc generates a similarity score can be compared with a threshold for veri cation I The FRIUFAR was 01 manufacture s threshold in 199 1 I A single trial reject rate is Characterized by 02 quot 100000 events are tested WVU 10 Feature Extraction Typical image black and lel Features nger length width thickness curvatures and relative location of features Feature Extraction mem D c p nn Disc plion mm of palm m Lhe base fuln ngers Thich ss of ugers at second phalanx Thickn ss of nge at Khiid phalanx An example feature set for hand geomeuy 4 Wu 12 Metrics Euclidean distance 61E Absolute distance dA d 239 qt ri K 8A 3921 Example User 2 71 63 70 61 74 56 56 52 281 362 268 278 243 136 USer 2 69 63 74 62 73 57 57 55 276 366 259 282 245 141 1515 55 56 63 53 60 47 48 47 249 303 258 268 241 152 1 User2 2141421 d A User2User2 42 758222 dA User2 User15 203 WVU 13 Market Access Control Used to access Health clubs Day care centers Laboratories Prisonsz etc Time amp Attendance Application ranges from coal mines to clean rooms Personal Identification Applications 70000 HandReaders are installed throughout the world The 1996 Olympic Games used HandReaders to protect access to 01 p 65000 people were enrolled 1 million transactions were handled over 28 da Since 1991 at San Francisco Airport HandReaders produced more than 100 million verifications 180 doors and 18000 employees In the United Kingdom Her majesty s Prisons rely on the HandReaders for prisoner and Visitor tracing Colleges eX University of Georgia use HandReaders for oncampus meal programs ard access to dormitories and protect their computer centers WVU 15 1 Privacy Issues Hand geometry is used to verify identity Templates cannot be reverse engineered to identify users 2 Operation by Disabled People Hand scanners can be used for scanning left hand palm up Could be enabled for blind persons to use WVU l6 Positives and Negatives 855 Hand pmr Thete scents to exist a certain amount of public acceptance for the hand bio metric because it is already used at Disney World NSPASS The United States meigraLion and Naturalization Service Passenger Accelerated Service Sys tem 90 and at various universities for veri cation in eg the University of Georgia I Itis claimed that hand geometry measuring is an easy doit yourself opera den 0 There exists a least one scenario evaluation of hand geometry as a biomete ric 13 which shows that hand is a good biometric for veri cation But little is known about error rates for hand geometry recognition 0 Hand geometry being a relatively weak biometric can be used for veri ca misuse of the data for identi cation cans 0 As with ngerprint hand geometry is measured when a subject presses the L 39 L L L quot quotV Such Contact may be cause for some public hygiene concerns and J r L V Therefore quot J Chapter 7 L 39 39 39 um m scum for pure identi cation 0 Moreover quot 39 39 39 39L E hiu metric There is debate as to whether hand geometry is truly a biometric ur Again some people do not have hands or measurable ngers or various rea SO S Available Databases 1 University of Bologna database httpbiascsruniboitresearchbiolabbiotreehtml 2 MSU hand geometry database project at WVU multirnodal biometrics References 1 Biometrics Personal Identification in Networked Society A Jain e 2 Hand give me five by D Sidlauskas in Vital signs of identity IEEE Specrtam February 1994 pp24 25 3 D P Sidlauskas 3D hand profile identi cation apparatus US Patent No 4736203 1988 4 A K Jain A Ross and Sh Pankanti A Prototype Hand Geometrybased Verification System Proc of 2nd Int l Conf on Audio and Videobased Biometric Person Authentication Washington DC pp 166171 March 2224 4 Bolle et a1 Guide to Biometrics Spring f NEW York 2004 PP 4547 recogsys com 1am quotbiometch two finger veri cation WVU 19 Prieprocessing IMAGE 100 200 300 400 500 100 200 300 400 500 ENAREEDIMAGET05 100 200 300 400 500 100 200 300 400 500 600 WVU IMAGEHBTOGRAM 15000 10000 5000 100 2390 50 250 HNAREEDIMAGETQ75 100 200 300 400 500 100 200 300 400 500 600 20 Feature Extraction REVERSED IMAGE MASK FOR COMPUTING HAND WIDTH 100 100 Note shown is the reversed mask 200 200 300 300 400 400 500 500 100 200 300 400 500 600 100 200 300 400 500 600 SUPERPOSITION OF MASK and BINARY IMAGE 21 Feature Extraction Matlab Code gtgt IM imread filename tiff read tifffile gtgt BW im2bWIMO75 binarization gtgt sizeIM provides info about image size gtgt mask zeros512640 creates image filled with zeros gtgt mask 26019OI450 l fills line with ones Feature lBWmask extracts feature 39 lengthQ ndFeature gt 0 finds feature length in pixels WVU 22 Face Recognition BIOM 426 Instructor Natalia A Schmid Imaging MOdalitieS Processing Methods October 307 2007 WVU l Applications Law enforcement mug shot identification Verification for personal identification driver s licenses passports etc Surveillance of crowd behavior Security applications US Visit program Face is a passive biometric Does not require cooperation E 1 banning nu quotW 105 515 R I I October 30 2007 WVU 2 I Mug shot Data Collection Environment well controlled frontal profile photographs uniformbackground gt Mugshot identicalposes similar illumination Canonical faces cropped size and position normalized minimum backgroun October 30 2007 WVU 3 Data Collection Face quot39 in more than 1 face can appear lighting conditions vary facial expressions different scale position orientation facial hair makeup glasses occlusion 11 Face recognition is a complex problem Detect face If multiple estimate location and size October 30 2007 WVU 4 Approaches Face Recognition Representation and Classification Criteria Variations Sensing 2 D intensity image color image infrared image 3 D range modality image combination of them Viewing angle Frontal Views profile Views general Views or a combination of them Temporal Static images time Varying image sequence may facilitate component face tracking expression identification etc l 39 J J rules statistical decision rules tools neural networks genetic algorithms etc October 30 2007 WW 5 Imaging Modalities Optical Camera color blackwhite Infrared Camera Laser radar new technology Image infrared image and Video sequence October 30 2007 W V 6 Face Biometric 39 Macro elements the mouth nose eyes cheekbones chin lips forehead and ears 39 Micro elements distances between the macro features or reference features and the size of features 39 Heat radiation Features derived from face images geometric and statistical Human can easily detect and identify individual s face in a scene Designing an automated system is hard October 30 2007 WVU 7 BlockDiagram Input Image Face Cropping Feature Face Detection Normalization Extraction Recognition Fatial landmark Features have to have Detection ex high discriminating Centers of the eyes power Cropping and Normalization Tasks facial region extraction minimizes in uence of other factors not related to face spatial normalization which aligns the centers of the eyes and xes the number of pixels between the eyes via rotation and scaling information Intensity normalization converts image into a vector and normalizes it to be zero mean variance one vector E October 30 2007 WVU 8 Performance Evaluation Factor I Measured using standard databases and objective performance statistics I The face recognition vendor test F RVT I The FRVT 02 reported 1 under normal indoor illumination the stateoftheart face recognition system achieve 90 verification rate at a false accept verification rate at a false accept rate 1 3 the 3D morphable models techniques improve nonfrontal face recognition I Illumination and pose are still challenging areas I The FRGC06 indoor illumination 95 verification rate at FAR01 2 under outdoor conditions the best vendor can get 50 October 30 2007 WVU 9 Face Detection Earlier methods correlation or template matching techniques matched filters subspace methods Recent methods are data driven learning based techniques Statistical modeling estimation of face 7 nonface patterns then apply pattern classifier Neural networkbased learning learn to discriminate face 7 non face patterns using training samples and the network structure Support Vector Machines Markov Random Field Color based detection October 30 2007 WVU 10 Face Recognition Requires I Lowdimensional representation to achieve data compression Usually starts from dimensionality reduction high dimensional space is almost empty I Enhanced discrimination abilities Achieving high separability between patterns October 30 2007 WVU ll Approaches Manually de ned features Geometric features such as distance and angles between geometric points eye corners mouth extremities nostrils chin top etci For profiles a set of characteristic points Locations of points can be extracted automaticallyi Problems Automatic extraction is not reliable The number of features is small The reliability of each feature is difficult to estimate October 30 2007 WVU 12 Approaches Automatically derived features Nonstatistical Methods Neural networks Statistical Methods Eigenfaces nonlinear deformations October 30 2007 WVU Representation and Recognition Techniques Representation methods PCA ICA Shape and texture Gabor wavelets Recognition methods BayesMAP FLDLDA Graph Matching October 30 2007 WVU Local Feature Analysis Based on macro features 1 Separation of face from background 2 Reference points are detected used the change in shading around features 3 Anchor points are tied in triangles 4i Angles are measured from each of anchor points 5 672bit template is generated 6 Change in lighting conditions or orientation leads to new templates 7 Live scan undergoes the same processing High percentage score results in match October 30 2007 genfaces Appearancebased approach Eigenface one s own face lnput 2D gray scale image Image is a highdimensional vector each pixel is a component Each image is decomposed in terms of other basis vectors eigenvectors N fZwkek k l Where N is the image dimension ek is the kth eigenfacei Template consists of weights Wk i The features of input image and database templates are compared using nearest neighbor rule lNN Euclidean distance October 30 2007 W V 16 Neural Network Detection Training Set N face images with identified macro features are fed into network o ner random images Other faces are entered with no identified macro features The unidentified faces are reentered into system with identified features The parts of ANN a face detection and framing b ANN input level 0 Receptive fields d Hidden units e Output Neural network October 30 2007 WVU l7 Neural Network Face Detection and Framing ace is separated from its background framed and transformed into appropriate size ANN input level face image is converted into pixels to correspond to array of input neurons Receptive fields the mapping is chosen to re ect general characteristics of face Hidden units have a onetoone neuronreceptive field relationship Hidden units determine if appropriate feature was detected Output a single output neuron that indicates positive or negative face match October 30 2007 WVU l8 Representation and Recognition 39 Principle Component Analysis PCA derives an orthogonal projection basis that leads directly to dimensionality reduction and feature selection Eigenfaces eigenvectors related to the largest eigenvalues PCA is optimal criterion for dimensionality reduction but does not always provide good discrimination Kirby90 Turk91 39 Solution Integrate PCA with the Bayes classifier Probabilistic Reasoning Models LinO 39 Shape and texture applies a twostage procedure Coding starts by marking important internal and boundary pointsi The points are aligned using translation scaling and rotation Calculate the average of control points 7 defines the mean shape Triangulate the marked face and warp each face into the mean shape The first stage yields the shape the second stage is related to texture Beymer95 Cootes98 Craw92 Edwards98 Lanitis97 Linl October 30 2007 WVU 19 Representation and Recognition 39 Classifiers used Mahalanobis distance classifier Lanitis97 from shape and texture Enhanced Fisher classifier Linl 39 The Gabor wavelets have desirable characteristics of spatial locality and orientation selectivityi Computation of Gabor features and use of exible graph matching Lades93 Gabor filters were used for twoclass categorization of gender race and facial expression LyonsOO Steps include registration of grid with the face using elastic matching and characterization of extracted points using L i GaborFisher classifier robust to illumination and facial expression variability Liu02 October 30 2007 WVU 20 10 Representation and Recognition The Bayes classifier is optimal when the probability distributions of data are owni The PRM method LinO integrates PCA and Bayes classifieri Fisher Linear Discriminant FLD LDA separates the mean of different classes as far as possible and compresses the same class data as much as possible Independent Component Analysis ICA a technique for blind source separation ICA analysis when carried out in the properly compressed and whitened space performs better than the Eigenface and Fisherface methods When augmented by additional criteria such as the MAP rule its performance degrades Graph matching achieves invariance to affine transformations or localized facial expressions Graph nodes have to be manually defined to find the corresponding nodes in the different graphs October 30 2007 WVU 21 KernelBased Methods Kernel PCA kernel FLD and SVM overcome the limitations of the linear approaches by nonlinearly mapping the input space into a highdimensional feature space Ti Cover Nonlinearly separable patterns in an input space are linearl separable With high probability if the input space is transformed nonlinearly to a highdimensional feature space 3D methods provide potential solutions to pose invariant recognition 3D models are often derived from laserscanned 3D heads range data or reconstructed using shape from shadingl Hsu and Jain HsuOl described the method of modeling 3D faces based on a triangular mesh model and individual facial measurements A potential solution to face recognition With variations in illumination pose and facial expressions Method by Zhao and Chellappa ZhaoOl applies a 3D model to synthesize a prototype image from a given image acquired under different lighting and viewing conditions October 30 2007 WVU 22 ll Face Pros and Cons Pros Used for manual inspection driver license passport Wide public acceptance for this biometric identifier The least intrusive from sampling point of view requiring no contact Face recognition can be used at least in theory for screening of unwanted individuals in a crowd in real time It is a good biometric identifier for smallscale verification applications October 30 2007 WVU Cons For robust identification face needs to be well lighted by controlled source Currently it performs poor in outdoor protoco Disguise is an obvious circumvention method Disguised person is not identifie There is some criminal association with face identifiers since it has been used by law enforcement agencies mugshots Privacy concenis Face Databases 39 The Olivetti ORL now ATampT database 40 subjects ten 92x112 pixels with a variety of lighting and facial expressions httpwwwukresearchattcomfacialrecognition 0 FERET 14126 images that includes 1199 subjects and 356 duplicate sets httpwwwdodcounterdrugcomfacialrecognition 0 FRVT 2002 120000 faces includes video of faces httpwwwfrvtorgFRVT2002defau1thtm 0 NIST 18 Mugshot Identi cation Database 32A8 mugshot images front images and pro les 500 dpi httpwwwnistgovsrdnistsd18htm 0 The MIT database 16 subjects 27 images per subject with varying illumination scale and head orientation ftpwhitechape1mediamitedupubimages 0 The Yale database 5850 imagesof 10 subjects each imaged under 576 viewng coditions 9 poses and 64 illumination conditions Size 640x480 256 grey levels httpcvcyaleeduprojectsyalefacesByalefacesBhtml 0 The Purdue database 4000 color images from 126 subjects imaged with different expressions illumination conditions and occlusion httprvll ecnpurdueedualeixaleixfaceDBhtml October 30 2007 WVU 24 12 References l Biametrics Persanal Identi catian in Netwarked Susiety A Jain et al Edt Ch 3 2 J Zhang Y Yan andM Lades Face Recognition Eigenface Elastic Matching and Neural Netsquot Praceealing afthe IEEE vol 85 no 9 pp 1423 7 1435 1997 3 R Chellappa C L Wilson S Sirohey Human and Machine Recognition of Faces A Survey Praceedings afthe IEEE vol 83 no 5 1995 pp 705 740 4 W Zhao and P J Phillips Face Recognition A Literature Survey NIST Techn Repart 2000 5 P N Belhurneur J P Hespanha and D J Kriegrnan Eigenfaces vs Fisherfaces Recognition Using Class Speci c Linear Projection IEEE Trans an Pattern Analysis and Machine Intelligence Vol 19 no7 pp 711 7 720 1997 6 M Kirby and L Sirovich Application of KarhunenLoeve Procedure for the Characterization of Human Face vol 12 no 1 pp 103 7 108 1990 October 30 2007 WVU 25 l3
Are you sure you want to buy this material for
You're already Subscribed!
Looks like you've already subscribed to StudySoup, you won't need to purchase another subscription to get this material. To access this material simply click 'View Full Document'