Learning Discriminative LBP-Histogram Bins

for Facial Expression Recognition

Caifeng Shan and Tommaso Gritti
Philips Research, High Tech Campus 36, Eindhoven 5656 AE, The Netherlands
{[Link], [Link]}@[Link]

Abstract

Local Binary Patterns (LBP) have been well exploited for facial image anal-
ysis recently. In the existing work, the LBP histograms are extracted from lo-
cal facial regions, and used as a whole for the regional description. However,
not all bins in the LBP histogram are necessary to be useful for facial repre-
sentation. In this paper, we propose to learn discriminative LBP-Histogram
(LBPH) bins for the task of facial expression recognition. Our experiments
illustrate that the selected LBPH bins provide a compact and discriminative
facial representation. We experimentally illustrate that it is necessary to con-
sider multiscale LBP for representing faces, and most discriminative infor-
mation is contained in uniform patterns. By adopting SVM with the selected
multiscale LBPH bins, we obtain the best recognition performance of 93.1%
on the Cohn-Kanade database.

1 Introduction
Machine analysis of facial expressions, enabling computers to analyze and interpret facial
expressions as humans do, has many applications such as human-computer interaction and
computer animation; so it has attracted much attention in last two decades [9, 3].
A vital step for successful facial expression analysis is deriving an effective facial rep-
resentation from original face images. Two types of features [15], geometric features and
appearance features, are usually considered for facial representation. Geometric features
deal with the shape and locations of facial components (including mouth, eyes, brows,
and nose), which are extracted to represent the face geometry [16]. Appearance features
present the appearance changes (skin texture) of the face (including wrinkles, bulges and
furrows), which are extracted by applying image filters to either the whole face or specific
facial regions [6]. The geometric features based facial representations commonly require
accurate and reliable facial feature detection and tracking, which is difficult to accommo-
date in real-world unconstrained scenarios, e.g., under head pose variation. In contrast,
appearance features suffer less from issues of initialization and tracking errors, and can
encode changes in skin texture that are critical for facial expression modeling. However,
most of the existing appearance-based facial representations still require face registration
based on facial feature detection, e.g., eye detection.
As an efficient non-parametric method summarizing the local structure of an image,
Local Binary Patterns (LBP) has been introduced for facial representation recently [1, 13]
(see Section 2 for details about LBP). The most important properties of LBP features
Figure 1: A face image is divided into sub-regions from which LBP histograms are ex-
tracted and concatenated into a single, spatially enhanced feature histogram.

Figure 2: The sub-regions selected by AdaBoost for each facial expression. From left to
right: Anger, Disgust, Fear, Joy, Sadness, and Surprise.

are their tolerance against monotonic illumination changes and their computational sim-
plicity. In the original LBP-based facial representation [1, 13], as shown in Figure 1,
face images are first equally divided into non-overlapping sub-regions to extract the LBP
histograms within each sub-region, which are then concatenated into a single, spatially
enhanced feature histogram. Possible criticisms of this method are that dividing the face
into a grid of sub-regions is somewhat arbitrary, as sub-regions are not necessary well
aligned with facial features, and that the resulting facial representation suffers from fixed
size and position of sub-regions. To address these, in [18, 12], by shifting and scaling a
sub-window over face images, many more sub-regions are obtained, and then Adaboost
[11] is adopted to select the most discriminative sub-regions in term of LBP histogram.
Figure 2 shows the selected sub-regions for each facial expression.
In most of the existing work, LBP histograms are extracted from local facial regions
as the region-level description, where the n-bin histogram is utilized as a whole. How-
ever, not all bins in the LBP histogram are necessary to contain useful information for
facial representation. It is helpful and interesting to have a closer look at the local LBP
histogram at the bin level, to identify the discriminative LBP-Histogram (LBPH) bins for
better facial representation. To our best knowledge, this problem has not been investi-
gated in the existing work. In this paper, we propose to learn discriminative LBPH bins
for the task of facial expression recognition. Adaboost (Section 3) is adopted to learn LBP
features at the bin level. Our experiments (Section 4) illustrate that the selected LBPH
bins provide a much more compact facial representation, reducing feature length greatly,
while producing better facial expression recognition performance. We experimentally
verify the validity of uniform patterns for facial representation from the point view of ma-
chine learning. We also evidently show that it is necessary to consider multiscale LBP for
facial representation. By adopting Support Vector Machine (SVM) with the selected mul-
tiscale LBPH bins, we obtain the recognition performance of 93.1% on the Cohn-Kanade
database, which is comparable to the best performance reported so far on the database.

Related Work — As a powerful means of texture description, LBP features have been
widely exploited in many applications (see a comprehensive bibliography related to LBP
methodology online1 ). For facial image analysis, LBP features have been extensively
exploited recently. Rodriguez and Marcel [10] proposed a generative approach for face
authentication, which considers local LBP histograms as probability distributions and
represents a generic face model by a collection of LBP histograms. Chan et al. [2]
presented to extract multi-scale LBP histograms from each local regions, which has shown
promising performance for face recognition. Liao et al. [5] proposed Multi-scale Block
LBP for face recognition, in which the computation is done based on average values
of block sub-regions, instead of individual pixels. Recently Zhao and Pietikäinen [19]
introduced the volume LBP and LBP from three orthogonal planes for dynamic texture
recognition, showing promising performance on dynamic facial expression recognition
by combining appearance and motion. More recently Tan and Triggs [14] introduced an
extension of LBP called Local Ternary Patterns (LTP) that is less sensitive to noise in
near-uniform regions.

2 Local Binary Patterns

The original LBP operator, introduced by Ojala et al. [7], labels the pixels of an image
by thresholding a 3 × 3 neighborhood of each pixel with the center value and considering
the results as a binary number. Formally, given a pixel at (xc , yc ), the resulting LBP can
be expressed in the decimal form as
7
LBP(xc , yc ) = ∑ s(in − ic )2n (1)
n=0

where n runs over the 8 neighbors of the central pixel, ic and in are the gray-level values
of the central pixel and the surrounding pixel, and s(x) is 1 if x ≥ 0 and 0 otherwise.
Ojala et al. [8] later made two extensions of the original operator. Firstly, the oper-
ator was extended to use neighborhood of different sizes, to capture dominant features
at different scales. Using circular neighborhoods and bilinearly interpolating the pixel
values allow any radius and number of pixels in the neighborhood. The notation (P, R)
denotes a neighborhood of P equally spaced sampling points on a circle of radius of R.
Secondly, they proposed to use a small subset of the 2P patterns, produced by the opera-
tor LBP(P, R), to describe the texture of images. These patterns, called uniform patterns,
contain at most two bitwise transitions from 0 to 1 or vice versa when considered as a
circular binary string. For example, 00000000, 001110000 and 11100001 are uniform
patterns. The uniform patterns represent local primitives such as edges and corners. It
was observed that most of the texture information was contained in the uniform patterns.
Labeling the patterns which have more than 2 transitions with a single label yields an
LBP operator, denoted LBP(P, R, u2), which produces much less patterns without losing
too much information.
After labeling an image with a LBP operator, a histogram of the labeled image fl (x, y)
can be defined as
Hi = ∑ I( fl (x, y) = i), i = 0, . . . , L − 1 (2)
x,y

1 [Link] bibliography
where L is the number of different labels produced by the LBP operator and I(A) is 1 if A
is true and 0 otherwise.

LBP based Facial Representation — Each face image can be seen as a composition of
micro-patterns which can be effectively detected by the LBP operator. Ahonen et al. [1]
introduced a LBP based face representation for face recognition. To consider the shape
information of faces, they divided face images into M small non-overlapping regions
R0 , R1 , . . . , RM (as shown in Figure 1). The LBP histograms extracted from each sub-
region are then concatenated into a single, spatially enhanced feature histogram defined
as:
Hi, j = ∑ I( fl (x, y) = i)I((x, y) ∈ R j ) (3)
x,y
where i = 0, . . . , L − 1, j = 0, . . . , M − 1. The extracted feature histogram describes the
local texture and global shape of face images. The face representation has also been
proved effective for facial expression recognition [13].
To address the limitations of the above scheme including arbitrary sub-region divi-
sion and fixed size/position of sub-regions, Adaboost later was adopted to learn the most
discriminative sub-regions (in term of LBP histogram) from a large pool of sub-regions
generated by shifting and scaling a sub-window over face images [18, 12]. Some exam-
ples of selected sub-regions for facial expressions are shown in Figure 2. In the existing
work, the LBP histograms are always extracted from sub-regions, and used as a whole
for the regional description. However, not all bins in the LBP histogram are discrimi-
native for facial representation. In the next section, we propose to learn discriminative
LBP-Histogram (LBPH) bins for better facial representation.

3 Learning Discriminative LBP-Histogram Bins

We adopt Adaboost to learn the discriminative LBPH bins. Adaboost [11] learns a small
number of weak classifiers whose performance is just better than random guessing, and
boosts them iteratively into a strong classifier of higher accuracy. The process of Ad-
aboost maintains a distribution on the training samples. At each iteration, a weak classi-
fier which minimizes the weighted error rate is selected, and the distribution is updated to
increase the weights of the misclassified samples and reduce the importance of the others.
Since Viola and Jones [17] introduced the Adaboost-based face detector, Adaboost has
been successfully used in many computer vision problems. For example, Adaboost was
adopted to select Gabor wavelet features [6] and spatial-temporal features [16] for facial
expression recognition.
As the traditional Adaboost works on two-class problems, the multi-class problem
here is accomplished by using the one-against-rest technique, which trains Adaboost be-
tween one expression with all others. For each Adaboost learner, the images of one ex-
pression were positive samples, while the images of all other expressions were negative
samples. The weak classifier is designed to select the single LBPH bin which best sepa-
rates the positive and negative examples. Similar to [17], a weak classifier h j (x) consists
of a feature f j which corresponds to each LBPH bin, a threshold θ j and a parity p j indi-
cating the direction of the inequality sign:
½
1 if p j f j (x) ≤ p j θ j
h j (x) = (4)
0 otherwise
4 Experiments
We carry out experiments on the Cohn-Kanade database [4], one of the most comprehen-
sive databases in the current facial expression research community. The Cohn-Kanade
database consists of 100 university students aged from 18 to 30 years, of which 65%
were female, 15% were African-American, and 3% were Asian or Latino. Subjects were
instructed to perform a series of 23 facial displays, six of which were based on descrip-
tion of basic emotions (i.e., Anger, Disgust, Fear, Joy, Sadness, and Surprise). Image
sequences from neutral to target display were digitized into 640×490 pixel arrays with 8-
bit precision for grayscale values. For our experiments, we selected 320 image sequences
of basic emotions from the database. The sequences come from 96 subjects, with 1 to
6 emotions per subject. For each sequence, the neutral face and three peak frames were
used for prototypic expression recognition, resulting in 1,280 images (108 Anger, 120
Disgust, 99 Fear, 282 Joy, 126 Sadness, 225 Surprise, and 320 Neutral). To test the al-
gorithms’ generalization performance, we adopt 10-fold cross-validation scheme in our
experiments, reported the average recognition results.
Following the existing work [13, 10], we scaled the faces to a fixed distance between
the two eyes. Facial images of 110×150 pixels were cropped from original frames based
on the two eyes location.

4.1 Experiment: Limited Sub-regions

In [13], face images of 110×150 pixels were divided into 42 (6×7) sub-regions of 18×21
pixels (as shown in the bottom right corner of Figure 4), and then the 59-label LBP(8, 2, u2)
operator was adopted to extract LBP features. These parameter settings were suggested
in [1]. We started our experiments with the same setting. Thus each face image was
described by a LBP histogram of 2,478 (42×59) bins. We adopted Adaboost to learn
discriminative LBPH bins and boost a strong classifier. We plot in the left side of Figure
3 the recognition performance of the boosted strong classifier as a function of the num-
ber of features selected. With the 200 selected LBPH bins, the boosted strong classifier
achieves recognition rate of 85.3%, which is much better than 79.1% produced by Tem-
plate Matching using all 2,478 bins in [13].

Spatial Distribution of Features Selected — Figure 4 shows the spatial distribution of

the top 200 features selected in the 10-fold cross-validation experiments. As can be ob-
served, different facial expressions have different distribution patterns. For example, for
“Disgust”, most discriminative features locate in the eye inner corners, while most dis-
criminative features for “Joy” are distributed in the mouth corners, which indeed reenforce
human observation. Overall, discriminative features for facial expression classification
mostly distribute in eyes and mouth regions.

Uniform Patterns — It was observed that most of the texture information was con-
tained in the uniform patterns [8], so uniform patterns was used to reduce the length
of LBP histograms. Based on this, the 59-label LBP(8, 2, u2) operator, instead of 256-
label LBP(8, 2), was widely used for facial representation. Here we verify the validity
of uniform patterns for facial representation from a point view of machine learning. By
using the LBP(8, 2) operator, each face image was represented by a LBP histogram of
10,752 (42×256) bins. We plot in the left side of Figure 3 the recognition performance
 0.9 0.9

0.85 0.85

0.8 0.8
LBP(8,2,u2) LBP(8, MultiScale, u2)
LBP(8,2) LBP(8,2,u2)
Average Recognition Rate

Average Recognition Rate

0.75 0.75

0.7 0.7

0.65 0.65

0.6 0.6

0.55 0.55

0.5 0.5
0 20 40 60 80 100 120 140 160 180 200 0 20 40 60 80 100 120 140 160 180 200
Number of Features NUmber of Features

Figure 3: Average recognition rate of the boosted strong classifiers, as a function of the
number of feature selected. Left: LBP(8, 2, u2) vs LBP(8, 2); Right: LBP(8, 2, u2) vs
LBP(8, Multiscale, u2).

of the boosted strong classifier. We can see that the boosted strong classifier of LBP(8, 2)
performs similarly with that of LBP(8, 2, u2), which illustrates that the non-uniform pat-
terns do not provide more discriminative information for facial expression recognition.
To further verify this, we took a closer look at the learned LBPH bins of LBP(8, 2), and
found that 91.1% of them are uniform patterns. Therefore, with boost learning we exper-
imentally verify that most of discriminative information for facial expression recognition
was contained in the uniform patterns.

Multi-scale LBP — By varying the sampling radius R, LBP of different resolutions can
be obtained. The multiscale LBP has provided better performance than single scale LBP
for texture classification [8] and face recognition [2]. Here we also investigate multi-
scale LBP for facial expression recognition. We applied the LBP(8, R, u2)(R = 1, · · · , 8)
to extract multiscale LBP features, resulting a LBP histogram of 19,824 (42×59×8) bins
for each face image. We then run Adaboost to learn discriminative LBPH bins from
the multiscale feature pool. We plot in the right side of Figure 3 the recognition perfor-
mance of the boosted strong classifier. As can be observed, the boosted strong classifier of
multiscale LBP(8, R, u2)(R = 1, · · · , 8) produces consistently better performance than that
of single scale LBP(8, 2, u2), providing recognition rate of 88.6% with the 200 selected
LBPH bins. Thus the multiscale LBP brings more discriminative information for facial
expression recognition. Figure 5 shows the scale distribution of the selected LBPH bins
in the 10-fold cross-validation experiments. We can see that the scale distribution of fea-
tures selected for different expressions are different. Overall, discriminative LBPH bins
distribute at all scales, especially scales R = 3, 4, 6, 7, 8. Our experimental results suggest
that multiscale LBP features should be considered for facial expression recognition.
We summarize our experiment results in Table 1, where we also include experimental
results reported in [13]. As can be observed, the boosted strong classifiers with 200
selected LBPH bins outperform the template matching method using all 2,478 bins [13].
The boosted strong classifier with 200 multiscale LBPH bins provides comparable result
to the SVM classifier using all 2,478 bins in [13].
 Feature Distribution: ANGER Feature Distribution: DISGUST Feature Distribution: FEAR

150 400 200

300 150
100
200 100
50
100 50

0 0 0

1 1 1
2 2 2
3 3 3
4 6 4 6 4 6
5 5 5 5 5 5
4 4 4
6 3 6 3 6 3
7 2 7 2 7 2
1 1 1

Feature Distribution: JOY Feature Distribution: SADNESS Feature Distribution: SURPRISE

400 200 200

300 150 150

200 100 100

100 50 50

0 0 0

1 1 1
2 2 2
3 3 3
4 6 4 6 4 6
5 5 5 5 5 5
4 4 4
6 3 6 3 6 3
7 2 7 2 7 2
1 1 1

Feature Distribution: NEUTRAL Feature Distribution: ALL EXPRESSIONS

200 1500

150
1000
100
500
50

0 0

1 1
2 2
3 3
4 6 4 6
5 5 5 5
4 4
6 3 6 3
7 2 7 2
1 1

Figure 4: Spatial distribution of the selected LBPH bins (an example face image divided
in sub-regions is included in the bottom right corner for illustration).

4.2 Experiment: More Sub-regions

In the above experiments, only 42 equally divided sub-regions were considered for fea-
ture selection. By shifting and scaling a sub-window over face images, we can get many
more sub-regions, which potentially contain more complete and discriminative informa-
tion about face images. We shifted the sub-window with the shifting step of 14 pix-
els vertically and 12 pixels horizontally. The sub-window was scaled as 14, 21, or 28
pixels (height) and 12, 18, or 24 pixels (width) respectively. In total 725 sub-regions
were obtained. By using multiscale LBP(8, R, u2)(R = 1, · · · , 8), a histogram of 342,200
(725×59×8) bins was extracted from each face image.
To improve computation efficiency, we adopted a coarse to fine feature selection
scheme: We first run Adaboost to select LBPH bins from each single scale LBP(8, R, u2),
then applied Adaboost to the selected LBPH bins at different scales to obtain final feature
selection results. We plot in the left side of Figure 6 the recognition performance of the
boosted strong classifiers as a function of the number of features selected. We can see
 Feature distribution at different scales
450 2500

Anger
Disgust
400
Fear
Joy
Sadness 2000
350
Surprise
Neutral

300

1500
250

200
1000

150

100
500

50

0 0
1 2 3 4 5 6 7 8 1 2 3 4 5 6 7 8
Scale Scale

Figure 5: Scale distribution of the selected LBPH bins. Left: distribution of each class;
right: overall distribution.

Methods Recognition Rate

LBP(8, 2, u2) + Adaboost 85.3%
LBP(8, 2) + Adaboost 85.9%
LBP(8, Multiscale, u2) + Adaboost 88.6%
LBP(8, 2, u2) + Template Matching [13] 79.1%
LBP(8, 2, u2) + SVM (Polynomial) [13] 88.4%

Table 1: Recognition performance of different methods using LBP features extracted from
equally divided sub-regions.

that the final boosted strong classifier of multiscale LBP provides better performance than
that of each single scale. Among strong classifiers of single scales, it seems that scales
(R = 3, 4, 5, 6) perform better, while the performance of scales (R = 1, 8) is poor.

Feature Distribution — The scale distribution of final selected multiscale LBPH bins is
shown in the right side of Figure 6. We can observe that most discriminative LBPH bins
come from scales (R = 3, 4, 5, 8). The scale distribution is different from that we obtained
in Figure 5, and this is because the features were selected from many more sub-regions.
Regarding the spatial distribution of final selected features, Figure 7 shows the top 20
sub-regions that contain most LBP bins selected for each facial expression. We can see
that each facial expression has its unique spatial distribution of selected features. As we
discussed before, most of discriminative features distribute in eyes and mouth regions.

SVM classification — Finally we adopted SVM to recognize facial expressions using the
selected LBPH bins. In [6], SVM using Gabor features selected by Adaboost (AdaSVM)
achieves the best performance (93.3%) reported so far on the Cohn-Kanade database. We
compare our LBP-based methods with the Gabor-based methods [6] in Table 2. We used
the SVM implementation in the library SPIDER1 , and the multi-class classification was
1 Public available at [Link]
 2500
0.9

0.85
2000

LBP(8, 1, u2)
0.8 LBP(8, 2, u2)
Average Recognition Rate

LBP(8, 3, u2)
LBP(8, 4, u2) 1500
0.75 LBP(8, 5, u2)
LBP(8, 6, u2)
LBP(8, 7, u2)
0.7
LBP(8, 8, u2)
LBP(8, Multiscale, u2)
1000

0.65

500
0.6

0.55 0
0 20 40 60 80 100 120 140 160 180 200 1 2 3 4 5 6 7 8
Number of Features

Figure 6: Left: Average recognition rate of boosted strong classifiers, as a function of the
number of feature selected. Right: Scale distribution of the selected LBPH bins

Figure 7: The top 20 sub-regions that contain most LBPH bins selected. From left to
right: Anger, Disgust, Fear, Joy, Sadness, Surprise, and Neutral.

accomplished by using the one-against-rest technique. It can be observed that the boosted
LBPH bins produce comparable results with the boosted Gabor features [6].

Features Recognition Rates

Adaboost AdaSVM(Linear) AdaSVM(RBF)
LBP 91.3% 93.0% 93.1%
Gabor[6] 90.1% 93.3% 93.3%

Table 2: Comparison between the boosted LBPH bins and Gabor wavelet features.

5 Conclusions
In this paper, we propose to learn discriminative LBP-Histogram (LBPH) bins for the
task of facial expression recognition using Adaboost. Our experiments illustrate that the
selected LBPH bins provide a compact and discriminative facial representation. We ex-
perimentally verify the validity of uniform patterns for facial representation. We also
evidently illustrate that it is necessary to consider multiscale LBP. By adopting SVM with
the selected multiscale LBPH bins, we obtain the recognition performance of 93.1% on
the Cohn-Kanade database.
References
[1] T. Ahonen, A. Hadid, and M. Pietikäinen. Face recognition with local binary patterns. In
European Conference on Computer Vision (ECCV), pages 469–481, 2004.
[2] C. Chan, J. Kittler, and K. Messer. Multi-scale local binary pattern histograms for face recog-
nition. In International Conference on Biometrics (ICB), pages 809–818, 2007.
[3] B. Fasel and J. Luettin. Automatic facial expression analysis: a survey. Pattern Recognition,
36:259–275, 2003.
[4] T. Kanade, J.F. Cohn, and Y. Tian. Comprehensive database for facial expression analysis. In
IEEE International Conference on Automatic Face & Gesture Recognition (FG), pages 46–53,
2000.
[5] S. Liao, X. Zhu, Z. Lei, L. Zhang, and S. Z. Li. Learning multi-scale block local binary
patterns for face recognition. In International Conference on Biometrics (ICB), pages 828–
837, 2007.
[6] G. Littlewort, M. Bartlett, I. Fasel, J. Susskind, and J. Movellan. Dynamics of facial expression
extracted automatically from video. Image and Vision Computing, 24(6):615–625, June 2006.
[7] T. Ojala, M Pietikäinen, and D. Harwood. A comparative study of texture measures with
classification based on featured distribution. Pattern Recognition, 29(1):51–59, 1996.
[8] T. Ojala, M. Pietikäinen, and T. Mäenpää. Multiresolution gray-scale and rotation invariant
texture classification with local binary patterns. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 24(7):971–987, 2002.
[9] M. Pantic and L. Rothkrantz. Automatic analysis of facial expressions: the state of art. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 22(12):1424–1445, 2000.
[10] Y. Rodriguez and S. Marcel. Face authentication using adapted local binary pattern his-
tograms. In European Conference on Computer Vision (ECCV), pages 321–332, 2006.
[11] R. E. Schapire and Y. Singer. Improved boosting algorithms using confidence-rated predic-
tions. Maching Learning, 37(3):297–336, 1999.
[12] C. Shan, S. Gong, and P. W. McOwan. Conditional mutual information based boosting for
facial expression recognition. In British Machine Vision Conference (BMVC), volume 1, pages
399–408, Oxford, UK, September 2005.
[13] C. Shan, S. Gong, and P. W. McOwan. Robust facial expression recognition using local
binary patterns. In IEEE International Conference on Image Processing (ICIP), volume 2,
pages 370–373, Genoa, Italy, September 2005.
[14] X. Tan and B. Triggs. Enhanced local texture feature sets for face recognition under difficult
lighting conditions. In IEEE International Workshop on Analysis and Modeling of Faces and
Gestures (AMFG), pages 168–182, 2007.
[15] Y. Tian, T. Kanade, and J. Cohn. Handbook of Face Recognition, chapter 11. Facial Expression
Analysis. Springer, 2005.
[16] M. Valstar and M. Pantic. Fully automatic facial action unit detection and temporal analysis.
In IEEE Conference on Computer Vision and Pattern Recognition Workshop, page 149, 2006.
[17] P. Viola and M. Jones. Rapid object detection using a boosted cascade of simple features.
In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 511–518,
2001.
[18] G. Zhang, X. Huang, S. Z. Li, Y. Wang, and X. Wu. Boosting local binary pattern (lbp)-based
face recognition. In Chinese Conference on Biometric Recognition, pages 179–186, 2004.
[19] G. Zhao and M. Pietikäinen. Dynamic texture recognition using local binary patterns with
an application to facial expressions. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 29(6):915–928, 2007.