Maj. Sanpachai Huvanandana Dept. of Electrical Engineering and Information Technology Chulachomklao Royal Military Academy
An Automatic Fingerprint Identification System using multiple classifiers
Abstract
Biometric applications have been wildly used nowadays. A fingerprint identification system is the most popular system due to the affordable price of a fingerprint scanner. An automatic fingerprint identification system (AFIS) using multiple classifiers is presented in this paper. The system consists of several steps: fingerprint enhancement, minutiae extraction, texture feature extraction, texture m a t c h i n g , a n d m i n u t i a e m a t c h i n g . T h e proposed matching system has been tested on a reasonably large fingerprint database and the e x p e r i m e n t a l r e s u l t s s h o w a n e f f e c t i v e performance.
1. Introduction
Among various fingerprint identification svstems. the traditional methods based on m i n u t i a e h a v e b e e n w i d e l y u s e d [ 3 , 5 , 8 ] because the topological structure of minutiae of a fingerprint is unique and invariant with aging. Jain et. al. proposed an novel alignment - based matching algorithm in [5] where the
54 zr:dr:rtitrnrl \.d. 2547
authors used ridges associated with minutiae to obtain transformation parameters and elastic matching had been used for matching.
T h e r e s u l t w a s a c c u r a t e b u t t h e s y s t e m required a large storage because all ridges i n f o r m a t i o n m u s t b e s a v e d . A n o t h e r fingerprint identification approach, so called a filterbank - based fingerprint matching [6]
used a bank of Gabor filters to capture the
details in fingerprint as a compact fixed length
F i n g e r C o d e . T h e m a t c h i n g i s b a s e d o n
Euclidean distance between two Finger Code,
hence, it is faster than traditional method but
r e p o r t e d a l o w e r p e r f o r m a n c e . In t h e
filterbank - based approach, convolution with
Gabor filters is the major computation to the
overall feature extraction time. In this papeq
we focus on a framework of an automatic
f i n g e r p r i n t i d e n t i f i c a t i o n u s i n g m u l t i p l e
c l a s s i f i e r s th a t c o n s i s t o f f i n g e r p r i n t
classification, texture matching, and minutiae
matching. The flow chart of our proposed
hybrid fingerprint identification system is
shown in Figure 1.
In Section 2, the fingerprint enhancement a n d m i n u t i a e e x t r a c t i o n f r a m e w o r k a r e explained. Section 3 discusses a fingerprint m a t c h i n g s y s t e m u s i n g a f i l t e r b a n k . T h e m u l t i c l a s s i f i e r fi n g e r p r i n t i d e n t i f i c a t i o n system is proposed in Section 4. The experi- mental result is shown in Section 5, followed by conclusions in Section 6.
2. Enhancement and Minutiae Extraction
2.1 LocalRidge Orientation (LRO ) Estimation
The optimal ridge direction is defined a s t h e d i r e c t i o n a l f i e l d t h a t s h o w s t h e dominant local structure orientation. We propose an orientation angle approximation
Texture snd fingerprint class database, and
distance matrix.
Relcct ldentificdpcrson
Figu,r e 7 : The proposed multiclassifier fingerprint identificatimt system.
via Wavelet Tiansform. Level - | of a 2 - D wavelet represents 4 sub - bands, called LL, HL, LH, and HH, which correspond to low
frequency, horizontal detail, vertical detail, and diagonal detail coefficients, respectively.
Let @ be an orientation angle, Jy and Jx are the vertical (LH) and horizontal (HL) w a v e l e t c o e f f i c i e n t s , r e s p e c t i v e l y . T h e estimate of the dominant orientation @ at the center (n,r) of a P x P non - overlap block is given by,
T h e a d v a n t a g e o f u s i n g w a v e l e t coefficients for orientation angle estimation is that the LH and HL wavelet coefficients, which can be viewed as smoothed gradient information, can be directly calculated and treated as partial derivative results. More detail can be found in [9].
Figure 2: (a) Original image (b) Enhanced fingerprint.
2.2 lmplementation of Spatial Filtering
The purpose of fingerprint enhancement filtering is to smooth (interpolate) the ridges along the same ridge line, characterize edges, and sharpened details in perpendicular of the ridge direction. The combination of any low - pass and high - pass filtering in perpendicular direction can be used to smooth and sharpen the images respectively. For simplicity, we choose a Gaussian kemel with zero mean and standard deviation (o) in our implementation.
The low - pass and high - pass Gaussian in spatial domain can be expressed as [{]:
j
y,
J J
2
a
" , r = O ' ' * ' [
) I
I
I( 1 )
Ware let 1'ransfom
Enhan**ment Alsaritim
; Rsferencc Pornt I Detetmrnation
I:tFuuurliaalr:vlnoarrn#r SS
( ' , , , . r ) ( 2 ) { ; , (x ) = = } e x p l x ' 1 2 o , 7
I Zlto;
(;,,( r). -E+e*p(- u'/zol,)- $ exp( y'f zol,l. 1j1
\ . l t t o i l z r o , 2
w h e r e A Z B , o , r > o r r .
H e n c e , f r o m t h e s e p a r a b l e p r o d u c t property, 2 - D filters can be realized by the multiplication of Eqs. (2) and (3). Take for example, without loss of generality, let us a s s u m e t h a t t h e r i d g e s a r e p e r f e c t l y h o r i z o n t a l . T h e 2 - D ( m x m ) f i l t e r t h a t enhances horizontal ridges is given by,
(;(,!.,v) = (;i-rX;1.v1,(- n r I |2 s.r,.r < {ot - l}l 2. un,l r tyoll. (4)
The same analysis can be applied for e n h a n c i n g o t h e r r i d g e d i r e c t i o n s w i t h different orientation filters. By applying the appropriate filters, each directional ridge can b e e n h a n c e d . B y r o t a t i n g t h i s h o r i z o n t a l G a u s s i a n f i l t e r m a s k , G ( x , y ) , w i t h appropriate direction, enhancement filters for different fingerprint orientations can be found. More specifically, let N be the number of ridge directions, the directional filter can be written as G, (x, )), q : the i'hrotation a n g l e , a n d i = {1 ,2,..,Nf. From Section 2.1, the estimation of the ridge is quantized into eight different directions, hence, eight direc- tional filters are needed in order to enhance t h e w h o l e f i n g e r p r i n t i m a g e s , 0 , = n i 18, { i = 0 , 1 , .. . , 7 } .
2.3 Enhancement
The enhanced fingerprint image, F",, (*, )), is determined by filtering fingerprint image, I(x, y), with an appropriate directional filter mask, G, (x, )), where 0 depends on the quantized ridge directions of the fingerprint image, @' (x, y). The enhancement can thus be expressed as:
m i 2 h l |
F-",(x,y)= t fcog,v)t(x-un,y-v4. (5)
2.4 Post - Processing
Post processing consists of binarization, t h i n n i n g , f e a t u r e e x t r a c t i o n , a n d f a l s e minutiae reduction. The details can be found
in [31.
After post - processing, we obtain fifty minutiae (endlng and bifurcation [1]) that have the highest coherency [4]. The estimate of the Coherency Cc at the center (n, r) of a P x P non - overlap block is given by,
where jx and Jy are the vertical (LH) and horizontal (HL) wavelet coefficients, respec- tively.
The feature vector is constructed based o n t h e t y p e ( e n d i n g a n d b i f u r c a t i o n ) , X - coordinate, Y - coordinate, and clrientation of minutiae, see Figure 4. At each minutiae point (called reference minutiae), we generate a feature vector by collecting five nearest neighboring minutiae and calculate distance, different of ridge orientation, and ridge count between the reference and neighbor minutiae pair. VUe also store orientation difference and distance between core and reference minutiae.
Hence, each fingerprint contains 50 vectors and each vector has 23 features, which are rotation and translation invariant. \7e can directly use them in minutiae matching stage.
3. Texture Feature Extraction
The texture feature extraction procedure is modified from [6]. Our texture feature extraction requires smaller computation and storage. Instead of using original fingerprint i m a g e , w e o b t a i n l o w f r e q u e n c y w a v e l e t coefficient (LL band) from Section 2.1. The texture feature extraction is as following:
1) Calculate a reference point from L o c a l R i d g e O r i e n t a t i o n ( L R O ) t h a t i s calculated in Section 2.1 using [3].
(6)
rrrdr:r4rlicrnr 7 vr.fi. zs47
2 ) D i v l d e a n L L b a n d o f W a v e l e t coefficients into 5 concentric bands with respect to the reference point. Each band has
16 sectors and has 10 pixels wide (See Fig' 3, fingerprint A1).
3 ) C o n v o l v e a n i m a g e w i t h 8 directional Gabor filters. We set the filter frequency equal to the average ridge frequency of LL band Wavelet coefficient (f=1/5). The f i l t e r m a s k s i z e o f 1 1 x 1 1 i s u s e d f o r o u r e x p e r i m e n t a n d a s t a n d a r d d e v i a t i o n o f Gaussian envelope o" and o, is set to 2.0.
4 ) R o t a t e a f e a t u r e v e c t o r 2 s t e P s clockwise and 2 steps counter' clockwise and o b t a i n a q u a n t i z e d r o t a t e d F i n g e r C o d e . Each step is corresponded to 72.5 degree.
Each fingerprint has five feature vectors that can be used in matching stage' The major difference between our method and [6]
is the use of LL band, the normalized value, and the reduced feature vector size. The E u c l i d e a n d i s t a n c e i s u s e d t o o b t a i n matching scores.
4 . T h e C o m p l e t e M a t c h i n s System
W e d i v i d e t h e c o m p l e t e m a t c h i n g system into 2 stages, a texture matching and a minutiae matching system.
4.1 Texture matching
We have observed statistically that the distance between two fingerprints, which have similar texture characteristics, is closer than f i n g e r p r i n t s t h a t h a v e d i f f e r e n t t e x t u r e c h a r a c t e r i s t i c s ( s e e Fig. 3). The texture matching in our paper has been adopted from [6]. The texture matching process based on our experiment using distance score is as following:
1) Given a set of fingerprint database, we categorize each fingerprint into classes, i.e., arch, left loop, right loop and double loop using the method in [7] and match each fingerprint against the others in the database.
2) The matching distances have been sorted, twenty percent of the fingerprints that have minimum distance have been obtained into the next stage.
t] c
Figure 3: The exomple of texture d'istance b e t w e e n i m a g e s . A l a n d A 2 i s t h e s a m e fingerprint with different impression. A7' A2 andB have similor charocteristic ridge.
4,2 Minttiae identif ication
The fingerprints obtained from Section 4 . 1 h a v e b e e n e n t e r e d in t o a m i n u t i a e matching system. Given two sets of minutiae feature vector (one from query fingerprint and the other from the matched fingerprint called template that has a minimum distance), we further match each vector in query against one in template using the concept of correlation m a t c h i n g [8] and count the number of s i m i l a r i t y . T o a v o i d a n i n e x a c t n e s s , t h e m a t c h i n g s h o u l d b e t o l e r a t e d t o s o m e bounding distance. The similarity value is incremented by one if all of the following criteria are met:
1) If a pair of reference and neighbor m i n u t i a e b e t w e e n q u e r y a n d t e m p l a t e fingerprint have the same type (ending and bifurcation).
2 ) I f t h e d i s t a n c e a n d o r i e n t a t i o n d i f f e r e n c e s b e t w e e n c o r e a n d r e f e r e n c e
T:s6uuuruiour:vlaaartn6r 57
minutiae from both query and template are in a bounding distance. In our experiment, we empirically chose 40 pixels for a bounding distance and 22.5 degrees for orientation difference.
3) lf a pair of reference and neighbor m i n u t i a e b e t w e e n q u e r y a n d t e m p l a t e fingerprint is in a bounding distance. \7e empirically chose 15 pixels for a bounding d i s t a n c e , 2 2 . 5 d e g r e e s f o r o r i e n t a t i o n difference, and 1 for ridge count.
T h e r e f e r e n c e m i n u t i a e t h a t h a s similarity value equal or greater than one is m a r k e d i n a c o r r e s p o n d i n g p a i r s . P o s e c l u s t e r i n g h a s b e e n u s e d t o c a l c u l a t e t r a n s f o r m a t i o n i n e v e r y p a i r o f m a r k e d m i n u t i a e i n q u e r y f i n g e r p r i n t t o t h e correspondence pair minutiae in template fingerprint. The computed transformation parameters X and Y translation and rotation
<p, are used to colrect the image transformation.
After transforming, we match two fingerprint minutiae against each other based on the b o u n d i n g b o x a n d c o u n t t h e n u m b e r o f matched minutiae, see Figure 5. The fingerprint that has a highest matching score has been selected. If the number of matching score is greater than a certain threshold, we identify as a matched fingerprint. If the number of matching score is not greater than a certain threshold, the system identifies as a reject and retrv has been requested.
Figure 4: The example of extracted feature (minutiae) with two fingerprints.
Figure 5:The exanple of matchedmhwtinewith the same minutiae type (solid) and different minutiae tyDe (dssh)
5. Experimental and Evaluation Result
For the experiments, fingerprint images were collected from 100 people with 8 images for each person. 500 images were selected and processed to construct database. The texture feature and minutiae feature are obtained as described in Section 2 and 3. 'We
construct a distance matrix and a matching system as in Section 4. The remaining 300 fingerprints that are not in the database were used to test t h e p r o p o s e d a l g o r i t h m . T h e s i m u l a t i o n results with different threshold are shown in Thble 1. Table 2 and 3 show the simulation results of an individual system. The lowest distance matching has been obtained from a texture matching and the highest matching score has been obtained from a minutiae identification.
Table 7: Simulation results using multiple classifier
Table 2: Simulation results using minutiae identification
'Iexture lhreshold
va.lue
Minutiae threshold
value
False Acceptanc€
lo/"\
False Reject
(o/"\
l 0 0.04 5
l ? 0.02 6 . 5
Texture threshold
value
Minutire threshold
value
False Acceptance
{o/"\
False Reject
{o/"\
l 0 0.04 6 . J
l 2 0.02 7.2
rr:dr:rr.rigrnr a tl.F1. 2547
Table 3: Simulation matching
results using texture
6. Summary and Conclusions
W e h a v e d e v e l o p e d a f i n g e r p r i n t matching system using multiple classifiers.
The system consists of several sequential steps:
f i n g e r p r i n t e n h a n c e m e n t a n d m i n u t i a e extraction, texture feature extraction using filterbank, texture matching, and minutiae verification. Our system distinguishes itself from other approach at proposed orientation estimation using wavelet coefficient, texture
feature extraction, and multiple classifiers s y s t e m , e . g . c o m b i n i n g t e x t u r e m a t c h i n g , fingerprint classifying, and minutiae verifica- t i o n . T h e e x p e r i m e n t a l r e s u l t s a r e v e r y convincing. The experimental results show better performance than a single classifier with a little time afford. The drawback of our s y s t e m i s t h e a c c u r a c y o f f i n g e r p r i n t c l a s s i f i c a t i o n . I n t h i s p a p e r , f i n g e r p r i n t classification is based on only the number and location of singularities. Singularities are not always detected completely if only a partial of fingerprint has been enrolled. The globalridge c h a r a c t e r i s t i c s s h o u l d b e i n c o r p o r a t e d . Misclassification results in slow convert, thus, achieving high classification rate can speed up our proposed system. i.;
Reference
[1] Sir Francis Galton, "Finger Prints," Macmillan and Co., 1892.
[2] L. Hong, Y. Wan and A.K. Jain, "Fingerprint Image Enhancement: Algorithms and Performance Evaluation", IEEE Tiansactions on PAMI, Vol. 20, No.8, pp. 777-789, August 1998.
l3l S. Huvanandana, Changick Kim and J. N. Hwang, "Reliable and Fast Fingerprint Identification for Security Applications", IEEE International Conference on Image Processing ( ICIP'2000), Vancouver, Canada, 2000.
I4l Bernd Jahne, Digital Image Processing: Concept, Algorithms, and Scientific Applications, Springer, 1997.
[5] A.K. Jain, L. Hong, and R. Bolle, "On-Line Fingerprint Verification" IEEE Tiansactions on PAMI, Vol. 19, No.4, pp. 302-314, April 1997.
16l A.K. Jain, L. Hong, and R. Bolle, "Filterbank-based Fingerprint Matching" IEEE Transactions on Image Processing, Vol. 9, No. 5, pp. 846-859, May 2000.
t7l K. Karu and A.K. Jain, "Fingerprint Classification", Pattern Recognition, Vol. 29, No. 3, p p . 3 B 9 - 4 0 4 , 1 9 9 6 .
t8l A. \Uahab, S.H.Chin, and E.C.Tan, "Novel approach to automated fingerprint recognition", IEE Proc.-Vis. Image Signal Process., Vol. 145, No. 3, pp. 160-166, June 1998.
[9] S. Huvanandana, S. Malisuwan, and K. Rosesokon, "Orientation Approximation via Wavelet Tiansform", The third international Symposium on Communication and Information Technologies (ISCIT2003), Sept 2003, Songkra, Thailand.
T:.:6ruu-raiaennvlaaorrn6'r S$
