Microsoft Word - brain_2_4.doc

Impact of Ethnic Group on Human Emotion Recognition Using

Backpropagation Neural Network

Nabil M. Hewahi
Computer Science Department

Bahrain University, Bahrain
nhewahi@gmail.com

Abdul Rhman M. Baraka
Computer Science Department

Islamic University of Gaza, Palestine
ambwafa@hotmail.com

Abstract
We claim that knowing the ethnic group of human would increase the accuracy of the

emotion recognition. This is due to the difference between the face appearances and expressions of
various ethnic groups. To test our claim, we developed an approach based on Artificial Neural
Networks (ANN) using backpropgation algorithm to recognize the human emotion through facial
expressions. Our approach has been tested by using MSDEF dataset, and we found that, there is a
positive effect on the accuracy of the recognition of emotion if we consider the ethnic group as
input factor while building the recognition model. We achieved 8% improvement rate.

Keywords: Emotion Recognition, Artificial Neural Networks, Human Computer
Interaction, Features Extraction.

1. Introduction
Obviously, the emotions is an important aspect in the interaction and communication

between people, and the human can express these emotions through facial elements, speech, body
movement, or etc.. Since the science of Artificial Intelligence (AI) is concerned with the automation
of intelligent behavior, we need to improve the interaction between computer and human by
recognition the human emotions by using face expressions [9] [12] [15] [20], speech [4] [16] or etc.
[11]. In additional, appearance of emotions are universal across individuals as well as human
ethnics and cultures[1] [2], but the way of emotion expression is that vary from one ethnic group to
another and the differences cross-cultural in the familiarity effect on recognition accuracy for
different types of expressions.

There are many of researches that focused on emotion recognition but regardless the ethnic
group as factor in their models.

In this paper, we propose an emotion recognition approach using facial expression
considering ethnic group based on Backpropagation neural network, to study the impact of ethnic
group on accuracy of emotion recognition models.

In next section we focus on some related works. Section three describes our proposed
approach. Section four discusses the implementation. Experiments and results are presented in
section five. Last section is the conclusion.

2. Related Works
In the early 1990’s the engineering community started to construct automatic methods of

recognizing emotion from facial expression in an image and videos [12], and many of computer
studies that focused on emotions recognition. Some studies use facial expressions to build a model
for emotion recognition, and others use another elements or factors to build the models like voice,
pulse, body movement and etc.

Matthew and Patterson [15] discuss a framework for the classification of emotional states,
based on still images of the face, using active appearance model (AAM) and get distance using n

N. M. Hewahi, A.R. M. Baraka - Impact of Ethnic Group on Human Emotion Recognition Using Backpropagation Neural
Network

Euclidean’s as features. To train and test the classifier they chose to use the facial expression
database known as “FEEDTUM” - Facial Expressions and Emotion Database [8] , and seven basic
emotions are used, happy, sad, angry, surprise, fear, disgust and natural state. The best results they
obtained are in happy, natural and disgust emotions at the rate 93.3%, fear at 90.0%, and 79.7% in
surprise, angry and sad at rate 63.9%. This study does not consider the ethnicity.

Raheja and Kumar [14] presented architecture for human gesture recognition, considering
color image with different gestures by using backpropagation Artificial Neural Networks (ANN or
NN). Four stages applied in the approach, face detection, image reprocessing, training network and
recognition model. The pre-processing stage contains three methods, histogram equalization, edge
detection, thinning and token generation. The ethnic group was not considered in this model. The
model was trained using the three different gesture images, happy, sad and thinking expressions of
faces. The model was tests with 100 images of three gestures, the results were 94.28% for happy,
85.71% for sad and 83.33% for thinking.

Karthigayan et.al [12] used Genetic Algorithms (GAs) and Artificial Neural Networks to
build human emotion classifier. This classifier detects six human emotions Neutral, Sad, Anger,
Happy, Fear, Disgust (or Dislike) and Surprise. They depend on two facial elements in the
classifier, eyes and lips. By applying some preprocessing methods and edge detection, they
extracted the eyes and lip regions, then extracted the features from these regions. Three feature
extraction methods are applied, projection profile, contour profile and moments. The GA is applied
to get the optimized values of the minor axes of an irregular ellipse corresponding to the lips and the
minor axis of a regular ellipse related to eye by using a set of new fitness functions. Finally, they
apply the results from GA on the ANNs model. Two architectures of ANNs models are proposed
with an average of 10 trials of testing. The achieved results of 3x20x7 and 3x20x3 of ANN
architecture were 85.13% and 83.57% of success rate respectively. The successful classification
even goes to the maximum of about 91.42% in the ANN model of 3x20x7 structure. A South East
Asian (SEA) race is only considered in this work.

3. Proposed Approach
Our proposed approach uses the face expression to detect the emotions through five steps

that shows in Figure 1.

INPUT

IMAGE PRE-
PROCESSING

POINTS SELECTION

FEATURES
EXTRACTION

ANN CLASSIFIER

Figure 1. A proposed approach.

BRAIN. Broad Research in Artificial Intelligence and Neuroscience
Volume 2, Issue 4, December 2011, ISSN 2067-3957 (online), ISSN 2068 - 0473 (print)

A. The Input
The input is an image of frontal human face which means that no bending the face to any

angle and no rotated face to any side.
B. Image Reprocessing
In this stage we apply certain sub processes on the image to achieve a good image with one

standard in terms of contrast and size. Because we depend on Euclidian distances to extract our
features, the image size must be standardized. Also we apply a contrast process for each image to
make the elements of the face distinguishable from other objects and also to be able to distinguish
between background and the face. This will help the user to select all the points clearly. Simple
algorithms are applied for resizing and contrast processes [5].

C. Point Selection
We chose 46 points which are distributed over human face image and use these points for

features extraction. The choice of these points is to determine the shape of each element of the face
(eyes, eyebrows and mouth), because the shape of these elements is changeable for each emotion,
but these changes are different for each race as shown in Figure 2.

Figure. 2. Examples of Angry from different races. (A,B) African. (C,D) Asian. (E,F) Caucasian.

The number of points and the position of points are not standardized, but it is depending on

the features that will be extracted, and used for the classifier. Many researches use various number
of points and positions based on their view about the feature to be considered [13] [18] [19]. Figure
3 shows the points we used.

Figure 3. 46 points are selected on face elements to describe the emotions.

N. M. Hewahi, A.R. M. Baraka - Impact of Ethnic Group on Human Emotion Recognition Using Backpropagation Neural
Network

D. Features Extraction
While human emotion is changing, face expression is also changing. This means that,

properties of face elements do change when face expression is being changed. As a consequence of
this, the distances between points are changing. The distances between certain points as will be
shown later describe the human emotion.

Based on what we stated above, we need to extract 28 features, which describe the distances
between certain points explained in the previous stage, these features are classified into six groups,
and each group describes the features of one face element. All features are a vertical distances
between two points. Group one contains seven features for mouth, groups two and three contains 14
features for eyes, groups four and five contain six features for eyebrows, and the last group has one
feature only which is the distance between the beginning of the eyebrow and the beginning of the
eye in same side, this is significant (from point 23 to 15) because it is used to measure the distance
of eyebrow from the eye. This feature is shown in Figure 4 by a line.

E. Classifier
For classification purpose of the emotions, we use ANN of supervised learning based on

backpropagation algorithm. Backpropagation neural network architecture is used with its standards
learning function with 28 inputs representing the extracted features and 6 outputs representing 6
emotions, happy, sad, angry, fear, shame and disgust. the emotions. We have also a hidden layer
with 16 nodes selected after various trails to obtain the best results. The used ANN is depicted in
Figure 5.

Figure 5. ANN Structure

4. Implementation

A. Face Features Extraction (FFE)
We used Microsoft Visual C#.net to develop a desktop application to apply the first of the

four stages of our approach. This application is called Face Features Extraction (FFE), and it is used
to perform four methods or functions as follow: Apply some pre-processing method, Selection of

Figure 4. Distance between eyebrow and eye.

BRAIN. Broad Research in Artificial Intelligence and Neuroscience
Volume 2, Issue 4, December 2011, ISSN 2067-3957 (online), ISSN 2068 - 0473 (print)

Points, Features Extraction and save all input information for the classifier (Features, Ethnic group,
Gender and emotion). Figure 6 shows the interface of FFE program.

B. Classifier
To build our classifier we use Mathworks Matlab using NNTool tool. This tool allows

building our specific ANN classifier, and use backpropagation method, and one hidden layers with
16 neurons.

C. Dataset
Montreal Set of Facial Displays of Emotion (MSFDE) dataset is selected to test and evaluate

our approach. This data set includes 224 gray images of females and males faces and four ethnic
groups African, Asian, Hispanic and Caucasian, these images describe seven facial expressions
nature, happy, sadness, angry, fear, shame and disgust. Each ethnic group contains two genders;
each gender contains four persons, each person has seven emotions, but each person has a one
image for nature emotion, so we ignored this emotion because we need more than one image for
each emotion to be able to train well the ANN. This means we use only six emotions, and Figure 6
shows samples for images of MSFDE dataset.

Figure 6. Interface of FFE program.

N. M. Hewahi, A.R. M. Baraka - Impact of Ethnic Group on Human Emotion Recognition Using Backpropagation Neural
Network

5. Experiments and results
We performed groups of experiments to study the impact of ethnic group (race) in the

accuracy of emotion recognition with three kinds of ethnic groups (Asian, Caucasian as African).
So we have three experiments, each experiment has a neural network as a classifier, and each neural
network has three layers where there are 16 neurons in the hidden layer except Asian network has
17 neurons (the best result with 17 neurons for Asians).

To study the accuracy of emotion recognition for our approach regardless the ethnic group
we used 108 images for the training representing six emotions of six persons. For testing, we used
36 images representing six emotions of six persons.

On the other hand, to study the accuracy of emotion recognition for our approach
considering the ethnic group we used 36 images for training for each ethnic group representing six
emotions of six persons, and test the classifier by using 12 images representing six emotions of six
persons.

Our experiments show that, the impact of ethnic group on the accuracy of emotions
recognition is a positive where the accuracy of emotion recognition considering ethnic group is
83.3% as shown in Figure 8, and we got 75% of accuracy regardless ethnic group as shown in
Figure 9.

African, Female and Angry

Asian, Male and Disgust

Caucasian, Male and Happy

Figure 7. Samples of MSFDE dataset.

BRAIN. Broad Research in Artificial Intelligence and Neuroscience
Volume 2, Issue 4, December 2011, ISSN 2067-3957 (online), ISSN 2068 - 0473 (print)

Also we got that, African race has the lowest accuracy, and the Caucasian race has the

highest accuracy of emotions recognition. In additional, the disgust emotion has best accuracy of
recognition and fear emotion has the worst accuracy of recognition.

6. Conclusion
Many psychological studies refer to that, the ethnic group is a considered factor in emotion

recognition, and may play an important role to increasing the accuracy of emotion recognition. In
this paper, we try to investigate the impact of ethnic group on accuracy of emotion recognition
based on face expressions by proposing an emotion recognition approach using backpropagation
neural networks. Our approach has five stages. To complete first four stages, we developed a
desktop program called FFE system, and we used Matlab program to build an ANN classifier. We
use MSFDE data set to test and evaluate our approach.

We demonstrated in this paper through our experiments positive results that the ethnic group
has a positive impact on the accuracy of identifying human emotions based on facial features. We
got 75% accuracy regardless ethnicity, and 83.3% accuracy considering ethnicity. However, this
effect is not very large. In additional, the fear emotion has the worst recognition accuracy, and the
disgust emotion has the best accuracy. Also, based on the race, the Caucasian has got the best
accuracy of emotion recognition, and African has the worst accuracy.

For future work we recommend that; develop a system that can extract the features from
human face image automatically, also finding more useful features that may lead to more accuracy
on recognition process.

References

[1] Elfenbein Hillary A. and Ambady Nalini, "Universals and Cultural Differences in
Recognizing Emotions". American Psychological Society. Volume 12, No 5, pp. 159-164. (2002).
[2] Beaupré Martin G. “Cross-Cultural Emotion Recognition among Canadian Ethnic Groups”,
Journal of Cross-Cultural Psychology, Vol. 36, 355-370, (2005).

Figure 9. Classification accuracy regardless the ethnic group (Total accuracy 75%)

Figure 8. Classification accuracy of emotions considering the ethnic group (Total accuracy is 83.3%)

N. M. Hewahi, A.R. M. Baraka - Impact of Ethnic Group on Human Emotion Recognition Using Backpropagation Neural
Network

[3] Castelli Fulvia. "Understanding emotions from standardized facial expressions in autism and
normal development". National Autistic Society, Vol 9. Issue: 4, pages 428-449 (2005).
[4] Douglas-Cowie E., Campbell N., Cowie R. and P. Roach, “Emotional Speech: Towards a
New Generation of Database” Speech Communication, Vol.40 (1-2). pp 33-60 (2003).
[5] Efford N.,”Digital Image Processing”, Personal Education Limited, (2000).
[6] Ekman Paul;Friesen Wallace V.;O'Sullivan Maureen;Chan Anthony;Diacoyanni-Tarlatzis
Irene;Heider Karl;Krause Rainer;LeCompte William Ayhan;Pitcairn Tom;Ricci-Bitti Pio E.;Scherer
Klaus;Tomita Masatoshi;Tzavaras Athanase . "Universals and Cultural Differences in the
Judgments of Facial Expressions of Emotion". Journal of Personality and Social Psychology.
Vol.53, No.4 pp.712-717 (1987).
[7] Ekman, P., Friesen, W. V., & Hager, J. C.”Facial Action Coding System. Salt Lake City”.
Research Nexus. (2002).
[8] FG-NET Database with Facial Expressions and Emotions Database (FEED),
http://cotesys.mmk.e-technik.tu-muenchen.de/isg/content/feed-database. Resources and Evaluation
Conference (LREC). (2006).
[9] Fischer Robert, "Automatic Facial Expression Analysis and Emotional Classification",
University of Applied Science Darmstadt (FHD), (2004).
[10] Hewahi N.,Olwan A., Tubeel N., El-Asar S. and Abu-Sultan Z., ”Age Estimation based on
Neural Networks using Face features”, Journal of Emerging Trends in Computing and Information
Sciences(CIS), Vol.1. No.2. pp. 61-68 (2010).
[11] Horlings Robert. "Emotion Recognition Using Brain Activity". Man-Machine Interaction
Group. TU Delft. (2008).
[12] Karthigayan M., Rizon M., Nagarajan R. and Yaacob S., “Genetic Algorithm and Neural
Network for Face Emotion Recognition”, University Malaysia Perlis (UNIMAP), (2008).
[13] Kim Moon Hwan, Joo Young Hoon, and Park Jin Bae. "Emotion Detection Algorithm
Using Frontal Face Image". Ministry of Information and Communication (MIC). (2005).
[14] Raheja J., Kumar U., “Human Facial Expression Detection from Detected in Captured
Image Using Backprobagation Neural Network”, International Journal of Computer Science &
Information technology (IJCSIT), Vol.2. No.1, (2010).
[15] Ratliff Matthew S. and Patterson Eric "Emotion Recognition Using Facial Expressions with
Active Appearance Models". Third International Conference on Human Computer Interaction.
ACTA Press Anaheim, CA, USA (2008).
[16] Rudra T., Kavakli, M. and Tien, D. “Emotion Detection from Male Speech in Computer
Games”. TENCON, IEEE Region 10 Conference. 4428779. pp1-4 (2008).
[17] Soto Jose´ A. and Levenson Robert W. "Emotion Recognition across Cultures: The
Influence of Ethnicity on Empathic Accuracy and Physiological Linkage". Department of
Psychology, Pennsylvania State University, University Park, PA 16802, USA (2009).
[18] Suzuki Kenji, Yamada Hiroshi and Hashimoto Shuji “A similarity-based neural network for
facial expression analysis”. Elsevier Science Inc. New York, NY, USA. (2007).
[19] Tsapatsoulis Nicolas, Karpouzis Kostas, Stamou George, Piat Frederic and Kollias Stefanos.
“A Fuzzy System for Emotion Classification based on the MPEG-4 Facial Definition Parameter
Set”. European Signal Processing Conference (EUSIPCO'00) (2000).
[20] Waller B., Cray .J. and Burrows A., “Selection for Universal Facial Emotion”, American
Psychological Association, 1528-3542/08, (2008).