Content uploaded by Ali Nozari Pour
Author content
All content in this area was uploaded by Ali Nozari Pour on Jul 22, 2018
Content may be subject to copyright.
I
nt. J. Applied Pattern Recognition, Vol. 2, No. 2, 2015 111
Copyright © 2015 Inderscience Enterprises Ltd.
A new method for automatic extraction of region of
interest from infrared images of dorsal hand vein
pattern based on floating selection model
Ali Nozari Pour*
Department of Electrical/Computer Engineering,
Hakim University,
Sabzevar, Iran
Email: Khayamboy@yahoo.com
*Corresponding author
Ehsan Eslami
Department of Electrical/Computer Engineering,
Khorasan University,
Mashhad, Iran
Email: Iranianexplorer89@gmail.com
Javad Haddadnia
Department of Electrical/Computer Engineering,
Hakim University,
Sabzevar, Iran
Email: Jhaddadnia@yahoo.com
Abstract: Personal identification based on vein pattern is one of the latest
biometric approaches that have ever attracted lots of attentions. The method of
personal identification suggested in this study utilises the individual’s dorsal
hand vein pattern. However, hand spin and relocation in different trials of
image acquisition is a limiting factor in application of this approach. We
introduce a new procedure for automatic selection of region of interest (ROI)
designated ‘floating ROI’ in which adjusting the lengths and angles of sides in
the ROI quadrant, the imaging process stays resistant against any hand
relocation. Moreover, a new method for the vein pattern extraction called
‘square thresholding’ is introduced that greatly improves the extraction of
vein-patterns. For this, the average of grey level of the pixels in a
5 × 5 neighbourhood is compared with 9 × 9 neighbourhood for any pixel.
To verify validity of the proposed methods, 1,200 images taken from
100 individuals is used. As a result, an identification rate with the accuracy of
96.41% is obtained.
Keywords: personal identification; biometrics; dorsal hand vein; pattern
recognition;; region of interest; ROI; thresholding; wavelet transform; artificial
neutral network.
112
A
. Nozari Pou
r
et al.
Reference to this paper should be made as follows: Nozari Pour, A., Eslami, E.
and Haddadnia, J. (2015) ‘A new method for automatic extraction of region of
interest from infrared images of dorsal hand vein pattern based on floating
selection model’, Int. J. Applied Pattern Recognition, Vol. 2, No. 2,
pp.111–127.
Biographical notes: Ali Nozari Pour received his Bachelor and Master degrees
in 2009 and 2013 from Sajad and Hakim Sabzevari universities respectively in
the field of electronic engineering. Since 2012, he has studied a lot on infrared
images and communicated with different scientific centres around the world.
His interested fields of research are pattern recognition, artificial neural
networks and fuzzy systems.
Ehsan Eslami received his Bachelor degree from Khorasan University in 2012
in the field of information and communications technology. His interested
fields of study include: artificial neural networks, RFID and artificial
intelligence.
Javad Haddadnia received his Bachelor and Master degrees in 1993 and 1996
from Amir Kabir University in the field of electronic engineering. Then, after
taking his PhD in the same major from the same university, he started to teach
and research at Hakim Sabzevari University. His interested fields of study
include: data mining, fuzzy systems, artificial neural networks and biomedical
image processing.
1 Introduction
For thousands of years, human beings have used simple personal features such as face,
voice, walking style and other modalities to identify each other. Findings and
achievements in recent decades utilise more precise biometric characteristic such as
fingerprints (Anil et al., 1997; Tong et al., 2005) and iris (Daugman, 2004) for identifying
and authenticating individuals. However, being outside the body, fingerprint is more
susceptible to forgery while the cost of device for the iris pattern identification is high
and its use is unfriendly and cumbersome (Li et al., 2007).
Today, identification methods, using subcutaneous vein patterns, is on the spotlight
since the respective device is capable of taking images without any contact with the body
and also for the veins are localised inside the body, this method is highly safe and
resistant to counterfeit (Wang et al., 2006).
One of the current issues of identification system is selection of a suitable region of
interest (ROI) for taking image in different trials so that they could be compared with the
authentic image previously registered in the database for each individual. In the dorsal
hand vein identification system, ROI should be selected as it includes the major veins at
every modes of hand spin. In other words, all ROIs should be taken from the same areas
of the back of the hand in order to carry out the comparison procedure. To prevent the
discrepancy due to hand spin and relocation in different trials, Badawi (2006) takes
advantage of a docking device to stabilise the hand before taking its image.
A
new method for automatic extraction of region of interes
t
113
In palm print and palm-dorsal vein verification, Lin et al. (2005) and Lin and Fan
(2004) eliminate the docking device but use two points, one between the small and ring
finger (P1) and another between the middle and index finger (P2). The putative line
between the two points forms the upper side of the ROI square to be exposed to the
camera. Han and Lee (2012) also use the same method for palm vein recognition. In a
similar fashion and using the same points for dorsal hand vein recognition, Wang et al.
(2007) implement similar rectangular region that is always exposed to the camera.
However, in all of mentioned studies there exist some restrictions for the position
where the participants place their hands below the camera while taking image.
Here to eliminate such restrictions, we introduce a new resistant method of ROI
selection against any hand spin and relocation. For this, by changing lengths and the
angles between sides of a quadrilateral ROI while the datum points remain unchanged,
the ROIs selected in different trials represent the same areas of anyone hand image.
Another concern of applying subcutaneous vein pattern is the lack of intensity
difference between the vessels and surrounding tissues which hardens the extraction of
the vein pattern by global thresholding. Wang et al. (2007) and Miura and Nagasaka
(2005) respectively use local thresholding and maximum curvature method for vein
pattern extraction. Here we introduce a new method designated square thresholding that
greatly improves the extraction of vein pattern from ROI.
2 Pre-processing of image
Since infrared ray does not penetrate uniformly in all skins, the contrast of images taken
from different individuals might not be the same. The contrast also varies based on the
noise and shades caused by muscles and bones (Ynag and Shi, 2012). For this reason,
before finalising the extraction process, it is necessary to increase the image contrast and
clear the noise as much as possible. In this paper, first a median filter of the size of
3 × 3 is used to clear the noise. Then to obtain uniform distribution of intensity in various
images, a normalisation procedure is implemented so that all taken images get the same
mean and variance. For the latter, the method used by Wang et al. (2012) is followed.
p1q1
i0j0
1
MI(i,j)
pq
−−
==
=×∑∑ (1)
()
p1q1
2
i0 j0
1
VAR I(i, j) M(I)
pq
−−
==
=−
×∑∑ (2)
()
()
2
0
0
2
0
0
VAR
MI(i,j)MI(i,j)M
VAR
G(i, j)
VAR
MI(i,j)MI(i,j)M
VAR
⎧+×− ≥
⎪
⎪
=⎨
⎪−×− <
⎪
⎩
(3)
Here ‘M’ and ‘VAR’ respectively represent mean and variance of input image and
M0 = 150, VAR0 = 255 are the mean and variance of Interest respectively.
114
A
. Nozari Pou
r
et al.
2.1 Extraction of ROI
The main purpose of selecting similar region in all images is to extract the entire vein
patterns from the same determined region in order to conduct the comparison function
properly. Since later performances are going to be applied on the ROI extracted in current
step, the selecting procedure of this step has a very important role in the final result. The
main issue that needs to be noted in this step is that extracted region should be stable as
much as possible in different situations and spins as it is extracted from a specific region
in all of individual samples.
In order to extract similar region in all images without using dock, a technique is
applied to define ROI using two datum points, one between the small and ring finger (P1)
and another one between the middle and index finger (P2) (Lin et al., 2005; Lin and Fan,
2004; Han and Lee, 2012; Wang et al., 2007).
To determine the two datum points, first the contour of hand is extracted using a
Sobel filter [Figure 1(b)]. Then for each point on the hand contour, the distance
between this point and the mid-point of the wrist is calculated. As seen in Figure 1(e),
there are five local maximums (corresponding finger tips) and four local minimums
(corresponding valley points between the fingers).
Figure 1 The location of ROI in methods introduced in Lin et al. (2005), Lin and Fan (2004),
Han and Lee (2012) and Wang et al. (2007), (a) the original NIR hand image
(b) hand contour (c) according to P1P2, the square region is located and denoted as ROI
for method used in Lin et al. (2005), Lin and Fan (2004) and Han and Lee (2012)
(d) according to P1P2, the rectangle region is located and denoted as ROI for method
used in Wang et al. (2007) (e) the distance distribution diagram between the contour
points and the mid-point of the wrist (see online version for colours)
A
new method for automatic extraction of region of interes
t
115
Now, both datum points (P1 and P2) are employed to locate the ROI. For this, a line
drawn with a distance (1/6 of P1P2 length) from P1P2 line to form the upper side of ROI
square. Based on this line, three other lines are also drawn to define the ROI boundaries
[Figure 1(c)] (Lin et al., 2005; Lin and Fan, 2004).
Han and Lee (2012) use the same ROI selection method in combination with
Gabor filter for palm vein recognition. Wang et al. (2007) also use the same
procedure with some slight modifications to extract the dorsal hand vein pattern. They
convert the square into a rectangle by elongating the side lines of the ROI quadrant
(lP1P3 = lP2P4 = 1.4 × lP1P2) [Figure 1(d)].
But both methods have weak points for dorsal hand vein. The first one refers to the
small selected region specially for square ROI (SROI) (Lin et al., 2005; Lin and Fan,
2004; Han and Lee, 2012) that contributes to losing many parts of dorsal hand veins but
because most of palm veins are in the centre of palm and are small and narrow patterns,
this size is seen suitable for palm vein. Another problem is ignoring the fact that in
normal state without hand spin, the points P1 and P2 are not in line with horizon, P2 is
higher than P1 and P1P2 makes an approximately ten degrees angle with horizon in most
hands. So if we consider the set of two points in a horizontal orientation, the hand will tilt
and the ROI will not exactly lie in the centre back of the hand or palm. The third and
most important problem is its weakness to adjust the ROI, when wrist spins.
Figure 2 Two left columns relate the two hand images in two spinning modes when ROI is
selected with the method of Lin et al. (2005), Lin and Fan (2004) and Han and Lee
(2012) and two columns on the right relate the same images when ROI is selected with
the method of Wang et al. (2007)
Notes: When the wrist spins, some selected points previously located in ROI, place out of
it in corresponding image. So in these methods the veins have less possibility of
wrist spin and this is because in all of them, ROI is stable.
116
A
. Nozari Pou
r
et al.
Figure 2 addresses this issue in two images of hand with two different rotational angles
for both methods. In two left columns based on which the ROI is selected with the SROI
method (Lin et al., 2005; Lin and Fan, 2004; Han and Lee, 2012) and two
columns on the right, ROI is selected based on the rectangle ROI (RROI) method
(Wang et al., 2007).
For better understanding of the weak points in both methods, on the veins of the first
image of every hand [Figures 2(a) to 2(d)], three points are designated and corresponding
these points on the rotated images [Figures 2(a1) to 2(d1)] is designated.
For example in Figure 2(a), the three points are located in SROI but in Figure 2(a1)
only one point (number 2) is located in SROI. It means, the area of the vein that points
number 1 and 3 locate in it, in Figure 2(a) situate in SROI but in Figure 2(a1) situate out
of SROI.
Moreover, these differences are shown with corresponding distance of two down
corner of ROI with edge of the hand. For example in Figure 2(a), the distance of left
down corner point to the edge is A (L = A) but this distance in corresponding figure
[Figure 2(a1)] is more than A (L > A).
This means that with wrist spin, areas are not selected from same region, so the
extracted vein patterns for different images of a hand are not the same. For this reason, in
most previous papers of this regard (Lin et al., 2005; Lin and Fan, 2004; Han and Lee,
2012; Wang et al., 2007) while collection of the image, the position where the
participants place their hands below the camera is restricted.
2.2 Proposed method for ROI extraction
The reason of difference in selected region of images in Figure 2 in the later methods
(SROI, RROI), is not noting of simple fact “when wrist spins, one side of hand is
compressed and the other side is extended”. For example, when the left hand rotates from
wrist to the left, the distance of small finger from the wrist decreases. Figure 3 shows this
fact for one hand in three modes of spin (normal, rotate to the left and rotate to the right).
Compared to Figure 3(a) that the hand is in normal state, when in Figure 3(b) the hand
turns to the left, the distance of two points (A and C) shortens and the distance of two
points (B and D) lingers.
Figure 3 One hand in three situations of spin, (a) hand in normal situation (b) hand spins to the
left, so compression happens in the left side of the hand and stretch in the right side
(c) hand spins to the right, so compression happens in the right side of the hand and
stretch in the left side
A
new method for automatic extraction of region of interes
t
117
So if the length and angle of the left side of ROI is fixed, a bigger area of the hand is
locates in the ROI and the more veins lie in it. According to what mentioned, one should
seek an approach by which the change of angle and length of the side removes the spin
impact. So the extracted region (ROI) is not necessarily a square or rectangle with a fixed
shape in all images, but should be quadrilateral that the length and angle of which
changes depending on the angle of hand spin.
In the method proposed here, in addition to P1 and P2, a third reference point (P3), is
added between the index finger and thumb and the line P2P3 is drawn. The angle made by
P1P2 and this new line is β and the angle that P1P2 makes with the horizon is called α,
where it defines the rate of hand rotation. The point (C1) is then designated on the line
P2P3 with 0.25 of the length of P2P3 distance from P2. C1 represents the first corner of
quadrilateral ROI [Figure 4(a)].
Figure 4 The process of ROI extraction in the proposed method (FROI), (a) finding the angle and
point C1 (b) finding the point C2 (c) finding the points C3 and C4 according to the rules
(d) Identifying ROI (e) extracting ROI (f) rotating ROI and placing the upper side along
with the horizon to obtain FROI
To designate the upper side of floating ROI (FROI), a line (d1) is drawn from C1 with the
length of 1.25 of P1P2 to form an angle equal to β-10 degree with the line P2P3
[Figure 4(b)]. So, the first and second problem in SROI and RROI procedures are done
(first problem relates the small area for ROI and second problem relates the different
angle between the line P1P2 and horizon that approximately equals ten degrees with
horizon in most hands).
118
A
. Nozari Pou
r
et al.
Now to define the side of FROI, the points C3 and C4 points are defined based on the
rotation angle of α and by the following rules:
1 If 0 < α < 10° (which happens when the left hand a little rotates to the right), the
angle of two lateral sides of FROI with its upper side is set to be 90 degrees, the
length of the left side is d1 (equal to the length of upper side of FROI) and the length
of right side is shorter than d1.
2 If α = 10° (when the hand does not rotate), the angle of two lateral sides of FROI
with its upper side is set to be 90 degrees and the length of two lateral sides is d1, in
other words, FROI is square.
3 If 10° < α < 20°(which happens when the left hand a little rotates to the left), the
angle of two lateral sides of FROI with its upper side is set to be 90 degrees, the
length of the right side is d1 and the length of the left side is shorter than d1.
4 If α > 20° (when the left hand rotates to the left), the left side of FROI should
become shorter and its angle with the upper side changes. The length of the right side
remains unchanged (equal d1) but its angle with the upper side changes.
5 If α < 0° (which happens when the left hand rotates to the right), the right side of
FROI is shorter and its angle with the upper side changes. On the left side, the length
remains the same (remains equal d1) but its angle with the upper side changes.
To quantify the change in lengths of FROI, the following mathematical equation is
designed.
()
21
100 10 log | 10 |
dd
100
⎛⎞
−α−
=×
⎜⎟
⎝⎠
(4)
where d2 represents the length of FROI side of the compressed area of the hand and d1 is
the length of the other unchanged side and the upper side of FROI that equals 1.25 of the
length of P1P2. α shows the angle that P1P2 makes with the horizon.
Here, log is used because the change of α is very more than change of d
1. As
mentioned before, as in normal state without any hand spin, P1P2 makes an approximately
ten degrees angle with the horizon in most hands, α-10 is used.
In order to obtain the angles that the two lateral sides make with the upper side of
FROI, the equation below is suggested:
()
θ90 0 20
θ180 10 log | 10 | oth
=<α<
⎧
⎨=−β+ α−
⎩ (5)
where θ is the left angle that lateral sides make with the upper side of FROI
[Figure 4(d)].
After finding the length of lateral sides and the angles they form with the upper side
of the FROI, we apply geometric equation to find coordinates of two other corners
[Figure 4(c)]. Finally, by connecting these points, FROI is shaped.
Figure 4 shows the entire process of FROI extraction based on the proposed
procedure.
In Figure 5, the result of ROI extraction with three methods [SROI, RROI and FROI
(our method)] is shown. For better comparison, on the veins of the first image of all
A
new method for automatic extraction of region of interes
t
119
hands (Sa, Sb, Ra, Rb, Fa and Fb), four points are signed and in the corresponding hand
image with different spin (Sa1, Sb1, Ra1, Rb1, Fa1 and Fb1), the location of such
corresponding points is marked. As shown in Figure 5, in SROI and RROI, the location
of these points in corresponding images has changed. For instance, in Figure 5(Sa),
numbers 2 and 4 are in and numbers 1 and 3 are out of SROI. While, in image of another
spin of the same hand [Figure 5(Sa1)] numbers 1, 2 and 3 are in and just number 4 is out
of SROI.
Figure 5 Comparing SROI and RROI methods with the proposed method (FROI)
Notes: First row relates the original two hand images in two spinning modes. In second,
third and fourth rows, ROI is selected with the SROI, RROI and FROI methods.
As a shown, by using SROI and RROI methods, selected ROI changes by hand
spin, but in proposed method, by changing the angle and length of ROI
quadrilateral sides, the effects of hand spin are controlled and ROIs in different
hand spins are similarly extracted.
120
A
. Nozari Pou
r
et al.
Moreover, this difference is easily found, looking at the distance of two down corners of
ROI with edge of the hand. For example in Figure 5(Sa), the distance between the left
down corner point and the edge is assumed D1 (L = D1) while in its corresponding image
in figure [Figure 5(Sa1)], this distance is less than D1 (L < D1).
So it can be expressed that using SROI and RROI methods, the extracted veins
change by hand spin, but in proposed method, by changing the angle and length of ROI
quadrilateral sides, the effects of hand spin are controlled and all vein patterns in different
hand spins are extracted from the same region. Figures 5(Fa), 5(Fa1), 5(Fb) and 5(Fb1)
signifies this.
3 Vein pattern extraction from FROI
After extracting FROI from the image, a 5 × 5 Gaussian filter is used with a standard
deviation of 0.8 to reduce the noise as much as possible. Then the vein pattern is
extracted from the FROI. As noted above, considering the inability of the current
thresholding methods for proper extraction of the vein patterns, we present a new and
innovative method of extracting the vein pattern, using square thresholding.
3.1 Square thresholding
Since the width of the veins in images is often about 3 to 5 pixels, first a ‘W’ mask with
dimensions of 9 × 9 is designed that is consisted of two ‘Q and P’ sub-masks. ‘P’ is a
5 × 5 sub-mask located in the centre of ‘W’ and ‘Q’ is a sub-mask composed of
remaining pixels surrounding the ‘P’ (Figure 6).
Figure 6 The designed mask for extracting the vein pattern from ROI by square thresholding
method
Now, a binary matrix T(x, y) with dimensions equal to that of FROI image matrix is
devised to serve as a template. Then, the central pixel of the mask (P13) is placed on each
one of the FROI pixels and each mask value of Qj and Pi is equal to the pixel intensity of
the covered pixel of the FROI image. Then, the sum average of the sub-masks P
(grey mask) and Q (dark mask) is calculated.
A
new method for automatic extraction of region of interes
t
121
56
j
j1
25
i
i1
1
QQ
56
1
PP
25
=
=
⎧=
⎪
⎪
⎨
⎪=
⎪
⎩
∑
∑
(6)
If P > Q, the pixel of original image located on the P13 pixel of W mask is identified as a
vein and its value ‘1’ (white) is assigned to the corresponding pixel on the template,
otherwise this pixel is identified as the background and is considered ‘0’ (black) in
corresponding pixel on the template.
This comparison is done for all pixels of the FROI image to identify each as the vein
components or background pixels. Finally, the binary matrix T(x, y), (the template),
represents the vein pattern extracted from FROI. Figure 7(b) shows the result in which all
veins are properly represented.
Figure 7 Different steps of the vein pattern extraction from FROI, (a) extracted ROI
by the proposed method (b) extracting pattern by square thresholding method
(c) removal of additional edges caused by non-rectangular areas of FROI
(d) vein pattern after applying closing algorithm (e) removal of small areas that have
been erroneously identified as the vein (f) vein pattern after thinning
3.2 Correction and thinning of the vein pattern
After extraction of the vein pattern and removal of the edges made by non-rectangular
areas of FROI [Figure 7(c)], and also after filling the broken points within the veins by
using closing algorithm with square structural elements (3 × 3), all images should be
converted to 54 × 66 [Figure 7(d)]. Then using thinning algorithm the width of extracted
vein pattern decreases to 1 pixel as no break will be visible [Figure 7(f)].
In Figure 8, this method (square thresholding) is applied to extract the vein patterns
from ROIs shown in Figure 5.
As revealed, in SROI and RROI in which ROI is stable in all states of hand spin;
some veins are missing and the vein patterns change in corresponding images with
different hand spin. While, in proposed method (FROI) the change of lengths and angles
of sides leads to control of hand spin. Also, as the vein patterns of different hand spins
are extracted from the same region, they are so similar.
122
A
. Nozari Pou
r
et al.
Figure 8 Extracted vein pattern from images of Figure 5
Note: Note that some veins are missing and the vein patterns is change in the
corresponding images with different hand spin produced by using SROI (row1),
RROI (row2) as compared to the ones produced by FROI (row3).
Therefore, in spite of other methods, while imaging, no precondition is needed for the
mode and spin of hands below the camera.
4 Feature extraction
In this essay, ‘wavelet transform’ is utilised for the feature extraction. As firstly,
a two-dimensional wavelet transform is taken from every image that results in four
horizontal, vertical, diagonal and overall approximation coefficient matrixes with the
same size of the image. Here, the bior6.8 wavelet transform from the biorthogonal
functions family is used (Figure 9).
Figure 9 Displaying a wavelet transform for Figure 8 (Fb), (a) approximation coefficient
(b) horizontal details (c) vertical details (d) diagonal details
(a) (b)
(c) (d)
A
new method for automatic extraction of region of interes
t
123
In following, each of the horizontal, vertical, diagonal and level-1 approximation images
is divided into six 27 × 22 sub-images and from which, variance and entropy is
calculated. So in sum, 48 features are extracted for every ROI image, considered as its
feature vector.
4.1 Preparing data for entrance to classifier
One of the most important things that should be done before applying classifier to a
feature vector is to select the best features and to make an entrance feature vector with the
fewer dimension. To this purpose, first normalisation is applied that the mean and
standard deviation are respectively mapped to zero and one (User Manual of
MATLAB 6.5, http://www.mathworks.com). One advantage of normalisation is to speed
up the neutral network training that is offered in next step for comparison.
After normalising the quantities, it is turn to reduce the dimension of feature vectors.
One of the most applicable methods that are used in this paper is principle component
analysis (PCA) (Theodoridis and Koutrombas, 1999). To use this method, all feature
vectors, regardless of their level, are assumed as one set. Using PCA leads to dimension
reduction of feature vectors without any noticeable fall of data information. It also leads
to the orthogonality of quantities of input vector and thus the quantities do not correlate
with each other. After apply PCA, components of feature vector have decreased to eight
elements with the highest distinction which speeds up the network train at the next step.
5 Comparison
As pointed out in previous sections, to compare the feature vectors of every image and
their classifying, neutral network is used. The network applied in this study is ‘multilayer
perceptron’ (MLP). As the goal of this essay is individual identification, in addition to
trained image recognition, network must be able to generate the class number of every
input image in the output if it belongs to one of trained data, otherwise appear a number
beyond the scope of class numbers or other than 1 and –1 in the output. The output
consists of seven neurons that in fact tell us the bits required to represent the binary class
number of the input image into the network.
Given, with an appropriate selection of neurons in hidden layer, any complexity in the
input data can be modelled by one hidden layer, so the hidden layer is assumed 1 in this
paper. Also, as the range of output variation is 1 and –1, so that the ‘tansig’ function is
used for the input, output and hidden layers.
To determine the number of neurons of hidden layer, first 5 to 85 neurons are
imposed by 5-step examination. With regard to optimal results in the range of 65 to 85,
this range is examined by the step length 1 and finally the best result is obtained when the
number of neurons of the hidden layer is equal to 73. So the number of neurons in hidden
layer is determined.
5.1 The examination results using MLP
To test the proposed method, 1,200 images of left hand of 100 different individuals
(12 images for each one) from database provided by Dr. B. Sankur from Bogazici
124
A
. Nozari Pou
r
et al.
University of Istanbul in Turkey are used (Yüksel et al., 2010), Nozari pour/Sankur,
personal communication.
Such images have been taken after for different modes of everyone’s left hand. First
under normal condition, three images are taken from hand. Again, three more images are
taken after having squeezed elastic ball repetitively (closing and opening the fist) for one
minute. Then, three images are taken after having carried a bag weighting 3 kg for one
minute and finally after having cooled the hand by holding an ice pack on the surface of
the back of the hand three more images are taken.
We randomly select six out of 12 images of the 70 individuals as gallery image
(420 images of 70 different classes) and remaining images of 70 individuals (familiar
images to system) plus 360 images belong to 30 more individuals (unfamiliar images to
system) are also used as the input for identification. So, the train and test patterns for the
network are respectively 420 (6 × 70 = 420) and 780 (420 + 360 = 780).
To examine the success of the proposed method, the receiver operating characteristics
(ROC) is used (Fawcett, 2006) that is the measure of success determining in
identification system and used in most papers of this regard.
It needs to be mentioned that as against many biometrics such as fingerprint
(He et al., 2003), face (James and Dimitrijev, 2010) and … there is no unique benchmark
database for hand veins, we have to apply SROI and RROI methods on our database for
comparison.
Figure 10 compares the results of SROI, RROI and proposed method (FROI). As
shown in Figure 10, the area under the curve above and close to 1 in proposed method
shows the success of this method. In comparison with two other methods RROI has better
results than SROI for the larger ROI for the bigger patterns of the dorsal hand veins than
the palm veins.
Figure 10 ROC curve for both stable-selection and proposed method (see online version
for colours)
A
new method for automatic extraction of region of interes
t
125
Table 1 Performance comparison for three methods (SROI, RROI, FROI) on Bogazici database
Number of samples Number of recognition
correctly
Number of recognition
incorrectly
Method
Familiar Unfamiliar
Familiar Unfamiliar
Familiar Unfamiliar
Recognition
accuracy (%) Notes
SROI 420 360 359 289 61 71 83.07 61 images of 11 individuals
submitted to gallery but not
recognized correctly in their
class and 71 images of
7 individuals, not submitted
to gallery but recognised in
one of the 70 classes.
RROI 420 360 375 321 45 39 89.23 45 images of 9 individuals
submitted to gallery but not
recognized correctly in their
class and 39 images of
5 individuals, not submitted
to gallery but recognised in
one of the 70 classes.
FROI 420 360 405 347 15 13 96.41 15 images of 3 individuals
submitted to gallery but not
recognized correctly in their
class and 13 images of
2 individuals, not submitted
to gallery but recognised in
one of the 70 classes.
126
A
. Nozari Pou
r
et al.
As previously described, all of such methods have weakness while wrist spin is happen,
according to the fixed ROI. However, such weakness is less apparent in palm vein when
rotation is little, due to the dense accumulation of palm veins in the centre and having
finer patterns than hand dorsal veins but while rotation is great this weakness is apparent
for palm veins. That is why in majority of studies previously carried out, a set of spinning
limitations have been applied.
Since, in the images of our database no limitation is applied for the spin and move of
hand while imaging, the results of the proposed method (FROI) are better than those in
two other SROI and RROI methods. It indicates more practicality of this method
compared to previous ones because in a real system, there should not exist any limitation
for the user hand mode.
In Table 1, the statistic results for three methods are shown. As shown, for SROI
method, from 420 images (six images for 70 individual) that was familiar to
system only 359 images recognition correctly. These images alluded to 59 people,
so 11 people that were familiar to system, do not recognised. Moreover from
360 images (12 images for 30 individual) that was unfamiliar to system only 289 images
recognition correctly that these images alluded to 23 people, so seven people that were
unfamiliar to system, recognised familiar. Also for RROI method, nine people
(45 images) that were familiar to system do not recognise and five people (39 images)
that were unfamiliar to system and do not submitted in gallery, recognised familiar.
Lastly, for proposed method (FROI), only three people (15 images) that were familiar to
system, do not recognised correctly and two people (13 images) that were unfamiliar to
system, recognised familiar.
Thus, by applying the proposed algorithm, we obtained to identification rate with the
accuracy of 96.41% that with attention to do not applied any spinning limitations for
images, this accuracy rate is good.
6 Conclusions
In this study, one of the most popular biometric techniques, ‘dorsal hand vein pattern’, is
used for personal identification. The major concern that exists in applying this technique
is selection of a proper ROI that is resistant against any hand rotation and move and is
extracted from a specific region. We introduce a new procedure called ‘Floating ROI’ for
automatic selection of ROI in which by variation of lengths and angles of the sides of
ROI based on the rate of hand spin, the effect of hand spin is neutralised.
With respect to extraction of vein pattern from ROI, the result of using global and
local thresholding method has proved to be unsatisfactory. Here we present a novel
procedure of square thersholding which compares the average of grey levels of the pixels
in a 5 × 5 neighbourhood with 9 × 9 neighbourhood for any pixel.
1,200 images from 100 subjects’ left hand are taken and divided into two groups
(70 individuals in the first group and 30 individuals in the second group). Six out of
12 images per individual from the first group randomly are stored in database; six
remained images from each individual from the first group plus all images in the second
group (420 + 360 = 780 images) are used for the test. For feature extraction and
compression, the wavelet transform and artificial neutral network are used. Finally,
applying the proposed method in which no spinning limitations are applied for hands in
image accusation period, only 15 out of 420 images familiar to the database are not
A
new method for automatic extraction of region of interes
t
127
identified; and only 11 out of 360 unfamiliar images are mistakenly recognised as
familiar. Thus, by applying the proposed algorithm, an identification rate with the
accuracy of 96.41% is obtained.
References
Anil, J., Lin, H. and Ruud, B. (1997) ‘On-line fingerprint verification’, PAMI, Vol. 19, No. 4,
pp.302–313.
Badawi, A.M. (2006) Hand Vein Biometric Verification Prototype: A testing Performance and
Patterns Similarity, Biomedical Engineering Department University of Tennessee, Knoxville,
TN, USA.
Daugman, J. (2004) ‘How iris recognition works’, Circuits and Systems for Video Technology,
IEEE Transactions on Circuits and Systems for Video Technology, Vol. 14, No. 1, pp.21–30.
Fawcett, T. (2006) ‘An introduction to ROC analysis’, Pattern Recognition Letters, Vol. 27,
pp.861–874.
Han, W.Y. and Lee, J.C. (2012) ‘Palm vein recognition using adaptive Gabor filter’, Expert
Systems with Applications, Vol. 39, pp.13225–13234 [online]
http://www.elsevier.com/locate/eswa.
He, Y., Tian, J., Luo, X. and Zhang, T. (2003) ‘Image enhancement and minutia matching in
fingerprint verification’, Pattern Recognition Letters, Vol. 24, pp.1349–1360 [online]
http://www.elsevier.com/locate/pr.
James, A.P. and Dimitrijev, S. (2010) ‘Inter-image outliers and their application to image
classification’, Pattern Recognition, Vol. 43, pp.4101–4112 [online]
http://www.elsevier.com/locate/pr.
Li, X.Y., Guo, S.X. and Gao, F.L. (2007) ‘Vein pattern recognitions by moment invariants’,
in IEEE Conference on Bioinformatics and Biomedical Engineering, pp.612–615.
Lin, C.L. and Fan, K.C. (2004) ‘Biometric verification using thermal images of palm-dorsavein
patterns’, IEEE Transactions on Circuits and Systems for Video Technology, Vol. 14, No. 2,
pp.199–213.
Lin, C.L., Chuang, T.C. and Fan, K.C. (2005) ‘Palmprint verification using hierarchical
decomposition’, Pattern Recognition, Vol. 38, No. 12, pp.2639–2652.
Miura, N. and Nagasaka, A. (2005) ‘Extraction of finger-vein patterns using maximum curvature
points in image profiles’, MVA2005 IAPR Conference on Machine Vision Applications,
May 16–18, Tsukuba Science City, Japan, pp.8–30.
Theodoridis, S. and Koutrombas, K. (1999) Pattern Recognition, Academic Press.
Tong, X., Huang, J., Tang, X. and Shi, D. (2005) ‘Fingerprint minutiae matching using the adjacent
feature vector’, Pattern Recognition Letters, Vol. 26, No. 9, pp.1337–1345.
User Manual of MATLAB 6.5 [online] http://www.mathworks.com.
Wang, K., Ma, H., Oluwatoyin, P. and Li, J. (2012) Finger Vein Recognition, Pattern Recognition
& Intelligent Systems Department, College of Automation, Harbin Engineering University
China.
Wang, K.J., Zhang, Y., Yuan, Z. and Zhuang, D.Y. (2006) ‘Hand vein recognition based on multi
supplemental features of multi-classifier fusion decision’, in Proceeding of the IEEE,
International Conference on Mechatronics and Automation.
Wang, L., Leedham, G. and Cho, S.Y. (2007) ‘Infrared imaging of hand vein patterns for biometric
purposes’, The Institution of Engineering and Technology, doi: 10.1049/iet-cvi: 20070009.
Ynag, J. and Shi, Y. (2012) ‘Finger-vein ROI localization and vein ridge enhancement’, Pattern
Recognition Letters, Vol. 33, pp.1569–1579.
Yüksel, A., Akarun, L. and Sankur, B. (2010) ‘Biometric identification through hand vein patterns’,
ICPR 2010:Int.Conf.on Pattern Recognition, The First International Workshop on Emerging
Techniques and Challenges for Hand-based Biometrics (ETCHB2010), Istanbul, August.
