Palm Vein Verification System based on SIFT matching

pass filtering. Then , some binarization is performed and for verification using this
image, two main streams can be fou nd: global matching of the reference and test
images through a pixel to pixel superposition [5] or after so me frequency analysis as
in Laplacian palm processing [8]. Another direction consists in performing local
matching of specific feature po ints extracted in the two references and test images [6].
The match ing algorithms in this las t approach are similar to those used for fingerprint
verification [7].
Our approach developed in this pap er falls in the last category. However, we use for
the matching of the feature points, the well known SIFT [9 ] algorithm which, to our
knowledge has never been used so far for hand vein verification. This algorithm
developed for graph matching was proven to be very efficien t for face v erificatio n
[10]. We tested the proposed system on a small home d atabase of video s o f the han d
palm of 24 persons containing two sessions.
We presen t in Section 2 the details of our app roach including a synopsis of the
developed system, the acquis ition device we used for acquiring the palm d atabase as
well as the preprocessing we made o n these images. We also explain our binarizatio n
procedure. In the following subsection, the detection of the feature points and the
SIFT procedure is briefly described. Due to the fact that some translations and
rotations are present between the two session s, we also p ropose a post pro cessin g
allowing a better separation of genuine and impostor’s scores. Finally, section 3
presents the experimental work, namely the pro tocols defin ed on each database an d
the results in terms of EER and ROC curves.
2 Description of t he proposed system
The proposed scheme o f the proposed method is shown in Figure 1 .
Figure 1: Synopsis of the developed method (verification step)
It is standard in biometrics with an enrollment and a verification step. One important
characteristic of our method is that only one imag e is n eeded for the enrollment step.
Moreover, a post-processing is add ed after the match ing phase in order to tackle
translation and rotation problems.
2.1 Acquisition
Image acquisition is done by 48 infrared LEDs and a CCD camera, which sensitive
range is between 400nm and 900nm but the in teresting wav elength is around 850nm.
We added to th is system one support in order to help the persons position ing their
hand and to therefore limit tran slations and rotations. The hand is approximately at
20cm of the camera’s lens (see Figure 2).
Figure 2: NIR acquisition device
At each acquisition, we acquired a short video wh ich provid ed u s a set of 30 grey-
scale pictures of size 640x320 pixels. We this way acquired the data of 24 persons in
two sess ions.
2.2 Pre-process ing
The first step of the pre-processing is the extraction of the region of interest (ROI).
Due to limited translation and rotation, th is is eased. Sizes are cut down to 232x280
pixels. Then, we apply the 5x5 box filter on the ROI in order to reduce the nois e.
After remov ing the high frequency no ise, we need to correct the brightn ess, which is
not uniform. A Gaussian low-pass 51x51 filter is applied on the ROI in order to
obtain the brightness image which is considered as low frequencies. Then, the
brightness is subtracted of the original ROI.
At this step, the con trast is still too bad. We therefore apply a normalization metho d
commonly used in fingerprint verification [7]. For each image I of size NxM, the
mean and variance (denoted µ and respectively) are calculated. Equation describes
the normalization process applied on the image with µ
and
, the desired value of
d
d
the mean and variance. For each pixel, we modify its gray-level with th e followin g
formula:
i
( ( , ) )? ( , )
IxyIxy
2
+>
d
d
2
Ixy
'( , )
=
(1)
i
( ( , ) )?
I x y Otherwise
2
d
d
2
Where
I(x,y)
corresponds to the gray-level for the pixel located at
(x,y)
for the original
image and
I’(x,y)
for the resultin g one after pre-processing. Figure 3 shows the
original image of the palm acquired with our sensor and the corresponding image
after pre-proces sing. For our experiments, we set emp irically the valu es of µ
to 12 8
d
and
to 40 .
d
Fig ure 3: (left) NIR image of palm, (righ t) Image after pre-p rocessing
2.3 Vein Pattern Ex traction
After n oise reduction and contrast normalization, the quality of the image is improved
as can be seen in Figure 3. To obtain th e vein pattern, it is necessary to extract the
veins from the background. In fact, the g rey lev el is lo w where the hemoglobin
absorbs the NIR. Therefore, the chosen ex traction algorithm is a local thresholdin g
one depen ding on the mean valu e of the neighborhood of each pixel.
Figure 4 shows two processing results for two images for the same individual. If we
would ask somebody to decide if these images correspond to the same indiv idual, he
would probably try to find similar areas between these two images. This is th e id ea of
the proposed methodology defined in th e n ext section .
Figure 4: Ex amples o f two bin arized images corresponding to the same indiv idual
2.4 Pattern definition
We chose to use a local description of the vein imag e in order to facilitate the
verification step. We used the SIFT descriptor, following of a comparative study [11]
showing that it is one of th e most efficient local ones.
The invarian t descrip tor developed in the SIFT algorithm described in [9] is applied
locally on key-points and is based on the image gradients in a local n eig hborhood.
The SIFT detector and descriptor are constructed from the Gaussian scale space of the
source image. Th e algorithm makes use of another scale space too, called differen ce
of Gaussian (DOG), which can be considered as the scale derivativ e of the Gau ssian
scale space. Extracted key-points are defined as points o f local extremum of the DOG
scale space.
The descriptor is created by sampling the magnitudes and orientations of the image
gradients in a neighborhood of each key-point and building smoo thed orientatio n
histograms that contain the important aspect of the neighborhood. Each local
descrip tor is composed on a 4x4 array (histogram). To each coordinate of th is array,
an 8 orientation v ector is associated. A 128-elemen ts vector is then built for each key-
point.
We used in this article the imp lementation provided by Lowe [9]. As illustration, we
obtained an average valu e of 800 detected k ey-points for the vein images at hand.
2.5 Matching similarity
Each individual is described by a set of invarian t features
Y(I) = {k
=(s
, x
, y
)},
i
i
i
i
i=1:N(I)
where
s
is the 128-elements SIFT invariant descriptor co mpu ted near
i
keypoint
k
,
(x
,y
)
its position in the original image
I
and
N(I)
the number of detected
i
i
i
keypoints for image
I
.
The verification prob lem for an individual given the set
Y(I)
correspond ing to the
biometric mod el of an individual in our case, is to measure the similarity with another
set of keypoin ts compu ted on the supposed v ein imag e of the ind ividual. We thus
have to compute a similarity b etween two sets of points
Y(I
)
and
Y(I
)
. We thus use
1
2
the following matching method which is a modified version of a decision criterio n
first proposed by Lowe [9]:
Given two points
x Y(I
)
and y
Y(I
)
, we say that
x
is associated to
y
iff :
1
2
(2)
d(x,y) = min
d(x,z) and d(x,y) C d(x,y')
{z Y (I2) }
Where
C
is an arbitrary thresho ld,
d(.,.)
denotes the Euclidean distance between the
SIFT descriptors and
y'
denotes any point of
Y(I
)
whose distance to
x
is minimal but
2
greater than
d(x,y)
:
(3)
d(x,y')=min
d(x,z)
{ z Y( I2), d (x ,z )>d (x ,y ) }
In other words,
x
is associated to
y
if
y
is the closest point from
x
in
Y(I
)
according to
2
the Euclidean distan ce between SIFT descriptors and if the second smallest value of
this distance
d(x,y')
is significantly greater than
d(x,y)
. The significance of the
necessary gap between
d(x,y)
and
d(x,y')
is en coded by the constant
C
. In the same
way , we say that
y Y(I
)
is asso ciated to
x
Y(I
)
iff
x
is the closest po int from
y
2
1
among
Y(I
)
according to the Euclidean distance b etween SIFT descriptors and if the
1
secon d smallest valu e of this distance
d(y,x')
satisfies
d(y,x)<C d(y,x')
. Then, we will
say that
x
is matched to
y
iff
x
is associated to
y
and
y
is associated to
x
. If
Y(I
)
and
1
Y(I
)
do not define any match ing point, we consider that both sets have an infinite
2
distance.
Figure 5 shows an examp le of matching results for an impostor and a genuine. We
obtained 23 detected associations for an impostor (left result) and 63 for the genuine
one. In the whole database, th e average value for impostors equals 6 (we show in
Figure 5 the comparison result hav ing the larger nu mber of associations for an
impostor) and 30 for genuine ones.
Figure 5 : Example of matching result for an impostor (o n the left) and with a genuine
(on the right)
The numb er of associations is used here as a similarity measure.
2.6 Post-processing
The post-pro cessing has for objectiv e to select among the detected asso ciations the
relevant ones. The acquisition can provide mainly small translation and rotatio n
artifacts. That means that the associatio ns for genuine images corresponding to
disp lacement vectors in Figure 5 shall b e relatively co llinear. In order to d etermine
relevant on es, we compute the direction of each disp lacement vector by considerin g
the k ey-points location in each image to match. We comp ute the histogram of
orientation of th ese vectors (with a step of 0.1 radian). The updated similarity value is
the number o f associations for the highest peak of the histogram.
Figure 6 sh ows the post-processing resu lt o f the association result given in Figure 5
(right). Note that we obtain ed a similarity value equals to 6 instead 23 and 51 instead
63 for the results p resented in Figure 3. This shows that we significantly increase the
difference between g enuine and impostor samples.
Figure 6: Post processing of the result of Figure 5 (right)
3 Experimental Results
3.1 Protocol
We realized a home database with 24 individuals. Two sessions have been made an d
videos of hand have been acq uired. We extracted from videos, 30 images for each
session. The total amount of images used for this experiment is 24x30x2 = 144 0
images.
We use one image for the enrollment and 5 9 images for the test for each indiv idual.
3.2 Performance evaluation
We used several measures to quantify the performan ce of the proposed system. In
order to comp ute th e False Rejection Rate (FRR) v alu e, we compare the biometric
reference with all samples of the same individual. Fo r the Fals e Acceptance Rate
(FAR) v alue, we compare th e biometric referen ce of an individ ual with all samp les
from different individuals.
Fo r out test database, the FRR value is therefore computed considerin g 24x59=1416
matching procedures and the FAR value uses 24x59x23= 32568 ones. We will u se the
Equal Error Rate (EER) measure that corresponds to the valu e of the biometric system
threshold for which FAR equals to FRR.
We will plot the disparities distribution between intraclass and interclass matchin g
resu lts and the ROC (Receiver Operating Ch aracteristic) curve.
3.3 Results
First, we plot th e distribu tion of the intraclass and interclass disparities on the test
datab ase using the proposed method with out and with post-processing. The benefit of
the post-processing is clear. It allows a goo d separatio n between g enuine an d
impostors.
Figure 7: Disparities distribution for the proposed method on the test datab ase
(left) without any post-processing (right) with post-processin g.
Computing the DET curve of the proposed method without any post-processing (see
Figure 8 ) we obtain an EER value equals to 0.14%. After using the pro posed post-
processing, we obtained an ideal EER value that is to say 0%.
The averag e computation time for verification with images of size 232x280 pixels an d
usin g a Matlab implementation is about 2.5 seconds.
Figure 8: DET curve of the proposed method without an y post-processing (b lue
dotted line) and with processing (red line)
4 Conclusions and perspectives
Han d vein verification is a promising techniqu e for the authentication of an indiv idual
as the expected perfo rmance is very good. The performance we can expect is very
good. This bio metric modality is also difficult to copy that makes it a good candidate
for many applications such as au thentication method for e-banking o perations.
In this paper, a comp lete biometric system based palm veins has b een developed. We
proposed an original method b ased on the u se of SIFT des criptors fo r the enrollment
and the verification steps. On e of the main benefit of the proposed approach is that
only one image is needed for the enrollment.
The efficiency of the proposed method is pro mising even if the test database used in
the experiments is quite small. In the future, we intend to define a more significant
datab ase confirming o ur first preliminary results.
References
1. S an g-K yun Im , Hy ung -M a n P a rk , You ng- Woo Kim , S a ng- C ha ng H an,
S oo- onW . K im , C hu l- He e K an g: An B iom e tr ic Id en tif ic a tion S y ste m
by E xtr ac tin g Ha nd Ve in P a tte r ns, Jou r na l of t he K or ea n P hys ic al
S oc iet y Vol . 38, No . 3, M a rc h ( 20 01) .
2. M oha m ed S hah in, Ahm e d B a da wi, Mo ha me d Ka m el : “ B iom e tr ic
Aut hen tic a tio n Using Fa st C or r e lat ion o f Ne a r I nf r ar e d Ha nd Ve in
P at te rn s” , In te rn at iona l jo ur na l of B iom e dic a l sc ie nc e s, Vo l. 2, ( 2007 ).
3. Zho ngb o Zha ng, S ili an g M a, Xi ao Ha n: M ulti sc al e F e a tur e E xtr a c tion
of F in ge r -V ein P a tte r ns B a sed on C ur ve le ts a nd Lo ca l In te r con ne c tion
S tru c tur e N eu ra l Ne two rk , P a tt er n R e cog nit ion, 18 th In te rn a tion al
C onf e re nc e on P a tte r n R e cog nit ion, V ol. 4, p p. 14 5- 148 , S e pte m be r
( 2006 ).
4. M as ak i Wa tan a be , T osh io End oh, M or ito S hi oha r a, a nd S hi ge ru
S as ak i: “P a lm v ei n a ut hen tic a tio n t ec hn olog y a nd it s app lic a tion s” .
P ro ce e ding s o f the B iom e tr ic C onso r tium C onf e r en ce , S e pte m be r 19th -
21 st, ( 2005 ).
5. S .K. Im , H. S . C hoi , S . W. Kim , “A D ir ec tio n- B as ed V asc ul ar P at te rn
Ext ra c tio n Algo rit hm fo r Ha nd V as cu la r P a tte rn V er i f ica tio n” , ETR I
J., vol . 25- 2, p p. 101 -1 08, ( 200 8) .
6. Lin gyu W ang , G r ah am Le ed ha m : “A T he rm a l Ha nd Vei n P at te rn
Ve r if ic at ion Sy ste m ”, In te rn at ion al Wor ksh op on P a tte r n R e co gni tion
f or C r im e P r ev en tion , S e cu r ity an d Su rv ei lla nc e , Vo l. 368 7, p p.58- 65 ,
( 2005 ).
7. L. H ong, Y . Wa n, a nd A. J ain , “F i nge r pr int Im a ge E nha nc em e nt:
Alg or ithm And P er f or m an ce Eva lu ati on” . I EE E T ra n sac t ions on
P at te rn Ana lys is a nd M a ch ine I nte lli ge nc e , 2 0( 8) , pp. 7 77- 789 , (1 998) .
8. J.- G. W an g, W .-Y. Ya u, A . S uwa ndy a nd E. S un g, “ P e rso n r e c ogni tion
by f using p alm p rin t and p alm ve in im ag es ba se d on ‘ Lap la ci an pa lm ’
r ep re se nt ati on,” P a tte rn R ec ogn itio n, vol . 41, pp . 153 1- 1544 , (20 08) .
9. Low e, D.: “ Dis tinc tiv e im a ge fe a tur e s fr om sc a le- in va ri an t k ey -
po ints” . In In te rn at iona l Jou rn a l o f C om put er Vi sion, 60 (2 ), pp. 91 -
11 0, (20 04) .
10 . R ose nbe r ge r, C. , B run , L.: S i mi la ri ty- B as ed M at ch ing fo r F a c e
Aut hen tic a tio n, In te rn a tion al C on fe r en c e on Pa tt er n R e c ogni tion
( IC P R ), ( 20 08) .
11 . M ikol aj cz yk, K., S c hm id , C .: A pe rf o rm a nc e ev alu at ion o f lo c al
de sc r ipto rs. In I EEE Tr an sa ct ion s on Pa tt er n A na lys is a nd M a chin e
I nte llig en c e, 27 (1 0) , pp. 16 15- 163 0, ( 2005 ) .