LOCAL LINE BINARY PATTERN FOR FEATURE EXTRACTION

ON PALM VEIN RECOGNITION

Jayanti Yusmah Sari, Chastine Fatichah, and Nanik Suciati

Department of Informatics Engineering, Faculty of Information Technology,

Institut Teknologi Sepuluh Nopember Surabaya
Keputih, Sukolilo, Surabaya 60111, East Java, Indonesia

E-mail: [email protected], [email protected], [email protected]

Abstract

In recent years, palm vein recognition has been studied to overcome problems in conventional
systems in biometrics technology (finger print, face, and iris). Those problems in biometrics includes
convenience and performance. However, due to the clarity of the palm vein image, the veins could not
be segmented properly. To overcome this problem, we propose a palm vein recognition system using
Local Line Binary Pattern (LLBP) method that can extract robust features from the palm vein images
that has unclear veins. LLBP is an advanced method of Local Binary Pattern (LBP), a texture
descriptor based on the gray level comparison of a neighborhood of pixels. There are four major steps
in this paper, Region of Interest (ROI) detection, image preprocessing, features extraction using LLBP
method, and matching using Fuzzy k-NN classifier. The proposed method was applied on the CASIA
Multi-Spectral Image Database. Experimental results showed that the proposed method using LLBP
has a good performance with recognition accuracy of 97.3%. In the future, experiments will be
conducted to observe which parameter that could affect processing time and recognition accuracy of
LLBP is needed

Keywords: Fuzzy k-NN, LLBP, Local Line Binary Pattern, palm vein.

Abstrak

Saat ini penelitian tentang pengenalan pembuluh darah pada telapak tangan (palm vein recognition)
telah banyak dilakukan untuk mengatasi masalah dalam teknologi biometrika yang lainnya
(pengenalan sidik jari, wajah dan iris) seperti ketidaknyamanan pengguna saat akusisi citra maupun
tingkat keakuratan pengenalannya. Namun masalah yang sering muncul dalam sistem pengenalan
pembuluh darah pada telapak tangan (palm vein) adalah fitur tekstur pembuluh darah (vein) yang
kurang jelas. Untuk mengatasi masalah tersebut, paper ini mengusulkan sistem pengenalan palm vein
menggunakan metode Local Line Binary Pattern (LLBP) yang telah teruji dapat mengekstraksi fitur
pembuluh darah pada gambar dengan jelas. LLBP merupakan metode pengembangan dari Local
Binary Pattern (LBP), metode yang menggunakan kombinasi nilai-nilai biner dari piksel
ketetanggaannya. Empat tahapan utama dalam penelitian ini yaitu deteksi Region of Interest (ROI),
preprocessing yang terdiri dari resize, penghilangan noise dan subtract citra, ekstraksi fitur
menggunakan metode LLBP dan pengenalan menggunakan Fuzzy k-NN. Metodologi yang diusulkan
telah diuji pada database citra palm vein CASIA Multi-Spectral. Hasil percobaan menunjukkan
bahwa metodologi yang diusulkan dapat mencapai akurasi sampai dengan 97.3%. Untuk penelitian
selanjutnya, diperlukan observasi untuk mengetahui parameter yang mempengaruhi waktu proses dan
akurasi pengenalan dari metode LLBP.

Kata kunci: Fuzzy k-NN, LLBP, Local Line Binary Pattern, palm vein

1. Introduction recognition systems, users have to touch the

surface of the input sensor by their finger and
Biometrics is the technology of identifying a palm. This can cause much inconvenience for the
person by using physical human features. There user and it is also possible to steal latent
have been several kinds of biometric recognition information from the fingerprint sensor. In
systems such as fingerprint, palm print, face, iris, addition, the condition of the finger surface (e.g.
etc. These conventional systems have some sweat, dryness) and skin distortion can cause
problems in terms of convenience and degraded recognition accuracy. For face
performance. In fingerprint and palm print recognition, performance highly depends on facial

116
 Jayanti Yusmah Sari, et al., Local, Line and Binary 117

expressions and illuminations, which can change. LLBP can extract robust features from the images
Iris recognition is most reliable in terms of with unclear veins it is more suitable to capture
accuracy, but the capturing device is expensive the pattern inside a palm vein image. Therefore,
and can be inconvenient compared to other recognition performance of palm vein will be
biometric systems [1]. To overcome these good (see table I). For recognition step, we
problems, vein patterns such as palm vein and considered the fuzzy k-NN [8] to be a suitable
finger vein have been studied. Vein recognition classifier since it does not need any learning
uses internal information from a person’s body algorithm so that it can decrease the processing
and vein patterns, which can be seen with infrared time. Moreover, due to the feature of palm vein
light illuminators and a camera. Compared to image that similar to each other, we choose fuzzy
finger vein, palm vein has more features that used k-NN method over k-NN method to avoid the
for recognition. Otherwise, it can show better false recognition of palm vein image.
result for recognition system. Palm vein
recognition technology is secure because the 2. Proposed Method
authentication data exists inside the body and so it
is very difficult to forge [2]. It is highly accurate. Figure 1 shows the block diagram of the
Palm vein recognition system consists of five key proposed method for palm vein recognition. The
steps: Infrared palm images captured, detection of method consists of five main stages: image
Region of Interest (ROI), pre-processing, feature acquisition (image is downloaded from CASIA
extraction and feature matching [3]. Database), ROI (Region of Interest) detection,
Most of the current available approaches for preprocessing, feature extraction by Local Line
palm vein recognition have similarities on the Binary Pattern (LLBP) and palm vein recognition
feature extraction method which utilized the by Fuzzy k-NN classifier.
features from the segmented blood vessel network
for recognition. In 2011, [5] proposed a method
for extract the texture feature of palm vein image
using Local Binary Pattern (LBP) with accuracy
of 93%. The experiment shows that unclear vein
can become the problem in palm vein recognition.
At the same year, [6] conducted an experiment for
finger vein recognition using a new variant of
LBP called local line binary pattern (LLBP) to
overcome the problem of unclear vein. LLBP
method is proposed for the first time by [7] in
2009 and it applied to face recognition.

TABLE I
EXPERIMENT IN RECOGNITION SYSTEM
Recognition Feature
Matching Result Figure 1. The overview of proposed method for palm vein
System Extraction
[5] recognition.
Palm Vein LBP ROC Curve Accuracy
93%
Finger Vein [6]
LLBP Hamming EER 2.1 Dataset
Distance 1.89%
Face[7] LLBP Local Accuracy The dataset used in this work is downloaded
Discriminant 100% from the CASIA Multi-Spectral Palmprint Image
Analysis
Database V1.0 [9]. This dataset consists of palm
vein images of 50 individuals (six samples per
Palm vein images are not always clear, individual), captured under six different NIR
therefore, segmentation errors can occur during illuminators. These six images were acquired
feature extraction process due to the low qualities from each user and these images were acquired in
of palm vein images. When the veins are not two different data acquisition sessions (three
segmented properly, the recognition accuracy may images in each session) with a minimum interval
be degraded. To overcome these problems, we of one month. Palm veins are most visible under
propose palm vein recognition system based on the illuminator at 940 nm wavelength. The sub-set
local feature using LLBP (Local Line Binary used here contains all samples from all individual
Pattern) method. The main difference between left hands under 940 nm illuminator [5]. Figure 2
LLBP and LBP is its neighborhood shape is a shows the sample of palm vein image used in this
straight line with length N pixel, unlike in LBP, research.
which is a square [6]. The straight-line shape of
118 Journal of Computer Science and Information, Volume 8, Issue 2, June 2015

edge of image (I) we maintaining high contrast

compared to the low contrast.

2.2.2 Image Contour

The principle of image contour is similar to

edge detection. But the image contour detected
the curve while the edge detection detected
edge. Detection of the curve is used as a straight
line to suffer the effects of discretization that
create zigzag lines in accordance with the tilt
Figure 2. The sample of palm vein image. angle. Image contour is done after filtering
using anisotropic diffusion filter by pulling on
2.2 ROI Detection the contour plot of the intensity of the image.
To increase the recognition accuracy and
reliability, it is important to extract the features
of vein patterns from the same region within
different palm vein images. This process is
known as extraction of region of interest (ROI).
Figure 3 shows the overview of ROI detection
step in this research, which consist of edge
detection using anisotropic diffusion filtering,
find the maximum image contour, find interest Figure 4. The result of image contour (left) and max
point and final region (ROI) detection. contour (right).

At first axes image contour detection is to

detect areas that contrast with the background
including hand edge and the edge of the palm
vein. Then max contour to thicken the layout
contour edges on the outside with maximizing
of x and y that are detected is the edge of the
hand. With the edge being detected we can
separate the foreground (hand) and the
background.

2.2.3 Interest Point

Figure 3. The overview of ROI detection.

2.2.1 Annisotropic Diffusion Filtering (ADF) P1

Annisotropic diffusion used in edge P2
detection algorithms. By running the diffusion
with an edge seeking diffusion coefficient for a
P3
certain number of iterations, the image can be
evolved towards a piecewise constant image
with the boundaries between the constant
components being detected as edges [10].
Annisotropic Diffusion make image more
smooth so in the next step edge can detected Figure 5. The interest points of palm vein image.
more specific. That is because the diffusion Interest point is some points in the hand
coefficient is chosen to vary spatially, in such a that have specific interest maximum point on
way, as to encourage intra region smoothing in the contour. Which the maximum is set to the 3
preference to inter region smoothing. As the highest contours. The three points are the point
region boundaries a high quality edge detector of the arch between the fingers. Those points
which successfully exploits global information. located between pinkie and ring finger (P1), ring
In this research we use 13 neighborhood pixels
for diffusion and because we want to detect
 Jayanti Yusmah Sari, et al., Local, Line and Binary 119

finger and middle finger (P2) and also between 2.4 Feature Extraction
the middle finger and index finger (P3).
After preprocessing, the next step is to
2.2.4 Final Region extract the feature of image. In this research, the
texture feature of palm vein image is extracted
so that vein can be seen clearly. We use the
Local Line Binary Pattern (LLBP) method to
V2 extract the feature and compare the result of
V1 V2 LLBP with the result of Local Binary Pattern
V1 (LBP) method.
V3 V4

Figure 6. The final region (ROI) detection.

Next step is to pull out line from P1 to the

furthest edge and set the middle point in
between (V1). The second line is to pull out
from P3 to the furthest edge in index finger area (a) (b) (c)
and also set the middle point in between (V2). Figure 8. (a) Extracted ROI, (b) resized image and (c)
Then, the points V1 and V2 showed in figure 6 subtracted imageP.
(left) can be located. With the points V1 and V2,
the ROI is defined as a rectangular region V1– The Local Binary Pattern (LBP) operator is
V2–V3–V4, where lV1V3 = 1.25 × lV1V2. Figure a texture descriptor based on the gray level
8(a) shows the result of the extracted ROI for comparison of a neighborhood of pixels [4]. The
the palm vein image in figure 6 (right). The original operator considers a 3×3 neighborhood
image with ROI detected is rotated then so that of 8 pixels around a center pixel . The
line V1V2 is horizontal. The rectangular region is threshold for these neighborhood pixels is
cropped and used for feature extraction to get the value of the center pixel and the result
the texture of vein. considered as a binary number or its decimal
equivalent [5]. LBP operator
2.3 Preprocessing showed in equation(1) and the threshold for the
neighborhood pixels showed in equation(2).

Motivated by LBP, Petpon and Srisuk [7]

proposed an LLBP operator for face
recognition. The operator consists of two
components: horizontal component (LLBPh) and
vertical component (LLBPv). The magnitude of
Figure 7. The overview of preprocessing step. LLBP can be obtained by calculating the line
binary codes for both components.
After ROI extraction, the next step is
preprocessing that showed in figure 7. First, (2)
resize the image using bicubic interpolation
which the dimension is 216 x 216 pixels, then
enhance the image using median filter with the The main difference between LLBP and
5 x 5 kernel. The last step of preprocessing is original LBP are as follows: 1) the LLBP
subtracting the original image from the operator has a straight line shape, this will
background. Figure 8(b) and figure 8(c) shows greatly assist LLBP operator in capturing the
the results of resizing and subtracting step.
120 Journal of Computer Science and Information, Volume 8, Issue 2, June 2015

change in image intensity, 2) the pattern of LLBPh, LLBPv and LLBPm are LLBP on
image at left and right side of the center pixel of horizontal direction, vertical direction, and its
the line are mirror because of the distribution of magnitude, respectively. N is the length of the
binary weight at left and right side are equal, line in pixel, hn is the pixel along with the
Thus, the number of pattern can be reduced. The horizontal line and vn is the pixel along with the
illustration of LLBP operator is shown in figure vertical line. c = N/2 is the position of the center
9, and its mathematic definitions are given in pixel, hc on the horizontal line and vc on the
equations(3)–(5). vertical line, and s(•) function defines a
thresholding function as in equation(2)

Figure 9. Example of LLBP operator [6].

(3)

(4)

(5)
sequence for horizontal (vertical) component is
defined from left (top) as 010111001111(2)
Employing equations(2) and (3), the (101001011101(2)). Hence, the binary code for
horizontal component of LLBP (LLBPh) extracts LLBP is 010111001111101001011101(2).
a binary code of N − 1 bits for each pixel. The
same numbers of bits are extracted by the
vertical component of LLBP (LLBPv) using
equations(2) and (4). Consequently, by Figure 10 shows the results of feature extraction
concatenating the binary codes from LLBPh and using LBP and LLBP, respectively.
LLBPv, the total binary code of LLBP for each
pixel is 2(N − 1) bits. In figure 9, the binary
 Jayanti Yusmah Sari, et al., Local, Line and Binary 121

testing data that each palm of a person will

has 5 images as database and 1 image as testing
data, respectively.
There are three experiments in this
research. Experiment 1 and 2 are conducted to
show how parameters k of fuzzy k-NN and N of
(a) (b) LLBP affect the verification accuracy of this
research, while experiment 3 conducted to
Figure 10. The extracted feature of image using (a) LBP compare the accuracy of LBP and LLBP
and (b) LLBP. method.
2.5 Matching (Recognition)
BEGIN
Input x, (testing data)
We use fuzzy k-NN classifier to match the Set k, 1 ≤ k ≤ m
extracted palm vein image from testing data Initialize i=1
with the one from training data. The basic DO UNTIL (KNN from x is found)
concept of this classifier is to assign Calculate distance from x to xi
membership as a function of the object’s IF (i ≤ k)
THEN add xi to E
distance from its k-nearest neighbors and the ELSE IF (xi is closer to x than any
memberships in the possible class l. The previous NN) THEN
pseudo-code of fuzzy k-NN classifier presented Delete the farthest neighbor
in figure 11. and include Xi in the set E
END IF
Consider W={w1, w2, ..., wm} a set of m END DO UNTIL
labeled data, x is the input for classification, k is Initialize i=1
the number of closest neighbors of x and E is DO UNTIL (x has membership value of all
the set of k nearest neighbors (NN). Let µi(x) is class i)
Calculate µi(x)
the membership of x in the class i, m be the
Increment i
number of elements that identify the classes l, END DO UNTIL
and W be the set that contain the m elements. To END
calculate µi(x), we use equation(6) [8].
Since we use fuzzy k-NN method, each Figure 11. Pseudo-code of fuzzy k-NN classifier.
element of x testing data is classified in more
than one class with membership value µi(x).
The decision to which class the elemen of x (7)
testing data belongs is made according to which
class the element of x testing data has the
highest membership value µi(x). Experiment 1: we used 4 different k values
in fuzzy k-NN: k=2, k=3, k=5, k=7 and compare
the accuracy of each k, k is the number of
closest neighbors that used when data is being
(6) classified. Motivated by [7], In this experiment,
we use the standard value of N, N=13, as the
neighborhood pixel for LLBP method. Since we
used 6-fold cross-validation procedure, the
predictive accuracies on the testing data of the 6
3. Experiment Result and Analysis runs of each k are averaged and reported as the
predictive accuracies. In table II, classification
The proposed method was implemented in result with predictive accuracy is reported for all
Matlab R2011b and evaluated using CASIA of k values. Accuracy is a comparison between
palm vein database [9]. There are 50 sets of left the number of correctly recognized data and
palm used with 6 sample images for each palm total number of data set. The accuracy is
of a person. In total, the database contains 300 calculated using equation(7).
images. The spatial and depth resolution of the As can be seen in table II, the highest
palm vein images were 768 × 576 pixel and 256 recognition result obtained from k=2 with
gray levels, respectively. 97.3% of mean accuracy while the lowest is
Using K-fold cross validation procedure obtained from k=7 with 94% of mean accuracy.
with K=6, we split data into 6 folds. In every K-
fold, data was divided into 250 database and 50
122 Journal of Computer Science and Information, Volume 8, Issue 2, June 2015

TABLE II different values of k for fuzzy k-NN: k=2, k=3,

THE RESULT OF EXPERIMENT 1
k=5, k=7. Table III shows the result of
Accuracy (%)
K-Fold experiment 2. The accuracy is the mean
k=2 k=3 k=5 k=7
1 98 96 90 90 accuracy of 6-fold cross-validation.
2 96 96 92 94
3 96 96 96 94
4 98 98 96 98 TABLE III
5 98 98 96 92 THE RESULT OF EXPERIMENT 2
6 98 98 96 96 Accuracy (%)
Mean (%) 97.3 97 94.3 94 N
k=2 k=3 k=5 k=7
9 97.3 97.0 94.3 94.0
Experiment with higher k shows the lower 13 97.3 97.0 94.3 94.0
value of mean accuracy. This result also 17 97.3 97.0 94.0 94.7
affected by the behavior of the data set used in 21 97.3 97.0 94.0 94.7
this experiment that has many classes (50
classes) but few instances (5 instances). As can be seen in table III, there is no
Considering the result of experiment 1, we use significant difference of mean accuracy for each
k=2 as the optimum value for experiment 3. N value of LLBP. This shows that the number
of neighborhood pixel used for LLBP not affect
the verification of accuracy in this research. But
FKNN KNN SVM according to [6], parameter N will affect the
100 processing time, the bigger the N value, the
longer the processing time for feature
Accuracy (%)

90 extraction. In this research, we do not conducted

experiment to observe how the N value can
affect the processing time.
80 Experiment 3: we also use the LBP method
to extract the feature in order to compare its
70 accuracy with LLBP method. According the
1 2 3 4 5 6 result of experiment 1, we used k=2 as the
K-Fold optimum value for recognition using fuzzy k-
NN. The result of experiment 3 showed in
Figure 12. Comparison accuracy of FKNN, FKNN and figure 13.
SVM. We have compared our proposed method
results with the result of LBP method using k=2
Due to the feature of palm vein image that for Fuzzy k-NN classifier. We use K-fold cross
similar to each other, we choose fuzzy k-NN validation procedure [11] to validate the
method over k-NN method to avoid the false accuracy for both methods. As can be seen in
recognition of palm vein image. To provide that figure 13, for all of K-fold, our proposed
FKNN is more suitable with the data on this method, LLBP method, has higher accuracy
experiment, we also compare the accuracy of compared to LBP method. From 6-fold cross
FKNN and the other classification method such validation, LLBP has 97.3% of mean accuracy
us KNN and SVM (support vector machine). whereas LBP method has 90.3% of mean
Figure 12 shows the comparison accuracy of accuracy.
palm vein recognition using FKNN, KNN and The LLBP operator has a straight line
SVM. shape, this will greatly assist LLBP operator in
As can be seen in figure 12, FKNN method capturing the change in image intensity.
has better accuracy than KNN method and SVM Therefore, LLBP can extract robust features
method. FKNN has 97.3% of mean accuracy from the images with unclear veins it is more
whereas KNN method has 95% of mean suitable to capture the pattern inside a palm vein
accuracy and SVM method has 78% of mean image. Vein feature in extracted image using
accuracy. Thus we conclude that FKNN method LLBP method is more distinct than vein feature
is more suitable than KNN and SVM method in extracted image using LBP method as can be
for classification the feature of palm vein image seen in figure 10. The more distinct feature
in this experiment. extracted, the higher the accuracy we get. Thus
Experiment 2: we used 4 different values of we conclude that LLBP method is more reliable
N that represent the number of neighborhood than LBP method for extract the feature on palm
pixel that used in LLBP: N=9, N=13, N=17 and vein recognition.
N=21 and we tested each N value with 4
 Jayanti Yusmah Sari, et al., Local, Line and Binary 123

[3] L. Mirmohamadsadeghi, & A. Drygajlo,

LLBP LBP “Palm vein recognition with local binary
100 patterns and local derivative patterns,” In
Biometrics (IJCB), 2011 International
95 Joint Conference on , pp. 1-6, 2011.
Accuracy (%)

[4] T. Ojala, M. Pietikainen, & T. Maenpaa,

90 “Multiresolution gray-scale and rotation
invariant texture classification with local
85 binary patterns,” Pattern Analysis and
Machine Intelligence, IEEE Transactions
80 on, vol. 24, pp. 971-987, 2002.
1 2 3 4 5 6 [5] L. Mirmohamadsadeghi, & A. Drygajlo,
“Palm vein recognition with local texture
K-Fold patterns,” IET Biometrics, vol. 3, pp. 198-
206, 2014.
Figure 13. Comparison accuracy of LBP and LLBP. [6] B. A. Rosdi, C. W. Shing, & S. A. Suandi,
“Finger vein recognition using local line
4. Conclusion and Future Work binary pattern,” Sensors, vol. 11, pp.
11357-11371, 2011.
This paper has proposed LLBP method, a [7] A. Petpon, & S. Srisuk, Face recognition
new approach of LBP, for reliable personal with local line binary pattern. In Image
identification using palm vein representations. and Graphics, 2009. ICIG'09. Fifth
The LLBP method can capture the pattern inside International Conference on, pp. 533-539,
a palm vein image with recognition accuracy of 2009.
97.3% from the left hand palm vein images of [8] J. M. Keller, M. R. Gray, & J. A. Givens,
the CASIA database. The result of this research “A fuzzy k-nearest neighbor algorithm.
shows that LLBP method is more reliable than Systems, Man and Cybernetics,” Systems,
LBP method for feature extraction on palm vein Man and Cybernetics, IEEE Transactions
recognition. In future, experiment conducted to on, vol. 15, pp. 580-585, 1985.
observe what parameter that can affect the [9] CASIA Multi-Spectral Palmprint Image
processing time and recognition accuracy of Database V1.0
LLBP is needed. http://biometrics.idealtest.org/dbDetailFor
User.do?id=6.
References [10] P. Perona, & J. Malik, “Scale-space and
edge detection using anisotropic
[1] K. R. Park, “Finger vein recognition by diffusion,” Pattern Analysis and Machine
combining global and local features based Intelligence, IEEE Transactions on, vol.
on SVM,” Computing and Informatics, 12, pp. 629-639, 1990.
vol. 30, pp. 295-309, 2012. [11] R. Kohavi, “A study of cross-validation
[2] Palm Vein Pattern Authentication and bootstrap for accuracy estimation and
Technology, Fujitsu white paper, 2005. model selection,” In IJCAI, vol. 14, pp.
1137-1145, 1995.