Upload
duongcong
View
236
Download
2
Embed Size (px)
Citation preview
1
1
2
3
Abstract — This paper introduces the establishment of PolyU Near-Infrared Face Database (PolyU-4
NIRFD) for face recognition. The PolyU-NIRFD contains images from 350 subjects, each 5
contributing about 100 samples with variations of pose, expression, focus, scale and time, etc. In total, 6
35,000 samples were collected in the database. The PolyU-NIRFD provides a platform for 7
researchers to develop and evaluate various near-infrared face recognition techniques under large 8
scale, controlled and uncontrolled conditions. Finally, we provide three protocols to evaluate the 9
baseline face recognition methods, including Gabor based Eigenface, Fisherface and LBP (local 10
binary pattern) on the PolyU-NIRFD database. 11
12
Index Terms — Near-infrared face recognition, face database, feature extraction 13
14
Corresponding author: Lei Zhang ([email protected]). Tel: 852-27767355. Fax: 852-27740842. Lei Zhang and David Zhang are with the Biometrics Research Center, Dept. of Computing, The Hong Kong Polytechnic University, Hong Kong. Baochang Zhang is with the School of Automation Science and Electronic Engineering, Beihang University, Beijing, China. Linlin Shen is with the Dept. of Electronic Engineering, Shenzhen University, Shenzhen, China.
Baochang ZHANG, Lei ZHANG, David ZHANG, Linlin SHEN, Zhenhua GUO
PolyU Near-Infrared Face Database
2
1. INTRODUCTION 15
Face recognition (FR) is a promising technology for automated personal authentication and it has a great 16
potential in applications of public security, video surveillance, access control and forensics, etc. [R. Chellappa 17
et al. 1995, W. Zhao et al. 2003]. Meanwhile, FR is one of the most active topics in the filed of computer vision, 18
and several large-scale face databases [W. Gao et al. 2008, Enrique et al. 2003, T. Sim et al. 2003, K. Messer, et 19
al. 1999, A. M. Martinez and R. Benavente 1998, A. S. Georghiades et al. 2001, K. C. Lee et al. 2005] are 20
publicly available for evaluating and comparing various FR methods. Generally speaking, FR in visible 21
spectrum has been mostly studied because it is convenient to implement in various environment and has a wide 22
range of applications [R. Chellappa et al. 1995, W. Zhao et al. 2003, Turk and Pentland. 1991, Belhumeur et 23
al. 1997]. Many FR algorithms have been proposed [R. Chellappa et al. 1995, W. Zhao et al. 2003], and the 24
large-scale face databases play an important role in evaluating and developing FR algorithms. 25
Face recognition technology (FERET) [P.J. Phillips et al. 2000], face recognition vendor test (FRVT) [P. J. 26
Phillips et al. 2002], and face recognition grand challenge (FRGC) [P. J. Phillips et al. 2005] have pioneered 27
both evaluation protocols and database construction. FRGC is more changeling than FERET and FRVT, as it 28
contains more uncontrolled variations and 3D images in its database. For example, in the most challenging set 29
of the FRGC v2 (Exp#4), the training set contains 10776 images from 222 subjects, while the query and target 30
sets contain 8,014 and 16,028 images respectively. The 3D training set of FRGC consists of controlled and 31
uncontrolled still images from 943 subject sessions, while the validation partition of FRGC contains images 32
from 466 subjects collected in 4007 subject sessions. Other publicly available face databases include the CAS-33
PEAL [W. Gao et al. 2008], BANCA [Enrique et al. 2003], CMU PIE [T. Sim et al. 2003], XM2VTSDB [K. 34
Messer, et al. 1999], AR [A. M. Martinez and R. Benavente 1998], Yale [Georghiades et al. 2001, Lee et al. 35
2005], etc. These face databases in visible spectrum provide a good evaluation platform for various FR 36
techniques, and in return they greatly facilitate the development of new FR methods. 37
In visual face recognition, the performance suffers from the lighting variations. To solve this problem, 38
3
traditional methods are mostly based on Lambertian model [T. Sim et al. 2003], which is too simply to describe 39
the real face surface under various illuminations. Recently, active near-infrared (NIR) FR was proposed to deal 40
with the illumination variations in different environments, and NIR based FR has shown promising performance 41
in real application, which uses different imaging sensors in invisible spectral bands to reduce the affection of 42
ambient light. X. Zou et al. proposed to use active NIR light to localize face areas in the images and then 43
recognize faces [X. Zou et al. 2005]. S.Z. Li et al. extracted the Local Binary Pattern (LBP) feature and used 44
Fisher analysis for NIR based FR, and they developed a complete NIR FR system which can perform face 45
detection, eye localization and face identification [Stan Li et al. 2007]. Pan et al. [Z.H. Pan et al. 2003] proposed 46
an NIR-based FR system which captures face images in wavelength of 0.7um-1.0um. Some other works using 47
NIR images for FR can be found in [J. Dowdall et al. 2003, A. S. Georghiades et al. 2001]. Zou et al. [X. Zou et 48
al. 2005] have shown that FR in NIR band has better performance than that in the visible band, and this was 49
also validated in S.Z. Li’s work [Stan Li et al. 2007]. However, so far there is not a large scale NIR face 50
database which is publicly available. There is a high demand to construct an open NIR FR database, on which 51
the researchers can test and compare their algorithms. In this paper, we will introduce such a database we 52
constructed in the Hong Kong Polytechnic University, and name the database as PolyU Near-infrared Face 53
Database (PolyU-NIRFD). 54
The face images in PolyU-NIRFD were collected from 350 subjects, each subject providing about 100 55
samples. The sample images involve various variations of expression, pose, scale, focus and time, etc. To 56
evaluate the performance of different FR methods on the PolyU-NIRFD, we provide three test protocols, 57
including the partition strategy of the training, gallery and probe sets and the baseline evaluation schemes. The 58
baseline algorithms we used for comparison are Eigenface, Fisherface, Local Binary Pattern (LBP) [T. Ojala et 59
al. 2002] and their Gabor filtering enhanced versions, which are well-known and representative methods in the 60
field of FR. Considering that the Gabor filtering can improve significantly the FR accuracy (e.g. Gabor 61
Fisherface and Gabor LBP are among the best FR algorithms), we also examine its performance on PolyU-62
NIRFD. 63
4
The rest of the paper is organized as follows. Section II introduces the developed NIR face acquisition 64
system and the establishment of PolyU-NIRFD. In Section III, the baseline algorithms are briefly described. 65
Section IV presents extensive experiments using three protocols to evaluate the performance of various FR 66
methods, including Gabor-Eigenface, Gabor-Fisherface, Gabor-LBP, on the PolyU-NIRFD. Conclusions are 67
drawn in Section V. 68
69
2. POLYU-NIRFD CONSTRUCTION 70
Different from the FR in visible band, which can simply use a common camera to capture face images, 71
FR in NIR band needs some additional hardware and special system design for image acquisition. This 72
section describes the NIR image acquisition system and the collection of the PolyU-NIRFD database. 73
74
2.1. Near-Infrared Face Image Acquisition 75
The hardware of our NIR face image acquisition system consists of a camera, an LED (light emitting diode) 76
light source, a filter, a frame grabber card and a computer. A snapshot of the constructed imaging system is 77
shown in Fig. 1. The camera used is a JAI camera, which is sensitive to NIR band. The active light source is in 78
the NIR spectrum between 780nm - 1,100 nm and it is mounted on the camera. The peak wavelength is 850nm, 79
and it lies in the invisible and reflective light range of the electromagnetic spectrum. An NIR LED array is used 80
as the active radiation sources, and it is strong enough for indoor use. The LEDs are arranged in a circle and 81
they are mounted on the camera to make the illumination on the face is as homogeneous as possible. The 82
strength of the total LED lighting is adjusted to ensure a good quality of the NIR face images when the camera-83
face distance is between 80cm-120cm, which is convenient for the users. When mounted on the camera, the 84
LEDs are approximately coaxial to the imaging direction and thus provide the best possible straight frontal 85
lighting. Although NIR is invisible to the naked eyes, many CCD cameras have sufficient response to the NIR 86
spectrum. The filter we used in the device is used to cut off the visible light, whose spectrum is shorter (780nm, 87
5
visible light). For the convenience of data collection, we put the imaging device into a black box of 19cm width, 88
19cm long, and 20cm high, as shown in Figure 1. 89
90
91 Fig. 1: The NIR face image acquisition device. ‘A’ is the NIR LED light source, and ‘B’ is the NIR sensitive camera 92
with a NIR filter. 93
94
2.2. PolyU-NIRFD Construction 95
By using the self-designed data acquisition device described in Section II-A, we collected NIR face images 96
from 350 subjects. During the recording, the subject was first asked to sit in front of the camera, and the normal 97
frontal face images of him/her were collected. Then the subject was asked to make expression and pose changes 98
and the corresponding images were collected. To collect face images with scale variations, we asked the 99
subjects to move near to or away from the camera in a certain range. At last, to collect face images with time 100
variations, samples from 15 subjects were collected at two different times with an interval of more than two 101
months. In each recording, we collected about 100 images from each subject, and in total 35,000 images were 102
collected in the PolyU-NIRFD database. The sample images in the PolyU-NIRFD are labeled as 103
‘NN_xxxxxx_S_D_****’, where “NN” represents the prefix of the label string, ‘S’ represents the Gender 104
information, ‘xxxxxx’ indicates the ID serial number of the subject, ‘D’ denotes the place where the image was 105
captured, and ‘****’ is the index of the face image. For example, “NN_200001_F_B_024” means that the 24th 106
image is from a Female subject collected in Beihang University. Figure 2 shows some face images of a subject 107
6
with variations of expression, pose and scale. Figure 3 shows the images of some subjects which were taken in 108
different times. 109
110
111 a) b) 112
113 c) d) 114
Fig. 2: Sample NIR face images of a subject. (a) Normal face image; and images with (b) expression variation; (c) 115 pose variation and (d) scaling variation. 116
117
118
119 Fig. 3: Sample NIR face images captured in more than two months. 120
121
To evaluate the performance of different methods on the PolyU-NIRFD, we design three types of 122
experiments, each of which contains a training set, a target (Gallery) set and a query (Probe) set. In Exp#1, the 123
used images include frontal face images as well as images with expression variations, scale changes (include 124
7
blurring), and time difference, etc. In this experiment, the sizes of the training set, target set and query set are 125
419, 574 and 2763, respectively. In Exp#2, we add more faces captured in uncontrolled conditions to make the 126
test more challenging. The sizes of the training set, target set and query set are 1876, 1159, and 4747 127
respectively. In Exp#3, we focus on the images with high pose variations and exclude the images with 128
expression, scale and time variations. The sizes of the training set, target set and query set are 578, 951, and 129
3648, respectively. In next section, we will present baseline face recognition algorithm, and then in Section IV 130
the three types of experiments will be conducted by different methods. 131
132
3. BASELINE ALGORITHMS 133
We choose both grey-level image and Gabor based Eigenface, Fisherface, and the LBP method as the 134
baseline algorithms. 135
3.1 Gabor Wavelet 136
In image processing and object recognition, Gabor features are widely used for image feature descriptors 137
extracted by a set of Gabor wavelets (kernels) which model the receptive field profiles of cortical simple cells. 138
They can capture salient visual properties in an image, such as spatial characteristics, because the kernels can 139
selectively enhance features in certain scales and orientations. The Gabor wavelets (kernels, filters) can be 140
defined as follows: 141
2 2 2 2, ,
2( || || || || / 2 ), / 2
, 2
|| ||ψ ( ) u v u viu v
u v e e eσ σ
σ− −⎡ ⎤= −⎣ ⎦
k z k zkz , (1) 142
where ( )xy=z , ( ) ( ) / 2cos
, sin max max, / 2 , , 0,..., -1, 0,..., -18
vjx v u
jy v u
k ku v k k v uk u v v u uφ
φππ φ= = = = = =k , v is the 143
frequency, u is the orientation, max max5, 8 and 2v u σ π= = = . In this paper, only magnitude part of the Gabor 144
feature is used to enhance the performance of Eigenface, Fisherface, and LBP. 145
146
8
3.2 Baseline Algorithm 147 148
Principal Components Analysis (PCA) is commonly used for dimensionality reduction in signal 149
processing, also known as Eigenface in face recognition. Eigenface serves as the baseline method due to its easy 150
implementation and reasonable performance. PCA chooses projection directions that maximize the total scatter 151
across all images of all faces in the training set, and the scatter matrix is calculated as: 152
_ _
1( )( )( )
CT
i i ii
p x x x x x=
= − −∑S (2) 153
Linear Discrimiant Analysis (LDA) is a widely used method for feature extraction and dimensionality 154
reduction in face recognition, which is the core part of the well-known Fisherface method. LDA tries to find the 155
discriminative project direction in which training samples belonging to different classes are separated. 156
Mathematically, it calculates the projection matrix in a way that the ratio of the determinant of the between-157
class scatter matrix of the projected samples and the within-class scatter matrix of the projected samples is 158
maximized. The between-class scatter matrix bS and within-class scatter matrix wS are defined as follows: 159
1( )( )( )
CT
b i i ii
p u u u uϖ=
= − −∑S , (3) 160
1( ) {(( )( ) ) | }
CT
w i i i i i ii
p E x u x uϖ ϖ=
= − −∑S , (4) 161
where 1
(1/ )in
i ijj
u n x=
= ∑ denotes the sample mean of class i , u is the mean of all training images, and ( )ip ϖ is the 162
prior probability. 163
In this paper, we also test the LBP based face recognition on the PolyU-NIRFD database. Derived from a 164
general definition of texture in a local neighborhood, LBP is defined as a gray-scale invariant texture measure 165
and is a useful tool to model texture images. LBP later has shown excellent performance in many comparative 166
studies, in terms of both speed and discrimination performance. The original LBP operator labels the pixels of 167
an image by thresholding the 3 3× neighborhood of each pixel with the value of the central pixel and 168
concatenating the results to form a number. The thresholding function ()f for the basic LBP can be formally 169
9
represented as: 170
00
0
0, ( ) ( )( ( ), ( ))
1, ( ) ( )i
ii
if I Z I Z thresholdf I Z I Z
if I Z I Z threshold− ≤⎧
= ⎨ − >⎩ (5) 171
where , 1,...,8iZ i = is an 8-neighborhood point around 0Z as shown in the central part in Fig. 4. A LBP can 172
also be considered as the concatenation of the binary gradient directions, and is called a micro-pattern. Fig. 5 173
shows an example of obtaining a LBP micro-pattern when the threshold is set to zero. The histograms of these 174
micro-patterns contain information of the distribution of the edges, spots, and other local features in an image. 175
1Z 2Z 3Z
8Z 0Z 4Z
7Z 6Z 5Z 176
Fig. 4. An example of 8-neighborhood around 0Z . 177 178
179
Fig. 5. An example of obtaining the LBP micro-pattern for the region in the black square. 180
181
4. EXPERIMENTS 182
The main objectives of the experiments in this section are to evaluate the performance of well-known face 183
recognition algorithms on the PolyU-NIRDF and investigate the strengths and weaknesses of the proposed 184
method and baseline algorithms. We choose the Eigenface (i.e. PCA) [Turk and Pentland. 1991], Fisherface 185
(i.e. LDA) [Belhumeur et al. 1997], LBP [T. Ahonen et al. 2006], Gabor-PCA, Gabor-LDA and Gabor-LBP as 186
the baseline algorithms. The Gabor-LDA and Gabor-LBP are among the best FR methods and they are 187
benchmarks to evaluate the performance of FR techniques. 188
10
4.1 Experiment 1 189
In Exp#1, we set up a subset from the whole PolyU-NIRFD database. In this subset, the training set contains 190
419 frontal images from 138 subjects, while the gallery set and probe set have 574 and 2763 images 191
respectively. No images in the probe and gallery sets are contained in the training set. The facial portion of each 192
original image is automatically cropped according to the location of the eyes. The cropped face is then 193
normalized to 64 64× pixels. The eight methods are then applied to this dataset to evaluate their performance. 194
For subspace based methods, the distance used in the nearest neighbor classifier is the cosine of the angle 195
between two feature vectors. For LBP histogram features, we use the histogram intersection similarity measure. 196
The sub-region size and the number of histogram bins for LBP are 8x8 and 32. Since the Gabor filtering is 197
performed on five scales and eight orientations, we need to downsample the response image to reduce data 198
amount. The downsampling factor is set to 4. Therefore, for Gabor-PCA and Gabor-LDA, the input signal size 199
is 264 64 5 8 4× × × ÷ =10240. 200
The FR results by the eight methods are illustrated in Figures 7 and 8. For subspace based methods PCA, 201
LDA, Gabor-PCA and Gabor-LDA, the curves of recognition rate versus feature vector dimensionality are 202
plotted in Figure 7. We see that the curves for PCA and LDA are flat when the dimension of feature vector 203
changes from 60 to 120. In this experiment PCA gets similar performance to LDA because the number of 204
training samples for each class is limited. From Figures 7 and 8 we can clearly see that using Gabor features can 205
improve greatly the performance of all the four methods. For example, the recognition rate of Gabor-LDA is 206
10% higher than that of LDA. Comparing Figure 7 with Figure 8, it can be seen that the LBP method achieve 207
better performance than PCA and LDA. 208
209
11
210 Fig. 7: Recognition results by PCA, LDA, Gabor-PCA and Gabor-LDA in Exp#1. 211
212
213 Fig. 8: Recognition results by LBP, Gabor-LBP in Exp#1. 214
215
4.2 Experiment 2 216
In Exp#2, we extracted from the whole database a much bigger subset than in Exp#1. In this subset, the training 217
set contains 1876 frontal images of 150 subjects, while the gallery and probe sets have 1159 and 4747 images, 218
respectively. The recognition results of PCA, LDA, and LBP are illustrated in Figure 9 and Figure 10. Different 219
from that in Exp#1, LDA achieves much better performance than PCA when using a small number of features. 220
Similar to those in Exp#1, Gabor based methods get much better performances than their original image based 221
12
counterparts. We can also observe that the best result of Gabor-LDA is 92.1%, which is not as good as in exp#1. 222
This is because the dataset in Exp#2 is more challenging by involving more variations of pose, expression and 223
scale than that in Exp#1. We can see that LBP has much better performance than PCA and LDA, which 224
validates again that LBP is effective way to model infrared face images. 225
226
227
Fig. 9: Recognition results by PCA, LDA, Gabor-PCA and Gabor-LDA in Exp#2. 228
229
230
Fig. 10: Recognition results by LBP, Gabor-LBP in Exp#2. 231
232
233
13
4.3 Experiment 3 234
Exp#3 is designed to evaluate the performance of the algorithms on the large variations of pose, expression, 235
illumination and scale, etc. In this subset, training set contains 578 images from 188 subjects, while gallery and 236
probe sets have 951 and 3648 images individually. The results are shown in Figures 11 and 12. 237
238
239 Fig. 11: Recognition results by PCA, LDA, Gabor-PCA and Gabor-LDA in Exp#3. 240
241
80
82
84
86
88
90
92
94
8 16 32 64 128
LBP
Gabor-LBP
242
Fig. 12: Recognition results by LBP, Gabor-LBP in Exp#3. 243
244
245
14
5. CONCLUSIONS 246
We introduced in this paper the establishment of PolyU near infrared face database (PolyU-NIRFD), which is 247
one of the largest NIR face databases so far. The main characteristics of the PolyU-NIRFD lie in two aspects. 248
First is its large scale. It consists of 35,000 images from 350 subjects. Second is its diversity of variations. It 249
includes variations of pose, expression, illumination, scale, blurring and the combination of them. Comparative 250
study of baseline algorithms was performed on the PolyU-NIRFD to verify its effectiveness. In the future we 251
will acquire a larger database under both controlled and uncontrolled environment, and perform more 252
experiments to investigate more effective NIR FR algorithms. 253
254
6. OBTAINING THE POLYU-NIRFD 255
The PolyU-NIRFD will be publically available soon. The information about how to obtain a copy of the 256
database will be found on the website (http://www.comp.polyu.edu.hk/~biometrics/polyudb_face.htm.) 257
258
ACKNOWLEDGMENT 259
The work is supported by the GRF grant of HKSAR, the central fund from Hong Kong Polytechnic 260
University, the Natural Science Foundation of China (NSFC) under Contract No. 60620160097, and 261
the National High-Tech Research and Development Plan of China (863) under Contract No. 2006AA01Z193 262
and 2007AA01Z195. 263
REFERENCES 264
R. Chellappa, C.L. Wilson, and S. Sirohey, 1995. Human and Machine Recognition of Faces: A Survey. 265 Proceedings of the IEEE, vol. 83, no. 5, pp. 705-740. 266 267 W. Zhao, R. Chellappa, and A. Rosenfeld, 2003. Face recognition: A literature survey. ACM Computing 268 Surveys, 35:399–458. 269 270
15
M.A. Turk and A.P. Pentland, 1991. Face Recognition Using Eigenfaces. Proceedings of IEEE Conference on 271 Computer Vision and Pattern Recognition, pp. 586-591. 272 273 P. Belhumeur, J. Hespanha, and D. Kriegman, 1997. Eigenfaces vs fisherfaces : Recognition using class specific 274 linear projection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19:711–720. 275 276 P.J. Phillips, H. Moon, S.A. Rizvi, and P.J. Rauss, 2000. The FERET Evaluation Methodology for Face-277 Recognition Algorithms. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 22, no. 10, pp. 278 1090-1104. 279 280 P. J. Phillips, P. Grother, R. J. Micheals, D. M. Blackburn, E. Tabassi, and J. M. Bone, 2003. Face recognition 281 vendor test 2002: Evaluation report. Nat. Inst. Standards Technol., Gaithersburg, MD, Tech. Rep. NISTIR 6965. 282 283 P.J. Philips, P.J. Flynn, T. Scruggs, K.W. Bowyer, Jin Chang, Kevin Hoffman, Joe Marques, Jaescik Min, 284 Willimam Worek, 2005. Overview of the face recognition grand challenges. in proc. of ICCV, pp.947-954. 285 286 W. Gao, B. Cao, S. Shan, X. Chen, D. Zhou, X. Zhang, D. Zhao, 2008. the CAS-PEAL Large-Scale Chinese 287 Face Database and baseline evaluation. IEEE Transactions on Systems, Man and Cybernetics, Part A, pp.149-288 161. 289 290 Enrique Bailly-Bailliére, Samy Bengio, Frédéric Bimbot, Miroslav Hamouz, Josef Kittler, Johnny Mariéthoz, 291 Jiri Matas, Kieron Messer, Vlad Popovici, Fabienne Porée, Belen Ruiz and Jean-Philippe Thiran, 2003. The 292 banca database and evaluation protocol. In International Conference on Audio-and Video-based Person 293 Authentication. 294 295 T. Sim, S. Baker, and M. Bsat, 2003. The CMU pose, illumination, and expression database. IEEE Transactions 296 on Pattern Analysis and Machine Intelligence. vol. 25, no. 12, pp. 1615–1618. 297 298 K. Messer, J. Matas, J. Kittler, J. Luettin, and G. Maitre, 1999. XM2VTSDB: The extended M2VTS database. 299 in Proc. Int. Conf. Audio Video-Based Biometric Person Authentication, pp. 72–77. 300 301 A. M. Martinez and R. Benavente, 1998. The AR face database. CVC, Barcelona, Spain, Tech. Rep. 24. 302 303 A. S. Georghiades, P. N. Belhumeur, and D. J. Kriegman, 2001. From few to many: Illumination cone models 304 for face recognition under variable lighting and pose. IEEE Transactions on Pattern Analysis and Machine 305 Intelligence. vol. 23, no. 6, pp. 643–660. 306 307 K. C. Lee, J. Ho, and D. J. Kriegman, 2005. Acquiring linear subspaces for face recognition under variable 308 lighting. IEEE Transactions on Pattern Analysis and Machine Intelligence. vol. 27, no. 5, pp. 684–698. 309 310 X. Zou, J. Kittler, and K. Messer, 2005. Face Recognition Using Active Near-IR Illumination. Proc. British 311 Machine Vision Conf.. 312 313 Stan Li, Rufeng Chu, Shengcai Liao, and Lun Zhang, 2007. Illumination Invariant Face Recognition Using 314 Near-Infrared Images. IEEE Transactions on Pattern Analysis and Machine Intelligence. vol.29, no.4, pp. 627-315 639 316 317 Z.H. Pan, G. Healey, M. Prasad, and B. Tromberg, 2003. Face Recognition in Hyperspectral Images. IEEE 318 Transactions on Pattern Analysis and Machine Intelligence. vol. 25, no. 12, pp. 1552-1560. 319 320
16
J. Dowdall, I. Pavlidis, and G. Bebis, 2003. Face detection in the near-ir spectrum. Image Vision and 321 Computing. vol. 21, pp.565–578. 322 323 A. S. Georghiades, P. N. Belhumeur, and D. J. Kriegman, 2001. From few to many: Illumination cone models 324 for face recognition under variable lighting and pose. IEEE Transactions on Pattern Analysis and Machine 325 Intelligence, vol.23(6), pp.643–660. 326 327 T. Ojala, M. Pietikäinen, and T. Mäenpää, 2002. Multiresolution Gray-Scale and Rotation Invariant Texture 328 Classification with Local Binary Patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence. vol. 329 24, no. 7, pp. 971-987. 330 331 T. Ahonen, A. Hadid, and M. Pietikäinen, 2006. Face Description with Local Binary Patterns: Application to 332 Face Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 28, no. 12, pp. 2037-333 2041. 334 335 L. Wiskott, J.M. Fellous, N. Krüger, and C. von der Malsburg, 1997. Face Recognition by Elastic Bunch Graph 336 Matching. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 19, no. 7, pp. 775-779. 337 338 C. Liu and H. Wechsler, 2002. Gabor Feature Based Classification Using the Enhanced Fisher Linear 339 Discriminant Model for Face Recognition. IEEE Transaction on Image Processing, vol. 11, no. 4, pp. 467-476. 340 341 Y. Gao and M.K.H. Leung, 2002. Face Recognition Using Line Edge Map. IEEE Transactions on Pattern 342 Analysis and Machine Intelligence, vol. 24, no. 6, pp. 764-779. 343