Template Aging Phenmenn in Iris Recognition

Received February 26, 2013, accepted May 3, 2013, date of publication May 16, 2013, date of current version May 21, 2013.

Digital Object Identifier 10.1109/ACCESS.2013.2262916

Template Aging Phenomenon in Iris RecognitionSAMUEL P. FENKER, ESTEFAN ORTIZ, AND KEVIN W. BOWYER (Fellow, IEEE)Department of Computer Science and Engineering, University of Notre Dame, Notre Dame, IN 46556, USA

Corresponding author: K. W. Bowyer ([email protected])

ABSTRACT Biometric template aging is defined as an increase in recognition error rate with increasedtime since enrollment. It is believed that template aging does not occur for iris recognition. Several researchgroups, however, have recently reported experimental results showing that iris template aging does occur.This template aging effect manifests as a shift in the authentic distribution, resulting in an increased false non-match rate. Analyzing results from a three-year time-lapse data set, we find ∼ 150% increase in the falsenon-match rate at a decision threshold representing a one in two million false match rate. We summarizeseveral known elements of eye aging that could contribute to template aging, including age-related changein pupil dilation. Finally, we discuss various steps that can control the template aging effect in typical identityverification applications.

INDEX TERMS Biometrics, iris recognition, error probability, false non-match rate, template aging.

I. INTRODUCTIONIris recognition technology began in the early 1990s with thework of John Daugman [1], [2], and has progressed quickly.The United Arab Emirates has used iris recognition success-fully in border control for the past decade [3]. India is usingiris recognition as part of its Aadhaar, or ‘‘Unique ID’’, pro-gram to give a unique ID number to each of about 1.2 billioncitizens. In 2013, the number of persons enrolled in India’sAadhaar program will surpass the size of the population ofthe United States [4]. In addition to these large governmentapplications, there are successful uses of iris recognition inbanks, hospitals, schools and other applications. Along withthe expanding range of applications for iris recognition, orperhaps because of it, iris recognition is also an active andexpanding research area [5], [6].

This paper is concerned with template aging in iris recog-nition. The ISO standard for biometric performance testingdefines template aging as follows – ‘‘Longer time intervalsgenerally make it more difficult to match samples to tem-plates due to the phenomenon known as ‘template aging’.This refers to the increase in error rates caused by time-related changes in the biometric pattern, its presentationand the sensor’’ [7]. Template aging receives substantialresearch attention in the face recognition research community(e.g., [8], [9]) and is documented in fingerprint matching [10].However, template aging has, until recently, been assumed tonot occur for iris recognition.

Section II reviews the literature related to template agingin iris recognition. Section III details our own recent

experimental results on iris template aging. Section IV out-lines how various elements of normal eye aging could con-tribute to an iris template aging effect. Section V describesnon-traditional enrollment schemes that could be imple-mented to reduce or eliminate the template aging effect.Section VI summarizes some of the issues arising in connec-tion with research on iris template aging.

II. LITERATURE REVIEWIn their 1987 patent, Flom and Safir [11] asserted that, ‘‘. . . thesignificant features of the iris remain extremely stable and donot change over a period of many years.’’ However, they alsoacknowledged that the iris does change over time, and that re-enrollment might be needed [11] – ‘‘Even features which dodevelop over time . . . usually develop rather slowly, so that anupdated iris imagewill permit ID for a substantial period . . . ’’.Daugman’s 1994 patent [1] asserted that the iris texture

is unchanging for life – ‘‘The iris of every human eyehas a unique texture of high complexity, which proves tobe essentially immutable over a person’s life.’’ Elsewhere,Daugman indicates that this conclusion was supportedthrough subjective evaluation of a dataset of ophthalmolo-gist’s images [2] – ‘‘The clinical database of iris images madeavailable to this author from ophthalmologists’ photographsspanning a 25 year period did not reveal any noticeablechanges in iris patterns for individual subjects’’.The view that iris texture is ‘‘essentially immutable over

a person’s life’’ became the prevailing view in the irisrecognition community. The assumption is often repeated in

266 2169-3536/$31.00 2013 IEEE VOLUME 1, 2013

S. P. FENKER et al.: Template Aging Phenomenon

the research literature, e.g., ‘‘. . . the iris is highly stable overa person’s lifetime . . . ’’ [12], ‘‘[the iris is] essentially stableover a lifetime’’ [13], and ‘‘the iris is highly stable over aperson’s lifetime’’ [14]. Popular references such asWikipediaoften expressed it as – ‘‘a key advantage . . . is . . . templatelongevity, as, barring trauma, a single enrollment can last alifetime’’ [15]. The idea that ‘‘a single enrollment can last alifetime’’ is challenged or disproved by recent research on iristemplate aging.

Tome-Gonzalez et al [16] compare matches betweenimages acquired in the same session with matches betweenimages acquired with one to four weeks of time lapse. Theyuse two different datasets, each acquired across four weeklysessions, using an LG 3000 sensor. They useMasek’s matcher[17], which does not have state-of-the-art performance. At afalse match rate (FMR) of 0.01%, they report a false non-match rate (FNMR) of 8.5% to 11.3% for within-sessionmatches, versus a FNMR of 22.4% to 25.8% for across-session matches. This is a same-session versus across-sessioncomparison, rather than a longitudinal template aging study,but it shows that iris match quality can depend on factors thatchange between acquisition sessions.

Baker et al [18] report on a study involving 26 irises(13 persons) with images acquired over 2004–2008 using anLG 2200 sensor. They used a version of the IrisBEE matcherthat was distributed in the Iris Challenge Evaluation [19].This matcher does not have state-of-the-art performance.They compared the authentic and impostor distributions forshort-term and long-term matches. Short-term matches werebetween two images taken in the same academic semesterbut not on the same day, in order to avoid same-session bias.Long-term matches were between an image taken in springof 2004 and one taken in spring of 2008. They found that theauthentic distribution for long-term matches shifted so as toincrease the FNMR. They reported that, ‘‘at a false accept rateof 0.01%, the false reject rate increases by 75% for long-time-lapse’’ [18]. They later expanded this study to include a largerdataset and additional matchers [46].

Rankin et al [20] studied variation in iris appearance overthree sessions at three-month intervals. They used visible-light illumination rather than near-infrared. We are not awareof any commercial iris recognition system that uses visiblelight. The quality of their images and iris matching softwarewas such that ‘‘Recognition failure was detected in 21% ofintra-class comparisons cases overall, taken at both threeand six month intervals’’ [20]. The combination of shorttime lapse, use of visible-light images, and such poor overallrecognition accuracy makes it impossible to draw any reliableconclusions related to iris template aging.

Sazonova et al [21] studied iris template aging usinga dataset acquired at Clarkson University with an OkiIrisPass-H sensor over the period 2005–2007 for 46 subjects.They analyzed results from the Masek [17] and VeriEye [22]matchers. They also looked at results based on using imagequality metrics to select a subset of ‘‘good quality’’ images.They found a template aging effect regardless of matcher

or image quality subset – ‘‘All verification rates decreasewith the increase of the time lapse’’ [21]. Results for thegood quality subset of the overall image dataset showeda more pronounced template aging effect (see Table 1 of[21]), ‘‘which suggests that while quality varies over time,it may be masking the effect rather than contributing to theeffect’’ [21].In our own earlier work, using data for 86 irises (43 per-

sons), we compared short-term matches, between two imagestaken on different days within one semester, with long-termmatches, between images taken in spring 2008 and spring2010 using an LG 4000 sensor. We found that the FNMRincreased for long-term matches. For the IrisBEE matcher,we found that ‘‘the increase in false reject rate ranges from157% at a threshold of 0.28 to 305% at 0.34’’ [23]. For theVeriEye matcher [22], we found that ‘‘The observed falsereject rate increases from short to long time-lapse by 195% ata threshold of 30 and up to 457% at a threshold of 100’’ [23].Later results [24] expanded on this study to include a thirdyear of time lapse, by presenting results for one, two and threeyears, and by using a bootstrap method [25] to compute 95%confidence intervals for the estimated change in the FNMR.This current paper builds on the results in [24] by explicatingseveral possible contributing factors in iris template aging,and also by discussing several possible means to control theeffects of iris template aging.Czajka [26] studied iris template aging with a dataset of

571 images representing 58 eyes, with up to eight years timelapse, acquired between 2003 and 2011. He considered per-formance of three different iris matchers, and concluded that[26] ‘‘Average values of the genuine scores obtained for allthe tested matchers may suggest that iris templates age, whatpartially supports earlier findings determined for differentmatchers and different databases, yet collecting samples witha shorter time lapse between captures than in this work’’.He also reports that [26] ‘‘The extent to which the templateageing phenomenon is observed is however uneven acrossdifferent matchers, in particular we observe a higher influenceof the time flow for more accurate methods’’.Ellavarason and Rathgeb [44] re-analyzed the two-year

time-lapse dataset of Fenker and Bowyer [23] using six dif-ferent iris feature extraction algorithms. The same weightedadaptive Hough and ellipsopolar transform [45] is used foriris segmentation for each of the six iris feature extractionalgorithms. The highest-performing of the six algorithmsis generally the 1-D log-Gabor filtering based on Masek’swork [17]. In general, for the six algorithms, they report clearevidence of template aging – ‘‘. . . the fact that the resultinggraph shows a gap between the FNMR rates for short andlong term shows the existence of ageing’’ [44].In summary, recent studies by different research groups

have found an iris template aging effect. These studies involvevarious different image datasets and various matching algo-rithms, various approaches to analysis, and explore lengthsof time from two to eight years. All report an increase inthe FNMR with increased time since enrollment. There are

VOLUME 1, 2013 267


also suggestions that template agingmay bemore pronouncedfor higher quality images [21] and for more accurate match-ers [26].

III. IRIS RECOGNITION TEMPLATE AGING EXPERIMENTIris images for this experiment were acquired from 2008through 2011 using an LG 4000 sensor. Example imagesof the same iris from 2008 and 2011 are shown inFigure 1. Of the 322 total subjects, 177 are male and 145 arefemale; and 243 are Caucasian, 37 are Asian, 24 are otherand 18 unknown. Subject age varies from 20 to 64 yearsold. We consider the dataset in terms of one cohort withthree years of time lapse (2008–2011), two cohorts with twoyears (2008–2010 and 2009–2011), and three cohorts withone year (2008–2009, 2009–2010 and 2010–2011). Table 1summarizes the numbers of subjects, images, and short- andlong-term authentic matches.

A. SELECTION OF IRIS MATCHERWe evaluated four iris matchers available to us in order toselect the best-performing one for use in this experiment.One is our own version of the IrisBEE matcher. The sec-ond is the commercial VeriEye SDK (version 2.4) [22].The third and fourth matchers are other commercial SDKs,

FIGURE 1. Images of the Same Eye from 2008 (top) and 2011 (bottom).Images were acquired using an LG 4000 iris sensor.

TABLE 1. Experimental data by period of time lapse

TimePeriod

Numberof

Subjects

Numberof Images

ShortTime-Lapse

Matches

Long Time-Lapse

Matches

08-09 88 4,553 11,986 30,47009-10 157 8,046 23,882 54,41710-11 181 11,734 29,120 97,87908-10 40 2097 5,829 14,28209-11 124 8,082 18,963 66,84908-11 32 2,338 5,244 20,888

which are not named here due to restrictions in their licenseagreements.Using the entire image dataset described above, an all-

versus-all matching experiment was performed with the fourmatchers. The results, illustrated in Figure 2, involve over285 million comparisons. Based on the VeriEye matcherperforming the best of the four matchers in this experiment,it was selected for use in the remaining experiments. It isworth noting that VeriEye was also found to be one of the top-performing matchers in the recent NIST IREX report [41].Unlike a matcher that reports a fractional Hamming distance,the VeriEye matcher reports a similarity score that rangesfrom 0 to 9443, where 0 is a non-match and 9443 is an exactmatch; that is, an image matched against itself.

B. EXPERIMENTAL METHODWe define a short-time-lapse match to be a comparison oftwo images acquired within a few months of each other,but on different days, so that there are no ‘‘same session’’matches. For our experiments, we have a set of short-time-lapse matches for images acquired in the spring semester ofeach of 2008, 2009 and 2010.We define a long-time-lapse match as a comparison of two

images acquired in different years. For example, we have aset of long-time-lapse matches between an image from spring2008 and an image from spring 2009. We have three differentone-year time-lapse datasets, two different two-year datasets,and one three-year dataset. The short-time-lapse spring 2008dataset is the baseline for the one-year 2008–2009 dataset,for the two-year 2008–2010 dataset, and the three-year

0 0.005 0.01 0.015 0.02 0.025 0.03

0.97

0.975

0.98

0.985

0.99

0.995

1

False Accept Rate

Tru

e A

ccep

t R

ate

IrisBee

VeriEye

Commercial 1

Commercial 2

FIGURE 2. Selection of a best-performing matcher to use intemplate-aging experiments.

268 VOLUME 1, 2013


2008–2011 dataset. The short-time-lapse 2009 dataset is thebaseline for the one-year 2009–2010 dataset and the two-year2009–2011 dataset. The short-time-lapse spring 2010 datasetis the baseline for the one-year 2010–2011 dataset.

We compared the impostor and authentic distributions foreach short-time-lapse baseline to those of each of its long-time-lapse comparisons. We found that the impostor distribu-tion for matches between images taken on average about onemonth apart does not differ substantially from the impostordistribution for one, two or three years of time lapse. How-ever, the authentic distribution for a long time-lapse datasethas a generally higher FNMR than its corresponding short-time-lapse distribution. To investigate this further, we com-pared the FNMR for short and long time lapse, across a rangeof decision threshold values. We also calculated a bootstrap95% confidence interval [25] for the estimated change inFNMR.

Statistics are noted for the various datasets for a particulardecision threshold value based on a false match rate (FMR)selected based on the experiment shown in Figure 2. For theVeriEye matcher, in general, at a low enough value for thedecision threshold, the FMR is 100% and the FNMR is 0%.As the decision threshold is increased, the FMRdecreases andthe FNMR increases. For the experiment shown in Figure 2,a decision threshold of 580 corresponds to an approximately1 in 2 million FMR. At this threshold, there is sufficient FNMdata in the various datasets to justify computing bootstrapestimates of the change in the FNMR. At the decision thresh-old corresponding to a 1 in 1 million FMR, there generallyare too few FNM results to allow a reliable estimation of theFNMR.

C. ONE-YEAR TIME-LAPSE RESULTSThe three one-year time-lapse datasets represent 226 irises(113 subjects) imaged in 2008–2009, 338 irises (169 sub-jects) imaged in 2009–2010, and 406 irises (203 subjects)imaged in 2010–2011. The average number of days betweenacquiring a pair of images in the short-time-lapsematch groupis approximately 39 days, and the average number of daysbetween acquiring a pair of images in the long-time-lapsematch group is approximately 360 days. Thus the average‘‘aging’’ between the two authentic distributions is approx-imately 11 months.

The FNMR as a function of the decision threshold forthe 2008–2009 dataset is shown in Figure 3. The FNMRstarts at zero for a low decision threshold value and increasesas the decision threshold increases. If no template agingeffect existed, then the FNMR curves for the short and longtime-lapse datasets should be the same. Figure 3 showsthat the FNMR curves for short and long time-lapse sepa-rate as soon as the FNMR is measurable and that the gapbetween the curves increases over the range of decisionthreshold values. At the decision threshold of 580, corre-sponding to approximately a 1 in 2million FMRon the overalldataset, the bootstrap mean increase in the absolute FNMR is0.035, with a 95% confidence interval of 0.00723 to 0.0627.

FIGURE 3. Comparison of Short- and Long-Term FNMR for 2008–2009.

This corresponds to a mean percent increase in FNMR of27%, with a confidence interval of 5% to 61%.The FNMR as a function of the decision threshold for the

2009–2010 dataset is qualitatively the same as that shown inFigure 3 for 2008–2009. Again, the FNMR curves separate assoon as the FNMR is measurable and the gap between themincreases as the decision threshold increases. At the 1-in-2M-FMR threshold, the mean absolute increase in the FNMR is0.086, with 95% confidence interval of 0.063 to 0.112. Thecorresponding percent mean increase in the FNMR is 60%,with a 95% confidence interval of 40% to 84%.The FNMR as a function of decision threshold for the

2010–2011 dataset is also qualitatively the same. Again, theFNMR curve for long time-lapse consistently runs abovethat for short time-lapse. At the 1-in-2M-FMR threshold,the mean absolute increase in FNMR is 0.063, with 95%confidence interval of 0.041 to 0.086. The correspondingpercent increase in the FNMR is 49%, with a 95% confidenceinterval of 29% to 73%.

D. TWO-YEAR TIME-LAPSE RESULTSWe have two datasets representing two-year time lapse. Theserepresent 90 irises (45 persons) in the 2008–2010 datasetand 270 irises (135 persons) in the 2009–2011 dataset. Theaverage number of days between a pair of images in the two-year time-lapse data is 732 days.The FNMR as a function of the decision threshold for

the two-year, 2008–2010 dataset is shown in Figure 4. As isthe case with the three one-year datasets, the FNMR curvefor long time-lapse runs above that for short time-lapse. Atthe decision threshold corresponding to a 1-in-2M FMR, themean absolute increase in the FNMR is 0.11, with a 95%confidence interval of 0.056 and 0.168. This corresponds toa mean percent increase in the FNMR of 82%, with a 95%confidence interval of 38% to 150%.For the two-year, 2009–2011 dataset, the FNMR for long

time-lapse is again consistently greater than that for shorttime-lapse. At the decision threshold for a 1-in-2M FMR, themean absolute increase in the FNMR is 0.13, with a 95%confidence interval of 0.095 to 0.167. This corresponds toa mean percent increase in the FNMR of 91%, with a 95%confidence interval of 63% to 127%.

VOLUME 1, 2013 269



For each of the two two-year time-lapse datasets, as was thecase with each of the three one-year time-lapse results, thereis a clear template aging effect. At the 1-in-2M FMR decisionthreshold, the mean FNMR for the two datasets is estimatedat 82% and 91%. Each of these two values is greater than thevalue for any of the three one-year datasets. And the lowerlimit of the 95% confidence interval does not include zerochange in the FNMR.

E. THREE-YEAR TIME LAPSE (2008–2011)There are 70 irises (35 subjects) in the 2008–2011 time-lapsedataset. The mean time between image acquisitions for thetwo images in a long-time-lapse match is 1,068 days. Thusthe average time lapse is about 34 months.

The FNMR as a function of the decision threshold for thethree-year, 2008–2011 dataset is shown in Figure 5. Again,the FNMR for long time-lapse is consistently greater than forshort time-lapse. At the 1-in-2M FMR decision threshold, themean absolute increase in the FNMR is 0.147, with a 95%confidence interval of 0.088 and 0.208. This corresponds toa mean percent increase in the FNMR of 153%, with a 95%confidence interval 85% to 307%. This mean percent increasefor three-year time lapse is greater than that for either of thetwo-year datasets.

F. CHANGE BETWEEN 2008–2009, 2008–2010 AND2008–2011The one-year 2008–2009 results, the two-year 2008–2010results, and the three-year 2008–2011 results share highlyoverlapping short time-lapse authentic distributions basedon images from spring 2008. The long time-lapse authenticdistributions represent the spring 2008 subjects who alsohad iris images acquired in spring 2009, 2010 and 2011,respectively. Consider these three results as a time sequence.For 2008–2009, the mean percent increase in FNMR is 27%,with a confidence interval of 5% to 61%. For 2008–2010,the mean percent increase in the FNMR is 82%, with a 95%confidence interval 38% to 150%. For 2008–2011, the meanpercent increase in the FNMR is 153%, with 95% confidenceinterval for the percent increase is 85% to 307%. This data,summarized in Table 2, shows a clear trend of increasing


FNMRwith increasing time lapse. This is consistent evidenceof an iris template aging effect. However, it is also importantto note that the 95% confidence interval indicates a broadrange for the estimate of the increase in the FNMR.

IV. EYE AGING AND IRIS TEMPLATE AGINGThe biometric signal for iris recognition is the texture of theiris. Of course, the features for iris recognition are computedfrom an image of the iris texture rather than the iris itself.This distinction is important because any eye aging effect thatchanges the image of the iris texture automatically has thepotential to cause template aging. Consider the structure ofthe eye as depicted in Figure 6. The iris is necessarily imagedthrough the cornea and the aqueous humor, and subject topossible occlusion in the image by the eyelids and eyelashes.The human eye is known to undergo many different age-

related changes [27]. We briefly outline how known aging-related effects involving the eyelids, the cornea, and the iriscould contribute to template aging. This list is not at allintended to be comprehensive. It is only to suggest the com-plexity of a comprehensive consideration of possible factorscontributing to template aging. Establishing a comprehensivelist of eye aging mechanisms that can contribute to iris tem-plate aging, and the relative importance of their contributions,would be a substantial research topic in itself.The eyelids can occlude part of the iris in an image. The

term ptosis refers to drooping of the eyelid. It is known thatthe drooping of the eyelid can be ‘‘caused by the normal agingprocess’’ [28]. Eyelid droop that increases with age can resultin increased iris occlusion in later images. This translates intoless of the iris in view in common between two images asthere is more elapsed time between the images. Less iris areain view in common between the images effectively meansfewer iris code bits for matching, which in turn means a largervariance for the match score and an increased probability oferror. Thus ptsosis, or eyelid droop, can potentially contributeto an iris template aging effect.The cornea is often idealized as having spherical-shaped

outer (anterior) and inner (posterior) surface. Models of irisimage formation that use such a 3D model of the corneahave recently been explored as a means to correct off-angle iris images [29]–[31]. As Kennell et al [30] point out,

270 VOLUME 1, 2013


TABLE 2. FNMR from 2008–2009 to 2008–2011

Time Period % Change in FNMR (95% CI)2008–2009 27% (5%, 61%)2008–2010 82% (38%, 150%)2008–2011 153% (85%, 307%)

FIGURE 6. Diagram of the Structure of the Eye. (Courtesy of National EyeInstitute, National Institutes of Health; Ref#: NEA05).

‘‘The shapes and measurements of the cornea and iris willvary from person to person, and indeed the topography of thecornea is irregular, and thus the cornea cannot be perfectlymodeled as a smooth spherical or ellipsoidal surface.’’ Notmentioned in [30], but important in the context of this paper,is that the shape of the cornea changes with age [32], [33].At younger ages, the cornea tends to have greater curvature

along the horizontal axis than along the vertical axis. At olderages, the opposite is true. This is referred to as a changefrom ‘‘with the rule’’ to ‘‘against the rule’’ astigmatism. Ina study involving over 700 persons from aged 20 to 80-plus,Hayashi et al [32] note, the ‘‘normal cornea becomes steeperand shifts from with-the-rule to against-the-rule astigmatismwith age’’. In a similar study, Vihlen and Wilson found[33] ‘‘a significant relationship between corneal toricity andage, i.e., corneal toricity shifts towards against-the-rule withage’’.

The distance from the corneal surface to the iris alsochanges with age. The anterior chamber is the part of the eyefrom the outer surface of the cornea to the iris. Atchison et al[34] reported that for the emmetropic eye (an eye not need-ing vision correction), ‘‘Despite considerable data scatter,we found significant age changes: anterior chamber depthdecreased 0.011 mm/year . . . ’’.

Changes in the shape of the cornea and in the distancefrom the cornea surface to the iris surface will cause imagesof the same iris taken at different ages to have differently

FIGURE 7. Sequence of Iris Images Ordered By Pupil Dilation Ratio. Thisvideo is formed from a sequence of iris images taken at different timesand sequenced in order of increasing pupil dilation ratio. It is clear thatthere are changes in iris texture based on pupil dilation, and that largerdifferences in texture occur for larger differences in dilation (please seemultimedia package to view video).

warped versions of the iris texture. Thus, age-related changesin cornea shape are a potential contributor to template aging.The iris itself undergoes known functional changes with

age [35]–[37]. In a study of pupil size across persons ofdifferent ages and at different illumination levels, Winn et al[36] concluded ‘‘pupil size decreased linearly as a functionof age at all illuminance levels’’. (In addition to the averagepupil size becoming smaller with age, the responsiveness ofthe iris to a given change in light intensity decreases withage [37].) Because average pupil size decreases with age,increasing time between images of the same iris results inincreasing average difference in pupil dilation. Studies in theiris recognition literature show that increased difference inpupil dilation results in an increased FNMR [38], [39]. SeeFigure 7 for an illustration of how iris texture changes withpupil dilation. Thus the normal aging effect of a decreasingaverage pupil size with increasing age is a contributor to iristemplate aging.Figure 8 shows a bar chart of the mean difference in pupil

dilation for the datasets whose pattern of increase in FNMRis summarized in Table 2. The difference in pupil dilationfor a pair of images is computed as the absolute value of thedifference in the ‘‘PupilIrisRatio’’ value given by the VeriEyeSDK. Note that the average difference in pupil dilation ratioincreases with increased time, just as is to be expected basedon what is known about iris aging [35], [36]. The increaseddifference in pupil dilation is then expected, based on the irisrecognition literature [38], [39], to degrade the match scores.The contribution of increased difference in pupil dilation

to template aging has been noted in previous work [23].

VOLUME 1, 2013 271


FIGURE 8. Difference in Pupil Dilation for 2008–2008, 2008–2009,2008–2010 and 2008–2011. The 2008–2008 value is the mean differencein dilation ratio between two images of the same eye acquired in 2008,in units of the VeriEye SDKŠs PupilIrisRatio value, which has a range of0 to 255; similarly with 2008–2009, 2008–2010 and 2008–2011.

While pupil dilation is not the only factor contributing totemplate aging, it appears to be an important factor. It is alsocurrently the most well-explored factor in template aging. Itis possible in principle to study the impact of eyelid occlusionfrom the image data. However, the appropriate parametersare not directly reported in the VeriEye SDK used in thiswork. Investigation of factors such as change in corneal shapedoes not appear to be possible using only the type of imagesacquired for iris recognition.

It is sometimes asserted that the iris texture does not changewith age and that therefore template aging does not occur.Setting aside the question of whether the iris texture changeswith age, there is something important to note about the agingeffects outlined above. Even if the iris texture at a givendilation ratio was perfectly constant throughout a person’slifetime, each of the mechanisms outlined above could stillcontribute to template aging. This demonstrates how asser-tions about the apparent constancy of iris texture over timecan simply miss the point when it comes to discussions abouttemplate aging.

V. APPROACHES TO CONTROLLING TEMPLATE AGINGHistorically, the standard practice for iris enrollment is toenroll a template computed from a single image that passeschecks for focus quality and percent of iris occluded. Thestandard practice for matching is to match a template froma single image. Given what is currently known about iristemplate aging, it appears that a combination of relativelysimple extensions to the normal iris enrollment and matchingprocedure could effectively control iris template aging formany identity verification applications.

One step to effectively eliminate the component of templateaging due to average pupil size decreasing with age is toenroll an iris with a set of templates from multiple imagesthat represent a range of dilation values. The initial enrollmentprocess might then have a more active element of varyingillumination in order to generate a range of pupil dilation, withspecial attention to obtaining images with smaller pupil sizes.

This makes it possible to limit the template aging effect dueto difference in pupil dilation by matching against multipletemplates representing varying degrees of pupil dilation.A second step to limit the performance degradation due to

various other sources of template aging is to automaticallyupdate an enrollment template based on a new iris image usedfor a successful verification. The template that is updatedshould be the one that represents a dilation ratio closest tothat of the new image. It may be that a more stringent set ofquality checks is required for a verification that generates atemplate update and that a less stringent set of checks allowsverification without a template update. The general effect ofthis approach is to limit the ‘‘aging period’’ in a practicalway. The general family of schemes in which a successfulverification can result in an updated enrollment template hasbeen termed ‘‘rolling re-enrollment’’.It is also possible that the iris sensor and / or the image

acquisition environment could use visible-light illuminationto attempt to control pupil dilation. The LG 4000 sensorused to acquire images for our study described above doesnot use visible-light illumination to attempt to control pupildilation. However, the Iris Guard AD 100 sensor [42] appearsto monitor pupil dilation during acquisition and if the pupildilation is too large, to switch on a visible light to induce asmaller pupil size. Thus the control of template aging due todifference in pupil dilation can have both algorithm-focusedand sensor-focused elements.

VI. SUMMARY AND DISCUSSIONWhile it was once believed that iris template aging did notoccur and so ‘‘a single enrollment can last a lifetime’’, mul-tiple studies now show that iris template aging does occur[18], [21], [23], [24], [26]. Specifically, the experimentalevidence shows that the FNMR increases with increasingtime between the enrollment image and the image to berecognized. The template aging effect found in these studiescan be controlled by relatively simple enhancements of thetraditional methods of enrolling and matching irises. Severalpoints seem important to make for a better understanding ofthe iris template aging issue.

1) NORMAL AGE-RELATED CHANGE IN PUPIL DILATION ISONE CONTRIBUTOR TO IRIS TEMPLATE AGINGIt is known that average pupil dilation decreases withage [35], [36]. It is also known that increased differencein pupil dilation increases the FNMR [38], [39]. Thus age-related change in pupil dilation is one contributing cause toiris template aging. Other possible contributing factors in iristemplate aging remain to be studied in detail.

2) RELATIVELY SIMPLE CHANGES TO ENROLLMENT ANDMATCHING CAN CONTROL TEMPLATE AGINGIris template aging is made more pronounced by the tradi-tional practice of enrolling a single template and assumingthat it is ‘‘for life’’. Enrolling an iris with multiple templatesrepresenting different degrees of dilation improves iris recog-

272 VOLUME 1, 2013


nition accuracy [43]. In situations where the biometric is usedfor access control, a ‘‘rolling re-enrollment’’1 scenario maybe an attractive solution. A user’s set of enrollment templatesmay be automatically updated after a verification in which theimage passes appropriate quality control checks.

3) TEMPLATE AGING CAN OCCUR EVEN IF THE IRIS TEXTUREREMAINS THE SAMEA change in the shape of the cornea can cause a change inthe image of the iris texture. A change in eyelid position cancause a change in the amount of iris texture imaged. A changein dilation can cause a change in the iris texture that appears inthe image. In all of these cases, the texture of the iris surfaceitself may remain the same, and a template aging effect canstill occur.

4) PERCEPTION OF STABLE IRIS TEXTURE OVER TIME DOESNOT IMPLY CLOSENESS OF BIOMETRIC MATCHIn one experiment, Hollingsworth et al [40] showed subjectspairs of iris images that were either from identical twins orfrom unrelated persons. In another experiment, the pairs ofimages were either from the left and right eyes from the sameperson, or from unrelated persons. In both cases, subjectswere accurate at categorizing pairs of images that belongtogether versus pairs that do not. This shows that humans canperceive a similarity in the iris texture between left and righteyes of the same person, or between eyes of identical twins.However, using the same images, the average iris recognitionscore was no closer for identical twins, or for the left and righteyes of same person, than it was for unrelated persons. Thisshows that humans perceive similarities in iris texture that arenot reflected in iris recognition match scores. Thus, humanperception of a stable iris texture pattern over time does notnecessarily imply anything about iris recognition templateaging. Studies of iris biometric template aging should bedone in the context of recognition match scores, not humanperception of texture similarity.

5) IRIS TEMPLATE AGING IS OBSERVED OR MEASUREDEXPERIMENTALLYAn estimate of the size of iris template aging is necessarilymade in the context of a particular combination of sensor,subject dataset, time interval and matcher version. Varyingany one of these elements will generally affect the esti-mated template aging effect. Additionally, an image dataset ofpoorer quality may make it harder to experimentally observea template aging effect [21]. Similarly, a matcher of poorerquality may make it harder to experimentally observe a tem-plate aging effect [26].

AcknowledgmentThe term ‘‘rolling re-enrollment’’ was suggested byJohn Daugman.

1The term ‘‘rolling re-enrollment’’ is due to John Daugman.

REFERENCES[1] J. G. Daugman, ‘‘Biometric personal identification based on iris analysis,’’

U.S. Patent 5 291 560, Mar. 1, 1994.[2] J. Daugman, ‘‘Recognizing persons by their iris patterns,’’ in Biometrics:

Personal Identification in Networked Society, A. K. Jain, R. Bolle, andS. Pankanti, Eds. Norwell, MA, USA: Kluwer, 1998.

[3] A. N. Al-Raisi and A. M. Al-Khouri, ‘‘Iris recognition and the challenge ofhomeland and border control security in UAE,’’ Telemat. Informat., vol. 25,no. 2, pp. 117–132, 2008.

[4] Unique Identification Authority of India (2013, May 15) [Online]. Avail-able: http://uidai.gov.in

[5] K. W. Bowyer, K. Hollingsworth, and P. J. Flynn, ‘‘Image understandingfor iris biometrics: A survey,’’ Comput. Vis. Image Understand., vol. 110,no. 2, pp. 281–307, May 2008.

[6] M. J. Burge and K. W. Bowyer, Handbook of Iris Recognition. New York,NY, USA: Springer-Verlag, 2013.

[7] Information Technology—Biometric Performance Testing and Reporting—Part 1: Principles and Framework, ISO/IEC Standard 19795-1:2006, 2006.

[8] A. Lanitis, ‘‘A survey of the effects of aging on biometric identity verifica-tion,’’ Int. J. Biometr., vol. 2, no. 1, pp. 34–52, 2010.

[9] N. Ramanathan, R. Chellappa, and S. Biswas, ‘‘Computational methodsfor modeling facial aging: A survey,’’ J. Vis. Lang. Comput., vol. 20, no. 3,pp. 131–144, Jun. 2009.

[10] Germany’s Federal Office for Information Security, ‘‘Evaluation of fin-gerprint recognition technologies—Biofinger,’’ Bundesamt fur Sicherheitin der Informationstechnik, Fraunhofer IGD, Darmstadt, Germany, PublicFinal Rep. 1.1, 2004.

[11] L. Flom and A. Safir, ‘‘Iris recognition system,’’ U.S. Patent 4 641 349,Feb. 3, 1987.

[12] J. Thornton, M. Savvides, and V. Kumar, ‘‘A Bayesian approach todeformed pattern matching of iris images,’’ IEEE Trans. Pattern Anal.Mach. Intell., vol. 29, no. 4, pp. 596–606, Apr. 2007.

[13] K. Miyazawa, K. Ito, T. Aoki, K. Kobayashi, and H. Nakajima, ‘‘Aneffective approach for iris recognition using phase-based imagematching,’’IEEE Trans. Pattern Anal. Mach. Intell., vol. 30, no. 10, pp. 1741–1756,Oct. 2008.

[14] D. Monro, S. Rakshit, and D. Zhang, ‘‘DCT-based iris recognition,’’ IEEETrans. Pattern Anal. Mach. Intell., vol. 29, no. 4, pp. 586–595, Apr. 2007.

[15] Iris Recognition, Wikipedia. (2011, Mar. 17) [Online]. Available:http://en.wikipedia.org/wiki/Iris_recognition

[16] P. Tome-Gonzalez, F. Alonso-Fernandez, and J. Ortega-Garcia, ‘‘On theeffects of time variability in iris recognition,’’ in Proc. IEEE Int. Conf.Biometr. Theory, Appl. Syst., Oct. 2008, pp. 1–6.

[17] L. Masek and P. Kovesi, ‘‘MATLAB source code for a biometric identi-fication system based on iris patterns,’’ B.E. thesis, School Comput. Sci.,Softw. Eng., Univ. Western Australia, Perth, WA, Australia, May 2003.

[18] S. Baker, K. W. Bowyer, and P. J. Flynn, ‘‘Empirical evidence for correctiris match score degradation with increased time-lapse between gallery andprobe matches,’’ in Proc. Int. Conf. Biometr., 2009, pp. 1170–1179.

[19] P. J. Phillips, W. T. Scruggs, A. J. O’Toole, P. J. Flynn, K. W. Bowyer,C. L. Schott, and M. Sharpe, ‘‘FRVT 2006 and ICE 2006 large-scaleexperimental results,’’ IEEE Trans. Pattern Anal. Mach. Intell., vol. 32,no. 5, pp. 831–846, May 2010.

[20] D. Rankin, B. Scotney, P. Morrow, and B. Pierscionek, ‘‘Iris recognitionfailure over time: The effects of texture,’’ Pattern Recognit., vol. 45, no. 1,pp. 145–150, 2012.

[21] N. Sazonova, F. Hua, X. Liu, J. Remus, A. Ross, L. Hornak, andS. Schuckers, ‘‘A study on quality-adjusted impact of time lapse on irisrecognition,’’ Proc. SPIE, vol. 8371, pp. 83711W-1–83711W-9, Apr. 2012.

[22] Neurotechnology. (2013, Feb. 4). VeriEye SDK, San Francisco, CA, USA[Online]. Available: http://www.neurotechnology.com/verieye.html

[23] S. P. Fenker and K. W. Bowyer, ‘‘Experimental evidence of a templateaging effect in iris biometrics,’’ in Proc. IEEE Comput. Soc. WorkshopAppl. Comput. Vis., Jan. 2011, pp. 232–239.

[24] S. P. Fenker and K. W. Bowyer, ‘‘Analysis of template aging in irisbiometrics,’’ in Proc. IEEE Comput. Soc. Workshop Biometr., Jun. 2012,pp. 45–51.

[25] M. E. Schuckers, Computational Methods In Biometric Authentication.New York, NY, USA: Springer-Verlag, 2010.

[26] A. Czajka, Template Ageing in Iris Recognition. Barcelona, Spain: BioSig-nals, 2013.

[27] C. A. P. Cavallotti and L. Cerulli, Age-Related Changes of the Human Eye.Totowa, NJ, USA: Humana Press, 2008.

VOLUME 1, 2013 273


[28] PubMed Health, U.S. National Library of Medicine. (2013).Eyelid Drooping [Online]. Available: http://www.ncbi.nlm.nih.gov/pubmedhealth/PMH0002013/

[29] J. R. Price, T. F. Gee, V. Paquit, and K. W. Tobin, ‘‘On the efficacy ofcorrecting for refractive effects in iris recognition,’’ in Proc. IEEE Conf.Comput. Vis. Pattern Recognit., Jun. 2007, pp. 1–6.

[30] L. R. Kennell, R. P. Broussard, R. W. Ives, and J. R. Matey, ‘‘Preprocessingof off-axis iris images for recognition,’’ Proc. SPIE, vol. 7119, pp. 711906-1–711906-9, Sep. 2008.

[31] H. J. Santos-Villalobos, D. Barstow, M. Karakaya, E. Chaum, andC. Boehnen, ‘‘ORNL biometric eye model for iris recogition,’’ inProc. IEEE 5th Int. Conf. Biometr., Theory, Appl. Syst., Sep. 2012,pp. 176–182.

[32] K. Hayashi, H. Hayashi, and F. Hayashi, ‘‘Topographic analysis ofthe changes in corneal shape due to aging,’’ Cornea, vol. 14, no. 5,pp. 527–532, 1995.

[33] F. S. Vihlen and G. Wilson, ‘‘The relation between eyelid tension, cornealtoricity, and age,’’ Investigat. Ophthalmol. Vis. Sci., vol. 24, no. 10,pp. 1367–1373, 1983.

[34] D. A. Atchison, E. L. Markwell, S. Kasthurirangan, J. M. Pope, G. Smith,and P. G. Swann, ‘‘Age-related changes in optical and biometric character-istics of emmetropic eyes,’’ J. Vis., vol. 8, no. 4, pp. 1–20, Apr. 2008.

[35] J. Birren, R. Casperson, and J. Botwinick, ‘‘Age changes in pupil size,’’J. Gerontol., vol. 5, no. 3, pp. 216–221, 1950.

[36] B. Winn, D. Whitaker, D. Elliot, and N. Phillips, ‘‘Factors affecting light-adapted pupil size in normal human subjects,’’ Investigat. Ophthalmol. Vis.Sci., vol. 35, no. 3, pp. 1132–1137, 1994.

[37] P. Bitsios, R. Prettyman, and E. Szabadi, ‘‘Changes in autonomic functionwith age: A study of pupillary kinetics in healthy young and old people,’’Age Age., vol. 25, no. 6, pp. 432–438, 1996.

[38] K. Hollingsworth, K. W. Bowyer, and P. J. Flynn, ‘‘Pupil dilation degradesiris biometric performance,’’ Comput. Vis. Image Understand., vol. 113,no. 1, pp. 150–157, Jan. 2009.

[39] P. Grother, E. Tabassi, G. W. Quinn, and W. Salamon, ‘‘NIST IREX I: Per-formance of iris recognition algorithms on standard images,’’ InformationAccess Division, NIST, Gaithersburg, MD, USA, Interagency Rep. 7629,Sep. 2009.

[40] K. Hollingsworth, ‘‘Genetically identical irises have texture similaritythat is not detected by iris biometrics,’’ Comput. Vis. Image Understand.,vol. 114, no. 11, pp. 1493–1502, 2011.

[41] P. Grother, G. W. Quinn, J.R. Matey, M. Ngan, W. Salamon, G. Fiumara,and C.Watson, ‘‘IREX III: Performance of Iris Identification Algorithms,’’Information Access Division, NIST, Gaithersburg, MD, USA, InteragencyRep. 7836, Apr. 2012.

[42] AD 100 Iris Camera, Iris Guard (2013, May 15) [Online]. Available:http://www.irisguard.com/

[43] E. Ortiz and K. W. Bowyer, ‘‘Dilation-award multi-image enrollment foriris biometrics,’’ in Proc. Int. Joint Conf. Biometr., Washington, DC, USA,2011, pp. 1–7.

[44] E. Ellavarason and C. Rathgeb, ‘‘Template ageing in iris biometrics: Aninvestigation of the ND-iris-template-ageing-2008-2010 database,’’ Bio-metrics and Internet-Security Research Group, Center for Advanced Secu-rity Research, Darmstadt, Germany, Tech. Rep. HDA-da/sec-2013-001,2013.

[45] A. Uhl and P. Wild, ‘‘Weighted adaptive hough and ellipsopolar transformsfor real-time iris segmentation,’’ in Proc. 5th Int. Conf. Biometr., 2012,pp. 283–290.

[46] S. Baker, K. W. Bowyer, P. J. Flynn, and P. J. Phillips, ‘‘Template agingin iris biometrics: Evidence of increased false reject rate in ICE 2006,’’in Handbook of Iris Recognition, M. Burge and K. W. Bowyer, Eds. NewYork, NY, USA: Springer-Verlag, 2013.

SAMUEL P. FENKER received the B.S. degree(cum laude) in computer science from the Uni-versity of Notre Dame, Notre Dame, IN, USA,in 2012. As an Ateyeh Undergraduate ResearchScholar during his career at Notre Dame, he hasco-authored and presented several works focusingon the effects of template aging in iris biometricsystems. He currently works in research and devel-opment for amajor electronic health record vendor.

ESTEFAN ORTIZ is currently pursuing the Ph.D.degree in iris biometric systems from the Depart-ment of Computer Science and Engineering, Uni-versity of Notre Dame, Notre Dame, IN, USA. Hereceived the Master of Science degree in electri-cal engineering from the University of Hawaii atManoa, Manoa, HI, USA, in 2006, and the Bache-lor of Science degree in electrical engineering andthe Bachelor of Arts degree in mathematics fromSt. Mary’s University, San Antonio, TX, USA, in

2003. He is a Dean’s Fellow with the University of Notre Dame. He hasserved as a Principal Research Engineer with the Research Corporation, Uni-versity of Hawaii. His current research interests include biometric systems,machine learning, computer vision, and pattern discovery.

KEVIN W. BOWYER is currently the Schubmehl-Prein Professor and a Chair of the Department ofComputer Science and Engineering, University ofNotre Dame, Notre Dame, IN, USA. He has servedas an Editor-in-Chief of the IEEE TRANSACTIONS

ON PATTERN ANALYSIS and MACHINE INTELLIGENCEand the IEEE Biometrics Compendium. He is thefounding General Chair of the IEEE InternationalConference on Biometrics Theory Applicationsand Systems series, having served as a General

Chair of the BTAS in 2007, 2008, and 2009 conferences. He served as aGeneral Chair of the 2011 International Joint Conference on Biometrics, andas a Program Chair of the 2011 Automatic Face and Gesture Recognitionconference. He is a fellow of the IAPR and a Golden Core Member of theIEEE Computer Society. His latest book is the Handbook of Iris Recognition.He received the Ph.D. degree in computer science from Duke University,Durham, NC, USA.

274 VOLUME 1, 2013

Template Aging Phenmenn in Iris Recognition

Documents

Transcript of Template Aging Phenmenn in Iris Recognition