Palmprint Identification using Boosting Local Binary Pattern

Xianji Wang, Haifeng Gong, Hao Zhang, Bin Li, Zhenquan Zhuang
Department of Electronic Science and Technology,
University of Science and Technology of China, Hefei, China
xjw@mail.ustc.edu.cn

Abstract identification, such as principle lines, wrinkles, ridges,

minutiae points, singular points, texture, etc [15].
Local Binary Pattern (LBP) is a powerful texture These features can be extracted at different image
descriptor that is gray-scale and rotation invariant [3]. resolutions. Of these features, texture is one of the
Because texture is one of the most clearly observable most clearly observable features in low-resolution
features in low-resolution palmprint images, we think palmprint images. This fact inspires us to develop a
local binary pattern based features are very palmprint recognition method based on powerful
discriminative for palmprint identification. In this texture descriptors.
paper, we propose a palmprint identification approach In this paper, we present a novel approach for
using boosted local binary pattern based classifiers. palmprint identification based on boosted statistical
The palmprint area is scanned with a scalable sub- local feature—Local Binary Pattern histogram—based
window from which local binary pattern histograms classifiers. Local Binary Pattern (LBP) is a powerful
[4] are extracted to represent the local features of a texture descriptor that is gray-scale and rotation
palmprin image. The multi-class problem is invariant [3]. In [4], Timo et al used local binary
transformed into a two-class one of intra- and extra- pattern histogram for face recognition and showed this
class by classifying every pair of palmprint images as method outperformed other well-known methods such
intra-class or extra-class ones[19]. We use the as PCA, EBGM and BIC on the FERET database. In
AdaBoost[18] algorithm to select those sub-windows their method, they limit the size and position of the
that are more discriminative for classification. Weak obtained features by equally dividing a face image into
classifiers are constructed based on the Chi square sub-windows from which the LBP histograms are
distance between two corresponding local binary extracted and only selecting one combination of the
pattern histograms. Experiments on the UST-HK sampling points P and the circle radius R for all of the
palmprint database show competitive performance. sub-windows. The weighted Chi square distance is
used to measure the dissimilarity of two different face
1. Introduction images. But the weights of Chi square distance are
chosen in a somehow coarse way. In our approach, we
Biometric personal identification/verification has improve their method in three aspects. First, by
been studied for a long time, which uses the scanning the palmprint image with a scalable sub-
physiological characteristics of human such as window, we obtain a more complete and agile
fingerprint, iris, face, palmprint, retina and hand description of the palmprint. Second, for each sub-
geometry or some behavioral features such as voice, window we extract LBP histograms of six
signature and giant[1]. Many biometric systems have combinations of P and R , then select the best one for
been developed for various commercial applications that sub-window. Last, the weights of Chi square
such as banking, airport security control and access distance are learned by applying the statistical learning
control. algorithm—AdaBoost. Experiments result on the
Compare to other biometric technologies, palmprint UST_HK palmprint database shows that the proposed
identification has a much shorter history. Even we can method yields the equal error rate of 2% that is
say that it is in its infancy. But in recent years, people comparable to that of the PamlCode method [12].
have paid much attention to this promising field. The rest of this paper is organized as follows:
Various palmprint representations have been adopted section 2 introduces the construction of LBP features.
for identification[5]-[14]. Section 3 proposes the method of using AdaBoost to
There are many discriminative features in a select more efficient LBP features. Experiment results
palmprint image that can be used for personal

The 18th International Conference on Pattern Recognition (ICPR'06)

0-7695-2521-0/06 $20.00 © 2006
 are reported in section 4. Section 5 summarizes this the palmprint image with a scalable sub-window, thus
paper. a sequence of sub-regions Ro , R1 ,! , Rm−1 is
generated and the spatially enhanced histogram is
2. Local Binary Pattern Based Features defined as:
The original LBP operator introduced by Ojala et al H i , j = ¦ I { f l ( x, y ) = i}I {( x, y ) ∈ R j } (2)
is a powerful texture descriptor [2]. The operator labels x, y

the pixels of an image with the binary thresholding In this histogram we effectively have a description
result of the 3×3 neighborhood of each pixel with the of the palmprint on two different levels of locality: the
center value. Then the histograms of the labels are used labels for the histogram contain information about the
as texture descriptor. Figure 1 is an illumination of the patterns on a pixel-level; the labels are summed a small
basic LBP operator. region to produce information on a region level [4].
Several possible dissimilarity measures have been
proposed for histograms. In this work, we use the
5 8 2 Threshold 1 1 0 following weighted χ 2 statistic:
3 4 6 0 1
1 2 9 0 0 1
m −1 n j −1 ( S j ,i − M j ,i )2
χ (S, M ) = ¦ ¦ wj
2
(3)
S j ,i + M j ,i
w
Binary: 11011000 j =0 i =0

Figure 1. The basic LBP operator Where w j is the weight of region j , n j is the number
th
Later this operator was extended to use of bins of the j histogram.
neighborhoods of different sizes and shapes using
bilinear interpolation [3]. The interpolation of intensity 3. Leaning the most Discriminative LBP
values of sampling points that are not in the center of Features
pixels allows any radius and number of pixels in the
neighborhood. By scanning the palmprint image with a scalable
Another extension to the original LBP operator is to sub-window, we can get thousands of sub-regions.
use so called uniform patterns [3]. A Local Binary Moreover, in this work, instead of fixing P and R for
Pattern is called uniform if it contains at most two all of the sub-regions like [4], we extract histograms of
bitwise transitions from 0 to 1 or 1 to 0, e.g. 00000000,
six combinations of P and R for every sub-region.
11100011 and 00011111. In [2], Ojala et al found in
These combinations are: (4, 1), (4, 2), (8, 1), (8, 2), (16,
their experiments with texture images that uniform
1), (16, 2). Thus the set of intra/extra features is an
patterns account for a bit less than 90% of all patterns
over-complete set and contains much redundant
when using the (8, 1) neighborhood and for about 70%
information. Here we use the AdaBoost learning
in the (16, 2) neighborhood.
algorithm to select the most significant LBP features
We use the following notation for the LBP operator:
from a large feature set.
LBPPu,2R , where P is the number of sampling points, AdaBoost [18] is one of the most commonly used
R is the radius, u 2 stands for using only uniform methods of combining classifiers that incrementally
builds linear combinations of weak classifiers to form a
patterns and labeling all remaining patterns with a
strong classifier. When used to feature selection,
single label.
AdaBoost constructs weak classifiers each of which is
After the labels sets are predetermined, a histogram
based on one of the selected features. The final strong
of the labeled image fl ( x, y ) can be constructed as: classifier is the linear combination of these weak
H i = ¦ I{ f l ( x, y ) = i}, i = 0,! , n − 1 (1) classifiers. Figure 3 shows the AdaBoost algorithm.
x, y
The AdaBoost learning procedure is aimed to derive
Where n is the number of different labels produced by α t and ht ( x) .
the LBP operator and IA = 1 if A is true, 0 otherwise.
This histogram contains information about the Input: N training examples ( x1 , y1 ),! , ( xN , yN ) where
distribution of local micro-patterns, such as edges, yi = −1, +1 for negative and positive examples
spots and flat areas, over the whole image. In [4],
Ahomen et al proposed to divide the whole image into respectively.
sub-regions to retain spatial information. Here, we scan

The 18th International Conference on Pattern Recognition (ICPR'06)

0-7695-2521-0/06 $20.00 © 2006
 Initialize: weights w1,i = 1/ 2m,1/ 2l for yi = −1, +1
respectively with m + l = N and i = 1,! , N .
For t = 1,!, T
N
1. Normalize the weights: wt ,i = wt ,i ¦w t, j
so that
j =1 Figure 3. The first five sub-windows from which the
Chi square distance between corresponding
wt is a probability distribution. histograms are calculated. The numbers in the
2. For each feature j with wt , train a classifier bracket are the corresponding sampling points P
and radius R.
h j and the error ε j = ¦ i wi h j ( xi ) − yi .
3. Choose the classifier ht with the lowest error ε t . We trained the other four classifier based on the
histograms of (16,2), (16,1), (8,2) and (8,1)
1− ei
4. Update the weights: wt +1,i = wt ,i β t , where ei respectively. Each classifier contains also 45 weak
classifiers like our classifier. The verification
equal 0 if example xi is classified correctly and 1 performance of five classifiers based on five types of
otherwise, β t = ε t 1 − ε t histograms is shown in Figure 4. From the figure we
can see that the performance of our classifier based on
Output: the final strong classifier:
T 1 T
all of the six types of combinations of P and R is
1 ¦ t =1α t ht ( x ) ≥ 2 ¦ t =1αt much better than the classifiers based on single
h( x ) = { combination of P and R . The equal error rate is about
0 otherwise
1 2% that is comparable to that of PalmCode method [12]
where α = log
on the same database as reported in [17].
t
βt

Figure 2. The AdaBoost Algorithm

4. Experiment Results
The proposed method was tested on the UST Hand
Image Database [20]. There are 5740 hand images
captured from 287 persons. Each person has 10 left and
10 right hand images. We randomly choose two
images of each class to construct the training set. All
images are preprocessed, cropped and scaled to 128
pixels high and 128 pixels wide. The training set yields
574 intra-class pairs and 657,804 extra-class pairs.
By shifting and scaling the sub-window, 7,021 sub-
regions are generated. For each sub-region, we extract
six types of LBP histograms. Thus a total of 42,126
histograms are extracted for each palmprint image. Figure 4. The receiver operator characteristic
And we get 42,126 candidate features for intra-extra (ROC) curves of five Boosting-LBP based classifiers
class classification by computing the Chi square based on five types of histograms respectively.
distance between corresponding histograms of each
image pair. Then, we apply AdaBoost on the positive
sample set of 574 intra-class pairs and the negative 5. Conclusion
sample set of 657,804 extra-class pairs to select the
most discriminative features. Finally, 45 Chi square This paper presents a new method for palmprint
distances of 45 corresponding LBP histograms pairs recognition by effectively extracting the texture using
are selected. The first five sub-windows, from which Local Binary Patterns and selecting the most
the LBP histograms are extracted, are shown in Figure discriminative Local Binary Patterns based features
3. using two-class AdaBoost algorithm. The multi-class
problem of palmprint recognition is transform into a
two-class one of intra- and extra- class by learning a

The 18th International Conference on Pattern Recognition (ICPR'06)

0-7695-2521-0/06 $20.00 © 2006
 similarity of every palmprint pair, as in [19]. Recognition and Artificial Intelligence, vol. 16, no. 4, pp.
Experimental results on the UST Hand Image Database 417-432, 2002.
show that our method can achieve the equal error rate [8] G. Lu, D. Zhang and K. Wang, “Palmprint Recognition
of 2% that is comparable to that of the PamlCode Using Eigenpalms Features”, Pattern Recognition Letters,
vol. 24, issues 9-10, pp. 1463-1467, 2003.
method [12] on the same database as reported in [17]. [9] C.C. Han, H.L. Cheng, K.C. Fan and C.L, Lin, “Personal
Authentication Using Palmprint Features”, Pattern
Acknowledgements Recognition, vol. 36, no. 2, pp. 371-381, 2003.
[10] J. You, W.K. Kong, D. Zhang and K. Cheung, “On
The work is supported by Talent Promotion Hierarchical Palmprint Coding width Multi-features for
Foundation of Anhui Province under grant No: Personal Identification in Large databases”, IEEE
2004Z026, and The Science Research Fund of MOE- Transactions on Circuit Systems for Video Technology, vol.
14, no. 2, pp. 234-243, 2004.
Microsoft Key Laboratory of Multimedia Computing [11] Lei Zhang and David Zhang, “Characterization of
and Communication under grant No.05071811. Palmprints by Wavelet Signatures via Directional Context
Modeling”, IEEE Trans. On SMC—B, vol. 34, no. 3, pp.
1335-1347, June 2004.
References [12] D. Zhang, W. Kong, J. You and M. Wong, “On-line
Palmprint Identification”, IEEE Trains. on PAMI, vol. 25, no.
[1] A.K. Jain, R.M. Bolle and P. Pankanti (eds.), 9, pp. 1041-1050, 2003.
[13] W.K. Kong and D. Zhang, “Feature-Level Fusion for
Biometrics: Personal Identification in Networked
Effective Palmprint Authentication”, Proceedings of 1st ICBA,
Society, Boston, Mass: Kluwer Academic Publishers, LNCS 3072, pp. 761-767, 2004.
1999. [14] W.K. Kong and D. Zhang, “Competitive Coding
[2] T. Ojala, M. Pietikäinen and D. Harwook, “A Scheme for Palmprint Verification”, Proceedings of the 17th
comparative study of texture measures with ICPR, vol. 1, pp. 520-523, 2004.
classification based on feature distribution”, Pattern [15] W. Shu and D. Zhang, “Automated Personal
Recognition, vol. 29, no. 1, pp. 51-59, 1996. Identification by palmprint”, Optical Engineering, vol. 37, no.
[3] T. Ojala, M. Pietikäinen and T. Mäenpää, 8, pp. 2659-2362, 1998.
“Multiresolution gray-sclae and rotation invariant [16] W. Shi, G. Rong, Z. Bain and D. Zhang, “Automatic
Palmprint Verification”, International Journal of Image and
texture classification with local binary patterns”, IEEE Graphics, vol. 1, no. 1, pp. 135-152, 2001.
Transaction on Pattern Analysis and Machine [17] Zhenan Sun, Tieniu Tan, Yunhong Wang and Stan Z. Li.
Intelligence, vol. 24, no. 7, pp. 971-987, 2002. “Ordinal Palmprint Representation for Personal
[4] T. Ahonen, A. Hadid and M. Pietikäinen, “Face Identification”. Proceedings of the IEEE CVPR, 2005.
Recognition with local binary patterns”, Proceedings of the [18] Y. Freund and R.E. Schapire, “A Decision-Theoretic
European Conference on Computer Vision, LNCS 3021, pp. Generalization of On-line Learning and An Application to
469-481, 2004. Boosting”, Journal of Computer and System Sciences, vol. 55,
[5] D. Zhang and W. Shu, “Two Novel Characteristics in no. 1, pp. 119-139, 1997.
Palmprint Verification: Datum Point Invariance and Line [19] B. Moghaddam, C. Nastar and A. Pentland, “A Bayesian
Feature Matching”, Pattern Recognition, vol. 32, no. 4, pp. Similarity Measure for Direct Image Matching”, Media Lab
691-702, 1999. Tech Report N0. 393, MIT, August 1996
[6] N. duta, A.k. Jain and K.V. Mardia, “Matching of [20] P. Viola and M. Jones, “Robust real time object
Palmprint”, Pattern Recognition Letters, vol. 23, no. 4, pp. detection”, IEEE ICCV Workshop on Statistical and
477-485, 2001. Computational Theories of Vision, July 2001.
[7] W. Li, D. Zhang and Z. Xu, “Palmprint Identification by [21] UST Hand Image Database, http://visgraph.cs.ust.
Fourier Transform”, International Journal of Pattern hk/biometrics/Visgraph_web/index.html.

The 18th International Conference on Pattern Recognition (ICPR'06)

0-7695-2521-0/06 $20.00 © 2006