Improving Iris Identification using User Quality and Cohort Information

Transcription

1 Improving Iris Identifiction using User Qulity nd Cohort Informtion Arun Pssi, Ajy Kumr Biometrics Reserch Lbortory Deprtment of Electricl Engineering, Indin Institute of Technology Delhi Huz Khs, New Delhi 0 06, INDIA Abstrct Iris is one of the most distinguishble fetures of humn body, which remins firly stble throughout the lifetime of n individul. This mkes iris recognition one of the most relible methods for biometric bsed identifiction. This pper investigtes new technique to improve the performnce of the system by using cohort informtion nd user-qulity s the weight in the mtching. The proposed pproch uses the cohort informtion t the decision stge s cscded clssifiers. However, the second stge is only used if the first stge clssifier is uncertin of its decision. The experimentl results from the decision-level clssifiers combintion re presented, which show tht the cscded clssifiction system significntly outperforms the single clssifier, especilly t lower vlue of which is most likely to be the operting point for ny system. This pper lso proposes new pproch to scertin the user-qulity (iris) nd illustrtes its usge in the performnce improvement.. Introduction The iris is the colored nnulr ring which surrounds the pupil. It hs extremely rich texture contributed by stripes, pits nd furrows which offer relible nd unique personl identifiction. The iris texture is highly unique, even in cse of identicl twins nd even between the left nd right eyes of n individul. Since the dimensionlity of iris texture is very high, recognition decisions cn be mde t confidence level high enough to support relible nd rpid serches throughout extremely lrge sized dtbses. 2. Prior work The most promising iris recognition pproch is presented by Dugmn [3]-[4] using 2D Gbor wvelets for encoding the iris. Wildes s [6] hs used the bnd-pss decomposition derived from ppliction of Lplcin of Gussin filters, implemented in the prctice by the Lplcin Pyrmid. The degree of similrity is evluted with bse on normlized correltion between the cquired nd dtbse representtions. The lgorithm proposed by Li M et l. [5] chrcterizes key locl vritions for iris recognition. The locl shrp vrition points, denoting the ppering or vnishing of importnt imge structure, were utilized to represent the chrcteristics of the iris. Tisse et l. [8] proposed grdient decomposed Hough trnsform nd 2D Hilbert trnsform to extrct pertinent informtion from iris texture. Boles [9] used the Wvelet trnsform zero crossings for extrcting fetures from imges of the iris nd representing them, by fine-to-corse pproximtions t different resolution levels, clculted on concentric circles in the iris, to generte D fetures. 3. Proposed pproch This pper presents new pproch (Figure ) to improve the performnce of trditionl iris identifiction systems, using the cohort informtion. The rigorous experimentl results presented in this pper illustrte significnt performnce improvement without ny noticeble increse in computtionl complexity. Secondly, this pper investigtes the performnce improvement with user qulity, unlike trditionl pproches using imge qulity. The quntifiction of user qulity is bsed on the user imge qulity s well s the confidence on user mtching score from the trining dt. We lso use different pproch for imge pre-processing s detiled in next section, for more relible nd ccurte extrction of iris from the cquired imge. Then we present our results on the lrgest dtbse ever used in the literture of iris recognition (2877 imges from 4 users) nd lso show the performnce vrition s the number of trining imges re vried, s opposed to previous pproches []-[4] where mostly 3 imges were chosen for trining. In the end we show tht the proposed pproch is invrint of the trining imge used for building the system. Figure : Overview of proposed pproch /07/$ IEEE

2 4. Imge preprocessing Extrction of relible fetures from the cquired iris imges requires creful imge-preprocessing. This preprocessing includes extrction of region of interest (ROI) nd the msking of the unwnted/noisy portions from the ROI. Moreover the ROI needs to be trnsltion, scle nd illumintion invrint for efficient nd sclble iris recognition. The detils of these pre-processing opertions re provided in the following subsections. Figure 2: Block digrm for extrcting Iris 4.. Iris locliztion The iris is the nnulr portion between the blck pupil nd while scler. Out of these two the pupil whose edge constitutes the inner boundry of the iris, chnges its size due to light vritions but outer iris boundry remins fixed. The inner nd outer boundry of the iris cn be pproximted s circles but these circles re usully not concentric Extrcting the pupil informtion The pupil in the cquired imge usully contins reflection from the source, which form some bright spots in the pupil, so if the pixel vlue inside the pupil is over prticulr threshold (200) then it is replced by pixel vlue of some neighborhood pixel. This opertion lmost fills the circles but this still it is not good enough to pply globl threshold for pupil circle estimtion. Therefore the imge is pssed through 7 7 medin filter (Figure 3b). The imge is next scnned row-wise nd number of consecutive pixels whose vlue is less thn certin threshold (65), re clculted. As this scn is done on ech row, the row contining the highest number of such consecutive pixels must correspond to the dimeter of the pupil, hlf of tht mximum vlue corresponds to the rdius of the pupil, y coordinte of the center of pupil is the row of the dimeter nd x coordinte is clculted by dding rdius of pupil to the column from where the consecutive pixels strted. This is the estimted rdius nd the center of pupil, which is used for the precise computtion of rdius/center s detiled in the next subsection Extrcting the iris informtion The outer boundry (with scler) is estimted by first medin filtering the originl imge, s this type of filter elimintes sprse noise while preserving imge boundries. The imge obtined is pssed through filter to improve the contrst of the imge nd through Gussin filter to further remove noise due to iris texture, then the edge detection is performed using the cnny edge detector. After the edge detection window is chosen in the edge detected imge round the center of the pupil. Then every pixel in this window is ssumed s the center nd the number of white pixels, tht re encountered t the perimeter of circle, with rdius vrying from 80 to 20 pixels, re computed. The winner, i.e., the rdius (mong pixel) nd the center (mong ll pixels) for which the mximum white pixels re encountered, is locted. Tht rdius r corresponds to the rdius of the iris nd pixel which ws chosen s the center p, which gve tht mximum count, is the center of the iris Imge Normliztion The vrint conditions of imge cpturing cn influence the size of the iris which must be processed in the system. To compenste the stretching of the iris texture s the pupil chnges in size, nd hve new model of iris which removes the non-concentricity of the iris nd the pupil, the iris imge is unwrpped into rectngulr region of This method lso stndrdizes the Imges obtined to be tested in the system. The imge is unwrpped using the following technique: (d) Figure 3: Originl Imge, Imge fter medin filter, Imge fter edge detection nd (d) Imge fter detecting outer iris circle. Figure 4: Imge normliztion model

3 In the figure 4, R is the iris rdius & r is the pupil rdius with i nd p s their respective centers. d is the distnce between the two centers. R = ib ; () r = pa ; (2) d = ip ; (3) Angle φ is the ngle between ip nd pc, Angle α is the ngle between ip nd pa nd for ny ngle θ it is clculted s α = φ θ; (4) Angle Ψ is clculted using the following eqution Sin( ψ ) Sin ( α ) = ip iβ (5) The distnce pb is clculted by the eqution Sin( α + ψ ) Sin( α ) = pb iβ (6) Which gives, AB = pb pa (7) Using this distnce AB for θ vrying from 0 to 360 the iris portion is unwrpped into rectngulr strip of size (Figure 5) Imge Enhncement The normlized iris imge hs low contrst nd my hve non-uniform brightness cused by the position of light sources. In order to obtin better distributed texture imge, removl of bckground illumintion (Figure 5b), is performed [5]. The imge so obtined is then put to histogrm equliztion to get the enhnced imge (Figure 5c). Such preprocessing improves the contrst of the imge, removes the non-uniform lightening nd mkes the finer texture properties of the iris clerer. Figure 5: Normlized imge, Men imge i.e. estimte of bckground illumintion, nd Enhnced imge Eyelsh nd Eyelid removl Eyelshes nd eyelids occlude the iris region which results in reduced ccurcy, especilly for persons hving smller eyes; therefore they hve to be removed from the iris region. For this purpose once the prmeters for Iris nd Pupillry circle re known the edge detected imge is scnned from bottom to top strting from the topmost point of Pupillry circle to topmost point of iris circle (leving some mrgins). If ny white pixel is encountered then window of 80 8 is scnned for white pixels, now if the count of the white pixels in tht region is more then the length of tht window, then it is considered s noise due to eyelsh nd eyelid. At the lowest row of such window, the iris imge is seprted s noisy nd noiseless prt of the imge. Sme process is pplied for lower eyelids but with different threshold (70). Figure 6: Edge detected imge, Imge with eyelsh & eyelid detected, Normlized Imge, nd (d) Imge with eyelsh nd eyelid msked. 5. Feture Extrction Fetures re extrcted using the phse informtion (Eq. 9) obtined by filtering the enhnced imge through Log- Gbor filter. The frequency response of log-gbor filter is given s; 2 (log( f / f o )) G( f ) = exp 2 (8) 2(log( σ / fo)) where f o represents the centre frequency, nd σ gives the bndwidth of the filter. The most importnt property of log-gbor filter is tht it hs n extended til t the high frequency end. The sttistics of nturl imges suggested by Field [8] indicte tht they hve mplitude spectr tht fll off t pproximtely /f. Therefore log-gbor filters hving extended tils should be ble to encode nturl imges more efficiently thn ordinry Gbor filters, which would over represent the low frequency components nd under represent the high frequency components in ny encoding. Therefore -D log-gbor filter, with center frequency f o equl to /8 nd bndwidth of 0.55f o ws used for extrcting the fetures. After the imge ws pssed through the filter the phse informtion ws used to form the feture vector. b r = if, Even { G ( f, σ ) * I } 0 b r = 0 if, Even { G ( f, σ ) * I } < 0 b i = if, Odd { G ( f, σ ) * I } 0 (9) b i = 0 if, Odd { G ( f, σ ) * I } < 0 where, I is the iris imge nd b r nd b i re the bits representing the feture vector. Therefore for every bit in normlized iris imge there re two bits in the feture vector. (d)

4 6. Feture mtching The feture vector is in binry formt nd therefore hmming distnce is used for mtching ny two feture vectors. In ddition to tht, only those bits re used for mtching tht re not occluded by the eyelshes nd eyelids s seprted out erlier, whose informtion is stored in msk for every imge. The hmming distnce is clculted using the following eqution. S = M N i= j= { ΑR( i, j) ΒR( i, j) +ΑI ( i, j) ΒI( i, j)} (0) The feture mtching comes up in three phses, clculting the User Qulity, the hmming distnce clcultion using User Qulity nd finlly locting the pproprite clss for ech user using the cohort informtion. 6.. Clculting the User Qulity The user qulity is mesure of distinctiveness of prticulr biometric. Previously reserchers hve focused minly on the quntifiction of iris imge qulity. However, in our pproch the user qulity, rther thn imge qulity, hs been effectively used to chieve the performnce improvement. During the trining phse, for every user, minimum imposter score is clculted. This score illustrtes tht how much given user is distinct from the others in the dtbse; hence we cll it s user qulity score. U = min{ S i } () 2N M Where, U is the user qulity of user nd imposter scores for tht user. i S re the 6.2. Hmming distnce with User Qulity The user qulity score clculted in trining is now used in mtching the test nd trining imges. For mtching ny two imges, the minimum of the two qulity scores corresponding to the clss the imges belong to, re used s the weight. In cse of genuine user the clss remins the sme nd so the minimum is the sme s the qulity score for tht user. After computing the usul hmming distnce score from test nd trining imge tht score is divided by the weight obtined for tht pir of imges. This essentilly mens tht if user is more probble to be ner ny other clss then its qulity score would be low nd dividing by tht vlue will put tht user wy from the other clsses. S ' = S (2) min( Ux, Uy) where, S is the new score, Ux is the user qulity of the climed clss nd Uy is the user qulity of imposter clss. The lrger is the minimum imposter mtching score, better is the user qulity s this offers lrge seprtion of user from the imposters Using the Cohort Informtion Once the finl scores re clculted, if the score is less thn the threshold then tht user is clssified s genuine but if the score is more thn the threshold then cohort informtion is used to clssify tht user s genuine or imposter. For this when the score is more thn threshold, the test feture vector is compred with the trining feture vectors for the cohorts nd if the score is less thn ll the cohort mtching scores the user is climed genuine but even if one cohort score is less thn the previous score then the user is climed imposter. 7. Experiments nd Results For evluting the performnce of lgorithm we hve firstly used the CASIA-I Iris dtbse consisting of 756 imges from 08 different subjects nd CASIA III dtbse with 200 imges from 300 different users. Four experiments were conducted on both dtbse, in which for ech user, lst, 2, 3 nd 4 iris imges were chosen for trining. (ROC) curve is used to report the performnce of the proposed methodology. The ROC curve is the Genuine Acceptnce Rte () versus Flse Acceptnce Rte () plot which mesures the ccurcy of mtching process nd shows the overll performnce of n lgorithm. Ech plot shows three ROC curves, one corresponding to norml mtching, second corresponding to mtching with user qulity nd third one corresponds to mtching with user qulity nd using the cohort informtion to clssify the user. As expected, when the number of trining imges is incresed the verifiction ccurcy increses nd the results in Figures 7-0 illustrte significnt performnce improvement with the usge of user qulity nd the cohort informtion. Another set of experiments utilized only one trining imge. However, we used permuttion of ll the imges of user s the trining imge nd then present our results (Figure 9) s the verge of ll the experiments. This verging ensures tht the improvement by the proposed lgorithm is not bised, i.e., does not depend upon the imge used for trining the system. Finlly, we show the performnce of our lgorithm (Figure 0) on dtbse of 2877 imges from 4 users (CASIA III dtbse) in which only one imge ws used for trining for ech user. This is the lrgest scle ever on which iris identifiction hs been illustrted in the literture. The performnce from 4 users illustrted in Figure 0, using one trining imge, confirms the dvntge over the conventionl pproches presented in the literture.

5 nd Cohorts 75 7 nd Cohorts nd Cohorts 5 nd cohorts (d) Figure 7: ROC curves on CASIA I dtbse of 08 users with trining imge, 2 trining imges, 3 trining imges nd (d) 4 trining imges nd Cohorts nd Cohorts nd Cohorts nd Cohorts (d) Figure 8: ROC curves on CASIA III dtbse of 300 users with trining imge, 2 trining imges, 3 trining imges nd (d) 4 trining imges.

6 nd Cohorts nd Cohorts Figure 9: ROC curves when trining imge ws used, nd for ech user ll imges were chosen for trining one t time nd verge of tht is plotted CASIA I dtbse nd CASIA III dtbse the Chinese Acdemy of Sciences Institute of Automtion for providing us the CASIA iris imge dtbse nd Cohorts Figure 0: ROC curve when trining imge ws used on dtbse of 4 users. 8. Conclusions The min contributions from this pper cn be summrized s follows: (i) This pper investigtes new technique to improve the performnce of the system by using cohort informtion (ii) new user-qulity mesure is proposed for the performnce improvement, (iii) the new pre-processing steps in sec.4 tht genertes more relible nd ccurte extrction of fetures. The experiments presented in this pper employed the fetures extrcted using the -D log-gbor filter. The usge of user qulity s weight t feture mtching level nd cohort informtion hs been successfully used to chieve the significnt performnce improvement. A comprison of the two pproches, one without ny user qulity nd cohorts nd other with user qulity nd cohorts hs been conducted on the CASIA Iris Dtbse, nd it is shown in Figures 7-8, tht the ltter one hs significnt improvement in verifiction ccurcy, especilly t lower, which will be most likely to be n operting point for ny system. Figures 9 lso illustrted tht the improvement in verifiction ccurcy tkes plce irrespective of the choice of trining imge. 9. Acknowledgement This work is prtilly supported by the reserch grnt from Ministry of Informtion nd Communiction Technology, Government of Indi, grnt no. 2(54)/2006- ESD. We thnkfully cknowledge Prof. Tieniu Tn from 0. References [] L. M, Y. Wng, nd T. Tn, Iris recognition using circulr symmetric filters, Proc. Interntionl Conference on Pttern Recognition, vol. II, pp , [2] L. M, Y. Wng, nd T. Tn, Iris recognition bsed on multi-chnnel Gbor filtering, Proc. 5th Asin Conference on Computer Vision, vol. I, pp , [3] J. Dugmn, How iris recognition works, IEEE Trns. CSVT, vol. I, pp. 2-30, [4] J. Dugmn, The importnce of being rndom: Sttisticl principles of iris recognition, Proc. Intl. Conference Pttern Recognition, vol. 36, pp , [5] L. M, T. Tn, Y. Wng nd D. Zhng. Efficient Iris Recognition by chrcterizing Key Locl Vritions, IEEE Trns. Imge Processing, vol. 3, pp , [6] R.P. Wildes, Iris Recognition: An Emerging Biometric Technology, Proc. of IEEE, vol. 85, pp , 997. [7] L. Shinyoung, K. Lee, O. Byeon. Efficient Iris Recognition through Improvement of Feture Vector nd Clssifier, ETRI Journl. vol. 23, pp. 6-70, June 200. [8] C. Tisse, L. Mrtin, nd L. Torres. Person Identifiction Technique using humn Iris Recognition. Proc. Intl. Conference on Vision Interfce. pp , My [9] W. W. Boles nd B. Boshsh. A Humn Identifiction Technique Using Imges of the Iris nd Wvelet Trnsform, IEEE Trns. Signl Processing. vol. 46, pp , Apr [0] Y. Chen, S. C. Dss, nd A. K. Jin. Loclized Iris Imge Qulity Using 2-D wvelets, Proc. Interntionl Conference on Biometrics, pp , Jn, [] D. M. Monro, S. Rkshit, D. Zhng. DCT Bsed Iris Recognition. IEEE Trns. Pttern Anlysis & Mchine Intell., vol. 29, pp , Apr [2] C. Boyce, A. Ross, M. Monco, L. Hornk nd X. Li, " Multi-spectrl Iris Anlysis: A Preliminry Study", Proc. of IEEE Computer Society Workshop on Biometrics, Jun [3] Z. Sun, Y. Wng, T. Tn, J. Cui. Improving iris recognition ccurcy vi cscded clssifiers, IEEE Trns. Systems, Mn & Cybern., vol. 35, pp , Aug [4] D. Fields, Reltions between the sttistics of nturl imge nd the response properties of corticl cells, J. Opticl Soc. Americ, vol. 4, no. 2, pp , 987.