1

Abstract

Palmprint verification is a relatively new but promising
personal authentication technique for its high accuracy
and fast matching speed. Two dimensional (2D) palmprint
recognition has been well studied in the past decade, and
recently three dimensional (3D) palmprint recognition
techniques were also proposed. The 2D and 3D palmprint
data can be captured simultaneously and they provide
different and complementary information. 3D palmprint
contains the depth information of the palm surface, while
2D palmprint contains plenty of textures. How to efficiently
extract and fuse the 2D and 3D palmprint features to
improve the recognition performance is a critical issue for
practical palmprint systems. In this paper, an efficient joint
2D and 3D palmprint matching scheme is proposed. The
principal line features and palm shape features are
extracted and used to accurately align the palmprint, and a
couple of matching rules are defined to efficiently use the
2D and 3D features for recognition. The experiments on a
2D+3D palmprint database which contains 8000 samples
show that the proposed scheme can greatly improve the
performance of palmprint verification.

1. Introduction
Accurate and robust personal authentication is a crucial
issue in the modern e-world, and biometric techniques are
promising solutions to this problem. Palmprint is a
relatively new but important member of biometric
characteristics. It contains a rich amount of stable texture
features, which lead to very high recognition accuracy.
Meanwhile, the merits such as low-cost, user friendliness
and high matching speed make it practicable to use in a
large scale. Many methods have been proposed for 2D
palmprint recognition [1-17]. In general, the current 2D
palmprint recognition methods can be classified into three
categories: coding-based methods [3,4,6,8,13], line-based
methods [10,12] and appearance-based methods [14-17].
Coding-based methods are the most influential ones in
palmprint recognition, and the representative algorithms
include PalmCode [3], Competitive Code [4], Ordinal Code
[6], Fusion Code [8], Robust Line Orientation Code [13],
etc. This kind of methods uses a bank of filters to enhance
and extract the phase and/or orientation features of
palmprint and then encode these features into binary codes
for fast matching. Line-based methods use some line or
edge detectors to explicitly extract the line information
from the palmprint and then use them for matching. The
representative algorithms include the Derivative of
Gaussian based line extraction [10], modified finite Radon
transform (MFRAT) based line extraction [12], etc.
Appearance-based palmprint recognition methods include
the PCA, LDA, and ICA based algorithms, etc [14-17].
Such methods do not use the prior knowledge of the object
but they need a training process.
Although 2D palmprint recognition has achieved high
accuracy, it still has some drawbacks. For example, 2D
palmprint is easy to be counterfeited, and the image quality
can be much affected by illumination changes and
scrabbling. To overcome these problems of 2D palmprint
recognition, 3D palmprint recognition techniques have
been recently developed [18], where the 3D palmprint data
are captured by using structured-light imaging [20]. In the
3D palmprint system [18], the depth information of the
palm inner surface is collected, and a curvature-based
method is used for 3D palmprint feature extraction and
matching. Since the 2D and 3D palmprint images can be
captured simultaneously, the 2D and 3D palmprint features
can be readily fused. In [18] and [19], some straightforward
fusing methods were used and better results than using 2D
or 3D features only were obtained.
This paper aims to develop an efficient 2D and 3D
palmprint matching scheme for more accurate and robust
personal authentication. To better use the 2D and 3D
palmprint information, we propose a couple of novel
feature matching rules. We classify the palmprint features
into three different levels: shape level, line level and texture

Efficient Joint 2D and 3D Palmprint Matching with Alignment Refinement

Wei Li
1, 2
, Lei Zhang
1
, David Zhang
1
, Guangming Lu
3
, Jingqi Yan
2

1
Biometrics Research Center, Dept. of Computing,
The Hong Kong Polytechnic University, Hong Kong, China
2
Institute of Image Processing and Pattern Recognition,
Shanghai Jiao Tong University, Shanghai, China
3
Biocomputing Research Center, Shenzhen Graduate School,
Harbin Institute of Technology, Shen Zhen, China
{cswli, cslzhang, csdzhang}@comp.polyu.edu.hk


2
level. Shape features are contained in 3D palmprint and
they are at the coarsest level. Principal line features exist in
both 2D and 3D palmprint images and they are at the
second level, which represent the structural information of
palmprint principal lines. Texture features are also
contained in both 2D and 3D palmprint images and they are
at the finest level, which represent the detailed features of
palmprint. Intuitively, the texture information is well suited
for palmprint discrimination, while the shape and line
information can be used for palmprint alignment, which is
an important step for robust palmprint recognition. Most of
the existing palmprint recognition methods assume that the
palmprint images have been well aligned before
performing feature extraction and matching. However, the
palm is not a rigid object and there are no sharp corner
points on it, which makes the accurate segmentation of ROI
(region of interest) very hard. In this paper, we will use the
shape and line features to effectively refine the alignment,
and propose a novel matching scheme to efficiently use the
three levels of features for accurate palmprint verification.
The rest of this paper is organized as follows. Section 2
introduces the three-level feature extraction of 2D and 3D
palmprint. Section 3 describes the alignment refinement of
texture feature maps. Section 4 presents the joint 2D and
3D feature matching scheme. Section 5 presents the
experimental results and Section 6 concludes the paper.
2. The Three Level Feature Extraction of 2D
and 3D Palmprint
2.1. Preprocessing


(a) (b)

(c) (d)
Figure 1. (a) The 2D palmprint image, the established coordinate
system and the ROI (the rectangle); (b) the extracted 2D ROI; (c)
the 3D palmprint image, whose cloud points have a one-to-one
correspondence to the pixels in its 2D counterpart; (d) the
obtained 3D ROI by grouping the cloud points corresponding to
the pixels in 2D ROI.

In data collection, the 3D palmprint and its associated 2D
palmprint images are captured simultaneously [18]. The
algorithm in [3] is used to extract the 2D ROI, and the 3D
ROI is then obtained by grouping the cloud points in
correspondence to the 2D ROI pixels, as shown in Fig. 1.
The ROI extraction process can align the palmprint
images to some extent. However, there are still some
translation and rotation in the images, and this is the main
reason for false rejection and recognition. Therefore, we
will propose an alignment refinement method in Section 3.
For 3D ROI, the curvature value is much more stable
than the original depth information [18]. We calculate the
mean curvature H of the 3D ROI as follows:
2 / 3 2 2
2 2
) ) ( ) ( 1 ( 2
) ) ( 1 ( 2 ) ) ( 1 (
y x
yy x xy y x xx y
h h
h h h h h h h
H
+ +
+ + +
=
(1)
where h is the height of the points in the palmprint w.r.t the
reference plane, h
x
, h
y
, h
xx
, h
yy
and h
xy
are the first, second
and hybrid partial derivatives of h. We then normalize the
mean curvature values to generate the mean curvature
image (MCI). Fig. 2 shows some examples of MCI.



Figure 2. 3D ROI (1
st
row) and the corresponding MCI (2
nd
row).
The left two columns are collected from one palm, while the right
two columns are collected from another palm.
2.2. 1
st
Level: Shape Feature Extraction
Shape features are contained in 3D palmprint, and they
are at the coarsest level of palmprint features. To remove
the noise and some trivial structures in the 3D data, we use
a 77 median filter and a 77 mean filter to smooth the
surface. Fig. 3 gives an example of the surface of 3D ROI
before and after filtering.


Figure 3. An example of shape surface of 3D ROI. The left one is
before filtering and the right one is after filtering.

After filtering, we down-sample the 3D ROI from 256
256 to 3232 to represent the shape surface. Then we use a
regular triangle mesh to describe the shape features, as
shown in Fig. 4. The structure of the triangle mesh can
greatly reduce the time consumption in the proposed
palmprint alignment refinement process (refer to Section 3)

3
by using the ICP (Iterative Closest Point) [22] algorithm.


Figure 4. Shape features described by regular triangle mesh. The
right image is the zoom-in of the up-right corner of the left mesh.
2.3. 2
nd
Level: Principal Line Feature Extraction




Figure 5. Examples of 2D ROI (1
st
column) and 3D MCI (3
rd

column) and the extracted principal line features of them (2
nd
and
4
th
columns respectively). The 1
st
row shows an example which
has clear principal lines on 2D ROI but not on 3D MCI; and the
2
nd
row shows an example which has clear principal lines on 3D
MCI but not on 2D ROI.

Line features exist in both 2D and 3D palmprint images.
For 2D palmmprint, the line features can be extracted from
the 2D ROI directly. For 3D palmprint, we extract the line
feature from the MCI image of the 3D ROI. Here we focus
on principal lines in the palm and ignore most of the
wrinkle lines. We adopt Huangs method [12], which is
based on a modified Radon Transform (RT) [23], for
principal line feature extraction.
Most of palmprint samples have clear principal line
features on both 2D ROI and 3D MCI images. However,
some samples have clear principal lines only on 2D ROI or
3D MCI. Some examples are shown in Fig. 5. In the worst
case, some samples do not have clear principal lines in
either 2D ROI or 3D MCI. However, they still have shape
information, which can be used for alignment refinement.
This is why we use the 1
st
level shape feature and 2
nd

principal line feature to do alignment refinement.
2.4. 3
rd
Level: Texture Feature Extraction
Texture features are the most discriminative features for
palmprint recognition. Since palmprint is full of line-like
features, the orientation coding technique is widely use for
texture feature extraction of palmprint. The representative
methods include Competitive Code [4], Ordinal Code [6]
and Robust Line Orientation Code [13]. We can extract the
orientation features from both 2D ROI and the 3D MCI. In
this paper, the Competitive Code method [4] is used. In
Competitive Code, the directional Gabor filter [21] is used
for orientation extraction
) (
2
) , , , (
2
) 4 (
8
2
2 2
2
2

+
= e e e y x
x i
y x
(2)
where
sin ) ( cos ) (
0 0
y y x x x + =
cos ) ( sin ) (
0 0
y y x x y + =

and (x
0
, y
0
) is the center of the Gabor function; is the
radial frequency in radians per unit length and is the
orientation of the Gabor functions in radians; and is
2 1
2ln 2
2 1

+
=

(3)
with being the half-amplitude bandwidth of the
frequency response.


(a) (b)
Figure 6. Example of orientation features in 2D ROI and 3D MCI.

Considering the accuracy and efficiency, we use six
orientation = 0, /6, 2/6, 3/6, 4/6, 5/6. Convolving
the six templates with 2D ROI and 3D MCI respectively,
and selecting the orientation which leads to the greatest
response, we get the orientation features of 2D ROI and 3D
MCI as shown in Fig. 6, from which we can see that the
extracted orientation can well represent the line direction in
a neighborhood.
3. Alignment Refinement of Feature Maps
As mentioned in the Introduction section, the texture
features (specifically, orientation features in this paper) are
well suited for palmprint discrimination, while the shape
and line features can be used for palmprint alignment. As
illustrated in Fig. 7, we estimate the translation and rotation
between two palmprint samples by using the ICP [22]
method according to 2D principal lines, 3D MCI principal
lines and 3D shape information, respectively. Three pairs of
parameters of translation T and rotation R will then be
obtained and used to align the texture feature maps for
palmprint matching. This is actually a palmprint alignment
refinement process.

4

Figure 7. Translation and rotation estimation between two
samples by ICP method.

For each point in the texture feature map, denote by
the orientation, and by (x, y) the location. Let
R=
co s s i n
s i n c o s
R R
R R





be the rotation matrix and
] [
y x
t t T = be the translation vector computed by ICP.
The corrected orientation and location ) , ( y x are
calculated as follows:
R
+ =
(4)
T
y
x
R
y
x
+

(5)
The corrected direction is then normalized to one of
the six orientations {
6 / 5 , 6 / 4 , 6 / 3 , 6 / 2 , 6 / , 0
}
using the nearest principle, and then we code them with
integer values 0, 1, 2, 3, 4 and 5, respectively. We
intuitively define the distance between parallel directions
as 0, the distance between perpendicular directions as 3, the
distance as 1 when the angle of the two directions is 6 /
or 6 / 5 , and the distance as 2 when the angle of the two
directions is 6 / 2 or 6 / 4 . Let D
d
and D
t
be the two
coded texture feature maps of two palmprint ROI images,
the distance between them can be defined as:
( )
1 1
1
( , ), ( , )
3
n m
D d t
i j
S F D i j D i j
nm
= =
=

(6)
where
( ) = 6 , min ) , ( F
, { } 5 , 4 , 3 , 2 , 1 , 0 ,
(7)
The smaller the distance S
D
is, the more similar the two
samples are. If the distance S
D
is less than a threshold, the
two palmprint images are recognized to be from the same
person. Note that each pairs of parameters T and R will lead
to a matching distance, and hence three distances will be
obtained. How to use them for decision making will be
discussed in next section.
4. Joint 2D and 3D Feature Matching
In Section 2.4, we extracted two types of orientation
features: one for 2D ROI and one for 3D MCI. In Section 3,
we estimated three pairs of translation and rotation
parameters. There are six possible combinations for the
alignment refinement of texture feature maps. Actually, it is
unnecessary to try all the combinations. Here we propose a
couple of matching rules for efficient feature matching.

0 0.5 1 1.5
0
2
4
6
Distance
F
r
e
q
u
e
n
c
e

%
Genuine
Impostor
t1 t2
Confident interal
Unconfid
-ent inte
-ral
Confid
-ent in
-terval

Figure 8. Confident and unconfident intervals in a typical
palmprint matching distance distribution.

Fig. 8 shows a typical matching distance distribution of
palmprint verification. We can divide the distance axis into
two parts. The unconfident interval is between threshold t
1

and t
2
, and the remaining part is the confident interval, as
illustrated in Fig. 8. If the matching distance falls into the
confident interval, we can easily and correctly decide
whether the test sample is a genuine or an imposter, and
accept this score as the final matching score. Thus the first
rule to end the matching process is:
Rule 1. If the matching distance falls into the confident
interval, end the matching process and output the result.
The feature map alignment refinement will reduce the
distance between two matching samples but for both
genuine and imposter matchings. Reducing the distance of
genuine matching is good for verification, but reducing the
distance of imposter matching is not what we want.
However, statistically the alignment refinement will bring
more benefit to genuine samples than imposter samples.
This can be supported by our experiments on a palmprint
database which contains 1000 samples collected from 100
palms. Fig. 9 shows the distribution of the drops of
ICP
Shifting: T
Rotation: R
Before ICP
After ICP
(a) Calculating the translation and rotation according to 2D principal palm lines.
ICP
Before ICP
After ICP
Shifting: T
Rotation: R
(b) Calculating the translation and rotation according to MCI principal palm
lines.
Before ICP
After ICP
ICP
Shifting: T
Rotation: R
(c) Calculating the translation and rotation according to shape information.
0 50 100
0
50
100
0 50 100
0
50
100
0 50 100
0
50
100
0 50 100
0
50
100

5
matching distances before and after alignment refinement
on the training database. From Fig. 9, we can see that most
of the decrease of imposter matching distance is close to 0,
while the decrease of genuine matching distance distributes
much more evenly. This implies that the alignment
refinement bring more benefit to genuine samples than
imposter samples.

0 0.1 0.2 0.3 0.4
0
10
20
30
Decreasing value of distance
F
r
e
q
u
e
n
c
y

%Genuine
Impostor

Figure 9. Distribution of the decrease of matching distances
before and after alignment refinement.

0 0.5 1
0
10
20
30
Decreasing scale
F
r
e
q
u
e
n
c
y

%Genuine
Impostor

Figure 10. Distribution of DS after alignment refinement.

To better classify the genuine and imposter, we define
the decreasing scale (DS) of matching distance as
( ) /( ) DS S S S C =
(8)
where S and S are the matching distances before and after
alignment refinement, and C is a constant. Here, we let C =
0.9. Fig. 10 shows the distribution of DS. We can see that
the genuine and imposter can be better separated by using
DS. Hence we define the following 2
nd
rule to end the
matching process:
Rule 2. If DS is greater than a preset threshold T
d
, end the
matching process and output the result.
By using the above two rules and the alignment
refinement, we propose a novel joint 2D and 3D palmprint
matching scheme, which is illustrated in Fig. 11. For each
pair of matching samples, we first compute the matching
score S
2D
by applying Competitive Code [4] to 2D ROI. If
the rule 1 is satisfied (i.e. the matching distance falls into
confident interval), then we take S
2D
as the final matching
distance. Otherwise, we apply Competitive Code to the 3D
MCI. If the matching distance by 3D MCI still falls into the
unconfident interval, we will use the extracted 2D principal
lines, 3D MCI principal lines and the 3D shape information
to do palmprint alignment refinement, and then perform
texture feature matching to check if rule 1 or rule 2 is fired.
If the two rules are still not satisfied, we simply accept the
score S
2D
as the final matching distance. With the proposed
matching strategy, most of the classification decisions will
be made in the first one or two steps, which can greatly save
the matching time. The alignment refinement process
makes the palmprint recognition system robust to
translation and rotation variations. After performing
experiments on the training palmprint database, we select t
1

=0.9 and t
2
=1.3 for rule 1, and T
d
= 0.35 for rule 2 by
experience.
5. Experimental Results
The joint 2D and 3D palmprint database used in this
paper is obtained via [25]. The database contains 8000
samples from 200 volunteers, including 136 males and 64
females. The youngest one is 10 years old and the oldest
one is 55 years old. A 3D palmprint image and the
corresponding 2D palmprint image were collected for each
subject in two separated sessions, and in each session 10
samples were collected from both the left and right hands of
each volunteer. The average time interval between the two
sessions is one month. The original spatial resolution of the
data is 768576. After ROI extraction, the central part
(256256) is used for feature extraction and recognition.
The experiments were performed on a PC with Core 2 CPU
@ 2.66GHz with 2GB RAM.
In the joint 2D and 3D palmprint verification, the class of
the input palmprint is known and each of the samples was
matched with all the other samples in the database. A
successful matching is called intra-class matching or
genuine if the two samples are from the same class.
Otherwise, the unsuccessful matching is called inter-class
matching or impostor. Using the HK-PolyU database [25],
there are 31,996,000 matchings in total. The verification
experiments were performed by using the proposed method
in comparison with the benchmark score level fusion
method which is called Matcher Weighting (MW) [24].
The EER values are listed in Table 1, where other relative
results are also listed. From Table 1, we can see that the
alignment refinement can improve much the verification
accuracy, while the proposed joint 2D and 3D matching
scheme can enhance greatly the performance. Please note
that the joint 2D and 3D fusion with the classical MW
method can only achieve a little improvement compared
with using the 2D feature alone. The reason may be that the
discrimination of 2D features and 3D MCI features are at
different levels, and the correlation between 2D image and
3D MCI also makes the MW method less effective. Fig. 12
shows the ROC curves by the different methods.
Table 2 lists the matching time by 2D ROI, 3D MCI, 2D
ROI and 3D MCI fusion by MW and the proposed method.
We can see that although the maximum time per matching
of the proposed method is longer than others, the average
matching time of it is even shorter than the classical MW
fusion method. This is because in the proposed joint 2D and
3D palmprint matching, most of the matching ends at the
first one or two steps, and only very few matchings go
through all of the steps in Fig. 11.

6

Table 1. The EER by different methods.
Methods EER
2D ROI 0.046%
3D MCI 0.294%
2D ROI with alignment refinement 0.033%
3D MCI with alignment refinement 0.248%
2D and 3D fusion by MW 0.045%
Proposed joint 2D and 3D matching 0.025%

Table 2. The minimum, maximum and average time per matching
by different methods.
Minimum Maximum Average
2D ROI 0.15ms 0.15ms 0.15ms
3D MCI 0.15ms 0.15ms 0.15ms
2D+3D fusion by MW 0.30ms 0.30ms 0.30ms
Proposed 0.15ms 3.60ms 0.28ms

10
-6
10
-4
10
-2
10
0
10
2
88
90
92
94
96
98
100
False Acceptance Rate
G
e
n
u
i
n
e

A
c
c
e
p
t
a
n
c
e

R
a
t
e
Proposed 2D and 3D fusion method
2D with alignment refinement
2D and 3D MCI fusion by MW
2D
3D MCI with alignment refinement
3D MCI

Figure 12. The ROC curves by different methods.
6. Conclusion
In this paper, we proposed a novel scheme to jointly use
palmprint 2D and 3D features for personal authentication.
We extracted three levels of 2D and 3D palmprint features:
shape features, principal line features and texture features.
By using the ICP method, we performed alignment
refinement to the texture feature map according to the
principal line features and shape features when necessary.
The alignment refinement reduces greatly the translation
and rotation variations introduced in the palmprint data
acquisition process. A couple of feature matching rules and
an efficient joint 2D and 3D feature matching scheme were
then proposed to fully use the 2D and 3D palmprint
information. The experimental results on the 2D+3D
palmprint database, which contains 8000 samples collected
from 200 volunteers, demonstrated that the proposed
method increases significantly the palmprint verification
accuracy. Meanwhile, its average matching time is even
less that by the classical weighting average fusing method.
Acknowledgement
The work is partially supported by the GRF fund from
the HKSAR Government, the central fund from Hong
Kong Polytechnic University. The NSFC
funds(60803090), the Natural Scientific Research
Innovation Foundation in Harbin Institute of Technology,
and the Key Laboratory of Network Oriented Intelligent
Computation, Shenzhen, China.
References
[1] W. Shu and D Zhang, Automated Personal Identification by
Palmprint, Optical Eng., vol. 37, no. 8, pp. 2659-2362,
1998.
[2] A. K. Jain and J. J. Feng, Latent Palmprint Matching, IEEE
Transactions on Pattern Analysis and Machine Intelligences,
vol. 31, Issue 6, pp. 1032 - 1047, June 2009.
[3] D. Zhang, A.W.K. Kong, J. You, and M. Wong, On-line
palmprint identification, IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 25, no. 9, pp.
1041-1050, 2003.
[4] A.W.K. Kong and D. Zhang, Competitive coding scheme
for palmprint verification, Proceedings of International
Conference on Pattern Recognition, vol. 1, pp. 520-523,
2004.
[5] Y.H. Pang, T. Connie, A. Jin and D. Ling, Palmprint
authentication with Zernike moment invariants,
Proceedings of the 3rd IEEE International Symposium on
Signal Processing and Information Technology, pp. 199-202,
2003.
[6] Z.N. Sun, T.N. Tan, Y. H. Wang, and S. Z. Li. Ordinal
palmprint representation for personal identification,
Proceeding of IEEE International Conference on Computer
Vision and Pattern Recognition, pp. 279-284, 2005.
[7] T. Connie, A. T. B. Jin, M. G. K. Ong, and D. N. C. Ling,
An automated palmprint recognition system, Image and
Vision Computing, vol. 23, Issue 5, pp. 501-515, 2005.
[8] A. Kong, D. Zhang, and M. Kamel, Palmprint identification
using feature-level fusion, Pattern Recognition, vol. 39, no.
3, pp. 478487, 2006.
[9] C. Poon, D.C.M. Wong and H.C. Shen, Personal
identification and verification: fusion of palmprint
representations, in Proceedings of International Conference
on Biometric Authentication, pp. 782-788, 2004.
[10] X.Q. Wu, K.Q. Wang, and D. Zhang, "Palmprint texture
analysis using derivative of Gaussian filters," International
Conference on Computational Intelligence and Security, vol.
1, pp. 751-754, 2006.
[11] M. G. K. Ong, T. Connie, and T. A. B. Jin, Touch-less palm
print biometrics: Novel design and implementation, Image
and Vision Computing, vol. 26, Issue 12, pp. 1551-1560,
December 2008.
[12] D. S. Huang, W. Jia, and D. Zhang, Palmprint verification
based on principal lines, Pattern Recognition, vol. 41, Issue
4, pp. 1316-1328, April 2008.
[13] W. Jia, D. S. Huang, and D. Zhang, Palmprint verification
based on robust line orientation code, Pattern Recognition,
vol. 41, Issue 5, pp. 1504-1513, May 2008.

7

Figure 11. Flow chart of joint 2D and 3D feature matching.


[14] Q. Li, Z. Qiu and D. Sun, Feature-level fusion of hand
biometrics for personal verification based on Kernel PCA,
International Conference on Biometrics, pp. 744-750, 2006.
[15] S. Ribaric, I. Fratric and K. Kis, A biometric verification
system based on the fusion of palmprint and face features,
Proceeding of the 4th International Symposium on Image,
Signal and Signal Processing and Analysis, pp. 15-17, 2005.
[16] L. Shang, D.S. Huang, J.X. Du and C.H. Zheng, Palmprint
recognition using FastICA algorithm and radial basis
probabilistic neural network, Neurocomputing, vol. 69, no.
13-15, pp. 1782-1786, 2006.
[17] S. Ribaric and I. Fratric, A biometric identification system
based on Eigenpalm and Eigenfinger features, IEEE
Transactions on Pattern Analysis and Machine Intelligence,
vol. 27, no. 11, pp. 1698-1709, 2005.
[18] D. Zhang, G. Lu, W. Li, L. Zhang, and N. Luo, "Palmprint
Recognition Using 3-D Information," IEEE Transactions on
Systems, Man, and Cybernetics, Part C: Applications and
Reviews, Volume 39, Issue 5, pp. 505 - 519, Sept. 2009.
[19] D. Zhang, V. Kanhangad, N. Luo, and A. Kumar, Robust
palmprint verification using 2D and 3D features, Pattern
Recognition, Volume 43, Issue 1, pp. 358-368, 2010.
[20] V. Srinivassan and H.C. Liu, Automated phase measuring
profilometry of 3D diffuse object, Appl. Opt., vol. 23, no.
18, pp. 3105-3108, 1984.
[21] T.S. Lee, Image representation using 2D Gabor wavelet,
IEEE Transactions on Pattern Analysis and Machine
Intelligence, vol. 18, no. 10, pp. 959-971, 1996.
[22] P. J. Besl and N.D. McKay, A method for registration of
3-D shapes, IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 14, no. 2, pp. 239 - 256, Feb.
1992.
[23] J. Radon, ber die bestimmung von funktionen durch ihre
integralwerte lngs gewisser mannigfaltigkeiten, Berichte
Schsische Akademie der Wissenschafter, Leipzig,
Math.-Phys. Kl (69), 262267, 1917.
[24] R. Snelick, U. Uludag, A. Mink, M. Indovina, and A. Jain,
Large-scale evaluation of multimodal biometric
authentication using state-of-the-art systems, IEEE
Transactions on Pattern Analysis and Machine Intelligence,
vol. 27, no. 3, pp. 450- 455, March 2005.
[25] HK-PolyU 2D+3D palmprint database: http://www.comp.
polyu.edu.hk/~biometrics/2D_3D_Palmprint.htm.