Content uploaded by Hongyu Li
Author content
All content in this area was uploaded by Hongyu Li on Jul 28, 2015
Content may be subject to copyright.
IEEE SIGNAL PROCESSING LETTERS, VOL. 19, NO. 10, OCTOBER 2012 663
Fragile Bits in Palmprint Recognition
Lin Zhang, Member, IEEE, Hongyu Li, and Junyu Niu
Abstract—Recent years have witnessed a growing interest in de-
veloping automatic palmprint recognition methods. Among them,
coding-based ones, representing the texture of a palmprint using
a binary code, are most prevalent and successful. We find thatnot
all bits in a code map generated by a specific coding scheme are
equally consistent. A bit is deemed fragile if its value changes across
code maps created from different images of the same palmprint.
In this paper, we first analyze the fragile bits phenomenon in a
state-of-the-art palmprint coding scheme, namely, binary orienta-
tion co-occurrence vector (BOCV). Then, based on our analysis,
we extend BOCV to E-BOCV by incorporating fragile bits infor-
mation in appropriate ways. Experiments conducted on the bench-
mark dataset demonstrate that E-BOCV can achieve the highest
verification accuracy among all the state-of-the-art palmprint ver-
ification methods evaluated. To our knowledge, this is the first work
investigating the fragile bits of coding-based palmprint recognition
approaches.
Index Terms—BOCV, fragile bits, palmprint.
I. INTRODUCTION
BOLSTERED by the requirements of numerous appli-
cations, such as access control, aviation security, or
e-banking, recognizing the identity of a person with high con-
fidence has become a topic of intense study. Biometrics based
methods are drawing more and more attention. As an important
member of the biometrics family, palmprint-based personal au-
thentication systems have been corroborated to have the merits
of high distinctiveness, robustness, high user-friendliness, and
cost effectiveness [1], [2]. In the past decade or so, a plethora
of algorithms have been proposed for palmprint recognition
[3], [4].
Among various palmprint recognition schemes, coding-based
ones are the most prevalent and appealing since in general they
have the advantages of high accuracy, robustness to illumina-
tion variation, and fast feature extraction and matching speed.
In a typical coding-based method, each field of the code map
is assigned a bit-wised code, based on the quantization result
Manuscript received June 11, 2012; revised July 18, 2012; accepted July 26,
2012. Date of publication August 03, 2012; date of current version August 13,
2012. This work was supported by the Fundamental Research Funds for the
Central Universities under Grant 2100219 033. The associate editor coordinating
the review of this manuscript and approving it for publication was Dr. Xiao-Ping
Zhang.
L. Zhang is with the School of Software Engineering, Tongji University,
Shanghai 201804, China (e-mail: cslinzhang@tongji.edu.cn).
H. Li is with the School of Software Engineeri ng, Tongji University, Shanghai
201804, China and also with the Department of Electronic Engineering, Fudan
University, Shanghai 200433, China (e-mail: hyli@tongji.edu.cn).
J. Niu is with the School of Computer Science, Fudan University, Shanghai
200433, China (e-mail: jyniu@fudan.edu.cn).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/LSP.2012.2211589
of the image’s responses to a set of specificfilters. PalmCode
proposed in [1] uses a single Gabor filter to extract the local
phase information. Its computational architecture is the same as
the IrisCode [5]. In [6], Kong and Zhang proposed the compet-
itive code (CompCode) scheme, which encodes the local orien-
tation field of a palmprint using symmetric Gabor filters along
six different orientations. In [7], Jia et al. proposed a different
coding method to extract the local orientation information of
palmprints, namely robust line orientation code (RLOC), which
isbasedonamodified finite Radon transform. In [8], Sun et
al. used differences between two orthogonal Gaussians to ex-
tract the local ordinal measures from palmprints. In [9], Guo et
al. proposed another coding method by binarizing a palmprint
image’s responses to the real Gabor filters along six different
orientations and they named their method as binary orientation
co-occurrence vector (BOCV). BOCV can get a state-of-the-art
verification accuracy so far. Recently, another two palmprint
verification methods were proposed [10], [11] and they could
achieve higher verification accuracies than BOCV; but both of
them have much higher computational complexities for feature
extraction and matching than BOCV.
As stated, a bit at a specificl
ocation of a code map generated
by a coding scheme discussed above is typically binarized from
the image’s response to a filter. Consider multiple images of the
same palmprint. Across all images, responses for that location
will be similar, but not exactly the same. Similarly, the bit from
the binary quantization could be the same across all code maps,
or it may differ in some of the code maps. A bit in a palmprint’s
code map is consistent if it has the same value for most images
of that palmprint; otherwise, it is considered as fragile. Actu-
ally, the concept of fragile bit and its effects on IrisCode have
been well studied in [12]–[14]. However, to our knowledge, no
investigation about fragile bits has been reported for palmprint
recognition in the literature.
In this paper, we attempt to give a thorough study on the
effects of fragile bits for palmprint recognition in the context
of a specificcod
ing scheme—binary orientation co-occurrence
vector (BOCV) [9], which is a state-of-the-art coding method
for palmprint recognition. To this end, at first, a straightforward
method for identifying fragile bits for BOCV is proposed. Then,
we propose to improve the verification performance of BOCV
by masking out the fragile bits when calculating the Hamming
distance. In addition, we believe that fragile bit patterns can
also bringinsomebeneficial information for palmprint recog-
nition. Thus, we propose a new metric fragile-bit pattern dis-
tance (FPD) to quantitatively measure the coincidence of the
fragile-bit patterns in two BOCV code maps. Then, FPD is fused
with the modified Hamming distance to further improve the
verification accuracy of the conventional BOCV scheme. Our
claims are validated by extensive experiments performed on the
PolyU palmprint dataset [15]. This paper focuses only on the
1070-9908/$31.00 © 2012 IEEE
664 IEEE SIGNAL PROCESSING LETTERS, VOL. 19, NO. 10, OCTOBER 2012
Fig. 1. Real Gabor filters along six different orientations.
feature extraction and matching stage of a palmprint recogni-
tion system. For the data preprocessing of palmprint images,
such as the ROI (region of interest) extraction, please refer to
[1] for details.
The remainder of this paper is organized as follows. Section II
presents a method for identifying fragile bits. Section III dis-
cusses how to extend the original BOCV by making use of
fragile bits information. Section IV reports the experimental re-
sults and Section V concludes the paper.
II. FRAGILE BITS IN BOCV
As pointed out in [13], [14], bit fragility generally occurs
when the inner product between a filter and a particular part of
the examined image produces a value with a small magnitude.
Hence, the fragility of each bit in a code map depends on a com-
bination of the palmprint structure at that particular location,
the filter adopted by the coding scheme, and the quantization
method for the filter response. We put our focus on investigating
the effects of fragile bits in the context of a specificcoding
method—BOCV [9], since it can achieve a state-of-the-art ver-
ification accuracy.
The neurophysiology-based Gabor filter proposed by Lee
[16] is used in BOCV. It is defined as
(1)
where , ,
,.In(1), is the radial frequency in
radians per unit length and is the orientation of the Gabor
filter. is defined by ,where is
the half-amplitude bandwidth of the frequency response. can
be determined by ,where is the standard deviation
of the Gaussian envelop. Examples of such real Gabor filters
are shown in Fig. 1 and their parameters are set as ,
.
The computational architecture of BOCV is rather simple and
straightforward. A BOCV code map consists of six bit-planes,
each of which is binarized from the image’s response to a real
Gabor filter along a specific orientation. Such a process is il-
lustrated in Fig. 2. Fig. 2(a) is a palmprint ROI image while
Fig. 2(b) are the 6 corresponding bit-planes binarized from ’s
responses to 6 Gabor filters with 6 different orientations.
For a filter applied to a specific location in a single image,
if the response has a large magnitude, then the corresponding
bit will likely be consistent. On the contrary, if the response is
close to zero, the corresponding bit will be very likely to be
fragile. Based on this analysis, we can use a simple method to
figure out the potential fragile bits in a single BOCV code map
approximately, similar as Hollingsworth et al. did for IrisCode
[14]. Given a palmprint ROI image, suppose that we have ob-
tained its response to the ()Gaborfilter.
By binarizing , we can get the BOCV bit-plane .At
Fig. 2. (a) is a palmprint ROI image ; (b) are the 6 bit-planes binarized from
’s responses to a set of Gabor filters defined by (1) with 6 different orientations.
Fig. 3. (a) and (b) are the different images taken from the same palmprint while
(c) is captured from a different palmprint; (d) is a fragility mask “slice” for (a);
(e) is a fragility mask “slice” for (b); (f) is a fragility mask “slice” for (c). Black
pixels are bits masked as fragile.
the same time, we can sort magnitude values contained in the
matrix to identify percent smallest ones. Then, we regard
the corresponding bits in binarized from these smallest mag-
nitudes as fragile. Fragility mask is stored in a separate matrix
. Consistent bits are represented as ones while fragile bits are
marked as zeros in the fragility mask. Thus, indicates
whether the bit is consistent or not.
Examples of fragility masks are shown in Fig. 3. In Fig. 3, (a)
and (b) are different images captured from the same palmprint
while (c) is an image captured from another different palmprint.
We convolve the images with . Then, for each image, the lo-
cations with the smallest 25% magnitudes are masked as fragile
(black pixels). Fragility masks generated from (a), (b), and (c)
are shown in (d), (e), and (f). It needs to be noted that each image
actually has 6 such fragility masks since BOCV uses 6 Gabor
filters . Then, these 6 mask “slices” form the
final fragility mask matrix.
III. EXTENDED BOCV
In this section, two ways to improve the verification accuracy
of the conventional BOCV scheme will be presented in detail.
A. Fragile Bit Masking
The original BOCV scheme adopts the normalized Hamming
distance to match two BOCV code maps. In the following, we
will use to represent the conventional Hamming distance.
weights all bits in a code map equally. However, actually
not all of the bits in a code map are equally useful. Instead,
fragile bits tend to magnify an intra-class matching distance.
According to this consideration, we propose to mask out fragile
ZHANG et al.: FRAGILE BITS IN PALMPRINT RECOGNITION 665
bits when computing the Hamming distance. With this modifi-
cation, the matching distance in a comparison is based on fewer
bits, but each bit used is more consistent.
The modified Hamming distance can be computed as the fol-
lowing. Suppose that and are two BOCV code maps. Their
fragility mask matrices are and , respectively. Then, the mod-
ified Hamming distance, denoted by ,isdefined as
(2)
where ()isthe bit-plane of (), represents the
bitwise “exclusive OR” operation, and means bitwise “AND”
operation.
B. Fragile-Bit Pattern Distance
From observation we find that the locations of fragile bits tend
to be consistent across different code maps of the same palm-
print while they quite disagree with each other in code maps
from different palmprints as illustrated by examples shown in
Fig. 3. This implies that further useful information could be ex-
tracted from fragile-bit patterns. To this end, we present a new
metric fragile-bit pattern distance (FPD) to quantitatively mea-
sure the dissimilarity of two fragile-bit patterns.
Consider two fragile bit pattern matrices, and . Their FPD
is defined as
(3)
where is the area of the code map. In practice, taking into ac-
count the possible translations in the extracted ROI sub-image
with respect to the one extracted in the enrolment, multiple
matches are performed by translating one set of features in hori-
zontal and vertical directions. And the minimum of the resulting
matching distances is considered to be the final matching dis-
tance. In such cases, is the area of the overlapping parts of the
two code maps.
Even though is not as powerful as , we can com-
bine them together in order to create a better classifier. We adopt
a simple weighted-average fusion rule to fuse and
together as
(4)
where is a parameter to control the contribution
of in the fusion. In the following, we refer the method
using to calculate the matching distance as Extended-BOCV,
or E-BOCV for short.
IV. EXPERIMENTAL RESULTS AND DISCUSSIONS
A. Dataset, Test Protocol, and Parameter Settings
Experiments were conducted on the benchmark PolyU palm-
print dataset [15], which contains 7,752 images captured from
384 different palms. In that dataset, sample images for each sub-
ject were collected in two sessions.
TAB L E I
VERIFICATION PERFORMANCE BY USING AND
IN THE CONTEXT OF BOCV
In our experiments, we took images collected in the first ses-
sion as the gallery set and images collected at the second session
as the probe set. Under such experimental settings, the gallery
set contained 3,889 images while the probe set contained 3,863
images. To obtain statistical results, each image in the probe set
was matched with all the images in the gallery set. If the two
images were from the same class, the matching between them
was counted as a genuine matching; otherwise it was counted
as an imposter matching. The equal error rate (EER), which is
the point where the false accept rate (FAR) is equal to the false
reject rate (FRR), is used to evaluate the verification accuracy.
Besides, by adjusting the matching threshold, a detection error
tradeoff (DET) curve, which is a plot of FRR against FAR for
all possible thresholds, can be created. Thus, the DET curves
obtained by methods evaluated will also be provided.
Parameters involved were tuned based on a sub-dataset
which contained the first 192 palmprints of the PolyU palm-
print dataset and the tuning criterion was that parameter values
that could lead to a lower EER would be chosen. As a result, the
parameters used in this paper were set as: , ,
,and .
Considering the imperfectness of the ROI extraction step,
we shifted the code maps vertically and horizontally in a small
range when matching. The minimal distance obtained by shift
matching was taken as the final matching distance. The shift
range was set as [ , 4] in the following experiments.
B. Effectiveness of Fragile Bit Masking
In this experiment, we compared the verification accuracies
obtained by using and respectively in the context of
BOCV. Their EERs are listed in Table I. From the results we can
see that by masking out fragile bits in code maps the EER could
be reduced from 0.03677% to 0.03365%. The drop of EER is
8.49% ( ). It demonstrates that the
palmprint verification accuracy could be improved by masking
out fragile bits.
C. Performance Evaluation of E-BOCV
Compared with the original BOCV scheme, the novelty of
E-BOCV lies in the matching method it adopts. Specifically,
it fuses the modified Hamming distance and the fragile-bit
pattern distance to compute the dissimilarity of two code maps.
In this experiment, its verification performance was evaluated
and compared with the other four state-of-the-art coding based
palmprint verification methods, including CompCode [6],
RLOC [7], OrdinalCode [8], and BOCV [9]. The results in
terms of EER are summarized in Table II. Fig. 4(a) shows the
DET curves generated by the five different palmprint verifica-
tion methods. Distance distributions of genuine matchings and
imposter matchings obtained by the proposed E-BOCV scheme
are plotted in Fig. 4(b).
666 IEEE SIGNAL PROCESSING LETTERS, VOL. 19, NO. 10, OCTOBER 2012
Fig. 4. (a) DET curves obtained by using various palmprint verification
methods; (b) distance distributions of genuine matchings and imposter match-
ings obtained by E-BOCV.
TAB L E I I
VERIFICATION PERFORMANCE OF DIFFERENT SCHEMES
From the results listed in Table II and the DET curves shown
in Fig. 4(a), we can see that E-BOCV performs the best in terms
of the verification accuracy among all the state-of-the-art palm-
print verification methods evaluated. With our experimental
settings, the EER achieved by the proposed E-BOCV scheme
was 0.0316%. Compared with the conventional BOCV method,
the drop of EER is 14.06% ( ),
which is quite significant. Therefore, the experimental results
clearly corroborated our claim that by appropriately making
use of fragile bits information, the performance of a palmprint
coding scheme could be largely improved.
V. C ONCLUSIONS
This paper is the first work discussing the fragility of bits
in a palmprint code map. We investigated the effects of fragile
bits in palmprint recognition in the context of a specificcoding
scheme, BOCV. We proposed to mask out the fragile bits when
computing the Hamming distance. Furthermore, we proposed a
new metric FPD to measure the dissimilarity of two fragile-bit
masks. By fusing the modified Hamming distance and the FPD
together, we extended the original BOCV to E-BOCV. The ef-
fectiveness of the proposed methods was corroborated by ex-
periments conducted on the benchmark palmprint dataset.
REFERENCES
[1] D. Zhang, W. Kong, J. You, and M. Wong, “Online palmprint identi-
fication,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 25, no. 9, pp.
1041–1050, Sep. 2003.
[2] A. K. Jain and J. Feng, “Latent palmprint matching,” IEEE Trans. Pat-
tern Anal. Mach. Intell., vol. 31, no. 6, pp. 1032–1047, Jun. 2009.
[3] A. Kong, D. Zhang, and M. Kamel, “A survey of palmprint recogni-
tion,” Pattern Recognit., vol. 42, no. 7, pp. 1408–1418, Jul. 2009.
[4] D.Zhang,W.Zuo,andF.Yue, “A comparative study of palmprint
recognition algorithms,” ACM Comput. Surv.,vol.4
4, no. 1, pp.
2:1–37, Jan. 2012.
[5] J.G.Daugman,“Highconfidence visual recognition of persons by a
test of statistical independence,” IEEE Trans. Pattern Anal. Mach. In-
tell., vol. 15, no. 11, pp. 1148–1161, Nov. 1993.
[6] W. K. Kong and D. Zhang, “Competitive coding scheme for palm-
print verification,” in Proc. IEEE Int. Conf. Pattern Recognit., 2004,
pp. 520–523.
[7] W.Jia,D.Huang,andD.Zhang,“Palmprintverification based on
robust line orientation code,” Pattern Recognit.,vol.41,no.5,pp.
1504–1513, May 2008.
[8] Z.Sun,T.Tan,Y.Wang,andS.Z.Li,“Ordinalpal
mprint represen-
tation for personal identification,” in Proc. IEEE Int. Conf. Computer
Vis. and Pattern Recognit., 2005, pp. 279–284.
[9] Z.Guo,D.Zhang,L.Zhang,andW.Zuo,“Palm
print verification using
binary orientation co-occurrence vector,” Pattern Recognit. Lett., vol.
30, no. 13, pp. 1219–1227, Oct. 2009.
[10] W. Zuo, Z. Lin, Z. Guo, and D. Zhang, “The Multiscale competitive
code via sparse representation for palmprint verification,” in Proc.
IEEE Int. Conf. Computer Vis. and Pattern Recognit., 2010, pp.
2265–2272.
[11] F. Yue, W. Zuo, and D. Zhang, “ICP registration using principal line
and orientation features for palmprint alignment,” in Proc. IEEE Int.
Conf. Image Process., 2010, pp. 3069–3072.
[12] R. M. Bolle, S. Pankanti, J. H. Connell, and N. Ratha, “Iris individu-
ality: A partial iris model,” in Proc. IEEE Int. Conf. Pattern Recognit.,
2004, pp. 927–930.
[13] K. P. Hollingsworth, K. W. Bowyer, and F. J. Flynn, “The best bits in
an iris code,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 31, no. 6,
pp. 964–973, Jun. 2009.
[14] K. P. Hollingsworth, K. W. Bowyer, and F. J. Flynn, “Improved iris
recognition through fusion of Hamming distance and fragile bit dis-
tance,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 33, no. 12, pp.
2465–2476, Dec. 2011.
[15] PolyU Palmprint Database 2006 [Online]. Available: http://www.
comp.polyu.edu.hk/~biometrics
[16] T. S. Lee, “Image representation using 2D Gabor wavelet,” IEEE
Trans. Pattern Anal. Mach. Intell., vol. 18, no. 10, pp. 957–971, Oct.
1996.