T
E
L
KO
M
NIK
A
, V
ol
.
17
,
No.
6,
Dec
em
be
r
20
1
9,
p
p.
2
97
5
~
2
98
2
IS
S
N: 1
69
3
-
6
93
0
,
accr
ed
ited
F
irst
Gr
ad
e b
y K
em
en
r
istekdikti,
Decr
ee
No: 2
1/E/
K
P
T
/20
18
DOI:
10.12928/TE
LK
OM
N
IK
A
.v
1
7
i
6
.
12948
◼
29
75
Rec
ei
v
ed
A
pril
17
,
20
1
9
; R
ev
i
s
ed
J
un
e 2
0
, 2
01
9
;
A
c
c
ep
te
d
J
ul
y
2
, 2
0
1
9
Reco
gn
iti
on
s
y
stem
fo
r
facial e
xpressio
n b
y
pr
ocessi
ng
imag
e
s
w
ith
de
ep learni
ng
neu
ral net
w
ork
Ho
lman
M
o
n
t
iel
A
r
iz
a
*1
,
H
enr
y
H
er
n
ánd
ez
M
ar
t
ínez
2
, Lu
z
A
n
d
r
ea
G
av
ir
i
a Roa
3
1
Uni
v
e
rs
i
d
a
d
Di
s
tri
ta
l
Fra
n
c
i
s
c
o
J
o
s
é
d
e
Ca
l
d
a
s
,
Fa
c
u
l
t
a
d
T
e
c
n
o
l
ó
g
i
c
a
,
Cll
6
8
D B
i
s
A
S
u
r
No.
4
9
F
–
7
0
,
Bo
g
o
t
á
D.C.,
Co
l
o
m
b
i
a
2
Uni
v
e
rs
i
d
a
d
Na
c
i
o
n
a
l
d
e
Co
l
o
m
b
i
a
,
Dep
a
r
ta
m
e
n
to
d
e
I
n
g
e
n
i
e
ría
d
e
Si
s
te
m
a
s
e
I
n
d
u
s
t
ri
a
l
,
Bo
g
o
tá
D.C.
,
Col
o
m
b
i
a
3
Fu
n
d
a
c
i
ó
n
Un
i
v
e
rs
i
t
a
ri
a
P
a
n
a
m
e
ri
c
a
n
a
,
Fa
c
u
l
ta
d
d
e
I
n
g
e
n
i
e
ría
,
Bo
g
o
tá
D.C.,
Co
l
o
m
b
i
a
*
Corre
s
p
o
n
d
i
n
g
a
u
th
o
r,
e
-
m
a
i
l
:
h
m
o
n
ti
e
l
a
@ud
i
s
tr
i
ta
l
.e
d
u
.c
o
Ab
strac
t
Th
e
re
c
o
g
n
i
ti
o
n
s
y
s
te
m
s
o
f
p
a
tt
e
rn
s
i
n
i
m
a
g
e
s
a
r
e
m
e
c
h
a
n
i
s
m
s
th
a
t
fi
l
te
r
th
e
i
n
fo
rm
a
ti
o
n
th
a
t
p
ro
v
i
d
e
s
a
n
i
m
a
g
e
to
h
i
g
h
l
i
g
h
t
t
h
e
a
re
a
o
f
i
n
t
e
re
s
t
fo
r
th
e
u
s
e
r.
Us
u
a
l
l
y
,
th
e
s
e
m
e
c
h
a
n
i
s
m
s
a
re
b
a
s
e
d
o
n
m
a
th
e
m
a
ti
c
a
l
tra
n
s
f
o
rm
a
ti
o
n
s
t
h
a
t
a
l
l
o
w
t
h
e
p
ro
c
e
s
s
o
r
to
p
e
r
fo
rm
i
n
te
rp
r
e
ta
t
i
o
n
s
b
a
s
e
d
o
n
t
h
e
g
e
o
m
e
tr
y
o
r
s
h
a
p
e
o
f
th
e
i
m
a
g
e
.
Howe
v
e
r,
th
e
s
tra
te
g
i
e
s
th
a
t
i
m
p
l
e
m
e
n
t
m
a
th
e
m
a
ti
c
a
l
tr
a
n
s
fo
rm
a
ti
o
n
s
a
re
l
i
m
i
te
d
,
s
i
n
c
e
th
e
e
ff
e
c
ti
v
e
n
e
s
s
o
f
t
h
e
s
e
t
e
c
h
n
i
q
u
e
s
i
s
re
d
u
c
e
d
b
y
c
h
a
n
g
i
n
g
th
e
m
o
rp
h
o
l
o
g
y
o
r
re
s
o
l
u
t
i
o
n
o
f
th
e
i
m
a
g
e
.
Th
i
s
p
a
p
e
r
p
r
e
s
e
n
t
s
a
p
a
r
ti
a
l
s
o
l
u
ti
o
n
to
t
h
i
s
l
i
m
i
t
a
ti
o
n
w
i
th
a
d
i
g
i
ta
l
i
m
a
g
e
p
r
o
c
e
s
s
i
n
g
t
e
c
h
n
i
q
u
e
b
a
s
e
d
o
n
a
d
e
e
p
l
e
a
rn
i
n
g
n
e
u
ra
l
n
e
two
r
k
(DN
N).
Th
i
s
te
c
h
n
i
q
u
e
i
n
c
o
rp
o
r
a
te
s
a
m
e
c
h
a
n
i
s
m
t
h
a
t
a
l
l
o
w
s
th
e
DN
N
to
d
e
te
rm
i
n
e
th
e
fa
c
i
a
l
e
x
p
r
e
s
s
i
o
n
o
f
a
p
e
rs
o
n
,
b
a
s
e
d
o
n
t
h
e
s
e
g
m
e
n
te
d
i
n
fo
rm
a
t
i
o
n
o
f
t
h
e
i
m
a
g
e
o
f
th
e
i
r
f
a
c
e
.
B
y
s
e
g
m
e
n
t
i
n
g
th
e
i
m
a
g
e
a
n
d
p
ro
c
e
s
s
i
n
g
i
ts
c
h
a
r
a
c
t
e
ri
s
ti
c
s
i
n
p
a
ra
l
l
e
l
,
th
e
p
ro
p
o
s
e
d
te
c
h
n
i
q
u
e
i
n
c
r
e
a
s
e
s
th
e
e
ff
e
c
ti
v
e
n
e
s
s
o
f
re
c
o
g
n
i
z
i
n
g
fa
c
i
a
l
g
e
s
t
u
re
s
i
n
d
i
ff
e
re
n
t
i
m
a
g
e
s
e
v
e
n
whe
n
m
o
d
i
fy
i
n
g
t
h
e
i
r
c
h
a
r
a
c
t
e
ri
s
ti
c
s
.
Key
w
ords
:
d
e
e
p
l
e
a
rn
i
n
g
n
e
u
ra
l
n
e
two
rk
,
f
a
c
e
re
c
o
g
n
i
ti
o
n
,
p
ro
c
e
s
s
i
n
g
i
m
a
g
e
s
,
r
e
c
o
g
n
i
ti
o
n
s
y
s
t
e
m
s
o
f
p
a
tt
e
rn
s
Copy
righ
t
©
2
0
1
9
Uni
v
e
rsi
t
a
s
Ahm
a
d
D
a
hl
a
n.
All
rig
ht
s
r
e
s
e
rve
d
.
1.
Int
r
o
d
u
ctio
n
T
he
di
gi
t
al
i
m
ag
e
proc
es
s
i
n
g
i
s
a
m
eth
od
o
l
og
y
t
o
i
m
pl
em
en
t
a
tec
hn
i
qu
e
tha
t
i
s
u
s
ed
i
n
order
to
i
m
prov
e
or
r
es
t
o
r
e
an
i
m
ag
e
tha
t
un
de
r
g
oe
s
s
om
e
c
ha
ng
e
d
urin
g
i
ts
ac
qu
i
s
i
t
i
on
proc
es
s
.
F
or
ex
am
pl
e,
r
ed
uc
i
ng
t
he
r
es
o
l
ut
i
o
n
of
a
s
i
gn
at
ure
w
h
en
s
c
an
n
i
n
g
a
do
c
um
en
t.
Due
to
th
e
l
arge
nu
m
be
r
of
tec
hn
i
qu
es
t
he
s
e
are
of
ten
c
o
nf
us
ed
,
bu
t
th
e
y
are
no
t
th
e
s
am
e,
s
i
nc
e
a
group
of
tec
hn
i
qu
es
i
s
de
di
c
at
ed
to
t
he
r
es
tora
ti
on
of
i
m
ag
es
an
d
are
us
ed
to
m
a
k
e
as
s
u
m
pti
on
s
a
bo
ut
m
i
s
s
i
ng
i
nf
orm
ati
on
.
T
he
oth
er
gro
up
c
on
t
ai
ns
i
m
pr
ov
em
en
t
tec
hn
i
qu
es
t
ha
t
al
l
o
w
th
e
us
er
to
e
l
i
m
i
na
t
e
or
hi
g
hl
i
gh
t
i
m
ag
e
i
nf
orm
ati
on
to
de
term
i
ne
area
s
or
r
eg
i
on
s
of
i
nte
r
es
t
[1,
2].
T
he
au
tom
ati
on
of
the
s
e
tec
hn
i
qu
es
h
as
al
l
o
wed
t
he
ge
ne
r
at
i
o
n
of
s
tr
ate
gi
e
s
an
d
al
g
orit
hm
s
f
or
the
proc
es
s
i
ng
of
v
i
de
o
or
s
eq
u
en
c
es
o
f
i
m
ag
es
i
n
a
c
o
nti
nu
o
us
m
an
ne
r
.
T
ha
t
i
s
,
the
s
am
e
l
og
i
c
a
l
s
tr
uc
tur
e
i
s
us
ed
t
o
proc
es
s
l
arge
v
ol
um
es
of
i
nf
or
m
ati
on
,
w
h
i
c
h
ha
s
en
c
ou
r
a
ge
d
the
d
ev
el
o
pm
en
t
of
ad
v
an
c
e
d
t
ec
hn
i
qu
e
s
s
uc
h
as
art
i
f
i
c
i
al
v
i
s
i
on
,
c
l
os
ed
s
ec
ur
i
t
y
s
y
s
t
em
s
,
the
an
al
y
s
i
s
of
hu
m
an
ga
i
t
or
the
l
oc
at
i
on
of
ge
om
etri
c
pa
tte
r
ns
[3
-
8
].
A
l
tho
u
gh
th
ere
i
s
a
l
ot
of
tec
hn
i
q
ue
s
to
proc
es
s
i
m
ag
es
,
t
he
s
e
h
av
e
s
om
e
l
i
m
i
tat
i
o
ns
,
be
c
a
us
e
t
he
y
de
p
en
d
on
the
c
a
ptu
r
e
de
v
i
c
e
a
nd
t
he
r
es
o
l
ut
i
on
of
th
e
i
m
ag
e
to
f
un
c
ti
o
n
i
n
an
ap
pro
pria
te
m
an
ne
r
.
T
he
l
o
w
er
the
r
es
ol
u
ti
on
of
th
e
i
m
ag
e,
the
m
ore
m
ath
e
m
ati
c
al
op
era
tors
, m
as
k
s
o
r
f
i
l
ters
us
ed
to
de
term
i
ne
t
he
r
eg
i
on
s
of
i
nt
eres
t
i
n
an
i
m
ag
e
wi
l
l
be
e
v
en
great
er.
I
n
oth
er
w
ords
,
b
y
i
nc
r
e
as
i
ng
the
nu
m
be
r
of
op
erat
ors
,
th
e
am
ou
nt
of
c
om
pu
tat
i
on
a
l
r
es
ou
r
c
es
to
d
ete
r
m
i
ne
t
he
i
nf
orm
ati
on
of
i
nte
r
es
t
al
s
o
do
es
[9
-
11
]
.
T
hi
s
l
i
m
i
tat
i
o
n
w
as
ad
dres
s
ed
i
n
d
i
f
f
erent
wa
y
s
i
n
di
f
f
erent
areas
of
e
ng
i
ne
erin
g
us
i
ng
i
nte
l
l
i
g
en
t
s
y
s
t
em
s
or
ex
pe
r
t
s
y
s
t
em
s
.
O
ne
of
the
c
o
m
m
on
l
y
us
ed
tec
h
ni
qu
es
i
s
t
he
V
i
o
l
a
J
on
es
al
g
orit
hm
,
s
i
n
c
e
i
t
i
m
pl
em
en
ts
a
l
e
arni
ng
a
l
go
r
i
thm
tha
t
i
de
nt
i
f
i
es
r
os
ters
i
n
s
ets
of
i
m
ag
es
[1
,
4]
.
A
l
t
ho
u
gh
,
t
hi
s
tec
h
ni
q
ue
i
s
ea
s
i
l
y
progr
am
m
ab
l
e
i
n
de
v
i
c
es
w
i
t
h
a
l
o
w
l
ev
el
of
proc
es
s
i
ng
,
i
t
do
es
n
ot
pe
r
f
orm
i
nte
r
pre
tat
i
on
s
of
the
m
oo
d
of
the
pe
r
s
on
[
12
-
17].
T
he
p
r
ob
l
e
m
of
i
de
nt
i
f
y
i
n
g
s
tat
es
of
m
i
nd
i
s
q
ui
t
e
c
om
pl
ex
an
d
i
s
n
ot
a
l
i
m
i
tat
i
o
n
of
the
a
l
g
orit
hm
of
V
i
o
l
a
J
o
ne
s
on
l
y
,
s
i
nc
e
the
i
de
n
ti
f
i
c
ati
on
of
f
ee
l
i
ng
s
f
r
o
m
pa
tte
r
ns
i
s
a
s
ub
j
ec
t
t
ha
t
i
s
u
nd
er
s
tu
d
y
f
or
the
d
ev
e
l
o
pm
en
t
o
f
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 16
93
-
6
93
0
T
E
L
KO
M
NIK
A
V
ol
.
17
,
No
.
6,
D
ec
em
be
r
20
19
:
29
7
5
-
2982
2976
s
ec
urit
y
s
y
s
tem
s
,
m
ed
i
c
al
t
r
ea
tm
e
nts
,
a
m
on
g
oth
ers
[
18
-
23]
.
I
n
th
i
s
pa
pe
r
th
ere
i
s
propos
e
d
a
s
tr
ate
g
y
to
d
ete
r
m
i
ne
m
oo
ds
f
r
o
m
f
ac
i
al
ex
pres
s
i
on
,
wi
th
a
l
o
w
c
os
t
c
om
pu
tat
i
o
na
l
a
l
go
r
i
thm
.
T
hi
s
al
go
r
i
thm
i
s
de
s
c
r
i
be
d
i
n
t
he
f
ol
l
o
wi
n
g
s
ec
ti
on
s
whi
c
h
are
org
an
i
z
ed
as
f
ol
l
o
w
s
:
s
ec
ti
o
n
2
c
on
tai
ns
a
d
es
c
r
i
pti
on
of
th
e
d
ev
el
op
e
d
t
ec
hn
i
qu
e.
I
n
s
ec
ti
on
3
,
th
ere
i
s
pres
e
nte
d
a
d
es
c
r
i
pti
on
of
the
ex
p
erim
en
t
tha
t
w
as
us
ed
t
o
t
es
t
th
i
s
al
go
r
i
th
m
.
In
s
ec
ti
on
4
,
th
e
r
es
ul
t
s
ob
ta
i
ne
d
wi
th
the
ex
pe
r
i
m
en
t
ar
e
d
es
c
r
i
be
d.
In
s
ec
ti
o
n
5
,
t
he
r
e
a
r
e
the
c
on
c
l
us
i
o
ns
th
at
wer
e
b
ui
l
t
f
r
om
the
r
es
u
l
ts
ob
t
ai
ne
d
.
2.
M
ateri
als
and
M
eth
o
d
s
A
s
i
t
w
as
s
a
i
d
i
n
the
pre
v
i
ou
s
s
ec
ti
o
n,
t
he
tec
hn
i
qu
e
s
of
i
m
ag
e
proc
es
s
i
ng
are
us
ua
l
l
y
ba
s
ed
on
m
ath
em
ati
c
al
tr
a
ns
f
or
m
ati
on
s
th
at
al
l
o
w
to
hi
g
hl
i
gh
t
or
i
nd
i
c
ate
r
eg
i
on
s
of
i
nt
eres
t
f
or
the
us
er.
A
t
pr
es
en
t,
i
t
i
s
tr
i
ed
to
i
m
prov
e
th
i
s
t
y
p
e
of
t
ec
hn
i
qu
es
b
y
i
m
pl
em
en
ti
ng
s
eg
m
en
tat
i
on
an
d
c
l
as
s
i
f
i
c
at
i
on
m
eth
od
s
to
r
ed
uc
e
t
he
am
ou
nt
of
pa
r
am
ete
r
s
ne
c
es
s
ar
y
wh
en
pe
r
f
orm
i
ng
an
o
pe
r
at
i
o
n.
Ho
we
v
er,
t
he
us
e
an
d
a
pp
l
i
c
at
i
on
of
c
l
as
s
i
f
i
ers
i
n
t
he
proc
es
s
of
pa
tte
r
n
r
ec
og
ni
t
i
o
n
i
n
a
n
i
m
ag
e
i
s
a
s
ub
j
ec
t
un
de
r
s
tud
y
,
du
e
to
th
e
l
arg
e
nu
m
be
r
of
ap
pl
i
c
a
ti
o
ns
t
ha
t
thi
s
t
y
p
e
of
tec
hn
i
qu
e
c
a
n
ha
v
e.
A
m
on
g
the
m
,
the
r
ec
og
n
i
ti
on
of
f
ac
i
al
ex
pres
s
i
on
s
a
nd
th
ei
r
r
el
at
i
on
s
h
i
p
wi
t
h
em
oti
on
s
f
r
o
m
an
i
m
ag
e
gi
v
e
n
b
y
t
h
e
us
er
i
s
a
n
i
s
s
ue
i
n
de
v
el
op
m
en
t,
s
i
nc
e,
c
on
v
e
nti
on
a
l
l
y
,
i
m
ag
e
pro
c
es
s
i
ng
i
s
us
e
d
to
i
d
en
t
i
f
y
pa
r
ts
of
the
hu
m
an
bo
d
y
a
nd
no
t
f
or
un
de
r
s
ta
nd
i
ts
op
erati
on
.
T
hi
s
art
i
c
l
e
pro
po
s
es
a
c
o
ntr
i
bu
t
i
o
n
t
o
t
hi
s
top
i
c
wi
th
th
e
de
v
e
l
op
m
en
t
of
a
c
ha
r
ac
ter
i
de
nt
i
f
i
c
ati
o
n
s
y
s
tem
ba
s
ed
on
a
d
ee
p
l
ea
r
n
i
ng
ne
ural
ne
t
wor
k
(
DNN)
,
w
h
i
c
h
i
de
nti
f
i
es
a
p
ers
on
's
f
ac
i
al
ex
pres
s
i
on
an
d
as
s
oc
i
ate
s
i
t
wi
th
a
f
ee
l
i
ng
t
ha
t
c
a
n
f
ee
l
.
F
i
n
al
l
y
,
s
tr
uc
ture of
th
e DN
N an
d i
t
s
prev
i
ou
s
pre
pa
r
at
i
o
n t
o
d
ev
e
l
o
p t
h
i
s
w
ork
i
s
de
s
c
r
i
be
d b
e
l
o
w
.
2.1
.
P
r
ep
ar
atio
n
and
P
r
o
ce
ss
i
n
g
o
f
a
Dig
it
al
Im
ag
e
In
r
el
a
ti
o
n
to
what
was
s
ai
d
b
ef
ore,
a
d
i
g
i
ta
l
i
m
ag
e
i
s
c
on
s
tr
uc
ted
w
i
th
a
n
u
m
eric
al
m
atri
x
w
hi
c
h
r
e
pres
en
ts
a
t
wo
-
di
m
en
s
i
on
al
i
m
ag
e.
T
he
di
m
en
s
i
on
s
of
the
m
atri
x
v
ar
y
d
ep
e
nd
i
ng
on
th
e
r
es
ol
uti
on
of
th
e
i
m
ag
e
a
nd
t
he
nu
m
be
r
of
m
at
r
i
c
es
tha
t
ar
e
us
ed
to
r
ep
r
e
s
en
t
the
s
am
e
s
c
en
e
c
ha
n
ge
s
d
ep
e
nd
i
ng
on
t
he
n
um
be
r
of
c
ol
ors
.
F
or
ex
am
pl
e,
a
s
i
ng
l
e
bi
na
r
y
c
o
ef
f
i
c
i
en
t
m
atri
x
i
s
r
eq
ui
r
ed
t
o repres
en
t a
bl
ac
k
an
d
w
hi
te
i
m
ag
e.
T
he
r
e
are
s
e
v
eral
w
a
y
s
t
o
ob
ta
i
n
a
d
i
g
i
ta
l
i
m
ag
e,
am
on
g
t
he
m
are
s
c
an
ne
r
s
a
nd
di
g
i
ta
l
c
a
m
eras
.
A
n
ad
v
a
nta
ge
of
the
s
e
de
v
i
c
es
i
s
t
ha
t
th
e
y
a
l
l
o
w
t
he
m
s
el
v
es
t
o
a
pp
l
y
tr
an
s
f
or
m
ati
on
s
to
m
od
i
f
y
t
he
i
m
ag
e
be
f
ore
s
tori
ng
i
t
,
s
uc
h
as
f
i
l
ters
t
o
el
i
m
i
na
t
e
b
ac
k
ground
l
i
gh
t,
c
r
op
or
r
ota
t
e
the
s
c
en
e.
Ho
wev
er,
the
p
r
oc
es
s
i
ng
c
a
pa
c
i
t
y
of
th
es
e
d
ev
i
c
es
i
s
l
i
m
i
ted
,
th
eref
ore,
proc
es
s
i
ng
on
c
e t
h
e a
c
q
ui
s
i
t
i
o
n o
f
th
e
i
m
ag
e i
n a
c
om
pu
ter i
s
t
he
m
o
s
t c
o
m
m
on
[1
].
T
r
ad
i
ti
on
al
c
om
pu
ter
i
m
ag
e
proc
es
s
i
ng
progr
am
s
al
l
o
w
tr
an
s
f
or
m
ati
on
s
or
be
au
t
i
f
i
c
a
ti
on
of
di
gi
tal
i
m
ag
es
on
l
y
,
du
e
to
th
i
s
,
s
of
tware
-
ba
s
ed
ap
pl
i
c
a
ti
o
ns
ha
v
e
be
e
n
d
ev
i
s
ed
to
i
nc
r
ea
s
e
the
am
ou
nt
of
o
pe
r
at
i
on
s
a
v
ai
l
a
bl
e
to
us
ers
t
o
un
de
r
s
tan
d
th
e
i
nf
orm
ati
on
t
ha
t
an
i
m
ag
e
prov
i
de
s
.
S
om
e
of
the
s
e
ap
p
l
i
c
at
i
on
s
al
l
o
w
i
nf
orm
a
ti
on
to
be
r
ec
o
v
er
e
d
b
y
r
ec
on
s
tr
uc
ti
ng
the
i
m
ag
e
ba
s
ed
on
as
s
um
pti
o
ns
or
e
l
i
m
i
na
t
i
ng
c
h
ara
c
teri
s
ti
c
s
tha
t
att
e
nu
ate
t
he
i
nf
orm
ati
on
of
i
nte
r
es
t
[2
-
5].
T
he
i
nf
orm
ati
on
of
i
n
teres
t
i
n
an
i
m
ag
e
i
s
us
ua
l
l
y
h
i
g
hl
i
gh
t
ed
or
ex
tr
ac
te
d
b
y
ge
o
m
etri
c
tr
an
s
f
or
m
ati
on
s
,
w
h
i
c
h
al
l
o
w
t
o
i
n
di
c
a
te
p
att
erns
or
gr
ou
ps
of
pi
x
e
l
s
tha
t
c
on
ta
i
n
c
ha
r
ac
teri
s
ti
c
s
prev
i
ou
s
l
y
d
ef
i
ne
d
b
y
the
u
s
er. A
m
on
g t
he
s
i
m
pl
es
t
i
s
the
s
y
s
tem
f
or dete
c
ti
ng
g
e
om
etri
c
f
i
gu
r
es
tha
t
i
s
ba
s
ed
o
n
the
n
um
be
r
of
po
i
nts
a
f
i
gu
r
e
c
a
n
ha
v
e,
an
d
am
on
g
the
m
os
t
c
om
pl
ex
are
the
s
y
s
t
em
s
th
at
i
de
n
ti
f
y
c
h
arac
teri
s
ti
c
s
of
an
i
m
ag
e u
s
i
ng
i
nt
el
l
i
ge
nt
s
y
s
tem
s
.
P
r
oc
es
s
i
ng
us
i
ng
i
n
tel
l
i
ge
n
t
s
y
s
tem
s
r
eq
ui
r
es
pre
-
pro
c
es
s
i
ng
of
t
he
i
m
ag
e,
b
ec
au
s
e
the
pa
r
t
i
c
ul
ar
c
ha
r
ac
teri
s
t
i
c
s
tha
t
on
e
want
s
to
f
i
n
d
when
i
m
pl
em
en
ti
ng
th
i
s
s
y
s
tem
m
us
t
b
e
po
i
n
ted
ou
t
.
O
ne
wa
y
to
d
o
thi
s
i
s
t
o
us
e
a
hi
s
t
og
r
am
to
tea
c
h
the
i
nt
el
l
i
ge
nt
s
y
s
te
m
the
r
el
ati
v
e
f
r
eq
ue
nc
y
wi
t
h
whi
c
h
gro
u
ps
of
c
ol
ors
ap
pe
ar
i
n
a
n
i
m
ag
e.
T
he
m
os
t
c
o
m
m
on
wa
y
to
es
ti
m
ate
the
v
al
u
es
of
a
hi
s
tog
r
am
i
s
ba
s
ed
on
:
de
c
r
ea
s
i
ng
th
e
nu
m
be
r
of
di
m
en
s
i
on
s
of
the
i
m
ag
e
b
y
c
on
v
ert
i
n
g
i
ts
c
o
l
or
f
orm
at
to
gra
y
s
c
a
l
e
an
d
es
ti
m
ati
ng
t
he
f
r
eq
u
en
c
i
es
wi
th
t
he
ex
pres
s
i
on
s
ho
w
n
i
n
(
1
)
.
W
he
r
e
the
di
m
en
s
i
on
s
are
w
(
wi
dth
)
a
n
d
h
(
he
i
gh
t)
,
n
r
ep
r
es
en
ts
the
gra
y
l
ev
el
s
and
i
s
th
e n
um
be
r
of
pi
x
el
s
[5
].
ℎ
(
)
=
.
ℎ
(
1)
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
Rec
og
n
i
ti
on
s
y
s
tem
for fac
i
al
ex
pres
s
i
on
by
proc
es
s
i
n
g i
m
ag
es
..
. (
Ho
l
ma
n
Mo
nt
i
e
l
A
r
i
z
a
)
2977
In
th
i
s
pa
pe
r
,
t
he
pre
-
proc
es
s
i
ng
of
the
i
m
ag
e
was
do
ne
i
n
th
e
f
ol
l
o
w
i
ng
wa
y
.
F
i
r
s
t,
the
nu
m
be
r
of
d
i
m
en
s
i
on
s
of
the
i
m
ag
e
i
s
r
e
du
c
ed
b
y
c
ha
n
gi
ng
the
c
o
l
or
f
orm
at
to
gr
a
y
s
c
al
e
.
T
he
n,
the
i
m
ag
e
i
s
s
e
gm
en
ted
i
nt
o
at
l
ea
s
t
f
ou
r
p
arts
t
ha
t
are
th
e
f
ac
e,
t
he
m
ou
th
an
d
th
e
e
y
es
(
the
g
l
as
s
es
c
ou
nt
as
on
e
e
y
e
)
.
F
i
na
l
l
y
,
e
ac
h
s
eg
m
en
t
of
the
i
m
ag
e
i
s
tr
a
ns
f
or
m
ed
i
nto
a
h
i
s
tog
r
am
an
d
s
tore
d
i
n
an
arr
a
y
tha
t
wi
l
l
l
at
er
b
e
proc
es
s
ed
b
y
th
e
D
NN
,
s
ee
F
i
g
ure
1.
T
he
pa
r
ti
c
u
l
ar
c
ha
r
ac
t
eris
ti
c
s
of
the
DNN
an
d
i
ts
r
e
l
at
i
on
wi
t
h
th
e
pre
-
proc
es
s
i
ng
of
the
i
m
ag
e
are de
s
c
r
i
b
ed
i
n t
he
f
ol
l
o
wi
ng
n
um
eral
.
F
i
g
ure
1
.
S
c
he
m
ati
c
of
i
m
a
ge
pre
-
pr
oc
es
s
i
ng
2.2
.
P
att
er
n
s
Re
cog
n
it
ion
w
it
h
a
DNN
Neura
l
ne
t
wor
k
s
of
de
ep
l
ea
r
n
i
n
g
(
DNN)
are
ex
ten
de
d
m
od
el
s
of
tr
ad
i
ti
on
a
l
ne
ura
l
ne
t
w
ork
s
,
bu
t
un
l
i
k
e
the
m
DNN
ge
ne
r
ate
m
od
el
s
t
o
r
ep
r
es
en
t
l
arg
e
v
ol
um
es
of
i
nf
orm
ati
on
i
n
a
s
i
m
pl
e
w
a
y
.
A
m
on
g
t
he
m
o
s
t
c
o
m
m
on
f
or
m
s
or
m
o
de
l
s
t
o
r
e
pres
en
t
grou
ps
of
da
t
a
wi
t
h
th
i
s
t
y
p
e
of
ne
t
w
ork
are
the
c
l
as
s
i
f
i
ers
an
d
the
a
pp
r
ox
i
m
ati
o
ns
b
y
r
e
gres
s
i
on
s
.
O
n
the
on
e
ha
nd
,
c
l
as
s
i
f
i
ers
are m
od
el
s
th
at
s
ol
v
e p
r
o
bl
em
s
o
f
c
l
as
s
es
i
n
w
h
i
c
h
i
t i
s
i
nt
en
d
ed
to
gro
up
o
bj
ec
ts
w
i
t
h
de
f
i
ne
d
c
h
arac
teri
s
ti
c
s
.
O
n
the
oth
er
ha
n
d,
r
e
gres
s
i
on
ap
prox
i
m
ati
on
s
are
nu
m
eric
al
r
ep
r
es
en
ta
ti
o
ns
g
en
er
ate
d
to
as
s
oc
i
ate
gro
up
s
of
nu
m
be
r
s
.
In
b
oth
c
as
es
,
t
he
m
od
e
l
i
s
a
b
l
ac
k
bo
x
,
t
ha
t
i
s
,
t
he
DNN
es
t
i
m
ate
s
an
o
utp
ut
v
a
l
u
e
f
r
om
c
ertai
n
i
np
u
t
i
nf
orm
ati
on
.
Ho
w
e
v
er,
the
us
er
n
ev
er
k
no
w
s
th
e
m
ath
em
ati
c
al
ex
pres
s
i
o
n
or
f
or
m
of
the
c
l
as
s
i
f
i
er
t
ha
t
m
ak
es
up
the
DN
N [1
4].
T
he
top
ol
og
y
or
f
or
m
of
th
e
DNN
d
ep
e
nd
s
on
c
ert
ai
n
pa
r
am
ete
r
s
de
f
i
ne
d
b
y
th
e
us
er,
am
on
g
w
h
i
c
h
ar
e
th
e
nu
m
be
r
of
en
tr
an
c
es
,
ex
i
ts
,
ne
uro
ns
an
d
hi
dd
e
n
l
a
y
er
s
,
the
f
orm
o
f
the
ac
ti
v
at
i
o
n
f
un
c
ti
on
an
d
the
al
go
r
i
t
hm
of
tr
ai
ni
ng
or
r
ed
uc
ti
o
n
of
the
err
or
[24
]
.
T
he
nu
m
be
r
of
i
np
uts
an
d
o
utp
uts
v
arie
s
d
ep
en
di
ng
on
t
he
grou
p
of
tr
ai
n
i
ng
da
ta,
tha
t
i
s
,
t
he
nu
m
be
r
of
i
np
uts
i
s
de
term
i
ne
d
b
y
the
i
nd
ep
en
de
n
t
v
ari
ab
l
es
t
ha
t
al
l
o
w
es
ti
m
ati
ng
th
e
o
utp
ut
v
a
l
ue
an
d
the
nu
m
be
r
of
ou
t
pu
ts
d
ep
en
ds
o
n
t
he
nu
m
be
r
of
v
ari
ab
l
es
t
ha
t
c
ha
n
ge
de
pe
n
di
ng
on
the
en
tr
a
nc
e.
T
he
nu
m
be
r
of
ne
ur
on
s
a
nd
hi
dd
en
l
a
y
ers
are
s
toc
h
as
ti
c
v
a
l
ue
s
,
s
ee
i
n
(
2),
de
term
i
ne
d
b
y
the
us
er
when
d
es
i
gn
i
n
g
the
n
et
w
or
k
,
i
t
s
ho
ul
d
be
tak
en
i
n
to
ac
c
ou
n
t
tha
t
i
nc
r
ea
s
i
ng
th
e
nu
m
be
r
of
ne
uro
ns
(
δ)
an
d
h
i
d
de
n
l
a
y
e
r
s
(
ε
)
i
nc
r
ea
s
es
the
DN
N
ac
c
urac
y
an
d
de
c
r
ea
s
es
the
p
erf
or
m
an
c
e
of
the
proc
es
s
or,
s
i
nc
e,
the
am
ou
nt
of
nu
m
eric
a
l
c
al
c
ul
ati
on
s
i
s
i
nc
r
ea
s
i
ng
.
{
,
}
∈
ℤ
|
{
,
}
≥
0
(
2)
T
he
r
e
i
s
n
o
wa
y
t
o
de
term
i
ne
t
he
ex
ac
t
nu
m
be
r
of
h
i
d
de
n
l
a
y
ers
an
d
ne
uro
ns
f
or
ea
c
h
DNN,
b
ec
au
s
e
ea
c
h
ap
p
l
i
c
ati
o
n
h
as
a
d
i
f
f
erent
tr
ai
ni
n
g
da
t
a
s
et.
T
hi
s
i
s
be
c
a
us
e
ea
c
h
ne
uron
s
tores
a
v
a
l
ue
c
al
l
e
d
wei
g
h
t,
w
h
i
c
h
i
s
r
es
po
ns
i
b
l
e
f
or
m
od
i
f
y
i
n
g
t
he
ou
t
pu
t
v
al
u
e
of
ea
c
h
n
eu
r
o
n
b
y
i
nc
r
ea
s
i
ng
or
de
c
r
ea
s
i
ng
th
e
i
np
ut
v
al
ue
.
In
a
dd
i
t
i
o
n,
th
e
wei
g
ht
i
s
ac
c
o
m
pa
ni
e
d
b
y
an
ac
ti
v
at
i
on
f
un
c
ti
o
n,
w
h
i
c
h
i
s
r
es
po
ns
i
b
l
e
f
or
l
i
m
i
ti
n
g
the
ou
t
pu
t
v
al
ue
of
ea
c
h
n
e
uron.
A
n
oth
er
f
ea
ture
of
the
wei
gh
ts
i
s
th
at
to
i
nc
r
ea
s
e
the
ac
c
u
r
ac
y
of
the
o
utp
u
t
v
al
ue
of
th
e
DNN,
d
i
f
f
erent
tr
ai
n
i
ng
a
l
g
orit
hm
s
are
us
ed
,
whi
c
h
a
uto
m
ati
c
al
l
y
m
od
i
f
y
t
he
w
e
i
gh
ts
of
ea
c
h
ne
u
r
on
to
r
ed
uc
e
the
m
argi
n o
f
err
or be
t
w
ee
n t
he
ou
t
pu
t
of
th
e n
et
w
ork
an
d t
he
tra
i
n
i
n
g d
a
ta
[
14
,
20
,
25
].
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 16
93
-
6
93
0
T
E
L
KO
M
NIK
A
V
ol
.
17
,
No
.
6,
D
ec
em
be
r
20
19
:
29
7
5
-
2982
2978
In
t
hi
s
pa
p
er
i
t
was
us
e
d
a
DNN
w
i
t
h
10
2
4
en
tr
i
es
,
on
e
(
1)
ou
tpu
t
,
12
0
0
n
e
urons
,
ten
(
1
0)
h
i
d
de
n
l
a
y
e
r
s
,
a
r
an
do
m
f
un
c
ti
on
w
i
th
un
i
f
orm
di
s
tr
i
bu
ti
on
to
as
s
i
gn
i
ni
ti
a
l
v
a
l
ue
s
to
the
wei
gh
ts
,
a
r
ec
t
i
f
i
ed
l
i
n
ea
r
u
ni
t
ac
ti
v
a
ti
o
n
f
un
c
ti
on
,
a
de
s
c
en
d
an
t
grad
i
e
nt
tr
ai
n
i
ng
f
un
c
ti
o
n
an
d
a
f
un
c
ti
on
of
s
i
m
i
l
arit
y
m
ea
s
urem
en
t
b
et
w
e
en
t
he
n
et
w
ork
ou
tp
ut
an
d
th
e
c
os
i
n
e
-
ba
s
ed
tr
ai
n
i
ng
da
ta
.
T
hi
s
m
ea
s
ure
of
s
i
m
i
l
arit
y
as
s
um
es
th
at
the
d
ata
groups
are
v
e
c
tors
an
d
the
i
r
ob
j
ec
ti
v
e
i
s
to
f
i
nd
an
a
ng
l
e
be
t
w
e
en
th
em
,
tha
t
i
s
,
the
ou
tp
ut
(
α)
of
(
3
)
v
ari
es
be
t
ween
-
1
a
nd
1
(
m
ea
ni
ng
t
he
s
am
e)
to
i
nd
i
c
ate
an
de
gree
of
c
orr
el
at
i
on
b
et
w
e
en
v
ec
tors
(
A
an
d
B
)
an
d
i
nd
i
c
ate
s
0
w
he
n t
h
e
v
ec
tor
s
are to
ta
l
l
y
di
f
f
erent (n
=
n
um
be
r
of
v
ec
tor c
om
po
ne
nts
)
.
α
=
c
os
(
)
=
∑
=
1
√
∑
2
=
1
√
∑
2
=
1
(
3)
T
he
tr
ai
ni
ng
of
the
n
et
wor
k
w
as
c
arr
i
ed
ou
t
wi
th
a
grou
p
of
s
am
pl
es
c
on
tai
ni
ng
the
i
nf
orm
ati
on
of
10
00
p
ho
tog
r
ap
hs
of
s
eg
m
en
ted
f
ac
es
as
m
en
ti
on
e
d
i
n
th
e
pre
v
i
ou
s
nu
m
eral
.
T
he
i
nf
or
m
ati
on
i
n
t
he
s
e
p
ho
to
graphs
w
as
gro
up
e
d
i
nto
a
v
ec
tor
w
h
ere
ea
c
h
g
r
a
y
c
om
po
ne
nt
es
ti
m
ate
d
w
i
t
h
t
he
hi
s
to
gra
m
be
c
o
m
es
an
att
r
i
b
ute
an
d
ea
c
h
grou
p
of
f
ou
r
(
4)
i
m
ag
es
b
ec
om
e
s
an
i
ns
t
an
c
e
s
ho
wn
i
n
T
ab
l
e
1.
Ho
w
e
v
er,
when
t
he
i
m
a
ge
c
an
n
ot
be
f
ul
l
y
s
e
gm
en
ted
,
a
v
a
l
ue
of
z
ero (
0)
i
s
as
s
i
g
ne
d
to
e
ac
h c
orr
es
po
nd
i
n
g a
ttr
i
bu
te.
B
y
gro
u
pi
ng
th
e
i
np
ut
da
t
a,
a
n
e
w
att
r
i
bu
t
e
was
c
r
ea
te
d
t
ha
t
i
s
as
s
oc
i
ate
d
t
o
e
ac
h
i
ns
tan
c
e
as
t
he
em
oti
on
tha
t
the
pe
r
s
on
f
ee
l
s
at
the
t
i
m
e
of
c
ap
turi
n
g
t
he
ph
o
tog
r
a
ph
.
T
hi
s
i
ns
tan
c
e
e
nc
od
es
t
he
em
oti
on
s
wi
th
f
ou
r
(
4)
i
nt
eg
ers
(
0
=
Neut
r
a
l
,
1
=
H
ap
p
y
,
2
=
S
ad
,
3
=
A
n
g
r
y
)
,
i
n
ord
er
t
o
c
on
v
ert
t
hi
s
prob
l
em
of
c
l
as
s
i
f
y
i
ng
p
ho
to
grap
hs
i
nto
a
po
l
y
n
om
i
al
ap
prox
i
m
ati
on
prob
l
em
.
T
hi
s
ap
pro
ac
h
w
as
ge
ne
r
a
ted
b
y
tr
ai
ni
n
g
the
D
NN
ac
c
ordi
ng
to
the
s
tr
uc
ture
of
F
i
gu
r
e
2
whi
c
h
h
as
10
2
4
en
tr
i
es
g
en
erat
ed
b
y
ea
c
h
hi
s
to
gra
m
an
d
a
us
er
de
f
i
ne
d
ou
t
pu
t.
F
i
na
l
l
y
,
th
e
c
om
bi
na
t
i
on
of
i
m
ag
e
s
eg
m
en
tat
i
on
s
tr
ate
g
i
es
an
d
th
e
ge
ne
r
at
i
on
of
the
ne
ura
l
ne
t
wor
k
al
l
o
wed
the
de
v
e
l
op
m
en
t
of
a
n
a
p
pl
i
c
a
ti
o
n.
T
hi
s
a
pp
l
i
c
at
i
o
n
a
l
l
o
w
s
the
us
er
to
i
m
po
r
t
groups
of
i
m
ag
es
an
d
ap
pl
y
the
Ha
arc
as
c
ad
e
al
go
r
i
thm
t
o
ea
c
h
i
m
ag
e
to
i
d
en
t
i
f
y
the
f
ac
e
an
d
ea
c
h
of
th
e
r
eg
i
on
s
of
i
nt
eres
t
[
5].
T
he
n
a
ge
om
etri
c
tr
an
s
f
or
m
ati
on
was
m
ad
e
to
ex
tr
ac
t
the
r
eg
i
o
ns
of
i
nte
r
e
s
t
as
i
nd
i
v
i
du
al
i
m
ag
es
an
d
l
ate
r
t
he
r
ou
t
i
n
e
es
ti
m
ate
s
the
h
i
s
tog
r
am
of
ea
c
h
s
eg
m
en
t.
O
nc
e
al
l
the
i
m
ag
es
ha
v
e
b
ee
n
en
c
od
ed
,
the
y
are
grou
pe
d
i
n
a
da
t
ab
as
e
an
d
the
DNN
i
s
tr
a
i
ne
d.
T
he
m
od
el
of
the
DNN
i
s
ex
p
orted
as
a
f
un
c
ti
on
,
w
h
i
c
h
wor
k
s
w
i
th
a
graph
i
c
al
i
nte
r
f
ac
e
tha
t
a
l
l
o
w
s
to
i
m
po
r
t
an
i
m
ag
e
or
tak
e
a
ph
oto
grap
h
to
k
no
w
th
e
m
oo
d
of
the
p
ers
on
.
T
ab
l
e
1.
G
r
o
up
i
ng
of
S
eg
m
en
te
d I
m
ag
es
A
t
t
r
ibu
t
e
s
Fac
e
M
o
u
t
h
R
igh
t
e
y
e
L
e
f
t
e
y
e
I
n
s
t
a
n
c
e
s
1
[
1
,
2
,
…
,
256
]
[
257
,
258
,
…
,
512
]
[
513
,
514
,
…
,
768
]
[
769
,
770
,
…
,
1
0
2
4
]
2
[
1
,
2
,
…
,
256
]
[
257
,
258
,
…
,
512
]
[
513
,
514
,
…
,
768
]
[
769
,
770
,
…
,
1
0
2
4
]
3
[
1
,
2
,
…
,
256
]
[
257
,
258
,
…
,
512
]
[
513
,
514
,
…
,
768
]
[
769
,
770
,
…
,
1
0
2
4
]
…
…
[
1
,
2
,
…
,
256
]
[
257
,
258
,
…
,
512
]
[
513
,
514
,
…
,
768
]
[
769
,
770
,
…
,
1
0
2
4
]
F
i
gu
r
e
2.
B
l
oc
k
di
a
g
r
am
of
t
he
DNN
tr
ai
ni
ng
pr
oc
es
s
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
Rec
og
n
i
ti
on
s
y
s
tem
for fac
i
al
ex
pres
s
i
on
by
proc
es
s
i
n
g i
m
ag
es
..
. (
Ho
l
ma
n
Mo
nt
i
e
l
A
r
i
z
a
)
2979
A
l
g
orit
hm
1.
A
pp
l
i
c
at
i
o
n
De
v
e
l
op
ed
1.
DNN Program()
2.
MA
Import image database ().
3.
MA
Apply Haarcascade algorithm (MA).
4.
MS
Segment the images (MA).
5.
HI
Create histograms (MS)
6.
DN
Define DNN
.
7.
DN
Train DNN (DN, HI)
.
8.
Export Trained Model (DN).
/***
Function to evaluate mood
***/
9.
Evaluate_Image ()
10.
DN
Import Trained Model ().
11.
IM
Read Image ().
12.
IM
Segment and create Histogram (IM).
13.
SA <
-
Evaluate Image using DNN
(EM, DN).
14.
SA <
-
Floor Function (SA).
15.
If SA = 0 then Print "Neutral"
16.
Else If SA = 1 then Print "Happy"
17.
Else if SA = 2
then Print "Sad"
18.
Else if SA = 3
then Print "Angry"
19.
End DNN
3.
E
xper
i
men
t
T
he
ap
p
l
i
c
at
i
o
n
pres
en
t
e
d
i
n
th
i
s
p
ap
er
w
as
m
ad
e
us
i
n
g
t
he
l
i
brar
i
es
K
E
RA
S
,
T
E
NS
O
RF
LO
W
an
d
O
P
E
N
CV
3.
4.0
of
P
Y
T
HO
N
3.5
.8
i
n
th
e
E
c
l
i
ps
e
IDE
4.9
.
0
i
n
terpr
ete
r
wi
th
the
h
el
p
of
the
P
y
D
e
v
th
i
r
d
pa
r
t
y
ad
d
-
on
.
A
no
the
r
f
ea
t
ure
of
the
a
pp
l
i
c
at
i
on
i
s
th
a
t
i
t
w
as
t
e
s
ted
on
a
c
om
pu
ter
w
i
th
an
In
tel
®
i
ns
i
de
CO
RE
T
M
i
3
pr
oc
es
s
or
an
d
8
G
B
of
RA
M.
I
n
a
dd
i
ti
on
,
the
a
pp
l
i
c
at
i
on
w
as
v
al
i
d
at
ed
us
i
ng
a
s
et
of
i
m
ag
es
av
a
i
l
ab
l
e
i
n
a
r
ep
os
i
tor
y
[2
6].
T
he
s
et
of
i
m
ag
es
c
on
s
i
s
ts
of
50
26
p
ho
to
graphs
of
pe
op
l
e
of
di
ff
erent
ag
es
,
ge
nd
er
a
nd
r
ac
e,
s
om
e
o
f
whi
c
h
us
e
ac
c
es
s
orie
s
s
uc
h
as
g
l
as
s
es
or
m
on
oc
l
es
t
o
r
e
du
c
e
the
ef
f
ec
ti
v
en
es
s
of
the
r
ec
og
n
i
t
i
on
a
l
go
r
i
thm
.
In
ad
di
t
i
o
n,
ea
c
h
pe
r
s
on
w
as
ph
o
tog
r
a
ph
e
d
s
e
v
er
al
ti
m
es
wi
th
an
or
i
en
t
ati
on
(
r
i
gh
t
s
ag
i
tt
al
,
l
ef
t
s
ag
i
tta
l
,
f
r
on
ta
l
)
a
nd
r
es
o
l
ut
i
on
(
32
x
30
,
64
x
60
,
1
20
x
1
28
)
di
f
f
erent
s
ho
w
n
i
n
T
ab
l
e
2.
In
th
i
s
pa
p
er
a
s
et
of
92
3
i
m
ag
es
w
as
us
ed
f
r
om
the
i
m
ag
e
da
tab
as
e
tha
t
ha
v
e
a
l
ab
el
th
at
i
nd
i
c
ate
s
the
m
oo
d
of
the
pe
r
s
on
,
a
f
r
on
tal
ori
en
t
ati
o
n
an
d
d
i
f
f
erent
r
es
ol
ut
i
on
.
F
r
om
the
9
23
i
m
ag
es
,
60
0
wer
e
us
e
d
t
o
tr
ai
n
the
DNN
an
d
th
e
r
es
t
w
ere
us
ed
to
c
he
c
k
th
e res
ul
ts
pro
v
i
d
ed
b
y
t
he
D
NN o
nc
e t
r
a
i
ne
d.
Ta
bl
e 2
.
E
x
am
pl
es
of
th
e
P
ho
to
graphs
of
th
e B
as
e o
f
G
r
a
y
s
c
a
l
e I
m
ag
es
3
2
x
3
0
6
4
x
6
0
1
2
0
x
1
2
8
1
2
0
x
1
2
8
S
a
d
A
n
g
r
y
N
e
u
t
r
a
l
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 16
93
-
6
93
0
T
E
L
KO
M
NIK
A
V
ol
.
17
,
No
.
6,
D
ec
em
be
r
20
19
:
29
7
5
-
2982
2980
4
.
Result
s
A
s
m
en
ti
on
ed
be
f
ore,
th
e
propos
e
d
s
tr
ate
g
y
s
e
g
m
en
ts
the
i
m
ag
e
(
r
eg
ardl
es
s
o
f
r
es
ol
ut
i
on
)
i
nt
o
s
e
v
era
l
pa
r
t
s
an
d
e
ac
h
o
ne
be
c
om
es
a
hi
s
to
gram
tha
t
r
ep
r
es
en
ts
the
gr
a
y
s
c
al
e
of
the
i
m
ag
e
,
s
ee
T
ab
l
e
3
.
A
l
s
o,
w
he
n
o
ne
c
r
ea
t
es
th
e
hi
s
t
og
r
am
c
an
al
s
o
as
s
i
g
n
a
l
ab
el
t
ha
t
i
nd
i
c
ate
s
the
m
oo
d
of
th
e
pe
r
s
on
.
F
i
na
l
l
y
,
s
ev
eral
c
on
f
i
gu
r
at
i
o
ns
of
th
e
ne
ura
l
ne
t
wor
k
w
er
e
tes
ted
to
ev
al
ua
te
whi
c
h
i
s
the
m
os
t
ap
propr
i
ate
wh
en
tr
y
i
ng
t
o
s
ol
v
e
t
hi
s
t
y
p
e
of
probl
em
s
.
In
t
ota
l
,
c
o
nf
i
gu
r
at
i
on
s
w
er
e
tes
t
ed
wi
t
h
i
n
c
r
ea
s
i
ng
of
on
e
h
un
dred
(
10
0)
i
n
o
ne
hu
nd
r
ed
(
10
0)
i
n
ea
c
h
ne
t
wor
k
tr
ai
ni
ng
th
e
nu
m
be
r
of
ne
urons
an
d
o
ne
(
1)
i
n
on
e
(
1)
nu
m
be
r
of
hi
dd
e
n
l
a
y
ers
.
In
t
ota
l
20
DNNs
wi
t
h
di
f
f
er
en
t
top
ol
og
i
es
wer
e
bu
i
l
t
a
nd
t
he
i
r
b
eh
a
v
i
or
i
s
pr
es
en
t
ed
i
n
F
i
g
ure
3
i
n
whi
c
h
t
h
e
s
o
l
i
d
l
i
ne
r
e
pres
en
ts
err
or
a
nd
t
he
d
o
tte
d
l
i
ne
r
ep
r
es
en
ts
th
e
m
argi
n
of
err
or.
T
he
s
e
l
i
ne
s
w
ere
c
a
l
c
ul
a
te
d
ba
s
ed
o
n
the
i
m
ag
es
fr
om
the
i
m
ag
e
da
tab
as
e
,
th
at
i
s
,
the
y
are
the
r
es
ul
ts
when
pe
r
f
or
m
i
ng
10
0
0
i
t
erati
on
s
of
tr
ai
n
i
n
g
an
d
e
v
al
u
ati
ng
th
e
DNN
wi
th
th
e
ba
s
e
of
i
m
ag
es
on
l
y
.
T
ab
l
e 3
.
P
r
oc
es
s
of
G
en
era
ti
on
of
th
e H
i
s
tog
r
am
s
an
d t
he
S
eg
m
en
ts
of
th
e I
m
ag
e
U
n
s
e
g
m
e
n
t
e
d
v
ie
w
Or
igin
a
l
H
a
a
r
c
a
s
c
a
d
e
S
e
g
m
e
n
t
e
d
v
ie
w
S
e
g
m
e
n
t
P
h
o
t
o
g
r
a
m
H
is
t
o
g
r
a
m
Fac
e
Mou
t
h
E
y
e
1
E
y
e
2
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
Rec
og
n
i
ti
on
s
y
s
tem
for fac
i
al
ex
pres
s
i
on
by
proc
es
s
i
n
g i
m
ag
es
..
. (
Ho
l
ma
n
Mo
nt
i
e
l
A
r
i
z
a
)
2981
F
i
gu
r
e
3.
B
eh
av
i
or of
th
e D
NN
w
i
t
h d
i
f
f
erent c
on
f
i
gu
r
at
i
on
s
5
.
Co
n
clus
ion
F
i
gu
r
e
3
s
h
o
w
s
tha
t
i
t
i
s
po
s
s
i
bl
e
f
or
the
pro
po
s
ed
s
tr
a
teg
y
to
r
ed
uc
e
the
m
argi
n
o
f
err
or
b
y
es
t
i
m
ati
ng
w
h
at
the
pe
r
s
on
i
n
the
p
ho
t
og
r
ap
h
i
s
f
ee
l
i
ng
wi
th
a
m
argi
n
of
err
or
c
l
os
e
t
o
t
w
e
nt
y
pe
r
c
en
t
(
20
%)
,
th
a
t
i
s
, t
ha
t
of
ev
er
y
ten
(
10
)
ph
oto
gr
ap
hs
c
an
no
t
pre
di
c
t
th
e
f
ac
i
a
l
ex
pres
s
i
o
n o
f
two
(
2).
T
hi
s
i
s
be
c
a
us
e
i
n
s
om
e
c
as
es
th
e
l
i
gh
t
a
ff
ec
ts
the
d
ete
c
t
i
on
of
th
e
gl
as
s
es
an
d
the
Ha
arc
as
c
ad
e
a
l
g
orit
h
m
i
s
no
t
pe
r
f
ec
t
an
d
d
et
ec
ts
m
ore
tha
n
t
wo
e
y
es
or
m
ore
tha
n
on
e m
ou
th.
O
ne
of
t
he
ad
v
an
tag
es
of
thi
s
em
oti
on
pred
i
c
ti
o
n
m
eth
od
ol
og
y
i
s
th
at
i
t
c
an
be
ge
ne
r
ate
d
f
or
ev
en
m
ore
att
r
i
bu
t
es
,
tha
t
i
s
,
i
f
on
e
h
av
es
a
broad
er
da
t
ab
as
e
t
he
n
c
ou
l
d
pre
-
te
l
l
m
ore
e
m
oti
on
s
.
I
t
c
an
b
e
s
ai
d
th
at
the
m
argi
n
of
err
or
i
s
c
om
pe
ns
ate
d
b
y
t
he
f
l
ex
i
bi
l
i
t
y
of
the
DNN
,
s
i
nc
e,
c
om
pa
r
ed
to
tr
a
di
t
i
o
na
l
c
l
as
s
i
f
i
ers
,
the
pre
di
c
t
i
on
b
y
m
ea
ns
of
po
l
y
n
om
i
al
s
al
l
o
w
s
es
ta
bl
i
s
hi
ng
a
ge
ne
r
al
i
z
e
d
m
od
el
of
DNN
t
o
s
ol
v
e
th
i
s
t
y
p
e
of
probl
em
s
.
In
ad
di
t
i
o
n,
the
DNN
c
ou
l
d
be
ad
j
us
t
ed
to
d
i
f
f
erent
i
m
ag
e
ac
qu
i
s
i
t
i
on
el
em
en
ts
(
c
a
m
eras
or
s
c
an
ne
r
s
)
,
be
c
au
s
e
du
r
i
ng
th
e
tr
a
i
n
i
n
g
th
e
wei
gh
ts
are
a
dj
us
ted
tak
i
ng
i
nt
o
ac
c
ou
nt
i
m
ag
e
s
wi
th
d
i
f
f
erent
r
es
ol
ut
i
on
s
w
h
i
c
h
r
ed
u
c
es
the
ef
f
ec
t
o
f
t
he
s
y
s
tem
ati
c
err
or
i
nd
uc
ed
b
y
the
e
nv
i
r
on
m
en
tal
c
on
d
i
t
i
on
s
.
A
s
c
an
be
s
ee
n
i
n
F
i
g
ure
3
the
DNN
arr
i
v
es
at
a
s
tea
d
y
s
t
ate
f
r
o
m
tr
ai
ni
ng
nu
m
be
r
1
3,
thi
s
m
ea
ns
tha
t
e
v
e
n
i
f
the
nu
m
be
r
of
hi
dd
en
l
a
y
ers
or
ne
uro
ns
i
s
i
nc
r
e
as
ed
,
i
t
i
s
no
t
p
os
s
i
bl
e
to
i
m
prov
e
the
p
erf
or
m
an
c
e
of
the
DNN.
O
n
the
on
e
ha
nd
,
i
t
i
s
po
s
s
i
bl
e
t
ha
t,
b
y
m
od
i
f
y
i
ng
the
f
un
c
ti
o
ns
of
ac
ti
v
at
i
o
n,
op
t
i
m
i
z
at
i
on
an
d
ge
ne
r
ati
o
n
of
the
i
n
i
ti
al
wei
gh
t
s
,
i
t
i
m
prov
es
the
pe
r
f
orm
an
c
e
of
the
D
N
N.
O
n
the
ot
he
r
ha
nd
,
the
o
bj
ec
ti
v
e
of
t
hi
s
pa
p
er
was
t
o
f
i
nd
th
e
m
os
t
ap
pro
pria
t
e
c
on
f
i
g
urati
on
,
s
o
tha
t
th
e
DNN
i
s
ex
ec
ut
ed
wi
th
f
ew
c
om
pu
tat
i
o
na
l
r
es
ou
r
c
es
an
d
am
on
g t
he
e
v
a
l
u
ate
d
c
on
f
i
gu
r
ati
on
s
th
e m
os
t a
de
q
ua
t
e
w
as
th
e o
ne
pr
es
en
te
d i
n
s
ec
ti
on
2.
Ref
er
en
ce
s
[1
]
Neh
ru
M
,
Pa
d
m
a
v
a
t
h
i
S
.
Il
l
u
m
i
n
a
ti
o
n
i
n
v
a
r
i
a
n
t
fa
c
e
d
e
te
c
t
i
o
n
u
s
i
n
g
v
i
o
l
a
j
o
n
e
s
a
l
g
o
r
i
th
m
.
2
0
1
7
4
th
In
te
rn
a
ti
o
n
a
l
Con
f
e
re
n
c
e
o
n
Ad
v
a
n
c
e
d
C
o
m
p
u
ti
n
g
a
n
d
Com
m
u
n
i
c
a
ti
o
n
Sy
s
te
m
s
(
ICACC
S)
.
Coi
m
b
a
to
re
.
2
0
1
7
:
1
-
4.
[2
]
Vi
k
ra
m
K,
Pa
d
m
a
v
a
th
i
S
.
F
a
c
i
a
l
p
a
r
ts
d
e
te
c
ti
o
n
u
s
i
n
g
Vi
o
l
a
J
o
n
e
s
a
l
g
o
r
i
t
h
m
.
2
0
1
7
4
t
h
In
te
rn
a
t
i
o
n
a
l
Con
fe
re
n
c
e
o
n
Ad
v
a
n
c
e
d
Com
p
u
t
i
n
g
a
n
d
Co
m
m
u
n
i
c
a
ti
o
n
Sy
s
te
m
s
(
ICACC
S)
.
Coi
m
b
a
to
re
.
2
0
1
7
:
1
-
4
.
[3
]
Al
y
u
s
h
i
n
M
V,
L
y
u
b
s
h
o
v
AA
.
Th
e
Vi
o
l
a
-
J
o
n
e
s
a
l
g
o
ri
t
h
m
p
e
rfo
rm
a
n
c
e
e
n
h
a
n
c
e
m
e
n
t
fo
r a
p
e
rs
o
n
'
s
f
a
c
e
re
c
o
g
n
i
ti
o
n
ta
s
k
i
n
th
e
l
o
n
g
-
w
a
v
e
i
n
fra
re
d
r
a
d
i
a
ti
o
n
ra
n
g
e
.
2
0
1
8
IEEE
Con
fe
r
e
n
c
e
o
f
Rus
s
i
a
n
Y
o
u
n
g
Res
e
a
r
c
h
e
r
s
i
n
El
e
c
tri
c
a
l
a
n
d
El
e
c
t
r
o
n
i
c
En
g
i
n
e
e
ri
n
g
(EI
Con
Rus
)
.
M
o
s
c
o
w
.
2
0
1
8
:
1
8
1
3
-
1
8
1
6
.
[4
]
Sh
a
m
i
a
D,
Ch
a
n
d
y
DA
.
A
n
a
l
y
z
i
n
g
th
e
p
e
rfo
rm
a
n
c
e
o
f
V
i
o
l
a
J
o
n
e
s
Fa
c
e
Det
e
c
to
r
o
n
th
e
L
DH
F
d
a
ta
b
a
s
e
.
In
te
rn
a
ti
o
n
a
l
Con
f
e
r
e
n
c
e
o
n
Si
g
n
a
l
Pro
c
e
s
s
i
n
g
a
n
d
Com
m
u
n
i
c
a
ti
o
n
(ICSPC)
.
C
o
i
m
b
a
to
re
.
2
0
1
7
:
3
1
2
-
315.
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 16
93
-
6
93
0
T
E
L
KO
M
NIK
A
V
ol
.
17
,
No
.
6,
D
ec
em
be
r
20
19
:
29
7
5
-
2982
2982
[5
]
Dja
m
a
l
u
d
d
i
n
D
,
In
d
ra
b
u
l
a
n
T
,
An
d
a
n
i
,
In
d
r
a
b
a
y
u
,
Si
d
e
h
a
b
i
SW
.
Th
e
s
i
m
u
l
a
t
i
o
n
o
f
v
e
h
i
c
l
e
c
o
u
n
t
i
n
g
s
y
s
te
m
fo
r
tra
f
fi
c
s
u
rv
e
i
l
l
a
n
c
e
u
s
i
n
g
Vi
o
l
a
J
o
n
e
s
m
e
th
o
d
.
2
0
1
4
M
a
k
a
s
s
a
r
In
te
r
n
a
ti
o
n
a
l
Co
n
fe
re
n
c
e
o
n
El
e
c
tr
i
c
a
l
En
g
i
n
e
e
ri
n
g
a
n
d
I
n
f
o
rm
a
ti
c
s
(M
ICEEI)
.
M
a
k
a
s
s
a
r
.
2
0
1
4
:
1
3
0
-
1
3
5
.
[6
]
G
u
o
j
i
n
C,
Y
o
n
g
n
i
n
g
L
,
M
i
a
o
fe
n
Z,
W
a
n
q
i
a
n
g
W
.
Th
e
i
m
a
g
e
a
u
to
-
fo
c
u
s
i
n
g
m
e
th
o
d
b
a
s
e
d
o
n
a
rt
i
fi
c
i
a
l
n
e
u
ra
l
n
e
two
r
k
s
.
2
0
1
0
I
E
EE
In
te
rn
a
ti
o
n
a
l
Con
fe
re
n
c
e
o
n
Com
p
u
ta
ti
o
n
a
l
I
n
te
l
l
i
g
e
n
c
e
fo
r
M
e
a
s
u
re
m
e
n
t
Sy
s
te
m
s
a
n
d
A
p
p
l
i
c
a
t
i
o
n
s
.
T
a
ra
n
to
.
2
0
1
0
:
1
3
8
-
1
4
1
.
[7
]
O
m
e
r
M
K,
Sh
e
ta
O
E,
Ad
re
e
s
M
S,
Sti
a
w
a
n
D,
Riy
a
d
i
M
A,
Bu
d
i
a
rto
R
.
Dee
p
Ne
u
ra
l
N
e
tw
o
rk
fo
r
He
a
rt
Dis
e
a
s
e
M
e
d
i
c
a
l
Pre
s
c
ri
p
ti
o
n
E
x
p
e
rt
Sy
s
te
m
.
In
d
o
n
e
s
i
a
n
J
o
u
rn
a
l
o
f
El
e
c
tri
c
a
l
E
n
g
i
n
e
e
ri
n
g
a
n
d
In
fo
rm
a
ti
c
s
.
2
0
1
8
;
6
(
2
):
2
1
7
-
2
2
4
.
[8
]
Zh
a
n
g
D,
Ha
n
X
,
Den
g
C
.
Re
v
i
e
w
o
n
th
e
re
s
e
a
r
c
h
a
n
d
p
ra
c
ti
c
e
o
f
d
e
e
p
l
e
a
rn
i
n
g
a
n
d
re
i
n
fo
rc
e
m
e
n
t
l
e
a
rn
i
n
g
i
n
s
m
a
r
t
g
ri
d
s
.
J
o
u
rn
a
l
o
f
Po
we
r a
n
d
En
e
r
g
y
S
y
s
t
e
m
s
.
2
0
1
8
;
4
(3
):
3
6
2
-
3
7
0
.
[9
]
G
u
o
j
i
n
C,
M
i
a
o
fe
n
Z,
Hon
g
h
a
o
Y
,
Y
a
n
L
.
Ap
p
l
i
c
a
t
i
o
n
o
f
Neu
ra
l
Net
wor
k
s
i
n
Im
a
g
e
Def
i
n
i
ti
o
n
Rec
o
g
n
i
ti
o
n
.
2
0
0
7
IEEE
In
t
e
rn
a
ti
o
n
a
l
Co
n
fe
re
n
c
e
o
n
S
i
g
n
a
l
Pro
c
e
s
s
i
n
g
a
n
d
Co
m
m
u
n
i
c
a
ti
o
n
s
.
Dub
a
i
.
2
0
0
7
:
1
2
0
7
-
1
2
1
0
.
[1
0
]
Ki
m
HI,
L
e
e
SH,
Ro
YM
.
Fa
c
e
i
m
a
g
e
a
s
s
e
s
s
m
e
n
t
l
e
a
r
n
e
d
wit
h
o
b
j
e
c
ti
v
e
a
n
d
re
l
a
ti
v
e
fa
c
e
i
m
a
g
e
q
u
a
l
i
ti
e
s
fo
r
i
m
p
ro
v
e
d
fa
c
e
re
c
o
g
n
i
ti
o
n
.
2
0
1
5
IEEE
In
te
rn
a
t
i
o
n
a
l
Con
fe
r
e
n
c
e
o
n
Im
a
g
e
P
ro
c
e
s
s
i
n
g
(ICIP)
.
Q
u
e
b
e
c
.
2
0
1
5
:
4
0
2
7
-
4
0
3
1
.
[1
1
]
P
Sri
n
i
v
a
s
a
,
R
K
Na
d
e
s
h
,
N
C
Se
n
th
i
l
.
Ro
b
u
s
t
Fa
c
e
Re
c
o
g
n
i
t
i
o
n
Us
i
n
g
En
h
a
n
c
e
d
L
o
c
a
l
B
i
n
a
r
y
Pa
tt
e
rn
.
B
u
l
l
e
ti
n
o
f
E
l
e
c
tri
c
a
l
E
n
g
i
n
e
e
ri
n
g
a
n
d
I
n
f
o
rm
a
ti
c
s
.
2
0
1
8
;
7
(1
):
9
6
-
101.
[1
2
]
Su
d
h
a
k
a
r
K,
Nit
h
y
a
n
a
n
d
a
m
P
.
An
Ac
c
u
r
a
te
Fa
c
i
a
l
Com
p
o
n
e
n
t
Det
e
c
ti
o
n
U
s
i
n
g
G
a
b
o
r
Fi
l
te
r.
Bu
l
l
e
ti
n
o
f
El
e
c
tri
c
a
l
E
n
g
i
n
e
e
r
i
n
g
a
n
d
I
n
fo
rm
a
ti
c
s
.
2
0
1
7
;
6
(3
)
:
2
8
7
-
2
9
4
.
[1
3
]
Ba
l
tru
š
a
i
t
i
s
T
,
Rob
i
n
s
o
n
P,
M
o
re
n
c
y
L
P
.
O
p
e
n
F
a
c
e
:
An
o
p
e
n
s
o
u
r
c
e
fa
c
i
a
l
b
e
h
a
v
i
o
r
a
n
a
l
y
s
i
s
to
o
l
k
i
t
.
2
0
1
6
I
EEE
W
i
n
te
r Co
n
fe
re
n
c
e
o
n
Ap
p
l
i
c
a
ti
o
n
s
o
f
Co
m
p
u
te
r
Vi
s
i
o
n
(
W
A
CV)
.
L
a
k
e
Pl
a
c
i
d
.
2
0
1
6
:
1
-
10.
[1
4
]
Fe
n
g
R,
L
e
u
n
g
CS,
Su
m
J
,
X
i
a
o
Y
.
Pro
p
e
rti
e
s
a
n
d
Pe
rfo
rm
a
n
c
e
o
f
I
m
p
e
rf
e
c
t
D
u
a
l
Ne
u
ra
l
Net
w
o
rk
-
b
a
s
e
d
$
k
$
W
T
A
Ne
tw
o
rk
s
.
IEEE
Tra
n
s
a
c
ti
o
n
s
o
n
Neu
ra
l
Net
wor
k
s
a
n
d
L
e
a
rn
i
n
g
Sy
s
te
m
s
.
2
0
1
5
;
2
6
(9
)
:
2
1
8
8
-
2
1
9
3
.
[1
5
]
Chu
W
S
,
De
l
a
T
o
rre
F,
C
o
h
n
J
F
.
S
e
l
e
c
ti
v
e
T
ra
n
s
fe
r
M
a
c
h
i
n
e
fo
r
Pe
r
s
o
n
a
l
i
z
e
d
Fa
c
i
a
l
E
x
p
re
s
s
i
o
n
An
a
l
y
s
i
s
.
IEEE
Tr
a
n
s
a
c
t
i
o
n
s
o
n
Pa
tt
e
r
n
An
a
l
y
s
i
s
a
n
d
M
a
c
h
i
n
e
I
n
te
l
l
i
g
e
n
c
e
.
2
0
1
7
;
3
9
(
3
):
5
2
9
-
5
4
5
.
[1
6
]
Ze
n
G
,
Po
rz
i
L
,
Sa
n
g
i
n
e
to
E,
Ric
c
i
E,
Se
b
e
N
.
L
e
a
rn
i
n
g
Pe
rs
o
n
a
l
i
z
e
d
M
o
d
e
l
s
fo
r
F
a
c
i
a
l
E
x
p
re
s
s
i
o
n
An
a
l
y
s
i
s
a
n
d
Ges
t
u
re
Re
c
o
g
n
i
ti
o
n
.
IEEE
Tra
n
s
a
c
ti
o
n
s
o
n
M
u
l
ti
m
e
d
i
a
.
2
0
1
6
;
1
8
(4
):
7
7
5
-
7
8
8
.
[1
7
]
Zh
i
q
i
Y
.
G
e
s
t
u
re
l
e
a
rn
i
n
g
a
n
d
re
c
o
g
n
i
ti
o
n
b
a
s
e
d
o
n
th
e
Ch
e
b
y
s
h
e
v
p
o
l
y
n
o
m
i
a
l
n
e
u
ra
l
n
e
t
work
.
2
0
1
6
IEEE
In
fo
r
m
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
,
Net
w
o
rk
i
n
g
,
El
e
c
tro
n
i
c
a
n
d
Au
to
m
a
t
i
o
n
Con
tro
l
Con
fe
re
n
c
e
.
Cho
n
g
q
i
n
g
.
2
0
1
6
:
9
3
1
-
9
3
4
.
[
18
]
L
e
e
D,
L
e
e
J
.
Eq
u
i
l
i
b
r
i
u
m
-
b
a
s
e
d
s
u
p
p
o
r
t
v
e
c
to
r
m
a
c
h
i
n
e
fo
r
s
e
m
i
s
u
p
e
rv
i
s
e
d
c
l
a
s
s
i
fi
c
a
t
i
o
n
.
IEE
E
Tra
n
s
.
Ne
u
r
a
l
Ne
tw
.
2
0
0
7
;
1
8
(
2
):
5
7
8
-
5
8
3
.
[
19
]
T
e
rre
n
c
e
J
.
T
h
e
De
e
p
L
e
a
r
n
i
n
g
Re
v
o
l
u
ti
o
n
.
T
h
e
M
IT
Pre
s
s
.
2
0
1
8
:
1
-
1
0
.
[2
0
]
Vi
g
n
e
s
w
a
ra
n
KR,
Vi
n
a
y
a
k
u
m
a
r
R,
So
m
a
n
KP,
Po
o
rn
a
c
h
a
n
d
ra
n
P
.
E
v
a
l
u
a
t
i
n
g
Sh
a
l
l
o
w
a
n
d
Dee
p
Neu
ra
l
N
e
two
rk
s
fo
r
Ne
two
rk
In
tru
s
i
o
n
Det
e
c
ti
o
n
S
y
s
t
e
m
s
i
n
Cy
b
e
r
Se
c
u
r
i
ty
.
2
0
1
8
9
th
In
t
e
rn
a
ti
o
n
a
l
Con
fe
re
n
c
e
o
n
Com
p
u
ti
n
g
,
Com
m
u
n
i
c
a
ti
o
n
a
n
d
N
e
tw
o
rk
i
n
g
T
e
c
h
n
o
l
o
g
i
e
s
(ICCC
NT
)
.
Ba
n
g
a
l
o
re
.
2
0
1
8
:
1
-
6.
[2
1
]
Sa
a
d
M
M
,
J
a
m
i
l
N,
Ha
m
z
a
h
R
.
Ev
a
l
u
a
t
i
o
n
o
f
Su
p
p
o
rt
Ve
c
to
r
M
a
c
h
i
n
e
a
n
d
Dec
i
s
i
o
n
T
re
e
f
o
r
Em
o
ti
o
n
Rec
o
g
n
i
ti
o
n
o
f
M
a
l
a
y
Fo
l
k
l
o
re
s
.
Bu
l
l
e
ti
n
o
f
El
e
c
tri
c
a
l
En
g
i
n
e
e
ri
n
g
a
n
d
I
n
fo
rm
a
t
i
c
s
.
2
0
1
8
;
7
(3
):
4
7
9
-
4
8
6
.
[2
2
]
Po
v
o
d
a
L
.
Se
n
ti
m
e
n
t
a
n
a
l
y
s
i
s
b
a
s
e
d
o
n
Su
p
p
o
rt
Ve
c
t
o
r
M
a
c
h
i
n
e
a
n
d
Bi
g
Dat
a
.
39
th
IEE
E
In
te
rn
a
ti
o
n
a
l
Con
fe
r
e
n
c
e
o
n
T
e
l
e
c
o
m
m
u
n
i
c
a
t
i
o
n
s
a
n
d
Si
g
n
a
l
Pr
o
c
e
s
s
i
n
g
(T
SP)
.
Vi
e
n
n
a
.
2
0
1
6
:
5
4
3
-
545.
[2
3
]
Ka
u
r
H,
M
a
n
g
a
t
V
.
A
s
u
r
v
e
y
o
f
s
e
n
ti
m
e
n
t
a
n
a
l
y
s
i
s
te
c
h
n
i
q
u
e
s
.
2
0
1
7
In
te
r
n
a
ti
o
n
a
l
Co
n
fe
re
n
c
e
o
n
In
I
-
SM
AC (Io
T
i
n
So
c
i
a
l
,
M
o
b
i
l
e
,
An
a
l
y
ti
c
s
a
n
d
Cl
o
u
d
(I
-
SM
AC)
.
In
d
i
a
.
2
0
1
7
:
9
2
1
-
925.
[2
4
]
Ad
e
g
e
AB
.
Ap
p
l
y
i
n
g
Dee
p
Neu
ra
l
Net
work
(DN
N)
fo
r
l
a
rg
e
-
s
c
a
l
e
i
n
d
o
o
r
l
o
c
a
l
i
z
a
ti
o
n
u
s
i
n
g
fe
e
d
-
fo
rward
n
e
u
r
a
l
n
e
two
r
k
(FF
NN
)
a
l
g
o
ri
t
h
m
.
2
0
1
8
IEE
E
In
te
r
n
a
ti
o
n
a
l
Co
n
fe
re
n
c
e
o
n
A
p
p
l
i
e
d
Sy
s
te
m
I
n
v
e
n
ti
o
n
(ICASI)
.
Ch
i
b
a
.
2
0
1
8
:
8
1
4
-
817.
[2
5
]
Ze
g
e
r
s
P,
Su
n
d
a
r
e
s
h
a
n
M
K
.
T
ra
j
e
c
to
ry
g
e
n
e
ra
t
i
o
n
a
n
d
m
o
d
u
l
a
t
i
o
n
u
s
i
n
g
d
y
n
a
m
i
c
n
e
u
r
a
l
n
e
tw
o
rk
s
.
IEEE
Tra
n
s
a
c
t
i
o
n
s
o
n
Ne
u
ra
l
Net
work
s
.
2
0
0
3
;
1
4
(3
):
5
2
0
-
5
3
3
.
[2
6
]
D
Dua
,
K
T
a
n
i
s
k
i
d
o
u
.
UC
I
M
a
c
h
i
n
e
L
e
a
rn
i
n
g
Rep
o
s
i
t
o
ry
[h
tt
p
:/
/
a
rc
h
i
v
e
.i
c
s
.
u
c
i
.e
d
u
/m
l
].
I
rv
i
n
e
,
CA:
Uni
v
e
rs
i
ty
o
f
Ca
l
i
f
o
rn
i
a
,
Sc
h
o
o
l
o
f
In
f
o
rm
a
ti
o
n
a
n
d
Co
m
p
u
te
r
Sc
i
e
n
c
e
.
2
0
1
7
.
Evaluation Warning : The document was created with Spire.PDF for Python.