JESC : January-June 2012, Volume 3, Number 1, pp. 109-123

1Computer Department, PDEA’S College of Engineering, Pune University,Ganesh Khind Road, 411005, Pune, India, E-mail: pmagarkar@rediffmail.com

2E&TC Dept., SGGS IET, Vishnupuri, Nanded 431606, India,Email: sntalbar@yahoo.com

1. INTRODUCTION

With introduction of Internet applications and automation, theamount of information is increasing far more quickly than our ability

to process it. Iris recognition is gaining large-scale focus in biometricsrecognition systems. The IRIS recognition system has rapidlyadvanced from prototype stage to the patented algorithm in recentyears due to higher accuracy and reliable recognition[1]. Theconventional iris recognition system has steps that include: imagecapture, pre-processing, iris detection and feature extraction, iriscode formation and database formation. This is followed by templatematching for iris recognition. One of the popular template matchingmethods is using Hamming Distance (HD) or HD method[1][2][3].The number of comparisons carried out for template matchingdepends first; upon the database size; and second, the bitcomparisons needed for HD calculation [2]. Due to broadspectrumreliability of iris recognition the ISO standard 19794-6 hasincorporated the iris data.

Wildes presented use of normalized correlation on iris imagethat can be used in identification process. Boles and Boashinvestigated two dissimilarity functions: learning and classificationresulting in fast processing but high error rate. Tan et al. proposednearest feature line with the constraint of eye lead occlusion problem.Ma et al. presented two different techniques such as XOR operatorand weighted Euclidean and hamming distances but proved failurein the case of occluded iris. Lim et al. presented LVQ neural networkfor matching purpose but resulted in poor recognition rate[3][4][5][6].

Figure 1: Conventional Iris Recognition System

The IRIS code is a binary information. Number of theory existsfor similarity matching of binary information. In information theory,the Hamming distance between two strings of equal length is thenumber of positions at which the corresponding symbols aredifferent. The correlation coefficient, ( ) indicates the strength anddirection of a linear relationship between two random variables. Ininformation theory and computer science, the Levenshtein distanceis a metric for measuring the amount of difference between twosequences. It can be considered a generalization of the Hammingdistance, which is used for strings of the same length and onlyconsiders substitution edits.

This research paper proposes a novel algorithm by organizingstored database information of IRIS codes that will reduce theredundant search iteration by method of elimination. The remainingpaper is organized in four sections. The section II gives the workoutline and section III discusses the research algorithm in details.The section IV compares the algorithm with HD algorithm andsection V presents the conclusions and future work.

Figure 2: Typical iris Sample

2. WORK OUTLINE

Conventional steps carried out by any iris recognition system areshown in Figure 1[4]. The applications using iris recognition systemsuch as identity verification system demand very large number ofcomparisons [3]. Equation (1) presents hamming distancecalculations that are done for determining the deviation of iristemplate with the database iris codes.

codeA codeB maskA maskBHD

maskA maskB� �

� (1)

With growth in technology and Internet data capture and storagehas become distributed. The bit granularity of the capture has setnew challenges of variable size IRIS codes. Variable size IRIS codesand voluminous size of database searching for IRIS code similaritymatching are becoming major challenges rather than areas of theiris that are obscured by eyelids, or by eyelashes or reflection on irisare detected and prevented and applicable research results frominfluencing the iris comparison through bit-wise mask function[5][7][8][9]. The iris code contains the phase data that is XORed todetect bit agreement and deviations. The bits to be considered arefirst selected by ( ) ANDing with mask functions associated withtest iris (A) and the template iris (B). The proposed algorithmreorganizes the iris code useful for fast comparisons. The irisdatabase can contain training set of iris codes for a person resultinginto huge number of iris codes in the database. The proposedalgorithm applies method of elimination. This results in eliminatingiris codes before comparison that are not favorite choices. The iriscode set shows regional changes due to very common disease likecataract. This is resulting in failure of conventional iris coderecognition. The proposed reorganization of iris code has applicationin recognition of iris with cataract. The selection of most favorableiris templates amongst number of available templates is the majorof the research. The research uses standard iris images from UBIRIS,Bath university iris database and CASIA iris database samples.Figure 2 Shows the typical iris sample. Iris structures are unique forevery person hence it is considered as very important biometricsmeasure. The iris structure is elliptical in shape. The Figure 3 shows

the elliptical model of iris with radius r. The iris being circular canhave 2 or 360 samples to form the iris code.

Figure 3: Model of the iris Structure

Figure 3 shows the elliptical model of iris with radius r and irisexpanded using 360 samples of width r, the radius of the circle. Aninner circle carries all low pass components. Let r1 be the radius ofthe inner circle then (2) shows the valid iris content of width r2 usefulfor the iris code. The r1 can vary with the intensity of the focus orlight falling on the iris. Depending upon the resolution the iris codesare extending from 80 bits to 140 bits. This is resulting into numberof very large comparisons.

r2 = r – r1 (2)

The proposed algorithm sub-divides the 2 radians iris band spacein to 8 sub-bands of /4 radians each as shown in Figure 4. Eachsample is tested for errors due to occluded due to eyelids, intensity

Figure 4: 8 Sub-Band Arrangement of Iris Code

focus points due to illumination sources or patches due to cataract.The samples then tagged with sample number and arranged in witherrorless sample at the top. This is done because if the errorless samplefail to correlate the remaining samples are not tested to confirm theiris recognition. This resulted in faster selection time and reduces thelead-time to select the correct iris template for comparison.

3. ALGORITHM

The proposed algorithm uses conventional steps of image processingsuch as image acquisition, iris localization, normalization sinceresearch is focused on IRIS code organization and adaptive searchingand associated similarity matching. Major focus of algorithm is IRIScode organization and searching rather than Iris capture andconditioning. Following section takes brief review of steps used inimplementation.

3.1 Image Acquisition and Preprocessing

This is one of the important factors for obtaining a good result. Theexisting iris databases internationally available are used for thepurpose of testing the algorithm. A good and clear image eliminatesthe process of noise removal and also helps in avoiding errors incalculation. In this case, computational errors are avoided due toabsence of reflections, and because the images have been taken fromclose proximity. The Infra-red light can be used for illuminating theeye, to avoid any specular reflections.

3.2 Iris Localization

The iris carries essential information required for biometricsrecognition It lies between the sclera and the pupil. The iris isseparated from the eye image and the pupil is filtered using lowpass filter. The iris inner and outer boundaries are located by findingthe edge image using the Canny edge detector. The Canny detectorworks with three stages, such as 1) finding the gradient, 2) non-maximum suppression and 3) hysteresis thresholding [3]. The Wilfesproposed thresholding in a vertical direction only to reduce theinfluence due to the eyelids. This can reduce the pixels on the circular

boundary. The Hough transform successfully localizes the boundary.Using the gradient image, the peaks are localized usingnonmaximum suppression. The next step, hysteresis thresholding,eliminates the weak edges below a low threshold. The thresholdvalues were found by trail and error, and were obtained as 0.3 and0.21. Edge detection is followed by the finding the boundaries ofthe iris and pupil. The Hough transform is used for detecting theparameters of geometric objects. The computations in Houghtransform are reduced having eightway symmetric sample pointson the circle for every search point and radius. The proposedalgorithm uses it to find the circles in the edge image.

3.3 Image Normalization

After segmenting the iris, it is normalized. This is done to enablegeneration of the iris code organized into 8 parts or sub-bands of /4 each. Since the iris orientation changes from person to person, it isrequired to normalize the iris image. The normalization givescommon boundaries of the image and makes easier to process. Thisprocess involves unwrapping the iris and converting it into its polarequivalent as shown in Figure 3. It is done using Daugmans Rubbersheet model. The center of the pupil is considered as the referencepoint and a remapping formula is used to convert the points on theCartesian scale to the polar scale [3].

3.4 Iris Adaptive Encoding

The final process is the generation of the iriscode. For this, the mostdiscriminating feature in the iris pattern is extracted. The phaseinformation in the pattern only is used because the phase angles areassigned regardless of the image contrast. Amplitude informationis not used since it depends on extraneous factors. The iris codegeneration is carried out on the texture-processed image for betterperformance as shown in Figure 4. Daugman has suggested the irisphase extraction information using 2D Gabor wavelets. It determineswhich quadrant the resulting phaser using the wavelet. Gabor filtersare used to extract localized frequency information. The log-Gaborfilters are more widely used for coding natural images. Statistics ofnatural images indicate the presence of high frequency components.

Since the ordinary Gabor filters under represent high frequencycomponents, the log filters become a better choice. The proposediris extraction method is more simpler than the LogGabor andefficient for frequency component processing. The iris image goesthrough the process shown in and sub banding is applied. TheFigure 5 shows the difference of Iris code generation output of theconventional iris recognition system and the proposed model.

Figure 5: Iris Code Formation Methodology

3.5 Tagged Index IRIS Code Organization

The iris codes are formed after sectoring or sub-banding the IRIS.IRIS code is a collection of binary bit sequence in the range 0, 1. LetBk = i0, i1, i2 ... in be a sequence of binary bits in a kth sector where k isin the bounds 0, 7 since 8 sub-band sectors are formed using rubberband model of IRIS data. The transition-state, in–1 to in is in the binaryrange [0, 1].

The Table 1 shows the transition states between two consecutivebits of iris code. The iris code being binary [0, 1] can take four statesas shown in Table 1. The transitions 0 0 and 1 1 show no changein state hence coded as 0 while transition state 0 1 shows positivechange or 1 0 shows negative change hence interpreted as 1 and– 1 respectively.

This concept has grown further to be used with three consecutivebits as show in Table 2. The quantization index is associated as a tagof the respective sector or sub-band. The IRIS codes are organizedin the IRIS code table using grouping based on coding quantizationindex such that index has a range [–2 –1 –1 0 0 1 1 2]. This is resultedin to a search tree balanced at [–2, 2] as shown in Figure 6. When theIRIS code is generated, the coding quantization index is calculatedand accordingly left branch or right branch is selected. This isresulted into 50% improvement over the sequential code search periteration using HD method. The algorithm searches only those IRIScodes having same Tagged index using HD method per IRIS sector.

Table 1Transition State Table

(n – 1)th State nth State Transition0 0 0

0 1 1

1 0 –1

1 1 0

Figure 6: IRIS Code Search Tree

F(Sn–1, Sn) = {0 : 0 = 0; 0 : 1 = 1; 1 : 0 = – 1; 1 : 1 = 0} (3)

Let f be a function that operate on bit data such that the Functionf returns code quantization index for each IRIS code sub-band asper (3). As shown in Figure 5 the coding quantization index of IRIScode sub-band is used as tagged index and before going for HDanalysis of sub-band. The tagged index are searched for mostfavorable IRIS sub-bands.

Table 2Tagged Index Organization Table

(n–1)th State nth State (n+1)th State Transition

0 0 0 00 0 1 10 1 0 10 1 1 21 0 0 –21 0 1 -11 1 0 -11 1 1 0

3.6 IRIS Code Sub-band Similarity Matching

Comparison of the bit patterns generated is done to check if the twoirises belong to the same person. Calculation of Hamming Distance(HD) is done for this comparison. HD is a fractional measure of thenumber of bits disagreeing between two binary patterns. Also, dueto varying bit granular density and distributed IRIS data captureresulting in variable size IRIS codes varying from 48 bits to 128 bits.Conventional HD method suffer heavily due to fixed size codes.Since this code comparison uses the iriscode data and the noisy maskbits, the proposed modified form of the HD algorithm is illustratedusing A and B templates. Let A and B are the templates andassociated masks for the iris code under test and the iris code fromthe data bank respectively. Let S be the sample space S = s0, s1, s2...s7. A(S) and B(S) are the sample spaces for iris codes A and Brespectively. In general, if A and B are identical then the function in(3) will return success or else it will show the failure.

7

0

( ( ), ( )) ( ) ( )i

i ii

F A S B S A S B S (4)

Practically, function f(A(s), B(s)) works till success and the worstcase of success is all 8 samples are required for testing the iris code.Equation (3) further extended to the hamming distance calculationsas shown in (4).

i i i i i

i

i i

S S S S S

S

S S

code A codeB maskA mask A maskBHD

maskA maskB

� � �

�(5)

Figure 5 shows the difference in methodology of iris codesubband organization for the iris under test. The iris code slice aresearched using sub-band tags associated with circular link list. Figure6 shows search tree organization. The iris code is tested slice wise.This helps eliminating the iris codes not eligible using (4). Themethod of the iris code sub-band matching and selection is shownin Figure 7.

Figure 7: Circular Ring List Organization for iris Sub-code

Figure 8: Test Data IRIS used for IRIS code Sub-banding

Table 3 shows the class descriptor used for IRIS code data.FOURCC is a 4 character data structure used to name the class. The”iris” is a container class which contains eight data classes in it, col0,col1,... col7 respectively which hold IRIS code subband data. Thedata classes hold IRIS codes to be searched. This class in algebraicterms of set theory is represented as shown in (6).

{ , " ", { }|0 7|}iS SIZE iris col i (6)

Being class multiple instances can coexist and the respectiveinstance destructor makes memory free hence removing the memorylocking overheads.

Table 3Tagged Index Organization Table Descriptor

IRIS tbl Descriptor Size 32 bitsIRIS tbl Signature FOURCC = iris<< TAG data >>

Descriptor Size 32 bitDecriptor Signature FOURCC = col0<< data >>

Descriptor Size 32 bitDecriptor Signature FOURCC = col1<< data >>

Descriptor Size 32 bitDecriptor Signature FOURCC = col2<< data >>

�Descriptor Size 32 bitDecriptor Signature FOURCC = col7<< data >>

4. DISCUSSION

The legacy methods for iris recognition mainly focus on featurerepresentation and matching. These methods do not undertakeresearch of selecting favorable iris code for matching to avoid

searching delays during finding the match. Also IRIS code data isnot stored as class descriptor to take advantages of object orientedtechnology, the recent advances in technology. Hence thesetechniques suffer in various ways of storage efficiency, searchefficiency and data communication efficiency. These techniques donot focus on distributed data organization to accommodate datadistributed all over the globe through Internet. Typically, data likeIRIS needs global distributed attention for important applicationslike security. Therefore, we only analyze and compare the accuracyand efficiency of feature representation and matching of thesemethods. The methods proposed by Daugman[1][2], Wildes et al.[3], Boles and Boashash [3] are probably the best-known. Thesecharacterize local details of the iris based on phase, texture analysisand zero-crossing representation respectively. Figure 8 present adetailed Test data used of IRIS code sub-banding and searchingmethodology of using trained images. The reorganization of CASIAIris Database to support distributted data in Iris information cloudused for the purpose of comparison. Figure 9 show the iris code bitsearch cost comparison results advocating the efficiency of theproposed algorithm.

Figure 9: Search Cost Comparison for Test Data Iris Codes on Learning Set Simulation

5. CONCLUSION

The proposed method performs better searching of IRIS codes hencespeedy iris recognition as compare to the conventional systems. Anovel organization of iris code presents the faster searching of iriscode over the conventional methods thereby saving on the searchingtimes. Being defined as a class descriptor, it has become portable andany service on Internet can import it. In recent world, security hasbecome a global challenge and speedy availability of secured datahas become a prime challenge. IRIS recognition is one of such primechallenge. The proposed IRIS code organization and searching beinga class descriptor supports distributed cloud architecture ”IRIS codecloud”. Being container class descriptor this data can be embedded inany MPEG-4 image providing high quality security andtransportability. The destination application can identify thedescriptor signature and encode the IRIS code information henceadaptive in nature. The resulting in easy instantiation and availabilityof IRIS code information. The applications proposed by Daugmanuse millions of iris code comparisons in a system such a reservationsystem, airport security system with central database rather than adistributed database [1][2]. The proposed algorithm since faster iniris code matching due to reduction in redundant searches isadvocated for such time critical systems. Due to slice organization ofiris code, even a person developed cataract can be efficiently detectedsince failure iris code slice is only 1 of 8. Hence the proposed algorithmis useful in adverse cases of impurity in iris codes.

References[1] J.G. Daugman, “Biometric Personal Identification System Based on

IrisAnalysis”. US Patent 5291560, 1994.

[2] J.G. Daugman, “Probing the Uniqueness and Randomness of Iris Codes:Results from 200 Billion iris Pair Comparisons”, in Proc. IEEE, 94(11), 2006,pp. 1927–1935.

[3] A. Poursaberi and B.N. Araabi, “Iris Recognition for Partially OccludedImages: Methodology and Sensitivity Analysis”, EURASIP Journal on Advancesin Signal Processing, 1, 2007.

[4] X. Li, and M. Xie, “A Novel Algorithm of Human Iris Recognition”, in Proc.ISCIT 2005, 2005, pp. 1190–1193.

[5] A.K. Jain, “Patric Flynn and Arun Ross”, Hand Book of Biometrics. SprngerPress ISBN-13 : 978-0-387-71040-2, pp. 71–90.

[6] Li Ma, Tienutan, Tunhong Wang and Dexin Zang, “Personal IdentificationBased of Texture Analysis”, IEEE Trans. on Pattern Analysis and MachineIntelligence, 25(12), pp. 1519–1533, 2003.

[7] K. Mayazawa, K. Ito, and T. Aoki, “An Efficient Iris Recognition AlgorithmUsing Phase Based Image Matching”, in Proc. Porc. IEEE 2005, 2005.

[8] X. Liu, K. Bowyer and P. Flynn, “Experimental Evaluation of Iris Recognition”,in Proc IEEE 2005 (CVPR05).

[9] S. Elsherief, M.Allam, and M. Fakur, “Biometric Personal Identification Baseon Iris Recognition”, in Proc. IEEE 2006, 2006, pp. 208–213.