Local Gray Code Pattern (LGCP): A Robust Feature Descriptor for Facial Expression Recognition

International Journal of Science and Research (IJSR), India Online ISSN: 2319-7064
Volume 2 Issue 8, August 2013
www.ijsr.net
Local Gray Code Pattern (LGCP): A Robust
Feature Descriptor for Facial Expression
Recognition
Mohammad Shahidul Islam
Atish Dipankar University of Science & Technology, School, Department of Computer Science and Engineering, Dhaka, Bangladesh.
Abstract: This paper presents a new local facial feature descriptor, Local Gray Code Pattern (LGCP), for facial expression recognition
in contrast to widely adopted Local Binary pattern. Local Gray Code Pattern (LGCP) characterizes both the texture and contrast
information of facial components. The LGCP descriptor is obtained using local gray color intensity differences from a local 3x3 pixels
area weighted by their corresponding TF (term frequency). I have used extended Cohn-Kanade expression (CK+) dataset and Japanese
Female Facial Expression (JAFFE) dataset with a Multiclass Support Vector Machine (LIBSVM) to evaluate proposed method. The
proposed method is performed on six and seven basic expression classes in both person dependent and independent environment.
According to extensive experimental results with prototypic expressions on static images, proposed method has achieved the highest
recognition rate, as compared to other existing appearance-based feature descriptors LPQ, LBP, LBPU2, LBPRI, and LBPRIU2.
Keywords: Expression Recognition, Image Processing, Local Feature, Pattern Recognition, CK+, JAFFE, Computer Vision
1. Introduction
FER (Facial Expression recognition) has gained substantial
importance in the applications of human-computer
interactions (C. Shan et al., 2009) as this is one of the most
effective, natural, and immediate means for human beings to
communicate their emotions and intentions (C. Shan et al.,
2005), (Y. Tian et al., 2003). It has attracted much attention
from behavioral scientists since the work of Darwin in 1872.
Expression recognition with high accuracy remains difficult
due to ethnicity and variation of facial expressions (G. Zhao
et al., 2009), though many works have been done with
automatic facial expression analysis. Extracting relevant
features from human face images is very important for any
successful facial expression recognition system. The
extracted features should retain essential information having
high discrimination power and stability which minimizes
within-class differences of expressions whilst maximizes
between-class differences (C. Shan et al., 2005).
1.1 Motivation
Facial features can be two types- global feature and local
feature. Global feature is a feature, which is extracted from
the whole face whereas local feature considers small local
region from the whole face. Some global feature extraction
methods are PCA (Principal component analysis), LDA
(Linear Discriminant Analysis) etc. Even they are popular
and widely used but their performance fluctuates with the
environment. Therefore, I have chosen local feature
methodology for feature extraction as it is robust in
uncontrolled environment. Some popular local feature
descriptors are Gabor filter, Local binary pattern, Local
phase quantization etc. Facial feature representations using
Gabor filter is time and memory intensive. S. Lajevardi et al.
2012 solved some limitations of Gabor-filter using log-
Gabor filter but the dimensionality of resulting feature vector
was still high. Local binary pattern is popular but sensitive to
non- monotonic illumination variation and shows poor
performance in the presence of random noise (T. Jabid et al.,
2010). LPQ (V. Ojansivu et al., 2008) is also very time and
memory expensive. Keeping all these sensitive issues in
mind, I have proposed LGCP, which overcomes almost all
those weaknesses.
1.2 Paper Review
Some surveys of existing research on facial expression
analysis can be found in B. Fasel et al., 2003 and M. Pantic
et al., 2000. Three types of facial feature extraction
approaches are there: the geometric feature-based system (Y.
L. Tian et al., 2001), the appearance-based system (Y. Tian
et al., 2003) and hybrid, which uses both the approaches.
Geometric feature vectors represent the shapes and spatial
locations of facial parts by encoding the face geometry from
the location, distance, angle, and other geometric relations
between these parts. A most commonly used facial descriptor
is the facial action coding system (FACS), in which, facial
muscle movements are encoded by 44 Action Units(AUs) (P.
Ekman, 1978). Y.Zhang et al. (2005) proposed IR
illumination camera for facial feature detection, tracking and
recognized the facial expressions using Dynamic Bayesian
networks (DBNs). M.Yeasin et al. (2006) used discrete
hidden Markov models (DHMMs) to recognize the facial
expressions. Z. Zhang et al. (1998) used 34 fiducial points as
facial features to present a facial image. Y.L. Tian et al.
(2001) proposed a multi-state face component model of AUs
and neural network (NN) for classification. I. Cohen et al.
(2003) employed Naive–Bayes classifiers and hidden
Markov models (HMMs) together to recognize human facial
expressions from video sequences. M.F. Valstar et al. (2005,
2006) used several fiducial points on face and mentioned that
geometric approaches are better in feature extraction than
appearance-based approaches. Geometric feature-based
methods need exact and accurate facial components
detection, in many situations, which is not possible (C. Shan
et al., 2009). Recent psychological research concluded that
spatial relations of the facial features from the full face could
be a good source of information for facial emotions (M.
Meulders et al., 2005, M. Zia Uddin et al., 2009). Principal
Component Analysis (PCA) is a holistic method widely used
to extract features from faces (A. Turk et al., 1991). PCA is
also very useful in reducing feature dimension. Lately,
413

www.ijsr.net
Independent Component Analysis (ICA) (M.S. Bartlett et al.,
2005) and Enhanced ICA (EICA) are used for feature
extraction. G. L. Donato et al. (1999) did a comprehensive
analysis of different techniques, including PCA, ICA, Local
Feature Analysis (LFA), Gabor-wavelet and local Principal
Components (PCs). Since then Gabor-wavelet representation
is widely used for feature extraction. However, the size of
filter bank needed for Gabor filters to extract useful
information from a face makes the Gabor representation time
and memory intensive. Afterwards log-Gabor filters
proposed by S. Lajevardi et al. (2012) overcame some
limitations of Gabor-filter but the dimensionality of resulting
feature vector was still high. Another popular feature
descriptor is Local Binary Pattern (LBP) (T. Ojala et al.,
2002) and its variants (G. Zhao et al., 2009). It is widely
used in many research works (C. Shan et al., 2005), (C. Shan
et al., 2009). Originally, LBP was introduced for texture
analysis. LBP labels each pixel of an image by thresholding
it’s P neighbors gray color intensity with the center gray
color intensity and derives a binary patter using Equation (1),
1Where gc and gp are the gray color intensity of the center
pixel (i, j) and p neighboring pixel respectively. A
comprehensive study of LBP can be found in T. Ojala et al.
(1996). Later he observed that binary patterns with less
transition from 1 to 0 and 0 to 1 occur more frequently in a
facial image. Therefore, patterns having more than two
transitions are discarded in LBPU2. He also proposed rotation
invariant LBP (LBPRI), in which he made a bit wise right
shift, eight times for a 8-bit binary pattern. He counted all
eight shifted patterns as a single bin (T. Ojala et al., 2002).
This method is widely adopted by many researchers, but it is
sensitive to non-monotonic illumination variation and shows
poor performance in the presence of random noise (T. Jabid
2010). To overcome this problem, T.Jabid et al. (2010)
proposed a facial descriptor, named as Local Directional
Pattern (LDP), which is more robust than LBP. LDP is
derived from the edge responses, which are less sensitive to
illumination changes and noises. H.Kabir et al. (2012)
extended LDP to LDPv by applying weight to the feature
vector using local variance and found it to be more affective
for facial expression recognition. V. Ojansivu et al. (2008)
proposed LPQ (Local Phase Quantization), a facial feature
extraction method that is blur insensitive. J. Li et al (2012)
extended LPQ to RI-LPQ along with SRC (Sparse
Representation-based Classification) classifier and obtained
better accuracy than LBP. K. Anderson et al. (2006) used the
multichannel gradient model (MCGM) to determine facial
optical flow. The motion signatures achieved and then
classified using Support Vector Machines.
1.3 Contribution
A novel feature extraction technique, LGCP (Local Gray
Code Pattern) is proposed, which characterizes both contrast
and texture information for more accurate facial expression
recognition performance. Figure 1 shows an overall flow of
the expression recognition system based on proposed LGCP
descriptor coupled with LIBSVM. The performance of
proposed method is evaluated using Multiclass Support
Vector Machine (LIBSVM) (C.C. Chang et al., 2011) with
different kernel setup. Proposed method LGCP is more
robust in extracting facial features, and has a better
recognition rate, as compared to LBP, Gabor-wavelet
features, and other appearance-based methods. LGCP
descriptor is stable in presence of noise. The rest of the paper
is organized as follows: The proposed LGCP feature is
described in section 2. Section 3 presents the experimental
setup used for evaluating the effectiveness of the proposed
feature representation, Section 4 lists the expression
recognition performances of LGCP compared with existing
representations. Finally, section 5 concludes the paper.
2. LGCP (Local Gray Code Pattern)
I have followed three steps to compute LGCP code from a
gray scale facial image.
Step 1: I have used Robinson Compass Mask (Robinson
1977) to maximize the edge value, which makes LGCP
descriptor stable in the presence of noise. It is a single mask
in eight major compass orientations: E, NE, N, NW, W, SW,
S, and SE as shown in Figure 2. The edge magnitude = the
maximum value found by the convolution of each mask with
the image. The mask that produces the maximum magnitude
defines the edge direction. The Robinson compass mask
computes the edge response values in all eight directions at
each pixel position and generates a code from the relative
strength magnitude. The obtained 3x3 matrix describes the
local curves, corners, and junctions, more stably and retains
more information. Given a central pixel in the image, the
eight directional edge response values {Ri}, i= 0, 1... 7 are
computed by
�� = � � ∗ ��
�
��
(2)
Where ‘a’ is a local 3x3 pixels region and mi is the Robinson
masks in eight different orientations centered on its position.
Figure 3 shows an example of obtaining edge response using
equation (2).
��,�� = � �� − �� × 2�
; �� = �
1, � < 0
0, � ≥ 0

�
�=1
(1)
Figure 2 : Robinson Edge response masks in eight
directions
Figure 1: Overview of the facial expression recognition
system based on LGCP representation
Gray Scale
Image
Face
Detection
Face Pre-
processing
Feature
Extraction
Feature
Dimension
Reduction
Multiclass
LIBSVM
Expression
Recognition
Apply Robinson
Compass Masks
414

www.ijsr.net
Step 2: The matrix with 8-edge response is then used to get
8-bit gray code pattern as shown in Figure 5. Gray code is a
reflected binary code where only one bit changes at a time.
For an example, in a three bit binary code 001(1) becomes
010(2) by changing two bits, but in case of gray code only
one bit changes at a time, so for gray code ‘1’=001 and
‘2’=011. The advantage of gray code is that it makes very
stable position digitizers, because only one bit changes at a
time, resulting in uncertainty of only one bit.
Step 3: in this step, I have computed the TF (Term
Frequency) of the gray code image using the following
formula, (Equation (3))
TF�bin�, I� =
��, ��
max��
; (3)
Where, ‘I’ is the image and i= 1,2,... ,n (number of bins). For
an example, if I have an image with four possible bins,
bin(1-4) and the bin counting value or histogram is 20, 10,
30, 25, then the TF for bin1 is 20/max(20,10,30,25)=2/3. 8-
bit gray code pattern can produces up-to 256 combinations.
Therefore, I have histogram of 256 bins. I have observed the
histograms of all the images and found only 15 gray code
patterns are mostly responsible for expression classification.
Rests of the patterns are arbitrary. The most prominent
patterns are
0,2,10,22,30,31,64,160,191,224,225,233,245,253 and 254 in
decimal value.
For the experiments, I have reduced the bin number from
256 to 15 by discarding the arbitrary patterns. Therefore, the
length of the TF is also reduced to 15 and the value range
from 1 to 0. After obtaining the term frequency for all
images, I have used it as a weight while building histogram
for each block using Equation (4).
��, �� = ��, �� × ��, �� (4)
Where i=1, 2... 81., x is the non-weighted bin value and y is
the weighted bin value.
LGCP gives more stable patterns in the presence of noise as I
have used Robinson compass masks to intensify the edges.
Figure 4 shows local 3x3 pixels region from an original gray
scale image and the corresponding image after adding
Gaussian white noise.
Figure 4: Noise sensitivity of LGCP, (a) 3x3 pixels region
from original Image, (b) same region with noise
It is clear from the figure that LBP pattern changes with
noise but LGCP remains stable.
3. Experimental Setup
There are several techniques for expression classification.
Among them support vector machine is popular due to its
input output data simplicity. In (C. Shan et al., 2009), the
author conducted experiments using few machine learning
techniques, namely Template matching, Linear Discriminant
Analysis, Linear programming, and Support Vector
Machine. He found SVM to be the best. Therefore, I have
decided to use SVM as the classifier in all through the
experiments. I have used Cohn-Kanade (CK+) (P. Lucey et
al., 2010) and Japanese Female Facial Expression (JAFFE)
start
4 6 8 3 5 7
end 7 4 5 7 5 6
3 2 1 4 3 0
(a) (b)
LBP = 11110001
LGCP= 10111111
LBP = 01110001
LGCP= 10111111
Figure 3: Detailed example of obtaining directional
edge response using Robinson Compass Mask
4 6 8 R1 R2 R3 12 16 12
7 4 5 R8 C R4 == 2 C -2
3 2 1 R7 R6 R5 -12 -16 -12
R1=(a).*m1=7x1+4x2+6x1+3x0+4x0+8x0+2x(-1)+1x(-2)+5x(-1)
R2=(a).*m2=4x1+6x2+8x1+7x0+4x0+5x0+3x(-1)+2x(-2)+1x(-1)
| | | | | | | | | |
R8=(a).*m8=4x1+7x2+3x1+6x0+4x0+2x0+8x(-1)+5x(-2)+1x(-1)
(b) Eight directional
edge response position.
(a) Local 3x3
pixels region
(c ) Obtained
matrix with edge
response from (a)
B1 B2 B3 B4 B5 B6 B7 B8
(8-bit Gray Code pattern)
Figure 5: Example of obtaining Gray code pattern value from 3x3 matrixes with edge response obtained from
Figure 3
12 16 12 R1 R2 R3
2 C -2 R8 C R4 10111111 191 213
-12 -16 -12 R7 R6 R5
(a ) Obtained
matrix with edge
response from
Figure 3(a)
(e) Corresponding
Gray Code
(b) Eight
directional edge
response
position.
(c) Binary
patterrn for
the matrix (b)
(d )
Corresponding
Decimal value
415

International Journal of Science and Researc
(J. Michael 1997) as the facial image datasets.
two datasets to clarify the outcome of
There are 326 posed images in the CK+. No
multiple instances for same expression in CK+. There are 7
prototype expression classes in it. I have not used
expression class for the experiment for this dataset.
However, JAFFE has multiple instances of same expression
from the same subject. It does not have contempt expression
class like CK+. Neutral class replaces the contempt class.
This dataset has total 213 images from 10 different women
with seven expression classes. Figure 6
vector building steps starting from raw image
Figure 6: Facial Feature Extraction
I have first converted all the images to gray scale if they are
in different format. Then I have used fdlibmex face detector
for detecting face area, which is free to use in
There is no publication for this detector. However, for the
experiments I have found it better than manual cropping.
After successfully detecting the face, I have re
detected face areas to 180x180 pixels for CK+ dataset and
99x99 pixels for for JAFFE dataset. I have conducted several
experiments with different combination of dimension and
found the above dimension as optimum one. Then
used an elliptical masking to remove some unnecessary areas
from the head corners and neck corners. The masked face
then divided into 9x9=81 blocks. I have done this to preserve
the spatial information from the facial area.
LGCP to extract feature from each block to build histogram
and concatenated them to obtain the final feature vector.
Therefore the length of the feature vector is, 81 x Number of
bins e.g. 15.
I have followed the above procedure for both training and
testing images. As a classifier, I have used LIBSVM (C.C.
Chang et al., 2011) with 10-fold cross validation and the
folds are not overlapping in a person dependent and
independent environment. The whole dataset is divided into
10 equal folds and each fold is validated against rest 90% of
the images in each fold validation. LIBSVM is a multiclass
classifier with variety of options that can be
input data types. I have set the kernel parameters for the
classifier to: s=0 for SVM type C
linear/polynomial/RBF kernel function respectively, c=100 is
the cost of SVM, g=0.000625 is the value of 1/ (length of
feature vector), b=1 for probability estimation. This setting
for LIBSVM is found to be suitable for CK+ dataset as well
as JAFFE dataset with both six seven classes of data.
4. Results and Analysis
Proposed method is tested against some popular methods
e.g., LBP (Local Binary Pattern) (T. Ojala
LBPRI (Rotation Invariant LBP) (T. Ojala
LBPU2 (Uniform LBP) (T. Ojala et al
(Rotation invariant and Uniform LBP) (T. O
and LPQ (V. Ojansivu et al., 2008), (J. Li
make the comparison more effective, I have used the same
Gray Scale
Image
Face
Detection
Masking
outsidethe
face
Dividing
theface
intoblocks
International Journal of Science and Research (IJSR), India Online ISSN: 2319
www.ijsr.net
facial image datasets. I have used
two datasets to clarify the outcome of proposed method.
There are 326 posed images in the CK+. No person has
multiple instances for same expression in CK+. There are 7
have not used neutral
experiment for this dataset.
However, JAFFE has multiple instances of same expression
subject. It does not have contempt expression
class like CK+. Neutral class replaces the contempt class.
This dataset has total 213 images from 10 different women
6 shows the feature
raw image.
Facial Feature Extraction
have first converted all the images to gray scale if they are
have used fdlibmex face detector
for detecting face area, which is free to use in MATLAB.
this detector. However, for the
have found it better than manual cropping.
have re-dimensioned
detected face areas to 180x180 pixels for CK+ dataset and
have conducted several
experiments with different combination of dimension and
found the above dimension as optimum one. Then I have
used an elliptical masking to remove some unnecessary areas
from the head corners and neck corners. The masked face is
have done this to preserve
the spatial information from the facial area. I have used
LGCP to extract feature from each block to build histogram
final feature vector.
ngth of the feature vector is, 81 x Number of
have followed the above procedure for both training and
have used LIBSVM (C.C.
fold cross validation and the
ing in a person dependent and
independent environment. The whole dataset is divided into
10 equal folds and each fold is validated against rest 90% of
the images in each fold validation. LIBSVM is a multiclass
classifier with variety of options that can be set depending on
have set the kernel parameters for the
classifier to: s=0 for SVM type C-Svc, t=0/1/2 for
linear/polynomial/RBF kernel function respectively, c=100 is
the cost of SVM, g=0.000625 is the value of 1/ (length of
ector), b=1 for probability estimation. This setting
for LIBSVM is found to be suitable for CK+ dataset as well
as JAFFE dataset with both six seven classes of data.
method is tested against some popular methods
e.g., LBP (Local Binary Pattern) (T. Ojala et al., 1996),
(Rotation Invariant LBP) (T. Ojala et al., 2002),
et al., 2002), LBPRIU2
(Rotation invariant and Uniform LBP) (T. Ojala et al., 2002)
., 2008), (J. Li et al., 2012). To
have used the same
experimental setup for all above methods as well as the setup
used by the authors.
Table 1 : Confusion matrices for facial expression
recognition systems using LGCP
Table 1 shows Confusion matrices for facial expression
recognition systems using LGCP on CK+
dataset.
Table 2 : Results (Classification Accuracy) comparison
using different methods on
Method CK+
LGCP 91.9
LBP 91.6
LBP(RI) 90.7
LBP(U2) 90.1
LBP(RIU2) 89.8
LPQ 80.2
Table 2 shows results (Classification
using different methods on CK+ and JAFFE Dataset
results of Table 1 and Table 2 are obtained using LIBSVM
with polynomial kernel. Experiments
with linear and RBF (radial basis function) kernel. The
results are shown in Table 3 on CK+ dataset:
Feature
Extraction for
each Block
Feature
Vector
Online ISSN: 2319-7064
experimental setup for all above methods as well as the setup
Confusion matrices for facial expression
recognition systems using LGCP on CK+ and JAFFE
Confusion matrices for facial expression
recognition systems using LGCP on CK+ and JAFFE
(Classification Accuracy) comparison
on CK+ and JAFFE Dataset
CK+ (%) JAFFE (%)
91.9 93.3
91.6 88.7
90.7 85.4
90.1 84.4
89.8 83.8
80.2 79.6
esults (Classification Accuracy) comparison
using different methods on CK+ and JAFFE Dataset. All the
results of Table 1 and Table 2 are obtained using LIBSVM
xperiments are also performed
with linear and RBF (radial basis function) kernel. The
are shown in Table 3 on CK+ dataset:
416

www.ijsr.net
Table 3: Results of proposed method and some other
popular methods in same experimental setup but different
SVM kernel setup on CK+ dataset (person dependent)
Methods Linear Kernel
Polynomial
kernel
RBF kernel
LGCP
(Proposed
method)
91.89% 91.89% 92.04%
LBP 91.59% 91.59% 91.88%
LBPRI 90.68% 90.68% 90.95%
LBPU2 90.12% 90.12% 90.65%
LBPRIU2 89.83% 89.83% 90.03%
LPQ 80.21% 80.21% 80.43%
Table 4 shows the results comparison on JAFFE dataset with
different SVM kernel setup:
popular methods in same experimental setup but different
SVM kernel setup on JAFFE dataset (person dependent)
Methods Linear Kernel
Polynomial
kernel
RBF kernel
LGCP 93.31% 93.31% 93.52%
LBP 88.72% 88.72% 89.31%
LBPRI 85.37% 85.37% 86.21%
LBPU2 84.43% 84.43% 85.06%
LBPRIU2 83.82% 83.82% 84.12%
LPQ 79.56% 79.56% 80.14%
I have also experimented on 6-class expressions by removing
‘contempt’ from CK+ dataset and ‘natural’ from JAFFE
dataset. This is because most of the previous works are done
on 6 expressions class. Therefore, to compare with those
results, I have removed one class from both the datasets.
Table 5 shows classification accuracy comparison on 6-class
expressions on Cohn-kanade dataset with different SVM
kernel setup. I have removed contempt class from CK+
dataset for proposed LGCP method.
popular methods in different SVM kernel setup on 6-class
expression (person dependent)
Methods
Linear
Kernel
(%)
Polynomial
Kernel
(%)
RBF Kernel
(%)
LGCP
(Proposed
Method)
94.9 94.9 95.9
Gabor
Feature(M.S.
Bartlett et al.,
2005)
89.4 89.4 89.8
LBP(C. Shan et
al., 2009)
91.5 91.5 92.6
LDP(T. Jabid et
al., 2010)
92.8 92.8 94.5
Table 6 shows the accuracy comparison of proposed method
with some other methods on CK+ dataset.
Table 6: Classification Accuracy Comparison on CK+
dataset in a person dependent environment
Author Classifier
Classification
accuracy (%)
LGCP
(Proposed Method)
Multi Class SVM(Poly) 92%
(S.W.Chew et al.,
2011)
SVM(RBF ) 75%
(G.Littlewort et al.,
2011)
SVM(Linear ) 90%
(L.A. Jeni et al.,
2012)
SVM 87%
(S. Naika et al.,
2012)
Multi Class SVM(RBF) 82%
Table 7 shows the accuracy comparison of proposed method
with some other methods on JAFFE dataset. I have done this
in person independent expression recognition environment.
Nine from ten subjects are chosen as the training sample and
the remaining one is chosen as test samples. Therefore,
training and testing subjects are different from each other. I
have repeated this process for all 10 subjects .The 10 results
are then averaged to get the final facial expression
recognition rate.
Table 7: Classification accuracy comparison in person
dependent environment; (SRC: Sparse Representation-based
Classification, GP: Gaussian Process classifier)
Author Classifier Classification accuracy (%)
LGCP
(Proposed Method)
LIBSVM 63.12%
(Y. Zilu et al., 2008) SVM 58.20%
(J. Li et al., 2012) SRC 62.38%
(F. Cheng et al., 2010) GP 55.00%
(J. Lyons et al., 1998) SVM 56.80%
Table 8 shows results comparison in a person dependent
environment on JAFFE dataset. I have followed a 10-fold
cross validation with non-overlapping folds. I have divided
the dataset into 10 random non-overlapping folds with nearly
equal number of instances e.g. 21/22 each. I have used one
fold for testing and rest 9 folds for training support vector
machine. I have repeated this for 10 times with each
independent fold and averaged the results to get the final
recognition rate.
Table 8: Comparison of classification accuracy of
proposed system with some other systems on JAFFE dataset
in a person dependent experimental environment (NN:
Neural Network, LDA: Local discriminant analysis)
Author Classifier
Classification
accuracy (%)
LGCP
(Proposed Method)
Multi Class
SVM(Poly)
93.3%
(K. Subramanian et al.,
2012)
SVM 88.09%
(J. B. M. J. Lyons et
al., 1999)*
LDA-based
classification
92.00%
(Z. Zhang et al., 1998) NN 90.10%
(G. Guo et al., 2003) Linear Programming 91.00%
*used a subset of the dataset
In all cases, proposed system outperforms all the existing
417

www.ijsr.net
systems. 6-class expression gives more accuracy than 7-class
expression. Reason for misclassifying is lack of proper face
registration and variety of face shapes.
5. Conclusion
In this paper, I have proposed a novel method for facial
feature extraction, which is insensitive to noise, and non-
monotonous illumination changes. The method encodes
spatial structure along with local contrast information for
facial expression recognition. Extensive experiments prove
that the proposed method is effective and efficient for
expression recognition. I are planning to incorporate a
effective face registration method in the preprocessing phase
and boosting technique with support vector machine e.g.
adaboost in the recognition phase in future.
References
[1] A. Turk, P. Pentland. "Face Recognition Using
Eigenfaces." in Proceedings of Computer Vision and
Pattern Recognition, 1991: 586-591.
[2] B. Fasel, J. Luettin. "Automatic Facial Expression
Analysis: A Survey." Pattrn Recognition 36, no. 1
(2003): 259-275.
[3] C. Shan, S. Gong & P.W. McOwan,. "Robust Facial
Expression Recognition Using Local Binary
Patterns,." Proc. IEEE Int’l Conf. Image Procession.
2005. 370-373.
[4] C. Shan, S. Gong, P.W. McOwan,. "Facial
expression recognition based on Local Binary
Patterns: A comprehensive study." Image and Vision
Computing 27, no. 6 (2009): 803-816.
[5] C.C.Chang and C.J.Lin.”LIBSVM:a library for
support vector machines ”ACM Transactions on
Intelligent Systems and Technology (2011).
[6] F. Cheng, J. Yu, and H. Xiong,. "Facial expression
recognition in JAFEE dataset based on Gaussian
Process classification." IEEE Transactions on Neural
Networks 21, no. 10 (2010): 1685–1690.
[7] G. Guo, C. R. Dyer,. "Simultaneous Feature
Selection and Classifier Training via Linear
Programming: A Case Study for Face Expression
Recognition." Edited by in Proc. of IEEE Conference
on Computer Vision and Pattern Recognition. 2003.
346-352.
[8] G. Zhao, M Pietkainen. "Boosted Multi-Resolution
Spatiotemporal Descriptors for Facial Expression
Recognition." Pettern recognition Letters 30, no. 12
(2009): 1117-1127.
[9] G.L. Donato, M.S. Bartlett, J.C. Hager, P. Ekman,
T.J. Sejnowski. "Classifying Facial Actions." IEEE
Transactions on Pattern Analysis and Machine
Intelligence 21, no. 10 (1999): 974-989.
[10] G.Littlewort, J. Whitehill, W. Tingfan,I. Fasel, M.
Frank, J. Movellan & M. Bartlett,. "The computer
expression recognition toolbox (CERT)" IEEE
International Conference on Automatic Face &
Gesture Recognition and Workshops (FG 2011).
2011. 298-305.
[11] H. Kabir, T. Jabid, and O. Chae."Local Directional
Pattern Variance (LDPv):A Robust Feature
Descriptor for Facial Expression Recognition." The
International Arab Journal of Information
Technology 9(4) (2012).
[12] I. Cohen, N. Sebe, S. Garg, L. S. Chen and T. S.
Huanga.Facial expression recognition from video
sequences: temporal and static modelling. Computer
Vision and Image Understanding, 91(2003), 160-
187.
[13] J. Li, Z. Ying. "Facial Expression Recognition Based
on Rotation Invariant Local Phase Quantization and
Sparse Representation." Edited by Computer,
Communication and Control IEEE Second
International Conference on Instrumentation &
Measurement. 2012. 1313-1317.
[14] J. Michael, Lyons, Miyuki Kamachi, Jiro Gyoba.
"Japanese Female Facial Expressions (JAFFE),."
(Database of digital images) 1997.
[15] K. Anderson and Peter W. McOwan. A Real-Time
Automated System for the Recognition of Human
Facial Expressions. IEEE Transactions on Systems,
Man, and Cybernetics, 36, 1 (2006), 96-105.
[16] K. Subramanian, S. Suresh and R. V. Babu. "Meta-
Cognitive Neuro-Fuzzy Inference System for Human
Emotion Recognition." Edited by IEEE World
Congress on Computational Intelligence. 2012. 10-
15.
[17] L A. Jeni, András Lőrincz, Tamás Nagy, Zsolt
Palotai, Judit Sebők, Zoltán Szabó & Dániel Takács.
3D shape estimation in video sequences provides
high precision evaluation of facial expressions.
Image and Vision Computing, 30, 10 (October 2012),
785-795.
[18] M. J. Lyons, J. Budynek, and S. Akamatsu.
"Automatic classification of single facial images."
IEEE Trans. Pattern Analysis and Machine
Intelligence 21, no. 12 (1999): 1357–1362.
[19] M. J. Lyons, S. Akamatsu, M. Kamachi, and J.
Gyoba. "Coding Facial Expressions with Gabor
Wavelets." Edited by Proc. of IEEE Int. Conference
on Automatic Face and Gesture Recognition. 1998.
200-205.
[20] M. Meulders, D. Boeck, V. Mechelen, A. Gelman.
"Probabilistic Feature Analysis of Facial Perception
of Emotions." Applied Statistics 54, no. 4 (2005):
788-793.
[21] M. Pantic and L. J. M. Rothkrantz. Automatic
analysis of facial expressions: the state of the art.
Intelligence, 22, 12 (2000), 1424–1445.
[22] M. Yeasin, B. Bullot and R. Sharma. Recognition of
Facial Expressions and Measurement of Levels of
Interest From Video. IEEE Trans. Multimedia, 8, 3
(2006), 500-508.
[23] M. Zia Uddin, J.J. Lee, T.S. Kim. "An Enhanced
418

www.ijsr.net
Independent Component-based Human Facial
Expression Recognition from Videos." IEEE Trans.
Consumer Electronics 55, no. 4 (2009): 2216-2225.
[24] M.F. Valstar, I. Patras , M. Pantic. "'Facial Action
Unit Detection using Probabilistic Actively Learned
Support Vector Machines on Tracked Facial Point
Data." IEEE Int'l Conf. Computer Vision and Pattern
Recognition 3 (2005): 76-84.
[25] M.F. Valstar, M. Pantic. "'Fully automatic facial
action unit detection and temporal analysis'." Proc.
IEEE Int'l Conf. on Computer Vision and Pattern
Recognition (CVPR'06), 2006: 149-156.
[26] M.S. Bartlett, G. Littlewort, I. Fasel & R. Movellan.
Real Time Face Detection and Facial Expression
Recognition: Development and Application to
Human Computer Interaction. In Proc. CVPR
Workshop Computer Vision and Pattern Recognition
for HumanComputer Interaction (2003).
[27] P. Ekman and W. Friesen. Facial Action Coding
System: A Technique for the Measurement of Facial
Movement. Consulting Psychologists Press, Palo
Alto, 1978.
[28] P. Lucey, J. F. Cohn, T. Kanade, J. Saragih, Z.
Ambadar & I. Matthews. The Extended Cohn-Kande
Dataset (CK+): A complete facial expression dataset
for action unit and emotion-specified expression.
Paper presented at the Third IEEE Workshop on
CVPR for Human Communicative Behavior Analysis
(CVPR4HB 2010) (2010).
[29] Robinson, G. "Edge detection by compass gradient
masks." Computer graphics and image processing, 6
(1977): 492-501.
[30] S. Lajevardi, Z. Hussain. "Feature extraction for
facial expression recognition based on hybrid face
regions." Advances in Electrical and Computer
Engineering 9, no. 3 (2012): 63-67.
[31] S. Naika C.L., S. Shekhar Jha, P. K. Das & S. B.
Nair,. "Automatic Facial Expression Recognition
Using Extended AR-LBP." Wireless Networks And
Computational Intelligence 292, no. 3 (2012): 244-
252.
[32] S.W.Chew, P.Lucey, S. Lucey, J. Saragih, J.F. Cohn
& S. Sridharan,. "Person-independent facial
expression detection using Constrained Local
Models." 2011 IEEE International Conference on
Automatic Face & Gesture Recognition and
Workshops (FG 2011),. 2011. 915-920.
[33] T. Jabid, H. Kabir, and O. Chae. "Local Directional
Pattern for Face Recognition." Edited by IEEE
International Conference on Consumer Electronics.
2010. 329-330.
[34] T. Ojala, M. Pietikäinen and D. Harwood. A
Comparative Study of Texture Measures with
Classification Based on Feature Distributions.
Pattern Recognition, 19, 3 (1996), 51-59.
[35] T. Ojala, M. Pietikäinen & T. Mäenpää.
Multiresolution Gray-scale and Rotation Invariant
Texture Classification with Local Binary Patterns.
Intelligence, 24/7(2002), 971-987.
[36] V. Ojansivu and J. Heikkilä. Blur Insensitive Texture
Classification Using Local Phase Quantization in
Proc. ICISP, Berlin, Germany (2008), 236-243.
[37] Y. L. Tian, T. Kanade and J. F. Cohn. Recognizing
action units for facial expression analysis. IEEE
Trans. Pattern Anal. Mach. Intell, 23, 2 (2001), 97-
115.
[38] Y. Tian, L. Brown, A. Hampapur, S. Pankanti, A.
Senior, R. Bolle. "Real-World real-Time Automatic
Recognition of Facial expressions." Edited by in
proceedings of IEEE Workshop on Performance
Evaluation of Tracking and Surveillance. 2003. 9-16.
[39] Y. Tian, T. Kanade, J. Cohn. "Facial Expression
Analysis." Handbook of Face Recognition, Springer,
2003.
[40] Y. Zhang and Q. Ji. Active and dynamic information
fusion for facial expression understanding from
image sequences. IEEE Trans. Pattern Anal. Mach.
Intel., 27, 5 (2005), 699–714.
[41] Y. Zilu, L. Jingwen and Z. Youwei. "Facial
Expression Recognition Based on Two Dimensional
Feature Extraction,." Edited by IEEE International
Conference on Signal Processing. 2008. 1440-1444.
[42] Z. Zhang, M. Lyons, M. Schuster, S. Akamatsu.
"Comparison between geometry-based and Gabor-
wavelets-based facial expression recognition using
multi-layer perceptron." Third IEEE International
Conference on Automatic Face and Gesture
Recognition, Proceedings. 1998. 454-459.
Author Profile
Mohammad Shahidul Islam received his
B. Tech. degree in Computer Science and
Technology from Indian Institute of
Technology-Roorkee (I.I.TR), Uttar
Pradesh, INDIA in 2002, M.Sc. degree in
Computer Science from American World
University, London Campus, U.K in 2005
and M.Sc. in Mobile Computing and Communication from
University of Greenwich, London, U. K in 2008. He has
successfully finished his Ph.D. degree in Computer Science
& Information Systems from National Institute of
Development Administration (NIDA), Bangkok, Thailand in
2013. Currently he is serving as an associate professor in the
department of computer science and Engineering at Atish
Dipankar University of Science & Technology, Dhaka,
Bangladesh. His field of research interest includes Image
Processing, Pattern Recognition, wireless and mobile
communication, Satellite Commutation and Computer
Networking.
419

Local Gray Code Pattern (LGCP): A Robust Feature Descriptor for Facial Expression Recognition

More Related Content

What's hot (16)

Viewers also liked (6)

Similar to Local Gray Code Pattern (LGCP): A Robust Feature Descriptor for Facial Expression Recognition (20)

More from International Journal of Science and Research (IJSR) (20)

Recently uploaded (20)

Local Gray Code Pattern (LGCP): A Robust Feature Descriptor for Facial Expression Recognition