Emotion Recognition from EEG Signals Using

Research ArticleEmotion Recognition from EEG Signals Using MultidimensionalInformation in EMD Domain

Ning Zhuang1 Ying Zeng12 Li Tong1 Chi Zhang1 Hanming Zhang1 and Bin Yan1

1China National Digital Switching System Engineering and Technological Research Center Zhengzhou 450002 China2Key Laboratory for NeuroInformation of Ministry of Education School of Life Science and TechnologyUniversity of Electronic Science and Technology of China Chengdu China

Correspondence should be addressed to Bin Yan ybspacehotmailcom

Received 27 March 2017 Revised 21 June 2017 Accepted 16 July 2017 Published 16 August 2017

Academic Editor Robertas Damasevicius

Copyright copy 2017 Ning Zhuang et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

This paper introduces a method for feature extraction and emotion recognition based on empirical mode decomposition (EMD)By using EMD EEG signals are decomposed into IntrinsicMode Functions (IMFs) automaticallyMultidimensional information ofIMF is utilized as features the first difference of time series the first difference of phase and the normalized energyTheperformanceof the proposed method is verified on a publicly available emotional database The results show that the three features are effectivefor emotion recognitionThe role of each IMF is inquired and we find that high frequency component IMF1 has significant effect ondifferent emotional states detectionThe informative electrodes based on EMD strategy are analyzed In addition the classificationaccuracy of the proposed method is compared with several classical techniques including fractal dimension (FD) sample entropydifferential entropy and discrete wavelet transform (DWT) Experiment results on DEAP datasets demonstrate that our methodcan improve emotion recognition performance

1 Introduction

Emotion plays an important role in our daily life and workReal-time assessment and regulation of emotionwill improvepeoplersquos life andmake it better For example in the communi-cation of human-machine-interaction emotion recognitionwill make the process more easy and natural Anotherexample in the treatment of patients especially those withexpression problems the real emotion state of patients willhelp doctors to provide more appropriate medical care Inrecent years emotion recognition from EEG has gainedmassattention Also it is a very important factor in brain computerinterface (BCI) systems which will effectively improve thecommunication between human and machines [1]

Various features and extraction methods have been pro-posed for emotion recognition from EEG signals includingtime domain techniques frequency domain techniques jointtime-frequency analysis techniques and other strategies

Statistics of EEG series that is first and second differencemean value and power are usually used in time domain [2]Nonlinear features including fractal dimension (FD) [3 4]

sample entropy [5] and nonstationary index [6] are utilizedfor emotion recognition Hjorth features [7] had also beenused in EEG studies [8 9] Petrantonakis and Hadjileon-tiadis introduced higher order crossings (HOC) featuresto capture the oscillatory pattern of EEG [10] Wang etal extracted frequency domain features for classification[11] Time-frequency analysis is based on the spectrum ofEEG signals then the energy power power spectral density(PSD) and differential entropy [12] of certain subband areusually utilized as features Short-time Fourier transform(STFT) [13 14] Hilbert-Huang transform (HHT) [15 16]and discrete wavelet transform (DWT) [17ndash19] are the mostcommonly used techniques for spectrum calculating It hasbeen commonly tested and verified that higher frequencysubband such as Beta (16ndash32Hz) and Gamma (32ndash64Hz)bands outperforms lower subband for emotion recognition[20 21]

Other features extracted from combination of electrodeare utilized too such as coherence and asymmetry of elec-trodes in different brain regions [22ndash24] and graph-theoreticfeatures [25] Jenke et al had done a research comparing the

HindawiBioMed Research InternationalVolume 2017 Article ID 8317357 9 pageshttpsdoiorg10115520178317357

2 BioMed Research International

EEG signals

IMF1

Feature extraction

SVMclassier

RecognitionresultsEMD

IMF2

IMF3

IMF4

IMF5

(5 s)

Figure 1 Block diagram of the proposed method

performance of different features mentioned above and got aguiding rule for feature extraction and selection [26]

Some other strategies such as utilizing deep networkto improve the classification performance have also beenresearched Zheng and Lu used deep neural network toinvestigate critical frequency bands and channels for emotionrecognition [27] Yang et al used hierarchical network withsubnetwork nodes for emotion recognition [28]

EMD is proposed by Huang et al in 1998 [29] UnlikeDWT which needs to predetermine transform base functionand decomposition level EMD can decompose signals intoIMF automatically These IMFs represent different frequencycomponents of original signals with band-limited charac-teristic By applying Hilbert transform to IMF we can getinstantaneous phase information of IMF So EMD is suitablefor analysis of nonlinear and nonstationary sequence such asneural signals

EMD is a good choice for EEG signals and we utilize itfor emotion recognition from EEG data Which feature iseffective for emotion recognition in EMD domain WhichIMF component is best for classification Is the performancebased on EMD strategy better compared to time domainmethod and time-frequency method or not All these havenot been researched yet and we investigate them in ourresearch

EMD has been widely used for seizure prediction anddetection but for emotion recognition based on EMD thereis not so much research Higher order statistics of IMFs [30]geometrical properties of the decomposed IMF in complexplane [31] and the variation and fluctuation of IMF [32]are used as features for seizure prediction and detectionFor emotion recognition Mert and Akan extracted entropypower power spectral density correlation and asymmetry ofIMF as features and then utilized independent componentanalysis (ICA) to reduce dimension of the feature set [33]The classification accuracy is computed with all the subjectsmixed together

In this paper we present an emotion recognition methodbased on EMD We utilize the first difference of IMF timeseries the first difference of the IMFrsquos phase and the nor-malized energy of IMF as features The motivation of usingthese three features is that they depict the characteristicsof IMF in time frequency and energy domain providingmultidimensional information The first difference of timeseries depicts the intensity of signal change in time domain

The first difference of phase measures the change intensity inphase and normalized energy describes the weight of currentoscillation componentThe three features constitute a featurevector which is fed into SVM classifier for emotional statedetection

The proposed method is studied on a publicly availableemotional database DEAP [20] The effectiveness of thethree features is investigated IMF reduction and channelreduction for feature extraction are both discussed whichaim at improving the classification accuracy with less compu-tation complexity The performance is compared with someother techniques including fractal dimension (FD) sampleentropy differential entropy and time-frequency analysisDWT

2 Method

To realize emotional state recognition the EEG signals aredecomposed into IMFs by EMD Three features of IMFs thefluctuation of the phase the fluctuation of the time series andthe normalized energy are formed as a feature vector whichis fed into SVM for classification The whole process of thealgorithm is shown in Figure 1

21 Data and Materials DEAP is a publicly available datasetfor emotion analysis which recorded EEG and peripheralphysiological signals of 32 participants as they watched 40music videos All the music video clips last for 1 minuterepresenting different emotion visual stimuli with gradefrom 1 to 9 Among the 40 music videos 20 are highvalence visual stimuli and 20 are low valence visual stimuliThe situation is exactly the same for arousal dimensionAfter watching the music video participants performed aself-assessment of their levels on arousal valence likingdominance and familiarity with ratings from 1 to 9 EEGwas recorded with 32 electrodes placing according to theinternational 10-20 system Each electrode recorded 63 s EEGsignal with 3 s baseline signal before the trial

In this paper we used the preprocessed EEG data forstudy with sample rate 128Hz and band range 4ndash45Hz EOGartefacts were removed as method in [20] The data was seg-mented into 60-second trials and a 3-second pretrial baselineremoved The binary classifications of valence and arousaldimension are considered We utilized the participantsrsquo self-assessment as label If the participantrsquos rating was lt5 the label

BioMed Research International 3

of valencearousal is low and if the rating was ge5 the label ofvalencearousal is high

Each music video lasts for 1 minute and 5 s EEG signalsare extracted as a sample So for each subject who watched 40music videos we acquire 480 labeled samples

22 Empirical Mode Decomposition EMD decomposes EEGsignals into a set of IMFs by an automatic shifting processEach IMF represents different frequency components oforiginal signals and should satisfy two conditions (1) duringthe whole data set the number of extreme points and thenumber of zero crossings must be either equal or differ atmost by one (2) at each point themean value calculated fromthe upper and lower envelope must be zero [29] For inputsignal 119909(119905) the process of EMD is as follows

(1) Set ℎ(119905) = 119909(119905) and ℎold(119905) = ℎ(119905)(2) Get local maximum and minimum of ℎold(119905)(3) Interpolate the local maximum and minimum with

cubic spline function and get the upper envelope119890max(119905) and lower envelope 119890min(119905)

(4) Calculate the mean value of the upper and lowerenvelope as

119898(119905) = (119890min (119905) + 119890max (119905))2 (1)

(5) Subtract ℎold(119905) with119898(119905)

ℎnew (119905) = ℎold (119905) minus 119898 (119905) (2)

If ℎnew(119905) satisfies the two conditions of IMF then the firstIMF component imf1 is gotten otherwise set ℎold(119905) =ℎnew(119905) and go to step (2) repeating steps (2)ndash(5) until ℎnew(119905)satisfies the two conditions of IMF Finally imf1 is gotten as

imf1 = ℎnew (119905) (3)

(6) If imf119899 is gotten set ℎold(119905) as

ℎold (119905) = ℎold (119905) minus imf119899 (4)

Go to step (2) and repeat steps (2)ndash(5) to get imf119899+1By the iterative process described above 119909(119905) can be

finally expressed as

119909 (119905) =119871

sum119899=1

imf119899 + 119903 (5)

It is a linear combination of IMF components and the residualpart 119903 Figure 2 shows a segment of original EEG signalscorresponding to the first five decomposed IMFs EMDworkslike an adaptive high pass filter It shifts out the fastestchanging component first and as the level of IMF increasesthe oscillation of IMF becomes smoother Each component isband-limited which can reflect the characteristic of instanta-neous frequency

23 Feature Extraction In this paper three features of IMFare utilized for emotion recognition the first difference oftime series the first difference of phase and the normalizedenergyThe first difference of time series depicts the intensityof signal change in time domain The first difference ofphase reveals the change intensity of phase representing thephysical meaning of instantaneous frequency Normalizedenergy describes theweight of current oscillation componentThemotivation of using these three features is that they depictthe characteristics of IMF in time frequency and energydomain utilizing multidimensional information

231 First Difference of IMF Time Series The first differenceof times series119863119905 depicts the intensity of signal change in timedomain Previous research has revealed that the variation ofEEG time series can reflect different emotion states [2] For anIMF component with 119873 points IMFimf1 imf2 imf119873the definition of119863119905 is

119863119905 =1119873 minus 1

119873minus1

sum119899=1

|imf (119899 + 1) minus imf (119899)| (6)

232 First Difference of IMFrsquos Phase Based on EMD EEGis decomposed into multilevel IMFs each IMF being band-limited and representing an oscillation component of originalEEG signals For an119873-point IMF IMFimf1 imf2 imf119873Hilbert transform is applied to it obtaining an analytic signal119911(119899)

119911 (119899) = 119909 (119899) + 119895119910 (119899) (7)

The analytic signal can be further expressed as follows

119911 (119899) = 119860 (119899) 119890119895120593(119899) (8)

where 119860(119899) = radic119909(119899)2 + 119910(119899)2 is the amplitude of 119911(119899) and120593(119899) = arctan(119910(119899)119909(119899)) is the instantaneous phase

First difference of phase119863119901 is defined as

119863119901 =1119873 minus 1

119873minus1

sum119899=1

1003816100381610038161003816120593 (119899 + 1) minus 120593 (119899)1003816100381610038161003816 (9)

which measures the change intensity in phase and representsthe physical meaning of instantaneous frequency

233 Normalized Energy of IMF For an 119873-point IMFIMFimf1 imf2 imf119873 the normalized energy 119864norm isdefined as follows

119864norm =sum119873119899=1 imf2 (119899)sum119873119899=1 1199042 (119899)

(10)

where 119904(119899) is the original EEG signal points So the numeratoris the energy of IMF and the denominator represents theenergy of original EEG data set The normalized energydescribes the weight of current oscillation componentWhenfed into the classifier log(119864norm) is taken as an element of thefeature vector according to [26]


0

minus1

0

1EE

G

minus05

0

05

IMF1

minus05

0

05

IMF2

minus05

0

05

IMF3

minus05

0

05

IMF4

minus05

05

IMF5

200 300 400 500 600 700 800 900 1000100Points

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

Figure 2 EEG signals and the corresponding first five IMFs

24 SVM Classifier The extracted features are fed into SVMfor classification SVM iswidely used for emotion recognition[34 35] which has promising property in many fields In ourstudy LIBSVM is implemented for SVM classifier with radialbasis kernel function and default parameters setting [36]

3 Performance Verification

In the following subsections we test our method on DEAPemotional dataset Training and classifying tasks were con-ducted for each subject independently and we utilized leave-one-trail-out validation to evaluate the classification perfor-mance Each subject watched 40 music video clips and everyvideo clips lasted 1 minute In our experiment we utilized theparticipantsrsquo self-assessment as label Every 5 s EEG signalsare extracted as a sample so for each subject we acquire 480labeled samples

In leave-one-trail-out validation for each subject 468samples extracted from 39 trails were assigned to trainingset and 12 samples extracted from the remaining one trailwere assigned to test set So there was no correlation betweensamples in the training set and the test set Among the total40 trails of one subject each trail will be assigned to the testset once as the validation dataThe 40 results from the 40 testtrails then can be averaged to produce a general estimation for

each subject The final mean accuracy is computed among allthe subjects

31 Effectiveness of the Features for Emotion Recognition Inorder to evaluate the effectiveness of the three features foremotion recognition we first use only one single featurefor classification each time All the experiments in thissubsection are under the condition that the first five IMFcomponents and total 32 electrodes are utilized for featureextraction The training and classifying for each subjectwere conducted respectively and the mean accuracy wascomputed among all the subjects

The mean classification accuracies of three features aregiven in Figure 3 It shows that all the three features candistinguish high level from low level on both valence andarousal dimension higher than random probability of 50For valence dimension the classification accuracy yields6827 6446 and 6107 with features 119863119905 119863119901 and119864norm respectively For arousal dimension the classificationaccuracy yields 6989 6756 and 6376 with features119863119905119863119901 and 119864norm respectively

32 IMF Reduction for Feature Extraction In this subsec-tion we did two experiments to investigate the role ofdifferent IMF components in emotion recognition In the


Table 1 Comparison of performance for different IMFs selected for feature extraction (32 channels) (standard deviation shown inparentheses)

Component Valence ArousalAccuracy () 119905-test (IMF1) Accuracy () 119905-test (IMF1)

IMF1 7041 (705) 119901 = 1 7210 (751) 119901 = 1IMF2 6347 (710) 119901 = 00002 6658 (936) 119901 = 00032IMF3 6145 (857) 119901 = 0 6456 (1052) 119901 = 00019IMF4 5955 (856) 119901 = 0 6399 (1096) 119901 = 00012IMF5 5574 (920) 119901 = 0 6238 (1223) 119901 = 0IMF1-2 6902 (700) 119901 = 04399 7047 (829) 119901 = 01940IMF1ndash3 6847 (669) 119901 = 02705 7008 (810) 119901 = 03116IMF1ndash4 6799 (658) 119901 = 01688 6960 (808) 119901 = 02107IMF1ndash5 6759 (658) 119901 = 01086 6900 (837) 119901 = 01293

Valence Arousal0

10

20

30

40

50

60

70

80

Accu

racy

()

Dt

Dp

EＨＩＬＧ

Figure 3 Classification accuracies of three single features For eachsubject one single feature was extracted from the first five IMFcomponents ldquo119863119905rdquo ldquo119863119901rdquo and ldquo119864normrdquo in the figure are correspondingto the three single features respectively The mean accuracies forall circumstances were computed among all the subjects Errorbars show the standard deviation of the mean accuracies across allsubjects

first experiment each time only one IMF component wasutilized for feature extraction and we analyzed which IMF iseffective for emotion recognition In the second experimentwe further verified whether the combination of multi-IMFswould improve the accuracy

Table 2 gives all the results in detail Standard deviationof the mean accuracies across all subjects is shown inparenthesis ldquoIMF1rdquo ldquoIMF2rdquo ldquoIMF3rdquo ldquoIMF4rdquo and ldquoIMF5rdquoare corresponding to single IMF component ldquoIMF1ndash3rdquo inthe table represents the first three IMFs corresponding toIMF1 IMF2 and IMF3 Similarly ldquoIMF1ndash4rdquo and ldquoIMF1ndash5rdquoare corresponding to the first four IMFs and the first fiveIMFs respectively

It shows that IMF1 yields the best performance 7041for valence and 7210 for arousal As the level increases theperformance decreases sharply The performance of IMF5 isonly 5574 for valence and 6238 for arousal We applied119905-test (120572 lt 005) to examine the performance between only

Table 2 Performance of 8 channels selected for feature extraction(Fp1 Fp2 F7 F8 T7 T8 P7 and P8) (standard deviation shown inparentheses)

PredictLabel

Valence ArousalHigh Low High Low

High 6664 2723 7493 2748Low 2024 3949 1555 35641198651 score 07374 07769Accuracy () 6910 (695) 7199 (777)

IMF1 utilized for feature extraction and other circumstancesThe null hypothesis is ldquothe performance is similarrdquo and if 119901value is larger than 120572 the null hypothesis is accepted Theresults of 119905-test in Table 1 show that the performance of IMF1is more splendid than other single components IMF2 IMF3IM4 and IMF5 with 119901 far less than 005 It also shows thatperformance of multi-IMF combinations is similar to onlyIMF1 utilized for feature extraction with 119901 larger than 005

IMF1 represents the fastest changing component of EEGsignals with the highest frequency characteristic As the levelincreases the oscillation becomes smoother with frequencybecoming lower and lower So we infer that the valence andarousal of emotion relate more tightly to high frequency It isalso coincided with the finding in [26] that Beta (16ndash32Hz)and Gamma (32ndash64Hz) bands are successfully selected moreoften than other bandsThese two bands are higher frequencysubbands of EEG signals

So combining the results of classification accuracy and 119905-test in practical use we just need to extract features fromIMF1 which will save vast time and relieve computationburden because only one level of EMDdecompositionneededto be done

33 Channel Reduction for Feature Extraction Form verifi-cation in Section 32 we know that using component IMF1will achieve good performance In this subsection we willinvestigate which electrodes are informative based on EMDstrategy


Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz

P4 P8 Po3 Po4 O1 Oz O2

minus004

minus003

minus002

minus001

0

001

002

003

004Dt

(a)

Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus003

minus002

minus001

0

001

002

003Dp

(b)

Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus0015

minus001

minus0005

0

0005

001

0015EＨＩＬＧ

(c)

Figure 4 Fisher distance of different channels with subject 1 Features are extracted from component IMF1 For each channel Fisher distanceis calculated among features extracted from 480 labeled emotion samples of subject 1 (a) Fisher distance under feature119863119905 (b) Fisher distanceunder feature119863119901 (c) Fisher distance under feature 119864norm

Fisher distance is an efficient criterion of divisibilitybetween two classes which is broadly used in patternrecognition It computes the ratio of between-class scatterdegree and within-class scatter degree between two classesLarger ratio means larger divisibility of the two classes Inour experiment we used fisher distance to mark importantelectrodes under condition that IMF1 is used for featureextraction For each channel fisher distance is calculatedamong features extracted from one subjectrsquos total 480 labeledemotion samples

Figure 4 gives fisher distance on valence dimension withsubject 1 Figure 4(a) shows that under feature119863119905 electrodesFp1 Fp2 FC6 Cp1 O1 andOz have larger values Figure 4(b)

shows that under feature 119863119901 Fp1 FC6 Cp1 Cp2 O1 OzP7 and P8 have larger values Figure 4(c) shows that underfeature119864norm F7 F8 T7 T8 P7 P8O1O2 andOzhave largervalues

Based on the analysis of all the subjects we selected thefollowing 8 electrodes Fp1 Fp2 F7 F8 T7 T8 P7 and P8for channel reduction verification Table 2 gives 1198651 score andclassification accuracy with 8 channels selected for emotionrecognition We see that 1198651 score is 07374 for valence and07769 for arousalThe classification accuracywith 8 channelsis 6910 for valence and 7199 for arousal slightly lowerthan accuracy with total 32 channels We also applied 119905-testto examine whether the performance of 8 channels is similar


Valence Arousal

FDSampEnDWT + DE (Beta)

DWT + DE (Gamma)Our method

0

10

20

30

40

50

60

70

80

Accu

racy

()

Figure 5 Classification accuracies of different methods ldquoFDrdquoldquoSampEnrdquo and ldquoDErdquo in the figure are corresponding to fractaldimension sample entropy and differential entropy respectivelyThe mean accuracy was computed among all the subjects Errorbars show the standard deviation of the mean accuracies across allsubjects

to total 32 channels The null hypothesis is ldquothe performanceis similarrdquo and if119901 value is larger than 120572 the null hypothesis isaccepted The 119905-test result shows that the performance under8 channels and 32 channels is similar with 119901 = 04194 forvalence and 119901 = 09521 for arousal

So in practical use we just need to extract features fromIMF1 with 8 channels Our offline experiment used every 5 sEEG signals as a labeled emotion sample This infers that ourmethod may provide a new solution for real-time emotionrecognition in BCI systems

34 Results Comparison with Other Methods In this subsec-tion we compared our proposed method with some classicalmethods including fractal dimension (FD) sample entropydifferential entropy and time-frequency analysis DWT Weused box counting for fractal dimension calculating Theparameter for sample entropy SampEn(119898 119903119873)was set as 119903 =02 119898 = 2 and 119873 = 128 We used ldquodb4rdquo decomposition torealize DWTThen the differential entropy of Beta (16ndash32Hz)and Gamma (32ndash64Hz) bands is extracted as features Ourmethodused IMF1 for feature extraction of119863119905119863119901 and119864normFor all the methods 8 selected channels FP1 FP2 F7 F8 T7T8 P7 and P8 are used for feature extraction

From Figure 5 and Table 3 we see that our methodyields the highest accuracy 6910 for valence and 7199for arousal We applied 119905-test (120572 lt 005) to examine the

performance between classical method and our method Thenull hypothesis is ldquothe performance is similarrdquo and if 119901 valueis larger than 120572 the null hypothesis is accepted The resultsof 119905-test in Table 3 show that the performance of our methodis more splendid than fractal dimension sample entropy anddifferential entropy of Beta band with 119901 far less than 005 Italso shows that the performance of ourmethod is similar andbetter than the differential entropy of Gamma band

EMD strategy outperforms time domainmethod includ-ing fractal dimension and sample entropy This is becausecompared to methods in time domain EMD has the advan-tage of utilizing more oscillation information Comparedto time-frequency method DWT EMD can decomposeEEG signals automatically getting rid of selecting transformwindow first The classification accuracy is also higher thanDWT So the experiment results infer that our method basedon EMD strategy is suitable for emotion recognition fromEEG signals

4 Discussion

Emotion recognition from EEG signals has achieved signifi-cant progress in recent years Previous methods are usuallyconducted in time domain frequency domain and time-frequency domain In this paper we propose a method offeature extraction for emotion recognition in EMD domaina new aspect of view By utilizing EMD EEG signals canbe decomposed into different oscillation components namedIMF automatically The characteristics of IMF are utilizedas features for emotion recognition including the first dif-ference of time series the first difference of phase and thenormalized energy

Compared to methods in time domain EMD has theadvantage of utilizing more frequency information Theexperiment results show that the proposed method outper-forms method in time domain such as fractal dimension in[3 4] and sample entropy in [5] Compared to time-frequencymethods such as STFT and DWT EMD can decomposeEEG signals automatically getting rid of selecting transformwindow first The classification accuracy is also higher thanDWT in [18]

We investigate the role of each IMF in emotion classifica-tion Features extracted from IMF1 yield the highest accuracyIMF1 is corresponding to the fastest changing componentof EEG signals so our study confirms the deduction thatemotion is more relative to high frequency componentThis consists with findings in [26] that Beta (16ndash32Hz) andGamma (32ndash64Hz) bands are successfully selected moreoften than other bands

Finally we selected 8 informative channels based onEMDstrategy namely FP1 FP2 F7 F8 T7 T8 P7 and P8 Ourproposed method just needs to extract features from IMF1with 8 channels whichwill save time and relieve computationburden Also in our experiment every 5 s EEG signals areextracted as a sample so it may provide a new solution forreal-time emotion recognition in BCI systems

Our limitation is that nowwe just test it onDEAP datasetso in the future we want to experiment it on more emotionaldatasets to verify the method comprehensively Also we will


Table 3 The mean accuracy of different kinds of methods (Fp1 Fp2 F7 F8 T7 T8 P7 and P8) (standard deviation shown in parenthesesstatistical analysis shown in column 119905-test)

Methods Valence ArousalAccuracy () 119905-test (our method) Accuracy () 119905-test (our method)

Fractal dimension 5308 (1914) 119901 = 0 5961 (2028) 119901 = 00034Sample entropy 5744 (1166) 119901 = 0 6296 (1382) 119901 = 00024DWT + differential entropy (Beta) 6087 (1174) 119901 = 00013 6466 (1159) 119901 = 00048DWT + differential entropy (Gamma) 6736 (661) 119901 = 03185 6855 (928) 119901 = 01189Our method 6910 (695) 119901 = 1 7199 (777) 119901 = 1

utilize more strategies such as feature smoothing and deepnetwork to improve the classification accuracy

5 Conclusion

In this paper an emotion recognitionmethod based on EMDusing three statistics is proposed An extensive analysis hasbeen carried out to investigate the effectiveness of the featuresfor emotion classification The results show that the threefeatures are suitable for emotion recognition Then the effectof each IMF component is inquired The results reveal thatamong the multilevel IMFs the first component IMF1 playsthe most important role in emotion recognition Also theinformative channels based on EMD strategy are investigatedand 8 channels namely FP1 FP2 F7 F8 T7 T8 P7 andP8 are selected for feature extraction Finally the proposedmethod is compared with some classical methods and ourmethod yields the highest accuracy

Conflicts of Interest

The authors declare that there are no conflicts of interestregarding the publication of this paper

Acknowledgments

This work was supported by the grant from the NationalNatural Science Foundation of China (Grant no 61701089)

References

[1] C-H Han J-H Lim J-H Lee K Kim and C-H Im ldquoData-driven user feedback an improved neurofeedback strategyconsidering the interindividual variability of EEG featuresrdquoBioMed Research International vol 2016 Article ID 3939815 7pages 2016

[2] K Takahashi ldquoRemarks on emotion recognition from multi-modal bio-potential signalsrdquo in Proceedings of the 2004 IEEEInternational Conference on Industrial Technology pp 1138ndash1143 Hammamet Tunisia December 2004

[3] O Sourina and Y Liu ldquoA fractal-based algorithm of emotionrecognition from EEG using arousal-valence modelrdquo in Pro-ceedings of the International Conference on Bio-Inspired Systemsand Signal Processing BIOSIGNALS 2011 pp 209ndash214 RomeItaly January 2011

[4] Y Liu and O Sourina ldquoReal-time subject-dependent EEG-based emotion recognition algorithmrdquo Lecture Notes in Com-puter Science (including subseries Lecture Notes in ArtificialIntelligence and Lecture Notes in Bioinformatics) vol 8490 pp199ndash223 2014

[5] X Jie R Cao and L Li ldquoEmotion recognition based on thesample entropy of EEGrdquoBio-MedicalMarerials andEngineeringvol 24 no 1 pp 1185ndash1192 2014

[6] E Kroupi A Yazdani and T Ebrahimi ldquoEEG correlatesof different emotional states elicited during watching musicvideosrdquo in in Procceding of the 2011 Interntionnal Conference onAffective Conputing pp 457ndash466 Memphis TN USA 2011

[7] B Hjorth ldquoEEG analysis based on time domain propertiesrdquoElectroencephalography and Clinical Neurophysiology vol 29no 3 pp 306ndash310 1970

[8] K Ansari-Asl G Chanel and T Pun ldquoA channel selectionmethod for EEG classification in emotion assessment based onsynchronization likelihoodrdquo in Proceedings of the 15th EuropeanSignal Processing Conference EUSIPCO 2007 pp 1241ndash1245Pozna Poland September 2007

[9] R Horlings D Datcu and L JM Rothkrantz ldquoEmotion recog-nition using brain activityrdquo in Proceedings of the InternationalConference on Computer Systems and Technology vol 25 pp 1ndash6 New York NY USA 2008

[10] P C Petrantonakis and L J Hadjileontiadis ldquoEmotion recogni-tion fromEEG using higher order crossingsrdquo IEEE Transactionson Information Technology in Biomedicine vol 14 no 2 pp 186ndash197 2010

[11] X W Wang D Nie and B L Lu ldquoEEG-based emotion recog-nition using frequency domain features and support vectormachinesrdquo in in Procceding of the International Conference onNeural Information Processing pp 734ndash743 Guangzhou China2011

[12] R-NDuan J-Y Zhu andB-L Lu ldquoDifferential entropy featurefor EEG-based emotion classificationrdquo inProceedings of the 20136th International IEEE EMBSConference onNeural EngineeringNER 2013 pp 81ndash84 New Jersey NJ USA November 2013

[13] G Chanel K Ansari-Asl and T Pun ldquoValence-arousal evalua-tion using physiological signals in an emotion recall paradigmrdquoin Proceedings of the 2007 IEEE International Conference onSystems Man and Cybernetics SMC 2007 pp 2662ndash2667Halifax NS Canada October 2007

[14] Y-P Lin C-H Wang T-P Jung et al ldquoEEG-based emotionrecognition in music listeningrdquo IEEE Transactions on Biomedi-cal Engineering vol 57 no 7 pp 1798ndash1806 2010

[15] S K Hadjidimitriou and L J Hadjileontiadis ldquoToward anEEG-based recognition of music liking using time-frequencyanalysisrdquo IEEE Transactions on Biomedical Engineering vol 59no 12 pp 3498ndash3510 2012


[16] S S Uzun S Yildirim and E Yildirim ldquoEmotion primitivesestimation from EEG signals using Hilbert Huang Transformrdquoin Proceedings of the IEEE-EMBS International Conference onBiomedical and Health Informatics pp 224ndash227 Hong KongChina January 2012

[17] M Murugappan M Rizon R Nagarajan and S Yaacob ldquoEEGfeature extraction for classifying emotions using FCM andFKMrdquo in Proceedings of the International Conference on AppliedComputer and Applied Computational Science vol 1 pp 21ndash25Venice Italy 2007

[18] Z Mohammadi J Frounchi and M Amiri ldquoWavelet-basedemotion recognition system using EEG signalrdquoNeural Comput-ing and Applications pp 1ndash6 2016

[19] M Murugappan ldquoHuman emotion classification using wavelettransform and KNNrdquo in Proceedings of the 2011 InternationalConference on Pattern Analysis and Intelligent Robotics (ICPAIRrsquo11) vol 1 pp 148ndash153 Putrajaya Malaysia June 2011

[20] S Koelstra C Muhl M Soleymani et al ldquoDEAP a databasefor emotion analysis using physiological signalsrdquo IEEE Trans-actions on Affective Computing vol 3 no 1 pp 18ndash31 2012

[21] I Wichakam and P Vateekul ldquoAn evaluation of feature extrac-tion in EEG-based emotion prediction with support vectormachinesrdquo in Proceedings of the 2014 11th International JointConference on Computer Science and Software Engineering(JCSSE rsquo14) pp 106ndash110 Chon Buri Thailand May 2014

[22] B Reuderink C Muhl and M Poel ldquoValence arousal anddominance in the EEG during game playrdquo International Journalof Autonomous and Adaptive Communications Systems vol 6no 1 pp 45ndash62 2013

[23] L Brown B Grundlehner and J Penders ldquoTowards wirelessemotional valence detection from EEGrdquo in Proceedings of the33rd Annual International Conference of the IEEE Engineering inMedicine and Biology Society EMBS 2011 pp 2188ndash2191 BostonMA USA September 2011

[24] V Rozgic S N Vitaladevuni and R Prasad ldquoRobust EEGemotion classification using segment level decision fusionrdquo inProceedings of the 2013 38th IEEE International Conference onAcoustics Speech and Signal Processing ICASSP 2013 pp 1286ndash1290 Vancouver BC Canada May 2013

[25] R Gupta K U R Laghari and T H Falk ldquoRelevance vectorclassifier decision fusion and EEG graph-theoretic features forautomatic affective state characterizationrdquoNeurocomputing vol174 pp 875ndash884 2016

[26] R Jenke A Peer andM Buss ldquoFeature extraction and selectionfor emotion recognition from EEGrdquo IEEE Transactions onAffective Computing vol 5 no 3 pp 327ndash339 2014

[27] W-L Zheng and B-L Lu ldquoInvestigating critical frequencybands and channels for eeg-based emotion recognition withdeep neural networksrdquo IEEE Transactions on AutonomousMental Development vol 7 no 3 pp 162ndash175 2015

[28] Y Yang Q M J Wu W L Zheng and B L Lu ldquoEEG-basedemotion recognition using hierarchical network with subnet-work nodesrdquo IEEE Transactions on Cognitive DevelopmentalSystems vol PP no 99 p 1 2017

[29] N E Huang Z Shen S R Long et al ldquoThe empirical modedecomposition and the hilbert spectrum for nonlinear and non-stationary time series analysisrdquo in Proceedings of the ProceedingsMathematical Physical and Engineering Sciences vol 454 pp903ndash995 1998

[30] S M S Alam and M I H Bhuiyan ldquoDetection of seizure andepilepsy using higher order statistics in the EMDdomainrdquo IEEE

Journal of Biomedical and Health Informatics vol 17 no 2 pp312ndash318 2013

[31] R B Pachori and V Bajaj ldquoAnalysis of normal and epilepticseizure EEG signals using empirical mode decompositionrdquoComputer Methods and Programs in Biomedicine vol 104 no3 pp 373ndash381 2011

[32] S LiW Zhou Q Yuan S Geng andD Cai ldquoFeature extractionand recognition of ictal EEG using EMD and SVMrdquo Computersin Biology and Medicine vol 43 no 7 pp 807ndash816 2013

[33] A Mert and A Akan ldquoEmotion recognition from EEG signalsby using multivariate empirical mode decompositionrdquo PatternAnalysis and Applications pp 1ndash9 2016

[34] K Takahashi ldquoRemarks on emotion recognition from multi-modal bio-potential signalsrdquo in Proceedings of the 2004 IEEEInternational Conference on Industrial Technology 2004 (IEEEICIT rsquo04) pp 1138ndash1143 Hammemet Tunsia

[35] G Chanel J Kronegg D Grandjean and T Pun ldquoEmo-tion assessment arousal evaluation using eegs and peripheralphysiological signalsrdquo in in procceding of the InternationalWorkshop on Multimedia Content Representation Classificationand Security pp 530ndash537 Istanbul Turkey 2006

[36] CCChang andC J Lin ldquoLIBSVMA library for support vectormachinesrdquo Acm Transactions on Intelligent Systems Technologyvol 2 no 3 p 27 2007

Submit your manuscripts athttpswwwhindawicom

Neurology Research International

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Alzheimerrsquos DiseaseHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

ScientificaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


BioMed Research International


Research and TreatmentSchizophrenia

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


Neural Plasticity


Parkinsonrsquos Disease


Research and TreatmentAutism

Sleep DisordersHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Neuroscience Journal

Epilepsy Research and TreatmentHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Psychiatry Journal


Computational and Mathematical Methods in Medicine

Depression Research and TreatmentHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Brain ScienceInternational Journal of

StrokeResearch and TreatmentHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Neurodegenerative Diseases


Journal of

Cardiovascular Psychiatry and NeurologyHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


EEG signals

IMF1

Feature extraction

SVMclassier

RecognitionresultsEMD

IMF2

IMF3

IMF4

IMF5

(5 s)

Figure 1 Block diagram of the proposed method

performance of different features mentioned above and got aguiding rule for feature extraction and selection [26]

Some other strategies such as utilizing deep networkto improve the classification performance have also beenresearched Zheng and Lu used deep neural network toinvestigate critical frequency bands and channels for emotionrecognition [27] Yang et al used hierarchical network withsubnetwork nodes for emotion recognition [28]

EMD is proposed by Huang et al in 1998 [29] UnlikeDWT which needs to predetermine transform base functionand decomposition level EMD can decompose signals intoIMF automatically These IMFs represent different frequencycomponents of original signals with band-limited charac-teristic By applying Hilbert transform to IMF we can getinstantaneous phase information of IMF So EMD is suitablefor analysis of nonlinear and nonstationary sequence such asneural signals

EMD is a good choice for EEG signals and we utilize itfor emotion recognition from EEG data Which feature iseffective for emotion recognition in EMD domain WhichIMF component is best for classification Is the performancebased on EMD strategy better compared to time domainmethod and time-frequency method or not All these havenot been researched yet and we investigate them in ourresearch

EMD has been widely used for seizure prediction anddetection but for emotion recognition based on EMD thereis not so much research Higher order statistics of IMFs [30]geometrical properties of the decomposed IMF in complexplane [31] and the variation and fluctuation of IMF [32]are used as features for seizure prediction and detectionFor emotion recognition Mert and Akan extracted entropypower power spectral density correlation and asymmetry ofIMF as features and then utilized independent componentanalysis (ICA) to reduce dimension of the feature set [33]The classification accuracy is computed with all the subjectsmixed together

In this paper we present an emotion recognition methodbased on EMD We utilize the first difference of IMF timeseries the first difference of the IMFrsquos phase and the nor-malized energy of IMF as features The motivation of usingthese three features is that they depict the characteristicsof IMF in time frequency and energy domain providingmultidimensional information The first difference of timeseries depicts the intensity of signal change in time domain

The first difference of phase measures the change intensity inphase and normalized energy describes the weight of currentoscillation componentThe three features constitute a featurevector which is fed into SVM classifier for emotional statedetection

The proposed method is studied on a publicly availableemotional database DEAP [20] The effectiveness of thethree features is investigated IMF reduction and channelreduction for feature extraction are both discussed whichaim at improving the classification accuracy with less compu-tation complexity The performance is compared with someother techniques including fractal dimension (FD) sampleentropy differential entropy and time-frequency analysisDWT

2 Method

To realize emotional state recognition the EEG signals aredecomposed into IMFs by EMD Three features of IMFs thefluctuation of the phase the fluctuation of the time series andthe normalized energy are formed as a feature vector whichis fed into SVM for classification The whole process of thealgorithm is shown in Figure 1

21 Data and Materials DEAP is a publicly available datasetfor emotion analysis which recorded EEG and peripheralphysiological signals of 32 participants as they watched 40music videos All the music video clips last for 1 minuterepresenting different emotion visual stimuli with gradefrom 1 to 9 Among the 40 music videos 20 are highvalence visual stimuli and 20 are low valence visual stimuliThe situation is exactly the same for arousal dimensionAfter watching the music video participants performed aself-assessment of their levels on arousal valence likingdominance and familiarity with ratings from 1 to 9 EEGwas recorded with 32 electrodes placing according to theinternational 10-20 system Each electrode recorded 63 s EEGsignal with 3 s baseline signal before the trial

In this paper we used the preprocessed EEG data forstudy with sample rate 128Hz and band range 4ndash45Hz EOGartefacts were removed as method in [20] The data was seg-mented into 60-second trials and a 3-second pretrial baselineremoved The binary classifications of valence and arousaldimension are considered We utilized the participantsrsquo self-assessment as label If the participantrsquos rating was lt5 the label








119898(119905) = (119890min (119905) + 119890max (119905))2 (1)


ℎnew (119905) = ℎold (119905) minus 119898 (119905) (2)


imf1 = ℎnew (119905) (3)





119909 (119905) =119871

sum119899=1

imf119899 + 119903 (5)




119863119905 =1119873 minus 1

119873minus1

sum119899=1

|imf (119899 + 1) minus imf (119899)| (6)


119911 (119899) = 119909 (119899) + 119895119910 (119899) (7)


119911 (119899) = 119860 (119899) 119890119895120593(119899) (8)



119863119901 =1119873 minus 1

119873minus1

sum119899=1

1003816100381610038161003816120593 (119899 + 1) minus 120593 (119899)1003816100381610038161003816 (9)



119864norm =sum119873119899=1 imf2 (119899)sum119873119899=1 1199042 (119899)

(10)



0

minus1

0

1EE

G

minus05

0

05

IMF1

minus05

0

05

IMF2

minus05

0

05

IMF3

minus05

0

05

IMF4

minus05

05

IMF5

200 300 400 500 600 700 800 900 1000100Points

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100














Valence Arousal0

10

20

30

40

50

60

70

80

Accu

racy

()

Dt

Dp

EＨＩＬＧ






PredictLabel








Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus004

minus003

minus002

minus001

0

001

002

003

004Dt

(a)

Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus003

minus002

minus001

0

001

002

003Dp

(b)

Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus0015

minus001

minus0005

0

0005

001

0015EＨＩＬＧ

(c)







Valence Arousal



0

10

20

30

40

50

60

70

80

Accu

racy

()








4 Discussion











5 Conclusion




Acknowledgments


References



















































Neural Plasticity










Psychiatry Journal









Journal of









119898(119905) = (119890min (119905) + 119890max (119905))2 (1)


ℎnew (119905) = ℎold (119905) minus 119898 (119905) (2)


imf1 = ℎnew (119905) (3)





119909 (119905) =119871

sum119899=1

imf119899 + 119903 (5)




119863119905 =1119873 minus 1

119873minus1

sum119899=1

|imf (119899 + 1) minus imf (119899)| (6)


119911 (119899) = 119909 (119899) + 119895119910 (119899) (7)


119911 (119899) = 119860 (119899) 119890119895120593(119899) (8)



119863119901 =1119873 minus 1

119873minus1

sum119899=1

1003816100381610038161003816120593 (119899 + 1) minus 120593 (119899)1003816100381610038161003816 (9)



119864norm =sum119873119899=1 imf2 (119899)sum119873119899=1 1199042 (119899)

(10)



0

minus1

0

1EE

G

minus05

0

05

IMF1

minus05

0

05

IMF2

minus05

0

05

IMF3

minus05

0

05

IMF4

minus05

05

IMF5

200 300 400 500 600 700 800 900 1000100Points

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100














Valence Arousal0

10

20

30

40

50

60

70

80

Accu

racy

()

Dt

Dp

EＨＩＬＧ






PredictLabel








Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus004

minus003

minus002

minus001

0

001

002

003

004Dt

(a)

Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus003

minus002

minus001

0

001

002

003Dp

(b)

Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus0015

minus001

minus0005

0

0005

001

0015EＨＩＬＧ

(c)







Valence Arousal



0

10

20

30

40

50

60

70

80

Accu

racy

()








4 Discussion











5 Conclusion




Acknowledgments


References



















































Neural Plasticity










Psychiatry Journal









Journal of



0

minus1

0

1EE

G

minus05

0

05

IMF1

minus05

0

05

IMF2

minus05

0

05

IMF3

minus05

0

05

IMF4

minus05

05

IMF5

200 300 400 500 600 700 800 900 1000100Points

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100

200 300 400 500 600 700 800 900 1000100














Valence Arousal0

10

20

30

40

50

60

70

80

Accu

racy

()

Dt

Dp

EＨＩＬＧ






PredictLabel








Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus004

minus003

minus002

minus001

0

001

002

003

004Dt

(a)

Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus003

minus002

minus001

0

001

002

003Dp

(b)

Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus0015

minus001

minus0005

0

0005

001

0015EＨＩＬＧ

(c)







Valence Arousal



0

10

20

30

40

50

60

70

80

Accu

racy

()








4 Discussion











5 Conclusion




Acknowledgments


References



















































Neural Plasticity










Psychiatry Journal









Journal of






Valence Arousal0

10

20

30

40

50

60

70

80

Accu

racy

()

Dt

Dp

EＨＩＬＧ






PredictLabel








Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus004

minus003

minus002

minus001

0

001

002

003

004Dt

(a)

Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus003

minus002

minus001

0

001

002

003Dp

(b)

Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus0015

minus001

minus0005

0

0005

001

0015EＨＩＬＧ

(c)







Valence Arousal



0

10

20

30

40

50

60

70

80

Accu

racy

()








4 Discussion











5 Conclusion




Acknowledgments


References



















































Neural Plasticity










Psychiatry Journal









Journal of



Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus004

minus003

minus002

minus001

0

001

002

003

004Dt

(a)

Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus003

minus002

minus001

0

001

002

003Dp

(b)

Fc5 Fc1 Fc2 Fc6

C3 Cz C4

Cp5 Cp1 Cp2 Cp6

Fp1 Fp2 Af3 Af4

F7 F3 Fz F4 F8

T7 T8

P7 P3 Pz


minus0015

minus001

minus0005

0

0005

001

0015EＨＩＬＧ

(c)







Valence Arousal



0

10

20

30

40

50

60

70

80

Accu

racy

()








4 Discussion











5 Conclusion




Acknowledgments


References



















































Neural Plasticity










Psychiatry Journal









Journal of



Valence Arousal



0

10

20

30

40

50

60

70

80

Accu

racy

()








4 Discussion











5 Conclusion




Acknowledgments


References



















































Neural Plasticity










Psychiatry Journal









Journal of







5 Conclusion




Acknowledgments


References



















































Neural Plasticity










Psychiatry Journal









Journal of





































Neural Plasticity










Psychiatry Journal









Journal of














Neural Plasticity










Psychiatry Journal









Journal of


Documents

Emotion Recognition from EEG Signals Using