Yanushkevich S.N., Wang P.S.P., Gavrilova M.L., Srihari S.N. (eds.) Image Pattern Recognition. Synthesis and Analysis in Biometrics
Подождите немного. Документ загружается.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
218 Synthesis and Analysis in Biometrics
contains a set of faces taken between April 1992 and April 1994 at the
Olivetti Research Laboratory in Cambridge, UK. There are 10 different
images of 40 distinct subjects. For some of the subjects, the images
were taken at different times, varying lighting slightly, facial expressions
(open/closed eyes, smiling/non-smiling) and facial details (glasses/no-
glasses). All the images are taken against a dark homogeneous background
and the subjects are in up-right, frontal position (with tolerance for some
side movement). Some sample images randomly chosen from this database
are shown in Fig. 8.3.
Fig. 8.1. Some sample images of ORL face database.
The proposed algorithm based-on nontensor product bivariate wavelet
is conducted as follows.
Algorithm for Nontensor-Product-Wavelet-Based face recognition
Step 1 Input training face images I, perform two level DNWT with
the new nontensor product wavelet transform constructed in
Section 8.2 on each face;
Step 2 Select low-frequency approximate components as the facial
feature representation and the PCA is applied for reduction of
computation complexity;
Step 3 Lengthen the feature matrix from columns and save them as
the input training data of the SVM;
Step 4 Input test face images, the same processing as Step 1 and
Step 2 are applied. Finally, input these face data into the
SVM systems and compute the correctness.
Here, we test the proposed method in both the performance against
the conventional PCA method
[
15
]
, the PCA+LDA+NN
[
28
]
method
and the recent face recognition scheme such as the DSWT+SVM and
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Nontensor-Product-Wavelet-Based Facial Feature Representation 219
DSWT+PCA+SVM
[
21
]
. To draw comparisons with others methods, the
same training and test sizes are used as in
[
21
]
, that is, five images chosen
randomly of each subject for training and the remaining for testing.
Table 8.1. Comparison of the computational time and recognition rate
among other algorithms.
Method Computation time (seconds) Recognition rate (%)
PCA+NN 11.5 89.88
PCA+LDA+NN 12.8 92.55
PCA+SVM 13.43 93.9
DSWT+SVM 7.28 96.5
DSWT+PCA+SVM 5.14 97.0
DNWT+PCA+SVM 4.40 98.6
As shown in Table 8.1, our algorithm produced a significant
improvement which includes a substantial advance in recognition rate and
in time of processing comparing the other illustrated algorithms. The result
shows that our approach: with a 98.6 recognition rate compares favorably
against the recent face recognition schemes that apply the same ORL
database for performance evaluation,such as the ULLELDA+NFL method
with a recognition rate 95.94
[
28
]
, the Gabor face+LDA+NFL method with
recognition rate 95.8
[
29
]
, and the KPCA method with recognition rate
97.75
[
16
]
.
Detailed evaluations are illustrated in Fig. 8.2. When the number of
features is 20, the proposed method gives the accuracy of 92.6, while
the DSWT+PCA+SVM, PCA+SVM and the PCA+NN give just 90.1,
92.5 and 88, respectively. When 85 features are used, the recognition
rate of the proposed method attains 96.5, and it is higher than that of
DSWT+PCA+SVM, PCA+SVM and PCA+NN. The good performance
of the proposed method demonstrate that DNWT+PCA+SVM is more
effective than other four methods in extracting and representing facial
features of real face images for face recognition.
In Fig. 8.3, the DSWT+PCA can represent the most features of real
face images. It is still inferior to the proposed the DNWT+PCA, perhaps
because it can reveal more facial features than tensor product ones do.
Figure 8.3 shows the effects of training sample size. The DSWT+PCA
obtains recognition rate 96.5 while the DNWT obtains the same recognition
rate at just 4 training faces. This is 2.1 lower than the accuracy of the
DNWT at 5 train samples. Especially, the DNWT can reach 100 when the
training sample is more than 7 while the DSWT never reaches.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
220 Synthesis and Analysis in Biometrics
20 30 40 50 60 70 80 90 100
85
90
95
100
Feature numbers
Recognition rate
DNWT+PCA+SVM
DSWT+PCA+SVM
PCA+SVM
PCA+NN
Fig. 8.2. Recognition rates versus the feature space dimension.
In a sense, it implies our method can obtains a perfect recognition rate
so long as the training sample is sufficient.
Many factors effect the correctness, such as feature representation,
training sample number. Besides these factors, the selection of classifier also
plays an important role. Conventional face recognition scheme commonly
adopted Nearest Feature Line algorithm (NFL) and Nearest Neighbor
algorithm (NN) as classifier. They have given a good evaluation under the
ORL testing framework
[
14,15,28,29
]
. However, in lately research work,
the SVM classifier has demonstrated to be better than the algorithms
mentioned above. In this paper, the SVM is applied to final classification.
The polynomial kernel function is often used and it was found that its
performance was better than Gauss kernel and other kernels in
[
21
]
.Here
we focus on the effects of SVM (Fig. 8.4) applied in the DSWT+PCA+SVM
method and our proposed method (DNWT+PCA+SVM). The influence of
the degree d upon polynomial kernel function is shown in Fig. 8.4 with
degree d =1, 2 and 3. It is illustrated that:
(a) The performance of the polynomial with d = 1 obviously outperform
d = 3 and a slightly better than d =2.
(b) Our scheme focused on the degree 1 which got the best recognition
rate.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Nontensor-Product-Wavelet-Based Facial Feature Representation 221
2 3 4 5 6 7 8
85
90
95
100
Training sample size
Recognition rate
DNWT+PCA+SVM
DSWT+PCA+SVM
DSWT+SVM
PCA+SVM
PCA+NN
Fig. 8.3. Recognition rates versus the train sample size.
Actually, we tried to take the high frequency sub-bands into account.
The detail information they carried may be those very discriminated. We
tried to plus one high frequency part to low frequency part after threshold,
and the entire compressed subband is represented as the facial feature,
but no better result has been obtained in such data set. It proves that
nontensor product wavelet transform reveals more facial features in face
image processing. The low-frequency components only are sufficient for
recognition.
8.4. Conclusions
Like other biometric systems, the primary stages employed by facial
recognition system to identify and verify subjects include facial image
capture, feature analysis, verification and identification. Feature analysis
is perhaps the most widely utilized facial recognition technology. Facial
feature extraction and representation are frequently used as the basis
of feature analysis. In this chapter, a new approach for facial feature
extraction and representation are developed based on the a constructed
nontensor product bivariate wavelet filter banks with linear phase, which
are constructed from the centrally symmetric matrices. By performing
nontensor product wavelet transform, the frequence components of the
wavelet subband are used for facial representation. It is demonstrated that
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
222 Synthesis and Analysis in Biometrics
20 30 40 50 60 70 80 90 100
85
90
95
100
Feature numbers
Recognition rate
DNWT+PCA+SVM polynomial(1)
DNWT+PCA+SVM polynomial(2)
DNWT+PCA+SVM polynomial(3)
DSWT+PCA+SVM polynomial(1)
DSWT+PCA+SVM polynomial(2)
DSWT+PCA+SVM polynomial(3)
Fig. 8.4. Recognition rates versus different kernels.
the proposed approach based the nontensor product wavelet transform is
more capable of representing facial image for recognition purpose than the
tensor product one. We also proposed a nontensor-product-wavelet-based
face recognition scheme by combining the technique of PCA and SVM.
Some compared experiments with main traditional methods are conducted.
The efficiency of our scheme has been demonstrated on the ORL database.
Acknowledgments
This work was partially supported by a grant (60403011) from the National
Natural Science Foundation of China and a grant (2003ABA012) and
(20045006071-17) from the Science & Technology Department, Hubei
province and Wuhan respectively, China.
Bibliography
1. Atalay, I. (1996). Face recognition using eigenfaces, M.sc. thesis, Institute of
Science and Technology, Istanbul Technical University, Istanbul, Jan.
2. Cavaretta, A. S., Dahmen, W. and Micchelli C. A. (1991). Stationary
subdivision, Mem. Amer. Math. Soc., 93, 453, pp. 1–186.
3. Qiuhui Chen, Charles A. Micchelli, Silong Peng, and Yuesheng Xu (2003).
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Nontensor-Product-Wavelet-Based Facial Feature Representation 223
Multivariate filter banks having a matrix factorization. SIAM Journal on
Matrix Analysis and its Application, 25,2, pp. 517–531.
4. Qiuhui Chen, Charles A. Micchelli, and Yuesheng Xu.(2005). Biorthogonal
wavelet filter banks from block centrally symmetric matrix. Linear Algebra
and Its Application, 402, pp. 111–125.
5. A. Cohen and I. Daubechies.(1993). Non-separable bidimensional wavelet
bases. Rev. Mat. Iberoamericana, 9,pp. 51–137.
6. W. Dahmen and C. A. Micchelli.(1997). Biorthogonal wavelet expansions.
Constr. Approx., 13,pp. 293–328.
7. I. Daubechies. Ten Lectures on Wavelets. (1992). Society for Industrial and
Applied Mathemathics, Philadelphia.
8. Hai-Hua Gao, Hui-Hua Yang, and Xing-Yu Wang.(2005). “Kernel PCA Based
Network Inrusion Feature Extraction and Dectection Using SVM”. In Proc.
IEEE NPL Conf.ICNC -2005- LNCS,3611, pp. 89–94.
9. W. He and M. J. Lai.(2000). “Examples fo bivariate nonseparable compactly
supported continuous wavelets”. IEEE Trans. Image Process, 9,8, pp. 949–
953.
10. R. Q. Jia and C. A. Micchelli.(1992). Using the refinement equation for the
construction of prewavelets v: extensibility of trigonometric polynomials.
Computing, 48, pp. 61–72.
11. J. Kovacevic and M. Vetterli.(1992). Nonseparable multidimensional perfect
reconstruction filter banks and wavelet bases for r
n
. IEEE Tran. on
Information Theory, 38, pp. 533–555.
12. Zhizheng Liang and Pengfei Shi.(2005). “Kernel direct discriminant analysis
and its theoretical foundation”. Elsevier Pattern Recognition Society 38,
pp. 445–447.
13. Hsien-Jen Lin.(2002). On Improving Linear Discriminant Analysis for Face
Recognition with Small Sample Size Problem. Master thesis, Department
of Computer Science and Information Engineering, National ChengKung
University.
14. L.J.Cao and W.K.Chong.(2002). “Feature Extraction in Support Vector
Machine: A Comparison of PCA,KPCA and ICA”. Pros. 9th Int. Conf. on
Neural Information Processing (ICONIP’02).
15. M.A.Turk and A.P.Pentland.(1991). “Face recognition using eigenfaces”. J.
Cognitive Neuroscience, 3,1 pp. 71–86.
16. M.H.Yang, N.Ahuja, and D.Kriegman.(2000). “Face recognition using kernel
eigenfaces”. Proc. IEEE Int. Conf. Image Processing. 1pp. 37-40
17. C. A. Micchelli.(2000).(1992). Using the refinement equation for the
construction of pre-wavelets vi: Shift invariant subspaces. Approximation
Theory, Spline Functions and Applications,(Maratea, 1991), NATO Adv. Sci.
Inst. Ser. C Math. Phys. Sci., 356, Kluwer Acad. Publ., Dordrecht.
18. C. A. Micchelli and T. Sauer.(1997) Regularity of multiwavelets. dv. Comput.
Math..
19. Vo Dinh Minh Nhat and Sungyoung Lee.(1997) “Line-Based PCA and LDA
Approaches for Face Recognition”. Proc. IEEE NPL Conf. ICNC -2005-
LNCS,3611, pp. 101–104.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
224 Synthesis and Analysis in Biometrics
20. S. D. Riemenschneider R. Q. Jia and D. X. Zhou.(1999). Smoothness of
multiple refinable functions and multiple wavelets. SIAM J. Matrix Anal.
Appl., 21, pp. 1–28.
21. Majid Safari, Meshrtash T.Harandi, and Babak N.Araabi.(2004). A svm-
based method for face recognition using a wavelet-pca representation of faces.
Pros. IEEE Int. Conf. on Image Processing, pp. 24-27.
22. Wendy S. Yambor. (2000). Analysis of PCA-based and fisher discriminant-
based image recognition. Technical Report CS-00-103, Computer Science
Department,Colorado State University.
23. M. Vetterli.(1986). Filter banks allowing perfect reconstruction. Signal Proc.,
10 (3):219–244.
24. Jian Yang, Alejandro F.Frangi, Jing yu Yang, David Zhang, and Zhong
Jin.(2005). “KPCA Plus LDA: A Complete Kernel Fisher Discriminant
Framework for Feature Extraction and Recognition”. IEEE Trans. Pattern
Anal. Mach. Intell., 27 (2).
25. Yanushkevich, S. N., Stoica, A., Shmerko, V. P. and Popel, D. V. (2005).
Biometric Inverse Problems. CRC Press/Taylor & Francis Group, Boca
Raton, FL.
26. Yanushkevich, S. N., Shmerko, V. P. and Stoica, A. (2006). Fundamentals of
Biometric-Based Training System Design, this issue.
27. BaiLing Zhang, HaiHong Zhang, and Shuzhi Sam Ge.(2004). “Face
Recognition by Applying Wavelet Subband Representation and Kernel
Associative Memory”. IEEE Trans. on Neural Networks., 15 (1).pp. 166–
177
28. Junping Zhang, HuanXing Shen, and Zhi-Hua Zhou.(2004) “Unified Locally
Linear Embedding and Linear Discriminant Analysis Algorithm (ULLELDA)
for Face Recognition”. Advances in Biometric Personal Authentication.,
LNCS 3338: pp.209–307.
29. Jianke Zhu, Mang I Vai, and Peng Un Mak.(2004) “Gabor Wavelets
Transform and Extended Nearest Feature Space Classifier for Face
Recognition”. In Proc.Third International Conference on Image and
Graphics, ICIG 2004., Hong Kong.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Chapter 9
Palmprint Identification by Fused Wavelet Characteristics
Guangming Lu
∗,‡,§
,DavidZhang
†,¶
,Wai-KinKong
†,
, Qingmin Liao
∗,∗∗
∗
Visual Information Processing Laboratory,
Graduate School at Shenzhen, Tsinghua University, Shen Zhen, China
†
Biometric Research Center,
Department of Computing, The Hong Kong Polytechnic University,
Kowloon, Hong Kong
‡
School of Computer Science and Engineering,
Harbin Institute of Technology, Harbin, China
§
lugmsz.tsinghua.edu.cn
¶
csdzhangcomp.polyu.edu.hk
cswkkongcomp.polyu.edu.hk
∗∗
liaoqmtsinghua.edu.cn
The wavelet theory has become a hot research topic in the last few
years for its important relative characteristics, such as sub-band coding,
multi-resolution analysis and filter banks. In this Chapter, we propose
a novel method of feature extraction for palmprint identification based
on the waveletransform, which is very efficient to handle the textural
characteristics of palmprint images at low resolution. In order to achieve
a high accuracy, four sets of statistical features (mean, energy, variance,
and kurtosis) are extracted based on the wavelet transform. Five
classifier combination strategies are then represented. The experimental
comparison of various feature sets and different fusion schemes shows
that the individual feature set of energy has the best classification ability
and the fusion schemes of Median rule as well as the Majority rule have
the best performance.
Contents
9.1. Introduction ................................... 226
9.2. Palmprint Images Collection and Preprocessing . . . . . . . . . . . . . . . 228
9.3. WaveletBasedFeatureExtraction....................... 231
9.4. PalmprintFeatureMatchingandDecisionFusion .............. 235
225
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
226 Synthesis and Analysis in Biometrics
9.5. ExperimentalResults.............................. 236
9.6. Conclusions ................................... 239
Bibliography ....................................... 239
9.1. Introduction
Biometrics based identification/verification technology has been greatly
emphasized and developed during the last few years. Lots of interesting and
meaningful research results have been achieved at the same time. Although
fingerprint identification is widely popular and extensively accepted by the
public, but some people do not have clear fingerprint because of their
physical work or problematic skin. Iris and retina recognition can provide a
very high accuracy, but they either suffer from the high cost input devices
or pose intrusion to users
[
1
]
,
[
2
]
,
[
3
]
. Recently, many researchers focus on
face and voice identification, nevertheless, their accuracy are still far from
being satisfactory. The accuracy and uniqueness of 3-D hand geometry is
still an open question
[
4
]
. Therefore, the investigation and development in
palmprint identification become particularly valuable.
First of all, the surface area of palmprint is very large when comparing
with fingerprint image. As the palm is the inner surface of the hand between
the wrist and the fingers, the stability of palmprint is high and the chance
of getting damaged is lower than that of the fingerprint
[
5
]
,
[
6
]
. The data
of palmprint would be helpful, as more information is available to verify
or identify the identity of a person. Many features of a palmprint can be
used to uniquely identify a person (as shown in Figure 9.1). In general, six
major types of features can be observed on a palm as follows.
Geometry Features. According to the palm’s shape, we can easily get the
corresponding geometry features, such as width, length and area.
Principal Line Features. Both location and form of principal lines
in a palmprint are very important physiological characteristics for
identifying individuals because they vary little over time.
Wrinkle Features. In a palmprint, there are many wrinkles which are
different from the principal lines in that they are thinner and more
irregular. These are classified as coarse wrinkles and fine wrinkles so
that more features in detail can be acquired.
Datum Points. Two end points called datum points are obtained by using
the principal lines (see Figure 9.1). These intersect on both sides of a
palm and provide a stable way to register palmprints. The size of a
palm can be estimated by using the Euclidean distance between these
end points.
April 2, 2007 14:42 World Scientific Review Volume - 9in x 6in Main˙WorldSc˙IPR˙SAB
Palmprint Identification by Fused Wavelet Characteristics 227
Fig. 9.1. Different features in the palm.
Delta Point Features. The delta point is defined as the center of a delta-
like region in the palmprint. Usually, there are delta points located in
the finger-root region. These provide stable and unique measurements
for palmprint authentication.
Minutiae Features. A palmprint is basically composed of the ridges,
allowing the minutiae features to be used as another significant
measurement.
Since
A palmprint contains different features including principal lines,
wrinkes, minutiae points, datum points and texture, which constitute
a set of multiscale features.
It motivates us to apply multi-resolution based approach to extract the
multiscale features. Different features require different image resolution
and have different applications. For minutiae points, and datum points,
extracting them requires a resolution of 400 dpi to 600 dpi
[
7
]
;for
principal lines and wrinkes, they can be extracted from images of
100 dpi
[
8
]
,
[
9
]
. In fact, we have tried to extract texture feature
from lower resolution images
[
10,11
]
. The high-resolution images are