Багатошарова МГУА-нейро-фаззі мережа на основі розширених нечітких нейронів та її використання для онлайн разпізнавання виразів обличчя
Real-time image recognition is required in many important practical problems. Interaction with users in online mode requires flexibility and adaptability from applications. The Group Method of Data Handling (GMDH) allows changing the model structure and adjusting the system architecture to the chara...
Збережено в:
| Дата: | 2020 |
|---|---|
| Автори: | , , , |
| Формат: | Стаття |
| Мова: | Англійська |
| Опубліковано: |
The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute"
2020
|
| Теми: | |
| Онлайн доступ: | https://journal.iasa.kpi.ua/article/view/221271 |
| Теги: |
Додати тег
Немає тегів, Будьте першим, хто поставить тег для цього запису!
|
| Назва журналу: | System research and information technologies |
| Завантажити файл: |  |
Репозитарії
System research and information technologies| _version_ | 1866302708034568192 |
|---|---|
| author | Bodyanskiy, Yevgeniy V. Zaychenko, Yuriy P. Hamidov, Galib Kulishova, Nonna Ye. |
| author_facet | Bodyanskiy, Yevgeniy V. Zaychenko, Yuriy P. Hamidov, Galib Kulishova, Nonna Ye. |
| author_sort | Bodyanskiy, Yevgeniy V. |
| baseUrl_str | http://journal.iasa.kpi.ua/oai |
| collection | OJS |
| datestamp_date | 2021-01-19T12:18:25Z |
| description | Real-time image recognition is required in many important practical problems. Interaction with users in online mode requires flexibility and adaptability from applications. The Group Method of Data Handling (GMDH) allows changing the model structure and adjusting the system architecture to the characteristics of each task under consideration. Moreover, the approximating properties of neo-fuzzy neurons used as elements of the system provide the high recognition accuracy under conditions of short data samples. This paper proposes a multilayer GMDH-neuro-fuzzy network based on extended neo-fuzzy neurons. The learning algorithm has filtering and tracking properties, guarantees the required speed important for real-time applications. The effectiveness of the proposed system is confirmed for the human emotions recognition. |
| doi_str_mv | 10.20535/SRIT.2308-8893.2020.3.05 |
| first_indexed | 2025-07-17T10:26:59Z |
| format | Article |
| fulltext |
Ye. Bodyanskiy, Yu. Zaychenko, G. Hamidov, N. Kulishova, 2020
66 ISSN 1681–6048 System Research & Information Technologies, 2020, № 3
TIДC
ТЕОРЕТИЧНІ ТА ПРИКЛАДНІ ПРОБЛЕМИ
ІНТЕЛЕКТУАЛЬНИХ СИСТЕМ ПІДТРИМАННЯ
ПРИЙНЯТТЯ РІШЕНЬ
UDC 519.8
DOI: 10.20535/SRIT.2308-8893.2020.3.05
MULTILAYER GMDH-NEURO-FUZZY NETWORK BASED ON
EXTENDED NEO-FUZZY NEURONS AND ITS APPLICATION
IN ONLINE FACIAL EXPRESSION RECOGNITION
Ye. BODYANSKIY, Yu. ZAYCHENKO, G. HAMIDOV, N. KULISHOVA
Abstract. Real-time image recognition is required in many important practical
problems. Interaction with users in online mode requires flexibility and adaptability
from applications. The Group Method of Data Handling (GMDH) allows changing
the model structure and adjusting the system architecture to the characteristics of
each task under consideration. Moreover, the approximating properties of neo-fuzzy
neurons used as elements of the system provide the high recognition accuracy under
conditions of short data samples. This paper proposes a multilayer GMDH-neuro-
fuzzy network based on extended neo-fuzzy neurons. The learning algorithm has fil-
tering and tracking properties, guarantees the required speed important for real-time
applications. The effectiveness of the proposed system is confirmed for the human
emotions recognition.
Keywords: Group Method of Data Handling, extended neo-fuzzy neuron, online
image recognition, facial expression recognition.
INTRODUCTION
Information technologies are actively being introduced into education, business,
healthcare, entertainment and other spheres of human life. This requires tech-
nology interactivity, to conduct continuous two-way cooperation between person
and computer or mobile device. One of the promising areas for such intellectual
interfaces’ development is the approach that uses the recognition of people, their
age, sex, state of health, emotional status on the real time video. This complex
technical problem already finds its own solutions [1–9]. Frequently, these deci-
sions use the machine learning and neuro-fuzzy approach.
As a technical problem, the task of a user emotional status recognition by
video is reduced to characteristic features detecting, and to the collected data clas-
sification. This problem is related to the fact that machine learning algorithms in
this task require the training data sets in which the samples number can be tens or
even hundreds of thousands. The forming of such sets is a serious, time-
consuming task, significantly increasing the projects developing cost and imple-
mentation duration.
The prospective methods of recognition by short datasets are fuzzy systems
and GMDH. Earlier it was proved that neural networks are universal approxima-
Multilayer GMDH-neuro-fuzzy network based on extended neo-fuzzy neurons …
Системні дослідження та інформаційні технології, 2020, № 3 67
tors and have some remarkable properties, such as parallel information processing,
ability to work with incomplete noisy input data and learning possibilities to
achieve the desired output.
The GMDH, from the other side, uses the principle of self-organization that
allows constructing an optimal structure of the mathematical model during the
algorithm operation. It’s very promising to combine advantages of these both
approaches for the solution of the problem — development an efficient model
structure. GMDH-neural networks whose nodes are active neurons [10–12],
N-adalines [13], R-neurons [14–16], Q-neurons [17] are known. At the junction
of the fuzzy GMDH [18] and neural networks, the GMDH-neuro-fuzzy system
[15, 19] and the GMDH-neo-fuzzy system [20] were created. These systems have
proven their effectiveness in solving a wide range of problems, but have lost the
main advantages of the original GMDH: a small number of evaluated parameters
in each node. In this regard, it seems promising to develop a GMDH-neo-fuzzy
system that combines the advantages of traditional GMDH and hybrid computa-
tional intelligence systems, and is trained using simple procedures to ensure high
speed of online image recognition.
The goal of the present paper is a synthesis of the GMDH neo-fuzzy system
for the online image recognition.
THE EXTENDED NEO-FUZZY NEURON AS A NODE OF GMDH-NEURO-
FUZZY SYSTEM
Takeshi Yamakawa and co-authors in [21–23] proposed the architecture of neo-
fuzzy neuron (NFN). The authors of the NFN admit among its most important
advantages, the high learning rate, computational simplicity, the possibility to find
the learning criterion global minima in real-time processing. Besides, NFN is
characterized by fuzzy linguistic “if-then” rules. The neo-fuzzy neuron is a non-
linear multi-input single-output system shown in Fig. 1.
It realizes the following mapping
h
j
ijijiii xwxf
1
)()(
and implements fuzzy inference
jijii wxx IS OUTPUT THE THEN IS IF ,
where jix is a fuzzy set with membership function )( iji x , jiw is a singleton
synaptic weight in consequent [2]. As it can be seen nonlinear synapse in fact
actualize Takagi–Sugeno fuzzy inference of zero order. The membership
functions )( iji x in the antecedent could be B-splines or triangular functions, for
example, like this
,otherwise ,0
;],[ if ,
;],[ if ,
,1,
,,1
,1
,,1
,1,
,1
ijiji
ijij
iij
ijiji
ijij
iji
ji ccx
cc
xc
ccx
cc
cx
Ye. Bodyanskiy, Yu. Zaychenko, G. Hamidov, N. Kulishova
ISSN 1681–6048 System Research & Information Technologies, 2020, № 3 68
where jic — the centers of membership functions, usually distributed on interval
[0, 1]. This contributes to simplify the fuzzy inference. An input signal ix
activates only two neighboring membership functions simultaneously and the sum
of the grades equals to unity, providing Ruspini partition:
1)()()()( ,1,1 iijijiijiiij xxxx .
The inference result can be produced by arbitrary defuzzyfication method.
Center-of-Gravity method gives output in the simple form:
)()()( ,1,1 iijijijijiii xwxwxf .
It is possible to improve approximating properties of such a system by using
a structural unit, called by authors as “extended nonlinear synapse” (ENSi) (Fig. 2)
and synthesized on its basis the “extended neo-fuzzy neuron” [24–27] (ENFN).
ENFN contains ENSi as elements instead of usual nonlinear synapses NSi.
xi(k)
fi(xi(k))
p
i
p
iiiiii xwxwxww 1
22
1
1
1
0
1
p
i
p
iiiiii xwxwxww 2
22
2
1
2
0
2
p
i
p
hiihiihihi xwxwxww 2210
φ1i(xi(k)
)
φ2i(xi(k)
)
φhi(xi(k)
)
ENSiс
Fig. 2. Extended non-linear synapse
w11
w21
w12
w22
wh2
whn
w1n
w2n
wh1
x1(k)
x2(k)
xn(k)
f1(x1(k))
f2(x2(k))
fn(xn(k))
y(k))
NS1
NS2
NSn
Fig. 1. Neo-fuzzy neuron
Multilayer GMDH-neuro-fuzzy network based on extended neo-fuzzy neurons …
Системні дослідження та інформаційні технології, 2020, № 3 69
By introducing the additional variables:
)...()()( 2210 p
i
p
liiliililiiliili xwxwxwwxxy ;
h
l
p
i
p
liiliililiiliii xwxwxwwxxf
1
2210 )...()()(
)(...)()( 111
1
11
0
1 ii
p
i
p
iiiiiiii xxwxxwxw
)(...)(...)( 222
0
2 ihi
p
i
p
hiii
p
i
p
iiii xxwxxwxw ;
T
2
0
21
1
1
0
1 ),...,,...,,,...,,( p
hi
p
ii
p
iiii wwwwwww ;
,))(),...,(),...,(),(),...,(),(()(~ T
22111 ihi
p
iii
p
iiiii
p
iiiiiiii xxxxxxxxxxx
we can write
),(~)( T
iiiii xwxf
,)(~~)(~)(ˆ
1
T
1
xwxwxfy T
n
i
ii
n
i
ii
where TTTT
1
T ),...,,...,(~
ni wwww ;
TTT
1
T
1 )~),...,(~,...,~()(~
nnii xxxx .
It’s easy to see that ENFN contains 1p hn adjusting synaptic weights and
fuzzy output, implemented by each ENSi, has the form:
IF ix IS lix THEN THE OUTPUT IS hlxwxww p
i
p
liilili ,...,2,1,...10 ,
i.e. essentially coincides with p-order Takagi–Sugeno inference.
Fig. 3 shows the architecture of an extended neo-fuzzy neuron.
THE NEO-FUZZY NEURON LEARNING ALGORITHM
As a goal function for NFN learning the local quadratic error function is used:
2
1 1
22 ))((
2
1
)(
2
1
)(ˆ)(
2
1
)(
n
i
h
j
ijiji kxwkykekykykE ,
Fig. 3. Extended neo-fuzzy neuron
fn(xn(k)
f2(x2(k))
f1(x1(k))
xn(k)
x2(k)
x1(k) ENS1
ENS2
ENSn
y(k)
Ye. Bodyanskiy, Yu. Zaychenko, G. Hamidov, N. Kulishova
ISSN 1681–6048 System Research & Information Technologies, 2020, № 3 70
where )(ky — external reference signal.
It is minimized in the gradient stepwise algorithm in the form:
))(()()1()( kxkekwkw ijijiji
))(())(()1()()1(
1 1
kxkxkwkykw iji
n
i
h
j
ijijiji
,
)(ke — learning error; — the scalar learning rate parameter.
To increase the speed of training process Kaczmarz–Widrow–Hoff one-step
learning algorithm [28–32] can be used:
))((
))((
))(()1()(
)1()(
2
T
kx
kx
kxkwky
kwkw
,
where )),...(()),...,(()),...,((())(( 2211111 kxkxkxkx hh
T)))(()),...,((..., kxkx nhniji ,
...),1(,...,)1(()1( 111 kwkwkw h
T
2 ))1(),...,1(),...,1(,... kwkwkw hnjih —
1nh — vectors generated by input variables.
The learning algorithm exponentially weighted form:
,10,))(()1()(
;))(()))(()1()()(()1()(
2
1
kxkrkr
kxkxkwkykrkwkw T
which has filtering and tracking properties can be effectively used in stochastic
and nonstationary situation.
THE NEURO-FUZZY SYSTEM AND ITS ARCHITECTURE OPTIMIZATION
USING THE GROUP METHOD OF DATA HANDLING
The neuro-fuzzy system under consideration is a multilayer feedforward architec-
ture that consists of extended neo-fuzzy neurons and shown on Fig. 4.
A )1( n -dimensional input signals vector
T
21 ),...,,( nxxxx arrives at
system zero receptive layer, and is transmitted then to first hidden layer containing
2
1 nn C neuron nodes with only two inputs. At the outputs of first hidden layer
nodes ]1[N , output signals 2]1[ )1(
2
,...,2,1,ˆ nl Cn
n
ly are formed. Further,
these signals are sent to selection block ]1[SB of first hidden layer, which selects
from output signals set ]1[ˆly the most accurate )( 1
*
1
*
1 nnn of them in the ac-
cepted criterion sense, most often the mean square error 2
]1[
ly
. From these *
1n
best outputs of first hidden layer
*]1[ˆly , 2n ( 2 2n n n usually) pairwise combi-
Multilayer GMDH-neuro-fuzzy network based on extended neo-fuzzy neurons …
Системні дослідження та інформаційні технології, 2020, № 3 71
nations are formed, which are fed to second hidden layer formed by neurons
]2[N similar to neurons ]1[N . From output signals of this layer
]2[ˆly , the selec-
tion block of the second hidden layer ]2[SB selects only those that exceed the
best signal of the first hidden layer
*]1[ˆly , for example, in terms of accuracy 2
]2[
ly
.
The third hidden layer generates signals that are superior in accuracy to the
best signal *]2[ˆly etc. The process of system forming occurs until the selection
block ]1[ sSB generates at its output only two signals *]1[
1ˆ
sy and *]1[
2ˆ
sy . These
signals that are applied to a single output neuron ][sN are used to calculate the
single system output signal ][ˆ sy .
As the nodes of the GMDH system, one can use various neurons types with
necessary approximating capabilities. However, at the same time, the main advan-
tage of the original GMDH method — the ability to real-time work in the pres-
ence of short training sets — may be lost.
This paper proposes to use the extended multidimensional neo-fuzzy neurons
[24–27] as GMDH system nodes. Its architecture is a Takagi–Sugeno–Kang neu-
ro-fuzzy system [33–35] with two inputs 1x and 2x , five sequentially connected
information processing layers and one output ˆly . A two-dimensional vector of
input signals T
21 ))(),(()( kxkxkx to be processed is fed to the input of the node
( 1,2,...,k N — the observation number in the training set or the current discrete
time). The node first layer contains 2h membership functions )(),( jpjipi xx ,
1,2,...,p h and implements fuzzification of input variables. The second layer
provides aggregation of membership levels calculated in the first layer, contains h
multiplication blocks and forms two-dimensional radial basis activation functions
)(),( jpjipi xx . The third layer is a layer of synaptic weights to be adjusted
during the training process, while layer outputs are values )(),( jpjipi
ij
lp xxw ,
and weights number is determined by number of membership functions at each
input h. The fourth layer is formed by two adders and calculates the sum of second
and third layers output signals. Finally, in the fifth neuron output layer normalization
is performed, as a result of which the node output signal lŷ is calculated.
N[2]N[1]
N[1]
N[1]
SB
[1
]
]1[
1ŷ
]1[
2ŷ
]1[
1
ˆ
ny
x1
x2
xn
N[2]
N[2]
SB
[2
]
]2[
2ŷ
]2[
2ŷ
]2[
2
ˆ
ny
*]1[
1ŷ
*]1[
2ŷ
*]1[
1
ˆny
*]2[
2ŷ
*]2[
2ŷ
*]2[
2
ˆ
ny
N[s]
][ˆ sy
*]1[ˆ sy
*]1[ˆ sy
SB
[s
–
1]
Fig. 4. GMDH-neuro-fuzzy system based on extended neo-fuzzy neurons
Ye. Bodyanskiy, Yu. Zaychenko, G. Hamidov, N. Kulishova
ISSN 1681–6048 System Research & Information Technologies, 2020, № 3 72
Thus, when a signal )( kx is fed to the neuron node input, elements of the
first layer calculate membership levels 1)(0,1)(0 jpjipi xx . As
membership functions, bell-shaped constructions with a not strictly local receptor
field are usually used. They avoid the appearance of “holes” in the fuzzified space
when using the scattered partition spaces of input variables [35]. The membership
functions of the first layer are Gaussian
2
2
2
))((
exp))((
i
pii
ipi
ckx
kx ,
2
2
2
))((
exp)(
j
pjj
jpj
ckx
x ,
where pic , pjc — parameters defining the centres of membership functions;
i , j — width parameters of these functions.
The h aggregated signals appear on second layer outputs
))(())(()(~ kxkxkx jpjipip ,
where
2
2
2
2
2
))((
exp
2
))((
exp)(~
j
pjj
i
pii
p
ckxckx
kx
2
2
2
)(
exp
pckx
, T),( pjpip ccc .
The third layer outputs are the values
)(~))(())(( kxwkxkxw p
ij
lpjpjipi
ij
lp ,
the fourth layer calculates the following two sums:
h
p
h
p
p
ij
lpjpjipi
ij
lp kxwkxkxw
1 1
)(~))(())(( ;
h
p
h
p
pjpjipi kxkxkx
1 1
)(~))(())((
and finally, at node output (fifth layer), a signal is formed
h
p
p
h
p
p
ij
lp
h
p
jpjipi
h
p
jpjipi
ij
lp
l
kx
kxw
kxkx
kxkxw
ky
1
1
1
1
)(~
)(~
))(())((
))(())((
)(ˆ
))(()())((
)(~
)(~
T
11
1
kxwkxw
kx
kx
w ijij
l
h
p
ij
p
ij
lp
h
p
h
p
p
pij
lp
,
where
1
1
))(())(())(())(())(((
h
p
jpjipijpjipi
ij
p kxkxkxkxkx ;
Multilayer GMDH-neuro-fuzzy network based on extended neo-fuzzy neurons …
Системні дослідження та інформаційні технології, 2020, № 3 73
T
21 ),...,,( ij
lh
ij
l
ij
l
ij
l wwww ;
T
21 )))(()),...,(()),((())(( kxkxkxkx ij
p
ijijij .
It can be noticed that the node implements a non-linear mapping of input
signals to output.
THE EXPERIMENTS
The task of a person's facial expression recognition is complex and multi-stage. It
includes pre-processing of the image and searching for the face area within the
image. After the face area is distinguished, it is possible to recognize the emotion
by the face features set. In the practice of faces expressions recognition, several
descriptor principles are used. The most common are adaptive appearance
models, which use the descriptions based on face image feature points and con-
tours. It is established that such descriptions convey complete information about
the person emotional state, even if it is expressed weakly.
Under the emotions influence, the facial muscles reduction leads to the dis-
placement of feature points and this movement can serve as an indicator of basic
facial actions. The most commonly used facial expressions are some basic emo-
tions (fear, sadness, happiness, anger, disgust, surprise) and neutral state.
As a base for the feature vector, it is proposed to use a set of 35 characteris-
tic points that can be localized in the facial area using contour detectors (Fig. 5).
The neo-fuzzy neurons has one output as the dimensionality of the output data
vector. Seven basic emotions were selected for recognition: anger, disgust, fear,
surprise, happiness, sadness, neutral expression. Therefore, the output values are
{1; 2; 3; 4; 5; 6; 7}. The character features vector contains the two-dimensional
coordinates of feature points position (fig. 5). So, the system input is vector
701}{ ix . The order of the polynomial in nonlinear synapse membership function
was chosen equal to 4, the number of synapses in the neo-fuzzy neuron was
equal to 5.
The proposed architecture ability to recognize individual emotions was in-
vestigated using photographs from two open bases — Psychological Image Col-
lection at Stirling (PICS) [36], partly from the Extended Cohn-Kanade (CK +)
database [37]. Some images are in public use as objects for recognition.
In this set of photographs, those were selected differ in the person emotional
state expression degree — from weakly noticeable to very noticeable.
Fig. 5. Examples of training images and position of characteristic points
Ye. Bodyanskiy, Yu. Zaychenko, G. Hamidov, N. Kulishova
ISSN 1681–6048 System Research & Information Technologies, 2020, № 3 74
In this task, special attention was paid to a learning data set small size. To
examine how the proposed architecture and learning algorithm will recognize fa-
cial expressions, small photo sets are used. Their dimensions are given in Table 1.
T a b l e 1 . Dimensions of training sets of photos for individual emotions
Emotion Anger Disgust Fear Happiness Sorrow Surprise Neutral
Data
set size
49 66 35 45 19 50 80
Then the architecture ability to learn from a mixed set was examined, and
sets total size was 344 photos. The number of unrecognized emotions is given
in Table 2.
T a b l e 2 . The number of unrecognized emotions as a result of GMDH-neuro-
fuzzy system based on extended neo-fuzzy neurons learning from a mixed set
Primary emotions Recognition
accuracy Anger Disgust Fear Happiness Sorrow Surprise Neutral
The percentage
of unrecognized
images,%
2,04 0 5,71 4,44 0 0 2,5
A GMDH-neuro-fuzzy system based on extended neo-fuzzy neurons con-
figured a task model from two rows. Learning error change is shown in Fig. 6.
CONCLUSIONS
The article proposes the GMDH system with extended neo-fuzzy neurons as
nodes. The system architecture allows modifying the process model structure in
real time due to proposed neo-fuzzy nodes-neurons synaptic weights adjusting. A
feature of the proposed architecture and its learning algorithm is the ability to
work with small training samples.
REFERENCES
1. A. Kołakowska, A. Landowska, M. Szwoch, W. Szwoch, and M.R. Wrobel, “Hu-
man-Computer Systems Interaction: Backgrounds and Applications”, ch. 3, Emotion
Recognition and Its Applications. Cham: Springer International Publishing, 2014,
pp. 1–62.
1 2 3 4 5 6 7 8 9 10
0
-0,5
1
-1,5
L
ta
rn
in
g
er
ro
r
Learning epochs number
The system second row error 10–3
1 2 3 4 5 6 7
4
3
2
1
0
L
ta
rn
in
g
er
ro
r
Learning epochs number
The system first row error 10–3
Fig. 6. A GMDH-neuro-fuzzy system based on extended neo-fuzzy neurons learning error
Multilayer GMDH-neuro-fuzzy network based on extended neo-fuzzy neurons …
Системні дослідження та інформаційні технології, 2020, № 3 75
2. Kaggle. Challenges in representation learning: Facial recognition challenge, 2013.
3. G.U. Kharat and S.V. Dudul, “Emotion Recognition from Facial Expression Using
Neural Networks”, in Human-Computer Systems Interaction. Advances in Intelligent
and Soft Computing, vol. 60, Z.S. Hippe, J.L. Kulikowski, Eds. Berlin, Heidelberg:
Springer, 2009.
4. C. Shan, S. Gong, and P.W. McOwan, “Facial expression recognition based on local
binary patterns: A comprehensive study,” Image and Vision Computing, vol. 27,
no. 6, pp. 803–816, 2009.
5. Ch.-Yi. Lee and Li-Ch. Liao, “Recognition of Facial Expression by Using Neural-
Network System with Fuzzified Characteristic Distances Weights”, IEEE Int. Conf.
Fuzzy Systems FUZZ-IEEE 2008 [IEEE World Congress on Computational Intelli-
gence], pp. 1694–1699, 2008.
6. N. Kulishova, “Emotion Recognition Using Sigma-Pi Neural Network”, Proc. of
2016 IEEE First International Conference on Data Stream Mining & Processing
(DSMP), Lviv, 2016, pp. 327–331.
7. A. Graves, J. Schmidhuber, C. Mayer, M. Wimmer, and B. Radig, “Facial Expres-
sion Recognition with Recurrent Neural Networks”, International Workshop on
Cognition for Technical Systems, Munich, Germany, October 2008.
8. S. Ouelett, “Real-time emotion recognition for gaming using deep convolutional
network features”, CoRR, vol. abs./1408.3750, 2014.
9. B. Kim, J. Roh, S. Dong, and S. Lee, “Hierarchical committee of deep convolutional
neural networks for robust facial expression recognition”, Journal on Multimodal
User Interfaces, pp. 1–17, 2016.
10. A.G. Ivakhnenko, G.A. Ivakhnenko, and J.A. Mueller, “Self-organization of the neu-
ral networks with active neurons”, Pattern Recognition and Image Analysis, vol. 4,
no. 2, pp. 177–18, 1994.
11. G.A. Ivakhnenko, “Self-organization of neuronet with active neurons for effect of
nuclear test explosion forecasting”, System Analysis Modeling Simulation, vol. 20,
pp. 107–116, 1995.
12. A.G. Ivakhnenko, D. Wuensch, and G.A. Ivakhnenko, “Inductive sorting-out
GMDH algorithms with polynomial complexity for active neurons of neural net-
works”, Neural Networks, vol. 2, pp. 1169–1173, 1999.
13. D.U. Pham and X. Liu, Neural networks for Identification, Prediction and Control.
London: Springer-Verlag, 1995, 238 p.
14. E. Lughofer, Evolving Fuzzy Systems – Methodologies, Advanced Concepts and Ap-
plications, Berlin-Heidelberg, Springer-Verlag, 2011.
15. T. Ohtani, “Automatic variable selection in RBF network and its application to neu-
rofuzzy GMDH”, Proc. Fourth Int. Conf. on Knowledge-Based Intelligent Engineer-
ing Systems and Allied Technologies, vol. 2, pp. 840–843, 2000.
16. Ye. Bodyanskiy, N. Teslenko, and P. Grimm, “Hybrid evolving neural network us-
ing kernel activation functions”, Proc. 17th Zittau East West Fuzzy Colloquium, Zit-
tau/Goerlitz: HS, 2010, pp. 39–46.
17. Ye. Bodyanskiy, O. Vynokurova, and I. Pliss, “Hybrid GMDH-neural network of
computational intelligence”, Proc. 3rd Int. Workshop on Inductive Modeling,
Krynica, Poland, 2009, pp. 100–107.
18. Yu. Zaychenko, “The fuzzy Group Method of Data Handling and its application for
economical processes forecasting”, Scientific Inquiry, vol. 7, no.1, pp. 83–96, 2006.
19. T. Ohtani, H. Ichihashi, T. Miyoshi, K. Nagasaka, and Y. Kanaumi, “Structural
learning of neurofuzzy GMDH with Minkowski norm”, Proc. 1998 Second Int. Conf.
on Knowledge-Based Intelligent Electronic Systems, vol. 2, pp. 100–107, 1998.
20. Ye. Bodyanskiy, Yu. Zaychenko, E. Pavlikovskaya, M. Samarina, and Ye. Viktorov,
“The neo-fuzzy neural network structure optimization using GMDH fog the solving
Ye. Bodyanskiy, Yu. Zaychenko, G. Hamidov, N. Kulishova
ISSN 1681–6048 System Research & Information Technologies, 2020, № 3 76
forecasting and classification problems”, Proc. Int. Workshop on Inductive Model-
ing, Krynica, Poland, 2009, pp. 77–89.
21. J. Miki and T. Yamakawa, “Analog implementation of neo-fuzzy neuron and its on-
board learning”, in Computational Intelligence and Applications, Ed. N.E. Mas-
torakis, Piraeus: WSES Press, 1999, pp. 144–149.
22. T. Yamakawa, E. Uchino, J. Miki and H. Kusanagi, “A neo-fuzzy neuron and its ap-
plication to system identification and prediction of the system behavior”, Proc. 2-nd
Int. Conf. on Fuzzy Logic and Neural Networks “IIZUKA-92”, Iizuka, Japan, 1992,
pp. 477–483.
23. E. Uchino and T. Yamakawa, “Soft computing based signal prediction, restoration
and filtering”, in Intelligent Hybrid Systems: Fuzzy Logic, Neural Networks and
Genetic Algoritms, Ed. Da Ruan, Boston: Kluwer Academic Publishers, 1997,
pp. 331–349.
24. Ye.V. Bodyanskiy and N.Ye. Kulishova, “Extended neo-fuzzy neuron in the task
of images filtering”, Radioelectronics. Computer Science. Control, no. 1(32),
pp. 112–119, 2014.
25. Z. Hu, Ye. Bodyanskiy, N. Kulishova, O.A. Tyshchenko, “Multidimensional Ex-
tended Neo-Fuzzy Neuron for Facial Expression Recognition”, International Journal
of Intelligent Systems and Applications (IJISA), vol. 9, no. 9, pp. 29–36, 2017.
26. Ye. Bodyanskiy, N. Kulishova and D. Malysheva, “The Extended Neo-Fuzzy Sys-
tem of Computational Intelligence and its Fast Learning for Emotions Online Rec-
ognition”, Proc. of the 2018 IEEE Second International Conference on Data Stream
Mining & Processing (DSMP), Lviv, Ukraine, August 21–25, 2018, pp.473 –478.
27. Ye. Bodyanskiy, N. Kulishova and O. Chala, “The Extended Multidimensional Neo-
Fuzzy System and its Fast Learning in Pattern Recognition Tasks”, Data, no. 3,
pp. 63–73, 2018.
28. S. Kaczmarz, “Angenaeherte Ausloesung von Systemen Linearer Gleichungen”,
Bull. Int. Acad. Polon. Sci, Let. A, pp. 355–357, 1937.
29. S. Kaczmarz, “Approximate solution of systems of linear equations, Int. J. Control,
vol. 53, pp. 1269–1271, 1993.
30. Ye. Bodyanskiy, I. Kokshenev and V. Kolodyazhniy, “An adaptive learning algo-
rithm for a neo-fuzzy neuron”, Proc. of the 3rd Conference of the European Society
for Fuzzy Logic and Technology, pp. 375–379, 2005.
31. V. Kolodyazhniy, Ye. Bodyanskiy and P. Otto, “Universal approximator employing
neo-fuzzy neurons”, Computational Intelligence: Theory and Applications, Ed. By
B. Reusch, Berlin-Heidelberg: Springer, 2005, pp. 631–640.
32. B. Widrow and Jr.M.E. Hoff, “Adaptive switching circuits”, 1960 URE WESCON
Convention Record, part 4. N.-Y.: IRE, 1960, pp. 96–104.
33. L.-X. Wang and J.M. Mendel, “Fuzzy basis functions, universal approximation and
orthogonal least-squares learning”, IEEE Trans. on Neural Networks, vol. 3, no. 5,
pp. 807–814, 1992.
34. L.-X. Wang, Adaptive Fuzzy Systems and Control. Design and Statistical Analysis.
Upper Saddle River: Prentice Hall, 1994, 256 p.
35. R. J.-S. Jang, C.-T. Sun, and E. Mizutani, Neuro-Fuzzy and Soft Computing: A Com-
putational Approach to Learning and Machine Intelligence. Upper Saddle River:
Prentice Hall, 1997, 640 p.
36. 2D face sets. Available: http://pics.psych.stir.ac.uk/2D_face_sets.htm
37. P. Lucey, J.F. Cohn, T. Kanade, J. Saragih, Z. Ambadar, and I. Matthews, “The Ex-
tended Cohn-Kanade Dataset (CK+): A complete dataset for action unit and emo-
tion-specified expression”, Proceedings of IEEE workshop on CVPR for Human
Communicative Behavior Analysis, San Francisco, USA, 2010.
Received 19.10.2020
Multilayer GMDH-neuro-fuzzy network based on extended neo-fuzzy neurons …
Системні дослідження та інформаційні технології, 2020, № 3 77
INFORMATION ON THE ARTICLE
Yevgeniy V. Bodyanskiy, ORCID: 0000-0001-5418-2143, Kharkiv National University
of Radioelectronics, Ukraine, e-mail: yevgeniy.bodyanskiy@nure.ua
Yuriy P. Zaychenko, ORCID: 0000-0001-9662-3269, Educational and Scientific
Complex “Institute for Applied System Analysis” of the National Technical University
of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”, Ukraine, e-mail:
zaychenkoyuri@ukr.net
Galib Hamidov, head of information technologies department of “Azershig”, Azerbaijan,
e-mail: galib.hamidov@gmail.com
Nonna Ye. Kulishova, ORCID: 0000-0001-7921-3110, Kharkiv National University of
Radioelectronics, Ukraine, e-mail: nokuliaux@gmail.com
БАГАТОШАРОВА МГУА-НЕЙРО-ФАЗЗІ МЕРЕЖА НА ОСНОВІ РОЗШИРЕНИХ
НЕЧІТКИХ НЕЙРОНІВ ТА ЇЇ ВИКОРИСТАННЯ ДЛЯ ОНЛАЙН РАЗПІЗНАВАННЯ
ВИРАЗІВ ОБЛИЧЧЯ / Є.В. Бодянський, Ю.П. Зайченко, Г. Гамідов, Н. Є. Кулішова
Анотація. Розпізнавання зображень в реальному часі потрібно в багатьох
практичних задачах. Взаємодія з користувачами в режимі онлайн потребує
гнучкості та адаптацїї від прикладних програм. Метод групового урахування
аргументів (МГУА) дозволяє змінювати структуру моделі та налаштовує її ар-
хітектуру відповідно до характеристик кожної задачі. Більш того, апроксимацій-
ні властивості нео-фазі нейронів як структурних елементів системи забезпе-
чують високу точність розпізнавання в умовах коротких вибірок даних.
Запропоновано багатошарову МГУА-нейро-фаззі мережу на основі розшире-
них нео-фаззі нейронів. Алгоритм навчання має фільтрувальні та відслідкову-
вальні властивості та гарантує необхідну швидкість для застосувань реального
часу. Ефективність запропонованої системи підтверджено в задачі розпізна-
вання людських емоцій.
Ключові слова: метод групового урахування аргументів, розширений нео-
фаззі нейрон, онлайн розпізнавання зображень, розпізнавання виразів обличчя.
МНОГОСЛОЙНАЯ МГУА-НЕЙРО-ФАЗЗИ СЕТЬ НА ОСНОВЕ РАСШИРЕННЫХ
НЕЧЕТКИХ НЕЙРОНОВ И ЕЕ ПРИМЕНЕНИЕ ДЛЯ ОНЛАЙН РАСПОЗНАВАНИЯ
ВЫРАЖЕНИЙ ЛИЦА / Е.В. Бодянский, Ю.П. Зайченко, Г. Гамидов, Н.Е. Кулишова
Аннотация. Распознавание изображений в реальном времени требуется в мно-
гих практических задачах. Взаимодействие с пользователями в режиме он-
лайн требует гибкости и адаптации от прикладных программ. Метод группо-
вого учета аргументов (МГУА) позволяет изменять структуру модели и
настраивает ее архитектуру в соответствии с характеристиками каждой зада-
чи. Более того, аппроксимационные свойства нео-фаззи нейронов как струк-
турных элементов системы обеспечивает высокую точность распознавания
в условиях коротких выборок данных. Предложена многослойная МГУА-
нейро-фаззи сеть на основе расширенных нео-фаззи нейронов. Алгоритм обу-
чения имеет фильтрующие и отслеживающие свойства и гарантирует необхо-
димую скорость для приложений реального времени. Эффективность предла-
гаемой системы подтверждена в задаче распознавания человеческих емоций.
Ключевые слова: метод группового учета аргументов, расширенный нео-фаззи
нейрон, онлайн распознавание изображений, распознавание выражений лица.
|
| id | journaliasakpiua-article-221271 |
| institution | System research and information technologies |
| keywords_txt_mv | keywords |
| language | English |
| last_indexed | 2025-07-17T10:26:59Z |
| publishDate | 2020 |
| publisher | The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" |
| record_format | ojs |
| resource_txt_mv | journaliasakpiua/c0/5425ee7d42c4e576823acbdfe31cccc0.pdf |
| spelling | journaliasakpiua-article-2212712021-01-19T12:18:25Z Multilayer GMDH-neuro-fuzzy network based on extended neo-fuzzy neurons and its application in online facial expression recognition Многослойная МГУА-нейро-фаззи сеть на основе расширенных нечетких нейронов и ее применение для онлайн распознавания выражений лица Багатошарова МГУА-нейро-фаззі мережа на основі розширених нечітких нейронів та її використання для онлайн разпізнавання виразів обличчя Bodyanskiy, Yevgeniy V. Zaychenko, Yuriy P. Hamidov, Galib Kulishova, Nonna Ye. Group Method of Data Handling extended neo-fuzzy neuron online image recognition facial expression recognition метод групового урахування аргументів розширений нео-фаззі нейрон онлайн розпізнавання зображень розпізнавання виразів обличчя метод группового учета аргументов расширенный нео-фаззи нейрон онлайн распознавание изображений распознавание выражений лица Real-time image recognition is required in many important practical problems. Interaction with users in online mode requires flexibility and adaptability from applications. The Group Method of Data Handling (GMDH) allows changing the model structure and adjusting the system architecture to the characteristics of each task under consideration. Moreover, the approximating properties of neo-fuzzy neurons used as elements of the system provide the high recognition accuracy under conditions of short data samples. This paper proposes a multilayer GMDH-neuro-fuzzy network based on extended neo-fuzzy neurons. The learning algorithm has filtering and tracking properties, guarantees the required speed important for real-time applications. The effectiveness of the proposed system is confirmed for the human emotions recognition. Распознавание изображений в реальном времени требуется в многих практических задачах. Взаимодействие с пользователями в режиме онлайн требует гибкости и адаптации от прикладных программ. Метод группового учета аргументов (МГУА) позволяет изменять структуру модели и настраивает ее архитектуру в соответствии с характеристиками каждой задачи. Более того, аппроксимационные свойства нео-фаззи нейронов как структурных элементов системы обеспечивает высокую точность распознавания в условиях коротких выборок данных. Предложена многослойная МГУА-нейро-фаззи сеть на основе расширенных нео-фаззи нейронов. Алгоритм обучения имеет фильтрующие и отслеживающие свойства и гарантирует необходимую скорость для приложений реального времени. Эффективность предлагаемой системы подтверждена в задаче распознавания человеческих емоций. Розпізнавання зображень в реальному часі потрібно в багатьох практичних задачах. Взаємодія з користувачами в режимі онлайн потребує гнучкості та адаптацїї від прикладних програм. Метод групового урахування аргументів (МГУА) дозволяє змінювати структуру моделі та налаштовує її архітектуру відповідно до характеристик кожної задачі. Більш того, апроксимаційні властивості нео-фазі нейронів як структурних елементів системи забезпечують високу точність розпізнавання в умовах коротких вибірок даних. Запропоновано багатошарову МГУА-нейро-фаззі мережу на основі розширених нео-фаззі нейронів. Алгоритм навчання має фільтрувальні та відслідковувальні властивості та гарантує необхідну швидкість для застосувань реального часу. Ефективність запропонованої системи підтверджено в задачі розпізнавання людських емоцій. The National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute" 2020-12-07 Article Article application/pdf https://journal.iasa.kpi.ua/article/view/221271 10.20535/SRIT.2308-8893.2020.3.05 System research and information technologies; No. 3 (2020); 66-78 Системные исследования и информационные технологии; № 3 (2020); 66-78 Системні дослідження та інформаційні технології; № 3 (2020); 66-78 2308-8893 1681-6048 en https://journal.iasa.kpi.ua/article/view/221271/223557 Copyright (c) 2021 System research and information technologies |
| spellingShingle | метод групового урахування аргументів розширений нео-фаззі нейрон онлайн розпізнавання зображень розпізнавання виразів обличчя Bodyanskiy, Yevgeniy V. Zaychenko, Yuriy P. Hamidov, Galib Kulishova, Nonna Ye. Багатошарова МГУА-нейро-фаззі мережа на основі розширених нечітких нейронів та її використання для онлайн разпізнавання виразів обличчя |
| title | Багатошарова МГУА-нейро-фаззі мережа на основі розширених нечітких нейронів та її використання для онлайн разпізнавання виразів обличчя |
| title_alt | Multilayer GMDH-neuro-fuzzy network based on extended neo-fuzzy neurons and its application in online facial expression recognition Многослойная МГУА-нейро-фаззи сеть на основе расширенных нечетких нейронов и ее применение для онлайн распознавания выражений лица |
| title_full | Багатошарова МГУА-нейро-фаззі мережа на основі розширених нечітких нейронів та її використання для онлайн разпізнавання виразів обличчя |
| title_fullStr | Багатошарова МГУА-нейро-фаззі мережа на основі розширених нечітких нейронів та її використання для онлайн разпізнавання виразів обличчя |
| title_full_unstemmed | Багатошарова МГУА-нейро-фаззі мережа на основі розширених нечітких нейронів та її використання для онлайн разпізнавання виразів обличчя |
| title_short | Багатошарова МГУА-нейро-фаззі мережа на основі розширених нечітких нейронів та її використання для онлайн разпізнавання виразів обличчя |
| title_sort | багатошарова мгуа-нейро-фаззі мережа на основі розширених нечітких нейронів та її використання для онлайн разпізнавання виразів обличчя |
| topic | метод групового урахування аргументів розширений нео-фаззі нейрон онлайн розпізнавання зображень розпізнавання виразів обличчя |
| topic_facet | Group Method of Data Handling extended neo-fuzzy neuron online image recognition facial expression recognition метод групового урахування аргументів розширений нео-фаззі нейрон онлайн розпізнавання зображень розпізнавання виразів обличчя метод группового учета аргументов расширенный нео-фаззи нейрон онлайн распознавание изображений распознавание выражений лица |
| url | https://journal.iasa.kpi.ua/article/view/221271 |
| work_keys_str_mv | AT bodyanskiyyevgeniyv multilayergmdhneurofuzzynetworkbasedonextendedneofuzzyneuronsanditsapplicationinonlinefacialexpressionrecognition AT zaychenkoyuriyp multilayergmdhneurofuzzynetworkbasedonextendedneofuzzyneuronsanditsapplicationinonlinefacialexpressionrecognition AT hamidovgalib multilayergmdhneurofuzzynetworkbasedonextendedneofuzzyneuronsanditsapplicationinonlinefacialexpressionrecognition AT kulishovanonnaye multilayergmdhneurofuzzynetworkbasedonextendedneofuzzyneuronsanditsapplicationinonlinefacialexpressionrecognition AT bodyanskiyyevgeniyv mnogoslojnaâmguanejrofazzisetʹnaosnoverasširennyhnečetkihnejronovieeprimeneniedlâonlajnraspoznavaniâvyraženijlica AT zaychenkoyuriyp mnogoslojnaâmguanejrofazzisetʹnaosnoverasširennyhnečetkihnejronovieeprimeneniedlâonlajnraspoznavaniâvyraženijlica AT hamidovgalib mnogoslojnaâmguanejrofazzisetʹnaosnoverasširennyhnečetkihnejronovieeprimeneniedlâonlajnraspoznavaniâvyraženijlica AT kulishovanonnaye mnogoslojnaâmguanejrofazzisetʹnaosnoverasširennyhnečetkihnejronovieeprimeneniedlâonlajnraspoznavaniâvyraženijlica AT bodyanskiyyevgeniyv bagatošarovamguanejrofazzímerežanaosnovírozširenihnečítkihnejronívtaíívikoristannâdlâonlajnrazpíznavannâvirazívobliččâ AT zaychenkoyuriyp bagatošarovamguanejrofazzímerežanaosnovírozširenihnečítkihnejronívtaíívikoristannâdlâonlajnrazpíznavannâvirazívobliččâ AT hamidovgalib bagatošarovamguanejrofazzímerežanaosnovírozširenihnečítkihnejronívtaíívikoristannâdlâonlajnrazpíznavannâvirazívobliččâ AT kulishovanonnaye bagatošarovamguanejrofazzímerežanaosnovírozširenihnečítkihnejronívtaíívikoristannâdlâonlajnrazpíznavannâvirazívobliččâ |