Finger-vein image recognition combining modified hausdorff distance with minutiae feature matching

doi:10.4236/jbise.2009.24040

J. Biomedical Science and Engineering, 2009, 2, 261-272

doi: 10.4236/jbise.2009.24040 Published Online August 2009 (http://www.SciRP.org/journal/jbise/

JBiSE

).

Published Online August 2009 in SciRes. http://www.scirp.org/journal/jbise

Finger-vein image recognition combining modified

hausdorff distance with minutiae feature matching

Cheng-Bo Yu, Hua-Feng Qin, Lian Zhang, Yan-Zhe Cui

Chongqing Institute of Technology, Chongqing 400050, China

Email: yuchengbo@cqit.edu.cn

Received 6 January 2009; revised 22 February 2009; accepted 26 February 2009.

ABSTRACT

In this paper, we propose a novel method for

finger-vein recognition. We extract the features

of the vein patterns for recognition. Then, the

minutiae features included bifurcation points

and ending points are extracted from these vein

patterns. These feature points are used as a

geometric representation of the vein patterns

shape. Finally, the modified Hausdorff distance

algorithm is provided to evaluate the identifica-

tion ability among all possible relative positions

of the vein patterns shape. This algorithm has

been widely used for comparing point sets or

edge maps since it does not require point cor-

respondence. Experimental results show these

minutiae feature points can be used to perform

personal verification tasks as a geometric rep-

resentation of the vein patterns shape. Fur-

thermore, in this developed method. we can

achieve robust image matching under different

lighting conditions.

Keywords: Biometrics; Finger-Vein Verification;

Gabor Enhancement; Minutiae Matching; Modified

Hausdorff Distance

1. INTRODUCTION

Biometrics is the science of identifying a person using

their physiological or behavioral features. Recently, vein

pattern biometrics has attracted increasing interest from

many research communities. Finger-vein recognition is a

new biometrica identification technology using the fact

that different person has a different finger-vein pattern

[2,3,4,5]. Compared with fingerprint recognition, the

advantages of the finger-vein recognition are [1]: 1) Do

not need to consider the condition of the skin surface and

can prevent the artificial finger; 2) Increase the forgery

difficulty by using the invisible features inside the hu-

man body which only appears under the infrared light; 3)

Non-contact recognition has no bad effect on public

health. The properties of uniqueness, stability and strong

immunity to forgery of the vein pattern make it become

a potentially good biometric which offers secure and

reliable for person identification. A typical vein pattern

biometric system consists of five processing stages [5]:

image acquisition, image enhancement, vein pattern

segmentation, feature extraction and matching. During

the image acquisition stage, vein patterns are usually

obtained using infrared imaging technologies. One

method is using a far-infrared camera to acquire the vein

pattern images of finger-vein [1,2,6]. In order to get the

shape representation of the pattern, after obtaining the

images, vein pattern is separated and extracted from the

background. Finally, a robust image similarity measure-

ment is imperative for matching images under different

conditions. There are a number of previous methods

based on Hausdorff distance function for image match-

ing [7,8,9,10,11]. One of the most distinguished benefits

of Hausdorff distance is that it does not require point

correspondences between two objects or two images.

Dubuisson and Jain [12] developed several modified

Hausdorff distances (MHD) for comparing the edge

maps computed from the gray-scale images. Paumard

proposed a censored Hausdorff distance (CHD) for com-

paring binary images [13]. Takacs introduced the

neighborhood function and associated penalties to ex-

tend the MHD for face recognition [14,24,25]. Guo et al.

proposed a new modified Hausdorff distance which is

weighted by a function derived from the spatial informa-

tion of human face [15]. Furthermore, Lin et al. pro-

posed modified Hausdorff distances with spatial

weighting determined by eigenface features [16]. The

eigenface-based weighting function provides more

weighting on important facial features, such as eyes,

mouth, and face contour. Zhu et al. employed an im-

proved Gabor filter for computing edge maps and ap-

plied a weighted modified Hausdorff distance (WMHD)

in a circular Gabor feature space for comparing images

[17]. LingyuWanga applied the modified Hausdorff dis-

tance based matching scheme to the interesting points

for comparing images [6].

This research is motivated by P. L. Hawkes et al [6].

262 R. Wirestam et al. / J. Biomedical Science and Engineering 2 (2009) 261-272

In this paper, we utilize the modified Hausdorff distance

(MHD) to analyze the spatial similarity between the mi-

nutiae feature sets. Experimental results indicate that

these minutiae feature points can be used to perform

personal verification tasks as a geometric representation

of the vein patterns shape. Furthermore we are able to

achieve robust image matching under different lighting

conditions.

2. PREPROCESSING OF INFRARED VEIN

PATTERN IMAGES

2.1. Finger-Vein Pattern Database

Our finger-vein image capture device mainly consists of

near-infrared light source, lens, cavum, light filter, im-

age ingesting equipment. Vein patterns cannot be ob-

served using normal, visible rays of light since they are

beneath the skin’s surface. However, vein patterns can

be viewed through an image sensor which is sensitive to

near-infrared light (wavelengths between 700 and 1000

nanometers), because near-infrared light passes through

human body tissues and are blocked by pigments such as

hemoglobin or melanin. As hemoglobin exists densely in

blood vessels, near-infrared light shining through causes

the veins to appear as dark shadow lines in the near-

infrared image. To obtain a stable finger-vein pattern ,our

light source adopts near-infrared light source with

wavelength of 890 m, the image ingesting part adopts

near-infrared CCD camera with wavelength of 900 m.

our finger-vein database comprises of 50 distinct sub-

jects, each subject having ten different images, taken at

different times and obtained from middle finger in the

right hand. The images are subject to variations such as

lighting. All the images are taken at a dark homogeneous

background and the finger is in upright, frontal position

(with a tolerance for some side movement). The images

are 256 grey levels per pixel and normalized to

376328



pixels. In our analysis, no pre-processing of

the images was conducted. Example image of five sub-

jects are shown in Figure 1(a).

2.2. The Lmage Normalization

The proportion of the vein area varies mostly at different

time. And for the convenience in further study, the di-

mension size normalization is done in this paper. The

vein image is defined as . That is to say the im-

age zooming will be done. For the difference of acquisi-

tion time, light intensity and the personal palm thickness,

the image gray scale distribution is different highly. If

the image difference is great, the difficulty of image

processing and matching will be increased. So the image

must be normalized. All of the images must be converted

to the standard image of the same mean and variance.

9664

(a) Inpuage t im

(b) Normalized image

Figure 1. The vein image.

R. Wirestam et al. / J. Biomedical Science and Engineering 2 (2009) 261-272 263

For dispelling the illumination effect, a method of

gray scale normalization is adopted.

 

255

,'

,

12

1



GG

Gjip

jip (1)

where is the gray scale value of original image;

is the gray scale value after converted; is

the minimum gray scale of original image; is the

maximum gray scale of original image. Figure 1(b)

shows the vein pattern image after enhancement with

noise reduction and normalization.



jip ,'





jip ,1

G

2

G

2.3. Orientation Image

The orientation image represents an intrinsic property of

the finger-vein images and defines invariant coordinates

for ridges and valleys in a local neighborhood. A number

of methods have been proposed to estimate the orienta-

tion field of fingerprint images [1,10,11,17]. Similarly,

by viewing a finger-vein image as an oriented texture,

we have improved orientation estimation algorithm.

Given a normalized image , the main steps of the

algorithm are as follows:

f

1) Compute the gradients and



ji

x,





ji

y, at

each pixel . Depending on the computational re-

quirement, the gradient operator may vary from the sim-

ple Sobel operator to the more complex Marr-Hildreth

operator.



ji,



2) Estimate the local orientation of each pixel





ji,.

a) Use the window to slide in the original

image, pixel gray value in the center of the window is

.

WW 



jif ,

b) Orientation of window is estimated and is

regarded as orientation of the pixel in the center

of the neighborhood (the size of the neighborhood is

decided by the actual conditions). The equations using to

estimate the local orientation is as follows:

WW 



ji,



 













 2

2

,,2,

W

i

W

iu

W

j

W

jv

yxx vuvujiV (2)

  















2

22 ,,,

W

i

W

iu

W

j

W

jv

yxy vuvujiV (3)

 















 jiV

jiV

ji

y

x

,

arctan

2

1

, (4)

where is the least square estimate of the local

ridge orientation at the block centered at pixel



ji,







ji,.

Mathematically, it represents the direction which is or-

thogonal to the dominant direction of the Fourier spec-

trum of the WW



window.

c) If every point in the image is passed, namely orien-

tation of all pixels are estimated, that is end. Otherwise,

repeating above Step.

3) Due to the presence of noise, corrupted ridge and

valley structures, minutiae, etc. in the input image, the

estimated local ridge orientation,



ji ,



, may not al-

ways be correct. Since local ridge orientation varies

slowly in a local neighborhood where no singular points

appear, a low-pass filter can be used to modify the in-

correct local ridge orientation. In order to perform the

low-pass filtering, the orientation image needs to be

converted into a continuous vector field, which is de-

fined as follows:





jij ,2cos 



i

x,





(5)

and





jij ,2cos 



i

y,





(6)

where





j,i

x



and





ji

y, are the

x

and com-

ponents of the vector field respectively. In the resulting

vector field, the low-pass filtering can be performed as

follows:

y













2

',,,

W

j

W

jv

xx vjuivuhji







2

W

iu

(7)

and

 















2

',,,

W

iu

W

j

W

jv

yy vjuivuhji (8)

where is a two-dimensional low-pass filter with unit

integral and

h



WW



specifies the size of the filter.

4) Compute the local ridge orientation at using



ji,



 















ji

ji

jiO

y

x

,

arctan

2

1

,'

'



(9)

With this algorithm, a fairly smooth orientation field

estimate can be obtained. Figure 2 shows the first ex-

ample of the orientation image estimated with our algo-

rithm in Figure 1.

2.4. Gabor Filter

The configurations of parallel ridges and valleys with

well defined frequency and orientation in a Finger Vein

image provide useful information which helps in re-

moving undesired noise. The sinusoidal-shaped waves of

264 R. Wirestam et al. / J. Biomedical Science and Engineering 2 (2009) 261-272

Figure 2. Orientation fields using and our method (20



W

and ).

10



W

ridges and valleys vary slowly in a local constant orien-

tation. Therefore, a band pass filter is used to tune the

corresponding frequency. Moreover, orientation can ef-

ficiently remove the undesired noise and preserve the

true ridge and valley structures. Gabor filters have both

frequency-selective and orientation-selective properties

and have optimal joint resolution in both spatial and

frequency domains. Therefore, it is appropriate to use

Gabor filters as band pass filters to remove the noise and

preserve true ridge/valley structures.

The circular Gabor filter is an effective tool for texture

analysis [20], and has the following general form:











sincos2exp

2

exp

2

1

,,,, 2

22

2

uyuxi

yx

x

uyxG



















(10)

where 1i, u is the frequency of the sinusoidal

wave,



controls the orientation of the function, and



is the standard deviation of the Gaussian envelope.

To make it more robust against brightness, a discrete

Gabor filter,







,,,, uyxG, is turned to zero DC (direct

current) with the application of the following formula:







2

12

,,,,

,,,,,,,,

~



 



n

uyxG

uyxGuysG

n

ni

n

nj





(11)

Here is the size of the filter. In fact, the imagi-

nary part of the Gabor filter automatically has zero DC

because of odd symmetry. The adjusted Gabor filter is used

to filter the preprocessed images. In our system, we ap-

plied a tuning process to optimize the selection of these

three parameters



2

12n



O



, , 5179.1u1116.0



.

2.5. The Vein Extraction

The image segmentation is very important in the whole

process of finger vein recognition and it is very difficult.

There are several image segmentation methods. The

classic methods are threshold method [26], region grow-

ing method [27], relaxation method [28], edge detection

method [29], split-merge algorithm [31] and so on. The

modern methods are NN method [30], fuzzy clustering

method [32] and so on. The different methods are

adapted in the different application fields. There is no

segmentation method fitting all images. So selecting the

suitable segmentation method is very important. There-

fore, we proposed a completely new method to segregate

the vein image, the principle of the algorithm is as fol-

lowing:

Step 1: Convolution (i=1, 2, ┄, 8) of each

pixel within the



iFgray

99



window in image were calcu-

lated by corresponding eight direction Operator (As

shown in Figure 3). Then get the largest convolution

in eight directions.

maxG



iFgrayMaxG

i





max (12)

Then, maximum is gray value of the point

maxG





max, GnmGray



(13)

Step 2 Threshold segmentation

Step 2.1 The first threshold segmentation







,,

,0

Graym nifGraym n

Graym notherwise



0













(14)

Step 2.2 The second threshold segmentation



NumGraysumGmean



 



,

,,

GmeanifGraym nGmean

Graym nGraym notherwise













(15)

Gmean is the average of non-zero elements in the

image.





Graysum and expresse the sum and

number of non-zero elements respectively.

Num

Step 2.3 Fuzzy enhancement

After the previous twice segmentation, then range of

the gray value is





Gmean,0, pseudo-vein characteris-

tics and noise are likely to exist in the region, cleared the

fuzzy part by using fuzzy enhancement operator. This

will further reduce the noise and removed pseudo-vein.

The algorithm are as following:

1) Calculate the degree of membership, namely:





1

,

,



K

nmGray

nmGrayGumn (16)

2) Calculate the means of all the elements

within

mn

T









1212







pp window

R. Wirestam et al. / J. Biomedical Science and Engineering 2 (2009) 261-272 265

0 0 0 0 0 0 0 00

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 00

0 0 0 0 0 0 0 0 0

3 0 -1 0 -4 0 -1 0 3

0 0 0 0 0 0 0 00







0 0 0 0 0 0 0 0 0

3 0 0 0 0 0 0 0 0

0 0 -1 0 0 0 0 0 0

0 0 0 0 -4 0 0 0 0

0 0 0 0 0 0 -1 0 0

0 0 0 0 0 0 0 0 3

0 0 0 0 0 0 0 0 0







3 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 -1 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 -4 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0-1 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 3

























































(a) 0º (Horizontal) orientation (b) 157.5º orientation (c) 135º orientation

0 0 3 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 -1 0 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 -4 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 0-1 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 3 0 0







0 0 0 0 3 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 -1 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 -4 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 -10 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 3 0 0 0 0







0 0 0 0 0 0 3 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 0 -1 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 -4 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0 0-1 00 0 0 0

0 0 0 0 0 0 0 0 0

0 0 3 00 0 0 0 0

























































(d) 112.5º orientation (e) 90º orientation (f) 67.5º orientation

0 0 0 0 0 0 0 0 3

0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 -1 0 0

0 0 0 0 0 0 0 0 0

0 0 0 0 -4 0 0 0 0

0 0 0 0 0 0 0 0 0

0 0-10 00 0 0 0

0 0 0 0 0 0 0 0 0

3 0 0 0 0 0 0 0 0







0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 3

0 0 0 0 0 0 -1 0 0

0 0 0 0 -4 0 0 0 0

0 0-1 0 0 0 0 0 0

3 000 00 0 0 0

0 0 0 0 0 0 0 0 0







(g) 45º orientation (h) 22.5º orientation

Figure 3. The direction of the operator.

3) Calculate pixel gray and gray degree

of membership adjusted within



', nmGray



'

mn

u







1212





pp

window, its mathematical expression is



2

'

2

1

11 1









 







 









r

mn

mnmn mn

mn

mnmn mn

mn

u

uifu

T

u

uifu

T



 

'

1', mn

uKnmGray

Step 2.4 The third threshold segmentation

T

(17)

Use









1212







pp window sliding in the original

image, pixel gray value in the window center is





', nmGray , then all the pixels within the window com-

pose of a set as the following:













pplklnkmfjiS ,,1,0,1,,,,,  

266 R. Wirestam et al. / J. Biomedical Science and Engineering 2 (2009) 261-272



JBiSE

Obtained the average of all pixels

in the window.



nmSAverage ,

Formula as follows













p

pm

p

pn

nmGray

pp

nmSAverage ',

1212

1

,

(18)

 











p

pm

p

pn

nmSAveragenmGray

pp

nm 2

,',

1212

1

,



(19)

  

nmnmSAveragenmT ,,,







 (20)

So each pixel of images has a threshold value. The

binary images obtained by the each respective threshold

values are







1,'

,' 0











ifGraymnTmn

Graym notherwise



,



(21)

Obtaining the final image only contain-

ing the vein characteristics.



', nmGray

It is obvious that the grain of the original image has

been got as Figure 4. The effect is comparatively ideal.

Median filtering method can eliminate burrs and make

the borderline smooth. In addition, because the result of

the new threshold dispose algorithm inducts massive

noises, these noises are wiped off according to the size

of them in this paper. The effect of filtering is as shown

in Figure 5.

2.6. Lmage Thinning

In this paper, we thin the vein image using the combina-

tion method of general conditional thinning and tem-

plates. Get rid of the special un-single pixel point after

the general conditional thinning.

a) The conditional thinning algorithm [22]

Region points are assumed to have value 1 and back-

ground points to have value 0. The method consists of

successive passes of two basic steps applied to the con-

tour points of the given region, where a contour point is

any pixel with value 1 and having at least one 8-

neighbor valued 0. With reference to 8-neighborhood

Figure 4. Segmentation. Figure 5. Filtering.

notation shown in Figure 6, Step 1 flags a contour point

for deletion if the following conditions are satisfied:

1

p



1

246

468

() 26

() 1

() 0

aNp

bTp

c ppp

dppp





















(22)

where





1

pN is the number of nonzero neighbors of

, that is,

1

p





98321 pppppN 





 (23)

And





1

pT is the number of 0-1 transitions in the

ordered sequence . For example,

23 892

,,,,,pp ppp





14



pN ,





13



pT in Figure 7.

In step 2, conditions (a) and (b) remain the same, but

conditions (c) and (d) are changed to

246

468

(') 0











cppp

dppp

(24)

Thus one iteration of the thinning algorithm consists

of: 1) applying step 1 to flag the remaining border points

for deletion; 2) deleting the flagged points; 3) applying

step 2 to flag the remaining border points for deletion; (4)

deleting the flagged points. This basic pro cedure is app-

9

p 2

p 3

p

8

p 1

p 4

p

7

p 6

p 5

p

Figure 6. Neighborhood Arrangement Used by the Thinning

lgorithm. A

Figure 7. Illustration of condition (a) and (b) in Eq.(22) (In

this case,





4

1



pN ,





3

1



pT ).

0 0 1

1 1

p

0

1 0 1

R. Wirestam et al. / J. Biomedical Science and Engineering 2 (2009) 261-272 267

lied iteratively until no further points are deleted, at

which time the algorithm terminates, yielding the skele-

ton of the region.

b) The improved thinning algorithm

The improved thinning algorithm is on the base of the

conditional thinning. On the conditional thinning image,

the template algorithm is added to get rid of the un-sin-

gle pixel point in this paper.

For the case of Figure 8, the un-single pixel point is

pointed by the red line. We must think of one method

to get rid of the pixels point, because they are the su-

perabundance points in the image. There are two

methods to get rid of it. One is that the image can still

be assured the connectivity after getting out of these

points. These points can be deleted according to the

connection of 4- neighborhood and 8-neighborhood.

The other uses the method of templates. The templates

are as follow Figure 9. The center of template is the

reference point.

The center of template is the reference point *. “1” is

the point of target image, “0”is the point of background

image, the points of Figure 6 which satisfy conditions

will be dispelled by the templates Me, Mf, Mg, Mh. Do

iteration based on the conditional thinning image until

the un-single pixel points of the first class are dispelled

completely.

c) Burr cutting

The noise and shadow of original finger vein image

will produce “Burr” in the skeleton image, and it will

affect the following processing. "Burr" can be dispelled

by recording the amount of the un-single pixel points

from each endpoint to the bifurcation points, and then

select a threshold. The value of the un-single pixel points

whose amount is less than threshold amounts 0, contrar-

ily, the value keeps constant. Image thinning is shown in

Figure 10.

3. EXTRACTION OF MINUTIAE POINTS

The branching points and the ending points in the vein

pattern skeleton image are the two types of critical

points to be extracted. To obtain the junction points from

the skeleton of vein patterns, we run the following

pixel-wise operation commonly known as the cross

number concept [21]: For a region (as follow

Figure 11), If is 1, and the number of transition

33

0

p

Figure 8. The un-single pixel points of the first kind.

Figure 9. The un-single pixel point templates of the first kind.

(a) General conditional thinning (b) Improved thinning algorithm (c) Burr cutting

Figure 10. Image Thinning.

268 R. Wirestam et al. / J. Biomedical Science and Engineering 2 (2009) 261-272

Figure 11. 33



region

trans

N

8

p

between 0 and 1 (and vice versa) from to

is greater than or equal to 6, then is a junction

1

p

0

p

point. Mathematically, this can be expressed with the

following equation:





 8

1

i

iitrans ppN , where

19 pp 

A similar approach can be applied to find the ending

points. The difference is that the number of transition

for an ending point is now exactly 2. The flow-

chart of the fingerprint enhancement algorithm is shown

in Figure 12 and experimental process is shown in Fig-

ure 13.

trans

N

Figure 12. A Flowchart of the proposed finger-vein processing algorithm.

(a) (b) (c) (d) (e) (f) (g) (h) (i)

(a) Input image (b) Orientation image estimation (c) Gabor filtering (d) Image segmentation (e) Filtering (f) General conditional thinning (g) Im-

proved thinning algorithm (h) Burr cutting (i) Minutiae points extraction

F

igure 13. Outputs of different stages of the algorithm.

1

p 2

p 3

p

8

p 0

p 4

p

7

p 6

p 5

p

Minutiae of the vein patte

r

n

(The bigger square for bifurcation points; the

smaller square for ending points)

Orientation image

estimation

Gabor filtering

Image segmentation

Image thinning

Minutiae points ex-

traction

Normalization

In

p

ut ima

g

e

R. Wirestam et al. / J. Biomedical Science and Engineering 2 (2009) 261-272 269

4. EXPERIMENT RESULT

Experiment on our finger-vein dataset: In this section,

three experiments conducted to demonstrate the per-

formance of the proposed method for object recognition.

Our finger-vein database comprises of 50 distinct sub-

jects, each subject having 10 different images, taken at

different times. The images are subject to variations such

as lighting. All the images are taken against a dark ho-

mogeneous background and the Finger are in upright,

frontal position (with a tolerance for some side move-

ment). The images are 256 grey levels per pixel and

normalized to pixels. These images are

pre-processed by our method before matching experi-

ment.

376328

4.1. Verification Using MHD

Many techniques can be applied to the analysis of minu-

tiae of vein pattern in a similar manner to those applied

to fingerprint minutiae. However, due to the fact that the

number of minutiae for the vein pattern is relatively

small compared to those for fingerprints, analysis based

on geometrical information is preferred to statistical

features. Since the vein pattern is represented as a set of

two-dimensional points, matching of a pair of such pat-

tern can be achieved by measuring the Hausdorff dis-

tance between the two minutiae sets.

The original Hausdorff distance was proposed for

comparing two binary images by Huttenlocher et al.[23].

The computation of Hausdorff distance does not require

point correspondences between the two point sets. They

proposed efficient computational algorithms for speed-

ing up the process of finding similar patterns in an image

by searching the regions with the smallest Hausdorff

distances in the image [6]. For two point sets





x

N

xxxxX ,,,, 321



and





y

N

yyyyY,,,, 321 , Eqs.

(25) and (26) give the definition for a Hausdorff distance,

where

H

D h and are the undirected and directed

Hausdorff distance for the two point sets, respectively.

The smaller the value of

H

D, the more similar the two

point sets are







 





XYdYXdYXHD ,,,max,



(25)





ji

Yy

Xx yxYXd

j

i

 

minmax, (26)

However, the original Hausdorff distance method is

sensitive to small perturbations in point locations for

shape alignment. To overcome the weakness of the

Hausdorff distance, the modified Hausdorff distance

(MHD) [12] for matching two objects is developed and

has the following desirable properties: 1) its value in-

creases monotonically as the amount of difference be-

tween the two sets of points increases, and 2) it is robust

to outlier points that might result from segmentation

errors. Unlike the original form, the undirected MHD is

defined as in Eq. (27)







Xx

ii

Yx

Xii

yx

N

YXdmin

1

, (27)

Figure 13 shows the false acceptance rate (FAR)

and false rejection rate (FRR) curves, from which it

can be seen that when the threshold value for the dis-

tance measure HD is 0.52, the equal error rate (EER)

is approximately 14.51%. An experiment with the

MHD was carried out on the same data set, and the

algorithm achieves 0.761% EER (as shown in Figure

14).

Genuine

%percentage

Imposter

FAR%

EER

HD distance FRR%

(a) (b)

(a) Genuine and imposter distributions and (b) Error rate for minutiae recognition using the original form of Hausdorff distance (EER =

14.51% where the threshold is observed to be 0.52)

Figure 13. Verification test results.

270 R. Wirestam et al. / J. Biomedical Science and Engineering 2 (2009) 261-272

Genuine

%percentage

FAR%

Imposter

EER

M

HD distance FRR%

(a) (b)

(a) Genuine and imposter distributions and (b)Error rate for minutiae recognition using MHD (EER = 0.761% where the threshold is observed to

be 0.43)

Figure 14. Verification test results.

4.2. Verification Using Shape of the Vein

Patterns

To verify that minutiae points can best represent the shape

of the vein patterns for vein recognition, in the following,

we show some experimental results of applying the pro-

posed robust image matching algorithm to the skeleton

representation of the vein pattern. The experiments utilize

the same MHD to measure the similarity. Figure 15 shows

FAR and FRR curves, from which it can be seen that when

the threshold value for the distance measure HD is 0.41,

the equal error rate (EER) is approximately 7%. Table 1

contains the results for the evaluation, which shows that

while the shape of the vein patterns can reach relatively

high recognition accuracy, the minutiae can further in-

crease the accuracy and require less time.

4.3. Two Types of Minutiae

To evaluate the usefulness of the two types of minutiae,

namely bifurcation and ending points, we carried out

three sets of personal verification experiments: firstly

with the bifurcation points only, and then with ending

points only, and finally the combination of the two

minutiae types. The three sets of experiments utilize

the same MHD to measure the similarity. Table 2 con-

tains the results for the evaluation, which shows that

while the bifurcation and ending points can reach rela-

tively high recognition accuracy, the combination of

the two minutiae can further increase the accuracy,

which is advantage especially when high feature dis-

criminating power is desired for a large population

group.

FAR%

%percentage

Imposter

Genuine

EER

MHD distance FRR%

(a) (b)

Figure 15. Error rate curves for minutiae recognition using MHD (EER = 7% where the threshold is observed to be 0.41).

R. Wirestam et al. / J. Biomedical Science and Engineering 2 (2009) 261-272 271

Table 1. Minutiae and shape of the vein patterns evaluation

using MHD.

Equal error

rate (%)

Threshold

value Time(s)

Shape of the

vein patterns 7 0.41 6.5

Combination of

both minutiae 0.761 0.43 2.2

Table 2. Minutiae evaluation using MHD.

Equal error

rate (%)

Threshold

value

Bifurcation points only 9.20 0.48

Ending points only 8.89 0.46

Combination of both

minutiae 0.761 0.43

5. CONCLUSIONS

This paper proposes a novel method applicable to fin-

ger-vein recognition. Firstly, we extract the features of

the vein patterns for recognition. Secondly, the minu-

tiae features are extracted from the vein patterns for

recognition, which include bifurcation points and end-

ing points. These feature points are used as a geomet-

ric representation of the shape of vein patterns. Finally,

the modified Hausdorff distance algorithm is proposed

to evaluate the discriminating between all possible

relative positions of the shape of vein patterns. Ex-

perimental results show the equal error rate (EER)

reaches 0.761% where the threshold value for the dis-

tance measure HD is observed to be 0.43. This result

indicates the minutiae features in the vein patterns can

be used as a feature sets in the personal verification

applications efficiently.

Though the current database is relatively small and it

is not adequate to draw any firm conclusion on the dis-

criminating power of vein patterns for a large population

(in terms of millions of users) group, the experiments do

show the potential of the minutiae of the vein patterns as

a biometric feature for personal verification applications

in a reasonable sized user group. The results presented

here indicate, as a new identity authentication technol-

ogy, the vein recognition has a better long term potential,

need to be studied further, and can be applied to the lives

of people better.

6. ACKNOWLEDGMENTS

This article belongs to science and technology re-

search item of Chongqing education committee

(KJ070620). The authors are also most grateful for

the constructive advice and comments from the

anonymous reviewers.

REFERENCES

[1] Z. B. Zhang, S. L. Ma, and X. Han, (2006) Multiscale

feature extraction of finger-vein patterns based on

curvelets and local interconnection structure neural net-

work, The 18th International Conference on Pattern Rec-

ognition (ICPR’06).

[2] K. Kanagawa, (2006) Finger vein authentication tech-

nology and its future junichi hashimoto, Information &

Telecommunication Systems Group, Japan IEEE.

[3] A. Jain, R. M. Bolle, and S. Pankanti, (1999) Biometrics:

Personal identification in networked society, Kluwer

Academic Publishers, Dordrecht.

[4] P. MacGregor and R. Welford, (1991) Veincheck: Imag-

ing for security and personnel identification, Adv. Imag-

ing 6 (7), 52-56.

[5] P. L. Hawkes and D. O. Clayden, (1993) Veincheck re-

search for automatic identification of people, in: Pre-

sented at the Hand and Fingerprint Seminar at NPL.

[6] L.Y. Wanga, G. Leedhamb, and D. S. Y. Choa, (2007)

Minutiae feature analysis for infrared hand vein pattern

biometrics, Pattern Recognition Society, Published by

[7] J. A. Gualtieri, J. L. Moigne, and C. V. Packer, (1992)

Distance between images, Fourth Symposium on the

Frontiers of Massively Parallel Computation, 216-223.

[8] J. You, E. Pissaloux, J. L. Hellec, and P. Bonnin, (1994)

A guided image matching approach using Hausdorff dis-

tance with interesting points detection, Proceedings of

the IEEE International Conference on Image Processing,

1, 968-972.

[9] V. D. Gesù and V. Starovoitov, (1999) Distance-based

functions for image comparison, Pattern Recognition Lett,

20 (2), 207-214.

[10] A. Ghafoor, R. N. Iqbal and S. Khan, (2003) Robust

image matching algorithm, Proceedings of the Fourth

EURASIP Conference Focused on Video/Image Proc-

essing and Multimedia Communications, 155-160.

[11] V. Perlibakas, (2004) Distance measures for PCA-based

face recognition, Pattern Recognition Lett, 25(6), 711-

724.

[12] M. P. Dubuisson and A. K. Jain, (1994) A modified

Hausdorff distance for object matching, Proceedings of

the 12th IAPR International Conference on Pattern Rec-

ognition, 1, 566-568.

[13] J. Paumard, (1997) Robust comparison of binary images,

Pattern Recognition Lett, 18(10), 1057-1063.

[14] B. Takács (1998), Comparing face images using the

modified Hausdorff distance, Pattern Recognition, 31(12),

1873-1881.

[15] B. Guo, K.-M. Lam, K.-H. Lin, and W. C. Siu, (2003)

Human face recognition based on spatially weighted

Hausdorff distance, Pattern Recognition Lett, 24(1–3),

499-507.

[16] K.-H. Lin, K.-M. Lam, and W. C. Siu, (2003) Spatially

eigen-weighted Hausdorff distances for human face rec-

ognition, Pattern Recognition, 36(8), 1827-1834.

[17] Z. Zhu, M. Tang, and H. Lu, (2004) A new robust circu-

lar Gabor based object matching by using weighted

272 R. Wirestam et al. / J. Biomedical Science and Engineering 2 (2009) 261-272

Hausdorff distance, Pattern Recognition Lett, 25(4),

515-523.

[18] L. Hong, Y.F. Wan, and A. Jain, (1998) Fingerprint image

enhancement: algorithm and performance evaluation,

IEEE Transactions on Pattern Analysis and Machine In-

telligence, 20(8).

[19] D. Marr, (1982) Vision, San Francisco, Calif.: W. H.

Freeman.

[20] J. G. Daugman, (1993) High confidence visual recogni-

tion of persons by a test of statistical independence, IEEE

Trans, Pattern Analysis and Machine Intelligence, 15(11),

1148-1161.

[21] U. Halici, L. C. Jain, and A. Erol, (1999) An introduction

to fingerprint recognition, in: L. C. Jain, U. Halici, I.

Hayashi, S. B. Lee, S. Tsutsui (Eds.), Intelligent Biomet-

ric Techniques in Fingerprint and Face Recognition, CRC

Press, Boca Raton, FL, 3-34.

[22] R. C. Gonzalez, (2005) Digital Image Processing, Pub-

lishing House of Electronics Industry, Beijing.

[23] D. P. Huttenlocher, G. A. Klanderman, and W. J. Ruck-

lidge, (1993) Comparing images using the Hausdorff dis-

tance, IEEE Trans. Pattern Anal. Mach. Intell, 15(9),

850-863.

[24] E. P. Vivek and N. Sudha, (2007) Robust Hausdorff dis-

tance measure for face recognition, Pattern Recognition,

Elsevier, 40(2), 431-442.

[25] E. P. Vivek and N. Sudha, (2006) Gray Hausdorff dis-

tance measure for comparing face images, IEEE Transac-

tions on Information, Forensics and Security, 1(3),

342-349.

[26] M. Sezgin and B. Sankur, (2004) Survey over image

thresholding techniques and quantitative performance

evaluation [J], Journal of Electronic Imaging, 11.

[27] S. Lachance, R. Bauer, and A. Warkentin, (2004) Appli-

cation of region growing method to evaluate the surface

condition of grinding wheels, International Journal of

Machine Tools and Manufacture, 44(7-8), 823-829.

[28] R. Saurel, F. Petitpas, and R. A. Berry, (2009) Simple and

efficient relaxation methods for interfaces separating

compressible fluids, cavitating flows and shocks in mul-

tiphase mixtures, Journal of Computational Physics,

228(5), 20 March, 1678-1712.

[29] C. C. Kang and W. J. Wang, (2007) A novel edge detec-

tion method based on the maximizing objective function,

Pattern Recognition, 40(2), 609-618.

[30] Z. B. Zhang, D. Y. Wu, S. L. Ma, and J. Ma, (2005) Mul-

tiscale feature extraction of finger-vein patterns based on

wavelet and local interconnection structure neural net-

works and brain,. ICNN&B’05, 2(13-15), 1081-1084.

[31] V. Barra, (2006) Robust segmentation and analysis of

DNA microarray spots using an adaptative split and

merge algorithm, Computer Methods and Programs in

Biomedicine, 81(2), 174-180.

[32] J. M. Lu, X. Yuan, and T Yahagi, (2006) A method of

face recognition based on fuzzy clustering and parallel

neural networks, Signal Processing, 86(8), 2026-2039.

Paper Menu >>

Journal Menu >>