Pattern Recognition Letters 33 (2012) 1520–1528

Contents lists available at SciVerse ScienceDirect

Pattern Recognition Letters

journal homepage: www.elsevier.com/locate/patrec

A novel biometric system based on palm vein image

Jen-Chun Lee ⇑
Department of Electrical Engineering, Chinese Naval Academy, Kaohsiung, Taiwan

a r t i c l e i n f o a b s t r a c t

Article history: Vein pattern recognition is one of the newest biometric techniques researched today. In this paper, one of
Received 15 July 2011 the reliable and robust personal identification authentication approaches using palm vein patterns is pre-
Available online 25 April 2012 sented. We consider the palm vein as a piece of texture and apply texture-based feature extraction tech-
niques to palm vein authentication in our work. A 2-D Gabor filter provides the optimized resolution in
Communicated by A. Heyden
both the spatial and frequency domains, thus it is a basis for extracting local features in the palm vein
recognition. In order to obtain effective pattern of palm vascular, we proposed an innovative and robust
Keywords:
directional coding technique to encode the palm vein features in bit string representation. The bit string
Biometrics
Palm vein recognition
representation, called VeinCode, offers speedy template matching and enables more effective template
Gabor filter storage and retrieval. The similarity of two VeinCodes is measured by normalized hamming distance.
Feature extraction A total of 4140 palm vein images were collected form 207 persons to verify the validity of the proposed
palm vein recognition approach. High accuracy has been obtained by the proposed method and the speed
of the method is rapid enough for real-time palm vein recognition. Experimental results demonstrate that
our proposed approach is feasible and effective for palm vein recognition.
Ó 2012 Elsevier B.V. All rights reserved.

1. Introduction vein pattern-extracting sensor, there is no contamination. Both

Wherever security is required, whether to restrict access to
buildings or to computer systems, it is necessary to employ an ac-
cess control system. Traditional personal verification methods rely
heavily on the use of passwords, personal identification numbers
(PINs), magnetic cards, keys, smart cards, etc. No matter which
method is employed, each of these offers only limited security.
Many biometric recognition systems have been proposed, based
on various human physiological features or behaviors including
facial images, hand geometry, fingerprints, palm prints, retinal
images, handwritten signature, and gait (Jain et al., 1999), to
improve the security of personal verification. Each of these systems
has merits and demerits. In the case of fingerprints, direct contact
of the finger with the fingerprint-image-extracting sensor causes
degradation in performance, especially in factory construction sites
where good-quality fingerprints are hard to obtain due to oil from
the finger, moisture, dirt, etc. For retina scanner, users must place
their eye close to the scanner, causing an uncomfortable feeling
and concerns of privacy. With hand-shape recognition devices,
problems may arise with users who suffer from arthritis or
rheumatism, leading to poor performance. Compared with the
other physical characteristics, palm vein authentication has been
developed to resolve these problems. Since it acquires a palm vein
pattern image without direct contact with the palm or with the
⇑ Tel./fax: +886 75829682.
E-mail address: i923002@gmail.com
user comfort and performance are improved, and stable operation
is expected. Palm vein recognition appeared in 1990s (MacGregor
and Welford, 1991), and has been popular since 2000 because of
the advantages it offers. A vein pattern reveals the vast network
of blood vessels underneath a person’s skin. Like fingerprints,
though it has never been proven in a strict scientific sense, the
shape of the vascular patterns of different individuals are believed
to be distinctive (MacGregor and Welford, 1991; Hawkes and Clay-
den, 1993), and very stable over a long period of time. In addition,
as the blood vessels are hidden underneath the skin and are mostly
invisible to the human eye, vein patterns are much harder for
intruders to copy compared to other biometric features. The prop-
erties of uniqueness, stability and strong immunity to forgery of
the vein pattern make it a potentially good biometric which offers
secure and reliable features for personal identity verification. As a
result of the Fujitsu research (Masaki et al., 2005) using data from
140,000 palms (70,000 individuals), Fujitsu has confirmed that the
FAR (false acceptance rate) is 0.00008% and the FRR (false rejection
rate) is 0.01%, with the following condition: a person must hold the
palm over the sensor for three scans during registration, and then
only one final scan is permitted to confirm authentication.
In this paper, we present a novel palm vein recognition method
for personal identification. The palm vein is an ideal part of the
body for biometrics; it normally does not have hair which can be
an obstacle for photographing the blood vessel pattern, and it is
less susceptible to a change in skin color, unlike a finger or the dor-
sal hand. There are many different features in palm vein images,

0167-8655/$ - see front matter Ó 2012 Elsevier B.V. All rights reserved.
http://dx.doi.org/10.1016/j.patrec.2012.04.007
 J.-C. Lee / Pattern Recognition Letters 33 (2012) 1520–1528
such as the geometry, the principal line, and the delta point. Fur-
thermore, hand veins show significant textural differences and a
lot of minutiae similar to the ridges and branches of fingerprints.
These features offer stable, unique and reliable biometrics for per-
sonal identification. Many verification technologies using biomet-
ric features of hand veins have been developed over the past
decade. Lin and Fan (2004) present personal verification results
using palm dorsal images acquired from a thermal infrared (IR)
camera operating in the 3.4–5.0 lm range. Their approach is based
on the combination of multi-resolution images obtained from the
pre-processed thermal vein images. Wang et al. (2007) present an-
other approach for personal authentication using hand vein images
acquired from the thermal imaging. The authors (Wang et al., 2007,
2008a) have employed Hausdorff distance to generate matching
scores between the extracted line patterns and illustrated promis-
ing results. Crisan et al. (2010) focused on improving the two
essential parts of a vein-scanning device: the hardware lighting
system and the feature extraction algorithms. Hsu et al. (2011)
has been successfully used the modified 2-directional 2-dimen-
sional principal component analysis ((2D)2PCA) to obtain eigenve-
ins which is a low dimensional representation of dorsal hand vein
patterns. However, there are rarely published papers about the
palm vein recognition. Most of the papers are white papers
published by Fujitsu Company (Sasaki et al., 2005, and Fujitsu
Laboratories Ltd, 2003). Fujitsu’s palm vein verification product
has high accuracy, but to the best of our knowledge, the features
used are not disclosed in any published research articles.
In the field of the palm vein recognition, some papers with differ-
ent technologies are described below. Wang et al. (2008b) proposed
a multimodal personal identification system where palmprint and
palm vein modalities were combined in a single image. Locality Pre-
serving Projection (LPP) was used to extract features of the fused
images, which they called ‘‘Laplacianpalm’’. Zhou and Kumar
(2011) employed the Hessian phase approach to feature extraction
and presented a matching approach that can effectively accommo-
date the potential image deformations, translational and rotational
variations by matching to the neighborhood of the corresponding
regions and generate more reliable matching scores. Zhang et al.
(2007) used a low cost CCD camera to capture the infrared palm
images and extracted the palm vein features by multi-scale filtering.
The experimental results demonstrate that the recognition rate of
their system is fine but not good enough to be a real system. Michael
et al. (2010) presented an innovative contactless palm print and
palm vein recognition system. The proposed system offers several
advantages like low-cost, accuracy, flexibility, and user-friendli-
ness. Although these authors claim the features they use to recog-
nize the vein patterns can attain a high level of accuracy, most of
them used small datasets to evaluate the performance of hand vein
recognition, so the practical feasibility of these algorithms has not
been demonstrated.
Up to now, there is currently no publicly available palm vein
pattern database provided to the research community. To demon-
strate the robustness of the proposed approach, we constructed
our own near-infrared (NIR) palm vein pattern image database,
which has 207 distinct participants and contains 4140 palm vein
pattern images. It is based on a contactless, non-invasive data
acquisition method and requires no injection of any agents into
the blood vessels. Therefore, it is by far the best-known non-inva-
sive for acquiring palm vein pattern images.
Gabor transform is first brought forward by Dennis Gabor in
1946. Daugman (1980) explains how 2-D Gabor filter can attain
the combined optimized resolution in space and frequency.
Namely, it can adjust the direction and frequency band width
when doing space localization. This characteristic is very useful
for analyzing the texture of images. The method has been success-
fully used in texture segmentation and classification (Dunn et al.,
1521
1994), fingerprint recognition (Lee and Wang, 1999), face recogni-
tion (Liu, 2004), palmprint recognition (Zhang et al., 2003), iris rec-
ognition (Daugman, 1993) and so on. Because the palm vein image
consists of many lines, it has stable and obvious direction which
makes it suitable to be viewed as a texture image. In this work,
the 2-D Gabor filter based palm vein recognition method is de-
scribed in detail. Experimental results on our own NIR palm vein
image database, including more than 200 classes, have demon-
strated that the proposed method is highly feasible and effective
for personal identification. The palm vein identification system
based on 2-D Gabor filter consists of two parts: a palm vein collec-
tion device for palm vein image acquisition and an algorithm for
real-time palm vein identification. There are three main issues to
be considered in the system:
(1) The palm vein images are captured by a NIR camera as the
input data. Then, a median filter is employed on the palm
vein images to remove some noises and an Otsu’s 1979
method is applied to select a suitable threshold to segment
the palm region. Finally, two baseline data points are found
automatically and a rectangular region of interest (ROI) is
obtained.
(2) 2-D Gabor filter was used on the palm vein pattern to extract
features. To obtain useful representation of the palm vein
modalities, we propose a new scheme named directional
coding method to encode the palm vein features in bit string
representation.
(3) The Hamming distance is employed to measure the similar-
ity of two VeinCodes and verify whether the template and
testing samples are captured from the same person or not.
The rest of this paper is organized as follows. We will briefly
introduce the palm vein image collection device and process in
Section 2. A detailed description of the proposed method for palm
vein recognition is given in Section 3. Experimental results are pre-
sented and discussed in Section 4, prior to conclusions in Section 5.
2. Palm vein images collection
In visible light, the vein structure of the palm is not always eas-
ily discernible. The visibility of the vein structure varies signifi-
cantly depending on such factors as age, level of subcutaneous
fat, ambient temperature and humidity and physical activity. In
addition a multitude of other factors including scars, pigmentation
and palm position can also effect upon the image acquisition. Up to
now, there is no publicly available palm vein pattern database
available to the research community. In order to perform a preli-
minary analysis on the features of the palm vein pattern, we con-
structed our own NIR palm vein image database. In this work, a
low-cost NIR charge-coupled device (CCD) camera that we used
to acquire palm vein images was a digital noise reduction (DNR)
DSP Camera. It possesses the benefits of high availability, uniform
and consistent good image quality, convenience and low cost.
Though principally designed for use in visible light, CCD cameras
are also sensitive to NIR wavelengths of the electromagnetic spec-
trum up to about 1100 nm. This is within the actinic IR range,
which covers the NIR spectrum from 700–1400 nm. This wave-
length also lies in the medical spectral window (700–900 nm) in
which the illumination penetrates deeper into the biological tis-
sues (Godik and Guljaev, 1991). In our work, the near IR illumina-
tion (LEDs) is evenly and circularly located around the camera,
producing a peak at 750 nm wavelength. In addition to the LEDs,
another important part of this system is an infrared filter. To elim-
inate the effect of visible light, an optical infrared filter was
mounted in front of the camera’s lens. For our data collection, a
1522 J.-C. Lee / Pattern Recognition Letters 33 (2012) 1520–1528
Hoya RM80 IR filter was used. Since there are no docking devices to
constrain the participant’s hand, it can be used for either the right
or left hand.
Using the NIR light source and the IR filter, the image on the
CCD sensor is a high quality image of palm vein. The captured palm
vein images are 8-bit gray images with a resolution of 320  240.
The reason for using such a low spatial resolution is that both
the data amount in palm vein images and noise sensitivity can
be reduced while the features of palm veins can still be preserved.
During the collection of images for the database, we restricted the
position where the participants placed their palms above the cam-
era. The distance between the hand and the camera was adjusted
to approximately 20 cm for the best acquisition. At this distance,
there is acceptable tolerance for positioning the palm within the
specified region. In our palm vein database, 207 volunteers were
enrolled in our study and palm vein images were captured from
their left hands at two different times after an interval of at least
one month. Ten palm vein images were acquired each time for
each person. The database constructed in this way includes possi-
ble variations of palm vein images under various conditions. It also
includes several racial groups, including Chinese, African and Cau-
casian. The age range of the volunteers was between 18 and
60 years, and their occupations ranged from university students,
professors and technicians to manual workers such as cleaners
and electricians. No distinction was made between male and fe-
male when the samples were collected. When seen on the com-
puter monitor, the mostly distinguishable component in the
database image is the superficial vein tree pattern. In our study,
the quality of the palm vein images does not change significantly
in an air-conditioned or outdoor environment.
3. Proposed method for palm vein recognition
In this paper, a palm vein recognition algorithm is presented,
which includes three basic processes: preprocessing, feature
extraction and matching. Detailed descriptions of these three steps
are introduced in the following sections.
3.1. Preprocessing and extraction of the ROI
Image segmentation is one of the most important steps in the
analysis of processed image data. Its main goal is to separate com-
ponents of the object image that have a strong relevance to the
analysis from the background. When a palm image is obtained,
the palm background is first segmented from the image. Binariza-
tion is used to segmenting the image into two levels; object (palm
region) and background. For palm image segmentation, Otsu’s
thresholding is applied to the palm image to estimate the palm
region. By comparing the segmented palm shapes with the of ori-
ginal palm images, we can see that they are almost indistinguish-
able as shown in Fig. 1(a) and (b). This shows that Otsu’s method is
quite effective in determining the threshold of palm images.
To increase verification accuracy and reliability, the features of
vein patterns extracted from the same region in different palm vein
images are compared for verification. The region to be extracted is
known as the ROI. For this reason, it is important to fix the ROI in
the same position in different palm vein images to ensure the sta-
bility of the principal extracted vein features. However, it is diffi-
cult to fix the ROI at the same position in different palm vein
images without the use of a docking device to constrain the palm
position when acquiring palm vein images. In this paper, we have
employed the two data points (p1 and p2, as shown in Fig. 2(d)), to
determine the approximate (not absolute) immovable ROI. The fol-
lowing processes are performed to determine the two data points
in binary palm vein images. First, the inner border tracing algo-
rithm (Sonka et al., 1999) is employed to find the palm border.
Fig. 1(c) shown an example illustrating how the extracted border
of a palm image which perfectly matches the original palm con-
tour. Then for each point on the palm contour, the distance be-
tween this point and the mid-point of the wrist is calculated.
Fig. 1f shows the distance distribution diagram. As can be seen,
there are five local maximums and four local minimums. The pat-
tern in the diagram is quite similar to the geometric shape of a
palm (see Fig. 1(a)), which also has five tips (local maximums)
and four finger-webs (local minimums). Experimenting on a wide
variety of palm vein images, we found that the four local minimum
locations in the distance distribution diagram are the same as fin-
ger-web locations and match between the two locations is very
close. Finally, two valley points, p1 (the valley point between the
small finger and ring finger) and p2 (the valley point between the
middle finger and the index finger), are selected as two key data
points, as shown in Fig. 1(d). These two data points (p1 and p2;)
are employed to locate the ROI. The procedure is described as
follows.
First, the straight line p1 p2 is formed by the points of p1 and p2
as shown in Fig. 1d. To eliminate the influences of palm rotation
and define the coordinates of ROI more conveniently, the palm im-
age is rotated by the angle h (Eq. (1)) between line p1 p2 and the
horizontal line.
h ¼ tan1 ðY p2  Y p1 Þ=ðX p2  X p1 Þ; ð1Þ
where ðX p1 ; Y p1 Þ is the coordinate of p1 and (X p2 ; Y p2 ) is the coordi-
nate of p2.
This makes the direction of line p1 p2 is horizontal, as shown in
Fig. 1(e). A rectangular region Rc1 c2 c3 c4 whose corners are c1, c2, c3
and c4 is located and denoted as the ROI. The length of line c1 c2
and linepffiffiffip1 p2 are equal. The Line c1 c2 and c1 c3 are vertical and
c1 c3 ¼ 2  c1 c2 . Finally, we redefine the original coordinate
(0, 0) at the upper-left point c1 of the ROI. The reason for extracting
the ROI in this manner is that this ensures all the ROIs reference
the same region in the palm vein image. In this way, we can over-
come the problem of not using docking devices while acquiring
palm vein images. High verification accuracy can simultaneously
be maintained.
3.2. Palm vein feature extraction by texture analysis
This section defines our palm vein feature extraction method,
which includes filtering and matching. The motivation for using a
Gabor filter in our palm vein research is first discussed.
3.2.1. Overview of Gabor functions
Marcelja (1980) and Daugman (1980, 1985) modeled the re-
sponses of the visual cortex by Gabor functions because they are
similar to the receptive field profiles in the mammalian cortical
simple cells. Gabor filters are extremely useful for texture analysis
because of the 2-D spectral specificity of texture as well as its var-
iation with 2-D spatial position. Daugman (1980, 1985) developed
the 2D Gabor functions (a series of local spatial bandpass filters),
which have good spatial localization, orientation selectivity, and
frequency selectivity. In addition to accurate 2D space and 2D spa-
tial frequency location, they also provide robustness against vary-
ing brightness and contrast of images (Zhang et al., 2003). 2-D
Gabor filter has the following general form:
( !)
2 2
1 xh k y h k
hðx; y; hk ; f ; rx ; ry Þ ¼ exp  þ 2  expði2pfxhk Þ; ð2Þ
2 rx ry
2
In Eq. (2), xhk ¼ x cos hk þ y sin hk and yhk ¼ x sin hk þ y cos hk , f is the
frequency of the sinusoidal plane wave, hk is the orientation of
the Gabor filter, and rx and ry are the standard deviations of
 J.-C. Lee / Pattern Recognition Letters 33 (2012) 1520–1528 1523

(a) (b) (c)

(c) (d) (e)

Fig. 1. The location of ROI is defined inside a palm vein image. (a) The original palm vein image. (b) The palm region is segmented by the Otsu’s method. (c) The palm border
(white pixels) is extracted by the inner border tracing algorithm. (d) Two data points p1 and p2 are selected. The palm image is rotated an angle h which is between line p1 p2
and the horizontal line. (e) According to p1 p2 , the rectangular region is located and denoted as the ROI. (f) The distance distribution diagram of the palm border.

Fig. 2. Example of directional code being applied on palm vein image. (a) is the original palm vein image, (b) to (e) show the imaginary part of the palm vein in the four
directions (h = 0, 45, 90, 135), and (f) is the result of taking the maximum gradient values obtained from the four responses.
1524
the Gussian envelope along the x -axis and y -axis, respectively.
To analysis the Gabor filter in terms of the real part and imag-
inary part, we express Eq. (2) in the complex form h = hRe + ihIm
where
( !)
2 2
1 xhk yhk
hRe ðx; y; hk ; f ; rx ; ry Þ ¼ exp  þ 2
2 rx ry
2
( !)
2 2
1 xhk yhk
hIm ðx; y; hk ; f ; rx ; ry Þ ¼ exp  þ 2
2 rx ry
2 are chosen to extract direction information. The direction of point
Such Gabor filters are widely used as tunable filters for extract-
ing the orientation or edge information from images. Lee (2003)
gave a good introduction to image representation by using Ga-
bor functions. To obtain high efficiency and better resolution
in both space and frequency domains, Daugman (1993) applied
the 2-D Gabor filter on iris recognition. Based on these proper-
ties, in this paper, we try to apply a Gabor filter to palm vein
authentication.
3.2.2. Filtering and feature extraction
Palm vein refers to the vascular pattern or blood vein patterns
recorded from underneath the human skin. Every person has un-
ique structure and position of veins, and this does not change sig-
nificantly from the age of ten (Biometric Technology Today,
2004). Because the palm veins contains vascular network which
also resembles line-like characteristic, it has stable and obvious
direction which makes it suitable to be viewed as a texture im-
age. Even though the line-based approach plays an important role
in palm vein authentication, it is difficult to extract them accu-
rately in practice. 2-D Gabor filter is widely used in extracting
the features of textured images. This motivates us to apply 2-D
Gabor filter to palm vein authentication. Considering that palm
vein also has obvious direction and local features, our palm vein
identification method uses 2-D Gabor filters with four different
orientations to filter the original ROI image. Then the direction
texture features are extracted. The original palm vein feature
information is to a large extent preserved. The proposed direc-
tional coding technique aims to encode the line pattern based
on the proximal orientation of the lines.
Since most local palm vein structures come with a well-de-
fined local frequency and orientation, we set f as the reciprocal
of the average inter-vein distance and m as the number of ori-
entations for calculating hk = p(k  1)/m, k = 1, . . ., m. After decid-
ing the parameters of the 2-D Gabor filters, the real and
imaginary parts at the sampling point (X, Y) can be defined as
follow:
PminðN;NþuÞ PminðM;Mþv Þ
i¼maxð1;1þuÞ j¼maxð1;1þv Þ ðP R ði
HDmin ¼ min
juj<U;jv j<V
w=21
X w=21
X where PR(QR) and PI(QI) are the real part and the imaginary part of P(Q),
g Re ðX; Y; hk ; f ; rx ; ry Þ ¼ IðX þ x; Y þ yÞ
x¼w=2 y¼w=2
 hRe ðx; y; hk ; f ; rx ; ry Þ
J.-C. Lee / Pattern Recognition Letters 33 (2012) 1520–1528
w=21
X w=21
X
g Im ðX; Y; hk ; f ; rx ; ry Þ ¼ IðX þ x; Y þ yÞ
x¼w=2 y¼w=2
 hIm ðx; y; hk ; f ; rx ; ry Þ ð6Þ
where I(,) denotes a w  w 256 gray-level image.
 cosð2pfxhk Þ; ð3Þ In the proposed method, we proposed a new scheme named
directional coding method to extraction the maximum gradient
values of the palm vein from the different orientation. Four filters
which have same frequency and variance but different directions
 sinð2pfxhk Þ; ð4Þ
(x, y) is defined as:
dRe ðx; yÞ ¼ arg max fg Re ðX; Y; hk ; f ; rx ; ry Þg ð7Þ
k¼1;2;3;4
dIm ðx; yÞ ¼ arg max fg Im ðX; Y; hk ; f ; rx ; ry Þg ð8Þ
k¼1;2;3;4
Fig. 2(b)–(e) shows the imaginary part of the palm vein in the four
directions (h = 0, 45, 90, 135).Fig. 2(f) is the result of taking the max-
imum gradient values obtained from the four responses. This image
depicts the strongest directional response of the palm vein and it
closely resembles the original palm vein pattern shown in Fig. 2a.
Finally, the each pixel in the maximum gradient image is coded to
two bits, called VeinCode (Vr, Vi), by the following inequalities.
V r ðx; yÞ ¼ 1 if dRe ðx; yÞ P 0; ð9Þ
V r ðx; yÞ ¼ 0 if dRe ðx; yÞ < 0; ð10Þ
V i ðx; yÞ ¼ 1 if dIm ðx; yÞ P 0; ð11Þ
V i ðx; yÞ ¼ 0 if dIm ðx; yÞ < 0; ð12Þ
Using this directional coding method, the direction information in
palm vein images is stored in the feature vector. Fig. 3d and e show
the real part feature of two palm veins belonging to the same per-
son with m = 4 (h = 0, 45, 90, 135). Fig. 3(d) and (f) show the real part
feature of two palm veins belonging to the different persons. From
Fig. 3, we find that the same persons real Gabor features are similar
and those of different persons are not. This reveals that the feature
based on 2-D Gabor filter can be used as the palm vein feature.
3.2.3. Palm vein matching
In order to describe clearly the matching process, each feature
vector is considered as two 2-D feature matrices, real and imagi-
nary. A normalized hamming distance (Daugman, 1993) is adopted
to determine the similarity measurement for VeinCode matching.
Let P and Q be two palm vein feature matrices. The normalized
hamming distance can be defined as
þ u; j þ v Þ  Q R ði; jÞ þ PI ði þ u; j þ v Þ  Q I ði; jÞÞ
; ð13Þ
2HðuÞHðv Þ
respectively; the Boolean operator, ‘‘’’, denotes Exclusive-OR, the size
of the feature matrices is N  M, the U = 3, and V = 3 determine the max-
ð5Þ imum allowable horizontal and vertical translation, H(u) = min (N,-
N + u) - max (1,1 + u), and H(v) = min (M,M + v) - max (1,1 + v). It is
 J.-C. Lee / Pattern Recognition Letters 33 (2012) 1520–1528
Fig. 3. (a)–(c) show the three original palm vein images. (d)–(f) show the real part Gabor features of the (a)–(c) palm vein images.
noted that HDmin is between 1 and 0. The hamming distance for perfect
matching is zero. Although we employed two data points to find the ROI
automatically in the preprocessing, there are still a few translations and
rotations during the location of ROI. In order to provide translation
invariance and rotation matching, the minimum HD value obtained
from the Eq. (13) which is considered to be the final matching score.
4. Experimental results
At present, only small data sets are used to evaluate recognition
performance for palm vein recognition by most algorithms. Only
the approaches proposed by Zhou and Kumar (2011) and Chen
et al. (2009) have been tested on large image sets involving more
than 200 subjects. To demonstrate the robustness of the proposed
approach, in our experiments, the total number of palm vein
images was 4140, which were collected from 207 persons each
with 20 palm vein images captured at two different times. The size
of each palm vein image was 320  240 resolution and 256 gray-
levels. The experiments are conducted in two modes: identification
(i.e., one-to-many matching) and verification (i.e., one-to-one
matching). In identification mode, the algorithm is measured by
correct recognition rate (CRR), the ratio of the number of samples
being correctly classified to the total number of test samples. In
verification mode, the receiver operating characteristic (ROC)
curve is used to report the performance of the proposed method.
Four kinds of experiment schemes were designed as follow: four
(six, eight or ten) images of each person were randomly selected
for training, and the other ten samples were used as the testing
images to verify the validity of the proposed approach. In this sec-
tion, two experiments conducted to demonstrate the performance
of the proposed method for palm vein recognition. The first test is
1525
to analyze the relationship between the recognition accuracy and
the different resolutions of ROI images. The second one is to iden-
tify the number of orientations hk selected for the proposed meth-
od and to further evaluate the real and imaginary Gabor features
for palm vein recognition, respectively. The following subsections
detail the experiments and results.
4.1. Performance evaluation of the proposed method
During the experiments, each ROI image is decomposed into
different levels (120  169, 60  84, 30  42 and 15  21 resolu-
tions) and the images at each level are been tested. Therefore, there
are four levels, the image with 120  169 resolution is the 1st level
and 15  21 resolution is the 4th level. The well-known statistical
pair, the false rejection rate (FRR) and the false acceptance rate
(FAR), is usually adapted to analysis evaluate the performance of
the experimental results. Figs 4(a)–(d) illustrate the FAR and FRR
distributions. It can be seen from those figures that there exists
two curves in the distributions at different resolution. One curve
corresponds to genuine matching and the other corresponds to
impostor matching. When the decision threshold is set as the
one corresponding to the intersection of genuine and impostor dis-
tribution curves, the total error reaches the minimum, and the cor-
responding threshold. According to the Fig. 4, equal error rates
(EERs) of the 3rd levels are much less than those of other levels.
In other words, the palm vein image with 30  42 resolution is
more suitable for feature extraction than the other resolutions. In
Table 1, it is evident that the image resolution can play an impor-
tant role in the matching process. Therefore, the image with higher
or lower resolution is not suitable for the method proposed in this
paper, and they can not offer higher recognition rate. High resolu-
1526 J.-C. Lee / Pattern Recognition Letters 33 (2012) 1520–1528

1st Level 2nd Level

(a) 0.2 (b) 0.2
FRR FRR
FAR FAR
0.15 0.15

Density

Density
0.1 0.1

0.05 0.05

0 0
0.3 0.4 0.5 0.6 0.7 0.8 0.9 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Matching Score Matching Score

3rd Level 4th Level

(c) 0.2
FRR
(d) 0.2
FRR
FAR FAR
0.15 0.15
Density

Density
0.1 0.1

0.05 0.05

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0 0.1 0.2 0.3 0.4 0.5 0.6
Matching Score Matching Score

Fig. 4. FAR and FRR distributions at each level: (a) 1st level, (b) 2nd level, (c) 3rd level, and (d) 4th level.

Table 1 ble 2. In addition, considering the information contains both real

The testing results of the each level with different training samples. and imaginary parts, we individually employed information from
Training samples Level each part for the recognition performance analysis. From Table 2,
1st (%) 2nd (%) 3rd (%) 4th (%)
we find that the accuracy for m = 8 is only slight better than that
for m = 4. This means that an adjusted Gabor feature, combined
4 94.278 95.640 95.275 93.128
6 96.073 97.136 96.954 95.942
real with imaginary, with four orientations is sufficient and would
8 97.186 98.429 98.822 97.971 lead to an efficient small-scale palm vein recognition system. Be-
10 97.840 98.749 99.184 97.191 sides, only the adjusted 2-D Gabor feature can successfully applied
Percentage values show recognition rate.
to palm vein recognition. Either using the real part Gabor features
or the imaginary part features alone would make the lower recog-
nition rate for successful palm vein recognition.

Table 2
The recognition rate of proposed methods obtained by tested 2-D Gabor with 4.2. Comparison and discussions
different orientations.

Methods Number of different orientations
m = 4 (%)
Only Real 94.64
Only Imaginary 91.47
Real + Imaginary 99.18
tion images are mixed with noise and low resolution images lose
too much information. The relationship between the verification
result and the different number of training samples are also shown
as Table 1. It reveals that as the number of training samples per
class is increased and the relative gain among those resolutions be-
comes more apparent.
In this paper, we proposed the directional coding method to ex-
tract maximum gradient values of the palm vein by combining the
real part with imaginary part Gabor features. However, there are
more than four orientations to select in the 2-D Gabor filter. There-
fore, in second test, we tested 2-D Gabor with different orienta-
tions, and then applied the proposed method to extract the
maximum gradient palm vein image. The result of recognition
rates with no rejection option for m = 4 and m = 8 are shown in Ta-
Among existing methods for palm vein recognition, those pro-
posed by Wang et al. (2008a), Wang et al. (2008b), and, Zhou and
m = 8 (%)
Kumar (2011), respectively, are the best-known. Moreover, they
95.67
characterize local details of the palm vein from different viewpoints,
93.15
99.35 i.e. the geometric-based approach, appearance-based approach, and
phase based representation method. To further prove the effective-
ness of the proposed method, we make detailed comparison be-
tween the proposed method and the above three methods (and
our previous work (Hsu et al., 2011)) on our own palm vein data-
base. We compare our experimental results of each algorithm with
its published results and find that they are consistent. This verifies
to a large extent the correctness of our implementations of other
methods. Table 3 and Fig. 5 detail the experimental results.
Table 3 illustrates that the CRR of each algorithm is greater than
90%. This implies the high accuracyof these methods. Fig. 5 illus-
trates the ROC curves on our palm vein database for the five algo-
rithms compared, namely, laplacianpalm, minutiae feature points,
Hessian phase, eigenvein and the proposed method. To better dis-
tinguish these well performing methods, the horizontal axis of
Fig. 5 is spread out using the logarithmic scale. From the results
shown in Table 3 and Fig. 5, we can find that proposed method
 J.-C. Lee / Pattern Recognition Letters 33 (2012) 1520–1528
Table 3
Performance comparison of several well-known methods on our palm vein database.
Methods Correct recognition Equal error
rate (%) rate (%)
Minutiae feature points (Wang 97.68 1.84
et al., 2008a)
Laplacianpalm (Wang et al., 2008b) 96.38 2.74
Hessian phase (Zhou and Kumar, 98.71 0.91
2011)
Eigenvein (Hsu et al., 2011) 98.33 1.12
Proposed 99.18 0.44
has the better performance, followed by the Zhou and Kumar
(2011) method and the previous work by Hsu et al. 2011, and
the methods described in Wang et al. (2008a, 2008b), respectively.
The experimental results demonstrate that our current method
is better than the geometric-based, appearance-based, and phase
based representation methods, respectively. Generally, the geomet-
ric-based approach (Wang et al., 2008a) extract local features such
as the locations and local statistics of the principal veins, minutiae
points, and ridge bifurcations. However, geometric features are dif-
ficult to be extracted, represented, and compared while the discrim-
inability of geometric features such as texture energy is not strong
enough for palm vein recognition. The Gabor-based method can
effectively analyze local variations of the intensity signals, which
can reflect most of the random shape information of the palm vein
and thus achieves much higher performance. The methods of our
previous work (Hsu et al., 2011) and Wang et al. (2008b) employed
the appearance-based method to represent of palm vein, such as
eigenvein (where features are extracted by the (2D)2PCA, and lapla-
cianpalm (where features are extracted by the LPP). While the
eigenvein method aims to preserve the global structure of the image
space, and the laplacianpalm method aims to preserve the preserve
the local structure of the image space, our proposed method aims to
preserve the global and local structure of the image space. That is,
their methods did not effectively exploit global and local features
of the palm vein which play an important role in recognition. In
addition, the observed performance from the Laplacianpalm meth-
od was not high performance on our palm vein database, and this
may due to the fact that it does not focus on extracting and matching
the local textured information, and lack strong mechanism has
strong mechanism to accommodate potentially large variations
which are more likely to exist in the contactless database. Unlike
the two appearance-based methods, the proposed method com-
bines the global and local structure of the image space to extract
the palm vein features and is thus expected to achieve better results.
The Hessian phase approach (Zhou and Kumar, 2011), extracts
palm-vein features by analyzing the eigenvalues of the local image
instead of filtering the image by pre-defined filters in different ori-
entations, also achieves reasonably superior performance, and at
the same time provides smaller template size as compared to other
methods. However, the Hessian phase based approach does not out-
perform the proposed method. Due to the fact that when the palm
veins in the image are not very clear, it is difficult to extract local fea-
tures with the Hessian phase representation. From Table 3 and
Fig. 5, we can see that the proposed method obtain better results
than other methods. It is because our method well characterizes
random shape features of the palm vein. We used directional coding
technique to extraction the maximum gradient values of the palm
vein image from the different orientation. It depicts the strongest
directional response of the palm vein and has a very close resem-
blance to the original palm vein pattern. The real and imaginary part
of Gabor features used in proposed method is in essence global
shape features of the palm vein. The proposed palm vein represen-
tation also reflects local shape information of the palm vein by
1527
Laplacian Palm
0.07
Minutiae Feature
Hessian Phase
0.06 Eigenvein
Proposed
0.05
False Reject Rate
0.04
0.03
0.02
0.01
0 -3 -2 -1 0
10 10 10 10
False Accept Rate
Fig. 5. Receiver operating characteristic from our palm vein database with different
approaches.
analyzing local variations of the intensity signals. This makes our
method slightly better than others.
In addition to validating the accuracy of the proposed method,
there is also a need to evaluate the speed of the system. The exper-
iments are running on the computing environment of 3.2 GHz PC
with 2 GB RAM using Matlab 7.0. The execution time for image pre-
processing, feature extraction (from a preprocessed image to a fea-
ture vector) and matching (from a pair of feature vectors to the
match result) are 513 ms, 46 ms, and 3.5 ms, respectively. Note
that the reported speed was the average time taken to process an
input/test image for a subject. The total execution time is less than
0.6 s, which is fast enough for real-time verification. In fact, the
program code is not yet optimized, and it is possible to further re-
duce computation time. The speed of the proposed system was fast
enough for potential practical implementation in a multimodal
biometric application.
Failure verifications occur in some palm vein images. They can
be categorized into two main groups by carefully observing these
failure palm vein images.
(1) In the preprocessing, we employed two data points to find
the ROI automatically. Although the palm vein images are
almost located at the same region, there is still a few unsta-
ble preprocessing, such as translation or rotation. These
translations and rotations result from both the misplace-
ment of palms during palm vein acquisition and the imper-
fect preprocessing. The matching algorithm given in Eq. (13)
increases the resistance of the developed algorithm against
rotation and translation and therefore results in a more
robust algorithm. It is worth noting that the proposed algo-
rithm is still sensitive to big amount of rotations and
translations.
(2) The thickness of the vein will be affected due to the seasonal
change, some principal vein features may be lost inside the
ROI region as a result. Moreover, the boundary of the regions
could form pseudo principal vein features. These issues
might lead to match difficulties, but only in rare and extreme
cases.
The conditions (1) and (2) mentioned above can not avoided
and is difficult to distinguish the palm vein images by applying
our approach. This problem can be alleviated by utilizing other
methods to extract vein features, or combining with other biomet-
ric verification methods.
1528 J.-C. Lee / Pattern Recognition Letters 33 (2012) 1520–1528
5. Conclusions
We have developed a reliable and robust palm vein identifica-
tion system for real-time personal identification by applying a
low-cost NIR CCD camera-based palm vein device to capture the
palm vein images. A preprocessing algorithm extracts a rectangle
area (ROI) from a palm vein image for feature extraction. To repre-
sent a low-resolution palm vein image and match different palm
vein images, we extend the use of 2-D Gabor filter to represent a
palm vein image using its texture feature, and apply a normalized
hamming distance for the matching measurement. In addition, we
proposed a new technique called directional coding to code the
palm vein features in two bits representation. This method repre-
sents the biometric features in bit string format which enable spee-
dy matching and convenient storage. Using this representation, the
total size of a palm vein feature is reduced to 2520 bits. In our palm
vein database of 4,140 palm vein images from 207 different palms,
we can achieve high recognition rate (> 99%), and its equal error
rate is 0.4%, which is comparable with all other hand-based bio-
metrics, such as hand geometry and fingerprint verification.
In summary, we conclude that our palm vein identification sys-
tem can achieve good performance in terms of speed and accuracy.
For further improvement of the system, we will focus on three is-
sues: (1) to reduce the size of the device for practical applications,
(2) to design a hand attachment frame for constraining the partic-
ipant’s palm for stable preprocessing, and (3) to combine the pro-
posed palm vein coding scheme with other texture feature
measurements such as texture energy for coarse level classification
to achieve higher performance.
6. Uncited references
(Gabor, 1946; Otsu, 1979; Vein recognition in Europe, 2004).
References
Chen, H., Lu, G., Wang, R., 2009. A new palm vein matching method based on ICP
algorithm. In: Proceedings of the International Conference on Interaction
Sciences pp.1207–1211.
Crisan, S., Tarnovan, I.G., Crisan, T.E., 2010. Radiation optimization and image
processing algorithms in the identification of hand vein patterns. Comput.
Stand. Inter. 32 (3), 130–140.
Daugman, J.G., 1980. Two-dimensional spectral analysis of cortical receptive field
profiles. Vision. Res. 20, 847–856.
Daugman, J.G., 1993. High confidence visual recognition of persons by a test of
statistical independence. IEEE Trans. Pattern Anal. 15 (11), 1148–1161.
Daugman, J.G., 1985. Uncertainty relation for resolution in space, spatial frequency
and orientation optimized by two-dimensional visual cortical filters. J. Optical
Soc. Am. 2 (7), 1160–1169.
Dunn, D., Higgins, W.E., Wakeley, J., 1994. Texture segmentation using 2D gabor
elementary functions. IEEE Trans. Pattern Anal. Mach Intell. 16 (2), 130–149.
Fujitsu Laboratories Ltd: Fujitsu Laboratories develops technology for world’s first
contactless palm vein pattern biometric authentication system, [Online March
2003], available at: <http://pr.fujitsu.com/en/news/2003/03/31.html>.
Gabor, D., 1946. Theory of communication. J. Inst. Electr. Eng. 93, 429–457.
Godik, E.F., Guljaev, Y.V., 1991. Functional imaging of the human body. IEEE Trans.
Eng. Med. Biol. 10, 21–29.
Hawkes, P.L., Clayden, D.O., 1993. Veincheck research for automatic identification of
people. Presented at the Hand and Fingerprint Seminar at NPL. pp. 230–236.
Hsu C. B., Hao S. S., Lee J. C., 2011. Personal Authentication through Dorsal Hand
Vein Patterns, Optical Engineering, 47, 067205–1–067205-10.
Jain, A.K., Bolle, R., Pankanti, S., 1999. Biometrics Personal Identification in
Networked Society. Kluwer Academic Publishers, Massachusetts.
Lee, C.J., Wang, S.D., 1999. Fingerprint feature extraction using Gabor filters.
Electronic Lett. 35 (4), 288–290.
Lee, T.S., 2003. Image representation using 2D Gabor wavelets. IEEE Trans. Pattern
Anal. Mach. Intell. 18 (10), 959–971.
Lin, C.-L., Fan, K.-C., 2004. Biometric verification using thermal images of palm-
dorsa vein patterns. IEEE Trans. Circuits Syst. Video Technol. 14 (2), 199–213.
Liu, C., 2004. Gabor-Based Kernel PCA with fractional power polynomial models for
face recognition. IEEE Trans. Pattern Analy. Mach. Intell. 26 (5), 572–581.
MacGregor, P., Welford, R., 1991. Veincheck: imaging for security and personnel
identification. Adv. Imaging 6 (7), 52–56.
Masaki, W., Toshio, E., Morito, S., and Shigeru S., 2005. Palm vein authentication
technology and its applications. In: The Biometric Consortium Conference. pp.
1–2.
Marcelja, S., 1980. Mathematical description of the responses of simple cortical
cells. J. Optical Soc. Am. 70 (11), 1297–1300.
Michael, G. K. O., Tee, C., Jin, A. T. B., 2010. Design and implementation of a
contactless palm print and palm vein sensor. In: 11th International Conference
on Control, Automation, Robotics and Vision, pp.1268–1273.
Otsu, N., 1979. A threshold selection method from gray-level histograms. IEEE
Trans. Syst. Man Cybernet. 9 (1), 62–66.
Vein recognition in Europe. 2004. Biometric Technology Today, 12(9), 6.
Wang, L., Leedham, G., Cho, D.S., 2007. Infrared imaging of hand vein patterns for
biometric purposes. Inst. Eng. Technol. Comput. Vision. 1, 113–122.
Wang, L., Leedham, G., Cho, D.S., 2008a. Minutiae feature analysis for infrared hand
vein pattern biometrics. Pattern Recogn. 41 (3), 920–929.
Wang, J.G., Yau, W.Y., Suwandy, A., 2008b. Person recognition by fusing palmprint
and palm vein images based on laplacianpalm feature. Pattern Recogn. 41 (5),
1514–1527.
Sasaki, S., Kawai, H., Wakabayashi, A., 2005. Business Expansion of Palm Vein
Pattern. Authentication Technology. Fujitsu Sci. Tech. J. 41 (3), 341–347.
Sonka, M., Hlavac, V., Boyle, R., 1999. Image Processing, Analysis, and Machine
Vision, 2nd ed. New York.
Zhang, D., Kong, W.K., You, J., Wong, M., 2003. Online Palmprint Identification. IEEE
Trans. Pattern Anal. Mach. Intell. 25 (9), 1041–1050.
Zhang, Y.B., Li, Q., You, J., Bhattacharya, P., 2007. Palm vein extraction and matching
for personal authentication. Lect. Notes Comput. Sci. 4781, 154–164.
Zhou, Y., Kumar, A., 2011. Human Identification using Palm-Vein Images. IEEE Trans.
Information Forensics and Security. In Press.