Você está na página 1de 4

Automatic Fingerprint Identification Using Cluster Algorithm

Ren Qun, Tian Jie, He Yuliang and Cheng Jiangang


AILAB, Institute of Automation, the Chinese Academy of Sciences, Beijing, 100080
Email: {qun.ren, jie.tian, yuliang.he, jiangang.cheng} @mail.ia.ac.cn

Abstract many matching problem as N times one-to-one matching.


Robert S. Germain et al.[1,2] proposed flash and geomet-
In this paper, a new fingerprint identification technique ric hashing algorithm, which create geometric keys for
is presented, which mainly consists of three modules: hashing tables based on rotation invariant properties of
enrolment module, identification module and feedback triangles having minutiae as vertices. Some biometric
module. In the identification module, clustering algorithm products in the market use a certain feature of fingerprint
is used to detect similar minutiae group from multiple image, such as average ridge width, to generate index
template images generated from the same finger and cre- table to speed access. All of those methods are inspiring.
ate the cluster core set. An algorithm compares the simi- However, those methods have not fully expressed charac-
arity level between the minutiae of test fingerprint and the teristic relationships among all the template images from
cluster core set and returns a likely list of candidates. In the same finger.
feedback module, we propose a path to learn and train the In this paper, we focus on application of clustering
cluster core vector based on the assessment of cluster analysis to fingerprint identification. Because matched
solution. The experiment results demonstrate that this minutia pair from the two matched images must have
similarity-searching approach proves suitable for one-to- similar expression patterns, while un-matched minutiae
many matching of fingerprints on large-scale databases. should have distinct patterns [3,4]. So, the clustering algo-
With the feedback module the proposed fingerprint identi- rithm can be used in fingerprint identification to detect
fication scheme has inspiring identification performance similar minutiae group from multiple template images
of application. generated from the same finger. We try to classify the
large number of minutiae generated from the same finger
to the succinct clusters, choose a representative ”core” for
each cluster, then we can get a “core feature vector” for a
1. Introduction finger. Thus we can determine the test image whether
come from the finger according its “core feature vector”.
Fingerprint has been used as a good means of personal Moreover, the statistical characteristics are calculated
identification for a long time due to the characteristic of using to modify the result of clustering. In our algorithm,
uniqueness and immutability. Fingerprint recognition is a we solve the identification problem using clustering ap-
rapidly evolving technology, which is being widely used proach to ensure high reliability of the fingerprint
in forensics and has the potential to be used in a large identification.
range of civilian application areas such as access control, In the following sections, we describe in detail our
financial security, information security, etc. Since manual method. Section 2 addresses the main steps of our algo-
fingerprint verification is extremely tedious and time- rithm. Section 3 gives the experimental results. Section 4
consuming, automatic fingerprint identification systems contains the conclusion and discussion.
(AFIS) are in great demand. Identification problem, a
particular case of “point pattern matching”, necessitates a 2. Fingerprint Identification
large database search of individuals to determine whether
a person is already in the database. Hence, the challenge In order to introduce our fingerprint identification
of fingerprint identification application is how to achieve method, a list of notations and some basic mathematical
one-to-many search over a limited period of time. Many background are given below.
fingerprint identification methods have appeared in litera-
ture over the years. The simplest way is to regard one-to- 2.1. Notation and definition

Supported by the National Natural Science Foundation of China under


Grant Nos. 60071002,60072007,69931010,60172057

1051-4651/02 $17.00 (c) 2002 IEEE


A test image, I, a gray-level fingerprint image, is a ac- clustering is application mainly in the following three
quired fingerprint images of the users who intend to ac- components:
cess the system. (1) Enrolment module: used to get the minutiae sets
Let Ti = {Ti , j , j = 1, L, n} , i = 1, L , m the template im- and stored template information in the database;
age set of the ith collected finger, where m is the number (2) Identification module: using clustering approach,
of the fingers in the image database and n is the number of identify which finger the test image is from.
template sample per finger. (3) Feedback module, we can modify the cluster result
Let M I = {M iI , i = 1, L LI } be the minutiae set of the test by statistic based on cluster assessment.

image, LI is the minutiae number of the test image. 2.3. Enrolment module
Let M iT, j = {M iT, j ,k , k = 1, L , LTi , j } , i = 1, L , m , j = 1,L, n be
fingerprint image Feature Template File
the minutiae set of the jth template image of ith finger, acquisition
Preprocess
Extraction (Feature Vector)

LTi , j is the minutiae number of the template image. Figure 2. a block diagram of the enrolment
Let N = {e1,L, en} be a set of n minutiae, and let module
Figure 2 shows a block diagram of the enrolment mod-
C = {C1 ,L , Cl } be a partition of N into disjoint subsets.
ule, which consist of fingerprint image acquisition, image
Each subset is called a cluster, and C is called a clustering pre-processing, feature extraction, and creation of tem-
solution, or simply a clustering. Two elements ei and plate file. The template file consists of three primary types
of data: the offset x and y value, (∆x , ∆y ) , using for image
e j are called mates with respect to C if they are mem-
segments, the position of singular point (Delta or Core
bers of the same cluster in C. point), and minutia information. Each minutia is associ-
ated with a real-valued vector, called its feature vector,
2.2. Outline of algorithm which contains some measurements on the minutiae. Gen-
Fingerprint Image
erally, the feature vector of the minutia is defined as:
M k = (x k y k θ k t k )
T
Captor (1)

where ( xk , y k ) is the coordinate of the minutiae, θ k is its


(1)Enollment Module
Test Image Template Image
Ti={Ti,1,...,Ti,n}
I t
Minutiae Minutiae local ridge direction and k the minutia type (bifurcation
Extrator Extrator or ridge ending).
(2)Identification Module
Template File
Minutiae Set
In the enrolment module, template file created from
Minutiae Set
MI MTi,j every template image will be saved in the system data-
base. The system database stores the template records of
Cluster
Analyzer System all the individuals that have access to the system.
Database
Cluster Solution
Ci={Ci,1,...,Ci,L} 2.4. Identification Module
Cluster Core Set (3)Feadback Module
KTi
alignment computering forming
Respectively clustering
template images core feature candidate list
Temp System replace Ch, KTh
Database to C'h,K'Th Figure 3. block diagram of enrolment module
C'h is better
than Ch
Figure 3 shows a block diagram of the identification
KTh module, which consists of images alignment, clustering,
Assessment of
Coarse Clustering computing “core feature vector”, and forming candidate
Matchor list. Algorithm is described as follow:
Cluster Solution
C'h={C'h,1,...,C'h,S} Input: the minutiae set of all the n template images of
Yes No Kernel Set K'Th
the ith finger M iT = {M iT, j , j = 1,L , n} , here i = 1, L , m
Add MTh to
Cadidate List
Un-match Cluster
Analyzer Output: a clustering solution Ci = {C1 , L , Cl }, and the
“core feature vector” set K iT = {K iT,1 ,L , K iT,l }
Yes
Fine Matchor
Minutiae Set Main steps:
MTh and MI Step 1: Alignment stage,
Figure 1. The flowchart of the fingerprint identi- Transformations such as translation, rotation and scal-
fication algorithm ing between the minutiae set M iT,1 and, j = 2, L, n are
The flowchart of the fingerprint identification algo-
rithm is shown in Figure 1. In our identification system, respectively estimated and all of the minutiae sets M iT, j

1051-4651/02 $17.00 (c) 2002 IEEE


j = 2, L, n are aligned with the minutiae set M iT,1 ac- Step 4: Stage of forming “cluster cores”
cording to the estimated parameters. (See figure 6) This is a post-processing phase should be added to the
Step 2: Stage of similarity measure. clustering algorithm to choose a representative “median
Define a good similarity measure function. (See Figure element” for each cluster, which will be helpful to the
one-to many-matching.
4). Set M i , M j is the minutia respectively in different
Set
template image, we define M i is similar with M j , if Ci = {M i , j , j = 1, L , L} , (3)
satisfied that Where, M i , j = (x i , j y i , j θ i , j t i , j )T , j = 1,L , L
(1) M i − M j < Thre and Then the vector K iT , defined as cores of the ith cluster Ci,
(2) Respectively exist M ' i and M '' i , M ' j and M '' j describes the cluster‘s local structure characteristic.
K iT = (x c ,i y c ,i θ c ,i ri ) (4)
T
located in the l-nearest neighborhood of M i , M j . s.t.
where,
M 'i − M ' j < T and M '' i − M '' j < T (2)
 L   L   L 
x c ,i =  ∑ xi , j  L , y c ,i =  ∑ y i , j  L , θ c ,i =  ∑ θ i , j  L
where both Thre and T are threshold parameters.  j =1   j =1   j =1 
Then measure similarity values between the minutiae
of different template samples from the same finger.
ri = λ ⋅ max
j ,k∈L and j ≠ k
(dis(M i, j , M i ,k )) , (5)

( λ is a parameter larger than one.)


θ r
T
T
i, b i , b
Step 5: Coarse matching stage,
With a good choice of a similarity measure, forming a
KiT,b
Ci , b Ci ,c set of “cluster cores” and then identify the test image is
KiT, c
very effective strategy. The coarse matchor is used to
determine whether the minutiae set MI is similar with the
every element of cluster core set K hT , that is, whether
most of the points in the minutiae set MI are dropped into
riT,a
the regions expressed by K hT . If yes, Add finger h in a
KiT,a likely list of candidates. Search whole element list of
Ci ,a cluster core set K T , then terminate identification and be
Figure 4. Clique graph out of this module.
Step 3: Clustering stage.
Create Clique Graphs using Ben-Dor and Yahkini’s 2.5. Feedback Module
approach [5], because this approach has the advantage in
compose MI with cluster modified core
the presence of a known similarity measure for the minu- MT
clustering
assessment feature
tiae. The basic idea of the method is that ideally, all minu-
tiae within each cluster should be similar to one another, Figure 6. a block diagram of feedback module
and not similar to any minutiae of other clusters. Hence, We propose a feedback module for cluster cores, which
we got the clustering solution from the minutiae sets of n is based on an analysis of the cluster assessment. The goal
template image. in the feedback module is to learn the characteristics of
Things we should notice that in a potent and reliable cluster core in every cluster. We believe that by modifica-
cluster the number of mate should larger than half of the tion the cluster core set data, some of the mistakes in the
number of the template images. And under the clique end-to-end sequential processing can be eliminated. Fig-
ure 6 shows a block diagram of the feedback module, The
graph model, noise points will tend to fall into extremely
small clusters (often singletons), which can subsequently main steps as follow:
be ignored. See figure 5, there are some kinds of possible Step 1: Using the fine matchor (one-to-one matched),
outcomes in clustering. match the minutiae set of test image with each minutiae
set in candidate list and search the best-matched finger
index (ith).
a b c d Step 2: Using Cluster analyzer computing the new
*
Figure 5. possible outcomes in clustering cluster solution Ci of the minutiae set consisted of
a. Mates in the same cluster
M I and M iT .
b. Minutiae with un-similar angle (non-mates)
c. Singletons Step 3: To clustering solution of ith finger, respectively
d. Minutia lying outside tolerance box of a measure homogeneity and separation of
cluster (non-mates) Ci ( H ave and S ave ) and those of Ci* *
( H ave *
and S ave ).In

1051-4651/02 $17.00 (c) 2002 IEEE


H ave is evaluated by the [1] R.S.Germain,”Large Scale Systems,” Biometrics: Personal
detail, for minutiae, homogeneity Identification in Networked Soc., A.Jain, R. Bolle, and S.
average similarity between the feature vector of a minutia Pankanti, eds., pp.311-325, 1998.
S
and that of its cluster. Separation ave is evaluated by the [2] Robert S. Germain, Andrea Califano, and Scott Colville,
Fingerprint Matching Using Transformation Parameter
weighted average similarity between cluster minutiae. In Clustering, IEEE Computer Science & Engineering, Oct.-
computing the above measures, singletons are considered Dec. 1997,pp. 42-49.
as additional one-member clusters. [3] A. K. Jain, R. P. W. Duin and J. Mao, Statistical Pattern
* Recognition: A Review, IEEE Transactions on PAMI,
Step 4: a solution improves if H ave > H ave and if
Vol. 22, No. 1, Jan. 2000.
*
*
S ave < S ave [5]. If the ratio of H ave *
and S ave is larger than [4] Amir Ben-Dor and Zohar Yakhini. Clustering gene expres-
sion patterns. In Proceedings of the Third Annual Interna-
the ratio of H ave and S ave , we added the minutiae M I to tional Conference on Computational Molecular Biology
the template minutiae sets, substitute Ci* for Ci and (RECOMB 99), 1999.
[5] R. Sharan, R. Elkon and R. Shamir.Cluster Analysis and its
reforming the cluster core using the method in last subsec- Applications to Gene Expression Data.
tion. http://www.math.tau.ac.il/~roded/schering.pdf

3. Experimental results
The performance of prototype fingerprint identification
system using clustering is evaluated in FVC2000 finger-
print image database. This database has four sub-
databases. Each sub-database is 110 fingers wide and 8
1_1.tif 1_2.tif 1_3.tif
samples per finger in depth. Figure 7 shows a cluster solu-
tion of the first finger of the subsets DB1-A. In this case,
18 clusters were labelled in sequence. See Figure 8, its
cluster core set is non-corresponding with the test image.
Research and development into refinements of the tech-
nology, including tests into large databases, is continuing.
We can’t compare our algorithm with those of others for
1_4.tif 1_5.tif 1_6.tif
several reasons. The main reason is that there is lack of
the standard database and measurements. Computing the
compatibility score and the identification rate are sug-
gested. The compatibility score for a clustering solution is
simply the number of pairs that are mates or non-mates in
both the true labelling and in the clustering solution. The
identification rate is the number of correctly identified
image divided by the total number of pairs. It is evident 1_7.tif
that in all cases performance was substantially on the
dataset.

4. Conclusion and Discussion


An automatic fingerprint identification system using
cluster algorithm has been designed based on three main
Clustering solution 1_8.tif
modules: enrolment module, identification module and
feedback module. In identification module, we use cluster Figure 7. Cluster sets
algorithm to show the coincidence of the features. In
feedback module, we modify cluster solution. All the
above makes our algorithm more able to distinguish im-
ages from different fingers. However, in reality, achieving
satisfactory fulfillment of all the discrepant requirements
is still important problem. The algorithms must be as fast
as possible for comfortable use in applications with large a b. 2_1.tif c
number of users. Figure 8. a .Clustering solution b. test image
2_1.tif of db1_a (FVC2000) c. identification
References

1051-4651/02 $17.00 (c) 2002 IEEE

Você também pode gostar