Analysis of Dilation in Children and its Impact on Iris Recognition
Priyanka Das, Laura Holsopple, Michael Schuckers, Stephanie Schuckers
AAnalysis of Dilation in Children and its Impact on Iris Recognition
Priyanka Das, Laura Holsopple, Stephanie SchuckersClarkson UniversityPotsdam, NY, USA [email protected],[email protected],[email protected]
Michael SchuckersSt. Lawrence UniversityCanton, NY, USA [email protected]
Abstract
The dilation of the pupil and it’s variation between amated pair of irides has been found to be an important fac-tor in the performance of iris recognition systems. Studieson adult irides indicated significant impact of dilation oniris recognition performance at different ages. However,the results of adults may not necessarily translate to chil-dren. This study analyzes dilation as a factor of age andover time in children, from data collected from same 209subjects in the age group of four to 11 years at enrollment,longitudinally over three years spaced by six months. Theperformance of iris recognition is also analyzed in presenceof dilation variation.
1. Introduction
Iris recognition is a well established field within biomet-ric recognition. In the last decade, efforts were targeted to-wards studying performance in systems longitudinally, i.e.as the time between the enrollment image and subsequentprobe image increases. Dilation has been identified as animportant factor contributing to the variability in the irisrecognition performance in multiple studies [18] [11] [8].Dilation is defined as the degree to which the pupil isdilated or constricted and has been represented as a dimen-sionless quantity of the pupil to iris ratio [1]. A high ra-tio indicates a high degree of pupil dilation and thus a lowusable iris area for analysis. A lower ratio indicates a con-stricted pupil that could complicate iris segmentation, im-pacting performance adversely. Intra-subject pupil size mayvary due to physiological factors like age, aging, in responseto emotional stimuli and environmental factors like illumi-nation and medical conditions and can vary stochasticallyover a time frame of few seconds to decades. Age, ag-ing and illumination are quantifiable factors. In a practicalscenario of iris recognition, illumination and environmentalfactors may vary. Identifying and quantifying the impact of age and aging on dilation and iris recognition performancecould help improve the robustness of the existing iris recog-nition technologies. Multiple studies reported the impact ofage on dilation and iris recognition performance in adults.However, the age group of children between 0 to 18 yearsremains unstudied. As indicated in a previous study [2], aschildren grow, they have a growth factor which impacts thesize of the pupil from birth to adolescence. Results fromstudies involving adults may not translate to children.This study looks into the impact of age and aging ondilation and the impact of variation in dilation on the irisrecognition performance in children in the age group of fourto 11 years at enrollment from 209 subjects over a period ofthree years.
2. State of Art : Dilation and Iris Recognition
To understand the effect of dilation on iris recognitionperformance, it is important to understand how iris recog-nition operates [6]. For the purpose of biometric recogni-tion, the pupillary boundary and the limbus boundary aredetected from the iris image and are segmented. The an-nular region representing the iris is projected into a pseudo-polar coordinate system drawing analogy to a homogeneousrubber sheet model. The elastic mesh work of the iris caus-ing the dilation and constriction is modelled by this coor-dinate system by drawing analogy from the topology of ahomogeneous rubber sheet annulus anchored along its outerperimeter with the tension controlled by an interior ring ofvariable radius [6]. The coordinate system has a polar vari-able i.e. the angle, θ , ranging between 0 and 2 π and a radialvariable, r, of the annular region which is always an unitinterval [0,1], both being dimensionless. Each point of theannular region i.e. the iris, irrespective of dilation, is repre-sented by a pair of variables [r, θ ]. Since the radial variableranges from the pupillary boundary to the limbus as a unitinterval, it inherently corrects the deformation in the iris dueto variable pupil dilation. This allows approximated com-parison between irides with variation in dilation, introduc-ing non-affine pattern deformation. However, this modelconsiders the deformation linear. There are studies affirm- a r X i v : . [ ee ss . I V ] S e p ng non-linear deformation pattern as well [23]. Thus, therubber sheet model, though widely used, does not alwaysprovides an absolute representation of the iris having de-formity as a function of pupil dilation. This has reflectedin various studies with poor biometric identification perfor-mance in cases with high degree of dilation variation, evento a point of false rejection during biometric identification[18] [11] [8]. Over three decades even though Daughman’srubber sheet retained its popularity, multiple alternatives tothe model has been published in literature [21] [20] [14][13] [15] [19] [6] [12] [9] [16] [3]. Most commercial algo-rithms however, remains a blackbox with no public infor-mation on the techniques adopted in their algorithm. Thus,the results of matching pair of iris images with varying dila-tion may be impacted differently with different algorithms.Pupil dilation or constriction is governed by the dilatorand sphincter muscles in the iris which are controlled bythe sympathetic and parasympathetic nervous system, re-spectively. The reflex action of dilation or constriction iseffected by various variable factors, including illumination,emotional and non-emotional factors and medical condi-tions. Physiological factors of age and aging is also relatedto dilation. Fluctuation in pupil dilation is highly correlatedwith emotion processing and non-emotional state of deci-sion making [17]. Many medical conditions may affect thepupil dilation. For opthalmological examination a chemi-cal compound atropine sulphate is used to dilate the pupilwhich remains effective for several hours. In 1950, Bir-ren et al. [4] reported significant reduction in pupil sizewith age in both light and dark conditions, from a study on222 subjects in the age group of 20 to 89 years of age. In1965 Alder reported from his medical research that pupilsare small in newborn babies and remain small until the firstyear of birth, reaching its maximum during childhood andadolescence, and then gradually decreases with advancingage [2]. More recently, in 1994 Winn et.al. investigatedpupil dilation at fixed illumination levels in 91 subjects inthe age group of 17 and 83 years and concluded a steadylinear decay in the dilation as a function of age for all illumi-nance levels [22]. In 2008 Hollingsworth et.al. [11] studiedthe performance of iris recognition with 18 adult subjectsand noted that, as the size of the pupil increases, the meanof the hamming distance gets closer to the non-match dis-tribution. The researchers also noted a larger difference inpupil size between enrollment and verification yields higherdissimilarity. In 2011, Fairhurst and Erbilek [8] in a studywith 632 images from 79 subjects in the age group of 18-73 years concludes a gradual decay in mean dilation fromthe age group <
25, to 25-60 to >
60 years. Aging affects theaccommodation capacity of the pupil, resulting in decreaseddilation with aging, keeping other factors constant. The re-searchers concluded that, in older adults (age >
60) the irisrecognition performance is less impacted by dilation as the change in dilation is decreased in this age group as opposedto that of ‘younger individuals’ (their study only includedadults). An increase in performance is noted with older agegroup with state of art segmentation. In 2013, Ortiz et al.[18] studied effect of age and aging on biometric perfor-mance using data from 955 subjects in the age group of 18to 64 years collected over 3 years. The age group between18 and 25 years dominated the subject count. They ob-served an increase in the hamming distance between matedpair of images with age and hypothesized that the increasedhamming distance is correlated with age as change in thesize of the pupil is impacted by age.It is important to note that the ‘effect of age’ and ‘effectof aging’ on any trait are two different concepts as definedbelow-•
Age study :
Impact of age of an individual on dilation is investigatedin this study by maximizing the use of our longitudinaldataset. We re-organized the data by age regardless ofwhat session they were captured, where each age betweenfour to 14 years is a cohort and performed a basic analysisof all data across age to understand dilation at differentages.•
Aging study or Longitudinal study :
Aging effects are typically investigated by comparing theperformance of a sample or samples at different points intime i.e. longitudinal study [5].Our literature review revealed that most prior researcheswere cross-sectional studies on the relationship between di-lation and age, predominantly on large groups of adult sub-jects and conclusions were drawn statistically on widelyspanned age groups. Only one research group, Ortiz et al.[18] correlated aging, dilation and iris recognition perfor-mance (hamming distance and match score) by designinga composite model and concluded a measurable degrada-tion in the performance metric due to dilation differencecaused by aging effect. We did not find any literature onthe effect of pupil size variation on iris recognition perfor-mance in children. This work concentrates on quantitativeanalysis of dilation as a factor of both age and aging andits effect on biometric performance in children in the agegroup of 4 to 14 years at a granular level for each age, from8612 samples collected from the same 209 unique subjectsover three years from seven collection sessions spaced byapproximately six months. This paper addresses mainly onthe three following questions for ages four to 14 years:•
Is there a relationship between age and dilation? • Does aging impact dilation over a period of threeyears? • How does the difference in dilation between matedpair of images impact iris recognition performance ina longitudinal scenario of three years? . Definitions and Acronyms
This section summarizes the terms and the formulas usedthroughout this paper.
Dilation or pupil dilation is a dimensionless quantitymeasuring the degree to which the pupil is dilated or con-stricted, measured as a ratio of pupil radius and iris radiusas defined below.
Dilation ( D ) = P upil radiusIris radius × (1) Difference in the pupil dilation between mated pair ofiris images is defined as Delta Dilation ( ∆ D) in this paperand the measure follows NIST work in [10] as below.
Delta Dilation (∆ D ) = 1 − − D − D (2)considering, D1 ≥ D2, where, D1 and D2 are the pupil di-lation of the first and the second iris images as estimatedusing equation 1. RI : Right Iris2. LI : Left Iris3. MS : Match Score4. G1 : Group 1 - All subjects who participated in at least two of the seven sessions5. G2 : Group 2 - All subjects who participated in Collec-tion 1 and Collection 7 with possible intermittent gap6. G3 : Group 3 - All subjects who participated in all sevensessions from Collection 1 to Collection 7
4. Data Collection Protocol and Statistics
The iris data used in this study is part of a larger lon-gitudinal dataset of multiple biometric modalities collectedfrom the same children over three years. This is a contin-uing study and till the point of analysis and preparation ofthis paper we had in total seven sessions of data collection.Data was collected from 239 subjects in the age group offour years to 11 years at the time of enrollment with sevenvisits subsequently spanned over three years, spaced by ap-proximately six months. 209 subjects participated in morethan one session, and data from these subjects were ana-lyzed.Researchers collaborate with the local school to identifysubjects for voluntary participation. An approved IRB pro-tocol requires an informed consent from parents and par-ticipants. Initial participation for the first session was open to children aged between 4 to 11 years. Henceforth, ev-ery year new subjects from Pre-K are added to the studywho are mostly in the age group of four to five years. Theequipment are setup in the school in an isolated room asprovided by the school for the entire collection week(s).The same equipment are used for each session to mini-mize variation in the data quality or properties. However,the room may vary at each session based on availability,which might affect the collection environment like lighting,temperature, noise which may impact the collected biomet-ric data. Measurements are taken to mitigate the variationin environmental factors effecting iris data collection. Theblinds in the room are drawn to minimize exposure fromexternal daylight and NIR from the sun which is the pri-mary illumination for iris capture. Factors such as medicalconditions, different collection rooms, weather conditions,time of year (fall or spring) cannot be eradicated. However,we provide each subject considerable time in the collectionroom to allow the eye to optimize and accommodate andadjust the dilation, to mitigate the variability due to varia-tion in illumination before coming for collection. Since thedata collections are done indoor we expect negligible im-pact due to weather and season. Participation on the day ofthe collection is voluntary; if a subject refuses to providedata on the day of the collection they are excused from par-ticipating. Participants are provided with an amicable en-vironment, with no emotional excitation/stimulating factorwhich might impact the data collection. However, any per-sonal emotional factor is not accounted in the data. A com-mercially available iris sensor, IG-AD100 Dual Iris Cameramanufactured by Iris Guard is used for data capture whichdetects and auto-captures iris in the NIR wavelength. Thequality of the camera and the images captured are compliantwith ISO/IEC 19794-6 [1]. In addition to the NIR illumi-nation, the camera also has a white LED flashing illumina-tion with the purpose of providing stimulus for stabilizingthe pupil dilation to maintain a stable intra-subject dilationacross sessions. The camera was donated for the purpose ofthe study of the impact on iris recognition in children.Four images were captured from each eye at each sessionwith some exceptions in first session and in sixth sessionwhen at least two images were captured. A few subjectsmay have more than four images captured per eye per ses-sion. Prior to the seventh session, the images collected werehighly correlated, due to the internal setup of the camerawhich captures images within a few seconds. The protocolwas modified in the seventh session when the four imageswere captured in two sets with a small time gap (less thantwo minutes) between the sets. A total of 8612 samples(LI: 4323, RI: 4289) were analyzed. Right and left iris im-ages were analyzed separately. The number of participantsin each session may vary due to new subjects added to thestudy every year, subjects moving out of the study and ab-igure 1: Number of subjects and image samples at differ-ent collection for different groups - G1, G2 and G3 (top tobottom) of study for both left and right irides (left and right,respectively)sentees on the day of the collection. The data has been stud-ied in three groups, G1, G2 and G3, as described in Section3.3. Subject and sample count in each visit for each groupis shown in Figure 1. 209 unique subjects have participatedin more than one collection. Thus in G1- 209 (209 for RI) ,in G2 - 105 (101 for RI) and in G3 - 63 (62 for RI) uniquesubjects were analyzed. The discrepancy in the number ofsubjects between LI and RI is due to the fact that for a fewsubjects both the irides could not be captured.
5. Results
The impact of age, ranging between four to 14 years, ondilation and longitudinal impact of aging over three yearsfor the age group of four to 11 years on delta dilation ( ∆ D)has been analyzed in this section. Further, the impact of ∆ Don longitudinal performance of match score has been ana-lyzed. All analysis is done based on groups (G1, G2, G3)as defined in Section 3.3. All attributes (pupil and iris ra-dius) has been extracted and matching has been performedwith commercially available software Verieye v11.1. Ver-ieye complies by ISO/IEC 19794-6 [1] guidelines in pro-cessing images. It is important to use a commercial ISOstandardized software to study aging. All further analy-sis has been performed in MATLAB 2018b and R studiov1.1.456.
To analyze the correlation between age and dilation, eachage between four to 14 years is considered as a cohort; im-ages in the dataset from each age are analyzed for each co-hort irrespective of the session the data was captured. Fig-ure 2 shows the dilation distribution in boxplots for each agecohort for all three groups of study (G1, G2, G3) for bothLI and RI and respective sample and subject count. For G1, data from a significant number of subjects areavailable for each age between four to 13 years. While themean does not appear to vary across ages, there is increasedvariability in ages 7-10 years compared to 4-6 and 11-14years, particularly between 50th percentile and the maxi-mum value of dilation, i.e. there was more spread for thosesubjects in the upper half of dilation. Said simply, the upperhalf of subjects in ages 7-10 had a higher dilation comparedto other ages. The 75th percentile changes from around 40for ages 4-6, to 45 for ages 7-10, and finally comes backaround 40 for ages above 10. Outliers are observed at al-most all ages. The mean dilation shows a partial patternacross ages- being minimum at age four and gradually in-creasing till the age group of 8-10 years and again graduallydecreasing or plateauing till the age of 14 years, particularlyprominent in G2 and G3. This pattern partially corroboratesthe observation by Alder in 1965 in [2] that the pupil reachits maximum size between first year of birth to childhoodand adolescence, and then gradually decreases with advanc-ing age. While this trend is not clearly evident in youngerages in this dataset given the large variability, there appearsto be a downward trend after age 10. The trends observedfor G1 are similar for both G2 and G3. It should also benoted that the subject count for age 4,5 and 14 years in G3is low and is not likely to produce reliable statistical plot.
This section performs longitudinal analysis (refer Sec-tion 2) i.e. change in dilation over time. Figure 3 and Fig-ure 4 shows boxplots of the variation in dilation over sevensessions for three years for each enrollment age betweenfour and 11 years for groups G2 and G3 respectively andthe subject count associated with each plot. G1 is elimi-nated from this analysis to simplify the segregation of dataand its analysis. It is important to note that the subject countin each age group for G3 is low. Two important points wereobserved in this analysis: (1) The variability in dilation issubstantially high over all 3 years for the enrollment agegroup of 7 years through 10 years in both G2 ad G3 forboth LI and RI. This observation is in line with our find-ing in Section 5.1 that high variability is observed in theage group of 7-10 years. However, this conclusion is notstrongly reflected in the age group of 6-9 years and 8-10years, where inconsistent variability is noted. Thus, this ob-servation remains inconclusive. (2) Considering the dilationvalue at 45 as the limit, the number of outliers increases asthe enrollment age increases for both groups- G2 and G3.
Delta dilation in terms of longitudinal data reflects thedifference in dilation between two mated pairs of imagescaptured at different time instances and this section ana-lyzes its impact on the match score, quantifying the per-igure 2: Boxplot of dilation for different age groups for all three groups of study - G1, G2, and G3 (top to bottom, for bothLI and RI(left and right, respectively). ’N’ denotes subject count and ’n’ denotes sample count for each boxplot.formance of iris recognition in a longitudinal scenario. Inour dataset ∆ D varies between 0 to 0.45 (0 to 45%). Therange of variation is high. Thus we looked at the distribu-tion of ∆ D at different time frames for all three groups ofstudy. Figure 5 shows the histogram of ∆ D with a bin sizeof 0.05 at different time frames for all three groups of studyfor left iris. The distribution is found to be consistent acrossall domain (time-frame, groups of study and left and rightiris). This observation supports consistency in our data col-lection. The majority of variation in ∆ D is between 5% to15% (0.05 to 0.15) with a small percentage of outliers at amaximum of 45% (0.45) ∆ D. We conclude that in average10% to 15% variation in ∆ D could be expected in a prac-tical scenario with factors affecting the collection like age,aging, partially controlled illumination and weather.Most previous studies indicated a linear relationship be- tween change in dilation and any metric determining thecorrelation between two samples from different time frame(Match score or MS in our case). Thus we tested the lin-ear relationship between the match score and delta dilationfor our dataset by fitting a linear statistical model for differ-ent time frames of study as shown in Figure 6 for all threegroups of studies. We note statistical significance in decayin MS with increased delta dilation for each individual timeframe analysis for all groups. A statistically significant neg-ative correlation between MS and ∆ D is noted across all do-mains (time frame, groups and left and right iris) of analy-sis. Overall, the estimate ranges from 328.4 ± ± p ≥ . ∗ e − , the slope of the model vary-ing at different time frame. However, we do not note anytrend in the decay in MS with aging (increase in time-framefrom 6 month to 36 month) with the linear model. Anotherigure 3: Dilation as a factor of aging for each enrollment age, over three years from seven collections and the number ofsubjects analyzed for each boxplot for G2 for both left and right iris (top and bottom respectively)significant observation was the fit of the model depicted asthe R-squared value, defined as the amount of variability inthe data captured by the model. We note that the designedlinear models only account for a minimum of 3.6% (3.0%in RI) to a maximum of 14.45% (18.36% in RI) and an av-erage of 8.5% (8.6% in RI) variation in match score acrossall groups and all time frames due to delta dilation.Assuming that ∆ D and MS are linearly related and themodel is the best representation of the relationship, ∆ D in-duces an approximated average of 8.5% variability in theMS. Aging is only one of many other factors affecting ∆ D.Thus it is safe to assume that a fraction of the 8.5% variabil-ity is induced due to aging. Thus, only a small percentage ofvariability in MS could be deduced due to dilation effectedas a factor of aging in children in the age group of 4 to 11 years at enrollment over a period of three years.No case of false rejection is noted due to high ∆ D. Overthree years, two cases of false rejection has been noted asbelow-• Case 1 - Left iris images rejected at 6 month time frame;the average ∆ D of the rejected images being 0.0901• Case 2 - Right iris images rejected at 18 month and 30month time frame; the average ∆ D of the rejected imagesbeing 0.1011 and 0.105 respectivelyIn both the cases the average ∆ D is comfortably below theaverage ∆ D in the population. And as such, the false rejec-tion do not appear to be the result of difference in dilationbetween the gallery and probe; the causes for the false re-jections are beyond the scope of this paper.igure 4: Dilation as a factor of aging for each enrollmentage, over three years from seven sessions for G3 for both LIand RI (top and bottom respectively). ‘N’ denotes subjectcount
6. Discussion and Conclusion
In an attempt to understand and quantify the impact ofage on dilation in children we conclude from our study thatthe dilation is minimum at age 4 and gradually increases andreaches it’s maximum at around 8-10 years age and thengradually decreases till age 14. We conclude, on average ∆ D varies between 10% to 15% irrespective of time framei.e. aging over a period of three years. No trend is notedin the variation in ∆ D with increasing time frame. No falserejection was noted due to ∆ D. Based on our linear model,on average ∆ D accounts for only 8.5% of variability of MS.Thus we conclude that aging impacts the match score bya fraction of 8.5%. This study cannot conclude anythingbeyond the studied age group of four to 14 years. Thoughdilation may vary across subjects and ages, we concludethat in a time period of three years the impact of aging oniris recognition performance is negligible for the age groupof 4 years to 11 years.It is extremely challenging to segregate the impact of dif-ferent factors- age, aging, illumination, weather and medi-cal factors on dilation. There is no publicly available datasetin this age group of four to 14 years concentrating specifi-cally on the impact of age and aging on dilation and delta di-lation for research. The data used for this study is collected from 209 individuals in the age group of four to 11 yearsover a period of three years spaced by approximately sixmonths. Measures has been adapted to collect data in a par-tially controlled environment to minimize the impact of thevariable environmental factors; however it is not completelyvoid of the other variability factors affecting dilation.Any data that identifies an individual is sensitive, moreso when it is biometric data of children. To protect the pri-vacy of the data and the subjects, the dataset is not pub-licly available presently. However, we understand the lackof data and the need for such datasets to advance researchin biometrics concerning children. In view of this, we aremaking efforts to provide access to the dataset for algorithmtesting for research purposes while protecting the privacy ofthe data and the subjects. Such efforts needs substantial re-sources and time. We plan to make it available by December2020.This paper analyzed dilation changes in children for dif-ferent ages and over time. State of the art software, Ver-iEye, a commercially available ISO standardized [1] soft-ware, was used for locating the pupil and outside of theiris in order to measure dilation and delta dilation. Thesemeasures are independent of the iris matching algorithmused. VeriEye was also used for iris matching. The ef-fect of dilation on iris matching performance may be im-pacted by the choice of algorithm and its robustness to vari-ation in dilation. Many techniques for iris recognition hasbeen proposed in the literature in the last three decades[7] [21] [20] [14] [13] [15] [19] [6] [12] [9] [16] [3]. Ofthese, Daugman’s iris recognition algorithm [7] is the old-est and is widely adapted. However, like most commer-cial systems, the algorithm used by VeriEye is a blackbox.There is no public information on what measures (if any)are taken to address the impact of dilation on recognitionperformance or what feature extraction and matching tech-niques are incorporated in the algorithm. However, the re-sults of this study are useful for practical applications whichwould likely to use a commercial algorithm like Verieye.We intend to study the performance of the dataset with ad-ditional algorithms, as well as allow organizations to uploadtheir own algorithms to test against this dataset in the nearfuture.
Acknowledgement
We extend our gratitude to the Potsdam Elementary andMiddle School, the staff, the participants and the parents ofthe participants for helping us to successfully create the in-valuable dataset for this study. We also thank all the collec-tors from our research team and other associates for theircontribution with their invaluable time. This material isbased upon work supported by the Center for IdentificationTechnology Research and the National Science Foundationunder Grant No. Clarkson 1650503.igure 5: Histogram of Delta Dilation, with bin width of 0.05, for different time frames of 6 to 36 months (left to right)longitudinally over 3 years for G1, G2 and G3 (top to bottom) for left iris. RI histogram is similar to that of LI and thus isnot included to accommodate spaceFigure 6: Linear modelling of the match score as a factor of change in dilation between enrollment and probe for eachdifferent time-frame (6 to 36 months) for G1, LI. Graphs of all other domain (groups, left and right iris and all times frames)show similar plots, and are not included due to space constraint. eferences [1] ISO/IEC 29794-6:2015 Information Technology - Biometric[3] M. Arsalan, H. G. Hong, R. A. Naqvi, M. B. Lee, M. C.Kim, D. S. Kim, C. S. Kim, and K. R. Park. Deep learning-based iris segmentation for iris recognition in visible lightenvironment.
Symmetry , 9(11):263, 2017. 2, 7[4] J. E. Birren, R. C. Casperson, and J. Botwinick. Age changesin pupil size.
Journal of Gerontology , 5(3):216–221, 1950.2[5] J. F. Corso. Sensory processes and age effects in normaladults.
Journal of Gerontology , 26(1):90–105, 1971. 2[6] J. Daugman. How iris recognition works. In
The essentialguide to image processing , pages 715–739. Elsevier, 2009.1, 2, 7[7] J. G. Daugman. High confidence visual recognition of per-sons by a test of statistical independence.
IEEE transactionson pattern analysis and machine intelligence , 15(11):1148–1161, 1993. 7[8] M. Fairhurst and M. Erbilek. Analysis of physical ageingeffects in iris biometrics.
IET Computer Vision , 5(6):358–366, 2011. 1, 2[9] A. Gangwar and A. Joshi. Deepirisnet: Deep iris represen-tation with applications in iris recognition and cross-sensoriris recognition. In , pages 2301–2305. IEEE, 2016. 2,7[10] P. J. Grother, J. R. Matey, E. Tabassi, G. W. Quinn, andM. Chumakov. IREX VI-Temporal stability of iris recog-nition accuracy. Technical report, 2013. 3[11] K. Hollingsworth, K. W. Bowyer, and P. J. Flynn. Pupil di-lation degrades iris biometric performance.
Computer visionand image understanding , 113(1):150–157, 2009. 1, 2[12] N. Liu, M. Zhang, H. Li, Z. Sun, and T. Tan. Deepiris: Learn-ing pairwise filter bank for heterogeneous iris verification.
Pattern Recognition Letters , 82:154–161, 2016. 2, 7[13] L. Ma, T. Tan, Y. Wang, and D. Zhang. Efficient iris recog-nition by characterizing key local variations.
IEEE Transac-tions on Image processing , 13(6):739–750, 2004. 2, 7[14] L. Ma, Y. Wang, and T. Tan. Iris recognition using circu-lar symmetric filters. In
Object recognition supported byuser interaction for service robots , volume 2, pages 414–417. IEEE, 2002. 2, 7[15] D. M. Monro, S. Rakshit, and D. Zhang. Dct-based irisrecognition.
IEEE transactions on pattern analysis and ma-chine intelligence , 29(4):586–595, 2007. 2, 7[16] K. Nguyen, C. Fookes, A. Ross, and S. Sridharan. Iris recog-nition with off-the-shelf cnn features: A deep learning per-spective.
IEEE Access , 6:18848–18855, 2017. 2, 7[17] M. Oliva and A. Anikin. Pupil dilation reflects the timecourse of emotion recognition in human vocalizations.
Sci-entific reports , 8(1):4871, 2018. 2[18] E. Ortiz, K. W. Bowyer, and P. J. Flynn. A linear regressionanalysis of the effects of age related pupil dilation change in Sample Quality -Part 6: Iris Image Data, 2015. 1, 3, 4, 7[2] F. H. Adler.
Physiology of the eye: Clinical Application .Mosby, 1965. 1, 2, 4iris biometrics. In ,pages 1–6. IEEE, 2013. 1, 2[19] Z. Sun and T. Tan. Ordinal measures for iris recognition.
IEEE Transactions on pattern analysis and machine intelli-gence , 31(12):2211–2226, 2008. 2, 7[20] C.-l. Tisse, L. Martin, L. Torres, M. Robert, et al. Personidentification technique using human iris recognition. In
Proc. Vision Interface , volume 294, pages 294–299, 2002.2, 7[21] R. P. Wildes, J. C. Asmuth, G. L. Green, S. C. Hsu, R. J. Kol-czynski, J. R. Matey, and S. E. McBride. A machine-visionsystem for iris recognition.
Machine vision and Applications ,9(1):1–8, 1996. 2, 7[22] B. Winn, D. Whitaker, D. B. Elliott, and N. J. Phillips.Factors affecting light-adapted pupil size in normal humansubjects.
Investigative ophthalmology & visual science ,35(3):1132–1137, 1994. 2[23] H. J. Wyatt. A minimum-wear-and-tearmeshwork for theiris.