emotion classification using eeg entropy · different classifiers namely quadratic discriminant...

��

�

�

Emotion classificationMandeep Singh

[1, 2, 3]Department of Electrical & Instrumentation Engineering,

[1] [email protected]

Abstract: Human emotions can be expressed as negativeto a particular stimulus. In this paperrecognition based on Electroencephalogram (EEG) signalselectrodes namely F3 and F4 for classificationInternational Affective Picture System (IAPS).Hz), theta (4–8 Hz), alpha (8–16 Hz), beta (16has been extractedfor every class of emotionscombinations of the Entropy attribute extracted from the different frequency bands. using LIBSVM(3 fold cross validation)with RBF kernel.accuracy remains consistently high whenthe combination of entropies extracted from all five frequency bandclassification. The maximum accuracy of 68.50Keywords: Emotions, Entropy, LIBSVM.

1. Introduction

Emotions play a central role in our interaction with outside world. It can be defined as a that occurs unexpectedly by an individual [1].making tasks are highly affected by people’s moodswerespecified by researchers. The first model was Ekman.Plutchikconsidered that emotions such as curiosity, disgust, joy, surprise, anger, joy, fear, acceptance, and sadness are eight basic emotions [2]. disgust, sadness and surprisearesix basic feeling of human beingand there is little consent on the meaning of the termrigorously suffering from stroke, paralysis, Amyotrophic lateral sclerosis (ALS) and other brain related disorders may lose their ability to speak or to communicate with environmentemotion classifier. For example, driver of bus driving on road in a bad moodother traveler in the same bus.Many researchers give literature on recognition of emotions based on facial expression, speech and gesture of human being but these features results into inaccurate emotional states. So emotion recognition from physiological signals such as Blood Volume PuGalvanic Skin Response (GSR) and Electroencephalogram (EEG) EEG signals can also be used for detecting epileptic activity of brain through which emotions can be recognized oepileptic person[7].Russell, J.A. (1980) demonstrated in a two dimensional space. According to his approach pleased was unpleased was characterizing along y-axis.words were closer while the opposites was arrangedwhere the two dimensions were differentiatedSome basic emotions namely happy, sad, surprised, angry, excited, frustrated, calm, shocked and naughty which can be seen in our daily routine shown in Fig

��

�� !!��""##$$��

motion classification using EEG EntropyMandeep Singh [1] Mooninder Singh[2] Money Goyal[3]

[1, 2, 3]Department of Electrical & Instrumentation Engineering, Thapar University, Patiala, India

[email protected] [2] [email protected][3] [email protected]

Human emotions can be expressed as a state of mind which can predict the response of a person whether positive or . In this paper, the aim of the project is to design an interactive and a smar

Electroencephalogram (EEG) signals. EEG signals are acquired from three subjects classification of emotions into two classes. The stimulus to subjects is provided

ystem (IAPS).The EEG data isdecomposed into five different frequency bands namely delta (016 Hz), beta (16–32 Hz) and gamma (32- 64Hz) by using filtering technique.

for every class of emotionsfrom the five frequency bands.The training and testing combinations of the Entropy attribute extracted from the different frequency bands. The classification of emotions

)with RBF kernel.When emotions are classified subject wise, the combination of entropies extracted from all five frequency band68.50% has been achievedonF3 electrode and 66.66% onF4 electrode.

Entropy, LIBSVM.

role in our interaction with outside world. It can be defined as a psychologicalby an individual [1].In our everyday activities, such as social communication and affected by people’s moods and different emotional states.Different models on emotions

The first model was specified by Darwin which was further carry on by Plutchik and considered that emotions such as curiosity, disgust, joy, surprise, anger, joy, fear, acceptance, and

]. Whereas Ekman considered thatemotions namely fear,aresix basic feeling of human being [3].There are more than 90 definitions of "emotion"

and there is little consent on the meaning of the term [4].Without the emotion identification, even the personrigorously suffering from stroke, paralysis, Amyotrophic lateral sclerosis (ALS) and other brain related disorders

lose their ability to speak or to communicate with environment[5]. Sothis necessitatesdesigninganFor example, driver of bus driving on road in a bad moodis not only risky to himself but also to

Many researchers give literature on recognition of emotions based on facial expression, speech and gesture of human being but these features results into inaccurate emotional states. So

physiological signals such as Blood Volume Pulse (BVP), Electrocardiogram (ECG), nd Electroencephalogram (EEG) is important for the study of emotions [6]

EEG signals can also be used for detecting epileptic activity of brain through which emotions can be recognized oRussell, J.A. (1980) projected a circumflex model of affect in which emotions were

in a two dimensional space. According to his approach pleased was characterizingaxis. The emotion describing words were so arranged that the like meaning

words were closer while the opposites was arranged diagonally [8]. Later this model was enhanceddifferentiated by valence and arousal [9].

Some basic emotions namely happy, sad, surprised, angry, excited, frustrated, calm, shocked and naughty which can igure 1.

��% % % �&'(�� ''�&��

��)�*��+��

using EEG Entropy

[email protected]

person whether positive or smart system for emotion

acquired from three subjects on two frontal stimulus to subjects is providedby images from

decomposed into five different frequency bands namely delta (0-4 64Hz) by using filtering technique. The entropy attribute

from the five frequency bands.The training and testing is performed on six of emotions is performed

, ithas beenfound that the combination of entropies extracted from all five frequency bands has been used for

F4 electrode.

psychological state of mind ommunication and decision

models on emotions which was further carry on by Plutchik and

considered that emotions such as curiosity, disgust, joy, surprise, anger, joy, fear, acceptance, and namely fear, happiness, anger,

There are more than 90 definitions of "emotion" Without the emotion identification, even the personwho is

rigorously suffering from stroke, paralysis, Amyotrophic lateral sclerosis (ALS) and other brain related disorders sitatesdesigningan affective

is not only risky to himself but also to Many researchers give literature on recognition of emotions based on facial

expression, speech and gesture of human being but these features results into inaccurate emotional states. So lse (BVP), Electrocardiogram (ECG),

study of emotions [6]. As EEG signals can also be used for detecting epileptic activity of brain through which emotions can be recognized of

a circumflex model of affect in which emotions were characterizing the x-axis and

The emotion describing words were so arranged that the like meaning enhanced by Lang et al

Some basic emotions namely happy, sad, surprised, angry, excited, frustrated, calm, shocked and naughty which can

��

�

�

2. Review

Kim, K.Het al. (2004) collected physiological database such as electrocardiogram and skin temperature variations from fifty participants. The emotions were classified into four emotional states namely sad, anger, stress, surprise build a user-independent system for recognizing emotions. The classification was performed by support vector machine as pattern classifier with ratios obtained wrespectively [10].Murugappan, M et al. (2009) acquired EEG data when stimulated by audiohealthy subjects. The classification of emotions was doneneutral. Using wavelet transforms, the EEG signal was decomposed into five different frequency bands (delta, theta, alpha, beta and gamma). Through the use of 5 fold cross validation withLinear discriminate analysis (LDANearest neighbor (KNN) classifier, classification was performed.maximum subset emotion rate of 73.75% on neutral emotions was achieved[11]. Petrantonakis, P.Cfrom three EEG channels, namely Fp1, Fp2 and a bipolar channel of F3 and F4 according to position of international 10–20 system. The emotions were classified into six badisgust, and sadness when novel emotion elicitation method based on the Mirror Neuron System was used to promote emotion stimulation. The extracted features using different classifiers namely quadratic discriminant analysis (QDA), support vector machines (SVMs) for recognition of emotions. The EEG data examined from from combined-channels results in accuracy of 62.3% using QDA and 8Anh, V.Het al. (2012) collected EEG data when images from IAPS dataset were used for evoking emotionsTwo approaches from Russell’s circumplex model were used. The algorithm used was Higuchi Fractal Dimension with Support Vector Machine as classifier. First approach was machine learning in which EEG signals of all the participants were taken under consideration and second approach windividual participants were taken. Since EEG signal of every subject has different characteristic so results showed that first approach was impossible to apply but second approach of individual subjects underand can recognize five states of human ememotions has been used by number of other researchers for recognition of emotions and different event related potential and average of event related potential. The classification was done using support vmachines as classifier [15-16]. Singh, M. et al. (2013) used the eNTERFACE 06 EEG database for classifying emotions into two classesDifferent attributes such as Power Spectral Density (PSD), ShorPotential (ERP), power, Entropy and varianceclassifiers such as ANN and Naïve Bayes

Happy

Fear

��

�� !!��""##$$��

Figure 1: Basic Emotions

(2004) collected physiological database such as electrocardiogram and skin temperature variations The emotions were classified into four emotional states namely sad, anger, stress, surprise

independent system for recognizing emotions. The classification was performed by support vector machine as pattern classifier with ratios obtained was 78% and 61.8% in case of three and four d

et al. (2009) acquired EEG data when stimulated by audio-visual stimuliThe classification of emotions was done into five classes such as disgust, happy, surprise, fear and

he EEG signal was decomposed into five different frequency bands (delta, theta, Through the use of 5 fold cross validation withLinear discriminate analysis (LDA

Nearest neighbor (KNN) classifier, classification was performed. An average accuracy of 79.14% was 73.75% on fear, 91% on disgust, 60% on surprise,88% on happy,

Petrantonakis, P.Cet al. (2010) collected EEG data from sixteen healthy subjects from three EEG channels, namely Fp1, Fp2 and a bipolar channel of F3 and F4 according to position of international

20 system. The emotions were classified into six basic emotions states namely happiness, surprisenovel emotion elicitation method based on the Mirror Neuron System was used to . The extracted features using higher order crossings (HOC) were classified with four

different classifiers namely quadratic discriminant analysis (QDA), k-nearest neighbor, Mahalanobis distance and support vector machines (SVMs) for recognition of emotions. The EEG data examined from

channels results in accuracy of 62.3% using QDA and 83.33% using SVM respectively [12(2012) collected EEG data when images from IAPS dataset were used for evoking emotions

’s circumplex model were used. The algorithm used was Higuchi Fractal Dimension with Support Vector Machine as classifier. First approach was machine learning in which EEG signals of all the participants were taken under consideration and second approach was the machine learning in which EEG signal of individual participants were taken. Since EEG signal of every subject has different characteristic so results showed that first approach was impossible to apply but second approach of individual subjects under consideration was used and can recognize five states of human emotion average accuracy 70.5% [14]. The IAPS dataset for evoking of emotions has been used by number of other researchers for recognition of emotions and different

ated potential and average of event related potential. The classification was done using support v

) used the eNTERFACE 06 EEG database for classifying emotions into two classesPower Spectral Density (PSD), Short Time Fourier Transform (STFT), e

, power, Entropy and varianceon the basis of time frequency domain were extracted and the classifiers such as ANN and Naïve Bayes were used for classification of emotions into two classes

Basic Emotions�

Sad

Surprised

Disgust

Fear

,�* -�

��% % % �&'(�� ''�&��

��)�*��+��

(2004) collected physiological database such as electrocardiogram and skin temperature variations The emotions were classified into four emotional states namely sad, anger, stress, surprise to

independent system for recognizing emotions. The classification was performed by support vector as 78% and 61.8% in case of three and four distinct emotions

visual stimuli from 20 disgust, happy, surprise, fear and

he EEG signal was decomposed into five different frequency bands (delta, theta, Through the use of 5 fold cross validation withLinear discriminate analysis (LDA) & K

An average accuracy of 79.14% was obtainedwith happy, and 87.5 % on

(2010) collected EEG data from sixteen healthy subjects from three EEG channels, namely Fp1, Fp2 and a bipolar channel of F3 and F4 according to position of international

surprise, anger, fear, novel emotion elicitation method based on the Mirror Neuron System was used to

higher order crossings (HOC) were classified with four nearest neighbor, Mahalanobis distance and

support vector machines (SVMs) for recognition of emotions. The EEG data examined from single-channel and 3.33% using SVM respectively [12].

(2012) collected EEG data when images from IAPS dataset were used for evoking emotions [13]. ’s circumplex model were used. The algorithm used was Higuchi Fractal Dimension

with Support Vector Machine as classifier. First approach was machine learning in which EEG signals of all the as the machine learning in which EEG signal of

individual participants were taken. Since EEG signal of every subject has different characteristic so results showed consideration was used

The IAPS dataset for evoking of emotions has been used by number of other researchers for recognition of emotions and different attributes such as

ated potential and average of event related potential. The classification was done using support vector

) used the eNTERFACE 06 EEG database for classifying emotions into two classes [17]. t Time Fourier Transform (STFT), event Related

on the basis of time frequency domain were extracted and the were used for classification of emotions into two classes [18-22].

��

�

�

Murugappan, M. et al.(2013) acquired EEG data when stimulated by audioThe emotion was performed into five classes such asextracted was spectral centroid and spectral entropy when EEG data was decomposed into four frequency bands namely alpha (8 Hz - 16 Hz), beta (16 Hz Fast Fourier Transform (FFT). The classification was carried out using kNeural Network (PNN). The highest accuracy of 91.33% on beta band was obtained which was indicating that KNN classifier was perform better than PNN 3. Data acquisition

For classification of emotions, EEG data vision and have no health issue. The MP150interfaced with EEG cap. EEG cap has totalused in unipolar or biopolar [24]. EEG cap frontal electrodes namely F3 and F4areSystem (IAPS) provided by national institute of mental health centre of emotions (NIMH) evoking emotions of participants. These images were collected by shown to subjectsare corresponding to High Valence High Arousal and High Valence Low Afor acquiring EEG signal iscomposed of EEG amplifiers, EEG gel used in electrode, earlobes attached to the participants. To make contact between electrodes, EEG gel corresponding to High Valence High Abackground for a period of 1.5 second. With suitable coordination between the data acquisitionsystem of images, EEG data has been collected at sampling rate of 500 samples per second The methodology followed for acquisition and

Figure 2:

4. Hardware

To capture EEG signal from human being for classification of emotion, there can be too many ways.used by Savran et al. (2006) was functional Near Infrared Spectroscopy (fNIRS) sensor to capture frontal brain activity, to capture activity in the rest of the brain and for acquiring peripheral body processes such as respiration belt,a plethysmograph (blood volume pressure) and GSR (Galvanic Skin Response)acquisition unit MP150is used which is a computer based acquisition system. The MP system namely MP150 or MP100 is a heart of data acquisition system. This MP150 unit takes incoming signals and converts this incoming signal into digital signal whichis processed by computer. Thismemory of computer which can be used for further exami 5. Conditioning of EEG signal

After the acquisition of EEG signal, conditioningresults. The acquired EEG signal is preprocessed for removal of any interference.This interference may be due to blink of eyes or any body movement of participantusing ACQ4.2 Software provided by BIOPACfrequency of 40 Hz and high pass IIR filter with cut off frequency of 0.5 Hz. The noise interference comb band pass filter with notch frequency of 50 Hz

EEG Data Acquisition

��

�� !!��""##$$��

(2013) acquired EEG data when stimulated by audio-visual stimuli from 20 healthy subjects. The emotion was performed into five classes such as disgust, happy, surprise, fear and neutral.Textracted was spectral centroid and spectral entropy when EEG data was decomposed into four frequency bands

16 Hz), beta (16 Hz - 32 Hz), gamma (32 Hz - 60Hz) and alpha to gamma (8 Hz st Fourier Transform (FFT). The classification was carried out using k-Nearest Neighbor (KNN) and Probabilistic

The highest accuracy of 91.33% on beta band was obtained which was indicating that KNN classifier was perform better than PNNclassifier [23].

For classification of emotions, EEG data is acquired from three healthy male subjects with vision and have no health issue. The hardware used for capturing brain signals isBIOPAC data

EEG cap has total of 20 electrodes as per International 10-20 system that can be EEG cap is placed over the head of participant whose data is to capture. Only two

are used for classification. The images from International Affective Picture ystem (IAPS) provided by national institute of mental health centre of emotions (NIMH) are

evoking emotions of participants. These images were collected by the scientist of university of Florida. The images High Valence High Arousal and High Valence Low Arousal

composed of EEG amplifiers, EEG gel used in electrode, earlobes attached to the To make contact between electrodes, EEG gel is used in the electrodes. First images shown t

corresponding to High Valence High Arousal for a period of 1second followed by plus symbol with a black background for a period of 1.5 second. With suitable coordination between the data acquisition

collected at sampling rate of 500 samples per second [25].

acquisition and classification of emotions was as shown in Figure

Methodologyused for classification of emotions

To capture EEG signal from human being for classification of emotion, there can be too many ways.used by Savran et al. (2006) was functional Near Infrared Spectroscopy (fNIRS) sensor to capture frontal brain

, to capture activity in the rest of the brain and for acquiring peripheral body processes such as respiration mograph (blood volume pressure) and GSR (Galvanic Skin Response) [17]. In this BIOPACdata

used which is a computer based acquisition system. The MP system namely MP150 or MP100 is a heart of data acquisition system. This MP150 unit takes incoming signals and converts this incoming

processed by computer. This processed signal is displayed on screen and stored in used for further examination [26].

onditioning of signal is very important aspect for obtaining better classification preprocessed for removal of any interference.This interference may be due to

blink of eyes or any body movement of participants under concern. The signal ispreprocessed in offline mode by Software provided by BIOPAC [27].The signal is filtered through low pass IIR filter with cut off

frequency of 40 Hz and high pass IIR filter with cut off frequency of 0.5 Hz. The noise interference comb band pass filter with notch frequency of 50 Hz. Figure 3 to 6shows filtering operations step by step as shown:

EEG Signal conditioning

Feature Extraction

��% % % �&'(�� ''�&��

��)�*��+��

visual stimuli from 20 healthy subjects. , fear and neutral.The attributes

extracted was spectral centroid and spectral entropy when EEG data was decomposed into four frequency bands 60Hz) and alpha to gamma (8 Hz - 60 Hz) using

KNN) and Probabilistic The highest accuracy of 91.33% on beta band was obtained which was indicating that

with normal to corrected data acquisition unit

20 system that can be to capture. Only two

nternational Affective Picture are used as stimuli for

scientist of university of Florida. The images rousal. The systemused

composed of EEG amplifiers, EEG gel used in electrode, earlobes attached to the First images shown to subject is

rousal for a period of 1second followed by plus symbol with a black background for a period of 1.5 second. With suitable coordination between the data acquisition system and display

ure 2.

To capture EEG signal from human being for classification of emotion, there can be too many ways. One method used by Savran et al. (2006) was functional Near Infrared Spectroscopy (fNIRS) sensor to capture frontal brain

, to capture activity in the rest of the brain and for acquiring peripheral body processes such as respiration ]. In this BIOPACdata

used which is a computer based acquisition system. The MP system namely MP150 or MP100 is a heart of data acquisition system. This MP150 unit takes incoming signals and converts this incoming

displayed on screen and stored in

is very important aspect for obtaining better classification preprocessed for removal of any interference.This interference may be due to

eprocessed in offline mode by filtered through low pass IIR filter with cut off

frequency of 40 Hz and high pass IIR filter with cut off frequency of 0.5 Hz. The noise interference is removed by shows filtering operations step by step as shown:

Classification

��

�

�

signal Figure 4: Low pass IIR filter

Comb band stop filter

6. Feature extraction

The preprocessed EEG data has beenwaveforms that fall primarily with five frequency bands. Different researches give different ways to work on these bands, the bands chosen lies in the following frequency manner: delta (0beta (16–32 Hz) and gamma (32- 64Hz). Feature extracted from these Entropy. It is considered as the strongest feature for emotion classification and signal [28]. EEG preprocessed signal with fivefrequency

EEG signal

band

band

7. Results and Discussion

The classification has been done into two classes’ namely high Arousal.Theclassification has beencarried out subject wise using LIBSVM (for classification of emotions along arousal axis each subject from five decomposed bands of EEG signal.

��

�� !!��""##$$��

Figure 3:Before preprocessing of Low pass IIR filter

Figure 5:High passIIR filter

has been usedforattribute extraction. EEG spectrum contains some characteristic waveforms that fall primarily with five frequency bands. Different researches give different ways to work on these bands, the bands chosen lies in the following frequency manner: delta (0-4 Hz), theta (4–8 Hz), alp

64Hz). Feature extracted from these five frequency bands of EEG signal It is considered as the strongest feature for emotion classification and defined as degree of randomness of

frequency bands as shown in Figure 7 to 12.

Figure EEG signal Figure 8:Gamma band

Figure 9:Deltaband

Figure 11:Beta band

done into two classes’ namely high Valence High Arousal and High Valence Low carried out subject wise using LIBSVM (3 fold cross validation

for classification of emotions along arousal axis [29].The training and testing has been carried out separately for each subject from five decomposed bands of EEG signal. For training,70% of the samples have

��% % % �&'(�� ''�&��

��)�*��+��.�

Before preprocessing of

IIR filter Figure 6:

EEG spectrum contains some characteristic waveforms that fall primarily with five frequency bands. Different researches give different ways to work on these

8 Hz), alpha (8–16 Hz), bands of EEG signal is

defined as degree of randomness of

Figure 7:Preprocessed

band Figure 10:Theta

bandFigure 12:Alpha

Valence High Arousal and High Valence Low ss validation)withRBF kernel

carried out separately for have been used and the

��

�

�

remaining 30% of the samples has beentraining the model for every subject. Using this best model, testing subject. The classification has been carried out on six combinations of the five sub bands such as beta, gamma of EEG signal.The classification accuracy at F3 and F4 electrode for

Table1:Classification a

Entropy combination Ed Et EaEbEg

Ed Et EaEb Et EaEbEg Ed EaEbEg Ed Et EaEg Ed Et EbEg

It can be seen for subject 1, when taking all frequency bands, an accuracy of 62.5% way when entropy determined by taking alpha beta theta delta combinations of EEG signalhas been obtained for F3 electrode, 46.electrode and 58.33% for F4 electrode when observation. For the entropy combination namely alpha beta delta gamma of EEG signalbeen achieved for F3 and F4 electrode. In the same way an accuracy of 45.83% for F3 electrode and 57.electrodehas been achievedwhen entropy signals. When entropy combination such as beta theta delta gamma of EEG signal for F3 electrode and 45.38% for F4 electrode From Figure 13 it can be seen that the best results are obtained when considering all signal for subject1.

Figure 13:Classification a In the same way for subject 2, the best accuracy of 63.42when all entropy combinations acquired by decomposing an EEG signal into five frequency bands The classification accuracy at F3 and F4 electrode for

/��0 ��..0�� 0�/��0

123��2��2��2��2*4 123��2��2��2��4�

��

�� !!��""##$$��

as been used for testing. The best values of cost factor and gamma training the model for every subject. Using this best model, testing has been performed on the test dataset

carried out on six combinations of the five sub bands such as The classification accuracy at F3 and F4 electrode for subject 1 is shown in Table 1.

Classification accuracy at F3 and F4 electrode for subject 1: Subject 1

Accuracy atF3 electrode (%) Accuracy atF4 electrode62.5 58.89

58.33 46.77 54.16 58.33 58.88 58.88 45.83 57.14 41.88 45.38

It can be seen for subject 1, when taking all entropy combinationsacquired by decomposing an EEG signal an accuracy of 62.5% has been obtained for F3 electrode and 58.89% for F4 electrode. In the same

by taking alpha beta theta delta combinations of EEG signal, an accuracy of 58.33% 46.77% for F4 electrode. An accuracy of 54.16 has been

electrode and 58.33% for F4 electrode when entropy combination such as alpha beta theta gamma For the entropy combination namely alpha beta delta gamma of EEG signal, an accuracy of 58.88%

achieved for F3 and F4 electrode. In the same way an accuracy of 45.83% for F3 electrode and 57.when entropy is determined by taking alpha theta delta gamma combination of EEG

When entropy combination such as beta theta delta gamma of EEG signal is taken, an accuracy of 41.88% for F4 electrode has been obtained.

13 it can be seen that the best results are obtained when considering all five frequency

Classification accuracy of Subject 1 for different Entropycombinations

best accuracy of 63.42% has been achievedfor F3 and 66.66% for F4 electrode all entropy combinations acquired by decomposing an EEG signal into five frequency bands

d F4 electrode for subject 2 is shown in Table 2.

��/0 ��0��.0 ��0�/��0

��..0 ��0 ��0

12��2��2��2*4 123��2��2��2*4 123��2��2��2*4 123��2��2��2*4

��

5. 5�

��% % % �&'(�� ''�&��

��)�*��+��

. The best values of cost factor and gamma is calculatedby the test dataset of that

carried out on six combinations of the five sub bands such as delta, theta, alpha, subject 1 is shown in Table 1.

F4 electrode (%)

EEG signal into five obtained for F3 electrode and 58.89% for F4 electrode. In the same

, an accuracy of 58.33% has been obtained for F3

h as alpha beta theta gamma is taken under , an accuracy of 58.88% has

achieved for F3 and F4 electrode. In the same way an accuracy of 45.83% for F3 electrode and 57.14% for F4 determined by taking alpha theta delta gamma combination of EEG

taken, an accuracy of 41.88%

frequencybands of EEG

combinations

for F3 and 66.66% for F4 electrode all entropy combinations acquired by decomposing an EEG signal into five frequency bands areconsidered.

��0 ��.0

123��2��2��2*4

��

�

�




From Figure 14 it can be seen that the best results are obtained when consideringsignal for subject2.

Figure 14:Classificationa For subject 3, better results has been obtained 68.5% for F3 electrode and 66.66% for F4 electrode. is shown in Table 3.




From Figure 15 it can be seen that the best results are obtained when considering all subject3.

/.��0 /�� 0//�//0��.0

123��2��2��2��2*4 123��2��2��2��4

��

�� !!��""##$$��


Accuracy atF3 electrode (%) Accuracy atF4 electrode63.42 66.66 61.9 52.38

66.66 57.14 57.42 57.14 61.90 61.9 59.89 41.38

14 it can be seen that the best results are obtained when considering all five frequency bands of EEG

Classificationaccuracy of Subject 2 for different Entropycombinations

obtained by taking all band combinations with highest accuracy achieved up to and 66.66% for F4 electrode. The classification accuracy at F3 and F4 electrode for


Accuracy atF3 electrode (%) Accuracy at F4 electrode68.5 66.66

54.16 52.38 45.83 57.14 57.14 57.14 61.90 45.38 41.88 61.9

15 it can be seen that the best results are obtained when considering all five band

//�//0��0 /�� 0��0 ��0 /�� 0

12��2��2��2*4 123��2��2��2*4 123��2��2��2*4

��

5. 5�

��% % % �&'(�� ''�&��

��)�*��+��

F4 electrode (%)

all five frequency bands of EEG

combinations

accuracy achieved up to classification accuracy at F3 and F4 electrode for subject 3

F4 electrode (%)

bands of EEG signal for

� � 0

��.0

123��2��2��2*4

��

�

�

Figure 15:Classification a It can be seenfrom Figure 16 when all fiveand F4 electrode, the accuracy is remaininginalong F3 electrode followed by F4 electrode as showF3 electrode,66.66% forF4 electrode.

Figure 16: Best

It shows that classification of emotions along the arousal can be performedextracted from all five frequency bands

8. Conclusion

In thisstudy, we extracted Entropy attributes from five frequency bands of EEG signal for classification of emotions along arousal axis into two classes. Among the six cthe best combination. The accuracy of classification along arousal axis achieved is 68.50% at F3 electrode and 66.66% at F4 electrode when all five frequency

9. Scope for Future Work

The emotion classifier designed in this study procedure for every subject which is actuallyclassification, so scope could be widened by increasing the number of electrodes.

/��0��/0

//�//0��.0

123��2��2��2��2*4 123��2��2��2��4�

/��0�� 0

6��(�&��

��

5.1�23��2��2��2��2*�4

��

�� !!��""##$$��

Classification accuracy of Subject 3 for different Entropy combinations

when all fivefrequency bands of EEG signal are compared for every subject along F3 is remainingin the range nearly 60% to 70%. The highest accuracy

followed by F4 electrode as shown in Figure16. An accuracy of 68.50% has been

Best classificationaccuracy obtained Subject wise

It shows that classification of emotions along the arousal can be performed best when combination of entropies extracted from all five frequency bands are used for classification.

ntropy attributes from five frequency bands of EEG signal for classification of emotions classes. Among the six combinations used in this study, Ed EtEaEbEghas been

The accuracy of classification along arousal axis achieved is 68.50% at F3 electrode and frequency bands of EEG signal are considered.

The emotion classifier designed in this study is subjective one not universal. It requires repetition of whole e for every subject which is actually a shortcoming. In addition to these only two electrodes

so scope could be widened by increasing the number of electrodes.

��.0��0 /�� 0��0 ��0

��.0

12��2��2��2*4 123��2��2��2*4 123��2��2��2*4

��

5. 5�

/.��0/��0//�//0

6��(�&�� 6��(�&��.

��

5.1�23��2��2��2��2*�4 5�1�23��2��2��2��2*�4

��% % % �&'(�� ''�&��

��)�*��+��/�

combinations

compared for every subject along F3 The highest accuracy has been obtained

has been achieved along

combination of entropies

ntropy attributes from five frequency bands of EEG signal for classification of emotions has been found to be

The accuracy of classification along arousal axis achieved is 68.50% at F3 electrode and

It requires repetition of whole only two electrodes are used for

��0

/�� 0

123��2��2��2*4

//�//0

6��(�&��.

��

�

�

References

[1] Emotions, available at http://www.thefreedictionary.com/emotion[2] R. Plutchik, “The Emotions: Facts, Theories and a New Model[3] P. Ekman, W.V. Friesen, M. O’Sullivan, A. Chan, I. Diacoyanni

LeCompte, T. Pitcairn, and P.E. RicciExpressions of Emotion,” J. Personality and Social Psychology, vol. 53

[4]Robert Horlings, “Emotion recognition using brain activity”, Man2008

[5] A. Choppin, “EEG-based human interface for disabled individuals: Emotion expression with neural networks”, Master's thesis, Tokyo Institute of Technology

[6] R. W. Picard, Affective Computing:[7] Mandeep Singh and SunpreetKaur “

Information Technology and Knowledge Management, vol.6, [8] J.A.Russell, “A Circumflex Model of Affect.” Journalof Personality and Social Psychol

1178, 1980 [9] M.M.Bradley and P.J.Lang “Measuring emotions: The self assement manikin and the semantic differential”

Journal of behavior therapy and experimental psychiarty, vol[10] K.H. Kim, S.W. Bang, S.R. Kim

signals." Medical and biological engineering and computing[11] M.Murugappan, R.Nagarajan, and S.Yaacob “Comparison of Different Wavelet Features from EEG Signals

Classifying Human Emotions” IEEE Symposium on Industrial Electronics and Applications,841, 2009

[12] P.C.Petrantonakis and L.J.Hadjileontiadis. "Emotion recognition from EEG using higher order crossings"Transactions on information Technology in Biomedicine, vol.

[13] M. M, Bradley and P. J. Lang “The International Affective Picture System (IAPS) in the Study of Emotion and Attention”, Handbook of Emotion Elicitation and Assessment

[14] V.H.Anh, M.N.Van, B.B.Ha and T.H.Quyetrecognition through EEG", IEEE pp.191-196, 2012

[15] Mandeep Singh, Mooninder Singh and AnkitaSignals” International Journal of Information Technology & Knowledge Management,vol. 7152, June 2014

[16] Mandeep Singh, Mooninder Singh Averaged ERP Signals” International Journal of Information Technology & Knowledge Management, vol. 7no. 2, pp. 153-161, June 2014

[17] A. Savran, K. Ciftci, G. Chanel, J.C. Mota, L.H. Viet“Emotion detection in the loop from brain signals and facial images” .IEE Computer society vol.2issue1 pp. 1831,Doi http://doi.ieeecomputersociety.org/10.110

[18]Mandeep Singh, Mooninder Singh and Nikhil Singhalbayes Classifier” International Journal of Information Technology & Knowledge Management, pp.51-55, Dec 2013

[19] Mandeep Singh, Mooninder Singh and NikhUsing EEG”, International Journal of Information Technology & Knowledge Management, vol.7no.1, pp.56Dec 2013

[20] Mandeep Singh, Mooninder Singh and Nikhil Signals”, International Journal of Information Technology & Knowledge Ma

[21] Mandeep Singh, Mooninder Singh and Classification” International Journal of Information Technology & Knowledge Management,10, Dec 2013

[22] Mandeep Singh, Mooninder Singh and Surbhi Gangwar Journal of Information Technology & Knowledge Managem

[23] M.Murugappan and S.Murugappan “(EEG) Signals Using Fast Fourier Transform (FFT)” and its Applications, 2013

��

�� !!��""##$$��

http://www.thefreedictionary.com/emotion R. Plutchik, “The Emotions: Facts, Theories and a New Model”, Random House, New York, 1962

[3] P. Ekman, W.V. Friesen, M. O’Sullivan, A. Chan, I. Diacoyanni-Tarlatzis, K. Heider, R. Krause, W.A. LeCompte, T. Pitcairn, and P.E. Ricci-Bitti, “Universals and Cultural Differences in the Judgments of Facial Expressions of Emotion,” J. Personality and Social Psychology, vol. 53, no. 4, pp. 712-717, Oct

Robert Horlings, “Emotion recognition using brain activity”, Man-Machine Interaction Group TU Delft

based human interface for disabled individuals: Emotion expression with neural networks”, Master's thesis, Tokyo Institute of Technology, 2000

: MIT Press, 1997 reetKaur “Epilepsy Detectionusing EEG anoverview” International Journal of

Information Technology and Knowledge Management, vol.6, no.1, pp.3-5, Dec 2102 ] J.A.Russell, “A Circumflex Model of Affect.” Journalof Personality and Social Psychology, vol.39,

] M.M.Bradley and P.J.Lang “Measuring emotions: The self assement manikin and the semantic differential” Journal of behavior therapy and experimental psychiarty, vol. 25,pp.49-59,March 1994 K.H. Kim, S.W. Bang, S.R. Kim "Emotion recognition system using short-term monitoring of physiological

Medical and biological engineering and computing, vol. 42, pp.419-427, 2004 [11] M.Murugappan, R.Nagarajan, and S.Yaacob “Comparison of Different Wavelet Features from EEG Signals

Classifying Human Emotions” IEEE Symposium on Industrial Electronics and Applications,

[12] P.C.Petrantonakis and L.J.Hadjileontiadis. "Emotion recognition from EEG using higher order crossings"information Technology in Biomedicine, vol.14, Issue.2 pp. 186-197, 2010

] M. M, Bradley and P. J. Lang “The International Affective Picture System (IAPS) in the Study of Emotion and Attention”, Handbook of Emotion Elicitation and Assessment, Oxford University Press, 2007

B.B.Ha and T.H.Quyet"A real-time model based support vector machine for emotion International Conference onControl, Automation and Information Sciences

Mandeep Singh, Mooninder Singh and AnkitaSandel“Emotion Classification along Valence Axis Using ERP International Journal of Information Technology & Knowledge Management,vol. 7

Mandeep Singh, Mooninder Singh and AnkitaSandel “Emotion Classification along Valence Axis Using International Journal of Information Technology & Knowledge Management, vol. 7

] A. Savran, K. Ciftci, G. Chanel, J.C. Mota, L.H. Viet, B. Sankur, L. Akarun, A. Caplier, and M. Rombaut “Emotion detection in the loop from brain signals and facial images” .IEE Computer society vol.2issue1 pp. 1831,Doi http://doi.ieeecomputersociety.org/10.1109/T-AFFC.2011.15,jan-march 2012

gh, Mooninder Singh and Nikhil Singhal“Emotion Recognition along valence Axis Using naïve bayes Classifier” International Journal of Information Technology & Knowledge Management,

Mandeep Singh, Mooninder Singh and Nikhil Singhal“ANN Based Emotion Recognition Along Valence Axis Using EEG”, International Journal of Information Technology & Knowledge Management, vol.7no.1, pp.56

ooninder Singh and Nikhil Singhal “Emotion Quantification along Valence Axis Using EEG Signals”, International Journal of Information Technology & Knowledge Management, vol.7, no.1, Dec

Singh and SurbhiGangwar “Feature Extraction from EEG for Emotion al Journal of Information Technology & Knowledge Management,

Mandeep Singh, Mooninder Singh and Surbhi Gangwar “Arousal Detection Using EEG Signal” International Journal of Information Technology & Knowledge Management, vol.7, no.1, pp.11-15, Dec 2013

Murugappan “Human Emotion Recognition through Short time Electroencephalogram (EEG) Signals Using Fast Fourier Transform (FFT)” 9th IEEE International Colloquium on

��% % % �&'(�� ''�&��

��)�*��+��

”, Random House, New York, 1962 Tarlatzis, K. Heider, R. Krause, W.A.

Cultural Differences in the Judgments of Facial 7, Oct 1987

hine Interaction Group TU Delft 27 March

based human interface for disabled individuals: Emotion expression with neural networks”,

” International Journal of

ogy, vol.39, pp.1161-

] M.M.Bradley and P.J.Lang “Measuring emotions: The self assement manikin and the semantic differential”

term monitoring of physiological

[11] M.Murugappan, R.Nagarajan, and S.Yaacob “Comparison of Different Wavelet Features from EEG Signals for Classifying Human Emotions” IEEE Symposium on Industrial Electronics and Applications, vol. 2, pp. 836-

[12] P.C.Petrantonakis and L.J.Hadjileontiadis. "Emotion recognition from EEG using higher order crossings" IEEE

] M. M, Bradley and P. J. Lang “The International Affective Picture System (IAPS) in the Study of Emotion and versity Press, 2007

time model based support vector machine for emotion International Conference onControl, Automation and Information Sciences,

Emotion Classification along Valence Axis Using ERP International Journal of Information Technology & Knowledge Management,vol. 7, no. 2, pp. 143-

Emotion Classification along Valence Axis Using International Journal of Information Technology & Knowledge Management, vol. 7,

, B. Sankur, L. Akarun, A. Caplier, and M. Rombaut “Emotion detection in the loop from brain signals and facial images” .IEE Computer society vol.2issue1 pp. 18-

“Emotion Recognition along valence Axis Using naïve bayes Classifier” International Journal of Information Technology & Knowledge Management, vol. 7,no.1,

“ANN Based Emotion Recognition Along Valence Axis Using EEG”, International Journal of Information Technology & Knowledge Management, vol.7no.1, pp.56-60,

along Valence Axis Using EEG nagement, vol.7, no.1, Dec 2013

Gangwar “Feature Extraction from EEG for Emotion al Journal of Information Technology & Knowledge Management,vol.7, no.1, pp.6-

“Arousal Detection Using EEG Signal” International 2013

ime Electroencephalogram 9th IEEE International Colloquium on.SignalProcessing

��

�

�

[24]V.Jurcak, D.Tsuzuki and I.Dan, “10.20, 10/10 and 10/5 systems revisited, their validity as based positioning systems” Neuro image, vol. 3

[25] Mandeep Singh, Mooninder Singh and Mclassification of emotions along arousal axis”Management,vol.8, no.2, 2015

[26] MP system hardware available at http://www.biopac.com/manuals/mp_hardware_guide.pdf[27] Acqknowledge 4 software guide, available at,http://www.biopac.com

/Manuals/acqknowledge%204%20software%20guide.pdf, pp.31[28]Mandeep Singh and HarleenKaur “ANN based Epilepsy detection using EEG

Information Technology & Knowledge Management, [29]Chang, Chih-chung, Lin, Chih

www.csie.ntu.edu..tw/~cjlin/libsvm/ ,2011 [June13,

��

�� !!��""##$$��

V.Jurcak, D.Tsuzuki and I.Dan, “10.20, 10/10 and 10/5 systems revisited, their validity as relative head surface based positioning systems” Neuro image, vol. 34, Issue. 4, pp.1600-1611, 2007

Singh and Money Goyal“Selection of attribute combinations of ERP’s for classification of emotions along arousal axis”International Journal of Information Technology & Knowledge

MP system hardware available at http://www.biopac.com/manuals/mp_hardware_guide.pdf] Acqknowledge 4 software guide, available at,http://www.biopac.com

0software%20guide.pdf, pp.31-95 HarleenKaur “ANN based Epilepsy detection using EEG” International Journal of

Information Technology & Knowledge Management, vol.7, no.1, pp.31-40, Dec 2013 Lin, Chih-Jen. “LIBSVM: A library for support vector machines”Internet:

www.csie.ntu.edu..tw/~cjlin/libsvm/ ,2011 [June13,2015]

��% % % �&'(�� ''�&��

��)�*��+��

relative head surface

“Selection of attribute combinations of ERP’s for International Journal of Information Technology & Knowledge

MP system hardware available at http://www.biopac.com/manuals/mp_hardware_guide.pdf ] Acqknowledge 4 software guide, available at,http://www.biopac.com

” International Journal of

Jen. “LIBSVM: A library for support vector machines”Internet:

emotion classification using eeg entropy · different classifiers namely quadratic discriminant...

Documents