Comparative analysis of Identification and

Classification of Face Emotions Using

Different Machine Learning and Deep
Learning Algorithms
Author1, Author2
Designation1
Designation2
ABSTRACT: Sentiments are significant and profound features of individual conduct. Examining
facial looks and acknowledging their expressive condition is stimulating job by extensive reaching
functions. Human Face expression Recognition is one of very influential and stimulating chores in
public interaction. Usually, face expressions are usual and straight ways for human beings for
communicate their emotions and intentions. Face expressions are the key characteristics of non-
verbal communication. Here, we introduce available dataset i.e., CK+, JAFFE and FER dataset that
are widely used in this work. This paper focuses upon facial expression recognition technique
founded on machine learning algorithms pair and also deep learning algorithms that will assist in
precise recognition and organization of human emotion.

KEYWORDS: FER, SVM, KNN, RNN, CNN, DBN;

INTRODUCTION
I.
Humans are expressive beings. Our expressive condition notifies how we perform by very
important procedures, to compound activities and hard choices. Our lives are in numerous paths
directed from our feelings, so understanding additional regarding feelings lets us to see more
regarding human conduct extra usually. It’s apparent that comprehending expressive condition of
human could be valuable for variety of uses by improving good comprehension of human
psychology, for examining behavior for better consumer practices, for improving useful publicity
canvasses, and outside.

Human emotions are categorized like: fear, contempt, disgust, and anger, surprise, sad, happy,
and neutral. These feelings are extremely delicate. Facial musculus frowns are extremely slight and
sensing these changes is extremely daring as even minor change consequences in changed
expressions [1]. Expressions of dissimilar or even same human may differ to same emotion as well,
as emotions are immensely situation supported [2]. Although one could emphasize on only those
parts of face that show most of feelings such as about mouth and eyes [3], how we obtain these
signs and classify them is yet an significant query. Neural networks and machine learning are
utilized to these chores and contain gotten decent outcomes. Machine learning algorithms have
established that they are extremely useful in pattern detection and categorization. Highly significant
features for any machine learning algorithm are characteristics.

Facial expression detection utilizes algorithm for detecting faces, codes facial expression and
spot expressive conditions. It performs this from examining faces in images or videos by cameras
fixed in laptops, phones, or computers. This recognizes diverse feelings on human face, business
images plus videos in present for examining video supplies. Facial expression detection discovers
uses in animated cinemas, check person’s pressure stage, mining demonstrations of psychiatric
patients. This is related in motorist’s sleepiness discovery as well. Intelligent automobiles can
1111
aware driver when he is sensing sleep from initially noticing his face and later eyes. This is
appropriate in emotion recognition in meeting for determining whether applicant’s character is
decent appropriate for work as well. This is utilized in analysis in video games as well. Throughout
examining stage, consumers are requested to compete in game for specified time and its response is
combined for making last creation.

FER (Face Expression Recognition) possesses significant phase is feature extraction and
cataloguing. Feature extraction comprises 2 kinds and these are symmetrical founded and look
founded. Cataloging is one of significant procedures as well where aforementioned expressions like
smirk, gloomy, fury, repugnance, astonishment, and distress are characterized. Geometrically
founded characteristic abstraction includes eye, mouth, nose, eyebrow, other facial parts plus look-
founded characteristic abstraction includes precise unit of face [4].

The overview of the Face expression recognition structure is demonstrated in Fig. 1. Face
expression recognition scheme comprises main phases like face image pre-treating, feature
abstraction and cataloging.

Fig. 1. Architecture of face expression detection structure

II. RELATED WORK

Problem of detection and fortitude of human feelings is always a significant study part from the
period of Charles Darwin who initially pointed connection amid induced expressive conditions and
typical expressions of human face. Darwin’s model is established from numerous various surveys
in following periods [5–7]. He presumed that emotional expressions are multimodal behavioural
outlines of person, and therefore shaped his personal full depictions of 40+ emotional conditions
[8]. Over last century, numerous diverse replicas for emotion arrangements, fluctuating from
generally felt elementary emotions to exclusive and compound ones were sensitively demarcated.
Two replicas investigated in arena of emotion detection [9, 10] are mainly utilized in preceding
period: elementary cataloguing of 6 emotional conditions from Ekman and Russell’s circumflex
prototype of emotions.

For present feeling detection structure, numerous methods are predictable. Nominal number of
AUs are used for detecting facial expressions [11, 12], even though changing numbers of AUs
utilized in evolving sharp facial expression detection structure. Delaunay triangulation technique is
utilized for connecting 68 facial activity parts in focussed face for detecting 7 facial expressions
1112
like contentment, fury, terrified, astonishment, grief, neutral, plus repugnance. Multi class support
vector machine (SVM) classifier executes fine having highest mean facial expression detection
degree of 84% founded on spatiotemporal traits in categorizing emotional expressions [13].

Kartali et al. have described outcomes of orthodox (SVM, Multi-Layer Perceptron (MLP)) and
deep studying approaches (Convolutional Neural Network (CNN), Alexnet CNN, Affdex CNN)
founded facial expression detection of 4 emotions (contentment, grief, fury, and distress) plus
attained highest detection precision of 85.05% utilizing Affdex CNN [14]. New vectored emotion
detection prototype is projected for identifying 3 main emotions: livid, pleased, plus neutral,
utilizing seventy facial vectors and deep neural network (DNN), and attained average precision of
84.33% [15]. In new works, researchers have utilized altitudinal and chronological data by input
video arrangements for classifying diverse facial expressions utilizing CNN, Collaborative Multi-
level CNN, and Long Short-term Memory (LSTM) [16–19]. Few usual subjects stated in previous
literatures because of absence of examples or information groups, minimal precision in
categorizing facial expressions, advanced computational difficulty (extra memory and power
needed to process information), not appropriate for present uses, and not easy method (limitations
in utilizing structure for diversity of uses) [20].

III.DATASET

In Present scenario Facial Expressions acts an significant part in all fields like Scientific, Medical,
Social, business, gaming, child psychology etc. In this work we have used CK+, JAFFE and FER
Dataset.

CK+: Lengthy Cohn-Kanade (known as CK+) facial expression folder [22] is community folder
for action part and emotion detection. This comprises posed and non-posed (impulsive) expressions
as well. CK+ includes whole of 593 series thru 123 subjects. In majority of preceding
compositions, final frame of these orders is selected and utilized for image

JAFFE: Japanese Female Facial Expression (JAFFE) folder is lab- regulated image folder which
comprises 213 models of modelled expressions by 10 Japanese ladies. Every person possesses 3˜4
images giving all 6 elementary facial expressions (fury, repugnance, distress, pleasure, grief, and
astonishment) and 1 image having neutral expression. Folder is difficult since this comprises some
instances per subject/expression. Classically, every image is utilized for leave-one-subject-out test.

FER: Facial expressions dataset with respect to perfect lighting conditions and lab environment.
Subjects photographed are between the age 5 to 70, school children, college students, professionals
and stage artists have participated in this process with the consent. Photography is done in various
places such as school, college and drama stages.

Method
Development of the FED Database. The participants were 49 Children (M : 12.65 years, SD : 0.57;
age range, 12–16; 57% Females), 41 Young (M : 20.24 years, SD : 1.15; age span, 19–31; 58%
women), middle aged 30 (M: 41.06 years, SD : 4.26; age span, 38–55; 46% women) Old Aged 58
and above ( M: 61.33 Years , SD:0.16 ;age span 58-80) extras, specials, or performers who only
had certain exclusions, Every contributors are from school, college and drama school.

Before to photo-shoot gathering, contributors were educated through mobile, e-mail, or fax
that goal of scheme was depositing record including portrayals of child, young, middle-aged, and
older adults showing 6 various facial expressions to utilize in technical studies. Contenders who are
proficient of convey every 6 various faces (by aid of face teaching and by backing of skilled
1113
practiced associate). Applicants were requested to approach with no cosmetics and attire regular
dress on photo-shooting gathering day.

Procedure and Materials.

Photo-shooting shifts happened at KSS amid FEB 2019 and May 2019 in photo studio precisely
arranged for this use. Skilled expert associate notified participants regarding overall goal of scheme
plus of day’s shift, and also regarding specific process in photo studio. Contributors were expressed
that they would be snapped numerous instances, displaying every 6 facial expressions. Photography
professional and photography associate aided contributors display these facial expressions by
afresh industrialized process which included 3 (partially mixed).

Validation of the Facial expression Database

Validation of Facial Expression Database is done to help the research people with help of assessors,
all the 600 images are perceived as child, young, middle aged and old aged.

Participants in the validation study many contributors were requested for validating face data,
they have asked for identifying the different emotions by rating them. Totally 600 images were
selected.

IV. MACHINE LEARNING ALGORITHMS TO RECOGNIZE AND CLASSIFY FACE EMOTIONS

Support Vector Machines (SVM)

Support Vector Machine (SVM) was initially perceived in 1992, presented from Boser, Guyon, and
Vapnik in COLT-92. These are group of linked regulated learning approaches utilized to classify
and regression. They are of group of comprehensive linear classifiers. In another words, this is
cataloguing and regression estimate device which utilizes machine learning model for maximizing
prognostic precision although unconsciously evading over-fit to info. These is outlined as structures
that utilize theory space of linear functions in high dimensional feature space, directed by learning
algorithm by optimization model which applies learning preconception resultant by statistical
learning model.

SVM are very influential cataloguing algorithms. The outline is finding ideal hyper plane
that splits 2 groups precisely. There is notion of boundary as well, that is meant to greatest among
both groups for avoiding further imbrication amid 2 groups [21]. Info that is not linearly divisible is
charted into advanced dimension for achieving healthier cataloguing outcomes. Kernel functions
like radial basis function (rbf) and polynomial are utilized for non-linear info [22].

1114
 In the event of emotion recognition, typically multi-class SVM is utilized in place of binary
for detecting emotions like fury, disdain, repugnance, distress, pleased, dejection and astonishment.
K-fold cross-authentication is utilized for removing somewhat alterations in folder and for
comparing diverse machine learning algorithms [23]. In k-fold cross authentication, folder is split k
times to k shares, and forecast outcomes are averaged over every iteration. Loconsole et al. [24]
utilized Principal component analysis (PCA) for characteristic set decrease and later nourishing
condensed characteristic fit to SVM. In PCA algorithm, image characteristic distance is altered to
eigen distance utilizing eigen matrix [25].

Alongside kernel arrangement, SVM possess approaches for changing limits such as C and
γ [26]. In this, C is penalty function for misclassification and gamma aids in optimising decision
boundary. Both limits disturb precision of classifiers and is altered for getting best consequences in
both binary and multi-class cataloguing.

K nearest neighbour (KNN)

KNN is plain nonlinear classifier utilized in numerous purposes, counting epilepsy
recognition, driver sleepiness recognition, emotion detection, seizure recognition, and numerous
extra difficulties. KNN is non-probabilistic learning algorithm utilized for classifying unknown trial
info founded on bulk of alike info amongst k-nearest neighbours nearby trial/unidentified info.
Diverse distance quantities can quantify space among trial info and every training info, like
Manhattan, Euclidean, Minkowski, and Chebyshev. Here, overhead 4 space capacities are utilized
for distinguishing facial emotional demonstrations, and average Precision of every space measure is
stated.

In here, k-Nearest Neighbours algorithm (or k-NN) is non-parametric technique that is

utilized to classify and regression. Input contains of k neighbouring training instances in
characteristics space. Output hangs on whether k-NN is utilized for regression or cataloguing. In k-
NN cataloguing, output is group associate. Object is categorised from bulk support of their
neighbour, with object being allocated to group very usual amongst their k adjoining neighbour (k
is +ve integer, k<1). If k=1, nevertheless entity is only allocated to group of that solitary adjoining
neighbour.

In K-NN regression, output is property quantity for the object. This value is mean of values
of their k adjoining neighbors. K-NN originates below example-founded learning, or lazy learning,
in which function is simply estimated locally and every assessment is delayed till cataloguing.
KNN algorithm is amongst modest of every machine learning algorithm regarding to cataloguing
and regression, this could be valuable to weight aids of neighbours, therefore close neighbours
donate extra to mean compared to additional far ones. For instance, regular weighing system
contains in providing every neighbour weight of 1/d, in which d is distance to neighbour.
Neighbours are occupied by group of items for which object property amount (for K-NN
regression) and class (for K-NN classification) is identified. This could be believed of as training
group for algorithm; however no clear training phase is obligatory.

In K-NN cataloguing, training designs are planned in d dimensional period, in which d is

quantity of characteristics existing. These designs are planned corresponding to their experiential
characteristics quantities plus are labelled corresponding to their recognized group. An unlabelled
test design is planned inside similar space and is categorised corresponding to extremely commonly
happening group amid its k extremely alike training designs; its adjoining neighbours. Very usual
resemblance amount for K-NN organization is Euclidian distance metric, outlined among feature
vectors and as:

1115
 In which, f signifies quantity of characteristics utilized for representing every form. Minor
space amounts signify bigger resemblance. Cataloguing happens subsequently classifying k utmost
alike training points to inquiry point. Relatively than utilizing normal voting system, algorithm
utilized here allocates class tags to query points utilizing prejudiced system founded upon every
neighbour’s immediacy to query point. Let d be distance amount, and x1, x2, x,...,xk be k adjoining
neighbours of x decided in escalating sequence of d(xi,x). So x1 is primary adjoining neighbour of
x. They suggest for assigning weight wi to i-th adjacent neighbor xi characterised as:

Pattern x is allotted to group having weights of envois amongst k adjoining neighbours

amount to highest quantity. This regulation was displayed to produce lesser fault degrees compared
to those gotten utilizing voting K-NN regulation.

V. DEEP LEARNING ALGORITHMS TO RECOGNIZE AND CLASSIFY FACE EMOTIONS

Convolutional neural network (CNN)

In neural networks, ConvNets or CNNs is a chief group to perform images detection, images
guiding. Items recognitions, detecting faces etc., are parts inwhich CNNs are extensively utilized.
CNN were stimulated from biotic procedure in which link relationship amid neurons looks like
configuration of human pictorial cortex. Computers perceives input image as collection of pixels
and this hangs on image resolution.

Figure A . depicting mathematical model of CNN

1116
 Founded on image resolution, this will perceive h x w x d (h = Height, w = Width, d =
Dimension). Elementary section illustration of CNN is exhibited in fig. 2. It contains 3 levels:
Convolution layer, Pooling Layer and Fully connected layer.

Fig. 2. Block diagram of CNN

1) Convolution Layer
This is primary layer for extracting characteristics by input image. Convolution conserves
association amid pixels from learning image characteristic utilizing little squares of input info. This
is scientific process which uses 2 inputs like image matrix plus sieve or kernel. Sizes are as ensues:

Filter; 𝑓ℎ×𝑓𝑤×d
Image matrix; h ×w ×d

Output; (h-𝑓ℎ+1) × (w-𝑓𝑤+1) × 1

This output is denoted as characteristics map. ReLU situates for Rectified Linear Unit for
non-linear process. Output is ƒ(x) = max (0,x).

2) Pooling Layer
This unit will decrease quantity of limits when images are excessively big. Spatial pooling
a.k.a subsampling or down sampling that declines dimensionality of every map nevertheless
possesses significant info.
Spatial pooling has numerous classes:
 Max Pooling
 Average Pooling
 Sum Pooling
Max. Pooling uses major section by corrected characteristic map. Captivating main component
might captivate mean pooling as well. Totality of every element in characteristic map is called as
sum pooling.

3) Fully Connected Layer

Layer a.k.a FC layer, we compressed our matrix as vector and supply that as completely
linked level as neural network. This performs like human neuron, that interlinks along one another
for broadcast of data. Every characteristic maps by pooling unit are interlinked for providing
trained output. Hereafter, output level will distinguish trained images. This trained image is utilized
to compare to novel image. Subsequently contrast appropriate expression is documented.

CNN is significantly utilized in varied computer vision uses, comprising FER. At start of
21st period, numerous lessons in FER literature [27], [28] discovered that CNN is strong for facing
position variations and scale differences and acts well compared to multilayer perceptron (MLP) in
circumstance of formerly hidden confront posture differences. [29] Working CNN for addressing
difficulties of topic freedom and conversion, revolution, and scale invariance in detection of facial
expressions also. CNN has 3 kinds of varied levels: convolutional levels, pooling levels, and
completely linked levels. Convolutional level contain group of learnable sieves for convolving
throughout entire input image and give numerous precise kinds of start characteristic maps.
Convolution process is related to 3 chief advantages: local connectivity, that studies associations
1117
amongst neighbouring pixels; weight sharing in similar characteristic map, that importantly
decreases amount of limits to be studied; plus shift-invariance to position of item. Pooling level
trails convolutional level and is utilized for reducing latitudinal volume of characteristic maps and
computational price of system. Regular pooling and max pooling are 2 utmost usually utilized
nonlinear down-sampling plans to conversion invariance. Completely linked level is regularly
comprised at close of system for ensuring that every neuron in level are completely linked to starts
in preceding level and for enabling 2D characteristic maps to be transformed to 1D characteristic
maps for more characteristic illustration and cataloguing. We catalogue formations and features of
few famous CNN prototypes which are used for FER in Table 3. Also these systems, numerous
famous resultant structures subsist as well. In [30], [31], region-based CNN (R-CNN) [32] was
used to learning characteristics for FER. In [33], Faster R-CNN [34] was utilized for identifying
facial expressions from producing great excellence area offers. Furthermore, Ji et al. projected 3D
CNN [35] for capturing motion info programmed in manifold nearby limits for action detection
through 3D convolutions. Tran et al. [36] projected well-made C3D, that uses 3D convolutions on
important managed training folders for learning spatio-temporal characteristics.

CNN is distinctive and extensively utilized prototype for deep learning. Deep learning goals
for inevitably study and abstract multilevel characteristic illustration by unprocessed info. Features
of CNN, like local assembly, weight distribution, and down selection process, making it promising
to efficiently reducing difficulty of system, decrease number of teaching limits, and give benefits of
tough sturdiness and error lenience, and being simple for training and optimizing also.

Deep belief network (DBN)

DBN projected from Hinton et al. [37] is graphical prototype which absorbs for extracting deep
graded illustration of training info. Conventional DBN is constructed using load of restricted
Boltzmann machines (RBMs) [38], that are 2 level reproductive stochastic prototypes comprised of
visible-unit level plus concealed component level. These 2 levels in RBM should shape two-part
graph with no side networks. In DBN, components in advanced levels are coached for learning
provisional dependences amongst components in adjoining inferior levels, bar upper 2 levels, that
contain directionless networks. Guidance of DBN comprises 2 stages: pre-training and fine-tuning
[39]. Initially, well-organized level by level avaricious learning strategy [40] is utilized for
initializing deep network in unsupported method, that may stop meagre limited best outcomes to
some degree deprived of obligation of huge quantity of branded info. Throughout this process,
contrastive divergence [41] is utilized for training RBMs in DBN for estimating estimate gradient
of log-likelihood. Before, limits of network plus anticipated output are perfected having humble
incline lineage below management.

Hinton projected quick and avaricious algorithm which learns profound, focused belief networks; 1
layer at 1 period. Upper 2 levels shape directionless associative memory [42]. DBN is probabilistic
reproductive prototype comprised of numerous levels of stochastic plus latent variables [43]. 2
utmost important characteristics of DBN are applying efficient, layer-by-layer studying process
plus implication obligatory for starting percept is together quick plus precise. Figure 3 displays
reproductive prototype of DBN having 1 noticeable level and 3 hidden levels’ network.
Reproductive network produces contenders by input info whereas discriminative network assesses
them. This is approved for generating exclusive and truthful facial imageries and additional.
Studying probability sharing of input info differentiates DBNs. In DBNs, Restricted Boltzmann
Machine (RBM) possess capability of representing info characteristics, therefore this is utilized for
building its elementary shape. RBM contains 2 levels; noticeable level and concealed level. Figure
4 displays elementary RBM prototype.

1118
 Fig 3: Generative model of DBN having one noticeable and three concealed levels

Fig 4: Basic model of Restricted Boltzmann machine

Recurrent neural network (RNN)

RNN is connectionist prototype which imprisons temporal data plus is highly appropriate to
consecutive information forecast by random distances. In supplement to coaching deep neural
network in solitary feed-forward way, RNNs comprise repeated ends which cover adjoining period
phases and distribute similar limits throughout every stages. Classic back transmission via time
(BPTT) is utilized for training RNN. Long-short term memory (LSTM), presented from Hochreiter
& Schmidhuber, is distinctive method of conventional RNN which is utilized for addressing
gradient disappearing and blast complications which are normal in guiding RNNs. Cell status in
LSTM is planned and structured from 3 gates: an input gate which lets or obstructs change of cell
status from input signal, an output gate which allows or stops cell status for affecting additional
neurons, and forget gate which controls cell’s self-repeated linking for accumulating or forgetting
their preceding status. By uniting these 3 gates, LSTM can mould long-term addictions in order and
is extensively used in video-founded expression detection duties.

1119
 Figure 5. Mathematical model of RNN
We presumed that Euclidean metric, normal space amid two spots in frame sequence archives order
of events. For example, in state of contentment or pleasure, there are swift eye flashing, crow’s feet
crinkles in side edge of eyes thrust up cheeks, association by muscle which ranges eye (eye, cheek,
chin…). Nevertheless, in situation of dejection, eyes have extremely sluggish eye flashing, sagging
higher eyelids, trailing emphasis in eyes, slender dragging low of lip corners. In contrast to
situation of repulse, there is distorting of face leftward or rightward, nose is warped up to crumpled
nose bridge, tapering eyes, dropped brows. Also this, in dread condition, there are elevated
eyebrows, strained reduce eyelids, eyebrows pulled collected, lips pulled parallel. Finally, in rage
condition, eyebrows are dragged low collected, eyes are wide-opened and evident, upper eyelids
are elevated in gaze, lips are fully unlocked to develop rectangle, and firmly shut by red margins of
lips developing thinner, and lips developing thinner. Exact repeated neural network classifier,
known as Long Short-Term Memory, was presumed to be utilized so as to obtain benefit of their
aptitude to utilize dynamic temporal behaviour of order for arrangement [44], symmetrical
descriptor of every edge was nourished to network face recognition phase later.

1120
 Fig 6: Sorting utilizing LSTM-RNN Classifier

VI. RESULTS
The results of this work are shown below:

Table 1: Comparison between machine learning algorithms i.e. KNN and SVM

Sl. No Algorithm Precision

1) KNN 95.5%

2) SVM 96.5%

SVM gives better accuracy in identification and classification of face emotions compared to KNN
as revealed in above Table 1

Table 2: Contrast amid deep learning algorithms i.e. CNN, RNN and DBN

Sl. No Algorithm Precision

1) CNN 97%

2) RNN 96%

3) DBN 95%

CNN gives better accuracy in identification and classification of face emotions compared to other
algorithms as shown in the above Table 2

1121
 Fig 3: Accuracy Graph
VII. CONCLUSION
Here, we presented accessible dataset i.e., CK+, JAFFE and FER dataset which extensively are
utilized in here. This paper also focused on facial expression detection technique founded on pair
of machine learning algorithms and also deep learning algorithms that aid us in precise recognizing
and cataloguing of human feeling. Conferring to numerous classifiers SVM classifier provides
improved detection precision and this delivers healthier cataloguing. In FER, SVM classifier is
extra usable comparison to further classifiers to recognize expression. Neural network founded
classifier CNN provides improved precision compared to various deep neural network founded
classifiers. CNN classifier is extra usable comparison to various classifiers for improved
cataloguing.

REFERENCES
[1].A. C. Le Ngo, Y. H. Oh, R. C. W. Phan and J. See, "Eulerian emotion magnification for
subtle expression recognition," 2016 IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP), Shanghai, 2016
[2].K. M. Rajesh and M. Naveenkumar, "A robust method for face recognition and face
emotion detection system using support vector machines," 2016 International Conference
on Electrical, Electronics, Communication, Computer and Optimization Techniques
(ICEECCOT), Mysuru, 2016
[3].H. Ebine, Y. Shiga, M. Ikeda and O. Nakamura, "The recognition of facial expressions with
automatic detection of the reference face," 2000 Canadian Conference on Electrical and
Computer Engineering. Conference Proceedings. Navigating to a New Era (Cat.
No.00TH8492), Halifax, NS, 2000, pp. 1091-1099 vol.2.
[4].Zhao, X., Zhang, S., 2016. A review on facial expression recognition: feature extraction and
classification. IETE Tech. Rev. 33, 505–517. https://doi.org/
10.1080/02564602.2015.1117403.
[5].Matsumoto, D.; Yoo, S.H.; Fontaine, J. Mapping expressive differences around the world. J.
Cross Cult. Psychol. 2008, 39, 55–74. [CrossRef]
[6].Røysamb, E.; Nes, R.B.; Czajkowski, N.O.; Vassend, O. Genetics, personality and
wellbeing. A twin study of traits, facets and life satisfaction. Sci. Rep. 2018, 8, 12298.
[CrossRef]
[7].Songa, G.; Slabbinck, H.; Vermeir, I. How do implicit/explicit attitudes and emotional
reactions to sustainable logo relate? A neurophysiological study. Food Qual. Prefer. 2019,
1122
 71, 485–496. [CrossRef]
[8].Keltner, D. Born to Be Good: The Science of a Meaningful Life; WW Norton & Company:
New York, NY, USA, 2009.
[9].Lövheim, H. A new three-dimensional model for emotions and monoamine
neurotransmitters. Med. Hypotheses 2012, 78, 341–348. [CrossRef]
[10]. Wundt, W.M. Outlines of Psychology (Classic Reprint); Fb&c Limited: London,
UK, 2017; ISBN 9780331795738.
[11]. Suk, M. and B. Prabhakaran, Real-time Mobile Facial Expression Recognition
System—A Case Study, in The IEEE Conference on Computer Vision and Pattern
Recognition (CVPR) Workshops. 2014, IEEE. p. 132-137
[12]. Ryan, A., et al., Automated Facial Expression Recognition System in 43rd Annual
2009 International Carnahan Conference on Security Technologies. 2009. p. 172-177.
[13]. Golzadeh, D. R. Faria, L. J. Manso, A. Eka´rt and C. D. Buckingham, “Emotion
Recognition using Spatiotemporal Features from Facial Expression Landmarks,” 2018
International Conference on Intelligent Systems (IS), Funchal—Madeira, Portugal, 2018,
pp. 789-794, https://doi.org/10.1109/IS.2018. 8710573
[14]. A. Kartali, M. Roglić, M. Barjaktarović, M. Đurić-Jovičić, M. M. Janković, Real-
time Algorithms for Facial Emotion Recognition: A Comparison of Different Approaches,
2018 14th Symposium on Neural Networks and Applications (NEUREL), Belgrade, 2018,
pp. 1-4.
[15]. G. Yang, J. S. Y. Ortoneda and J. Saniie, “Emotion Recognition Using Deep Neural
Network with Vectorized Facial Features,” 2018 IEEE International Conference on
Electro/Information Technology (EIT), Rochester, MI, 2018, pp. 0318-0322.
[16]. Hai-Duong Nguyen, Soonja Yeom, Guee-Sang Lee, Hyung-Jeong Yang, In-Seop
Na, Soo-Hyung Kim, Facial Emotion Recognition Using an Ensemble of Multi-Level
Convolutional Neural Networks, International Journal of Pattern Recognition and Artificial
Intelligence, 33(11), 1940015, 2019 https://doi.org/ 10.1142/S0218001419400159.
[17]. Liliana, Emotion recognition from facial expression using deep convolutional neural
network, Journal of Physics: Conference Series,1193, 012004, 2019.
[18]. M. Mohammadpour, H. Khaliliardali, S. M. R. Hashemi and M. M. AlyanNezhadi,
“Facial emotion recognition using deep convolutional networks,” 2017 IEEE 4th
International Conference on Knowledge Based Engineering and Innovation (KBEI),
Tehran, 2017, pp. 0017-0021.
[19]. Ko BC. A Brief Review of Facial Emotion Recognition Based on Visual
Information. Sensors (Basel). 2018 Jan 30; 18(2):401. https://doi.org/10.3390/s18020401
PMID: 29385749
[20]. Mehta D., Siddiqui M., Javaid A. Y. (2019). Recognition of Emotion Intensities
Using Machine Learning Algorithms: A Comparative Study. Sensors (Basel, Switzerland),
19(8), 1897. https://doi.org/10.3390/ s19081897 PMID: 31010081

[21]. X. Jiang, "A facial expression recognition model based on HMM," Proceedings of
2011 International Conference on Electronic & Mechanical Engineering and Information
Technology, Harbin, Heilongjiang, China, 2011
[22]. W. Swinkels, L. Claesen, F. Xiao and H. Shen, "SVM point-based real-time emotion
detection," 2017 IEEE Conference on Dependable and Secure Computing, Taipei, 2017.
[23]. J. M. Saragih, S. Lucey and J. F. Cohn, "Real-time avatar animation from a single
image," Face and Gesture 2011, Santa Barbara, CA, USA, 2011
[24]. C. Loconsole, C. R. Miranda, G. Augusto, A. Frisoli and V. Orvalho, "Real-time
emotion recognition novel method for geometrical facial features extraction," 2014
International Conference on Computer Vision Theory and Applications (VISAPP), Lisbon,
Portugal, 2014
[25]. Neerja and E. Walia, "Face Recognition Using Improved Fast PCA Algorithm,"
1123
 2008 Congress on Image and Signal Processing, Sanya, Hainan, 2008
[26]. K. M. Rajesh and M. Naveenkumar, "A robust method for face recognition and face
emotion detection system using support vector machines," 2016 International Conference
on Electrical, Electronics, Communication, Computer and Optimization Techniques
(ICEECCOT), Mysuru, 2016 50
[27]. B. Fasel, “Robust face analysis using convolutional neural networks,” in Pattern
Recognition, 2002. Proceedings. 16th International Conference on, vol. 2. IEEE, 2002, pp.
40–43.
[28]. ——, “Head-pose invariant facial expression recognition using convo lutional neural
networks,” in Proceedings of the 4th IEEE International Conference on Multimodal
Interfaces. IEEE Computer Society, 2002. p. 529.
[29]. M. Matsugu, K. Mori, Y. Mitari, and Y. Kaneda, “Subject independent facial
expression recognition with robust face detection using a convolu tional neural network,”
Neural Networks, vol. 16, no. 5-6, pp. 555–559, 2003.
[30]. B. Sun, L. Li, G. Zhou, X. Wu, J. He, L. Yu, D. Li, and Q. Wei, “Combining
multimodal features within a fusion network for emotion recognition in the wild,” in
Proceedings of the 2015 ACM on Interna tional Conference on Multimodal Interaction.
ACM, 2015, pp. 497–502.
[31]. B. Sun, L. Li, G. Zhou, and J. He, “Facial expression recognition in the wild based
on multimodal texture features,” Journal of Electronic Imaging, vol. 25, no. 6, p. 061407,
2016.
[32]. R. Girshick, J. Donahue, T. Darrell, and J. Malik, “Rich feature hierarchies for
accurate object detection and semantic segmentation,” in Proceedings of the IEEE
conference on computer vision and pattern recognition, 2014, pp. 580–587.
[33]. J. Li, D. Zhang, J. Zhang, J. Zhang, T. Li, Y. Xia, Q. Yan, and L. Xun, “Facial
expression recognition with faster r-cnn,” Procedia Computer Science, vol. 107, pp. 135–
140, 2017.
[34]. S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object
detection with region proposal networks,” in Advances in neural information processing
systems, 2015, pp. 91–99.
[35]. S. Ji, W. Xu, M. Yang, and K. Yu, “3d convolutional neural networks for human
action recognition,” IEEE transactions on pattern analysis and machine intelligence, vol. 35,
no. 1, pp. 221–231, 2013.
[36]. D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri, “Learning
spatiotemporal features with 3d convolutional networks,” in Computer Vision (ICCV), 2015
IEEE International Conference on. IEEE, 2015, pp. 4489–4497.
[37]. G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm for deep belief
nets,” Neural computation, vol. 18, no. 7, pp. 1527–1554, 2006.
[38]. G. E. Hinton and T. J. Sejnowski, “Learning and releaming in boltz mann
machines,” Parallel distributed processing: Explorations in the microstructure of cognition,
vol. 1, no. 282-317, p. 2, 1986.
[39]. G. E. Hinton, “A practical guide to training restricted Boltzmann machines,” in
Neural networks: Tricks of the trade. Springer, 2012, pp. 599–619.
[40]. Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle, “Greedy layer wise training
of deep networks,” Advances in neural information processing systems, 2007, pp. 153–160.
[41]. G. E. Hinton, “Training products of experts by minimizing contrastive divergence,”
Neural computation, vol. 14, no. 8, pp. 1771–1800, 2002.
[42]. G. E. Hinton, S. Osindero, and Y. W. Teh, “A fast learning algorithm for deep belief
nets”, Neural Computation, Vol. 18, No. 7, pp. 1527-1554, 2006.
[43]. G. E. Hinton, “Deep belief networks” Scholarpedia, Vol. 4, No. 5, p. 5947, 2009
[44]. F. A. Gers, N. N. Schraudolph, and J. Schmidhuber, “Learning precise timing with
lstm recurrent networks,” Journal of machine learning research, vol. 3, no. Aug, pp. 115–
1124
143, 2002.

1125