This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
730 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 3, MARCH 2010
Improving Biometric Authentication PerformanceFrom the User Quality
Ajay Kumar and David Zhang
Abstract—The effectiveness of a biometric measurement andsensing system is directly related to the performance generatedfrom sensed data. This paper investigates a new approach toquantify the quality of sensed data from the user templates. Theobjective is to incorporate the quality of sensed data to generate areliable estimate on the matching scores. The proposed method ofextracting user quality is based on the confidence of generatingreliable matching scores from the user templates. We simulta-neously extract the palmprint and hand-shape images from thesingle hand image and ascertain the performance improvement forthe individual trait. The experimental results from the proposedapproach are also presented when the biometric measurementsfrom the finger knuckles are employed. The experimental resultspresented in this paper show significant improvement in perfor-mance while incorporating the proposed method of user quality inthe matching stages. The proposed user-quality-based fusion of thetwo biometric modalities also achieves promising improvement inperformance.
Index Terms—Biometrics, biometric quality, finger knuckle,hand geometry, palmprint, personal authentication, user quality.
I. INTRODUCTION
B IOMETRIC measurement is a key component of severalpersonal authentication systems that only render services
to legitimately enrolled users. The performance and the useracceptance are the two primary criterions for the selection ofsuch biometric measurement systems for real deployment. Thebiometric measurement systems that employ hand image datahave high user acceptance and have invited lot of attentionin the literature [1]–[8]. The peg-free hand imaging is moreuser friendly, as compared with the imaging setup employingpegs to constrain the hand pose, but generates high variationsin the acquired images. Therefore, the performance of a hand-based biometric measurement system using peg-free imagingis quite low and requires further efforts to make such systemsready for real deployment. The acquisition of hand images thatcan deliver palmprint and hand shape information is easy andhas also been demonstrated in our earlier work [5], [8]. In thecontext of recent work on quality-based fusion in [9]–[14] andthe current popularity of hand-based biometrics system, the
Manuscript received March 7, 2009; revised July 6, 2009. First publishedOctober 16, 2009; current version published February 10, 2010. This workwas supported in part by the Department of Computing, The Hong KongPolytechnic University, under Grant 4-Z0F3. The Associate Editor coordinatingthe review process for this paper was Dr. George Giakos.
Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TIM.2009.2028773
quality-based analysis for a hand-based system deserves carefulevaluation and is the focus of this paper.
A. Prior Work
The hand-based biometric measurement systems have at-tracted a lot of attention in the literature. The matching ofpalmprint measurements derived from Gabor phase features [2],ordinal filters [4], and minutiae features [1] has shown to offerpromising results for civilian and forensic applications. Theextraction of hand geometry requires low-resolution imagingand has attracted several studies in the literature [3], [5], [8].There has been recent trend to investigate the use of peg-free imaging for hand authentication, primarily due to thehigh user convenience in civilian applications, and it has beenshown to offer promising results. The quality-dependent fusioncan offer significant improvement in performance over fusionwithout quality and has extensively been investigated for iris[14], face [13], and fingerprint [10] modalities. Recently, therehave also been interesting efforts [18] to achieve promisingimprovement in performance using quality-based normalizationof the matching scores. However, there has not been anyeffort to exploit the quality-based performance improvementfor palmprint- or hand-geometry-based biometric systems. Itmay be noted that, unlike the textured gray-level patterns forfingerprint, iris, or face images, it is very difficult to ascertainthe quality of a shape, i.e., binary (hand geometry) images.Therefore, one has to develop new approaches to ascertain thequality of such images, possibly from the nature of the extractedfeatures or the response of the employed matcher to the ex-tracted features.
II. USER QUALITY
The quality of a biometric sample is not only associatedwith the images but also with the interaction of the biometricwith the employed feature extractor and the matching criteriaused for the decision. The user quality in this paper [6] isdefined from the associated biometric sample and quantified asa measure of confidence of user biometric sample with its owntemplates. Therefore, the user quality is computed from the usertemplates (acquired during the registration). The researchers inthe prior work have presented promising efforts on the quan-tification of image quality. However, in our approach, insteadof individually estimating the quality of the query images fromvarious integral transforms, a single quality measure for eachuser is estimated from its genuine training matching scores. Theestimated quality of the biometric samples in such a way can
KUMAR AND ZHANG: IMPROVING BIOMETRIC AUTHENTICATION PERFORMANCE FROM THE USER QUALITY 731
efficiently be employed to achieve the performance improve-ment. This type of quality is the quality for the biometric of theuser, rather than the quality of an image, and hence, it is termedas user quality. The user quality is defined as
Qp = min {Spi } ∀ i = 1, . . . , zC2 (1)
where Qp is the user quality of the pth user, Spi ’s are the genuine
matching scores, and z represents the number of templatesavailable for the pth user.
The user quality score estimated during the training phase isused in matching the test and training images. The matchingof two images uses the maximum of the two quality scores,corresponding to the class the images belong to, as the weight[see (2)]. In case of genuine user, the class remains the same,and so the maximum is the same as the quality score forthat user. After computing the usual matching score from thetest and training image, that score is multiplied by the weightobtained for that pair of images. This essentially means that if auser is more probable to be near any other class, then its qualityscore would be high, and multiplying by that value will put thatuser away from the other classes, e.g.,
S ′ = S ∗ max(Qa, Qb) (2)
where S ′ is the new or quality-weighted score, Qa is the userquality of the claimed class, and Qb is the user quality of theimposter class. The smaller is the genuine matching score Qp,the better is the user quality, as this offers high confidenceon the generated matching scores or larger separation of userfrom the imposters.
III. FEATURE EXTRACTION
The performance improvement using hand-based biometricmeasurement systems is the focus of our efforts in this pa-per. We firstly employed hand images as samples of typicalbiometric measurement. The acquired hand images are firstlysubjected to image normalization for the reliable segmentationof the region of interest, i.e., palmprint and hand geometry im-ages. The employed steps for image normalization and featureextraction are the same as detailed in [5]. As detailed in [5],the hand binarized shape images are firstly obtained for theextraction of effective features based on geometrical informa-tion or geometry plus interior content. The 17 features that cancharacterize every hand-shape image, i.e., perimeter, four fingerlengths, eight finger widths, palm width, palm length, handarea, and hand length, are extracted. The distance between thefeature vector from an unknown hand shape fx and the featurevector from the known class x is computed from the Euclideannorm (‖.‖)
h(f, fx) =∑
i
∣∣f i − f ix
∣∣ . (3)
The simultaneously extracted fixed-size palmprint images aresubjected to discrete Cosine transform (DCT) decompositionfor the characterization of palmprint texture. DCT is highly
Fig. 1. Hand geometry and palmprint image samples from two subjectsemployed in experiments.
computationally efficient1 and, therefore, suitable for any on-line hand identification system. Each of the 300 × 300 pixelsegmented palmprint images are divided into 24 × 24 pixeloverlapping blocks. The extent of this overlapping has empiri-cally been selected as 6 pixels [5]. Thus, we obtain 144 separateblocks from each palmprint image. The DCT coefficients fromthe decomposition of each of these L square block pixels, i.e.,f(j, k), is obtained as
C(u, v) = ε(u) ∗ ε(v) ∗L−1∑j=o
L−1∑k=0
f(j, k)
∗ cos[π.u
2.L(2j + 1)
]cos
[π.v
2.L(2k + 1)
](4)
where u, v = 0, 1, . . . , L − 1, ε(u) = ε(v) =√
2/L for u �= 0,and ε(u) = ε(v) =
√1/L for u = 0.
The standard deviation of the DCT coefficients in each of theblocks, i.e., C(u, v), forms the 1 × 144 feature set from eachof the extracted palmprint images. The high degree of intraclassvariability in palmprint features, mainly resulting from the peg-free imaging, is reduced using Z score normalization.
IV. EXPERIMENTS AND RESULTS
To study the relative merits of employing the user quality,we performed several experiments on the hand image databasefrom 100 subjects, which has been employed in [5] earlier. Thisdatabase has five left-hand images acquired in one session andfive images acquired in the second session. The estimation ofuser quality is performed during the training phase, and theestimated user quality is used to evaluate the performance fromthe test images (second session). The typical hand geometryand palmprint image samples resulting from the automatedsegmentation of acquired hand images is shown in Fig. 1.The distribution of user quality from the palmprint and handgeometry images is shown in Fig. 2. The comparative receiveroperating characteristics (ROCs) from the test data, with andwithout the use of user quality, are illustrated in Figs. 3 and 4.These ROCs illustrate that the integration of the user qualityhas significantly improved the individual performance usingpalmprint (Fig. 3) and hand- geometry (Fig. 4) matching. Theequal error rate (EER) for palmprint is reduced from 5.6% to3.6%, whereas it decreases from 6.4% to 5.8% for the handgeometry. Therefore, the achieved performance improvementis higher for the palmprint images than from the hand geometry
1DCT is the basis of JPEG and several other standards (MPEG-1, MPEG-2for TV/video, and H.263 for video phones).
732 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 3, MARCH 2010
Fig. 2. Distribution of estimated user quality for (a) the hand geometry and (b) the palmprint.
Fig. 3. Comparative ROCs from the palmprint.
Fig. 4. Comparative ROCs from the hand geometry.
Fig. 5. Comparative ROCs for the score-level combination.
TABLE IIMPROVEMENT IN EER USING USER QUALITY
images and more pronounced at lower values of false accep-tance rate (FAR). The palmprint and hand geometry matchingscores were combined using hyperbolic product combinationto ascertain the performance from the combination of twomodalities. The hyperbolic product combination generates thecombined score from tanh(sp ∗ sh), where sp is the normalizedpalmprint score, and sh is the normalized hand geometry score.This combination was empirically evaluated against the sum,product, and weighted sum and found to generate better per-formance. The comparative performance, with and without theuse of user-quality-based matching scores, from this nonlinearcombination is illustrated in Fig. 5. Table I summarizes thecomparative EER obtained from each of the experiments.
KUMAR AND ZHANG: IMPROVING BIOMETRIC AUTHENTICATION PERFORMANCE FROM THE USER QUALITY 733
Fig. 6. Distribution of user quality for (a) the PolyU Palmprint database and (b) the comparative ROCs.
We also performed the experiments on the PolyUpalmprint v1 database [15] from 108 users to ascertain thequality-based performance improvement. However, the ordinalrepresentation, as detailed in [4], was employed for the ex-traction of the reliable features since this representation hasshown to offer better performance than those from the Gaborphase features employed in [2]. The feature extraction usingordinal representations employed the same parameters as usedin [4]. We employed the first four images for the training, i.e.,generation of user quality, and the rest of the images were usedto ascertain the performance. The distribution of user qualityfrom this palmprint data set is illustrated in Fig. 6(a), andthe corresponding comparative performance can be ascertainedfrom the ROCs displayed in Fig. 6(b). The performance im-provement from the user-quality-based matching is significant,particularly at lower FAR, which is a preferred operating pointfor majority of the applications.
Another set of experiments were performed using the fingerknuckle biometric image samples. Automatically extracted fin-ger knuckle images of size 80 × 100 pixels, from the left-handmiddle fingers, were employed to ascertain the performanceimprovement using user quality. Fig. 7 shows a sample imagefrom the middle finger and the automatically extracted knuckleregion employed for the biometric measurement. The methodof extracting the finger knuckle is same as detailed in [16].The database from 105 subjects was employed for performanceevaluation, and the three enrolment images are used for thetraining, i.e., to ascertain user quality, and the rest of thetwo images were employed for the testing phase. The ordinalrepresentation was employed for the representation of knucklefeatures, and matching scores were generated using the stepsdetailed in [4]. The distribution of user quality for the fingerknuckle biometric image samples is illustrated in Fig. 8(a). TheROCs for the test data, with and without the use of user quality,are illustrated in Fig. 8(b). The EER of 5.8% was achievedfrom the test samples, which reduced to 4.22% when the user
Fig. 7. Sample image of a middle finger and automatically extracted knuckleregion.
quality was employed in the authentication. The experimentalresults shown in Fig. 8(b) suggest significant improvementin performance for the finger knuckle biometric measurementsystem when the user quality is incorporated in the decision-making process.
V. CONCLUSION
This paper has investigated a new approach to achieve theperformance improvement for the hand-based biometric mea-surement systems by incorporating user quality during thematching stage. The estimation of user quality is based onthe confidence of generating a reliable matching score fromthe user templates. The experimental results illustrated in
734 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 59, NO. 3, MARCH 2010
Fig. 8. Distribution of user quality for (a) the middle finger knuckles and (b) the comparative ROCs.
Figs. 3, 4, 6, and 8 consistently suggest significant improvementin the performance from the use of quality-based matching.In addition, the experimental results in Fig. 5 suggest that thequality-based fusion of palmprint and hand geometry scoresachieves promising improvement in performance. In this paper,five user templates were employed for the palmprint and handgeometry experiments (Table I), since these samples wereacquired during the registration stage. The number of usertemplates acquired during the registration is usually smaller. Inthe worst case, when only one user template is available, theuser quality scores can be generated by the matching scoresascertained from the scaled and rotated templates with itself.
The quality of biometric measurement also depends on theimage quality, which is often linked to imaging resolution. Incase of palmprints, civilian applications typically use 100-dpiimaging, whereas about 400 dpi is typically needed to acquirepalmer ridge or minutiae features. The quantification of userquality introduced in this paper can also be generalized forhigh-resolution palmprint images [1] and can be useful forforensic applications. It should be noted that an increase inimaging resolution does not necessarily results in higher userquality since the quality and the background of the palmerridge structure is user specific. Future efforts should be fo-cused on evaluating the user-quality-based personal recognitionon large biometric databases, i.e., samples from more than500 subjects. Recently, Hao et al. [17] have illustrated the utilityof multispectral palmprint images for personal identification.The comparison and combination of the integrated palmprintimage quality with the user quality measure suggested in thispaper would be highly useful and suggested for further work.
REFERENCES
[1] A. K. Jain and J. Feng, “Latent palmprint matching,” IEEE Trans. PatternAnal. Mach. Intell., vol. 31, no. 6, pp. 1032–1047, Jun. 2009.
[2] D. Zhang, W. K. Kong, J. You, and M. Wong, “On-line palmprint identifi-cation,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 25, no. 9, pp. 1041–1050, Sep. 2003.
[3] A. Kumar and D. Zhang, “Hand geometry recognition using entropy-based discretization,” IEEE Trans. Inf. Forensics Security, vol. 2, no. 2,pp. 181–187, Jun. 2007.
[4] Z. Sun, T. Tan, Y. Yang, and S. Z. Li, “Ordinal palmprint representationfor personal identification,” in Proc. CVPR, 2005, pp. 279–284.
[5] A. Kumar and D. Zhang, “Personal recognition using hand shape andtexture,” IEEE Trans. Image Process., vol. 15, no. 8, pp. 2454–2461,Aug. 2006.
[6] A. Kumar and D. Zhang, “Incorporating user quality for performanceimprovement in hand identification,” in Proc. 10th ICARCV , Hanoi,Vietnam, Dec. 2008, pp. 1133–1136.
[7] X. Jiang, W. Xu, L. Sweeney, Y. Li, R. Gross, and D. Yurovsky,“New directions in contact free hand recognition,” in Proc. ICIP, 2007,pp. 389-392.
[8] A. Kumar, D. C. M. Wong, H. Shen, and A. K. Jain, “Personal authentica-tion using hand images,” Pattern Recognit. Lett., vol. 27, no. 13, pp. 1478–1486, Oct. 2006.
[9] J. Kittler, N. Poh, O. Fatukasi, K. Messer, K. Kryszczuk, J. Richiardi, andA. Drygajlo, “Quality dependent fusion of intramodal and multimodalbiometric experts,” Proc. SPIE, vol. 6539, pp. 653 903-1–653 903-14,Apr. 2007.
[10] J. Fierrez-Aguilar, Y. Chen, J. Ortega-Garcia, and A. K. Jain, “Incorpo-rating image quality in multi-algorithm fingerprint verification,” in Proc.ICB, 2006, pp. 213–220.
[11] K. Nandakumar, Y. Chen, S. C. Dass, and A. K. Jain, “Quality-basedscore level fusion in multibiometric systems,” in Proc. ICPR, Hong Kong,Aug. 2006, vol. 4, pp. 473–476.
[12] J. Fierrez-Aguilar, J. Ortega-Garcia, J. Gonzalez-Rodriguez, and J. Bigun,“Kernel-based multimodal biometric verification using quality signals,”Proc. SPIE, vol. 5404, pp. 544–554, 2004.
[13] N. Poh and S. Bengio, “Improving fusion with margin-derived confidencein biometric authentication,” in Proc. AVBPA, vol. 3546, LNCS, 2005,pp. 474–483.
[14] Y. Chen, S. Dass, and A. K. Jain, “Localized iris image quality using 2Dwavelets,” in Proc. ICB, Hong Kong, Jan. 2006, pp. 373–381.
[15] The PolyU Palmprint Database, ver. 1.0. [Online]. Available:http://www.comp.polyu.edu.hk/~biometrics
[16] A. Kumar and C. Ravikanth, “Personal authentication using fingerknuckle surface,” IEEE Trans. Inf. Forensics Security, vol. 4, no. 1,pp. 98–110, Mar. 2009.
[17] T. Y. Hao, Z. Sun, and T. Tan, “Comparative studies on multispectral palmimage fusion for biometrics,” in Proc. ACCV , vol. 4844, LNCS, Tokyo,Japan, 2007, pp. 12–21.
[18] N. Poh, T. Bourlai, and J. Kittler, “Improving biometric device interoper-ability by likelihood ratio-based quality dependent score normalization,”in Proc. IEEE Conf. Biometrics: Theory, Appl. Syst., Washington D.C.,2007, pp. 1–5.
KUMAR AND ZHANG: IMPROVING BIOMETRIC AUTHENTICATION PERFORMANCE FROM THE USER QUALITY 735
Ajay Kumar received the Ph.D. degree from TheUniversity of Hong Kong, Hong Kong, in May 2001.He completed his doctoral research with The Uni-versity of Hong Kong in a record time of 21 months(1999–2001).
In 1993, he was with the Indian Institute of Tech-nology, Kanpur, India, and the Indian Institute ofTechnology, Delhi, India, before joining the IndianRailway Service of Signal Engineers. From 2001to 2002, he was a Postdoctoral Researcher withthe Department of Computer Science, Hong Kong
University of Science and Technology. He was awarded The Hong KongPolytechnic University Postdoctoral Fellowship from 2003 to 2005 and waswith the Department of Computing. From 2005 to 2008, he was an AssistantProfessor with the Department of Electrical Engineering, Indian Institute ofTechnology, Delhi. He was the Founder and Lab In-Charge of the BiometricsResearch Laboratory, Indian Institute of Technology, Delhi. Since 2009, hehas been an Assistant Professor with the Department of Computing, TheHong Kong Polytechnic University, Hong Kong. His research interests includepattern recognition with emphasis on biometrics and computer-vision-baseddefect detection.
David Zhang received the degree in computer sci-ence from Peking University, Beijing, China, theM.Sc. degree in computer science and the Ph.D.degree from the Harbin Institute of Technology,Harbin, China, in 1982 and 1985, respectively, andthe Ph.D. degree in electrical and computer engineer-ing from the University of Waterloo, Waterloo, ON,Canada, in 1994.
From 1986 to 1988, he was a Postdoctoral Fellowwith Tsinghua University, Beijing, China, and thenan Associate Professor with Academia Sinica,
Beijing. He is currently the Head of the Department of Computing anda Chair Professor with The Hong Kong Polytechnic University, Kowloon,Hong Kong, where he is the Founding Director of the Biometrics TechnologyCentre (UGC/CRC) supported by the Hong Kong SAR Government in 1998. Heis the Founder and Editor-in-Chief of the International Journal of Image andGraphics, a Book Editor for the Springer International Series on Biometrics,and an Organizer for the International Conference on Biometrics Authentica-tion. He is the author of more than ten books and 200 journal papers.
Dr. Zhang is a Croucher Senior Research Fellow, a Distinguished Speaker ofthe IEEE Computer Society, and a Fellow of the International Association forPattern Recognition. He is an Associate Editor of more than ten internationaljournals including IEEE TRANSACTIONS and Pattern Recognition and theTechnical Committee Chair of IEEE Computational Intelligence Society.