Beruflich Dokumente
Kultur Dokumente
LOCAL FEATURES
i=1
P(c
j
|y
i
) (2)
In our case, posterior probabilities are directly
estimated by k-Nearest Neighbours. Let k
ij
the
number of neighbours of y
i
belonging to class c
j
.
Using this estimate in (2), our classication rule
becomes:
c = arg max
1jC
m
Y
i=1
k
ij
(3)
That is, a class c with the largest number of
votes accumulated over all the vectors belong-
ing to the test image is selected. This justies why
techniques of this type are often referred to as vot-
ing schemes.
2.4 Ecient approximate search of
matching feature vectors
An important feature of a practical palmprint
recognition method is speed. The preprocessing
and feature extraction steps described in section 2.1
and 2.2 respectively, have been carefully chosen to
be simple and fast, but the retrieval of a large num-
ber of high-dimensional vectors for each test image
and the search of the k-NN among a huge pool of
vectors obtained from the reference images seems
an intractable issue. A large reference set is essen-
tial for the proposed method to be eective. There-
fore, a fast search algorithm has to be applied to
perform the complete process in a reasonable time.
To this end, we have adopted the well known kd-
tree data structure.
If a guaranteed exact solution is not needed, as
can be assumed in our case, the backtracking pro-
cess associated to the exact kd-tree search can be
aborted as soon as a certain criterion is met by the
current best solution. In [4], the concept of (1 +)-
approximate nearest neighbour query is introduced.
A point p is a (1 + )-approximate nearest neigh-
bour of q if the distance from p to q is less than 1+
times the distance from p to its nearest neighbour.
This concept has been used here to obtain an e-
cient approximate search that can easily cope with
very large sets of reference vectors.
3 EXPERIMENTS
In this work we have used the PolyU Palmprint
Database, created by the Biometric Research Cen-
ter of Hong Kong [1]. This database contains
600 grayscale palmprint images from 100 dierent
palms. There are 6 images for each palmprint and
we have selected 3 images for training and 3 for test-
ing. So, both training and testing sets contain 300
palmprint images from 100 dierent subjects, 3 im-
ages from each one. The database is more detailed
in [1] and a bigger version is used and described
in [19].
Figure 3: Examples of images and its correspond-
ing selection (they are not in the same scale).
The images are correctly placed and neither ro-
tation nor scaling normalization are needed. On
the other hand, the here proposed local features
approach is invariant to translations. The position
of the local features is not stored and then, this
position is not used anyway to consider the nearest
neighbor of a local feature.
Despite other works [8, 9], where geometrical
hand features are also used, we only use the palm-
print features. Original database images are 384
284 sized. After reducing the size to its half, we
select an area of 75 75 pixels. The selected area
center is the mass center of the whole hand image,
excluding ngers. By this way, we select the palm-
print image area with more information. We only
work with this selected area and the preprocess-
ing approaches are applied only to this region of
interest. Two examples of original images and its
corresponding selections are shown in gure 3.
In experiments we have compared the error rate
obtained by each of the 3 preprocessing methods
showed in section 2.1. In all cases we have ex-
tracted local features from preprocessed images as
it is shown in section 2.2 and we have used the vot-
ing scheme showed in section 2.3. We have varied
30
35
40
45
50
55
60
100 150 200 250 300 350 400 450 500
E
r
r
o
r
r
a
t
e
(
%
)
Number of selected local features (n)
w=11
w=15
w=19
Figure 4: Results with no Equalisation preprocess-
ing.
10
15
20
25
30
35
40
100 150 200 250 300 350 400 450 500
E
r
r
o
r
r
a
t
e
(
%
)
Number of selected local features (n)
w=11
w=15
w=19
Figure 5: Results with Global Equalisation pre-
processing.
dierent parameters values: the local feature win-
dow size (w), the number of local features extracted
for each image (n) and in the case of local equali-
sation also the crop window size (v). However, for
clarify, in this work we only shown the best results.
They are shown in gures 4, 5 and 6. In case of
local equalisation, those results are obtained with
v = 9.
As can be seen in gures 4, 5 and 6, in all cases a
greater number of extracted local features (n) im-
proves the classication accuracy. Regarding to the
local feature window sizes (w), while for no equal-
isation and global equalisation there is not a sig-
nicant accuracy improvement, for the local equal-
isation method this parameter has an important
inuence on the error rate. The best results are
obtained when local equalisation is used, more con-
cretely a 0 % error rate is achieved for local equal-
isation with w = 19 and n = 300.
4 CONCLUSIONS
A system for palmprint recognition using local fea-
tures has been presented. For this purpose we have
used local features and we have compared 3 dif-
ferent preprocessing methods. The results for local
0
1
2
3
4
5
100 150 200 250 300 350 400 450 500
E
r
r
o
r
r
a
t
e
(
%
)
Number of selected local features (n)
w=11
w=15
w=19
Figure 6: Results with Local Equalisation prepro-
cessing.
equalisation are the best. On the other hand, in fu-
ture works we will report our system performance
by means of verication rates instead of classica-
tion rates.
References
[1] PolyU Palmprint Database.
http://www.comp.polyu.edu.hk/biometrics/.
[2] J. Daugman. The importance of being ran-
dom: Statistical principles of iris recognition.
Pattern Recognition, 36(2):279291, 2003.
[3] Nicolae Duta, Anil K. Jain, and Kanti V. Mar-
dia. Matching of palmprints. Pattern Recog-
nition Letters, 23:477485, 2002.
[4] S. Arya et al. An optimal algorithm for ap-
proximate nearest neighbor searching. Journal
of the ACM, 45:891923, 1998.
[5] Anil K. Jain and David Maltoni. Handbook of
Fingerprint Recognition. Springer-Verlag New
York, Inc, 2003.
[6] Anil K. Jain, Arun Ross, and Sharath
Pankanti. A prototype hand geometry-based
verication system. In Proc. of 2nd Intl Con-
ference on Audio- and Video-based Biometric
Person Authentication (AVBPA), Washington
D.C., pages 166171, March 22-24 1999.
[7] D. Keysers, R. Paredes, H. Ney, and E. Vidal.
Combination of tangent vectors and local rep-
resentations for handwritten digit recognition.
In In SPR 2002, International Workshop on
Statistical Pattern Recognition, 2002.
[8] A. Kumar, D. C. M. Wong, H. C. Shen, and
A. K. Jain. Personal verication using palm-
print and hand geometry biometric. In Proc.
of 4th Intl Conf. on Audio- and Video-Based
Biometric Person Authentication (AVBPA),
pages 668678, Guildford (UK), June 9-11
2003.
[9] Ajay Kumar and David Zhang. Integrating
shape and texture for hand verication. In
Proc. of Third International Conference on
Image and Graphics (ICIG04), pages 222
225, Hong Kong (China), 2004.
[10] Stan Z. Li and Anil K. Jain, editors. Handbook
of Face Recognition. Springer, 2004.
[11] K Messer, J Kittler, M Sadeghi, A Kostin, and
R. Paredes et al. Face authentication test on
the banca database. In J.Kittler, M Petrou,
and M Nixon, editors, Proc. 17th Intern. Conf.
on Pattern Recognition, volume IV, pages 523
529, Los Alamitos, CA, USA, August 2004.
IEEE Computer Society Press.
[12] R. Mohr, A. Picard, and C. Schmid. Bayesian
decision versus voting for image retrieval. In
Proc of the CAIP-97, 1997.
[13] R. Paredes, J. C. Perez, A. Juan, and E. Vi-
dal. Local Representations and a direct Voting
Scheme for Face Recognition. In Proc. of the
Workshop on Pattern Recognition in Informa-
tion Systems (PRIS 01), Set ubal (Portugal),
July 2001.
[14] R. Paredes, E. Vidal, and F. Casacuberta.
Local features for speaker recognition. In
SPR 2004.International Workshop on Statisti-
cal Pattern Recognition. LNCS 3138 of Lecture
Notes in Computer Science, pages 10871095,
2004.
[15] A. Ross, S. Dass, and A. K. Jain. A deformable
model for ngerprint matching. Pattern Recog-
nition, 38(1):95103, 2005.
[16] M. Hatef R.P Duin J. Kittler and J. Matas. On
combinig classiers. IEEE Trasn. on PAMI,
1998.
[17] U. Uludag, A. Ross, and A. K. Jain. Biometric
template selection and update: A case study in
ngerprints. Pattern Recognition, 37(7):1533
1542, 2004.
[18] Jane You, Wenxin Li, and David Zhang. Hi-
erarchical palmprint identication via multi-
ple feature extraction. Pattern Recognition,
35:847859, 2002.
[19] David Zhang, Wai-Kin Kong, Jane You, and
Michael Wong. Online palmprint identica-
tion. IEEE Transaction on Pattern Analy-
sis and Machine Learning, 25(9):10411050,
September 2003.