8/7/2019 icarcv419

1/6

On Fingerprint Template Synthesis

Wei-Yun Yau, Kar-Ann Toh, Xudong Jiang, Tai-Pang Chen and Juwei Lu

Cent re for Signal ProcessingSchool of El ectr ical & Elect ronic Engineeri ng

Nanyang Technological University

Abstract

In this paper, we address the false rejection problem dueto small solid state sensor area available for fingerprintimage capture. We propose a minutiae data synthesisapproach to circumvent this problem. Main advantagesof this approach over existing image mosaicing approachinclude low memory storage requirement and low com-putational complexity. Moreover, possible overhead onthe search engine (for fingerprint matching) due to data

redundancy could be reduced. Extensive experimentswere conducted to determine the best transformationsuitable for minutiae alignment. We demonstratethe idea of synthesis with an example using physicalfingerprint images. The proposed synthesis system is alsofound to improve (lower) the number of false rejects dueto the use of different fingerprint regions for matching.

keywords:-Biometrics, image database, fingerprint identification,fingerprint verification, minutiae points, template match-ing.

1 I n t r od uct i on

In general, an automatic fingerprint identification or ver-ification (see e.g. [1, 2, 3, 4]) system consists of threemain processing stages namely, image acquisition, fea-ture extractionand matching. In image acquisition, queryand template database images are acquired through var-ious input devices. Development over the years has seenthrough means that mechanically scan the ink based fin-gerprints into the computer system, and means which di-rectly capture the fingerprints using more sophisticatedsolid state sensors. With fingerprint images which couldbe distorted or contaminated with noise, the automatedsystem seeks to extract characteristic features which arediscriminating and yet invariant with respect to imageorientation. The final stage offingerprint identification orverification is to search and verify matching image pairs.

The use of inkless sensors has advanced the data ac-quisition aspects in an automatic fingerprint identifica-tion/verification system. This includes optical and solidstate devices. By means of CCD array and laser technolo-gies, the optical sensors offer a cost effective solution forfingerprint image capture. Conforming to the regulationby National Institute of Standard and Technology (NIST,USA), conventional optical sensors have a sensing area of1-inch by 1-inch. The solid state sensors, which adopt ca-pacitance, electric field, pressure or temperature sensingtechnologies, offer a more compact means for fingerprintimage capture with additional features to detect presenceof fingers such as locally adjustable automatic gain con-trol [5]. However, due to manufacturing limitation andcost factors, most solid state sensors do not come with

large sensing area (e.g. VeridicomsiTouch has a sensingarea of 1.5cm by 1.5cm and Infineons FingerTip has a

sensing area of11.1mm by 14.3mm). Moreover, the imag-ing area for the finger is further restricted to the area in

contact with the sensor. This, as compared to conven-tional ink based rolled fingerprint impression, possessesa much smaller information area. A consequence of thiscan be seen in using different partial areas of the samefinger for matching, which causes false rejection. For thisreason, during enrollment of a person in a database, arolled fingerprint would be preferred over a plain touchimpression.

Apart from requiring the individual user to ensuregood placement of fingerprint area during image ac-

quisition process, few automatic fingerprint identifica-tion/verification system has addressed the problem offalse rejection cause by using different image regions formatching. While acquiring a few separate fingerprint im-ages during registration could simply handle the false re- jection problem, much of the acquired information couldbe redundant (due to much common regions) and hencetakes up unnecessarily large storage space. Moreover,overheads of the multi-modal search would increase due tothe much larger number of records available for matching.In [6], an image mosaicing technique is developed for con-structing a rolled fingerprint from an image sequence ofpartial fingerprints. The proposed fingerprint mosaicingalgorithm consists of four stages namely, (i) segmentationof foreground and background areas in each frame; (ii)weighting of each images contribution using a foregroundmask; (iii) stacking of the weighted gray scale frames tocompute the mosaic gray scale image; (iv) stacking ofthe foreground masks to compute a confidence index atevery pixel. Although mosaicing technique possesses thecapability of acquiring a larger area of fingerprint image,it is at the expense of large storage requirement for amuch larger synthesized image. Moreover, as seen fromthe pixel level computation which is applied directly tothe acquired image, the computational cost is high.

In this paper, we propose a minutiae based syn-thesis method for an automatic fingerprint identifica-tion/verification system. The proposed methodology notonly synthesize necessary information for fingerprint iden-tification, but also possesses several desired features:

1. no restriction on the hardware sensor area;

2. small storage requirement since the synthesized datacontains only the necessary minutiae informationneeded for matching. This is especially useful forsearch within a large database in fingerprint identi-fication;

3. low computational complexity.

The paper is organized as follows: in section 2 we pro-vide a brief overview on fingerprint identification and ver-ification problems related to the subject matter. This isfollowed by section 3 where our representation system forminutiae data and our methodology for minutiae synthe-sis are presented. Section 4 provides an outline on minu-tiae detection, minutiae alignment and several transfor-

mations available for minutiae alignment. In section 5,we perform experimental study to decide upon the besttransformation for our application. Then, on top of aminutiae synthesis example, we provide an evaluation of

8/7/2019 icarcv419

2/6

the synthesis system in aspects of false rejection. Someconcluding remarks are drawn in section 6.

2 Fi nger pr int I dent ifica-

t ion/ Ver ification

While fingerprint identification refers to the process ofmatching a query fingerprint against a template finger-print database to establish the identity of an individual,fingerprint verificationrefers to determination of whethertwo fingerprints are from the same finger or not. Sinceverification is a comparison of a queryfingerprint againstan enrolled template fingerprint, it is also termed as one-to-one matching. Identification, on the other hand, istermed as one-to-many matching.

In view of effective search within a huge database inan identification problem, fingerprint matching is usuallycarried out at two levels namely, the coarse level and thefine level. The coarse level matching is also referred to asclassification (see e.g. [7, 8, 9, 10]) where fingerprints aregrouped into few classes so that a fine level search can be

performed within the matched class.As it is difficult to use the raw digital fingerprint imagedirectly for matching in an automatic fingerprint identifi-cation/verification system, a suitable computer represen-tation is essential. Several desirable properties for suchrepresentation can be identified as:

1. retention of the discriminating power of each finger-print (information content),

2. stable with respect to noise and distortion,

3. small in data storage size, and

4. ease in manipulation (e.g. adding new data points).

Here, we note that large information content may pose aconstrain to small data storage size requirement.

According to [11], eighteen different types of finger-

print features can be identified. These features includeridge endings, ridge bifurcations, short ridges and ridgecrossovers which are collectively termed as minutiae. Ithas been widely accepted that local ridge structures (acollection of minutiae details) from twofingerprints matcheach other if the fingerprints are from the same source[11, 12]. Hence, the problem of fingerprint verificationcan be reduced to a point pattern matching problem whenthese local structures are considered.

3 M inut iae D at a Synt hesis

Our representation for the fingerprint consists of a globalstructure and a local structure [13]. The global struc-

ture consists of positional and directional information ofridge endings and ridge bifurcations. The local structureconsists of relative information of each detected minutiawith other neighboring minutiae. Since the local struc-ture contains relative information which is insensitive torotation and translation, the main issue concerning minu-tiae data synthesis is to establish the relationship betweenthe global structures of two fingerprints acquired withcommon regions.

Let

M = {(xi , yi ,i , ti )}, i = 1, 2, ...n (1)

be the set of minutiae containing the positional informa-tion (x, y), directional information () and minutiae typeinformation (ti = 0 indicates a ridge ending and ti = 1indicates a bifurcation) for n minutiae elements in theglobal structure.

Suppose we have a total of m number of minutiae datasets from m partial fingerprints of the same finger, then

we can write for the kt h minutiae data set as:

Mk = {(xi , yi ,i , ti )}k , i = 1, 2, ...nk , k = 1, 2,...,m.(2)

Among these minutiae data sets, there would b e com-mon regions whereby information is redundant. If it isto search through each individual minutiae data sets formatching, it would not be cost effective since these redun-

dant information are being searched through more thanonce. Moreover, the geometrical relationships amongthese minutiae data sets are no longer preserved sincethese data sets are treated as separate entity. In orderto save data storage space with respect to redundancy aswell as to provide a good overall picture about the minu-tiae sets, a synthesis with consideration to relationshipbetween data sets is needed.

For fingerprint images with common regions, we canexpress the resultant synthesized information as:

Mm1

=m!

k = 1

{fk (xi , yi ,i , ti )}k , i = 1, 2, ...nk (3)

where fk , k = 1, 2,...,m denote the necessary topologicaltransformations for aligning the different sets of minutiaedata. Suppose o1 is the number of overlapping points

between M1 and M2 . Then the number of minutiae pointsin M

21

can be expressed as n(M21

) = n1 + n2o1 . Now

let o2 be the number of overlapping points between M21

and M3 . And the number of minutiae points in M31

is

n(M31

) = n1 + n2 + n3 o1 o2 . In short, the total

number of minutiae points in the synthesized minutiaedata set Mm

1can be written as:

n(Mm1

) =m"

k = 1

nk

m1"k = 1

ok , (4)

where ok is the number of overlapping minutiae pointsbetween M

k1

and Mk + 1 . Hence, if each ok is of consid-

erable size, the total number of minutiae in the synthe-sized data set (n(Mm

1)) can be significantly smaller than#

mk = 1 nk .Upon acquiring two images for synthesis, the task im-

mediately after minutiae detection is to find correspon-dence between these two images so that the global minu-tiae information between the two images can be aligned.A match between these two images is performed utilizingboth local and global minutiae information. Two minu-tiae data sets are considered matched if a weighted scorebetween the local and the global information exceeds acertain threshold value. The match shall reject the inputimage if the intersecting region is too small for good cor-respondence. We shall discuss our alignment method ina separate section.

Let Mj and Mk , j "= k, j,k {1, 2,...,m}, be two fin-gerprint minutiae data sets to b e synthesized. Suppose

there are p corresponding points (minutiae coordinates)between the two images. Denote this set of p correspond-ing points by C. Then, a topological transformation fcan be determined relating Mj and Mk from

xj = f(x k ) (5)

where xj = {(xi , yi )}j and x k = {(xi , yi )}k for all i C.Since the transformation will be used for aligning thosenon-corresponding minutiae points, a careful study on itssensitivity with respect to noise and deformation is nec-essary. We shall discuss various transformation modelsfor image points alignment in the following section.

4 M i nu t i ae D et ect i on an d

AlignmentFor minutiae detection, we adopt an adaptive ridge trac-ing algorithm which is evolved from [14]. Our approach

8/7/2019 icarcv419

3/6

adaptively traces the gray-level ridges of the fingerprintimage and applies adaptive oriented filters to the imageonly at those regions that require to be smoothed. A longtracing line will be obtained when there is little variationin contrast and when the bending level of the ridge islow. Main advantage of our approach is that tracing isby adaptive piece wise linear approximation of the ridgeswhich speeds up the process of ridge detection as com-

pared to other methods which adopt either pixel wise orfixed step tracing [15]. The tracing is only performedwithin the region of interest. The region of interest issegmented based on the local certainty level c(x, y) atpixel (x, y) on image I .

Several stopping criteria for ridge tracing which deter-mines detection of minutiae are adopted as:

1. Tracing exits from region of interest. In this case,minutiae extraction will not be performed.

2. Tracing ridge line intersects another already tracedskeleton ridge line. Under this condition, a bifurca-tion minutiae is detected.

3. Tracing ridge line ends when the tracing line isshorter than a threshold value and when the next

traced point lies on another ridge.In addition to minutiae detection, post processing is per-formed to remove spurious minutiae.

Having the minutiae extracted for two images, it isnecessary to align the two sets of data so that they forma larger picture of the fingerprint. Several problems whichare inherent to this alignment process are enumerated asfollows:

1. Translation and rotation variance between the twofingerprint images.

2. Some minutiae may be dropped and some spuriousminutiae may be detected.

3. Deformation of the fingerprint images which induceslocation errors.

Together with an indication of ridge ending and ridgebifurcation, the notation in (1) provides a global descrip-tion of the minutia. Since this feature vector is not rota-tion and translation invariant, we construct a local featurevector for our alignment purpose. In what follows, only abrief outline on the local structure for alignment match-ing will be provided. The interested reader is referred to[13] for greater details.

Let mj , j = 1, 2,...,l be the jt h nearest neighbour withrespect to m0 . The distance between m0 and mj can thenbe expressed as

dj 0 =$

(xj x0 )2 + (yj y0 )2 . (6)

Denote by 0 the direction ofm0 , the relative radial anglefor mj with respect to m0 is given by

j 0 = tan1

%yj y0

xj x0

& 0, j 0 . (7)

Let cj 0 , j = 1, 2,...,l be the ridge count between m0 andmj , then together with corresponding minutiae type tj 0we pack the local feature vector as

Fl j = [dj 0 , j 0, cj 0 , tj 0 ]T , j = 1, 2, ...,l. (8)

It is obvious that this local structure is rotation and trans-lation invariant since it contains only relative information.Hence, it can be used directly for preliminary local align-ment matching.

A match weighting the similarity between the local fea-ture vectors from the two images is performed so that

a common reference can be established. Once the pre-liminary correspondence b etween these local features isestablished, the transformation required for global align-ment can be found. To validate good correspondences

for this transformation, a further match combining bothlocal and global information is adopted.

Generally, transformation means can be classified inlinear form and nonlinear form. Consider two sets ofimage points: x = (x, y) and X = (X, Y). The prob-lem here is to find the best transformation f that relatesthese two sets of image points, i.e. x = f(X ). For lineartransformations, we have x = T X where T denotes the

transformation matrix.Affine geometry compares distances only on the same

line or on parallel lines. As compared to the Euclideangeometry, affine geometry relaxes the requirement on per-pendicularity. Hence transformation under affine geome-try is more general as compared to that under Euclideangeometry.

Projective geometry is the result of relaxing the restric-tions preserving parallel lines but require that straightlines remain straight lines for any changes we might im-pose on the figure.

Topology encompasses the projective, affine andEuclidean geometries. An even smaller set of properties isinvariant under topological transformation (e.g. preserv-ing only closed curves, order and connectivity). Here,only the quadratic type of topological transformation isinvestigated.

5 E x per i m ent s

5.1 Transform at ion st udy

In this section, we perform experimental study, usingphysical fingerprint data, to determine a suitable trans-formation for minutiae synthesis. We collect 5 imagescorresponding to 5 different areas (centre, top-left, top-right, bottom-left and bottom-right) for each finger forthe experiment. A total of 200 images were captured us-ing the Veridicom Sensor for 40 fingers.

Matching was first performed to obtain the correspond-

ing coordinates between two images which are to be syn-thesized. We used the centre area as the base image tomatch with one of the other areas (top-left, top-right,bottom-left and bottom-right) of the same finger. Thematched image pairs with 10 or more corresponding minu-tiae coordinates were then used for the following trans-formation study. As a result, only 50 matched pairs werefound to have 10 or more matched points.

To assess the accuracy of each transformation dis-cussed in previous section, 3/4 of the matched pointswere used for identifying the transformation parameters(fitting) and the rest of 1/4 were used for extrapolationtest (testing). The distribution of the sum of squared er-rors (SSE) for these matched pairs are plotted in Fig. 1and Fig. 2 for fit data and test data respectively. Thecontinuous line ( ) corresponds to SSE distribution for

affine transformation. The dashed line (- -) and thedotted line (...) correspond to projective transformationand topological transformation respectively. The meanvalue and the standard deviation (STD) for these errorsare also tabulated in Table 1 to reflect an overall view ofthese results.

As seen from Fig. 1, the topological (quadratic) trans-formation provides the best fit since the dotted curve fallsbelow the other two curves for all samples. This is alsoreflected in Table 1 since the mean SSE and the stan-dard deviation (STD) are the smallest among the threetransformations. As for test data not included in the fit-ting process, results from Fig. 2 and Table 1 show thataffine transformation gives the best result, in the senseof lowest mean SSE and lowest STD. It is important to

note that both the mean SSE and STD for the other twotransformations (projective and topological) are consider-ably huge as compared to those by affine transformation.Main reason being that coordinate warping according to

8/7/2019 icarcv419

4/6

0 5 10 15 20 25 30 35 40 45 500

100

200

300

400

500

600SSE distribution

sample number

SSE

Figure 1: Sum of squared error distribution forfit data

0 5 10 15 20 25 30 35 40 45 500

500

1000

1500

2000

2500 SSE distribution

sample number

SSE

Figure 2: Sum of squared error distribution fortest data

the fit data (interpolation) may not necessary fit well thetest data (extrapolation). Base on this study, the affinetransformation is adopted for alignment in our minutiaesynthesis system.

5.2 A minut iae synt hesis example

In this part, we show an example of synthesizing threefingerprint images. As shown in Fig. 3 through Fig. 5,three fingerprint images are captured from three differ-ent portions of the same finger. Minutiae points (shownin circles in Figures 3-5) are detected from these finger-print images using the ridge tracing algorithm. A visualexamination on these figures shall reveal that the minu-tiae information extracted in each image contains similarpoints (found in common regions) and dissimilar points(found outside common regions). It is also observed thateven within the common region, some minutiae detectedin one image may not be detected in another image due todifferent image qualities. Due to these reasons, when anytwo of these three images are used for matching in a fin-gerprint identification or verification system, false rejec-tion would occur when the threshold related to the totalnumber of matched minutiae pairs is set rather high.

Fig. 6 shows the synthesized minutiae points collectedfrom Figures 3-5, using Fig. 3 as the background image.

Table 1: Sum of Squared Errors for fit and testdata

Sum of squared error for fit dataAffine Projective Nonlinear

Mean 3 1.1304 62.2254 16.4531STD 25.3905 87.7418 19.4167

Sum of squared error for test dataAffine Projective Nonlinear

Mean 28.2344 227.7379 464.9472STD 29.7245 428.2317 1559.9000

The circles in the figure indicates the original detectedminutiae points from Fig. 3, whereas the plus and starsindicate those additional minutiae points transferred fromFig. 4 and Fig. 5 respectively. As seen from Fig. 6, theseadditional minutiae points have found correct correspon-dences on the fingerprint image (Fig. 3) which are notdetected in the original capture. A match comparing a

query image data with minutiae data from Fig. 6 will havea higher matching count.

50 100 150 200 250 300

50

100

150

200

250

300

Figure 3: Fingerprint sample 1 with detectedminutiae

5.3 Perform ance evaluati on

In this experiment, we show that the fingerprint synthesismethod can improve performance in terms of False Rejec-tion caused by using different regions of fingerprints formatching. A test sample consisting of 115 query imagesand 6115 template images were used for matching eval-uation. The query images were randomly acquired fromdifferent partial regions of a finger of each individuals.The first five sets (labeled as (a)-(e)) of template imagesconsist of different partial regions (i.e. centre, top-left,top-right, bottom-left, bottom-right) from each enrolledfinger. The last set (label as (f)) of templates consistsof synthesized data which are obtained by merging those

corresponding data from the samefi

nger of thefi

rstfi

vesets. As such, the last set of templates contains the samenumber of records as those in the first five sets, but withricher information.

8/7/2019 icarcv419

5/6

50 100 150 200 250 300

50

100

150

200

250

300

Figure 4: Fingerprint sample 2 with detectedminutiae

In Table 2, the percentages of match between the queryimage data set with each of the six template data setsare shown. It is seen that the synthesized data set (f)has provided the highest percentage of match for similarfingers.

The number of matched minutiae pairs are also plot-ted in Fig. 7 for all test samples. In this figure, the solidline represents the distribution of the number of matchedminutiae pairs corresponding to the synthesized templateset (f) while all other dotted lines represent the distrib-

ution corresponding to template sets (a)-(e). It is seenfrom this figure that the number of matched minutiaepairs (solid line) provides almost a covering envelope forthe synthesize template set over all other template sets.This indicates that most of the synthesized data set hassuccessfully captured the required minutiae informationfrom individual template data set. Those uncoveredcases corresponds to much distorted information due toincorrect as well as inaccurate transformation as a resultof the matching process. It is thus noted that obtainingas much matching minutiae pairs before synthesizing thedata could possibly help to improve the situation.

Table 2: Percentage of match for each template

setData sets

(a) (b) (c) (d) (e) (f)57 .39% 50.88% 5 4.9 5% 4 1.12% 36.79% 80.0 0%

6 Conclusion

In view of the limitation in solid state image sensor area,we propose, in this paper, a method to synthesize fin-gerprint data. The method is advantages over existingmosaicing technique in terms of low computational cost

and low memory storage requirements. Several transfor-mation models were compared for minutiae points align-ment. The affine transformation, which was found toprovide good interpolation and extrapolation capabilities,

50 100 150 200 250 300

50

100

150

200

250

300

Figure 5: Fingerprint sample 3 with detectedminutiae

was adopted for minutiae data synthesis. The synthesizedtemplate data set was found to improve matching perfor-mance in the sense of reducing false rejection which wascaused by using different fingerprint regions of the samefinger for matching.

References[1] Anil Jain, Lin Hong, and Ruud Bolle, On-line finger-

print verification, I EEE Tr ans . Pat t er n Anal y s i s and

M ac hi ne I nt el l i gence, vol. 19, no. 4, pp. 302313, 1997.

[2] Anil K. Jain, Lin Hong, Sharath Pankanti, and RuudBolle, An identity-authentication system using finger-prints, in Proceedings of the IEEE, 1997, pp. 13651388.

[3] Nalini K. Ratha, Kalle Karu, Shaoyun Chen, andAnil K. Jain, A real-time matching system for largefingerprint databases, I E E E T r a n s. P a t t e r n A n a l ysi sand Mac hi ne I nt el l i genc e, vol. 18, no. 8, pp. 799812,1996.

[4] U. Halici, L. C. Jain, and A. Erol, Introduction tofingerprint recognition, in I nt el l i gent Bi omet r i c T ec h-niques in Fingerprint and Face Recognition, L. C. Jainand et. al., Eds., pp. 334. The CRC Press internationalseries on computational intelligence, 1999.

[5] Lawrence OGorman, Fingerprint verification, in B i o -met r i c s : Per s onal I dent ifi cati on i n N etworked Society,Anil K. Jain, Ruud Bolle, and Sharath Pankanti, Eds.,1999, pp. 4364.

[6] Nalini K. Ratha, Jonathan H. Connell, and Ruud M.Bolle, Image mosaicing for rolled fingerprint construc-tion, in 14t h I nt er nat i onal Conf er enc e on Pat t er nRecognition, 1998, vol. 2, pp. 16511653.

[7] Raffaele Cappelli, Alessandra Lumini, Dario Maio, andDavide Maltoni, Fingerprint classification by di-rectional image partitioning, I E E E T r a n s. P a t t e r nAnal y s i s and Mac hi ne I nt el l i genc e, vol. 21, no. 5, pp.402421, 1999.

[8] Anil K. Jain and Lin Hong, A multichannel approachto fingerprint classification, I E E E T r an s. P a tt er nAnal y s i s and Mac hi ne I nt el l i genc e, vol. 21, no. 4, pp.348359, 1999.

[9] Kalle Karu and Anil K. Jain, Fingerprint classifica-

tion, Pattern Recognition , vol. 29, no. 3, pp. 389404,1996.

[10] Ugur Halici and Guclu Ongun, Fingerprint classifica-tion through self-organizing feature maps modified to

8/7/2019 icarcv419

6/6

50 100 150 200 250 300

50

100

150

200

250

300

Figure 6: Fingerprint sample with synthesizedminutiae (o: from samlpe 1; +: from sample 2;*: from sample 3)

treat uncertainties, in Proceedings of the IEEE , 1996,pp. 14971512.

[11] U.S. Government Printing Office, Washingtion, D. C.,T he Science of F ingerpri nts: Classifi cation and Uses,1984.

[12] H. C. Lee and R. E. Gaensslen, Eds., Adv anc es i n Fi n-gerprint Technology, Elsevier, New York, 1991.

[13] Xudong Jiang and Wei Yun Yau, Fingerprint minutiaematching based on the local and global structures, in

15th Internaional Conference on Pattern Recognition,1999.

[14] D. Maio and D. Maltoni, Direct gray-scale minutiaedetection in fingerprints, I E E E T r a n s. P a t t er n A n a l y -s i s and M ac hi ne I nt el l i genc e, vol. 19, no. 1, pp. 2740,1997.

[15] Xudong Jiang, Wei Yun Yau, and Wee Ser, Minutiaeextraction by adaptive tracing the gray level ridge of thefingerprint image, in I EEE I nt er nai onal Conf er enc eon I mage Pr ocess i ng ( I CI P 99), 1999, pp. 852856.

0 20 40 60 80 100 1200

10

20

30

40

50

60

70

80

sample number

numberofmatchedminutiae

Figure 7: Distribution of number of matchedminutiae for each test sample