Regional Approach for Iris Recognition

Regional Approach for Iris Recognition

DONATO IMPEDOVOUniversity of Bari ”A. Moro”

Department of Computer Sciencevia Orabona, 4, 70125 Bari

[email protected]

GIUSEPPE PIRLOUniversity of Bari ”A. Moro”


[email protected]

LORENZO SCIANATICOUniversity of Bari ”A. Moro”


[email protected]

Abstract: Recent studies have explored a various set of techniques, methods and algorithms for iris recognition.Segmentation algorithms, normalization processes and well-defined distances measures are employed as decisionmethods to compose the outline of all currently deployed systems, including commercial ones. So far, all systemsmake use of the entire iris pattern of any individual, assuming that a fixed threshold applied on the distance canpartially avoid problems like noise of information and acquisition errors. In this paper a new approach is presentedthat uses only selected sectors of the iris and a simple analysis strategy, for verification purposed. Therefore, itdiffers from all previous approaches that process the entire iris pattern. In addition, three Bayes-based similaritymeasures have been considered for iris image matching: the Whitened Cosine (WC), the PRM Whitened CosineTransform (PWC) and the Whitin-class Whitened Cosine Transform (WWC). The experimental results, carried outusing the CASIA database, are encouraging and demonstrate that the proposed approach supports several directionsfor further research.

Key–Words: Biometrics, Iris, Recognition, Biometrics, Ssim, Regional Analysis

1 Introduction

Recently, biometrics is becoming more and more im-portant in modern society and a multitude of biometricsystems have already been developed. The advantageof biometric technologies resides in his capability toimprove systems like PINs, passwords, keys or othertoken-based identification systems, with the identifi-cation of humans by their traits. Moreoverm, biomet-rics have increased their importance even in biomed-ical images sector. Two types of biometric traits canbe defined: behavioral and physiological. The firstone refers to the way a person performs certain tasks,like signature, keystroke dynamics or gait. The sec-ond one is a measure taken on a precise part of thehuman body, like fingerprints, face and DNA [1, 2].Human eyes offers a wide set of biometric traits thathave been considered for personal verification, likeretina, sclera and iris [1, 2, 3, 4]. The blood vesselstructure of human retina can be used in a biomet-ric technology, obtaining very good results. Unfortu-nately, retinal scan has some disadvantages. A retinalscanner is very expensive, more than an iris scanner,and this technology is less user-friendly than the sim-ilar one. Moreover, there are other problems in retinalscanning.

Measurement accuracy can by affected by dis-eases like cataracts, and also by severe astigmatism.

Figure 1: The intricate pattern of an iris

Moreover this technology is perceived by some asinvasive, because it is required to the subject to stayvery close to the instrument [2, 3, 4]. The net resultsthat iris is one of most attractive and interestingpart of human eye for personal verification. Facerecognition is perceived as less intrusive than irisrecognitin, but it is also affected by problems offlase match provided by natural aging of individualsand by the increasing use of plastic surgery. Irisis almost stable over years, and sometimes it wassubject os some speculations like ”iridology”, apseudo-science that claims to obtain medical diag-nosis analysing human iris, which was proved as a

WSEAS TRANSACTIONS on INFORMATION SCIENCE and APPLICATIONSDonato Impedovo, Giuseppe Pirlo, Lorenzo Scianatico

E-ISSN: 2224-3402 61 Volume 11, 2014

medical fraud [1]. On the contrary, retinal scan canprovide help to medical diagnosis of some disease [3].

Iris is located under the cornea. Eyelashes andeyelids protects iris from external agents, like dustor similar. At the center of iris is located the pupil,that controls the quantity of light passing througheye. Usually, iris is not only characterized by its color(that can be rare like blue and green or more commonlike hazel and dark) but even by its pattern, which isusually very complex and rich of details. Thereforeit is not surprising that since many years iris patternshave been rightly considered useful for biometricmeasurement [1].

Due to intensive research, several applica-tions of iris scanner were deployed in last yearsto ensure security in some critical situations. Thefirst example is UK’s I.R.I.S. (Iris RecognitionImmigration System), which started operating in2004 but which was closed to new registrations in2011 [5]. Google uses this technology to controlaccess to their datacenters [6]. The most interestingapplication is the recognitions of Gula, the famous”Afghan Girl” (figure 2) portrayed by Mc.Curry forNational Geographic, in 1984 and traced 18 yearslater. The system that recognized Afghan Girl wasimplemented by Daugman himself [7]. In future, itis planned to substitute authentication for differentweb platforms like Facebook and eBay with irisrecognition[8]. It is important to remind that thesetechnologies aim to integrate PINs and passwords,because biometrics requires further development.An example is the well-known smartphone SamsungGalaxy SIII and SIV which permits to unlock devicewith a combination of face and voice recognition,but reporting this as ”low protection”, while PIN andpassword are reported as ”medium-high protection”.In one of the last models of Apple iPhone wasincluded fingerprint recognition to unlock device, butthe system was hacked few days after his presentation.

At the state-of-the-art, there are several ap-proaches for iris recognition. Wildes [9] uses for seg-mentation a technique based on the Hough Transform,which is a standard algorithms in computer vision todetect circular objects in any image, include as defaultin various computer vision products like MATLab orOpenCV. Wildes [9] obtains normalization with animage registration technique which aligns a new im-age acquired with a selected image in a database. Fea-ture encoding is performed using a Laplacian of aGaussian filter, then normalized correlation is com-puted to match two images. An example of variationof this strategy is given by Boles [10]. Instead of using

Figure 2: Sharbat Gula, ”Afghan Girl”

image registration technique, iris images are scaled tohave a constant diameter that is chosen from a refer-ence image. Once the images are normalized to thesame diameter, feature extraction is performed. Thisis accomplished by extracting intensity values of dif-ferent parts of iris along different concentric circles.Iris normalization allows ensures that the same num-ber of points is extracted from all iris images. So,the encoding is done using the zero-crossing of the1D wavelet. In his work, Zhu et al.[11] proposed an-other metric to evaluate different pattern similarity, theWeighted Euclidean Distance:

WED(k) =

n∑i=1

(fi − f (k)i )2

δ(k)i

(1)

Where an unknown iris template is matched withall n examples of the database and match is found at aminimum k. The most important iris recognition tech-nique is given by Dougman [12]. Dougmans system isthe basis of all commercial systems developed so far.First, image is segmented using iteratively the follow-ing integro-differential operator [13]:

max(r,x0,y0)

∣∣∣∣∣Gσ(r) ∗∮(r,x0,y0)

I(x, y)

2πrds

∣∣∣∣∣ (2)

where I is the image and G is a Gaussian smooth-ing function. After normalization, a 2D version of Ga-bor filter is used for feature extraction, whereas theHamming Distance is used for matching two iris tem-plates [14, 15]:

HD =1

N

N∑i=1

Xi Y Yi (3)

The most important important phase is segmen-tation. In fact, a bad segmented image will not pro-vide good results in terms of false positive and falsenegative values. For this reason, several different ap-proaches have been proposed for segmantation step.


E-ISSN: 2224-3402 62 Volume 11, 2014

A first alternative approach proposes to guide in someway Daugman’s integro-differential operator. In fact,a global approach can be very expensive in terms ofcomputational costs. An alternative way is to use theHough transform to search eye’s position in any im-age with less accuracy and then ,when a probably lo-cation of the iris is obtained, refine the search withDaugman’s operator[17]. The author himself also pro-poses the use of Independent Component Analysis forfeature extraction and coding, while for the match-ing phase Euclidean distance is used. Further researchand other tests in less controlled enviroments revealedanother problem in Daugman’s operator: the integro-differential operator fails in segmentation in presenceof some artifacts in image, artifacts due to some in-strumental problem. The presence of these problemsreveals that using a bad-calibrated instrument or in-appropriate instruments can bring to mismatched irislocalizations and can compromise recognition quality[18]. For this reason, it is better to combine Daug-man’s operator with Hough transform, which can helpto solve this problem, with other methods of pre-processing. Once image is normalized with Daug-man’s Rubbersheet model, the encoding is done inTisse’s work with Gabor’s filter and Hilbert transform.The template obtained is compared with another tem-plate with Hamming distance. The entire coding in Clanguage of this system also increases system’s speed.Another approach that was successfully tested con-sist in apply to the entire iris image, correctly ex-tracted, normalized and scaled, a diadic wavelet rep-resentation [19]. Even in this approach, differencebetween templates is evalued with some distancemethos, like Hamming distance, euclidean-based dis-tance or wavelet-derived measures.In his work, Lim [20] make use of Haar wavelet totransform a normalized iris image, obtaining a tem-plate that is only 87 bit long. The learning and de-cision strategy is the LVQ algorithm, basis of SOMneural networks.A very similar strategy were analyzed by Sondhi andothers [21]. In his work, Sondhi used a more com-plex iris localization, paying a lot of attention to pre-processing steps to eliminate artifacts and noise onimage, then recognition step is performed with a Self-organizing Map. In his work were given accurate de-tails on various settings of the neural network and onhis learning system. His work also values system per-formance considering the machines that executes theprogram, to give an evaluation of the computationalcosts of the procedure.Another important tecnique that lets to train a systemwith a few number of examples is multichannel Gaborfiltering [22]. It is demonstrated that extracting an ar-ray with this filter and using a nearest neighboor clas-

Figure 3: Iris and template

sificator that uses weighted euclidean distance raisessystem’s performance from 80% with a train set madeby a single example to 99% with a train set made byfive examples per individual.Haar transform, with Hamming distance, can be usedto perform an iris template comparison with a multi-scale analysis in two dimensions [23]. Even in thiscase, it is demonstrated that a template made by 87bit is sufficient to distinguish a person from an im-postor with a precision equal to 97%. Segmentationis performed starting from the center of the pupil andproceeding to the edge of the iris, excluding eyelids,eyelashes and sclera. Moreover, it is proven that theuse of optical lenses does not affect the system.There is another great problem: memory waste. It isdemonstrated that 87 bit are sufficient to recognize aperson, but the various images that are outputted fromevery step of computation are saved in extended for-mat (TIFF, for example), and memory waste can be alittle matter for few images, but can be a great problemwhen a lot of images are stored. Moreover, it is im-portant to remark that a lot of systems adopt a client-server architecture, then an extended format can beexpensive in a network system. This is another issueof these system. Daugman himself, which gave a lotof contributes for iris recognition, also has consideredimage compression methods [24]. In details, whenimage are saved, an uniform background is applied tothe images to suppress everything external the eyelids,then image is saved using the JPEG and JPEG2000format, as in figure 3. Notice that these formats uses alossy compression. Daugman’s work shows that evenwith an heavy compression a system lowers his per-formance about 2-3%, with minimum influence on re-sults. However, Daugman says that these standardsare open standards, so they are not subject to any lim-itation.

More recently, Liu et al. [16] presented an irissegmentation approach based on Hough Transformthat is effective for determining the pupil regionas well as the limbic boundary and iris region. In


E-ISSN: 2224-3402 63 Volume 11, 2014

Figure 4: An Iris from CASIA database

addition it also allows improved eyelid detectionusing linear Hough Transform.All these evolutions helped to realize fast and ef-ficient methods, with the aim to use biometrics onlow performance an capacity devices. An importantresearch where given with a proposal of iris recogni-tion system on smartphone. Smartphone technologytoday is very evolved, and is very important to equipthese devices with data protections systems, becausenowaday on smartphones people stores not onlynumbers and SMS, but also e-mail, credit cards andso on. The most recent iPhone made by Apple hasa fingerprint recognition system, originally availableon LG-KP3800 made by LG Electronics, that wasviolated few days after presentation. Voice and facerecognition was inserted in last Samsung devices, assaid before. Even iris recognition were adapted fora mobile version [25]. In their work, Park and Rheeanalyze speed of a system on a low power processor(only 200 MHz). They choose iris recognition insteadfingerprint because the latter one needs dedicateddevices for acquisition and processing. In this casealso they adapted the system to execute it without anyproblems on an ARM CPU.Other contributes of similar types exists [26]. Anaccurate analysis of different solutions brings todefinition of a lot of important characteristics forthis system, non only considering the algorithm, butalso considering limitations of devices. It is alsointeresting that this study can be adapted to otherbiometrics.

In this paper a new approach for iris recognitionis proposed, based on a novel feature encoding andmatching strategy. Instead of using entire iris tem-plate, images will be processed to find most stableregions of an iris, considering twenty image for au-

thor. Then, a ranking of these regions will be used toevaluate stability of various iris sectors, and a certainnumber of sectors will be used for matching phase,according to a well-defined similarity measure. Thepaper is organized as it follows. Section 2 reports sys-tem description. Section 3 presents the experimentalresults, carried out on the CASIA database [29].The conclusion of the work and some considerationabout future directions for research are discussed inSection 4.

2 System DescriptionThe system proposed in this paper is schematically de-scribed in Figure 5. This section will address eachphase of the system, that will be described in detail.This section is organized as it follows: subsection2.1 shows the chosen segmentation procedure; sub-section 2.2 shows Daugmans Rubbersheet model fornormalization. Subsection 2.3 shows how Ssim wasemployed to evaluate stability analysis. Subsection2.4 shows the matching function adopted and the de-cision rule. Figure 5 shows an outline of the systememployed for this study.

2.1 SegmentationIn the segmentation step, the annulus of the iris im-age must be isolated from the rest of the eye image.For the purpose, it is necessary to localize the portionof the iris image derived from inside the limbus (theborder between the sclera and the iris) and outside thepupil. It is also important to include the portion of iristhat is not covered by eyelids (both upper and lower).In fact, eyelids can cover part of the limbus, and thepupillary boundary can be far less well defined. Ofcourse, pupil is typically darker than the iris. Depend-ing on pigmentation of skin and iris, eyelid contrast isquite variable. Eyelashes can make eyelids boundaryirregular. Therefore, any iris localisation method hasto be sensitive to a wide range of edge contrast androbust to irregular borders. In this paper, followingthe approach of Wildes [9], the Hough Transform hasbeen used for iris detection in eye image, according toequation:

x2c + y2c − r2 = 0 (4)

that is the well-known equation of a circle.This involves employing Canny edge detection to gen-erate an edge map. Gradients were biased in the ver-tical direction for the outer iris/sclera boundary [27].Figure 6 shows an example of segmentation. It is pos-sible to apply methods for handling both eyelids andeyelashes, to suppress and eliminate them from image.


E-ISSN: 2224-3402 64 Volume 11, 2014

Figure 5: The system implemented

Figure 6: Iris Segmentation by Hough Transform

Figure 7: Another example of an iris

Figure 7 also shows that another Hough Transform isrequired to detect and remove pupil from the images[28]. Sometimes Hough transform does not reach cor-rectly in his task, as shown in figure 8.

2.2 NormalisationNormalization keeps image unaltered and simplifiessuccessive steps, transforming original image in a dif-ferent one. As described by Daugman [12], patternmust be invariant to size and must consider size ofpupil and is position in iris image. Moreover, it isnecessary for correct execution of the recognition sys-tem to consider the non-linear transformation of pupilthat stretches iris. This is a very critical task. In factit is worth noting that pupil is not located at the centerof the iris, but it is slightly shifted towards nose. Thismeans that any mathematical model to normalize irismust consider this extremely important aspects. In lit-erature, several models exist [10, 12]. In this paper we

Figure 8: Example of mismatched Transform

Figure 9: Dougman’s Rubbersheet Model

consider Daugman’s Rubbersheet Model, representedby figure 9, that is the most complete and effectivemodel. The model uses projected pseudo polar coor-dinates. In this case the grid must not be necessar-ily concentric, in consideration of pupil displacement.The angle - the polar variable is clearly dimension-less, as the radial variable because, as stated by Daug-man, it ranges from pupil boundary to the limbus al-ways as a unit interval [0, 1]. Dilation and constric-tion of iris is modeled by this coordinates system asthe stretching of a homogeneous rubber sheet with thetopology of an annulus anchored on his outer perime-ter with tension controlled by an interior ring of vari-able radius. With this coordinates system corrects theelastic pattern deformation due to variations of pupildimension.

The rubber sheet model devised by Daugmanremaps each point within the iris region to a pair ofpolar coordinates. The remapping of the iris regionfrom (x, y) Cartesian coordinates to the normalised


E-ISSN: 2224-3402 65 Volume 11, 2014

Figure 10: An illustration of pupil displacement

Figure 11: Three examples of normalised irises

non-concentric polar representation is modelled as:

I(x(r, θ), y(r, θ)) −→ I(r, θ) (5)

with:

x(r, θ) = (1− r)xp(θ) + rxl(θ) (6)

y(r, θ) = (1− r)yp(θ) + ryl(θ) (7)

where:

• I is iris image,

• (x, y) are the original Cartesian coordinates,

• (r, θ) are the corresponding normalized polar co-ordinates,

• xp, yp and xl, yl are the coordinates of the pupiland iris boundaries along the θ direction.

It is important to notice that Dougman’s modelalso considers pupil’s displacement. Pupil is not per-fectly concentric with iris, as shown in figure 10.Pupil displacement in figure is accentuated respect thereal one.

Therefore, as Figure 11 shows, this model allowsthe remapping of an iris image on a fixed-size rectan-gular template, which is effective for feature matchingby Hamming distance [12] or other distance measures[9, 11].

2.3 SSIM Index for Region SelectionIn the literature iris recognition is performed by con-sidering the entire iris image. The idea of this In thispaper a new approach is considered which finds anduses for personal verification only the most effectivesectors extracted from the iris image. To evaluate thequality of sectors images, the low level statistical in-dex proposed by Wang and al. [30] is considered. Thisindex is called Ssim, and will be adopted to evaluatesimilarity between different images. Figure 12 showsan outline of the Ssim index, that evaluates every pixelof the image by considering it on the basis of lumi-nance, contrast and structure.

The luminance comparison is evaluated as it fol-lows:

l(x, y) =2µxµy + C1

µ2x + µ2y + C1(8)

Contrast comparison is evaluated by:

c(x, y) =2σxσy + C2

σ2x + σ2y + C2(9)

Structure comparison is evaluated as:

s(x, y) =σxy + C3

σxσy + C3(10)

whereC1,C2 andC3 are constants added to avoidzero values. These three formulas (equations 8, 9 and10) can be combined to obtain the Ssim index as fol-lows:

SSIM(x, y) =(2µxµy + C1)(2σxσy + C2)

(µ2x + µ2y + C1)(σ2x + σ2y + C2)(11)

This statistic is calculated on window sizes of8×8. The window can be displaced pixel-by-pixel onthe image but the authors propose to use only a sub-group of the possible windows to reduce the complex-ity of the calculation. This index has a range between-1 and 1, which is reachable only with identical im-ages. In this paper, Ssim is employed to create a rankof irises sectors. More precisely, for every author, theiris images are divided into sixteen equal-size sectors.Successively, for every sector the average value of theSsim index is computed for each sector and used torank the various sectors. It is worth noting that thisapproach gives a low rank to iris sectors that are usu-ally covered by eyelids and/or eyelashes. On the con-trary, other sectors receive a high rank. Than, select-ing a certain number of sectors, it is possible to applya similarity measure for the matching step.


E-ISSN: 2224-3402 66 Volume 11, 2014

Figure 12: SSIM scheme

2.4 Bayes decision rule induced measuresRecently, new measures for image matching havebeen derived from the Bayes decision rule and theireffectiveness has been demonstrated for some specificbiometric applications [31]. In order to use the mea-sures, some preconditions need to be satisfied:

• The conditional probability density functions ofall the classes are multivariate normal.

• The prior probabilities of all the classes areequal.

• The covariance matrices of all the classes areidentical to the covariance matrix of all samplesregardless of their class membership.

• The whitened pattern vectors in the Bayes deci-sion rule are normalized to unit norm.

It is possible to demonstrate that if these condi-tions are all satisfied, then the Bayes decision rule be-comes the Whitened Cosine (WC) [31]:

δ(U, V ) =(W tU)t(W tV )

‖W tU‖‖W tV ‖(12)

Moreover, it is worth noting that the fourth as-sumption is the most important. In fact, if this as-sumption is not satisfied, then the Bayes decisionrule became the Mahalanobis distance, that is a wellknown classifier, but with lower performance than theWC. The whitening transform is defined as [31]:

W = ΦΛ−12 (13)

where Φ is an eigenvector matrix and Λ is theeigenvalue diagonal matrix, obtained by factorizationvia SVD of the covariance matrix of all examplesΣ = ΦΛΦ′.In his work, Liu [31] kept the forth assumption andmodified the third one, defining two new similaritymeasure, modifying the whitening transformation ma-trix. The two new measures are:

• The PRM Whitening Cosine Transform, that isnamed PWC, is defined as:

Wp = Φ∆−12 (14)

where:

∆ = diag{σ21, σ22, . . . , σ2d} (15)

with:

σ2i =1

L

L∑i=1

1

Nk − 1

Nk∑j=1

(y(k)ij −mki

)2(16)

being L the number of classes that needs to beclassified.

• The Whitin-class Whitened Cosine Transform,named WWC, is obtained by factorizing thewithin-class scatter matrix with SVD, as follows:

Ww = Φw∆− 1

2w (17)

The two measures along the Whitened Cosinewere adapted to be used on a binary classificationproblem instead of a multiclass problem.


E-ISSN: 2224-3402 67 Volume 11, 2014

Figure 13: The experimental process

3 Experimental ResultsThe experiment tests were carried out on the iris im-ages of the section ’Lamp’ of the CASIA database[29]. Each iris image has been processed as previ-ously described, according to Figure 13. Then, thestability analysis has been considered to sort regionsaccording to their stability ranking.

In the tests, the four most effective (top-ranked)sectors of eyes were used for the matching phase. Forthe purpose, the effectiveness of the three measures(WC, PWC and WWC) were compared by consider-ing the Area Under Curve (AUC) of the Receiver Op-erating Characteristic (ROC). An example is shown inFigure 14, 15 and 16.Figure 14 and figure 15 report two examples of ROCtests, excellent case and good case, while figure 16report an example of poor test.

From AUC values reported in Table 1, it resultsthat WWC reaches in almost cases an area that is notstatistically different from 0.50, which indicates a per-formance similar to that of a random classifier. Thismeans that WWC fails to classify correctly data. Con-versely, WC and PWC are much more effective thanWWC. In some cases they provide very good classi-fication results. A more detailed comparison amongWC and PWC is performed in terms of FAR and FRRvalues.

From AUC values reported in Table 1, it resultsthat WWC reaches in almost cases an area that is notstatistically different from 0.50, which indicates a per-formance similar to that of a random classifier. Thismeans that WWC fails to classify correctly data. Con-versely, WC and PWC are much more effective than

Figure 14: An example of excellent test

Author WC PWC WWC1 0.68 0.62 0.502 0.89 0.89 0.533 0.95 0.95 0.994 0.71 0.71 0.505 0.62 0.63 0.506 0.61 0.61 0.637 0.81 0.82 0.538 0.57 0.58 0.529 0.77 0.76 0.5110 0.50 0.50 0.50

Table 1: AUC values of ROC curves. An AUC valueclose to 1 represents an excellent case of test, while avalue close to 0.50 represents a fair test, similar to arandom classifier.


E-ISSN: 2224-3402 68 Volume 11, 2014

Figure 15: An example of good test

Figure 16: An example of poor test

WWC. In some cases they provide very good clas-sification results, while author number 10 does notreach tests different than the failed one. A more de-tailed comparison among WC and PWC is performedin terms of FAR and FRR values.

Table 2 reports FAR values while Table 3 reportsFRR values for WC and PWC. As it is possible to ob-serve, variations between values suggest that probablythere is no difference in terms of discriminant powerbetween WC and PWC. This consideration was con-firmed by the T-test, that revealed that there is no sta-tistical significant difference between the two mea-sures, both in terms of FAR and FRR. It is possibleto conclude that the Whitened Cosine classifier andthe PRM Whitened Cosine have both similar perfor-mances on regional analysis of irises sectors. Also,in figure 17 and 18 are reported boxplot for FAR and

Author WC PWC1 23.3 20.32 5.7 5.73 6.9 6.94 25.4 25.45 37.4 36.56 48.9 48.97 27.3 26.58 26.3 25.49 24.8 25.7

Table 2: FAR table.

Author WC PWC1 53.5 50.02 15.5 15.53 10.0 10.04 31.5 31.55 41.0 40.06 30.5 30.57 20.0 21.18 45.0 44.19 24.0 24.9

Table 3: FRR table.

FRR values. In next section it is reported a short dis-cussion about experimental results of this study.

4 ConclusionIn this work a regional approach for iris recognitionis presented. Iris recognition is performed only con-sidering a reduced part of the iris image. For the pur-pose, a new technique is proposed to select the mostvaluable, user-dependent, regions of iris images forrecognition purposes. In addition , three Bayes-basedsimilarity measures have been considered for iris im-age matching: the Whitened Cosine (WC), the PRMWhitened Cosine Transform (PWC) and the Whitin-class Whitened Cosine Transform (WWC) measures.The experimental results demonstrate the foundationof the proposed approach, although much more re-search is necessary to evaluate its validity under realworld conditions.Another interesting enhancement for this new way ofrecognition should be located in the stability analysisstep. For example, it should be evaluated the use ofcoregistration techniques to find the best alignment ofnormalised irises patterns, then stable sectors searchwill be performed. Moreover, other indexes like SSIMcan be employed, or other distances measures.


E-ISSN: 2224-3402 69 Volume 11, 2014

Figure 17: FAR values boxplot

Figure 18: FRR values boxplot

Acknowledgements: The research was supported bythe University of Bari and in the case of the third au-thor, it was supported for his Master’s Degree Thesis.

References:

[1] L. Berggren, Iridology: A critical review, ActaOphthalmologica 63(1): pp. 1-8, 1985.

[2] http://www.ccip.govt.nz/newsroom/information-notes/2005/biometrics.pdf

[3] C. Ostaff, Retinal Scans Do More Than LetYou In The Door, available at: http://phys.org/news6134.html

[4] Z. Zhi, E. Y. Du, N.L. Thomas, E. J. Delp,A New Human Identification Method: ScleraRecognition, Systems, Man and Cybernetics,Part A: Systems and Humans, IEEE Transac-tions, May 2012, Vol. 42, Issue 3, pp. 571-583.

[5] M. Keegan, Manchester Airport eye scan-ners scrapped over delays, available at:http://www.manchestereveningnews.co.uk/

[6] http://www.youtube.com/watch?v=1SCZzgfdTBo

[7] http://www.cl.cam.ac.uk/

˜jgd1000/afghan.html

[8] W. Lance, Iris recognition gadget eliminatespasswords, CNET, May 2011

[9] R. Wildes, Iris recognition: an emerging bio-metric technology, Proceedings of the IEEE, Vol.85, No. 9, 1997, pp. 1348-1363.

[10] W. Boles, B. Boashash, A human identificationtechnique using images of the iris and wavelettransform, IEEE Transactions on Signal Pro-cessing, Vol. 46, No. 4, 1998, pp. 1185-1188.

[11] Y. Zhu, T. Tan, Y. Wang, Biometric personalidentification based on iris patterns. Proceedingsof the 15th International Conference on PatternRecognition, Spain, Vol. 2, 2000, pp. 2801-2804.

[12] J. Daugman, How Iris Recognition Works, Pro-ceedings of 2002 International Conference onImage Processing, Vol.1, 2002, pp. 1-10.

[13] S. Mahboubeh, B. S. Puteh, B. I. Subariah,R.K. Abdolreza, Fast Algorithm for Iris Lo-calization Using Daugman Circular Integro Dif-ferential Operator, Proc. of the 2009 Interna-tional Conference of Soft Computing and PatternRecognition (SOCPAR ’09), 2009, pp. 393-398.


E-ISSN: 2224-3402 70 Volume 11, 2014

http://www.ccip.govt.nz/newsroom/information-notes/2005/biometrics.pdf



http://phys.org/news6134.html

http://phys.org/news6134.html

http://www.manchestereveningnews.co.uk/news/greater-manchester-news/manchester-airport-eye-scanners-scrapped-682100

http://www.youtube.com/watch?v=1SCZzgfdTBo

http://www.youtube.com/watch?v=1SCZzgfdTBo

http://www.cl.cam.ac.uk/~jgd1000/afghan.html

http://www.cl.cam.ac.uk/~jgd1000/afghan.html

[14] J. Daugman, Results from 200 billion iriscross-comparisons, University of Cambridge,Computer Laboratory, UCAM-CL-TR-635, June2005, pp. 1-8.

[15] M. Vatsa, R. Singh, P. Gupta, Comparison ofIris Recognition Algorithm, IEEE ICISIP, 2004,pp. 354-358.

[16] C. Liu, A. Verma, J. Jia, Iris recognition basedon robust iris segmentation and image enhance-ment, Int. J. Biometrics, Vol. 4, No. 1, 2012, pp.56-76.

[17] Y. Huang, S. Luo, E. Chen, An efficientiris recognition system, Proceedings of the FirstInternational Conference on Machine Learningand Cybernetics, Beijing, 4-5 November 2002,450-454

[18] C. L. Tisse, L. Martin, L. Torres, M. Robert,Person identification technique using human irisrecognition. In Proceedings of Vision Interface,May 2002 (pp. 294-299)

[19] D. de Martin-Roche, C. Sanchez-Avila, R.Sanchez-Reillo, Iris recognition for biometricidentification using dyadic wavelet transformzero-crossing. In Security Technology, 2001IEEE 35th International Carnahan Conference,October 2001, pp. 272-277.

[20] S. Lim, K. Lee, O. Byeon, T. Kim, Efficientiris recognition through improvement of featurevector and classifier. ETRI journal, 23(2), June2001, 61-70.

[21] S. Sondhi, S. Vashisth, A. Gaikwad, A. Garg,Iris pattern recognition using Self-OrganizingMap, MPGI National Multi Conference 2012(MPGINMC-2012) ,Advancement in Electronics& Telecommunication Engineering,Proceedingspublished by International Journal of ComputerApplications (IJCA), 7-8 April, 2012, pp. 12-17

[22] L. Ma, Y. Wang, T. Tan, Iris recognition basedon multichannel Gabor filtering. In Proc. FifthAsian Conf. Computer Vision, January 2002,Vol. 1, pp. 279-283.

[23] J. M. Ali, A. E. Hassanien, An iris recogni-tion system to enhance e-security environmentbased on wavelet theory. AMO-Advanced Mod-eling and Optimization, 5(2), 2003, 93-104.

[24] J. Daugman, C. Downing, Effect of se-vere image compression on iris recognition per-formance. Information Forensics and Security,IEEE Transactions on, 3(1), March 2008, 52-61.

[25] K. R. Park, D. W. Rhee, Real-time iris lo-calization for iris recognition in cellular phone.In Software Engineering, Artificial Intelligence,

Networking and Parallel/Distributed Comput-ing, 2005 and First ACIS International Work-shop on Self-Assembling Wireless Networks.SNPD/SAWN 2005. Sixth International Confer-ence, May 2005, pp. 254-259.

[26] H. Lu, C. R. Chatwin, R. C. Young, Iris Recog-nition on Low Computational Power Mobile De-vices, Biometrics - Unique and Diverse Appli-cations in Nature, Science, and Technology, Dr.Midori Albert (Ed.), ISBN: 978-953-307-187-9,InTech, 2011.

[27] P. Kovesi, MATLAB Functions for Com-puter Vision and Image Analysis availableat: http://www.cs.uwa.edu.au/pk/Research/MatlabFns/index.html

[28] L. Masek, Recognition of Human Iris Patternsfor biometric Identification, BEng Thesis, TheUniversity of Western Australia, Western Aus-tralia, 2003.

[29] CASIA Iris Image Database, available at:http://biometrics.idealtest.org/

[30] Z. Wang, A.C. Bovik, H.R. Sheikh, E.P. Si-moncelli, Image quality assessment: From errorvisibility to structural similarity, IEEE Transac-tions on Image Processing, vol. 13, no. 4, Apr.2004, pp. 600-612.

[31] C. Liu, The Bayes Decision Rule Induced Sim-ilarity Measures, IEEE Transactions on PatternAnalysis and Machine Intelligence, vol. 29, no.6, June 2007, pp. 1086-1090.


E-ISSN: 2224-3402 71 Volume 11, 2014

http://www.cs.uwa.edu.au/ pk/Research/MatlabFns/index.html

http://www.cs.uwa.edu.au/ pk/Research/MatlabFns/index.html

http://biometrics.idealtest.org/

Documents

Regional Approach for Iris Recognition