Content uploaded by Yingbo Zhou
Author content
All content in this area was uploaded by Yingbo Zhou
Content may be subject to copyright.
Figure 1: Block diagram for the human identification using finger knuckles.
Abstract—The usage of finger knuckle i mages for personal
identification has shown promising results and generated lot of
interest in biometrics. In this work, we investigate a new
approach for efficient and effective personal identification using
KnuckleCodes. The enhanced knuckle images are employed to
generate KnuckleCodes using localized Radon transform that
can efficiently characterize random curved lines and creases.
The similarity between two KnuckleCodes is computed from the
minimum matching distance that can account for the variations
resulting from translation and positioning of fingers. The
feasibility of the proposed approach is investigated on the finger
knuckle database from 158 subjects. The experimental results,
i.e., equal error rate of 1.08% and rank one recognition rate of
98.6%, suggest the utility of the proposed approach for online
human identification.
I. INTRODUCTION
ESEARCH efforts for human identification using peg-free
hand imaging have shown high interest in the biometrics
literature. Palmprint [1], hand geometry [2], palm vein [3],
palm dorsal vein patterns [4] have been investigated as an
alternative (or compliment) to the more popular fingerprint
identification. However, the research efforts to investigate
the utility of finger knuckle patterns for personal
identification have been very limited. As a result, there is no
known use of knuckle pattern identification in commercial or
civilian applications.
The anatomy of human hands is quite complicated and
intricate but highly responsible for the individuality of
hand-based biometrics. The finger skin surface is highly rich
in texture which consists of palmer friction ridges and palmer
flexion creases [5]. The palmside hand surface regions are
highly popular for personal identification and employed to
acquire fingerprint, palmprint and hand shape features. The
back surface of hands is plainer as it has no specialized
functions. While fingers bend forward and assist in grasping
the objects against palm, its backward motion is resisted.
.
There are tendons running along the back of the fingers that
assist in extending the fingers when required. However, if
fingers are attempted to be bent past certain points, the
tendons will hyper-extend, injuring the fingers. The
unidirectional bending of fingers generates highly unique
pattern formations, on the finger surface joining proximal
phalanx and the middle phalanx bones, and employed for the
finger knuckle based identification.
A. Related Work
The human identification using 3D and 2D finger surface
details have generated lot of research interest and promising
results have been detailed in the literature [6]-[10]. Woodard
and Flynn [6] have demonstrated that finger knuckle surface
is quite distinctive 3D surface and detailed the user
identification from range images using curvature and
shape-based index. The usage of range images to extract
distinctive 3D finger geometrical features is also detailed in
[2], [7]. Recently, reference [8] has detailed the usage of
finger knuckle surface for online user identification using
combination of sub-space features. This paper also describes
the simultaneous extraction and usage of geometrical features
for performance improvement. The image details from the
palmside finger surface can also be employed for the user
identification as detailed in [9] and [10]. However, the texture
details/patterns that can be acquired from low-resolution
palmside finger surface imaging are very limited but useful in
improving the performance from the simultaneously acquired
palm surface (palmprint) as detailed in [9]-[10].
B. Selecting a Finger
Our approach in the selection of one of the fingers for the
human identification is motivated by the anatomical studies
[11]. The forefinger and little finger have very high mobility
and agility. The ring finger is regarded as the clumsiest and
stiffest of all fingers. Therefore the ring
Human Identification Using KnuckleCodes
Ajay Kumar, Yingbo Zhou
Department of Computing
The Hong Kong Polytechnic University
Hung Hom, Kowloon, Hong Kong
Email: ajaykr@ieee.org, 08509610g@polyu.edu.hk
R
Figure 2: (a) Finger image, (b) segmented finger knuckle
image, (c) mean bi-cubic image, (d) enhanced knuckle image.
(a)
(b) (c) (d)
†finger offered high user inconvenience even with the
peg-free imaging considered in this work. The middle finger
offered higher surface area and higher stability in acquiring
knuckle patterns. Therefore middle fingers images were
selected for evaluating knuckle identification. It may also be
noted that the comparative evaluation of performance from
the four finger knuckle images in [8] suggested that middle
finger achieves the best performance. It is also possible to
acquire four fingers simultaneously, as in [8], and improve
the performance from the combination since four knuckle
patterns present quite distinct texture.
II. BLOCK DIAGRAM
The block diagram for the proposed personal identification
system using finger knuckles is shown in figure 1. The
backside (knuckle side) fingers images are conveniently
acquired from a digital camera for the identification. Each of
these images requires localization of region of interest for the
feature extraction. This region of interest is the region having
maximum knuckle creases and automatically extracted using
the edge detection based approach detailed in reference [8].
The extracted knuckle images have non uniform
illuminations and therefore require image enhancement. The
enhanced knuckle images are employed for the feature
extraction. The feature extraction approach employed in this
work is detailed in section III and generates binarized codes
KnuckleCodes for each of the localized pixel positions. The
extracted knuckles are used to match with those in the
database stored during the user registration. The extracted
knuckle images have wide translational and rotational
variations which have to be accounted during matching
stages. Therefore the matching distance employed in this
work is based on comparison in the neighborhood and is
detailed in section IV. The user identity is ascertained from
the identity associated with the KnuckleCodes that generates
minimum matching distance.
III. KNUCKLE IMAGE ENHANCEMENT
The finger surface is highly curved and results in uneven
reflection which also generates shadow. The knuckle images
therefore have low contrast and uneven illuminations. These
undesirable effects are reduced in the pre-processing step
using nonlinear image enhancement. The image enhancement
steps employed in this work are summarized as follows:
(a) Each of the knuckle images is divided into 10 × 10
pixels sub-blocks, and the gray-level in each of the blocks is
computed by summing up all pixel values in this block and
then divided by a constant N. Selecting N = 256 achieved the
desirable enhancement performance in our experiments. The
gray-level in each block represents the average background
illumination from the corresponding sub-block.
(b) The estimated blockwise average from (a) is used to
construct a background illumination image, of the same size
† The ring finger is also known as toe finger since it has no more mobility
than a toe.
as original knuckle image, using bi-cubic interpolation.
(c) The background illumination image from (b) is
subtracted from the original image for the normalization of
uneven illumination. The resulting normalized image is
subjected to the histogram equalization to obtain the final
enhanced knuckle image.
Figure 2 shows a sample of finger knuckle image and the
corresponding steps to obtain the enhanced image. It can be
noticed from figure 2(d) that the employed steps for the image
enhancement have been quite successful in improving the
illumination and the contrast of the finger knuckle images.
IV. FEATURE EXTRACTION
The enhanced knuckle image mainly consists of curved
lines and creases. Therefore the feature extraction approaches
that are highly successful in the characterization of such
random textured biometric traits (e.g., iris, palm, etc.) deserve
investigation. There were two key factors in the selection of
feature extraction approach for our system. Firstly, the
localized estimation of features, rather than the global
estimation, was preferred as such approaches have shown to
deliver better performance. Secondly, the computational
complexity another factor as the developed system is
intended for the online user identification.
The feature extraction approach is focused on the
detection and characterization of curved knuckle lines and
creases from the enhanced finger knuckle images. The Radon
transform can effectively accentuate such line features by
summation of image pixels along several directions and is
highly computationally efficient. However the lines and
creases in the images are curved rather than linear along
image dimensions, therefore localized Radon transform was
employed for the feature extraction.
A. Localized Radon Transform
The localized Radon transform (LRT) for a discrete image
g[m ,n] on a finite grid
can be defined as:
Figure 3: Computing localized Radon transform in a 14 × 14
pixel region in the directions of 0o, π/6, π/3, π/2, 2π/3, 5π/6
and the Xp is 2 pixels wide.
Μ ∑,
, (1)
where 0, 1, …, 1, q is a positive integer, and
is centred at ,
. The represents set of points on
such that
,:
, (2)
where p denotes the slope of , i.e. slope of line passing
through the centre ,
of
. The LRT is not an
invertible transform but useful to represent line and crease
like features. The line width of can be empirically selected
corresponding to the width of the observed knuckle lines in
the acquired finger images. In this work, this line width is
therefore empirically selected as two pixels. Figure 3 shows
the computation of LRT for the width of equals 2, when
the localized region of 14 × 14 pixels (l = 14) is selected in 6
(D = 6) directions. The main objective of employing LRT in
our approach is to efficiently and effectively ascertain the
orientation p of knuckle lines and creases in a localized
region.
B. KnuckleCodes Using LRT
The orientation of curved lines and creases is estimated
from the localized magnitude of LRT. This orientation for
every pixel centred at ,
on
is estimated from the
summation of pixels along the line of given slope p. It should
be noted that the curved knuckle lines (figure 2) appear as
dark lines, i.e., identified from the pixel gray-levels which are
smaller than the background. Therefore minimum, rather than
maximum, magnitude of is of interest. The index of the
dominant direction at every pixel forms the feature and is
computed as follows:
,
arg
Χ , 1,2, …, (3)
where the ,
represents the line direction or the
dominant index of pixel ,
. This operation is repeated
as the centre of lattice
moves over all the pixels in the
image. The dominant direction at every pixel is binary
coded using Z binary bits and is referred to as KnuckleCode in
this paper. The binarized encoding of orientation index from
every pixel is preferred mainly for the efficient template
storage and matching. In order to effectively handle the pixels
localized at the edges of images, zero padding is employed for
such boundary pixels. As long as the padding pixel values are
same, the dominant direction ,
will not be
influenced by such padding. This guarantees the correctness
of the algorithm in computing ,
for boundary pixels.
V. MATCHING KNUCKLECODES
The matching of two KnuckleCodes, extracted from two
different fingers, should be robust to handle the translation
and rotational variations in the localized knuckle images.
These variations can be either due the translation and/or
rotation of presented fingers (due to peg-free imaging) and/or
due to the inaccuracies in the localization of knuckle image
regions. These inaccuracies can also be due to the influence of
the shadows resulting from the highly curved 3D finger
surface. Therefore, the approach employed in this work
attempts to generate the best possible matching score while
considering the translation and rotation of fingers that
sometimes present partially matching knuckles.
The matching score between two Z bit KnuckleCodes R
and T, obtained from two corresponding knuckle images, is
generated as follows:
, min,, ,∑∑
,
,
,
(4)
where
represents the KnuckleCodes (acquired during user
registration) with the width and height expanded to m2w
and n 2h, while
w floor
, h floor
, (5)
x,y
,
,, ,
, (6)
,
0
1 (7)
where b = 1, 2, ..Z which denotes binary bits for the Z bit
KnuckleCodes while m and n denotes the width and height of
KnuckleCodes, i.e., size of R and T. It may be noted that the
size of KnuckleCodes is significantly smaller than the original
knuckle image size and depends on the local region size
(one fourth of the knuckle image size for 2) employed
in this work. The KnuckleCodes can be considered similar to
FingerCodes [13] or IrisCodes [14] as they also represent the
localized texture information.
VI. EXPERIMENTS
The proposed approach for human identification using
knuckle images is rigorously evaluated on finger image
database from 158 subjects. This database was acquired over
a period of 11 months and each subject/volunteer contributed
five image samples which resulted in total of 790 images.
These images were acquired using a digital camera in an
indoor environment using unconstrained (peg-free) setup as
detailed in [8]. The middle finger images from each of the
subjects are employed to automatically extract 80 × 100 pixel
knuckle region using the segmentation method detailed in [8].
Figure 2(a) shows a sample of acquired middle finger image
and correspondingly segmented knuckle image in figure 2(b).
The segmented images have low contrast and may suffer from
Figure 4: Finger knuckle image samples fro
m
(a), and the corresponding enhanced image
s
Figure 5: Gray level representation
generated for knuckle image in (a) using L
R
even Gabor filters in (c).
(a)
(b)
(a)
(b)
non-uniform illumination. Therefore eac
h
images were enhanced as detailed in sec
t
knuckle image samples from the four subj
e
figure 4(a) and the corresponding enhanc
e
are reproduced in figure 4(b).
The enhanced knuckle images are subje
c
extraction using LRT as detailed in secti
number of candidate directions (D) fo
r
empirically fixed to 8. The performan
c
achieved by 5-fold cross validation an
d
experimental results is presented. This r
e
realistic experiments, similar to as in [9]
images have large variations within the sa
m
from shadows, illumination and pose chang
e
representation of KnuckleCodes generated
f
image sample in figure 5(a) is shown in fig
u
m
the five users in
s
amples in (b)
of KnuckleCodes
R
T in (b), and using
Figure 6: The ROC curves
(c)
h
of the knuckle
tion II. The five
e
cts are shown in
e
d image samples
c
ted to the feature
on III. The total
r
every pixel is
c
e evaluation is
d
the average of
e
presents a more
]
, as the knuckle
m
e class resulting
e
s. The gray level
f
rom the Knuckle
u
re 5(b)-(c). The
receiver operating characteristics (
R
genuine and imposter 124030 (15
scores is shown in figure 6. The co
m
results from our approach, with
approach, i.e., eigenknuckles and
f
i
s
[8] on the database employed in th
i
figure 6. Table 1 summarizes the
b
e
s
rate from the experiments. We als
o
for the recognition and the co
r
cumulative match characteristics (
C
7.
Another possible approach fo
r
features is to employ real part
o
ascertain the orientation at every p
filtered response. Such an approach
the palmprint data in [15] and ac
h
Therefore the generation of
K
nuck
l
filters is also investigated and the
shown in figure 6-7. The twelve re
a
15 mask size, centred at frequ
e
employed to achieve the best perfo
r
number of filters and mask sizes ge
n
The experimental results in figure 6
the performance from the
K
nuck
l
LRT, i.e., KnuckleCodes (Radon), i
s
to those from real Gabor filter base
d
Equal
E
EER (%)
KnuckleCodes
(Radon)
KnuckleCodes
(Gabor)
Mean 1.08 2.66
Std
deviation 1.08 1.81
Table 1: Comparative performanc
from the experiments.
R
OC) using 790 (158 × 5)
8 × 157 × 5) matching
m
parison of experimental
the appearance based
herknuckles employed in
i
s work is also shown in
t case average equal error
o
performed experiments
r
esponding comparative
MC) are shown in figure
r
extracting orientation
o
f Gabor functions and
ixel using the maximum
has been investigated on
h
ieves promising results.
l
eCode using such Gabor
comparative results are
a
l Gabor filters, with 15 ×
e
ncy of 1/2√2 were
r
mance. Smaller or larger
n
erated poor performance.
a
nd figure 7 suggests that
l
eCodes generated using
far superior as compared
d
encoding.
E
rror Rate
EigenKnuckles Fisherknuckles
13.92% 12.66%
1.24 1.27%
e from the experiments.
Figure 7: The CMC curves from the experiments.
The experimental results from the LRT using 8 directions
(D = 8) for various parameters, i.e. (figure 3, equation 2)
and l (summation length, figure 3), are illustrated in table 2. In
order to clearly define the centre of the LRT parity of line
width (width of ) and length (l) must be the same, i.e. both
are even or odd, so while the line width is equal to 1 and 3 line
length 13 and 15 are used. Table 3 illustrates the comparative
experimental results from the KnuckleCodes generated using
LRT and even Gabor filters. The number of even Gabor filters
at every pixel defines the extraction of possible orientation
information (directions D) that can be acquired for every pixel
in the knuckle image. The experimental results in table 3
suggests that the best results are achieved when the total
number of even Gabor filters are fixed to 12. However, little
better results are obtained when the total number of directions
are fixed to 8 (as compared to 12), for the case when
KnuckleCodes are generated using LRT.
A. Discussion
It may also be noted that the generation of KnuckleCodes
using Gabor filters is highly computationally demanding as it
requires convolution operation at every pixel and orientation
as compared to simple sum in LRT. Therefore, the
KnuckleCodes generated from LRT are also favourably
suitable for online user identification. It should be noted that
reference [8] simultaneously employs hand geometry features
while reference [9] [10] employed palmside finger/palm
features. Therefore any direct comparison of our results, that
employed only middle finger knuckle images, with [8]-[10] is
difficult. The accuracy of segmenting knuckle images from
the presented fingers highly influences the matching scores
between the corresponding KnuckleCodes. In order to handle
the rotational and translational variations in the segmented
knuckle images, we employed minimum of matching score
(4) generated from the translation of respective templates in
the region that extended to one third of length and width of
the templates. The finger knuckle image database employed
in this work is being made available [17] to encourage further
research efforts in knuckle biometrics.
. The
equation
effectively
comore from
all the
translated
versions templates.
VII. CONCLUSION
This paper has investigated a new approach for human
identification using finger knuckle images. The orientation of
curved finger knuckle lines and creases are extracted as
template, referred to as KnuckleCodes, and employed for the
human identification. The KnuckleCodes extractied using
LRT achieves the best results as compared to various
approaches investigated in this paper. The advantage of using
Radon transform based KnuckleCodes, lies not only in
significantly higher performance, but also in the template
storage and matching (the size/dimension of KnuckleCodes is
one fourth of the corresponding dimension of knuckle image
or 25 × 20). In summary, the experimental results from the
finger knuckle identification approach investigated in this
paper achieves significantly promising results, i.e., average
rank-one recognition rate of 98.6% and equal error rate of
1.08% on the database of 158 persons. These results can be
attributed to extraction of reliable orientation features using
LRT, usage of robust knuckle image enhancement technique,
and importantly to the generation of reliable matching
distances using equation (4)-(7) that can also account for the
translation of finger knuckles.
The orientation of palm-lines and creases has been
exploited for the personal identification using Gabor
functions [15] and Radon transform based features [12].
However, the finger knuckles present more attractive
Equal Error Rate
1 2 3 4
l 13 15 14 13 15 14
Mean (%) 1.15 1.15 1.08 2.78 2.53 6.96
Std deviation (%) 1.57 1.57 1.08 0.96 1.48 0.9
Table 2: Performance analysis using Localized Radon Transform.
Equal Error Rate
D (Intervals in 0-π ) 6 8 10 12 14 16
KnuckleCodes (Radon) Mean (%) 2.03 1.08 1.29 1.14 1.27 1.29
Std deviation (%) 1.37 1.08 1.59 1.37 1.60 1.24
KnuckleCodes (Gabor) Mean (%) 4.18 11.14 5.82 2.66 3.29 7.59
Std deviation (%) 2.31 1.88 1.04 1.81 2.26 2.24
Table 3: Comparative performance analysis with the variations in D.
alternative than palmprints, mainly because of its smaller
surface area, think/rich lines and creases, and importantly due
to large possibility of higher user acceptance. The user
acceptance for employing finger knuckle in human
identification is expected to be very high as there is no stigma
of personal information (fortune telling beliefs, i.e. life-line,
heart-line, head-line, etc.) associated with finger knuckle
lines/creases.
This paper has offered promising and computationally
smaller alternative for the finger knuckle identification.
However, much more needs to be done and investigated as
there exits lot of potential from finger knuckles for the human
identification. There have not been any studies to ascertain
the stability of finger knuckles with age, time, varying
medical and environmental conditions. Further work is also
required to ascertain the individuality of finger knuckles in
large population (say more than 1000 subjects) and on the
dermatoglyphics analysis to ascertain the variation of knuckle
patterns across the population.
REFERENCES
[1] A. Kumar, “Incorporating cohort information for reliable palmprint
authentication,” Proc. ICVGIP, Bhubaneswar, India, pp. 583–590, Dec.
2008.
[2] V. Kanhangad, A. Kumar, D. Zhang, “Combining 2D and 3D hand
geometry features for biometric Verification,” Proc. CVPR’09
Biometrics Workshop, Miami, Florida, Jun. 2009.
[3] J.-G. Wang, W.-Y. Yau, A. Suwandy, and E. Sung, “Person recognition
by fusing palmprint and palm vein images based on “Lapacianpalm”
representation,” Pattern Recognition, vol. 41, pp. 1531–1544, 2008.
[4] L. Wang, G. Leedham and Siu-Yeung Cho, “Minutiae Feature Analysis
for Infrared Hand Vein Pattern Biometrics,” Pattern Recognition, 41
(3), pp. 920-929, 2008.
[5] A. K. Jain and M. Demirkus, “On latent palmprint matching, MSU
Technical Report, May 2008
[6] D. L. Woodard and P. J. Flynn, “Finger surface as a biometric
identifier”, Computer Vision and Image Understanding, vol. 100, no. 3,
pp. 357-384, Dec. 2005.
[7] S. Malassiotis, N. Aifanti, and M. G. Strintzis, “Personal authentication
using 3-D finger geometry,” IEEE Trans. Inf. Forensics Security, vol. 1,
no. 1, pp. 12–21, Mar. 2006.
[8] 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
[9] S. Ribaric, and I. Fratric, “A Biometric identification system based on
eigenpalm and eigenfinger Features”, IEEE Trans. Pattern Anal. Mach.
Intell., vol. 27, no. 11, Nov. 2005.
[10] Y. Hao, T. Tan, Z. Sun and Y. Han, “Identity verification using
handprint,” Proc. ICB 2007, Lecture Notes Springer, vol. 4642, pp.
328-337, 2007.
[11] R. W. Young, “Evolution of the human hand: the role of throwing and
clubbing,” J. Anat., 202(1), pp. 165-174, Jan. 2003.
[12] W. Jia, D.-S. Huang, and D. Zhang, “Palmprint verification based on
robust line orientation code,” Pattern Recognition, vol. 41, pp.
1504-1513, 2008.
[13] A.K. Jain, S. Prabhakar, L. Hong, and S. Pankanti, “FingerCode: A
filterbank for fingerprint representation and matching,” Proc. CVPR,
vol. 2, pp. 187-193, 1999.
[14] J. Daugman, “Probing the uniqueness and randomness of IrisCodes:
Results from 200 billion iris pair comparisons,” Proc. IEEE, vol. 94,
no. 11, pp 1927-1935, 2006.
[15] A. W. K. Kong and D. Zhang, “Competitive coding scheme for
palmprint verification”, Proc. 17th ICPR, Washington, DC, pp.
1051-4651, 2004.
[16] V. Onana, E. Trouvé, G. Mauris, J.-P. Rudant, and E. Tonyé,
“Detection of linear features in synthetic-aperture radar images by use
of localized radon transform,” Applied Optics, vol. 43, no. 2. Jan. 2004.
[17] http://web.iitd.ac.in/~biometrics/knuckle/iitd_knuckle.htm
