EEG Channel Correlation Based Model for Emotion

Recognition

Author
Islam, Md Rabiul, Islam, Md Milon, Rahman, Md Mustafizur, Mondal, Chayan, Singha, Suvojit
Kumar, Ahmad, Mohiuddin, Awal, Md Abdul, Islam, Md Saiful, Moni, Mohammad Ali

Published
2021

Journal Title
Computers in Biology and Medicine

Version
Accepted Manuscript (AM)

DOI
https://doi.org/10.1016/j.compbiomed.2021.104757

Copyright Statement
© 2021 Elsevier. Licensed under the Creative Commons Attribution-NonCommercial-
NoDerivatives 4.0 International Licence (http://creativecommons.org/licenses/by-nc-nd/4.0/)
which permits unrestricted, non-commercial use, distribution and reproduction in any medium,
providing that the work is properly cited.

Downloaded from
http://hdl.handle.net/10072/406838

Griffith Research Online

https://research-repository.griffith.edu.au
EEG Channel Correlation Based Model for Emotion
Recognition

Md. Rabiul Islam, Md. Milon Islam, Md. Mustafizur Rahman, Chayan Mondal, Suvojit Kumar Singha,
Mohiuddin Ahmad, Md. Abdul Awal, Md. Saiful Islam and Mohammad Ali Moni*

Md. Rabiul Islam Md. Milon Islam

Electrical and Electronic Engineering Computer Science and Engineering
Bangladesh Army University of Engineering & Technology Khulna University of Engineering & Technology
Natore-6431, Bangladesh Khulna-9203, Bangladesh
rabiulnewemail@gmail.com milonislam@cse.kuet.ac.bd

Md. Mustafizur Rahman Chayan Mondal

Electrical and Electronic Engineering Electrical and Electronic Engineering
Jashore University of Science and Technology Khulna University of Engineering & Technology
Jashore-7408, Bangladesh Khulna-9203, Bangladesh
mustafizur.170710@gmail.com chayan.eee.92@gmail.com

Suvojit Kumar Singha Mohiuddin Ahmad

Electrical and Electronic Engineering Electrical and Electronic Engineering
Khulna University of Engineering & Technology Khulna University of Engineering & Technology
Khulna-9203, Bangladesh Khulna-9203, Bangladesh
singha10.suvojit@gmail.com ahmad@eee.kuet.ac.bd

Md. Abdul Awal Md. Saiful Islam

Electronics and Communication Engineering School of Information and Communication
Khulna University Technology, Griffith University,
Khulna-9208, Bangladesh Gold Coast, Australia
m.awal@ece.ku.ac.bd saiful.islam@griffith.edu.au

Mohammad Ali Moni

School of Health and Rehabilitation Sciences
The University of Queensland
St Lucia, QLD 4072, Australia
m.moni@uq.edu.au

*Corresponding author:
Mohammad Ali Moni (Email: m.moni@uq.edu.au).

1
Abstract: Emotion recognition using Artificial Intelligence (AI) is a fundamental prerequisite to improve Human-
Computer Interaction (HCI). Recognizing emotion from Electroencephalogram (EEG) has been globally accepted in
many applications such as intelligent thinking, decision-making, social communication, feeling detection, affective
computing, etc. Nevertheless, due to having too low amplitude variation related to time on EEG signal, the proper
recognition of emotion from this signal has become too challenging. Usually, considerable effort is required to
identify the proper feature or feature set for an effective feature-based emotion recognition system. To extenuate the
manual human effort of feature extraction, we proposed a deep machine-learning-based model with Convolutional
Neural Network (CNN). At first, the one-dimensional EEG data were converted to Pearson's Correlation Coefficient
(PCC) featured images of channel correlation of EEG sub-bands. Then the images were fed into the CNN model to
recognize emotion. Two protocols were conducted, namely, protocol-1 to identify two levels and protocol-2 to
recognize three levels of valence and arousal that demonstrate emotion. We investigated that only the upper
triangular portion of the PCC featured images reduced the computational complexity and size of memory without
hampering the model accuracy. The maximum accuracy of 78.22% on valence and 74.92% on arousal were obtained
using the internationally authorized DEAP dataset.
Keywords: Emotion, Convolutional Neural Network, Feature Extraction, EEG, Pearson's Correlation Coefficient,
Complexity.

I. Introduction
Emotion is a feeling that implies how we act for a particular instance. Now in the era of technological development, the
research of emotion recognition from EEG has become very popular due to its non-invasive feature. The remarkable
development has occurred on emotion-based non-invasive Brain-Computer Interface (BCI) and signal processing
technology. Therefore, the analysis of EEG-based emotion recognition has become a very attractive issue in affective
computing. Various information and important data can easily be extracted from the human brain and analyzed for
knowing the inner truth of humans.
Nowadays machine especially robot has been utilized in many industries, hospitals and even in household applications.
People are setting higher expectations for robots as they become more prevalent in many parts of daily life. For better
human-machine interaction, it is hoped the super ability of decision making, self-thinking, emotion sensing. The
assurance of emotion recognition is an inevitable need to make a robot more practical for real-life applications. The
patient's affective information, including emotional state, is considered a key factor of his/her mental and physical
status. A patient's emotional state has a significant effect on the treatment management process [1]. Therefore clinics,
hospitals, and other healthcare providers should keep a proper system for emotional state assessment to strengthen
healthcare. Since human behavior is mainly dependent on emotion and psychologist wants to analyze the condition of
the human mind; recognition of emotion has also become a hot topic in psychology. Many psychological experiments
demonstrated the relation between emotions with EEG signals [2], [3]. Even emotion is now treated as a governor of
music and video recommendation. For instance, in a sad mood, it recommends a funny event for making a person
happy. In gamming, the level may become harder in positive emotional states. In court, during the criminal's statement,
emotion can also be recorded to judge his authenticity and reliability. Accordingly, emotion recognition has become an
indispensable part of our daily life. Consequently, computer scientists, AI specialists, physiologists, and biologists are
continuously generating enormous studies using computer technology to understand human emotions [4].
The research on emotion recognition demands the emotionally aroused signal of the human being. Humans remain
emotionally excited in their real-life activities. However, it is too tough to accumulate the aroused data on that situation.
For research, emotion-relevant EEG data is usually collected by listening music, watching videos, and playing
emotion-related games. The audiovisual stimuli create the proper excitement on human brain signals among the
different types of stimuli as audiovisual stimuli affect the human mind by both audio and video context [5]. To
recognize human emotion, many scholars used various types of raw signals [6]. Many used EEG signals [2], [7]–[9]
and facial expressions [10], [11]; and few used gesture, speech signals [12], autonomous nervous signals [13]. Subjects
need to express emotion explicitly when facial expressions and speech signals are being used. Moreover, from the
facial expression, only these types of emotion can be classified whose effect will change the human face structure.

2
However, subjects who feel happy internally but do not express it by their face cannot be classified. So facial
expression is not considered as a good emotion recognizer raw signal. For speech signals, the subject's emotion can be
classified with the help of the voice's intensity. For that reason, this type of emotion recognition method is not
applicable to autistic persons. Moreover, the people who are not able to speak remain out of this type of emotion
recognition method. Some researchers tried to classify emotion from gesture [14] and body movement [15]. People
with physically handicapped are not able to present body language and gestures to express emotion. Furthermore, the
autonomous nervous system is not too applicable because of its complex acquisition technique. Interestingly, subjects
cannot control the spontaneously generated EEG signal. When we extract emotion from EEG, then the emotion of the
people who cannot speak or unable to express themselves by gesture and posture can also be recognized. Besides, the
EEG signal acquisition approach is not as tricky as multimodal signal acquisition. Moreover, the low cost, wireless
flexible and portable acquisition medium makes it more popular. Consequently, the best and peerless signal for
emotion recognition is EEG for any kind of person at any time in any place.

Emotional
Video

Stimuli EEG CNN Emotion

Figure 1: Emotion recognition from EEG signal using CNN.

Different researchers have applied multiple methods and approaches in recognizing and classifying emotion [16]–
[18]. Recently, the wavelet transform has become a prevalent analysis method due to its good performance both in the
time and frequency domain [8]. EEG signal is a non-linear, non-stationary, and temporal asymmetry type signal on the
microvolt range. The computation and analysis of this type of signal are too challenging. Some researches on emotion
recognition have been performed by selecting proper features and shallow machine learning algorithms like Support
Vector Machine (SVM) [19]–[23], k Nearest Neighbor (kNN) [19], Decision Tree (DT), or Multi-Layer Perceptron
(MLP) etc. Soleymani et al. developed an emotion recognition method based on user-independent analysis. They
achieved a good result in terms of accuracy, like 68.5% for valence and 76.4% for arousal upon the classification of 3
classes [24]. Zheng et al. investigated stable EEG patterns using machine learning and systematically evaluated the
performance of different feature extraction methods, selection, and smoothing [25]. Atkinson and Campos suggested a
valence arousal-based method combined with mutual information-based feature selection methods and kernel
classifiers [26]. They proposed an EEG-based Brain-Computer Interface to explore a set of emotional types and
incorporate additional features relevant for signal preprocessing and classification. In the conventional feature
extraction methods, several features or feature set are being considered to train the recognition system. In contrast, deep
learning eliminates the difficulties of selecting the useful and significant features. It allows the machine to learn the
feature from the input data set automatically and transfer its learning to the classifier.
Längkvist et al. proposed a Deep Belief Network (DBN) architecture to reduce the complexity and necessity of
multimodal sleep data [27]. They also recovered the fact of time consumption in the classification stage. Li et al. used
differential entropy feature with a novel DBN and achieved 11.5% and 24.4% improvements on the task of affective
state recognition [28]. Martinez et al. classified four different emotional states using CNN, considering the skin
conductance and blood volume pulse signal [29]. The researcher Wen et al. [30] proposed a method using channel
correlation, but they did not use the emotion-relevant sub-band data. Another researcher used deep and convolutional
neural networks on mean, median, standard deviation, and many other time-domain features [31], but time-domain
features are not too helpful. The author of [32] calculates the PCC of multichannel EEG data considering four sub-
bands.
This paper proposed an emotion recognition method of lower complexity from audio-visual stimuli-based EEG
signals using CNN. The architectural flow diagram is shown in the following Fig. 1. Emotion is highly related to beta
and gamma sub-bands, moderately related to alpha sub-band, and very poorly connected to theta sub-band. Therefore,

3
we have considered only alpha, beta gamma sub-bands despite considering four sub-bands. That makes our information
more significant and reduces computational complexity. Again, we proved that only the upper triangular portion of the
PCC featured images was substantial to recognize emotion. Our work used a deep architecture-based Convolution
Neural Network as it could automatically extract internal features. Thus, the main contributions of the proposed
research are as follows:
1) The development of an improved emotion recognition method with lower computational complexity, lower
memory requirement, lower time consumption for processing; without hampering the performance.
2) Developing a novel procedure to construct a matrix of significant two-dimensional data based on channel
correlation from the one-dimensional EEG signals is proposed.
3) The development of a CNN-based modified model that can recognize emotion using a single feature
Pearson's Correlation Coefficient of EEG sub-bands.
This work is the updated version of our previous work [33]. Here, the computation complexity is significantly
reduced by reducing input images' size considering only the upper triangular matrix of PCC featured images.
Moreover, the convolution neural network model is modified according to the size of our newly generated PCC
featured images that applied as input.
The rest of the paper is organized as follows. Section II demonstrates the necessary dataset collection and
description with data preprocessing including some preprocessing techniques. In addition, the development of
network architecture is also described in this section. The experimental results for two different protocols with the
DEAP dataset are illustrated in section III. The discussions along with the comparative study with the current works

EEG data Decomposition

δ θ α β γ
Preprocessed EEG data
video×channel×data Not Used
40×40×8064

Segmentation
Reshaping

40 to 32 Channels PCC Feature

63s to 60s data

video×channel×data
40×32×7860
Upper
Triangular
Matrix
23×23
Interchanging the
Channel Correlation Matrix Featured Image
dimensions

channel×data×video
32×7860×40 Classification
Classified Emotion CNN

Figure 2: The successive process of our proposed method includes reshaping, decomposition, segmentation, PCC featured image extraction, and
classification to recognize emotion from raw EEG data.

4
 are depicted in section IV. Finally, the conclusion with some significant future works is discussed in section V.
II. Materials and Methods
The proposed system takes EEG data as input and generates classified emotion as output. The system diagram of our
experiment is illustrated in Fig. 2. Later the individual portion of the system is described separately. Firstly, the raw
EEG data were preprocessed. Secondly, just one feature named Pearson's Correlation Coefficient between every
possible combination of two channels among 32 channels was determined. Thus, we formulated the PCC featured
images of channel correlation. Afterward, the PCC featured images containing two-dimensional channel correlation
data were taken as input into our designed CNN classification algorithm to recognize emotion.

A. Dataset
We have considered the internationally accredited 'DEAP' dataset [34] in our study to perform the experiment and
measured our model's performance. There remain sixteen different emotional class-based labeled EEG data of 32
participants. 'DEAP' stands for Dataset for Emotion Analysis using Physiological signals. It is a commonly useable
open dataset for emotion analysis that contains EEG, physiological and video signals.
Table I
Information about 'DEAP' Dataset
Attributes Details Information
EEG acquisition system Biosemi ActiveTwo
Channels of recorded signals 32 EEG channel (512 Hz)
12 Peripherals
1 Status channel
3 Unused channels
Total of 48 channels
EEG electrodes AF3, AF4, C3, C4, Cz, CP1, CP2, CP5, CP6, F3, F4, F7, F8, Fz,
FC1, FC2, FC5, FC6, Fp1, Fp2, O1, O2, Oz, P3, P4, P7, P8 Pz,
PO3, PO4, T7 and T8
Rated parameter Valence, arousal, dominance, liking, and familiarity
Range of rating 1-9, except familiarity (1-5)
Available data format video×channel×data = 40×40×8760

Researchers used the dataset to evaluate the performance of their own designed system of emotion recognition. The 16
male and 16 female participants (age ranges from 19 to 37 years) were watched forty different emotional music videos
and the corresponding physiological recordings and participant ratings were stored. Each video had a duration of 1 min
long. They recorded EEG signals using the International 10/20 electrode placement system. The relevant information
about the 'DEAP' dataset is shown in Table I.
As the EEG signal is very low amplitude signals, the extraction of information from this type of signal is too
complicated. Moreover, an Electrooculogram (EOG) and another noise signal may hamper the original signal's
effectiveness if it is not preprocessed. Firstly, the raw data were downsampled to 128 Hz from 512 Hz to reduce the
volume. Later, EOG artifacts were removed. Finally, it was passed through a bandpass filter of frequency 4~45 Hz. We
collected the preprocessed data from the official website of 'DEAP'.

B. Reshaping and Permutation

In the emotion recognition task, the downloaded data shape format was like that video×channel×data = 40×40×8064. It
illustrated 40 different videos, 40 channels, and 8064 data points (63s data, sampling frequency 128Hz, data
points=128×63). Among them, only the first 32 electrodes (channels) were used for the EEG signals recording.
Therefore, we extracted these 32 channel data only. Also, the last 60s data were original. We reshaped the data as
video×channel×data = 40×32×7680. Afterward, the reshaped data were permuted to arrange the data as following
format channel×data×video = 32×7680×40. The data structure before and after reshaping and permutation has
demonstrated in Fig. 3.

5
 Figure 3: Dimension 3D of data before and after reshaping and permutation.

C. Decomposition
The activity of human work affects fully on brain waves. As a result, the mental states and conditions can be extracted
from the EEG signals. Generally human EEG is a composite type signal that consists of five different types of brain
waves of different frequency namely delta (1 Hz < f < 4 Hz), theta (4 Hz < f < 8 Hz), alpha (8 Hz < f < 12 Hz), beta
(12 Hz < f < 30 Hz) and gamma (30 Hz < f < 60 Hz). The different sub-bands are incorporated with different mental
states and activities. For an instance, the delta sub-band is related to relaxing or calm activities like deep sleep and
unconsciousness. The theta and alpha sub-bands are related to low-level excitement like drowsiness, imagination,
closing the eye etc. The medium-level excitements including thinking, anxiety and stressed are incorporated with the
beta sub-band signal. Lastly, the gamma sub-band signals affect hyperactivity, such as alertness, agitation, object
matching and sensory processing. In this part of our work, the EEG signals were decomposed into the five sub-bands
using Discrete Wavelet Transform. Later, as delta and theta sub-band associated with relaxing to low-level brain
activity, we ignored it for emotion recognition. The original EEG signal and our decomposed alpha, beta, and gamma
sub-band signals of channel Fp1, of the first participant (p1) for the first video (v1), are shown in Fig. 4.
Main EEG signal
50
0
-50
0 20 40 60
Amplitude (µV)

Alpha Sub-band
50
0
-50
0 20 40 60

Beta Sub-band
50
0
-50
0 20 40 60

Gamma Sub-band
50
0
-50
0 20 40 60

Time (s)
Figure 4: Decomposition of main preprocessed EEG signal (of p1, v1, Fp1) into alpha, beta, and gamma sub-band signals.

D. Segmentation
As CNN requires many training data for an exclusive performance, the EEG data have to be segmented. One EEG
signal's length is the 60s, which means the number of total data points is 7680. Before calculating the PCC of different
channels, we need to segment the data. Here the EEG signals were segmented into 20 parts in which every segment

6
contains 384 data points (data of 3s). The segmented part of a single channel EEG is shown in Fig. 5. Accordingly, for
a single participant, the total segmented partition number will be 20 times 32 times 40, i.e., 25,600.

1 2 3 ……………….……….……………………….………………………………. 20
50
0
-50
0 10 20 30 40 50 60

Figure 5: Every decomposed signal of the 60s is segmented into 20 portions.

E. PCC Featured Image Formation

To recognize emotion from EEG signal, many researchers used various types of the method such as Short Time Fourier
transform (STFT), Common Spatial Pattern (CSP), Discrete Wavelet Transform, Statistical feature, Higuchi fractal
dimension (HFD), Magnitude Squared Coherence Estimate (MSCE), Power Spectral Density (PSD), Fast Fourier
Transform (FFT), Higher-Order component (HOC), Differential Entropy (DE), Differential Asymmetry (DASM),
Rational Asymmetry (RASM), Affective Signal Processing (ASP), etc. We used Pearson's Correlation Coefficient to
generate the PCC-featured effective images for emotion recognition.
Pearson's correlation coefficient is the scale of indication that represents the statistical relationship between two
continuous variables. It is based on covariance; it is the best method of measuring the relationship between two
variables. The value of Pearson's Correlation Coefficient ranges from +1 to -1, where +1 indicates a perfect positive
correlation and -1 is the opposite of this. The calculation of PCC for two series of the dataset a and b is

cov( a, b)
ab  (1)
 a b
where a = (a1, a2, a3 …. an) and b = (b1, b2, b3, …. bn). Here cov(a,b) indicates the covariance between a and b, and σa,
σb implies the standard deviation of the dataset a and b, respectively. In our work, we calculated the PCC directly by
using a user-defined function in MATLAB by (2).

ni 1 (aibi )  i 1 ai i 1 bi
n n n

ab  (2)
ni 1 ai 
n 2
 a 
n
i 1 i
2
ni 1 bi 
n 2
 b 
n
i 1 i
2

F. Convolutional Neural Network Model

Convolutional Neural Network is a branch of Deep Neural Networks (DNN), which has proven outstanding image
classification and computer vision performance. The main structure of CNN is very similar to the connectivity pattern
of the neuron of the human brain. One of the main advantages of CNN is that it requires a little bit of preprocessing or
sometimes needs no preprocessing like other traditional classification algorithms. It can automatically learn many
features from the training dataset and use them to predict the test data. The first operation of a CNN is making
convolution with the images and a filter. In the neural network discussion, the filter and kernel are the same terms. The
convolutional layer is the primary layer of any CNN network. This layer performs matrix multiplication between a
particular portion of input images and a specific shape of the kernel.
The kernel moves through an image and makes a new convolved image. Our proposed CNN architecture from the
PCC-based images to the classified emotion is shown in Fig. 6. In the architecture, there remain three sets of

7
Figure 6: The proposed CNN model for emotion recognition consists of three sets of convolution, Rectified Linear Unit (ReLU), and pooling
layers in addition to one flatten and one dense layer. The size of images, number of filters, padding, and stride, etc., on each level, are shown.

Convolution, ReLU, and Pooling layers. Afterward, we used a flatten layer. Lastly, the dense layer was connected.
Here we used a dropout of 0.25 for reducing the network complexity.
After a single convolution layer, the dimension of the output images will be nH×nW×nC, considering the size of input
images, filters, weight, and bias are as similar as expressed in (3), (4), (5), and (6), respectively.

size _ of _ input  n[Hl 1]  nW

[ l 1]
 nC[l 1] (3)

size _ of _ filter  f H[l ]  fW[l ]  nC[l 1] (4)

 
size _ of _ weight  f H[l ]  fW[l ]  nC[l 1]  nC[l ] (5)

size _ of _ bias  111 nC[l ] (6)

The output layer at level l can be calculated as follows.

n[Hl 1]  2 p [l ]  f [l ]
nH  1 (7)
S [l ]

size _ of _ output  n[Hl ]  nW

[l ]
 nC[l ] (8)

The calculation of nw is precisely similar to nH like (7). Here, p[l] and S[l] represent the padding and stride values,
respectively at a level l. The filters of size 3×3 were used in the 1st, 2nd, and 3rd convolution layers. The number of
filters is shown in Fig. 6. Since we did not want to lose the information from the pixels of a corner of the input PCC-
based image, we used the same convolution (padding layer, p=1) for the 1st and 3rd convolution layers. A valid

8
convolution (i.e. no padding, p=0) method was used for the 2nd convolution layer to reduce computational complexity.
The number of padding (p) and stride (s) in every layer is indicated clearly in Fig. 6.
In the classification task for normal images, the most popular type of pooling is max pooling. Since the PCC-
featured images are texture-type images, we used average pooling despite using max pooling. It may be noted that a
texture-type image does not contain any particular pattern like an ordinary or natural image. The 2×2 sized filters with
stride 2 were used in every pooling layer. The major hyper-parameters and their corresponding information, such as the
value or the type of them of our trained CNN model, are shown in Table II.
Table II
The Value or Types of the Major Hyper-parameters of Our Trained CNN Model
Model Hyper-parameter Value or Type
Learning rate 0.001
Momentum 0.80
Number of epochs 50
Batch size 512
Dropout 0.25
Pooling method Average pooling
Activation function ReLU and sigmoid (for protocol 1)
ReLU and softmax (for protocol 2)
Optimization algorithm Adam

G. Activation Functions
In our model, the ReLU activation function is used that best fits CNN architecture and faster. The ReLU has a positive
feature compared to 'sigmoid' and 'tanh' that is never saturated with a significant value of x. Besides, it is more reliable
and accelerates the convergence by six times. The mathematical expression of ReLU is given in (9).
f ( x)  max( 0, x) (9)

As our input data consists of values from -1 to +1, normalization is not necessary here. We used l2 type kernel
regularizer to remove the overfitting in dense layers where the regularization parameter is 0.001.
In protocol 1, we made a binary classification. For doing this, we used the 'sigmoid' activation function in the last
layer. The 'sigmoid' is a smooth non-linear activation function that returns the probabilities of a class, and since the
probability ranges from 0 to 1, its range is also the same. The equation of the 'sigmoid' activation function is given in
(10). Protocol 2 is not a binary classification task because we classified valence and arousal into three classes.
Therefore, 'sigmoid' is not suitable for this purpose. For multiclass classification 'softmax' activation function is perfect.
The 'softmax' activation function returns the probability of every class, and lastly, it targets the class that belongs to the
most considerable probability. The 'softmax' function is defined in (11).
1
 ( z)  (10)
1  e z
1
h ( x)  (11)
1  exp(  T x)
Here, hθ is the scalar output of 'softmax' in the range of hθ(x)∈ R and 0<hθ(x)<1. The θ and x are the vectors of weights
and input values, respectively.

H. Optimization Algorithm and Cost Function

In a neural network, updating the model parameter like weight and bias values needs an optimization algorithm. The
optimization algorithm runs a procedure of finding the optimum or satisfactory solution. An optimization algorithm's
primary purpose is to minimize the loss or error of a neural network model. Some popular examples of the optimization
algorithm are Stochastic Gradient Descent (SGD), Batch Gradient Descent (BGD), Nadam (Nesterov Adam optimizer),
Adagrad (Adaptive Gradient), Adadelta (Adaptive Delta), Adam (Adaptive Moment Estimation), Adamax, RMSprop,

9
etc. The Adaptive Moment Estimation (Adam) gradient descent algorithm was being used to optimize our neural
network. The main target of a perfect neural network is not to increase the accuracy but to reduce the loss. The
optimization algorithm always updates the values of weights and biases to reduce the value of the loss.
In protocol 1, where the number of output classes M=2, the cross-entropy was calculated by using (12). Similarly, in
the 2nd protocol, as the number of output classes is three, the categorical cross-entropy was calculated by (13).

cost p 1  [ y log( p)  (1  y ) log( 1  p)] (12)

M
cost p 2   yo ,c log( Po ,c ) (13)
c 1

where 'M' indicates the number of output classes, 'y' is the binary indicator (0,1) that defines the correct or incorrect
classification, and 'Poc' implies the predicted probability on observation 'o' of class 'c'.
Table III
Cost Function of the Binary and Multi-class Classification Problem
Types of Configuration of the output
Cost function
Problem layer
Binary One node with a 'sigmoid' Binary Cross-Entropy
Classification activation function.
Multi-Class One node for each class using the Categorical Cross-Entropy
Classification 'softmax' activation function.

Table IV
Shape and Parameter Values of Our CNN Model in Protocol 1

Layer (type) Output Shape Parameter

Input 23, 23, 1 -
Conv_1 (Conv2D) 23, 23, 64 640
AvgPool_1 (AveragePooling) 11, 11, 64 0
Conv_2 (Conv2D) 9, 9, 128 73856
AvgPool_2 (AveragePooling) 4, 4, 128 0
Conv_3 (Conv2D) 4, 4, 256 295168
AvgPool_3 (AveragePooling) 2, 2, 256 0
Flatten_1 (Flatten) 1024 0
Dense_1 (Dense) 512 524800
Dropout_1 (Dropout) 512 0
Dense_2 (Dense) 2 1026
Total parameter: 895,490
Trainable parameter: 895,490
Non-trainable parameter: 0

Table V
Shape and Parameter Values of Our CNN Model in Protocol 2
Layer (type) Output Shape Parameter
Input 23, 23, 1 -
Conv_1 (Conv2D) 23, 23, 64 640
AvgPool_1 (AveragePooling) 11, 11, 64 0
Conv_2 (Conv2D) 9, 9, 128 73856
AvgPool_2 (AveragePooling) 4, 4, 128 0
Conv_3 (Conv2D) 4, 4, 256 295168
AvgPool_3 (AveragePooling) 2, 2, 256 0
Flatten_1 (Flatten) 1024 0
Dense_1 (Dense) 512 524800
Dropout_1 (Dropout) 512 0
Dense_2 (Dense) 3 1539
Total parameter: 896,003
Trainable parameter: 896,003
Non-trainable parameter: 0

10
 We used the 'Google-Colab' user interface to execute the code. The summary of the proposed emotion recognition
model is shown in Table IV and Table V, respectively. It may be noted that the two architectures were almost the same
except for the output shape of the last dense layer. As we classified 2 and 3 different classes in protocols 1 and 2,
respectively, the last dense layer's shape was 2 and 3.

III. Result
In our experiment, we used the 'DEAP' dataset of the EEG signals to classify emotion. The emotion-related EEG
signals were firstly converted into PCC featured images. Here, we calculated the correlations of different channels data
for the same emotional video and the same segmentation. This channel correlation matrix contains the PCC between
the combinations of every two channels for similar segmentation. Thus, the matrix of sized 32×32 could be found for
every segmentation. As in every 60s the EEG signals were segmented into 20 segments, and a participant showed 40
different emotional videos; the number of the total square matrices will be 20×40=800. Thus for 32 participants,
800×32, i.e., 25600 square matrices of channel correlation were generated for a single sub-band. The channel
correlation-based images were formulated with the highest correlation coefficient 1 as yellow color and the lowest
correlation coefficient -1 as blue color. Some samples

(a) (b)

p1, v1, s1 p1, v1, s2

(c) (d)

symmetric

p15, v20, s7 p31, v20, s7

Figure 7: Random samples of PCC featured images in which every pixel indicates the PCC of the relevant two channels data. (a) Image for
participant number 1, video number 1, segmentation number 1 (b) Image for participant number 1, video number 1, segmentation number 2 (c)
Image for participant number 15, video number 20, segmentation number 7 (d) Image for participant number 31, video number 20, segmentation
number 7.

of PCC-based images are shown in Fig. 7, where 'p' indicates the certain participant, 'v' is for video, and 's' represents
the segment number. Afterward, these images were fed into a CNN-based classification algorithm. In protocol-1, we
distinguish low and high levels of valence and arousal. Here, the levels are distinguished by following the scale as low:
0≤value≤4.5 and high: 4.5<value≤9. In protocol-2 low, medium, and high levels of valence and arousal are
differentiated by following the range of low: 0≤value≤3, medium: 3<value≤6 and high: 6<value≤9.
Alarcao and Fonseca [7] and Zheng and Lu [9] published their investigation that emotion was closely related to the beta
and gamma band and moderately associated with the alpha band. Furthermore, the theta band had a slight activity
depending on the emotion. On the contrary, as delta band was correlated with dreamless and deep sleep, unconscious
mind, and very

11
 (a) (b)

(c) (d)

Figure 8: Model accuracy and loss curves for two-class classification: (a) accuracy in valence, (b) loss in valence, (c) accuracy in arousal, (d) loss
in arousal.

(a) (b)

(c) (d)

Figure 9: Model accuracy and loss curves for three-class classification: (a) accuracy in valence, (b) loss in valence, (c) accuracy in arousal, (d)
loss in arousal.

poorly related to emotion. Therefore, we considered the three sub-bands except the delta, theta and main EEG data to
extract emotion using CNN.
Since we used a set of 25600 PCC featured images of the main frequency and three sub-band frequencies, there
remains a total of 25600×4=102,400 labeled PCC featured images. The data were divided into training (90%),
validation (5%), and testing (5%) data. During the training period, the model validated the accuracy of the data selected
in the validation portion. The accuracy and loss of 2 classes and 3 classes classification of valence and arousal are
shown in Fig. 8 and Fig. 9. From these graphs, it is clear that the minimum loss occurs between the 45 to 50 epochs for

12
both of the two protocols. As in the machine learning approach, the minimization of loss is more desirable than the
maximization of accuracy, we used to train the system up to 50 epochs.

Low Medium High Low Medium High

Predicted Class 0 1 2
0 1 2

High Medium Low

High Medium Low
Low High Low High

0
0
0 1 0 1
Actual Class

Low

Low

1
0

1
High

High

2
2
1

(a) (b) (c) (d)

Figure 10: Confusion Matrices (a) two-class valence (b) two-class arousal (c) three class valence, (d) three class arousal.

The confusion matrices of our model are illustrated in Fig. 10. The four separate confusion matrices are generated
for the two-class valence, two-class arousal, three-class valence, and three-class arousal classification task. In every
confusion matrix, the last row represents the value of precision, and the last column represents the value of recall.
Besides, the last row and the last column data indicates the percentage accuracy of the specific task.

Table VI
Classification Outcomes of Our Model
Valence /
Class Precision Recall F1 score support
Arousal
Protocol 1: 0= Low, 1= High
0 0.74 0.75 0.74 2168
Valence
1 0.81 0.81 0.81 2952
0 0.80 0.68 0.74 2234
Arousal
1 0.78 0.87 0.82 2886
Protocol 2: 0=Low, 1=Medium, 2= High
0 0.63 0.70 0.66 816
Valence 1 0.69 0.64 0.66 2072
2 0.74 0.76 0.75 2232
0 0.70 0.59 0.64 858
Arousal 1 0.70 0.63 0.66 2055
2 0.70 0.81 0.75 2207

The overall classification report containing the value of precision, recall, F1 score, and accuracy of every class of
two different protocols is given in Table VI. The high value of the F1 score indicates that the importance of precision
and recall were balanced and satisfactory for both of the two protocols.
Here, only the inter-channel correlations were considered. The inter-channel correlations were more meaningful as
the human brain remains aroused by any single emotional stimuli in only a few seconds (1s-4s). On the other hand, the
inter-sample correlations may be another option for future work. It was important as it allowed one to consider the
variation of temporal characteristics of EEG signals for different emotional states. In addition, the inter-sample
correlation technique will enlarge the volume of data that promotes the deep machine learning algorithm performance
for large-sized data.

13
IV. Discussion
To recognize emotion firstly, the channel correlation-based PCC featured images were generated, as stated earlier. The
PCC-featured images of size 32×32 are consisted by the diagonal symmetrical matrices that were showed clearly in
Fig. 7. Since the diagonal elements represented the correlation between the same channels, the value will always be 1
for all of the PCC matrices. As a result, in every image, the color of the diagonal element was yellow.

Figure 11: Complexity reduction considering only the upper triangular matrix.

Since the upper or lower triangular matrix data of a single PCC-based image were the same, it was unnecessary to
use the same data in double time. Therefore, before sending the PCC matrices to our CNN model, the lower triangular
data were removed. Thus, the data points were reduced from 1024 (32×32) to 528 (1+2+…..+32) as explained in Fig.
11. An extra 0 data was added with this 528 data points to formulate a new shape PCC featured image of size 23×23.
Eventually, only 51.1% of data points of the square matrix of PCC-based images were fed to the CNN model because
these points consist of significant information for emotion recognition. Using this technique, the size of memory and
the computational complexity had been reduced without degrading the emotion recognition accuracy.
The data points of input images were reduced from 1024 to 529, indicating a 48.34% memory reduction only in this
step of execution. The reduction of complexity in protocols 1 and 2 has been given in Table VII considering some
specific factors. In protocol-1, the total trainable parameter was reduced to 895,490 from 1,550,850, which tends to a
42.59% reduction in computation for the whole program. We used the GPU of ‘Google Colab’ with a 5 Mbps internet
connection through the laptop of configuration Intel(R) Core(TM) i5-8250U CPU, 1.60GHz, 7.88GB DDR4 RAM,
256GB SSD, 64-bit operating system for executing our programing code. The new model also needed less time for
execution by using the same processor and same internet speed. The total time for 50 epochs was reduced to 207s from
351s. Similarly, in protocol-2, (see Table VII); the new model shows the 48.34%, 42.24%, 66.56% reduction of
complexity in the input size, number of total parameters and execution time respectively.
Table VII
Computational Complexity Reduction based on Input Size, Parameters, Execution Time in New Model
Factors for reducing Number of total Execution time
Input size
complexity trainable parameters (for 50 epochs)
Protocol-1
Old Model 32×32=1024 1,550,850 351s
New Model 23×23=529 895,490 207s
Complexity Reduction (%) 48.34% 42.59% 41.03%
Protocol-2
Old Model 32×32=1024 1,551,363 302s
New Model 23×23=529 896,003 101s
Complexity Reduction (%) 48.34% 42.24% 66.56%

The overall accuracy of our proposed model is shown in Table VIII. In Table VIII, it is represented that the accuracy
for two-class valence and arousal recognition tasks is 78.22% and 74.92%, respectively. In the three-class
classification, the accuracy is 70.23% and 70.25% for valence and arousal recognition tasks. Another important fact
that the percentage accuracy of protocol-1 is comparatively higher than the accuracy

14
 Table VIII
Comparison of Different Emotion Recognition Methods
Number of Classified
Author Extracted Features Classifier Accuracy (%) Dataset
emotion
Koelstra et al. Multimedia Content Gaussian Naive Bayes 4 (HVHA, HVLA, Valence=57.6
DEAP
(2011) [34] Analysis (MCA) classifier LVHA, LVLA) Arousal=62.0
Jirayucharoensak Power Spectral Density Deep learning network with Valence=49.5
3 class DEAP
et al. (2014) [35] (PSD) SAE Arousal=46.0
Ackermann et al. Average
Statistical feature SVM and Random Forest 3 (anger, surprise, etc.) DEAP
(2016) [4] Accuracy= 55
Valence=81.4
2 (low, high)
Tripathi et al. Statistical features in time Convolutional Neural Arousal=73.4
DEAP
(2017) [31] and frequency domain Networks (CNN) Valence=66.7
3 (low, medium high)
Arousal=57.6
Dynamical Graph Valence= (86.2±12.3)
Song et al. PSD with
Convolutional Neural 2 (low, high) Arousal= (84.5±10.8) DREAMER
(2018) [36] graph
Networks (DGCNN) Dominance= (85±10.3)
Valence= (97.69±1.22)
DEAP
2D frame sequence Arousal= (97.53±1.52)
Cheng et al.
considering spatial Deep Forest 2 (low, high) Valence= (89.03±5.56)
(2019) [37]
position Arousal= (90.41±5.33) DREAMER
Dominance= (89.89±6.19)
Power Spectral Density
Fang et al. Multi-Feature Deep Forest 5 (angry, happy, sad,
(PSD) and Overall accuracy = 71.05 DEAP
(2020) [38] (MFDF) pleasant, and neutral)
Differential Entropy (DE)
Our
Valence=78.22
Method, 2 (low, high)
Arousal=74.92
P1 Pearson’s Correlation Convolutional Neural
DEAP
Our Coefficient (PCC) Networks (CNN)
Valence=70.23
Method, 3 (low, medium high)
Arousal=70.25
P2

of protocol-2. The cause is that whenever the number of classes to be categorized increases, then the loss will also
increase. As a result, the accuracy will decrease. From the comparison table of work like ours, the accuracy of our
proposed method is indeed satisfactory. The accuracy is not the highest, then how can it be acceptable? Our target is
not to design a model of maximum accuracy, the target was to develop a model of lower computational complexity and
lower memory and time required. For instance, the model of Tripathi et al. [31] works with lots of features, whereas we
used only one feature named PCC. Besides, the accuracy in just 2 class valence is only higher; on the contrary, the
remaining accuracy is lower than our method. The model of Song et al. [36] is very effective, but as the dataset is not
the same as ours, the proper comparison is not too meaningful here. Recently, some deep learning-based models, such
as the deep forest-based model [37], [38] and rhythm-specific deep learning-based model [39] perform very well in the
context of emotion recognition. Cheng et al. [37] used a deep forest algorithm on the 2D frame sequence considering
the spatial position across channels. They applied it on DEAP and DREAMER datasets and achieved top-ranked results
in terms of accuracy stated in Table VIII. Fang et al. [38] used Multi-Feature Deep Forest as a classifier on PSD and
DE features and classified emotion into five classes. The model was tested using the DEAP dataset and achieved
71.05% overall accuracy, which is the highest for many classes classification. However, we focused on developing a
model of lower computational complexity with a satisfactory level of classification accuracy. Our work's novelty is to
use the technique of using only the upper triangular matrix data that reduces the size of input PCC-featured images
from 32×32 to 23×23. It drastically reduces the computational complexity and time of training and operation.
Moreover, we used only the PCC featured images with the CNN model; except for this feature, no additional effort was
required to find many features. It reduces the trouble to extract features manually.

V. Conclusion
In this paper, we have used the Convolutional Neural Network model to recognize emotion from EEG signals. Despite
using raw EEG data a systematically developed PCC featured images were considered for lower computation
complexity and short operation time. The whole process of emotion recognition was completed by following the rule of

15
classification using logistic regression. Here, two different protocols were used; in both protocols, we observed the
accuracy, model loss, and classification report. We achieved 78.22% and 70.23% accuracy in the valence classification
task and 74.92% and 70.25% accuracy in the arousal classification task for two and three classes classification
respectively. More accuracy and real-time operation compatibility are the fundamental prerequisites to enjoy the
advantages of emotion recognition in the practical field. Consequently, more research and studies are essential in
channel reduction, significant feature extraction, and deep network optimization in the future.

References
[1] P. Romaniszyn-Kania et al., “Affective state during physiotherapy and its analysis using machine learning methods,” Sensors, vol. 21,
no. 14, p. 4853, 2021, doi: 10.3390/s21144853.
[2] D. Mathersul, L. M. Williams, P. J. Hopkinson, and A. H. Kemp, “Investigating models of affect: relationships among EEG alpha
asymmetry, depression, and anxiety.,” Emotion, vol. 8, no. 4, p. 560, 2008.
[3] D. Sammler, M. Grigutsch, T. Fritz, and S. Koelsch, “Music and emotion: electrophysiological correlates of the processing of pleasant
and unpleasant music,” Psychophysiology, vol. 44, no. 2, pp. 293–304, 2007.
[4] P. Ackermann, C. Kohlschein, J. Á. Bitsch, K. Wehrle, and S. Jeschke, “EEG-based automatic emotion recognition: Feature extraction,
selection and classification methods,” in 2016 IEEE 18th International Conference on e-Health Networking, Applications and Services,
Healthcom 2016, 2016, pp. 1–6, doi: 10.1109/HealthCom.2016.7749447.
[5] M. R. Islam et al., “Emotion Recognition From EEG Signal Focusing on Deep Learning and Shallow Learning Techniques,” IEEE
Access, vol. 9, pp. 94601–94624, 2021, doi: 10.1109/access.2021.3091487.
[6] M. M. Rahman et al., “Recognition of Human Emotions Using EEG Signals: A Review,” Comput. Biol. Med., p. 104696, 2021, doi:
https://doi.org/10.1016/j.compbiomed.2021.104696.
[7] S. M. Alarcão and M. J. Fonseca, “Emotions recognition using EEG signals: A survey,” IEEE Trans. Affect. Comput., vol. 10, no. 3, pp.
374–393, 2019, doi: 10.1109/TAFFC.2017.2714671.
[8] M. R. Islam and M. Ahmad, “Wavelet Analysis Based Classification of Emotion from EEG Signal,” 2019, doi:
10.1109/ECACE.2019.8679156.
[9] W. L. Zheng and B. L. Lu, “Investigating Critical Frequency Bands and Channels for EEG-Based Emotion Recognition with Deep
Neural Networks,” IEEE Trans. Auton. Ment. Dev., vol. 7, no. 3, pp. 162–175, 2015, doi: 10.1109/TAMD.2015.2431497.
[10] E. Barsoum, C. Zhang, C. C. Ferrer, and Z. Zhang, “Training deep networks for facial expression recognition with crowd-sourced label
distribution,” in Proceedings of the 18th ACM International Conference on Multimodal Interaction, 2016, pp. 279–283.
[11] M. Soleymani, S. Asghari-Esfeden, Y. Fu, and M. Pantic, “Analysis of EEG Signals and Facial Expressions for Continuous Emotion
Detection,” IEEE Trans. Affect. Comput., vol. 7, no. 1, pp. 17–28, 2016, doi: 10.1109/TAFFC.2015.2436926.
[12] S. R. Kadiri, P. Gangamohan, S. V. Gangashetty, and B. Yegnanarayana, “Analysis of excitation source features of speech for emotion
recognition,” in Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH, 2015,
vol. 2015-January, pp. 1324–1328.
[13] M. T. Valderas, J. Bolea, P. Laguna, M. Vallverdú, and R. Bailón, “Human emotion recognition using heart rate variability analysis
with spectral bands based on respiration,” in 2015 37th Annual International Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC), 2015, pp. 6134–6137.
[14] S. Piana, A. Stagliańo, F. Odone, and A. Camurri, “Adaptive Body Gesture Representation for Automatic Emotion Recognition,” ACM
Trans. Interact. Intell. Syst., vol. 6, no. 1, pp. 1–31, 2016, doi: 10.1145/2818740.
[15] A. Vaskinn, K. Sundet, T. Østefjells, K. Nymo, I. Melle, and T. Ueland, “Reading emotions from body movement: A generalized
impairment in schizophrenia,” Front. Psychol., vol. 6, no. JAN, p. 2058, 2016, doi: 10.3389/fpsyg.2015.02058.
[16] J. Shukla, M. Barreda-Angeles, J. Oliver, G. C. Nandi, and D. Puig, “Feature Extraction and Selection for Emotion Recognition from
Electrodermal Activity,” IEEE Trans. Affect. Comput., vol. 5, no. 3, pp. 327–339, 2019, doi: 10.1109/TAFFC.2019.2901673.
[17] K. Schaaff and T. Schultz, “Towards emotion recognition from electroencephalographic signals,” in Proceedings - 2009 3rd
International Conference on Affective Computing and Intelligent Interaction and Workshops, ACII 2009, 2009, pp. 1–6, doi:
10.1109/ACII.2009.5349316.
[18] Z. Mohammadi, J. Frounchi, and M. Amiri, “Wavelet-based emotion recognition system using EEG signal,” Neural Comput. Appl., vol.
28, no. 8, pp. 1985–1990, 2017, doi: 10.1007/s00521-015-2149-8.
[19] S. K. Hadjidimitriou and L. J. Hadjileontiadis, “EEG-based classification of music appraisal responses using time-frequency analysis
and familiarity ratings,” IEEE Trans. Affect. Comput., vol. 4, no. 2, pp. 161–172, 2013.
[20] M. Li and B.-L. Lu, “Emotion classification based on gamma-band EEG,” in 2009 Annual International Conference of the IEEE
Engineering in medicine and biology society, 2009, pp. 1223–1226.
[21] L. He, D. Hu, M. Wan, Y. Wen, K. M. Von Deneen, and M. C. Zhou, “Common Bayesian Network for Classification of EEG-Based
Multiclass Motor Imagery BCI,” IEEE Trans. Syst. Man, Cybern. Syst., vol. 46, no. 6, pp. 843–854, 2016, doi:

16
 10.1109/TSMC.2015.2450680.
[22] Y. Velchev, S. Radeva, S. Sokolov, and D. Radev, “Automated estimation of human emotion from EEG using statistical features and
SVM,” in 2016 Digital Media Industry & Academic Forum (DMIAF), 2016, pp. 40–42.
[23] M. Chen, J. Han, L. Guo, J. Wang, and I. Patras, “Identifying valence and arousal levels via connectivity between EEG channels,” in
2015 International Conference on Affective Computing and Intelligent Interaction, ACII 2015, 2015, pp. 63–69, doi:
10.1109/ACII.2015.7344552.
[24] M. Soleymani, M. Pantic, and T. Pun, “Multimodal emotion recognition in response to videos (Extended abstract),” 2015 Int. Conf.
Affect. Comput. Intell. Interact. ACII 2015, vol. 3, no. 2, pp. 491–497, 2015, doi: 10.1109/ACII.2015.7344615.
[25] W. L. Zheng, J. Y. Zhu, and B. L. Lu, “Identifying stable patterns over time for emotion recognition from eeg,” IEEE Trans. Affect.
Comput., vol. 10, no. 3, pp. 417–429, 2019, doi: 10.1109/TAFFC.2017.2712143.
[26] J. Atkinson and D. Campos, “Improving BCI-based emotion recognition by combining EEG feature selection and kernel classifiers,”
Expert Syst. Appl., vol. 47, pp. 35–41, 2016, doi: 10.1016/j.eswa.2015.10.049.
[27] M. Längkvist, L. Karlsson, and A. Loutfi, “Sleep stage classification using unsupervised feature learning,” Adv. Artif. Neural Syst., vol.
2012, 2012.
[28] K. Li, X. Li, Y. Zhang, and A. Zhang, “Affective state recognition from EEG with deep belief networks,” in 2013 IEEE International
Conference on Bioinformatics and Biomedicine, 2013, pp. 305–310.
[29] H. P. Martinez, Y. Bengio, and G. Yannakakis, “Learning deep physiological models of affect,” IEEE Comput. Intell. Mag., vol. 8, no.
2, pp. 20–33, 2013, doi: 10.1109/MCI.2013.2247823.
[30] Z. Wen, R. Xu, and J. Du, “A novel convolutional neural networks for emotion recognition based on EEG signal,” in 2017
International Conference on Security, Pattern Analysis, and Cybernetics (SPAC), 2017, pp. 672–677.
[31] S. Tripathi, S. Acharya, R. D. Sharma, S. Mittal, and S. Bhattacharya, “Using Deep and Convolutional Neural Networks for Accurate
Emotion Classification on DEAP Dataset,” in Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, 2017, pp.
4746–4752, [Online]. Available: https://www.aaai.org/ocs/index.php/IAAI/IAAI17/paper/view/15007/13731.
[32] H. Mei and X. Xu, “EEG-based emotion classification using convolutional neural network,” in 2017 International Conference on
Security, Pattern Analysis, and Cybernetics (SPAC), 2017, pp. 130–135.
[33] M. R. Islam and M. Ahmad, “Virtual Image from EEG to Recognize Appropriate Emotion using Convolutional Neural Network,”
2019, doi: 10.1109/ICASERT.2019.8934760.
[34] S. Koelstra et al., “DEAP: A database for emotion analysis; Using physiological signals,” IEEE Trans. Affect. Comput., vol. 3, no. 1,
pp. 18–31, 2012, doi: 10.1109/T-AFFC.2011.15.
[35] S. Jirayucharoensak, S. Pan-Ngum, and P. Israsena, “EEG-Based Emotion Recognition Using Deep Learning Network with Principal
Component Based Covariate Shift Adaptation,” Sci. World J., vol. 2014, 2014, doi: 10.1155/2014/627892.
[36] T. Song, W. Zheng, P. Song, and Z. Cui, “EEG emotion recognition using dynamical graph convolutional neural networks,” IEEE
Trans. Affect. Comput., vol. 11, no. 3, pp. 532–541, 2018.
[37] J. Cheng et al., “Emotion Recognition from Multi-Channel EEG via Deep Forest,” IEEE J. Biomed. Heal. Informatics, vol. 25, no. 2,
pp. 453–464, 2021, doi: 10.1109/JBHI.2020.2995767.
[38] Y. Fang, H. Yang, X. Zhang, H. Liu, and B. Tao, “Multi-Feature Input Deep Forest for EEG-Based Emotion Recognition,” Front.
Neurorobot., 2020.
[39] D. Maheshwari, S. K. Ghosh, R. K. Tripathy, M. Sharma, and U. R. Acharya, “Automated accurate emotion recognition system using
rhythm-specific deep convolutional neural network technique with multi-channel EEG signals,” Comput. Biol. Med., vol. 134, p.
104428, 2021.

17
 EEG based emotion recognition model is proposed using Convolutional Neural Network
architecture
 Pearson’s Correlation Coefficients (PCC) of alpha, beta and gamma sub-bands are taken
into consideration for emotion recognition
 A novel method of emotion recognition focusing on lower computational complexity based
on memory requirement and computational time
 Low, medium and high level of valence and arousal based emotion recognition with only
one feature of data named PCC.