Content uploaded by Mahsa Arab
Author content
All content in this area was uploaded by Mahsa Arab on Apr 28, 2020
Content may be subject to copyright.
978-1-7281-5350-6/19/$31.00 ©2019 IEEE
Finger Knuckle Surface Print Verification using
Gabor Filter
Mahsa Arab
Faculty of Biomedical Engineering
Amirkabir University of Technology
Tehran, Iran
Mahsa.Arab@aut.ac.ir
Saeid Rashidi
Faculty of Medical Sciences and Technologies
Islamic Azad University, Science and Research Branch
Tehran, Iran
Rashidi.Saeid@srbiau.ac.ir
Abstract— The need for reliable user verification methods has
increased due to severe security concerns. Hand-based biometrics
plays an important role in providing security in real-time
environments and are more successful in speed and accuracy.
Finger knuckle images can also be used in forensic and criminal
verification applications. This paper investigates an approach for
personal verification using finger knuckle surface images. In this
paper, after applying the pre-processing and noise reduction of
finger knuckle images, by using Gabor filter extracting textural
features from both proximal and distal phalanx knuckle regions.
The textural features obtained from the Gabor filter are combined
with the features of the gray-level co-occurrence matrix and finally
classified by using K-nearest neighbor classifier and fuzzy K-
nearest neighbor classifier. In the finger knuckle images database
of 1435 Finger Knuckle print samples from 287 Fingers, we
achieved an accuracy of 97.7% with fuzzy K-nearest neighbor
classifier.
Keywords—biometric; finger knuckle; Gabor filter; gray-level
co-occurrence matrix; k-nearest neighbor
I. Introduction
A long time ago, man was looking for a way to personalize
information. Along with advances in technology, man has come
up with ways to achieve his goal of securing information. Today,
many business applications such as e-banking require security
applications with high-speed, high-precision and accurate
authentication. The convenience and ease of use of these
systems is an important task for many organizations. Personal
authentication plays an important role in society. One must be
able to determine the people identification or verification of an
individual claim whenever required. This process is known as
person authentication. A person can be recognized based on the
following three basic methods: (a) what he remembers, (b) what
he possesses and (c) who he is intrinsically [1]. The first method
is also called the knowledge-based approach and the second
method is also called the token-based approach. Knowledge-
based approaches are hard to remember and token-based
approaches are time-consuming and expensive to replace. Third
method is also called biometric traits. Biometrics involves the
automatic identification of an individual based on his
physiological or behavioral traits [2].
Biometrics, offers a more reliable solution to the problem of
person recognition [1]. Different modalities are used to
authentication: fingerprint, hand geometry, palm Hand, hand
vein patterns and finger knuckle. Identification with finger
knuckle patterns are not known as well as fingerprint and
palmprint [3]. By growing the number of the users, hand
geometry-based methods are not adequate enough. Also, palm
print based methods captures large area with lack of dominant
features like principle lines and etc. for identification. Hand vein
patterns systems are not high density and have expensive
cameras which is too costive to be used for authentication
systems [4].
In this paper, we proposed a method for verification by using
Gabor filter and gray-level co-occurrence matrix (GLCM). After
extracting minor and major regions from the finger dorsal
images, feature extraction has been done. Then, the results from
the Gabor filter and the GLCM were combined. Finally, these
results are classified by k-nearest neighbor (KNN) classifier
with referencing and fuzzy k-nearest neighbor (FKNN)
classifier.
In this paper, a user verification system using finger knuckle
images was investigated. The image preprocessing was detailed
in Section III. This section also considers, image enhancement,
binarization, determine the reference point in finger images,
normalization and segmentation. The features were extracted of
dorsal finger knuckles were illustrated in Section IV. This
section also considers two different methods for feature
extraction. The experiments and results from this paper were
summarized in Section V. Finally, the main conclusions from
this paper were summarized in Section VI.
A. Related Work
The use of finger knuckle surface images for biometric
identification has been increased in the literature. Woodard and
Flynn first to use 3D finger knuckle surface images for personal
identification and proposed it as a biometric identifier in 2005.
In their work, the finger knuckle surface images were acquired
by means of a 3D sensor. The sensor dimensions are 213 mm ·
413 mm · 271 mm and it weighs about 11 kg. In this paper, the
feature extraction process is performed using geometrical
analysis of finger knuckle surface images. The complexity of 3D
data processing is the main drawback of this design [5].
Later in 2009, Kumar et al. have proposed a new approach
for personal authentication using finger knuckle surface
imaging. The texture features of finger knuckle surface bending
are unique. The finger knuckle surface images from each user
are normalized and rotational variations in the finger knuckle
images. In their work, Peg-free imaging was used in proposed
new user authentication. Moreover, a robust approach that is
adapted to the rotation of hand-pose was also developed by
resulting due to the appearance of the rings. An important aspect
of this research is the simultaneous extraction of finger geometry
features that are used to achieve better performance [6].
AlMahafzah et al. proposed to improve the performance of
Finger Knuckle Print by use of multi-algorithm fusion level
feature. Locality preserving projection (LPP), principal
component analysis (PCA), local phase quantization (LPQ) and
log-Gabor filters (LG) have been used to extract finger knuckle
surface features. After feature extraction, four methods of Min-
Max, Z-score, Median and Median Absolute Deviation and
Tanh-Estimator have been used to normalize the features. In this
paper, results were indicated that using a multi-algorithm
verification approach improves performance than using any
single algorithm [7]. In 2014, Kumar at Hong Kong polytechnic
university to investigate the features of the finger knuckle
surface to people identification. Regardless of previous
biometrics studies, only the 'major' finger knuckle patterns were
investigated in the literature. This paper investigates the
possibility of using 'minor' finger knuckle patterns. The database
which was used in this study of 250 middle finger dorsal images
acquired from 50 subjects. The results indicate that the
simultaneous use of major and minor finger knuckle surface
improve the performance of the identification or verification
system. These references in the literature have however
exploited minor and major finger knuckle images that capture
patterns formed on the finger dorsal surface [3,8].
Later in 2015, Narishige Abe and Takashi Shinzaki at Fujitsu
Laboratory in Japan, a new method was used to record the finger
knuckle surface. The main purpose of this study was to segment
online videos of the hand and to obtain major finger knuckle
surface images. One of the advantages of this proposed method
is the use of inexpensive devices to obtain hand images. Also, in
this system, users can get finger knuckle surface prints without
holding hands while taking the image. [9].
Usha and Ezhilarasan investigated combined method for
personal identification by using finger knuckle surface. In their
work, angular geometric analysis-based feature extraction
method (AGFEM) and contourlet transform based feature
extraction method (CTFEM) have been used. [3]. In this study,
results indicate combination of texture feature with shape-
oriented features obtained high accuracy. In other work, were
proposed method for personal authentication by using finger
knuckle surface texture features. In their work, texture feature
extraction methods (TFEM), completed local ternary pattern
(CLTP) method, 2D log Gabor filter (2DLGF) method and
fourier–scale invariant feature transform (F-SIFT) method have
been used. Experimental results indicate that combination of
textural features reduced the error rate by 27%, compared to
other methods. [10].
Akku Anna Rajan and Vipin V, proposed a new approach for
identification by using minor finger knuckle images. In their
work, feature extraction techniques like 1-D log Gabor filter,
local binary patterns (LBP) and three patch local binary patterns
(TLBP) have been used. The texture pattern of finger knuckle
will not change in during of the time and unique for personal
authentication. The database which was used in this study for
Hong Kong polytechnic university. In this study, experimental
results have been investigated the possibility of using minor
finger knuckle images for human authentication. [11].
This paper has investigated biometric verification capability
by using major and minor regions of finger knuckle surface
images for humans and has proposed algorithms for the
automated segmentation of the region of interest finger knuckle,
image normalization and image enhancement. These steps are
detailed in the following sections.
II.
T
HE
P
ROPOSED
S
YSTEM
D
ESIGN
This paper proposes a method for personal verification using
finger knuckle surface, as a biometric identifier. The block
diagram of the proposed system is shown in Fig. 1. In the first
phase, the Finger Knuckle Surface images are preprocessed and
regions of interest (ROI) were extracted for further processing.
In the second phase, knuckle feature regions were identified and
features information were obtained from ROI by means of
Gabor filter and Gray Level Co-occurrence Matrix (GLCM). As
mentioned in Part I, using Gabor filter in finger knuckle surface
identification is a common solution which were used in literature
before (AlMahafzah et al, Usha and Ezhilarasan). Finally, the
texture features information was extracted of images. This
features information was classified by the KNN classifier and
FKNN classifier.
Database for finger knuckle surface has been acquired in the
Hong Kong polytechnic university campus and IIT Delhi
campus by using a contactless setup that simply uses a hand held
camera during 2006-2013. finger knuckle images database is
contributed from the male and female volunteers. all the images
are in bitmap (*.bmp) format and this database has 2515 finger
dorsal images from the middle finger of 503 subjects. In this
database, 88% of the subjects are younger than 30 years. This
dataset acquired after a very long interval (4 to 7 years) to
ascertain the stability of the knuckle crease and curved lines
[12]. The Finger dorsal images are shown in Fig. 2.
Fig. 1:
System design of the proposed personal verification
Fig. 2:
Finger dorsal images
III.
P
REPROCESSING
A. Image Enhancement
Finger knuckle surface images of the dataset received with
noise and each image has some roughness. In this paper, the
gaussian bandpass filter was used. Due to the low pass filter,
sudden information changes in the image that is considered
noise are eliminated [13]. The two-dimensional Gaussian filter
is a product of the same Gaussian function in one dimension.
The two-dimensional Gaussian filter
defined as follows:
(,)=1
2
(1)
Where σ is the standard deviation for the Gaussian curve
elongation.
B. Binarization
Each acquired finger knuckle surface image is first subjected
to thresholding operation to obtain the binarized image. The
magnitude of the thresholding limit is computed 200, this value
the measure of separability between the two classes of pixels in
the image. The resulting binarized finger knuckle surface image
contains small sporadic dots that cause incorrect results in the
extraction of finger knuckle geometry features. Also, in some
images, the boundary between the two classes is not well
defined. Therefore, morphological operations are employed on
the binarized images to enhance images and remove sporadic
dots. These two operators also help to smooth the image border
[13]. These results are shown in Fig. 3.
Fig. 3
:
(a) Finger dorsal image, (b) binarized image, (c) binarized image
with morphological operations
C. Determine the Reference Point
In this paper, the ROI are major and minor regions of finger
knuckle surface images. The first step for extracting these areas
is to determine the fingertip point in images. Knowing the
position of this point is a guide to pinpointing the location of
the major and minor finger knuckle surface. First, we specify a
reference point in the images. By finding the focal point of the
fingertip and the maximum distance to the fingernail, we find
the reference point. To do this, find the first, last and middle
white pixels. Then, using the coordinates provided by the finger
border, we find the maximum distance from this point and
specify the reference point.
D. Normalization
In this paper, the finger directions are not the same in all
images. Therefore, the location of ROI changes as the image
changes. This problem, causes the error rate to increase
significantly. For all fingers to be in the vertical direction, each
image should rotate at different rates. To do this, we used the
image rotation around the reference point. For each finger, the
same process is repeated again. After the images have been
normalized, their reference points change and the process of
finding the points again needs to be reset. These results are
shown in Fig. 4.
E. Segmentation
Each finger image needs to be segmented into major and
minor regions of the finger knuckle surface. The main idea of
this section is based on the inner width of the fingers. The minor
and major regions of finger knuckle surface are at "3/4.3" and
"2/3" lengths of tip of the finger. Based on these distances, given
the width of each image, we can extract the ROI. These results
are shown in Fig. 5 and Fig.6.
Fig. 4:
(a) Determine the reference point in the binarized image, (b)
Normalized image, (c) Background removal, (d) Determine the reference point
after the Normalized image
.
Fig. 5:
(a) Determine the inner width of the fingers, (b) Minor finger
knuckle, (c) Major finger knuckle.
IV.
F
EATURE
E
XTRACTION
A. 2D Gabor filter
We used Gabor filter to extract features from the detected
major and minor regions of the finger knuckle surface. The
most important advantage of Gabor filter is having settings that
can be applied to images in different directions and scales. In
addition, Gabor filters are robust against photometric
disturbances, such as image noise and brightness changes. A
two-dimensional Gabor filter is a Gaussian kernel function
modulated by a complex sinusoidal plane wave called the
carrier, defined as:
(,)=(,)
(,)(2)
and
are complex sine and two-dimensional Gaussian
functions, respectively.
The basic mathematical relation of the Gabor filter is
defined as follows:
(,)=
exp−
+
2
exp(2
+)
= +(3)
=−+
In (3), f is the frequency of the sinusoid, represents the
direction is a Gabor function and its values range from 0 to 180
degrees,
isthephaseoffsetofthecosinefactoristhe
Gaborfunction,whichisselectedindegreesfrom-180to
180degrees,
is the standard deviation of the Gaussian
envelope and γ the spatial dimension that determines the width
of the Gaussian factor. The appropriate values for this variable
are between 0.2 and 1 [14].
We have employed twelve Gabor filters in four scales and
three orientations as shown in Fig.7.
Fig. 6:
Finger surface images in (a)-(d), corresponding minor
finger knuckle region identified for segmentation during fine
segmentation in (e)-(h), segmented minor finger knuckle images in (i)-
(l), corresponding major finger knuckle region identified for
segmentation during fine segmentation in (m)-(p), segmented major
finger knuckle images in (q)-(t).
Fig. 7: Gabor filter in three scales and four orientations.
B. Extract Statistical Features
The output is a ROI images by Gabor filters, the images of
the same size, except that the textural features are prominent.
From these images we can get many statistical features. In this
paper, for each output image of Gabor filter, local energy, mean,
variance and standard deviation were obtained. Then the
absolute, real and imaginary part of the resulting matrix is
obtained and classified as a matrix of features.
C. Gray Level Co-occurrence Matrix (GLCM)
In statistical texture analysis, texture features are usually
divided into three types: type I, type II and higher. The GLCM
is a method for extracting second-order statistical texture
features. The GLCM is a matrix where the number of rows and
columns is equal to the number of gray levels used in the image.
GLCM is a symmetric matrix and which is defined by the joint
probability density of two different position pixels. For a digital
image of size M in N Which is indicated as I(x,y), Gray-level is
defined as (,|,).
is particular angle and
d
is
particular displacement distance. Elements in the GLCM are
defined as follows [15, 16]:
(,|,)=(,|,)
∑∑(,|,)
(4)
In this proposed method, twenty-two features of major and
minor finger knuckle images were selected. In our proposed
technique, the texture feature was extracted by the GLCM.
V.
E
XPERIMENTS AND
R
ESULTS
In this paper, absolute, real, and imaginary part of the output
of Gabor filter for major and minor finger knuckle surface
images were extracted as the statistical features. These
extracted features were examined and combined, separately.in
the next step, the results of the GLCM algorithm were extracted
from images. Then the GLCM and Gabor filter results were
studied independently and combinedly. The results were
classified by referenced KNN classifier and FKNN classifier,
individually. In this paper, each image of finger knuckle
contains 467 features. First, 2D matrix (467 × 1435 features)
are arranged by the features which were extracted in the
previous step. Each five rows of the 1
st
dimension of matrix
related to each individual and the 2
nd
dimension of matrix
related to the extracted features. In this study 287 individual
have been studied, totally. Referenced KNN classifier uses
difference of the features instead of features themselves.
After setting the 2D matrix and selecting 20 fake samples,
one reference from the real samples was randomly selected.
Then, the euclidean distance of 4 real residuals and fake samples
were calculated and the final results were placed in a new matrix.
Finally, based on the new matrix, the process of computing
neighbors and class selection were done. These results are
shown in Table I.
To achieve better results, a novel feature selection method
called 'Mahalanobis' distance based forward feature selection
was used. Speed enhancement was the dominant goal of this
proposed method. Thus, the usage of this forward selection
method increasingly improved accuracy. According to this
method, 10 and 20 superior features were extracted which were
used for classification. The results are shown in Table II and
Table III. Also, the details of the extracted features are shown in
Table IV.
TABLE I. C
OMBINATION OF
GLCM
A
ND
A
LL
P
ARTS OF
G
ABOR
F
ILTER
Combination of GLCM And All Parts of Gabor Filter
Accuracy%
Train % 30 50 70 90
Major% 87.26 87.07 87.80 88.82
Minor% 86.45 87.80 87.51 88.96
Combination of
Major and
Minor%
87.39 88.21 88.86 89.84
TABLE II. R
ESULTS WITH THE
10
SUPERIOR
F
EATURES OF THE
KNN
AND
FKNN
C
LASSIFIER
Results with the 10 superior Features of the KNN Classifier
Train % 10 20 30 40 50 60 70 80 90
Mean
(EER%) 3.79 3.29 2.80 3.04 2.91 2.91 2.48 2.45 2.53
Std
(EER%) 0.67 0.57 0.72 0.29 0.45 0.33 0.81 0.70 0.70
Results with the 10 superior Features of the FKNN Classifier
Train % 10 20 30 40 50 60 70 80 90
Mean
(EER%) 3.65 3.33 2.98 2.74 2.61 2.53 2.55 2.48 2.52
Std
(EER%) 0.38 0.24 0.29 0.21 0.32 0.17 0.35 0.37 0.58
TABLE III. R
ESULTS WITH THE
20
SUPERIOR
F
EATURES OF THE
KNN
AND
FKNN
C
LASSIFIER
Results with the 20 superior Features of the KNN Classifier
Train % 10 20 30 40 50 60 70 80 90
Mean
(EER%) 5.02 4.42 4.13 3.98 4.11 4.01 3.67 3.81 4.07
Std
(EER%) 0.83 0.81 1.01 0.99 0.87 1.11 1.00 1.06 1.03
Results with the 20 superior Features of the FKNN Classifier
Train % 10 20 30 40 50 60 70 80 90
Mean
(EER%) 3.52 2.97 2.81 2.74 2.52 2.64 2.30 2.79 2.50
Std
(EER%) 0.38 0.39 0.22 0.23 0.23 0.39 0.24 0.48 0.46
TABLE IV.
CAPTIONS OF TWENTY SUPERIOR FEATURES
Mean
minor
G
11a
Var
ma
j
or
G
13
Mean
ma
j
or
G
22
Energy
ma
j
or
G
13
Var
ma
j
or
G
12
Mean
minor
G
21
Std
minor
G
11
Energy
ma
j
or
G
12
Var
ma
j
or
G
11
Std
minor
G
13
Mean
ma
j
or
G
11
Std
ma
j
or
G
11
Energy
minor
G
21
Energy
absMinor
G
13
Energy
ma
j
or
G
11
Mean
minor
G
22
Std
ima
j
Ma
j
or
G
12
Energy
ima
g
Minor
G
14
Mean
absMa
j
or
G
11
Var
minor
G
14
a.
The mean of Gabor filter at scale 1 and orientation 1
The results obtained from classification with referencing are
promising. Since the difference of the features are more reliable
than the features themselves, more accurate results are being
expected. Experimental results indicate the effectiveness of this
method. In this paper, the points of strength are using the easy
methods and simple classification by using KNN classifier with
referencing and FKNN classifier. In this paper, we have tried to
reduce error rate with simple methods and our results are shown
that we achieved this requested. Comparison different methods
are shown in Table V.
TABLE V. C
OMPARISON DIFFERENT METHODS
Ref Features EER (%)
A.Kumar [3] Local Binary Patterns and 1D
log-Gabor filters 1.04
K. Usha a nd M.
Ezhilarasan [4]
Angular geometric analysis
based feature extraction method
(AGFEM) and contourlet
transform based feature
extraction
method (CTFEM)
1.04
A.Kumar and
Ch. Ravikanth
[6]
Geometrical features 2.35
N. Abe [9] Local Derivative Pattern 7
K. Usha a nd M.
Ezhilarasan
[10]
Texture feature extraction
methods (TFEM), completed
local ternary pattern (CLTP)
method, 2D log Gabor filter
(2DLGF) method and fourier–
scale invariant feature transform
(F-SIFT) method
decrease in error rate by
27% (in average) when
compared to the existing
benchmark system taken
A.Anna R ajan
[11]
Localbinarypattern,
1-DlogGaborfilter
and three patch LBP
12
Propose d 2D-Gabor filter and Gray-
Level Co-Occurrence Matrix 2.30
VI.
C
ONCLUSIONS
This paper, has investigated the possibility of major and
minor finger knuckle surface images for biometric verification.
In this study, was used to extract the features of finger knuckle
surface images by Gabor filter, which is the most popular tool
for extracting texture features. In the next step, the statistical
features calculated as the final features of the Gabor filter
output. Then, the difference of features is placed in features'
matrix instead of the features itself, so-called referencing. In
this study, the results of the GLCM algorithm are also
investigated. Finally, the features are classified by KNN
classifier and FKNN classifier. The results of this study indicate
the use of referencing in classification is very useful.
Accurate segmentation of major and minor finger knuckle
regions is simultaneously important for more accurate matching
and it can control the achievable accuracy verification from the
finger knuckle surface images. In this paper, the results indicate
that the human verification using minor and major finger
knuckle images can constitute a promising addition to the
biometrics security, especially covert verification of suspects,
surveillance and forensics applications using finger surface
images. In future we will concentrate on increased recognition
accuracy by combining finger knuckle surface with other
biometrics systems.
REFERENCES
[1] A. K. Jain, A. A. Ross and K. Nandakumar, Introduction to
Biometrics,Springer, 2011.
[2] D.D. Zhang, "Introduction", Palmprint Authentication, Springer Science
& Business Media, p.p. 3, 2004.
[3] A. Kumar,"Can We Use Minor Finger Knuckle Images to Identify
Humans?",Department of Computing, The Hong Kong Polytechnic
University, Hung Hom, Kowloon, Hong Kong, IEEE, 2012.
[4] K. Usha and M. Ezhilarasan "Fusion of geometric and texture features for
finger knuckle surface recognition", Alexandria Engineering Journal,
Available online, pp.683-697 ,19 November 2015.
[5] L. Woodard and J. Flynn, "Finger surface as a biometric identifier",
Computer Vision and Image Understanding, vol 100, p.p.357–384, 2005.
[6] A. Kumar and Ch. Ravikanth, “Personal authentication using finger
knuckle surface,” IEEE Trans. Info. Forensics & Security, vol. 4, no. 1,
pp. 98-110, Mar. 2009.
[7] H. AlMahafzah,"Multi-algorithm Feature Level Fusion Using Finger
Knuckle Print Biometric", Springer-Verlag Berlin Heidelberg 2012, pp.
302–311, 2012.
[8] A. Kumar,"Importance of Being Unique From Finger Dorsal Patterns:
Exploring Minor Finger Knuckle Patterns in Verifying Human Identities,"
IEEE Transactions On Information Forensics And Security, Vol. 9, No.
8, August 2014.
[9] N. Abe and T. Shinzakia,"on the fly finger knuckle print authentication",
Fujitsu Laboratories Ltd, 2015.
[10] K. Usha and M. Ezhilarasan,"Robust personal authentication using finger
knuckle geometric and texture features", Ain Shams Engineering Journal,
2016.
[11] A.Anna Rajan," Performance Evaluation of Texture Based Feature
Extraction Techniques for personal Authentication using minor finger
knuckle Matching", International Journal of Current Engineering and
Technology, vol. 5, No. 5, 2015.
[12] The Hong Kong Polytechnic University Contactless Finger Knuckle
Image Database, Version 1.0, October 2012.
http://www4.comp.polyu.edu.hk/~csajaykr/fn1.htm.
[13] Rafael C.Gonzalez, Digital_Image_Processing_3ndEd, 2008.
[14] M. Haghighat, S. Zonouz and M. Abdel-Mottaleb, "Cloud ID:
Trustworthy cloud-based and cross-enterprise biometric verification",
Expert Systems with Applications, vol 42, p.p.7905–7916, 2015.
[15] Md. B. Islam and B. Kaliyaperumal, " Feature-based Intelligent Image
Retrieval Algorithms: A Comparison", Global Science and Technology
Journal, Vol 1, No. 1, p.p..23-40, 2013.
[16] F.Albregtsen, " Statistical Texture Measures Computed from Gray Level
Co-ocurrence Matrices", Image Processing Laboratory Department of
Informatics University of Oslo, p.p. 1-14, 2008.