Proceedings of the 2011 International Conference on Wavelet "\nalysis and Pattern Recognition, Guilin, 10-13 July, 2011

A Novel Matching Technique for Fingerprint Recognition by Graphical

Structures.

Zhifan Gaol, Xinge Youl, Long Zhou2, Wu Zeng2,3

1 Department

of Electronics and InfonnationEngnineering,Huazhong University of Science and Technology,Wuhan, China 2Department ofElectric InfonnationEngineering , Wuhan Polytechnic University,Wuhan, China

3State Key Laboratory of InfonnationEngineering in Surveying, :rvrapping and Remote Sensing,Wuhan University,Wuhan, China E-l\AIAL: [email protected],[email protected],[email protected], [email protected]

Abstract
Fingerprint matching is an important issue in auto matic fingerprint identification systems. There are dif ficulties about fingerprint matching based on neighbor hood. One is the size of the neighborhood can not be determined readily, the other is the feature in the neigh borhood can be affected by the noise. To deal with these problem, we developed a novel algorithm for fingerprint matching based on local structures to efficiently extract neighboring minutiae features. Neighboring features present the information of peripheral minutiae which directly connect with the central minutiae on topology. We use one feature vector to present neighboring fea tures from different samples. The samples considered as the same class can make the proposed algorithm robust to rotation and translation of fingerprint images. The experiments are conducted on FVC2002, and the results illustrate the effectiveness of the proposed algorithm.

and for evidence of the crime all over the world. In the work of fingerprint matching, the characteristic features is the core. around this topic: tures. :rvrany studies have been conducted how to capture features, how to ex

tract features, how to select features, and how to use fea There are several useful features categorized into three levels[6]. Level 1 features, such as pattern class and ridge pattern, provide the macro details of the fingerprint impressions. Minutiae belongs to Level 2 feature. Level 3 features are micro patterns,including ridge width,edge con tour, pores and other permanent details. The level 1 feature has no uniqueness,however it can provide the general clas sification of the fingerprint in statistical analysis. Utilizing the level 2 feature is the main method to discriminate fin gerprints. It provides sufficient infonnation[ll] to enlarge intra-class distance and reduce inter-class distance until fin gerprints from different people could be recognized. So far, Level 3 feature has been used as a supplementary method to provide extra discriminatory information, partly due to the instruments of capturing fine level 3 feature are expensive. :rvrany approaches about fingerprint matching have been shown in literature, and the most popular fingerprint match ing algorithms are based on minutiae. It is necessary to So as to capture accurate information about minutiae.

Keywords:
Fingerprint matching; graphical structure; biomet rics; pattern recognition

Introduction
With the advent of electronic connection in our society,

nonnal-quality fingerprint images, the performance of these approaches relying on minutiae is better than most other method[lO]. Though minutiae are divided into many cat egories, it usually adopts two types of minutiae: bifurcation and ridge ending, shown in Fig.l. ridge Tico and automatic personal identification has been becoming an im portant issue in many field. However, traditional idenfica tions widely used, like ID card and password, cannot meet the need of more accurate and safe personal identification. Fingerprint recognition is a popular technique to overcome the disadvantages of traditional identifications, and provide a more accurate and private solution,With the development of Automatic fingerprint identification system(AFIS), fin gerprint identification is widely used in identification units

Kuosmannen[14] proposed a method to create a minutiae descriptor, which sampled a set of local orientation val ues unifonnly around the minutiae point. Cappelli[3] in troduces another minutiae point called Minutia Cylinder Code(MCC) to combine nearest neighbor-based methods and fixed radius-based methods to utilize their advantages

978-1-4577 -0282-2/11/$26.00 2011 IEEE 77

Proceedings of the 2011 International Conference on Wavelet "\nalysis and Pattern Recognition, Guilin, 10-13 July,

2011

descriptor, to combine the preceding advantages and over come these drawbacks. First, the descriptor is defined and its characteristics are presented. Then the descriptor is ap plied on the topology of fingerprint minutiae and extract features, so each central point can be represented as a fea ture vector, and these feature vectors can be converted into feature values. Finally, we use feature vector sets from dif ferent fingerprints to finish verification. The rest of this paper is organized as follows: Section 2 defines the proposed descriptor and presents its characteris tics. Section 3 describes the process of matching algoritllm. The experiments are presented in section 4. Section S shows the conclusion and future work.

Figure 1. A par t of a figure impression with r idge bifurcation and ridge ending. The circle presents the ridge bifurcation, and the rect angle presents the ridge ending.

and eliminate their drawbacks. Chang[4] combined Hough transform and 2-D cluster appproach to correctly identify the missing patterns and pseudo-patterns. The researches of Starink[12], Le[16], Tan[13] are around the topic of en ergy minimization. Bishnu[2] studied the 2-D partial point set pattern matching, and introduces some simple assump tions for approximate match. Weber[l7] used a template based algorithm to recognize fingerprint simply, that was simpler than Hough transform. Udupa[IS] improved We ber's method, to consider the problem of fingerprint match ing as comparing between two corresponding feature sets. Moreover, point pattern matching in fingerprint is Global minutiae matching has two also considered as global minutiae matching and local minutiae matching. disadvantages[3]: computationally demanding and lack of robustness. The state-of-the-art idea is about how to make the performance of local matching reach the global level. Based on neighborhood, local minutiae matching is very efficient, and tolerant against spurious and missing feature points. However, they have their own shortcomings. First, if the radius of neighborhood is fixed, as a threshold, the value of radius is not easy to determined, because location of minutiae is inaccurate and fingerprint images have non linear distortion, and the minutiae near the neighborhood border may be lost. Second, if the number of the points in neighborhood is fixed, the methods are sensitive to missing and spurious minutiae. In this paper, we introduce a novel minutiae-based local Let T and tively.

Figure 2. (a)an original fingerprint image. (b)The minutiae set. (c)The topology net.

The Proposed Methods

I be

two fingerprint minutiae vectors, corre

sponding to the template and the input fingerprint, respec Unlike other matching techniques, the representa tion of minutiae vectors are considered as feature vectors whose elements are the fingerprint minutiae[lO]. The ele ments of the vectors can be described by some fingerprint features, like type, orientation, location of minutiae and so on. Most popular minutiae matching approaches consider each minutiae as a triplet
m
=

{.T, y, B},

where the

are x-coordinate and y-coordinate respectively and each minutiae as a couple

m

x, y B is the

minutiae angle , however, in the following, we only consider

=

{x, y },

and

x, y means the

78

Proceedings of the 2011 International Conference on Wavelet "\nalysis and Pattern Recognition, Guilin, 10-13 July, 2011

the first quadrant to the fourth quadrant.

{:

Xi - Xik :::: 0, Yi - Yik :::: + 1800 Xi - Xik < S + 3600 Xi - Xik :::: 0, Yi - Yik <

(6)

Now the neighboring features are presented as the differ ent distances and the different angles to the central point. In order to process these information, the different neighbor ing points are mapped to be on a feature circle neighborhood.

C, and the
(7)

center of this circle is the central point corresponding the

C
Let r be the radius of

MAPC(N)
number of the point

Figure 3. There has some marked point on this topology net, which is from A to J. Let A be the central point. The neighboring points are B, C, D, E, F, G, H, I, while J does not be longs to the neighborhood of A, because the connection between A and J is not direct, re lying on F or other points.

C, and the

NP

on the circumference of C is 8r, shown in Fig.4. One neigh boring point is mapped to one point on the circumference of

C.

There mapping function

MAPC from the neighboring

X

point to circumference point is divided into two parts,

Cweight

startingweight

(8)

(9) coordinates in Cartesian coordinates: The formula(8), is a part of

MAPC

about the weight of

each neighboring point, where startingweight is the start

(1)

ing weight of each point in the neighborhood, and position

T
wheremi
=

is the weight after mapping. The formula(9) is about the

=

Cweight

{mi,m, ... ,mj}

=

(2)

Cposition

of neighboring point on the circle

1, 2, ... ,NT,

1, 2, ... , NI, {Xj,Yj},i {Xi,Yi},mj NI and NT are the number of points in

Next, the feature circle

shown in Fig.5,

C. C is extended to a feature vector where the point in the circle C, corre V , while
the point corresponding to

I and T respectively.

sponding to the small angle, is mapped to the low position of the feature vector feature vector the large angle, which is mapped to the high position of

is the point in the neighborhood which has the direct con nectivity to the central point mi ,that is, the connection be tween mi and mik does not rely on any other point.

Let N be the neighborhood of minutiae mi,and the mik

V.

In feature vector

V , the

non-zero ele

ments present the neighboring point. The element position

Vposition
(3)

and value

tance d of the neighboring point, respectively.

Vweight present the angle 8 and the dis C position Cweight

(10) (11)

where k point

1, 2, ... ,NP, NP is the number of neighboring

Vposition Vweight
The feature vector

points, the function Con presents the map from the central

mi

to its neighboring point. Fig.3 shows the direct

connectivity in topology net. Let d be the Euclidean distance between the neighbor hood point and the central point, and S be the slope from the neighborhood point to the central point.

is sensitive to the rotation of the

fingerprint image. In the several capture of fingerprint from the same person, the fingerprint images have obvious differ ences in vision, one of which is the rotation. This difference leads the absolute locations of neighboring minutiae to be variable, resulting in the feature circle and the feature vec tor, while the pattern of one fingerprint will be changed.

Yi - Yik Xi - Xik

(5)

However, the relative positions between the neighboring points and the central point are invariant. vector So the feature

Then we transform the slope S to the corresponding an gle 8. the range of 8 is

[0,360)

is changed to overcome the rotation in following,

by counterclockwise from

and the new feature vector is

V min.

79

Proceedings of the 2011 International Conference on Wavelet "\nalysis and Pattern Recognition, Guilin, 10-13 July, 2011
W is
(1 3) where j value
=

weight vector
7 4 5 6 3 0 7 2 1 8 8 9 10 11 12 13 14 0 6 5 4 3 2 1 16 15

0,1,2... , Sr - 1,
a

N+ is the positive integer col Then the weighted Last

lection and

WV of every feature vector in SV is calculated.

is any positive integer.

the pattern corresponding to the smallest weighted value in ogy net, we can get a feature vector

Vmin. Vmin is depend on the num ber of vertex. WV

WV is presented by Vmin.

For every vertex in the topol The num ber of

I-1 t-8r

I-

Vweight WT '\:"' (i-I) i Vweight L- a

(1 4)

Figure 4. N P is the number of points which have the distance r from the central point. The central point is denoted by 0, and other digits pres ents the point related to N P. When radius R is N P is

S min be the vector set of all the Vmin in one finger print image. in which, the V";'in presents the Vmin corre
Let sponding to the ith minutiae, and the number of minutiae.

where i

1,2, ... , Sr.

minutiaenum presents

1,

N P is 8; When radius R is 2,

16.

When radius R is r, N P is 8r, ex The template fingerprint minutiae vector T and the input

analogia.

Now the operation of circular shift, left shift or right shift, is used to deal with the feature vector be the feature vector set translated from V:

V.

Let

SV

SV
where i
=

{VilVi

CircularShijt(V)}

(1 2)

S min each, STmin and SImin, and the corresponding Vmin is presented as VT min and VImin. min SImin [VIImin' VI2 in' .. " V .1.rmin uti aenU m j (1 6) m STmin [VT;'in, VT';'in, ..., VT,:l:utiae nu mj (1 7)
template fingerprint minutiae vector I have a denoted by
= =

ments on V and VOis V itself, which is shown in Fig.6.

circular shift, and Viis the feature vector after shifting i ele

0 , 1 , 2...Sr-1, CircularShijt is the function of

SV

presents the various patterns caused by rotation. In order to make the fingerprint matching algorithm robust to rotation, instead of patterns in

VI:nin is composed of VIeight and VI;osition VT';' in is composed of VT;ei ht and VTiosition The matching is divided into locaf matching and global matching. Considering VI:nin and VT";'in, we can calcu
where ,while late the local matching score. Let the starting score be zero. the position of the element the distance sector

SV ,

it uses one pattern

Vmin.

First

we add the weights to the elements of the feature vector, the

c small angle V 8r

VTj.n is be traversed. If a non-zero value element is found, mt .

2 .-

8r-l

8r-2

high position

DS

DS is presented as
(Bb, Ub) (Bb,Sr)

VTgosition is taken down.

Bb> O,Ub < Bb> O,Ub:;:' Bb:::; O,Ub <

Then

Sr Sr Sr
(1 8)

(0,Ub)
upper boundary of

8r 8r2 8rl

where Bb and Ub represent the below boundary and the

3 2 1

DS,

Bb

large angle

low position

VTj ton + DT H and DTH is the threshold determining POSt t

whether two minutiae from different fingerprmts are at the If It only has elements of zero value in
.

VT;;"sition - DTH, Ub

same position of the feature vector. Figure 5. The transform process from a fea ture circle

to a feature vector

V.

there is no matching point in

of VI:nin, VI:nin and VT';'in. Otherwise the elements of non-zero value VTn

DS

DS

between

80

Proceedings of the 2011 International Conference on Wavelet "\nalysis and Pattern Recognition, Guilin, 10-13 July, 2011

vious methods are all finished on FVC2002 database. Table 1 shows the results. method Tico[14] Chen[5] Benhammadi[l] Lumini[9] Liu[8] Zsolt[7] Our method DBl(%) 4. 0 4. 6 4. 2 4. 2 4. 3 4. 3 3. 5 DB2(%) 4. 2 5. 3 2. 6 3. 9 4. 0 4. 8 3. 9 DB3(%) 7.1 8. 9 10. 6 15.0 10. 1 7. 5 5.6 DB4(%) 7. 7 6. 7 5. 1 6. 7 4. 6 4. 6 4. 2

Figure

6.

The transform process from a fea

Table

1. EER

ture circle C to a feature vector V.

4
in VT in and the elements V I ;::in corresponding in V I ;"in the VT and VI '!,ight is described by DW, there is a threshold called VT H, if the absolute value of DW is are compared using the weight. The difference between

Snmmary and Future Work

In this paper, we have developed a novel topology-based

:ight

representation technique for fingerprint verification. It uti lizes neighbor structure information to match the point pat terns at principle preliminarily, and carries on global match ing at last. The proposed representation is invariant to rota tion and translation by analyzing the relationship of minu tiae(genuine and spurious), not capturing more information about fingerprints. As the matching method by relationship of point sets, the structures are robust to missing and spuri ous minutiae. Then, the structures are not influenced by the resolution of fingerprint images, which is convenient to deal with the fingerprint matching at different resolutions. In ad dition, local noise only affects the structures in a local field, resulting from the structure has a good structure stability. If the noise affect the whole image, like white noise, this problem should be solved in preprocessing stage, and it has been given some attracting solution in many previous work. Experimental results show that the performance of the pro posed algorithm is efficient. In future work, we would like to optimize our algorithm further and try to achieve the bet ter performance, which includes the consideration to deal with the large difference of two feature sets in waiting com parison. Then we will consider affect to the structures by the nonlinear transformations of fingerprint images. Con sidering the structures as a "tension" draw on analogy with

smaller than VT H, the element is matched, and the local matching scoreLMSbetween the two feature vector adds one to itself.

One feature vector VT ';'in in STmin is compared with every feature vector in SImin. The feature vector which has the biggest LMS is the matching vector to the

VT ';'in. Every feature vector in STmin has a matching vec

tor in SImin, and the two feature vectors are considered as a matching pair, each of which has aLMS. The number of local matching scores is equal to the element num ber of STmin. Then the sum ofLMSis the global matching score GMSbetweenSTmin andSlmin. Let N umberT is the number of non-zero value element inSTmin. If matched. image T and I are matched. Otherwise, T and I are not

N;:::r,TT is smaller than VT H, the fingerprint

Experiments
We perform experiments on

DB4A

DBIA, DB2A, DB3A,

physics, we want to measure the non-linear transformation through the change of the structures.

of the public databases FVC2002, which contains

3200 fingerprints from 400 different fingers, to evaluate the performance of our algorithm. The test is performed with Matlab under Windows XP on Celeron M, 1.6GHz and 1536M machine. The result is measured by the value of equal error rate (EER) experimented on

ACKNOWLEDGMENT
This work was supported by the grant 60973154 and

DBIA, DB2A, DB3A, DB4A,

61075015 from the NSFC, NCET-07-0338 from the Min istry of Education, China. This work was also partially supported by the Fundamental Research Funds for the Cen tral Universities, HUST:201OZD025, and Hubei Provin-

and compares our algorithm with previous works at evalu ating the performance of the proposed methods. These pre-

81

Proceedings of the 2011 International Conference on Wavelet "\nalysis and Pattern Recognition, Guilin, 10-13 July, 2011

cial Science Foundation under Grant 201OCDA006 and 20 lOCDB06601,China.

[12] JP Pascual Starink and E. Backer. Finding point corre spondences using simulated annealing. PatternRecog

nition, 28(2):231-240,1995.

References
[1] F. Benhammadi,

[13] X. Tan and B. Bhanu.

Fingerprint matching by ge

netic algorithms. Pattern Recognition, 39(3):465-477,

J\![]\J

Amirouche,

H.

Hentous,

2006. [14] M. Tico and P. Kuosmanen. Fingerprint matching

K Bey Beghdad, and M. Aissani. Fingerprint match

ing from minutiae texture maps. Pattern Recognition, 40(1):189-197,2007. [2] Arijit Bishnu, Sandip Das, Subhas C. Nandy, Bhargab B., and Bhattacharya. Simple algorithms for partial point set pattern matching under rigid motion. Pattern

using an orientation-based minutia descriptor. IEEE

Transactions onPattern Analysis andMachine Intelli gence, pages 1009-1014,2003.

[15] R. Ddup D, G. Garg, and P Sharma. curate fingerprint verification. 197. Springer, 2001. [16] T. Van Le, KY. Cheung, and M.H. Nguyen. A fin gerprint recognizer using fuzzy ev olutionary program ming. In P roceedings of the 34th Annual Hawaii Fast and ac

Recognition, 39(9):1662-1671,2006.
[3] R. Cappelli, M. Ferrara, and D. Maltoni. Minutia

In Audio-and Video

Based Biometric Person Authentication, pages 192-

Cylinder-Code: A New Representation and Matching Technique for Fingerprint Recognition. IEEE Trans

actions onPattern Analysis andMachine Intelligence,

2010. [4] S.H. Chang, F.H. Cheng, W.H. Hsu, and G.Z. Wu. Fast algorithm for point pattern matching: invariant to translations, rotations and scale changes. Pattern

International Conference on System Sciences, 2001,

page 7, 2001. [17] DM Weber. A cost effective fingerprint verification

Recognition, 30(2):311-320, 1997.

[5] Y. Chen, S. Dass, and A Jain. Fingerprint quality

algorithm for commercial applications. In Proc. 1992

South African Symposium on Communication and Sig nalProcessing, pages 99-104,1992.

indices for predicting authentication performance. In

Audio-and Video-BasedBiometric Person Authentica tion, pages 160-170. Springer,2005.

[6] AK Jain, Y. Chen, and M. Demirkus. ridges: Pores and

High-resolution fingerprint matching using

level 3 features. IEEE Transactions onPattern Analy

sis and Machine Intelligence, pages 15-27, 2007.

[7] ZM Kovacs-Vajna. A fingerprint verification system

based on triangular matching and dynamic time warp ing. Pattern Analysis andMachine Intelligence, IEEE

Transactions on, 22(11):1266-1276,2002.

[8] N. Liu, Y. Yin, and H. Zhang. A fingerprint matching algorithm based on Delaunay triangulation net. 2005. [9] A Lumini and L. Nanni. Two-class fingerprint

matcher. Pattern Recognition, 39(4):714-716,2006. [10] D. 11altoni, D. Maio, AK Jain, and S. Prabhakar.

Handbook offingerprint recognition. Springer-Verlag

New York Inc, 2009. [11] S. Pankanti, S. Prabhakar, and AK Jain. On the in dividuality of fingerprints. IEEE Transactions onPat

tern Analysis and Machine Intelligence, 24(8):10101025,2002.

82