Journal of Signal and Information Processing, 2011, 2, 270-273
doi:10.4236/jsip.2011.24038 Published Online November 2011 (http://www.SciRP.org/journal/jsip)
1
The Study of Multi-Expression Classification
Algorithm Based on Adaboost and Mutual
Independent Feature
Liying Lang, Zuntao Hu
Department of Electrical and Information, Hebei University of Engineering, Handan, China.
Received June 20th, 2011; revised August 13th, 2011; accepted August 22nd, 2011.
ABSTRACT
In the paper conventional Adaboost algorithm is improved and local features of face such as eyes and mouth are sepa-
rated as mutual independent elements for facial feature extraction and classification. The multi-expression classifica-
tion algorithm which is based on Adaboost and mutual independent feature is proposed. In order to effectively and
quickly train thresho ld values of weak classifiers of fea tures, Sample of training is carried out simple improvement. We
obtain a good classification results through experiments.
Keywords: Adaboost Multi-Expression Classification Algorithm, Local Feature, Feature Extraction, Sample Training
1. Introduction
Due to increasingly wide application, Facial multi-ex-
pression classification which is widely concerned by gov-
ernments and research institutes is becoming a challeng-
ing research topic in field of pattern recognition. Cur-
rently there are many main methods used for facial ex-
pression classification, such as Euclidean distance, Sup-
port vector machine (SVM), Neural network(NN), Hid-
den Markov(HMM), Adaboost and Linear discriminant
analysis (LDA), etc. Adaboost algorithm which owns
higher speed and higher detection rate has been success-
fully applied in the field of face detection. Most of facial
expression changes exist in eyes and mouth, so features of
eyes and mouth are treated as mutual independent ele-
ments [1]. The method can greatly reduce redundancy and
improve the speed of training threshold values. However,
if we want to use the algorithm to facial multi-ex pression
classification, there is a main problem. When error recog-
nition rate of weak classifiers of Adaboost is higher,
overall recognition rate of the algorithm will reduce to
zero exponentially, key problem is to train a weak classi-
fier threshold value accurately and fast. So we improve
training samples, and negative samples is proposed in this
paper [2]. Threshold value of weak classifier is the most
crucial part. Positive samp les contain on ly images of eyes
or mouth, while negative samples are removed eyes and
mouth. So the multi-expression classification algorithm
which is based on Adaboost and mutual independent fea-
ture is proposed. Experimental results prove that false
recognition rate is almost close to zero.
2. Adaboost Multi-Expression Classification
Algorithm
Basic idea of Adaboost algorithm is to use a large num-
ber of weak classifiers to add up together to form a very
strong classification through a certain method. In the
algorithm, each training sample is assigned a weight, and
it demonstrates a probability of some weak classifiers
which can be selected into training set. If a sample is
accurately classified by the current weak classifier, its
weight will be reduced. On the contrary, if a sample is
not properly classified, the weight is to be raised accord-
ingly. In this way, Adaboost algorithm can focus on more
difficult samples [2,3].
A weak classifier


j
H
x consists of the following
three parts: rectangle feature value

j
f
x, a classifica-
tion threshold value
j

and a direction sign
j
p (
j
p=
0,1). They are in line with the following relation ship [4]:



,, 1,2,3,4,5,6,7
0,
jj jj
j
jpf xpj
Hx else
 



 (1)
We obtain the minimum threshold value

min
f
i

and
the maximum threshold value

max
f
i

of eyes and mo-
uth, making
 
min max
j

f
ifi



.

j

H
xis one of seven

The Study of Multi-Expression Classification Algorithm Based on Adaboost and Mutual Independent Feature271
facial expressions .Positive samples of human eyes and
mouth contain a variety of gestures, such as eyes open or
closed, wearing glasses, mouth open or closed, etc. Neg-
ative samples do not contain any of eyes and mouth. The
algorithm is showed as follows:
The first step:
Giving samples
  


112 2
,,,, ,
nn
x
yxy xy, i
x
is in-
put training sample, indicates positive sam-
ple or negative sample, L is the number of positive sam-
ples ,m is the number of negative samples. T is the num-
ber of iterations of strong classifier. t is the probabil-
ity distribution of sample weigh, and making

0,1
i
y
D

1,17,1,,,1,, 7nlDil ni .
n is the total number of samples, j is one of seven types
of expression [5]

1,
1,
i
iy
yl iy


 (2)
The second step:
1) We train a weak classifier of each feature of sample
i
x
, whose output is multi-class.
2) In weight distribution t
D, we select the best clas-
sifier t
h from various weak classifiers, making the clas-
sification error rate minimum. The following formula ob-
taining the maximum:
 


'
,,
ttiti
il
rDilylhx

,l
(3)

 


'
,
max max,,
tt tit
il
rr Dilylhx


i
l
(4)



,
tti
hx hxl (5)
3) According to the classification performance, the
weight of classifier will be changed.
1 1ln
21
t
tt
r
ar






(6)
4) Sample weights will be updated and normalized.
 





1
,exp ,
,ttii
tt
Dil aylhil
Dil z


 (7)
t
z is normalization constant
5) For the normalization constant t, weight distribu-
tion is made a probability density.
z
1t
D
 




,,exp ,
tt titi
il
zDil aylhxl
 (8)
The third step:
The final strong classifier can be obtained as follows:


1
:1T
ii
t
t
iHx yc
n


The algorithm error rate fits the following inequality
on the training set at this time [4,6].
z
,
(9)
 
1
,
T
tt
t
H
xlsignahxl






(10)
3. Training and Testing of Local Features
3.1. Training
We need a large number of samples in sample training,
which is an important characteristic about Adaboost al-
gorithm. Selection of sample is very important, which
determines the effect of the classification. We divide sa-
mples into positive and negative. Eyes and mouth are
treated as mutual independent feature elements. In order
to easily obtain the weak classifiers threshold values, we
train the eyes and mouth respectively. When we get the
minimum threshold value or the maximum threshold
value of kinds of expression, chenges of eyes and mouth
will lead to huge variations. So classifiers are very diffi-
cult to get threshold values accurately and effectively.
Therefore, specific training process is divided into four
steps [7]. The first step, we use the positive samples to
train, obtaining the threshold values quickly. The second
step, in order to adjustment the threshold values appro-
priately, we use whole face images to train. As the false
detection always occurs in eye or mouth, we use eye im-
ages and mouth images to further adjust the threshold
parameters in third step. The fourth step is to use the new
negative sample proposed in the paper to reduce the false
detection rate.
3.2. Detection
The specific detection process is as follows: when there
is a image to detecte, we use different rectangular boxes,
whose sizes are generally from small to large, to scan the
whole image. The size of the smallest rectangle box is
normalized, For example 24 × 24 pixel. Each rectangular
box move a pixel from right to left until it reaches edge
of the image. When rectangular box scans the whole im-
age at this level, rectangular box enlarge a certain pixel
to next scanning. Therefore the largest rectangle box is
several times larger than the smallest rectangular box.
During the scan, each rectangular area can be carried out
classification decision. The purpose of scan is to find a
specific facial feature region. Therefore, if detection area
can be adopted by the classification, it proves to find a
person’s eyes or mouth, else detection will be stopp ed. If
any of rectangular area can not be adopted by the classi-
fier, the region does not exist any eyes or mouth. Of
course, when dimensions of the rectangular area enlarge
in proportion, threshold parameters of weak classifier
also amplified by the same proportion.

The Study of Multi-Expression Classification Algorithm Based on Adaboost and Mutual Independent Feature 272
4. Experimental Results and Analysis
4.1. Facial Expression Image Database
We selected Japanese women expression database of the
Kyushu University JAFFE and Our self-built facial ex-
pression database in experiments. The JAFFE database is
made up 213 individual images of ten persons, and each
person shows anger, disgust, fear, happiness, sadness,
surprise and neutral. There are 2 - 4 images for every
face expression, and images are all 256 × 256 grayscale
images. Our self-built database composes of seven indi-
viduals, each person also shows seven basic facial ex-
pressions, and pictures are also 128 × 128 grayscale ima-
ges [3,8]. As images obtain from two different databases,
there are very huge difference. To be able to quickly and
accurately classify the facial expression, images need
appropriate pretreatment. In this paper, we use geometric
normalization and the intensity normalized to improve
the image quality. We get the mouth images, the eye im-
ages and the whole face images. Figure 1 shows the im-
ages selected from the JAFFE database and self-built
database. Figure 2 shows the pretreatment image and the
negative sample.
4.2 Experimental Results
We selected 120 facial expression images of six persons
from JAFFE database and selected 42 images of three
persons from self-built database in experiments. Table 1
shows different experimental results between JAFFE da-
tabase and self-built datab ase. Figure 3 shows the recog-
nition rates of the seven different facial expressions(1 is
on behalf of happiness, 2 is neutral, 3 is sadness, 4 is sur-
prise, 5 is disgust, 6 is angry, 7 is fear). It is can be seen
from the Table 1 that recognition rates in JAFFE are
much more higher than the self-built database. The rea-
son is that the facial expression in self-built database is
not exaggerated, resulting in a huge difference between
the experimental results.
4.3. Experimental Analysis
In experiments, the number of iteration is set 60. With
the number of iterations changing, the error recognition
rate will be very different. We compared two different
expressions in a group, the experimental results show
that recognition rate of happiness and surprise are more
higher than disgust, sadness and neutral. The reasons are
that facial expression changes of happiness or surprise is
much more obvious, feature extraction is more easier,
and classification get a small error. What’s more, it can
be seen from Figure 4 that the error recognition rate of
the first 30 iterations reduces more faster than the 30
times later. Especially you can see that when the iteration
is to 50, the change of error rate is very small. According
Figure 1. Parts of samples from JAFFE and our self-built
database.
Figure 2. The pretreatment image and the negative sample.
Table 1. The detection results of the method.
Database Total
sample numberWrong detection number Correct detection rate
JAFFE 120 13 89.16%
Self-built
database 42 7 83.33%
Figure 3. The results of seven types of expression recogni-
tion rate.
to the characteristics of this ch ange, without affecting the
recognition rate, we can reduce the number of iterations
to accelerate the classification speed.
5. Conclusions
In the paper, Adaboost algor ithm, which has been succe-
ssfully applied in the field of face detection, applies to
the facial expression classification. We introduce the ba-
sic principle of Adaboost milti-expresion classification
algorithm and demonstrate the process of training and
testing in detail. Because changes of facial expression
mainly exist in eyes and mouth, we treat eyes and mouth
as mutual independent elements, which improved the
speed of training threshold value. The negative samples
is proposed and used in training and testing. The ex-
perimental results demonstrated the feasibility of the
method, which obtained a good recognition result.

The Study of Multi-Expression Classification Algorithm Based on Adaboost and Mutual Independent Feature
273
Figure 4. Error recognition rate changing with the iteration
number.
6. Acknowledgements
This work is supported by Department of Electrical and
Information, Government of Handan, China. Authors are
thanking to them for their sponsorship to do this work.
REFERENCES
[1] X. Z. Zhao and X. F. Zhu, “Adaboost Algorithm Based
on Detection of the Human Eye of Sample Selection,”
Computer Technology and Development, Vol. 20, No. 2,
2010, pp. 12-16.
[2] H. Y. Wang and X. Li, “Facial Expression Recognition
Based on Improved Gabor Transformation and Adaboost
Algorithm,” Journal of Liaoning University of Technol-
ogy (Natural Science Ed it i o n ), Vol. 30, No. 1, 2010.
[3] M. Qiao and Y. J. Chen, “Feature Extraction Methods on
Facial Expression Recognition,” Journal of Chongqing
Institute of Technology, Vol. 22, No. 6, 2008, pp. 118-
122.
[4] G. L. Yang, Z. L. Wang and J. X. Ren, “Adaboost Algo-
rithm Is Used for Facial Expression Recognition,” Com-
puter Engineering, Vol. 25, No. 4, 2005, pp. 78-83.
[5] Y. Ma, X. Q. Ding and Z. E. Wang, “Robust Preciseeye
Location under Probabilistic Framework,” Republic of
Korea, Seoul, 2004, pp. 339-344.
[6] J. Zhao and L. A. Xu, “Target Detection Based on Ada-
boost Al g or i t h m , ” Computer Engineering, Vol. 30, No. 2,
2004, pp. 78-83.
[7] Y. Huang and Z. L. Ying, “Classifier Combination Based
Expression Recognition,” Computer Engineering and Ap-
plications, Vol. 43, No. 25, 2007.
[8] P. S. Aleksic and A. K. Katsaggelos, “Automatic Facial
Expression Recognition Using Facial Animation Parame-
ters and Muhistream HMMs,” IEEE Transactions on In-
formation Forensics and Security, Vol. 1, No. 1, 2006, pp.
3-11. doi:10.1109/TIFS.2005.863510