I
AE
S
I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
Ar
t
if
icial
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
Vol.
14
,
No.
4
,
Augus
t
2025
,
pp.
3287
~
3299
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
4
.
pp
32
87
-
3299
3287
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
C
la
ssi
f
ic
at
io
n
of
T
asi
k
m
al
aya
b
at
ik
m
ot
ifs u
si
n
g
c
o
n
vol
u
t
io
n
al
n
e
u
r
al
n
e
t
w
o
r
k
s
T
e
u
k
u
M
u
f
izar
1
,
As
o
S
u
d
iarj
o
1
,
E
vi
De
wi
S
r
i
M
u
lyan
i
1
,
Agus
A
h
m
ad
Wak
ih
2
,
M
u
h
am
m
ad
Akb
ar
Kas
yf
u
r
r
ah
m
an
1
,
L
u
t
h
f
i
A
d
il
al
M
ah
b
u
b
1
1
D
e
pa
r
tm
e
nt
of
I
nf
or
ma
ti
c
s
E
ngi
ne
e
r
in
g, F
a
c
ul
ty
of
E
ngi
ne
e
r
in
g, P
e
r
ju
a
nga
n T
a
s
ik
ma
la
y
a
U
ni
ve
r
s
it
y, T
a
s
ik
ma
la
ya
, I
ndone
s
ia
2
D
e
pa
r
te
me
nt
of
P
r
im
a
r
y E
duc
a
ti
on, F
a
c
ul
ty
of
E
duc
a
ti
ona
l
S
c
i
e
nc
e
s
, P
e
r
ju
a
nga
n
T
a
s
ik
ma
l
a
ya
U
ni
ve
r
s
it
y,
T
a
s
ik
ma
la
y
a
, I
ndone
s
ia
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
S
e
p
29,
2024
R
e
vis
e
d
J
un
9,
2025
Ac
c
e
pted
J
ul
9,
2025
T
h
i
s
p
a
p
er
p
res
e
n
t
s
a
s
t
u
d
y
o
n
t
h
e
cl
a
s
s
i
fi
ca
t
i
o
n
o
f
t
ra
d
i
t
i
o
n
a
l
T
as
i
k
ma
l
ay
a
b
at
i
k
m
o
t
i
fs
u
s
i
n
g
co
n
v
o
l
u
t
i
o
n
al
n
eu
ra
l
n
e
t
w
o
rk
s
(CN
N
).
T
h
e
e
x
p
er
i
men
t
s
rev
eal
e
d
t
h
a
t
t
h
e
h
i
g
h
co
mp
l
ex
i
t
y
o
f
b
at
i
k
mo
t
i
f
s
s
i
g
n
i
fi
ca
n
t
l
y
i
m
p
act
e
d
mo
d
e
l
p
erfo
rma
n
ce,
as
t
h
e
h
a
n
d
l
i
n
g
o
f
each
c
l
as
s
i
n
fl
u
e
n
ced
t
h
e
o
v
eral
l
res
u
l
t
s
.
In
i
t
i
al
e
x
p
er
i
men
t
s
w
i
t
h
t
h
e
o
r
i
g
i
n
al
d
a
t
as
e
t
d
em
o
n
s
t
ra
t
ed
s
u
b
o
p
t
i
ma
l
p
erfo
rma
n
ce,
ch
aract
er
i
zed
b
y
accu
rac
y
an
d
v
a
l
i
d
at
i
o
n
cu
rv
e
s
i
n
d
i
ca
t
i
n
g
o
v
erf
i
t
t
i
n
g
,
w
i
t
h
o
n
l
y
7
5
%
accu
rac
y
ach
i
ev
e
d
at
a
l
earn
i
n
g
rat
e
o
f
0
.
0
0
1
,
a
b
at
c
h
s
i
ze
o
f
3
2
,
an
d
5
0
ep
o
c
h
s
.
T
o
en
h
a
n
ce
p
erfo
rma
n
ce,
w
e
i
mp
l
eme
n
t
e
d
d
at
a
s
eg
me
n
t
a
t
i
o
n
,
d
at
a
a
u
g
me
n
t
a
t
i
o
n
,
o
p
t
i
m
i
zed
t
h
e
ch
o
i
ce
o
f
t
h
e
b
e
s
t
o
p
t
i
mi
zer,
u
t
i
l
i
zed
an
o
p
t
i
ma
l
arch
i
t
ect
u
re,
an
d
c
o
n
d
u
ct
ed
h
y
p
er
p
aramet
e
r
t
u
n
i
n
g
.
T
h
e
b
e
s
t
-
p
erfo
rm
i
n
g
mo
d
el
w
a
s
t
ra
i
n
e
d
o
n
d
a
t
a
s
u
b
j
ec
t
ed
t
o
s
p
eci
f
i
c
p
rep
r
o
ces
s
i
n
g
f
o
r
each
cl
a
s
s
,
u
s
i
n
g
t
h
e
A
d
am
o
p
t
i
m
i
zer
w
i
t
h
h
y
p
er
p
arame
t
er
t
u
n
i
n
g
s
et
t
o
a
l
earn
i
n
g
rat
e
o
f
0
.
0
0
1
,
a
b
at
c
h
s
i
ze
o
f
3
2
,
an
d
5
0
e
p
o
c
h
s
.
I
n
t
h
e
h
y
p
er
p
aramet
er
t
u
n
i
n
g
e
x
p
er
i
men
t
w
i
t
h
t
h
e
v
i
s
u
al
g
e
o
m
et
r
y
g
r
o
u
p
n
et
w
o
r
k
(
V
G
G
N
et
)
arch
i
t
ect
u
re,
i
t
w
as
s
h
o
w
n
t
h
at
t
h
er
e
i
s
an
i
mp
r
o
v
em
en
t
i
n
t
h
e
p
re
d
i
c
t
i
o
n
o
f
t
h
e
ku
m
el
i
cl
a
s
s
,
ach
i
ev
i
n
g
an
accu
ra
cy
o
f
1
0
0
%
.
K
e
y
w
o
r
d
s
:
B
a
ti
k
mot
if
s
C
las
s
if
ica
ti
on
C
onvolut
ional
ne
ur
a
l
ne
twor
ks
C
omput
e
r
vis
ion
I
mage
pr
oc
e
s
s
ing
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
E
vi
De
wi
S
r
i
M
ulyani
De
pa
r
tm
e
nt
of
I
nf
or
mat
ics
E
nginee
r
ing
,
F
a
c
ult
y
of
E
nginee
r
ing
,
P
e
r
juanga
n
T
a
s
ikm
a
laya
Unive
r
s
it
y
P
E
T
A
S
t,
No
177
,
T
a
s
ikm
a
laya
,
W
e
s
t
J
a
va
,
I
ndon
e
s
ia
E
mail:
e
viaja
de
c
h
@gmail.
c
om
1.
I
NT
RODU
C
T
I
ON
B
a
t
ik
,
a
t
r
a
d
i
ti
on
a
l
I
nd
on
e
s
i
a
n
te
xt
i
le
a
r
t
,
is
r
e
no
wne
d
f
o
r
i
ts
in
t
r
i
c
a
te
de
s
ig
ns
a
nd
c
u
l
tu
r
a
l
s
ig
n
if
ic
a
nc
e
,
e
s
pe
c
ia
l
ly
i
n
the
T
a
s
ik
ma
lay
a
r
e
gi
on
,
w
hi
c
h
is
c
ha
r
a
c
t
e
r
iz
e
d
b
y
u
n
iq
ue
mo
t
if
s
t
ha
t
r
e
f
lec
t
l
oc
a
l
he
r
it
a
g
e
a
nd
a
r
ti
s
t
r
y
.
As
g
lo
ba
li
z
a
ti
on
c
ha
l
le
nge
s
th
e
p
r
e
s
e
r
va
t
ion
o
f
s
u
c
h
c
u
lt
u
r
a
l
i
de
nt
it
ies
,
le
ve
r
a
gi
ng
te
c
h
no
lo
gy
b
e
c
om
e
s
e
s
s
e
n
ti
a
l
to
s
a
f
e
gu
a
r
d
a
n
d
p
r
o
m
ot
e
th
e
s
e
a
r
t
f
o
r
m
s
.
R
e
c
e
n
t
a
d
va
n
c
e
men
ts
i
n
ma
c
h
in
e
le
a
r
ni
ng
,
pa
r
ti
c
u
la
r
ly
c
on
vo
lu
t
io
na
l
n
e
u
r
a
l
ne
tw
or
ks
(
C
NN
s
)
,
o
f
f
e
r
p
r
omi
s
i
ng
s
o
lu
ti
ons
f
o
r
a
u
to
ma
ti
ng
t
he
c
la
s
s
i
f
ica
ti
on
o
f
c
om
p
lex
b
a
t
i
k
m
ot
i
f
s
,
t
he
r
e
by
e
n
ha
n
c
i
ng
th
e
a
c
c
e
s
s
ib
il
i
ty
a
nd
a
p
p
r
e
c
ia
ti
on
o
f
t
hi
s
c
u
l
tu
r
a
l
t
r
e
a
s
u
r
e
.
De
e
p
l
e
a
r
n
ing
may
be
a
ki
nd
o
f
a
n
a
r
t
if
i
c
ial
in
tell
i
ge
nc
e
t
ha
t
ha
s
r
e
c
e
nt
ly
a
c
hieve
d
f
a
ntas
ti
c
s
uc
c
e
s
s
in
c
las
s
i
f
ica
t
ion
a
n
d
s
e
gmen
tati
on
tas
ks
[
1]
–
[
3]
,
a
tt
r
a
c
ti
n
g
a
lo
t
of
a
tt
e
n
ti
on
in
the
tr
e
a
tm
e
n
t
o
f
n
umer
ous
dis
or
de
r
s
[
4
]
.
C
NN
ha
s
be
e
n
us
e
d
i
n
p
r
e
v
ious
s
t
udi
e
s
a
nd
ha
s
be
c
o
me
popu
lar
in
im
a
ge
c
las
s
if
ica
t
ion
t
a
s
ks
due
to
it
s
a
bil
it
y
to
a
u
tom
a
ti
c
a
ll
y
lea
r
n
hie
r
a
r
c
hic
a
l
f
e
a
tur
e
r
e
p
r
e
s
e
nta
ti
o
ns
f
r
om
r
a
w
da
ta
,
s
ig
ni
f
ica
n
tl
y
i
mpr
o
vin
g
a
c
c
ur
a
c
y
a
n
d
e
f
f
icie
nc
y
[
5
]
,
but
i
n
s
o
me
c
a
s
e
s
r
e
s
e
a
r
c
h
r
e
qui
r
e
s
s
e
ve
r
a
l
t
e
c
hni
que
s
to
im
pr
ove
it
s
a
c
c
ur
a
c
y
,
inclu
ding
a
ugme
ntat
ion
a
n
d
hype
r
pa
r
a
me
ter
tec
hn
i
que
s
.
I
n
th
is
s
tudy
,
ba
s
e
d
on
t
he
c
omp
lexi
ty
o
f
th
e
i
mage
,
the
C
NN
a
l
go
r
it
hm
is
a
ble
to
c
las
s
i
f
y
ba
t
ik
mot
if
s
.
T
he
r
e
s
ul
ti
ng
m
ode
l
wi
ll
be
opt
im
ize
d
wit
h
a
ugm
e
ntati
on
a
nd
h
ype
r
pa
r
a
mete
r
tec
hn
iques
.
C
NN
is
ve
r
y
e
f
f
e
c
ti
ve
f
or
i
mage
pr
o
c
e
s
s
ing
tas
ks
[
6
]
,
i
nc
lu
di
ng
t
hos
e
invol
vin
g
c
olo
r
im
a
ge
s
[
7
]
,
c
olo
r
pr
oc
e
s
s
ing
i
n
c
ludi
ng
the
e
x
tr
a
c
ti
o
n
of
i
nf
o
r
ma
ti
o
n
a
bou
t
the
s
pe
c
t
r
a
l
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14,
No.
4,
Augus
t
2025
:
328
7
-
3299
3288
pr
ope
r
t
ies
o
f
t
he
ob
jec
t's
s
u
r
f
a
c
e
a
n
d
l
ook
f
o
r
t
h
e
b
e
s
t
s
i
mi
la
r
i
ty
o
f
a
s
e
t
o
f
de
s
c
r
ip
ti
o
ns
whic
h
h
a
ve
be
e
n
known
to
do
a
n
i
nt
r
oduc
ti
o
n
[
8]
.
C
N
Ns
a
s
t
he
c
las
s
if
ica
ti
o
n
a
l
go
r
it
h
m
,
i
t
s
pe
c
if
i
e
s
the
s
a
mp
l
e
c
ha
r
a
c
te
r
is
ti
c
s
o
f
the
a
c
qui
r
e
d
i
mage
s
,
a
nd
the
s
o
f
twa
r
e
de
ve
lope
d
to
ge
ne
r
a
te
da
tas
e
ts
o
f
dif
f
e
r
e
nt
c
h
a
r
a
c
ter
is
ti
c
s
[
9]
.
T
he
im
a
ge
c
las
s
if
ica
ti
on
pr
oc
e
s
s
r
e
quir
e
s
the
c
ompl
e
tene
s
s
of
the
f
e
a
tur
e
s
of
the
im
a
ge
whic
h
f
or
m
a
n
inf
or
mative
im
a
ge
pa
tt
e
r
n
s
o
that
inf
or
mation
f
r
om
the
im
a
ge
c
a
n
be
dis
playe
d
[
10]
,
[
11]
.
T
h
i
s
s
tudy
a
im
s
to
inves
ti
ga
te
the
e
f
f
e
c
ti
ve
ne
s
s
of
C
NN
a
r
c
hit
e
c
tu
r
e
s
in
c
las
s
if
ying
T
a
s
ikm
a
laya
ba
ti
k
mot
if
s
.
B
y
e
xplor
ing
va
r
ious
model
c
onf
igu
r
a
ti
ons
a
nd
opti
mi
z
a
ti
on
te
c
hniques
,
we
s
e
e
k
to
identif
y
the
mos
t
e
f
f
e
c
ti
ve
a
ppr
oa
c
h
f
or
a
c
hieving
high
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y.
T
hr
oug
h
thi
s
r
e
s
e
a
r
c
h,
we
hope
to
c
ontr
ibut
e
to
the
p
r
e
s
e
r
va
ti
on
of
c
ult
ur
a
l
he
r
it
a
ge
while
pr
ovidi
ng
a
pr
a
c
ti
c
a
l
tool
f
o
r
a
r
ti
s
a
ns
a
nd
s
take
holder
s
in
the
ba
ti
k
indus
tr
y
.
2.
M
E
T
HO
D
F
or
the
c
las
s
if
ica
ti
on
of
T
a
s
ikm
a
laya
ba
ti
k
mot
if
s
,
im
a
ge
da
ta
ha
s
be
e
n
c
oll
e
c
ted
f
r
om
va
r
ious
ba
ti
k
a
r
ti
s
a
ns
a
nd
s
hops
in
T
a
s
ikm
a
laya
,
including
a
r
e
a
s
s
uc
h
a
s
S
ingapa
r
na
in
T
a
s
ikm
a
laya
R
e
ge
nc
y,
kn
own
f
or
it
s
S
uka
pur
a
ha
nd
-
dr
a
wn
ba
ti
k
,
a
s
we
ll
a
s
the
c
it
y
o
f
T
a
s
ikm
a
laya
it
s
e
lf
.
T
he
r
e
a
r
e
f
our
di
s
ti
nc
ti
ve
T
a
s
ikm
a
lay
a
ba
ti
k
mot
if
s
that
will
s
e
r
ve
a
s
c
las
s
e
s
in
the
c
las
s
if
ica
ti
on:
pay
ung
,
k
ume
li
,
k
ujang
,
a
n
d
me
r
ak
ngibi
ng
.
I
n
tot
a
l
,
ther
e
a
r
e
163
r
e
c
or
ds
,
whic
h
a
r
e
pr
e
s
e
nted
in
the
T
a
ble
1.
F
i
gu
r
e
1
s
hows
t
he
r
e
s
e
a
r
c
h
s
ta
ge
s
in
c
l
a
s
s
i
f
yi
ng
t
yp
ica
l
T
a
s
ik
ma
la
ya
ba
t
i
k
m
ot
i
f
s
.
T
h
e
d
ia
gr
a
m
i
l
lus
t
r
a
tes
a
s
ys
te
ma
ti
c
wo
r
kf
l
ow
f
o
r
c
l
a
s
s
i
f
yi
ng
t
r
a
d
i
ti
on
a
l
T
a
s
ik
ma
la
ya
ba
t
ik
m
o
ti
f
s
,
b
e
g
in
nin
g
w
it
h
i
de
n
ti
f
y
i
ng
th
e
r
e
s
e
a
r
c
h
p
r
ob
le
m
,
r
e
vie
wi
ng
r
e
l
e
va
nt
l
it
e
r
a
tu
r
e
,
a
nd
c
o
l
lec
ti
ng
mo
t
i
f
da
ta
f
r
o
m
f
o
u
r
c
las
s
e
s
.
T
he
p
r
oc
e
s
s
c
o
nt
i
nue
s
wi
t
h
p
r
e
p
r
oc
e
s
s
i
ng
,
in
c
l
ud
in
g
r
e
s
iz
in
g
i
ma
ge
s
,
s
e
gm
e
n
ti
ng
m
ot
i
f
s
us
in
g
m
e
t
ho
ds
l
ik
e
c
a
nn
y
e
dg
e
de
tec
ti
on
a
nd
th
r
e
s
ho
ld
i
ng
,
a
n
d
a
u
gm
e
n
t
in
g
d
a
t
a
th
r
ou
gh
tec
hn
iq
ue
s
s
u
c
h
a
s
r
a
n
do
m
c
r
o
pp
in
g
,
r
o
tat
i
on
,
f
l
i
pp
in
g
,
a
f
f
i
ne
t
r
a
ns
f
o
r
m
a
t
io
n
,
a
nd
pa
dd
in
g
.
O
nc
e
p
r
e
pa
r
e
d
,
th
e
i
ma
ge
da
ta
is
us
e
d
to
t
r
a
in
a
n
d
va
l
i
da
te
a
C
N
N
,
w
i
th
pe
r
f
o
r
man
c
e
e
va
lua
te
d
us
i
ng
a
c
o
nf
us
io
n
ma
t
r
i
x
.
T
h
e
m
od
e
l
is
t
he
n
o
pt
im
iz
e
d
by
s
e
le
c
t
in
g
t
he
be
s
t
o
p
ti
mi
z
e
r
(
s
uc
h
a
s
Ad
a
m
o
r
s
to
c
ha
s
t
ic
g
r
a
d
ie
nt
de
s
c
e
n
t
(
S
G
D
)
)
,
r
e
f
i
ni
ng
t
he
C
NN
a
r
c
hi
tec
t
ur
e
,
a
nd
t
un
i
n
g
h
y
pe
r
pa
r
a
m
e
te
r
s
li
ke
lea
r
ni
ng
r
a
te
,
ba
tch
s
iz
e
,
a
n
d
e
poc
h
c
ou
nt
.
F
i
na
ll
y
,
m
od
e
l
pe
r
f
o
r
ma
nc
e
s
a
r
e
c
om
pa
r
e
d
,
c
on
c
l
us
i
ons
a
r
e
d
r
a
wn
a
b
ou
t
t
he
mos
t
e
f
f
e
c
ti
ve
a
pp
r
oa
c
h
,
a
n
d
th
e
c
las
s
i
f
i
c
a
ti
on
s
ys
te
m
f
o
r
T
a
s
ik
ma
la
ya
ba
t
ik
m
o
ti
f
s
is
c
o
mp
le
ted
.
I
n
T
a
bl
e
1
is
th
e
d
a
tas
e
t
ge
ne
r
a
te
d
f
r
o
m
t
he
i
mag
e
d
ig
i
ti
z
a
t
i
on
p
r
oc
e
s
s
b
a
s
e
d
o
n
4
c
las
s
e
s
o
f
ba
ti
k
m
o
ti
f
s
.
T
a
ble
1.
Da
tas
e
t
pr
oc
e
s
s
ing
r
e
s
ult
C
la
s
s
R
e
c
or
d
P
ay
ung
24
K
um
e
li
50
K
uj
ang
23
M
e
r
ak
N
gi
bi
ng
66
F
igur
e
1.
E
xpe
r
im
e
ntal
methods
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
C
las
s
if
ication
of
T
as
ikm
alaya
B
ati
k
motif
s
us
ing
c
onv
olut
ional
ne
ur
al
ne
tw
or
k
s
(
T
e
uk
u
M
ufi
z
ar
)
3289
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
3.
1.
P
r
e
p
r
oc
e
s
s
in
g
3.
1.
1.
Re
s
izin
g
T
he
pur
pos
e
of
r
e
s
izing
im
a
ge
s
to
150
×
150
pixel
s
dur
ing
the
p
r
e
pr
oc
e
s
s
ing
s
tage
is
to
s
tanda
r
dize
the
input
s
ize
,
e
ns
ur
ing
that
a
ll
im
a
ge
s
ha
ve
uni
f
or
m
dim
e
ns
ions
f
o
r
c
ons
is
tent
p
r
oc
e
s
s
ing
by
the
mo
de
l.
T
his
u
nif
or
mi
ty
is
c
r
uc
ial,
a
s
mos
t
de
e
p
lea
r
ning
a
r
c
hi
tec
tur
e
s
r
e
quir
e
f
ixed
input
s
ha
pe
s
.
R
e
s
izing
a
ls
o
r
e
duc
e
s
c
omput
a
ti
ona
l
load
a
nd
s
pe
e
ds
up
tr
a
ini
ng
by
mi
nim
izing
memor
y
us
a
ge
,
a
ll
owing
the
model
to
it
e
r
a
te
thr
ough
the
da
tas
e
t
mor
e
e
f
f
icie
ntl
y
[
12
]
,
[
13]
.
A
ddit
ionally,
maintaining
a
c
ons
is
tent
im
a
ge
s
ize
e
nha
nc
e
s
model
pe
r
f
or
manc
e
by
e
na
bli
ng
it
to
f
oc
us
on
the
e
s
s
e
nti
a
l
f
e
a
tur
e
s
of
ba
ti
k
mot
if
s
without
be
ing
d
is
tr
a
c
ted
by
va
r
iations
in
s
ize
a
nd
s
ha
pe
,
ult
im
a
tely
c
ont
r
ibu
ti
ng
to
mo
r
e
a
c
c
ur
a
te
c
las
s
if
ica
ti
on
outcome
s
.
3.
1.
2.
S
e
gm
e
n
t
at
ion
an
d
a
u
gm
e
n
t
at
ion
I
n
the
da
ta
pr
e
pr
oc
e
s
s
ing
pha
s
e
of
ou
r
ba
ti
k
mot
if
c
las
s
if
ica
ti
on
pr
ojec
t,
we
i
mpl
e
mente
d
im
a
ge
da
ta
s
e
gmenta
ti
on
a
s
a
c
r
uc
ial
s
tep.
T
his
pr
oc
e
s
s
invol
ve
d
or
ga
nizing
the
im
a
ge
da
tas
e
t
int
o
dis
ti
nc
t
c
a
tegor
ies
,
f
a
c
il
it
a
ti
ng
e
f
f
e
c
ti
ve
tr
a
ini
ng
a
nd
e
va
luation
of
t
he
mac
hine
lea
r
ning
model.
T
he
pur
pos
e
of
us
i
ng
da
ta
s
e
gmenta
ti
on
is
to
e
nha
nc
e
f
e
a
tur
e
e
xtr
a
c
ti
on,
i
mpr
ove
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y,
a
nd
f
a
c
il
it
a
te
be
tt
e
r
da
ta
r
e
pr
e
s
e
ntation
[
14]
,
[
15
]
.
T
he
tec
hniques
us
e
d
f
or
t
his
ba
ti
k
da
ta
a
r
e
th
r
e
s
holdi
ng
a
nd
c
a
nny
e
dge
de
te
c
ti
on.
De
e
p
lea
r
ning
models
r
e
quir
e
lar
ge
da
ta
s
e
ts
to
r
e
c
ognize
im
a
g
e
s
a
c
c
ur
a
tely,
da
ta
a
ugmenta
ti
on
tec
hniques
c
a
n
be
a
ppli
e
d
to
e
xpa
nd
the
da
tas
e
t
by
modi
f
ying
e
xis
ti
ng
im
a
ge
s
to
incr
e
a
s
e
da
ta
diver
s
it
y
[
16
]
,
[
17
]
.
T
he
ne
xt
s
tage
is
to
pe
r
f
o
r
m
da
ta
a
ug
me
ntati
on
.
T
he
mea
n
ing
f
ul
da
ta
a
u
gmen
tati
on
c
a
n
a
c
c
ompl
is
h
the
h
ighes
t
a
c
c
ur
a
c
y
w
it
h
a
low
e
r
e
r
r
or
r
a
te
o
n
a
ll
da
tas
e
ts
by
us
in
g
t
r
a
ns
f
e
r
lea
r
nin
g
mo
de
ls
[
1
8]
,
[
19
]
.
T
his
pr
oc
e
s
s
is
c
r
uc
ial
f
or
r
e
c
ognizing
r
ich
a
nd
c
ompl
e
x
pa
tt
e
r
ns
,
s
uc
h
a
s
ba
ti
k
mot
if
s
,
with
the
a
im
s
o
f
incr
e
a
s
ing
da
ta
va
r
iabili
ty,
r
e
duc
ing
ove
r
f
it
ti
ng
,
a
nd
im
pr
oving
model
r
obus
tnes
s
.
T
he
int
e
nti
ons
be
hind
thi
s
a
r
e
to
s
tr
e
ngthen
lea
r
ning
,
e
nha
nc
e
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y,
a
nd
opti
mi
z
e
the
us
e
of
li
mi
ted
da
tas
e
ts
.
T
he
da
ta
a
ugmenta
ti
on
tec
hniques
a
ppli
e
d
include
R
a
ndomR
e
s
i
z
e
dC
r
op(
p=
0.
1)
,
R
otate
(
li
mi
t=10,
p=
0.
5)
,
S
hif
tS
c
a
leR
otate
(
s
hif
t_l
im
it
=
0.
1
,
s
c
a
le_lim
it
=
0.
2,
r
o
tate
_li
mi
t=30,
p=
0.
7)
,
Hor
izonta
lF
li
p
(
p=
0.
5)
,
Ve
r
ti
c
a
lF
li
p(
p=
0.
5)
,
Af
f
ine(
s
he
a
r
=
20,
p=
0.
5)
,
a
nd
P
a
dI
f
Ne
e
de
d(
mi
n_he
ight
=
300
,
mi
n_wi
dth=300,
bor
de
r
_mode=
0,
va
lue
=
0,
p=
0.
1)
.
F
igur
e
2
s
hows
the
r
e
s
ult
s
o
f
da
ta
s
e
gmenta
ti
on
a
nd
da
ta
a
ugment
a
ti
on
f
o
r
or
igi
na
l
im
a
ge
in
F
igur
e
2
(
a
)
,
a
f
ter
s
e
gmenta
ti
on
i
n
F
igur
e
2(
b
)
,
a
nd
a
f
ter
s
e
gmenta
ti
on
a
nd
a
ugmenta
ti
on
in
F
igur
e
2(
c
)
.
(
a
)
(
b)
(
c
)
F
igur
e
2.
R
e
s
ult
s
of
da
ta
s
e
gmenta
ti
on
a
nd
da
ta
a
u
gmenta
ti
on
of
(
a
)
o
r
igi
na
l
i
mage
,
(
b)
a
f
ter
s
e
gment
a
ti
on
,
a
nd
(
c
)
a
f
ter
s
e
gmenta
ti
on
a
nd
a
ugmenta
ti
on
B
a
ti
k
mot
if
s
e
xhibi
t
a
diver
s
e
r
a
nge
of
pa
tt
e
r
ns
,
m
a
king
the
c
las
s
if
ica
ti
on
pr
oc
e
s
s
c
ompl
e
x.
P
r
e
vious
s
tudi
e
s
ha
ve
dis
c
us
s
e
d
how
im
a
ge
s
iz
e
,
im
a
ge
qua
li
ty,
a
nd
pa
tt
e
r
n
c
ha
r
a
c
ter
is
ti
c
s
a
f
f
e
c
t
the
c
las
s
if
ica
ti
on
of
ba
ti
k
[
20
]
.
T
his
f
indi
ng
is
a
ls
o
e
vide
nt
in
thi
s
r
e
s
e
a
r
c
h,
whic
h
c
onc
ludes
that
a
c
hieving
good
pe
r
f
or
manc
e
r
e
quir
e
s
s
pe
c
if
ic
tr
e
a
tm
e
nts
f
o
r
e
a
c
h
ba
ti
k
mot
if
to
e
ns
ur
e
opti
mal
model
pe
r
f
o
r
manc
e
.
T
he
t
r
e
a
tm
e
nt
of
e
a
c
h
mot
if
du
r
ing
the
pr
e
pr
oc
e
s
s
ing
s
tage
is
il
lus
tr
a
ted
i
n
the
T
a
ble
2.
T
a
ble
2
.
Da
tas
e
t
pr
oc
e
s
s
ing
r
e
s
ult
C
la
s
s
O
r
ig
in
a
l
S
e
gme
nt
a
ti
on
A
ugme
nt
a
ti
on
P
ay
ung
K
um
e
li
K
uj
ang
-
M
e
r
ak
N
gi
bi
ng
-
-
F
igur
e
3
il
lus
tr
a
tes
the
outcome
s
of
a
pplyi
ng
di
f
f
e
r
e
nt
c
ombi
na
ti
ons
o
f
da
ta
s
e
gmenta
ti
on
a
nd
da
ta
a
ugmenta
ti
on
tec
hniques
.
F
igur
e
3(
a
)
s
hows
the
pe
r
f
or
manc
e
r
e
s
ult
s
us
ing
s
e
gmenta
ti
on
a
nd
a
ugme
ntation
.
F
igur
e
3(
b)
s
hows
the
pe
r
f
or
manc
e
r
e
s
ult
s
us
ing
a
ugmenta
ti
on
without
s
e
gmenta
ti
on
.
F
igur
e
3(
c
)
s
hows
the
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14,
No.
4,
Augus
t
2025
:
328
7
-
3299
3290
pe
r
f
or
manc
e
r
e
s
ult
s
us
ing
s
e
gmenta
ti
on
a
nd
with
out
a
ugmenta
ti
on
.
F
igur
e
3
(
d)
s
hows
pe
r
f
or
manc
e
r
e
s
ult
s
without
s
e
gmenta
ti
on
a
nd
without
a
ugmenta
ti
on.
(
a
)
(
b)
(
c
)
(
d)
F
igur
e
3.
P
e
r
f
or
manc
e
r
e
s
ult
s
of
da
ta
s
e
gmenta
ti
on
a
nd
da
ta
a
ugmenta
ti
on
of
(
a
)
with
s
e
gmenta
ti
on
a
nd
with
a
ugmenta
ti
on,
(
b
)
without
s
e
gmenta
ti
on
a
nd
with
a
ugmenta
ti
on,
(
c
)
with
s
e
gmenta
ti
on
a
nd
without
a
ugmenta
ti
on
a
nd
(
d)
without
s
e
gmenta
ti
on
a
nd
wi
thout
a
ugmenta
ti
on
3.
2.
M
od
e
li
n
g
3.
2.
1.
T
r
ain
in
g
an
d
v
a
li
d
at
io
n
T
he
da
ta
tr
a
ini
ng
pr
oc
e
s
s
invol
ve
s
s
pli
tt
ing
the
da
tas
e
t
int
o
two
pa
r
ts
:
tr
a
ini
ng
da
ta
a
nd
tes
ti
ng
da
ta
,
with
a
r
a
ti
o
of
80%
f
or
t
r
a
ini
ng
a
nd
20%
f
o
r
tes
t
ing.
I
n
the
modeling
pha
s
e
,
the
f
i
r
s
t
t
r
a
ini
ng
p
r
oc
e
s
s
us
e
s
the
or
igi
na
l
da
tas
e
t
without
s
e
gmenta
ti
on
a
nd
a
ug
menta
ti
on,
e
mpl
oying
a
C
NN
(
opti
mi
z
e
r
:
Ada
m,
lea
r
ning
r
a
te
:
0
.
0
01
,
ba
tc
h
s
iz
e
:
3
2
,
a
n
d
e
po
c
hs
:
20
)
.
T
he
a
r
c
h
i
tec
tu
r
e
c
ons
is
ts
o
f
th
r
e
e
c
on
vo
l
ut
io
na
l
la
ye
r
s
(
32
,
6
4
,
1
28
)
,
thr
e
e
max
pooli
ng
laye
r
s
(
2
,
2
)
,
a
nd
one
de
ns
e
laye
r
(
128
)
.
T
he
model
a
c
hieve
d
a
n
a
c
c
ur
a
c
y
of
75%
.
How
e
ve
r
,
the
a
c
c
ur
a
c
y
gr
a
ph
s
hows
that
t
r
a
ini
ng
a
c
c
ur
a
c
y
incr
e
a
s
e
s
while
va
li
da
ti
on
a
c
c
ur
a
c
y
de
c
r
e
a
s
e
s
a
s
the
e
poc
hs
pr
ogr
e
s
s
,
indi
c
a
ti
ng
that
the
model
is
li
ke
ly
e
xpe
r
ienc
ing
ove
r
f
it
ti
ng
.
T
o
im
p
r
ove
the
model's
pe
r
f
or
manc
e
,
it
is
ne
c
e
s
s
a
r
y
to
make
s
ome
a
djus
tm
e
nts
.
Among
de
e
p
lea
r
ning
types
,
C
NN
a
r
e
the
mos
t
c
omm
on
types
of
de
e
p
lea
r
ning
models
uti
li
z
e
d
f
or
medic
a
l
im
a
ge
diagnos
is
a
nd
a
na
lys
is
.
How
e
ve
r
,
C
NN
s
u
f
f
e
r
s
f
r
om
high
c
omput
a
ti
on
c
os
t
to
be
i
mpl
e
mente
d
a
nd
may
r
e
quir
e
to
a
da
pt
huge
nu
mber
o
f
pa
r
a
mete
r
s
[
21]
,
to
e
nha
nc
e
the
model's
pe
r
f
o
r
manc
e
,
da
ta
will
be
pr
oc
e
s
s
e
d
us
ing
s
e
gmenta
ti
on
a
nd
a
ugmenta
ti
on
tec
hniques
.
S
e
c
ondly,
the
s
e
lec
ti
on
of
a
C
NN
opti
m
i
z
e
r
will
be
c
onduc
ted,
a
nd
thi
r
dly,
hype
r
pa
r
a
mete
r
tuni
ng
will
be
pe
r
f
or
med
to
f
ur
the
r
im
p
r
ove
the
model's
p
e
r
f
or
manc
e
[
22
]
.
3.
3.
Op
t
im
az
at
io
n
m
o
d
e
l
3.
3.
1.
Op
t
im
ize
r
c
om
p
ar
is
on
F
r
om
the
e
xpe
r
im
e
nts
c
onduc
ted
with
f
our
opt
im
i
z
e
r
s
(
Ada
m,
S
GD
,
r
oot
mea
n
s
qua
r
e
pr
opa
ga
ti
on
(
R
M
S
pr
op
)
,
a
nd
a
da
pti
ve
g
r
a
dient
a
lgo
r
it
hm
(
Ad
a
Gr
a
d
)
)
us
ing
a
lea
r
ning
r
a
te
of
0
.
001,
a
ba
tch
s
ize
of
32,
a
nd
50
e
poc
hs
,
the
im
a
ge
a
bove
s
hows
that
a
ll
f
our
opti
mi
z
e
r
s
pe
r
f
or
m
e
f
f
e
c
ti
ve
ly
.
T
he
gr
a
phs
o
f
model
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
C
las
s
if
ication
of
T
as
ikm
alaya
B
ati
k
motif
s
us
ing
c
onv
olut
ional
ne
ur
al
ne
tw
or
k
s
(
T
e
uk
u
M
ufi
z
ar
)
3291
a
c
c
ur
a
c
y
a
nd
va
li
da
ti
on
a
c
c
ur
a
c
y
indi
c
a
te
that
both
tr
a
ini
ng
a
nd
va
li
da
ti
on
a
c
c
ur
a
c
ies
incr
e
a
s
e
a
s
the
e
poc
hs
pr
ogr
e
s
s
.
F
igu
r
e
4
p
r
e
s
e
nts
the
pe
r
f
or
manc
e
r
e
s
ult
s
obtaine
d
us
ing
the
o
r
igi
na
l
da
tas
e
t,
s
pe
c
if
ica
ll
y
il
lus
tr
a
ti
ng
the
model
a
c
c
ur
a
c
y
in
F
igur
e
4
(
a
)
a
nd
t
he
c
onf
us
ion
matr
ix
in
F
igur
e
4
(
b)
.
(
a
)
(
b)
F
igur
e
4.
P
e
r
f
or
manc
e
r
e
s
ult
wi
th
or
igi
na
l
da
ta
of
(
a
)
model
a
c
c
ur
a
c
y
(
b
)
c
onf
us
ion
matr
ix
F
igur
e
5
s
hows
the
r
e
s
ult
of
model
a
c
c
ur
a
c
y
us
ing
4
opti
mi
z
e
r
s
,
na
mely
Ada
m
opti
mi
z
e
r
in
F
igur
e
5
(
a
)
,
S
GD
opti
mi
z
e
r
in
F
igur
e
5
(
b)
,
R
M
S
pr
op
opti
mi
z
e
r
in
F
igu
r
e
5
(
c
)
,
a
nd
Ada
Gr
a
d
opti
mi
z
e
r
in
F
igur
e
5
(
d
)
.
Ne
xt,
we
tes
ted
the
model's
pe
r
f
o
r
ma
nc
e
us
ing
the
f
ou
r
op
ti
mi
z
e
r
s
.
F
r
om
the
c
on
f
us
ion
matr
ix
s
hown,
we
f
ound
that
the
highes
t
pe
r
f
or
mi
ng
mode
l
us
e
d
the
Ada
m
opti
mi
z
e
r
,
whic
h
a
c
hieve
d
80%
a
c
c
ur
a
c
y.
F
igur
e
6
is
the
r
e
s
ult
o
f
the
c
o
nf
us
ion
matr
ix
us
ing
4
opti
mi
z
e
r
s
,
na
mely
Ada
m
opti
mi
z
e
r
in
F
igu
r
e
6(
a
)
,
S
GD
opti
mi
z
e
r
in
F
igur
e
6
(
b)
,
R
M
S
pr
op
op
ti
mi
z
e
r
in
F
i
gur
e
6(
c
)
,
a
nd
Ada
Gr
a
d
opti
mi
z
e
r
in
F
igur
e
6
(
d)
.
(
a
)
(
b)
(
c
)
(
d)
F
igur
e
5.
M
ode
l
a
c
c
ur
a
c
y
with
4
opti
mi
z
e
r
s
of
(
a
)
Ada
m,
(
b
)
S
GD
,
(
c
)
R
M
S
pr
op,
a
nd
(
d
)
Ada
Gr
a
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14,
No.
4,
Augus
t
2025
:
328
7
-
3299
3292
F
igur
e
6.
C
onf
us
ion
matr
ix
with
4
opti
mi
z
e
r
s
of
(
a
)
Ada
m,
(
b)
S
GD
,
(
c
)
R
M
S
pr
op,
a
nd
(
d
)
Ada
Gr
a
d
3.
3.
2.
Ar
c
h
it
e
c
t
u
r
e
c
o
m
p
ar
is
on
C
NN
s
ha
ve
ga
ined
r
e
mar
ka
ble
s
uc
c
e
s
s
on
ma
ny
im
a
ge
s
’
c
las
s
if
ica
ti
on
tas
ks
in
r
e
c
e
nt
ye
a
r
s
.
How
e
ve
r
,
the
pe
r
f
or
manc
e
o
f
C
NN
s
highl
y
r
e
li
e
s
upon
their
a
r
c
hit
e
c
tur
e
s
[
23]
.
T
he
ne
xt
s
tep
to
im
p
r
ove
the
model's
pe
r
f
or
manc
e
is
to
s
e
lec
t
the
be
s
t
C
NN
a
r
c
hit
e
c
tur
e
,
I
n
the
C
NN
a
r
c
hi
tec
tur
e
,
a
n
a
ve
r
a
ge
-
pooli
ng
laye
r
a
nd
a
max
-
pooli
ng
laye
r
a
r
e
c
onne
c
ted
in
pa
r
a
ll
e
l
in
o
r
de
r
to
boos
t
c
las
s
if
ica
ti
on
pe
r
f
or
manc
e
[
24]
,
f
or
thes
e
ba
ti
k
mot
if
s
.
I
n
th
is
e
xpe
r
im
e
nt,
thr
e
e
a
r
c
hit
e
c
tur
e
s
will
be
tes
ted,
de
tailed
a
s
in
T
a
ble
3
.
T
a
ble
3.
T
hr
e
e
a
r
c
hit
e
c
tur
e
s
model
C
NN
M
ode
l
O
pt
omi
z
e
r
L
e
a
r
ni
ng
r
a
te
B
a
tc
h
s
iz
e
E
poc
h
VG
G
N
e
t
A
da
m
0.001
32
10
R
e
s
N
e
t
A
da
m
0.001
32
10
G
oogL
e
N
e
t
A
da
m
0.001
32
10
B
a
s
e
d
on
the
e
xpe
r
im
e
nts
with
thr
e
e
C
NN
a
r
c
hit
e
c
tur
e
(
vis
ua
l
ge
ometr
y
gr
oup
ne
two
r
k
(
VG
GN
e
t
)
,
r
e
s
idual
ne
twor
k
(
R
e
s
Ne
t
),
a
nd
GoogL
e
Ne
t
)
,
we
f
ound
the
g
r
a
ph
s
hows
that
the
a
c
c
ur
a
c
y
a
nd
v
a
li
da
ti
on
a
c
c
ur
a
c
y
of
e
a
c
h
model
incr
e
a
s
e
a
s
the
e
poc
h
s
pr
ogr
e
s
s
,
indi
c
a
ti
ng
that
the
models
de
mons
tr
a
te
good
pe
r
f
or
manc
e
.
How
e
ve
r
,
hype
r
pa
r
a
mete
r
tuni
ng
i
s
ne
e
de
d
to
c
ontr
oll
ing
the
lea
r
ning
pr
oc
e
s
s
,
pr
e
ve
nti
ng
ove
r
f
it
ti
ng/under
f
it
ti
ng
a
nd
i
mpr
ov
ing
a
c
c
ur
a
c
y
a
nd
ge
ne
r
a
li
z
a
ti
on.
W
it
hout
pr
ope
r
hype
r
pa
r
a
mete
r
t
uning,
a
model
may
f
a
il
to
a
c
hieve
opti
mal
pe
r
f
or
manc
e
,
e
v
e
n
if
the
a
lgor
it
hm
it
s
e
lf
is
a
dva
nc
e
d.
3.
3.
3.
T
u
n
n
i
n
g
h
yp
e
r
p
ar
a
m
e
t
e
r
Hype
r
pa
r
a
mete
r
tuni
ng
is
the
pr
oc
e
s
s
of
opti
mi
z
ing
the
s
e
tt
ings
of
hype
r
pa
r
a
mete
r
s
,
whic
h
a
r
e
pa
r
a
mete
r
s
not
lea
r
ne
d
by
the
model
du
r
ing
t
r
a
i
ning
.
Hype
r
pa
r
a
mete
r
tuni
ng
is
e
s
s
e
nti
a
l
in
t
r
a
in
ing
s
uc
h
m
o
de
ls
a
n
d
s
ig
ni
f
ica
n
tl
y
i
mp
a
c
ts
t
he
i
r
f
i
na
l
p
e
r
f
o
r
m
a
n
c
e
a
nd
t
r
a
in
i
ng
s
pe
e
d
[
2
5
]
.
F
i
gu
r
e
7
s
h
ows
t
he
r
e
s
ul
t
of
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
C
las
s
if
ication
of
T
as
ikm
alaya
B
ati
k
motif
s
us
ing
c
onv
olut
ional
ne
ur
al
ne
tw
or
k
s
(
T
e
uk
u
M
ufi
z
ar
)
3293
m
o
de
l
a
c
c
u
r
a
c
y
us
in
g
th
e
A
da
m
op
ti
m
ize
r
:
V
GG
Ne
t
op
t
im
ize
r
i
n
F
i
gu
r
e
7
(
a
)
,
R
e
s
Ne
t
op
ti
mi
z
e
r
in
F
i
gu
r
e
7(
b
)
,
a
nd
Google
Ne
t
opti
mi
z
e
r
in
F
igu
r
e
7
(
c
)
.
F
igu
r
e
8
s
hows
the
c
onf
us
ion
matr
ix
o
f
e
a
c
h
a
r
c
hit
e
c
tur
e
u
s
ing
the
Ada
m
opti
mi
z
e
r
:
VG
GN
e
t
a
r
c
hit
e
c
tur
e
in
F
igur
e
8(
a
)
,
R
e
s
Ne
t
a
r
c
hit
e
c
tur
e
in
F
igur
e
8(
b)
,
a
nd
GoogL
e
Ne
t
a
r
c
hit
e
c
tur
e
in
F
igu
r
e
8(
c
)
.
(
a
)
(
b)
(
c
)
F
igur
e
7.
M
ode
l
a
c
c
ur
a
c
y
r
e
s
ult
s
us
ing
A
da
m
opti
mi
z
e
r
of
(
a
)
VG
GN
e
t,
(
b
)
R
e
s
Ne
t
,
a
nd
(
c
)
GoogL
e
Ne
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14,
No.
4,
Augus
t
2025
:
328
7
-
3299
3294
(
a
)
(
b)
(
c
)
F
igur
e
8.
C
onf
us
ion
matr
ix
us
ing
A
da
m
opti
mi
z
e
r
of
(
a
)
VG
GN
e
t,
(
b)
R
e
s
Ne
t
,
a
nd
(
c
)
GoogL
e
Ne
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
C
las
s
if
ication
of
T
as
ikm
alaya
B
ati
k
motif
s
us
ing
c
onv
olut
ional
ne
ur
al
ne
tw
or
k
s
(
T
e
uk
u
M
ufi
z
ar
)
3295
F
igur
e
9
s
hows
the
r
e
s
ult
s
of
hype
r
pa
r
a
mete
r
tu
ning
on
the
VG
GN
e
t
a
r
c
hit
e
c
tur
e
by
incr
e
a
s
ing
the
number
of
e
poc
hs
to
50
f
o
r
1
st
e
xpe
r
i
ment
i
n
F
igur
e
9(
a
)
a
nd
2
nd
e
xpe
r
im
e
nt
in
F
igu
r
e
9(
b)
.
I
t
c
a
n
be
obs
e
r
ve
d
that
the
a
c
c
ur
a
c
y
gr
a
ph
indi
c
a
tes
i
mpr
ove
ment,
with
bo
th
t
r
a
ini
ng
a
c
c
ur
a
c
y
a
nd
v
a
li
da
ti
on
a
c
c
ur
a
c
y
incr
e
a
s
ing
c
los
e
ly
togethe
r
a
s
the
num
be
r
of
e
poc
hs
incr
e
a
s
e
.
T
his
de
mons
tr
a
tes
that
t
he
model
pe
r
f
or
ms
we
ll
.
(
a
)
(
b)
F
igur
e
9.
M
ode
l
a
c
c
ur
a
c
y
r
e
s
ult
s
tunni
ng
hype
r
pa
r
a
mete
r
(
a
)
1
st
e
xpe
r
i
ment
a
nd
(
b)
2
nd
e
xpe
r
i
ment
F
igur
e
s
10
s
hows
the
r
e
s
ult
s
of
the
c
onf
us
ion
matr
ix
f
r
om
the
hype
r
pa
r
a
mete
r
tun
ing
e
xpe
r
im
e
nt
on
the
VG
GN
e
t
a
r
c
hit
e
c
tur
e
.
T
he
number
of
e
poc
hs
wa
s
incr
e
a
s
e
d
to
50
f
or
1
st
e
xpe
r
im
e
nt
in
F
igur
e
1
0(
a
)
a
nd
2
nd
ex
pe
r
im
e
nt
in
F
igur
e
10(
b)
.
T
he
r
e
s
ult
s
indi
c
a
te
that
the
model
pe
r
f
or
ms
we
ll
,
with
a
n
im
pr
ove
m
e
nt
in
the
pr
e
diction
of
the
k
ume
li
c
las
s
.
(
a
)
(
b)
F
igur
e
10.
C
onf
us
ion
matr
ix
r
e
s
ult
s
tunni
ng
hype
r
pa
r
a
mete
r
(
a
)
1
st
e
xpe
r
im
e
nt
a
nd
(
b)
2
nd
e
xpe
r
im
e
n
t
3.
4.
P
e
r
f
or
m
an
c
e
c
o
m
p
ar
is
on
Af
ter
c
onduc
ti
ng
e
xpe
r
im
e
nts
to
im
pr
ove
mod
e
l
pe
r
f
or
manc
e
,
the
c
ompar
is
on
r
e
s
ult
s
c
a
n
be
de
ter
mi
ne
d.
T
he
be
s
t
-
pe
r
f
or
mi
ng
model
wa
s
a
c
hieve
d
us
ing
the
Ada
m
opti
mi
z
e
r
with
hype
r
pa
r
a
met
e
r
s
s
e
t
to
a
lea
r
ning
r
a
te
of
0.
001
,
a
ba
tch
s
ize
of
32,
a
nd
50
e
poc
hs
,
r
e
s
ult
ing
in
a
n
a
c
c
ur
a
c
y
of
80%
.
T
he
pe
r
f
or
manc
e
r
e
s
ult
s
f
or
the
Ada
m
opti
mi
z
e
r
c
a
n
be
s
e
e
n
in
T
a
bl
e
4
.
T
a
ble
5
s
hows
the
a
c
c
ur
a
c
y
va
lues
f
o
r
e
a
c
h
op
ti
mi
z
e
r
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
14,
No.
4,
Augus
t
2025
:
328
7
-
3299
3296
T
a
ble
4.
T
he
pe
r
f
or
manc
e
o
f
opti
mi
z
e
r
Ada
m
C
la
s
s
P
r
e
c
i
s
io
n
R
e
c
a
ll
F1
-
s
c
or
e
S
uppor
t
P
ay
ung
0.88
0.70
0.78
10
K
um
e
li
0.71
1.00
0.83
17
K
uj
ang
0.75
0.75
0.75
4
M
e
r
ak
n
gi
bi
ng
0.93
0.68
0.79
19
T
a
ble
5.
Ac
c
ur
a
c
y
o
pti
mi
z
e
r
O
pt
im
iz
e
r
A
c
c
ur
a
c
y
(%)
A
da
m
80
S
G
D
72
R
M
S
pr
op
76
A
da
G
r
a
d
72
T
a
ble
6
s
hows
the
a
c
c
ur
a
c
y
va
lue
s
f
r
om
e
xpe
r
im
e
nts
us
ing
thr
e
e
a
r
c
hit
e
c
tur
e
s
:
VG
GN
e
t,
R
e
s
Ne
t
,
a
nd
GoogL
e
Ne
t.
T
he
highes
t
a
c
c
ur
a
c
y
w
a
s
a
c
hi
e
ve
d
GoogL
e
Ne
t
with
the
Ada
m
opti
mi
z
e
r
,
r
e
a
c
hing
100%
.
How
e
ve
r
,
hype
r
pa
r
a
mete
r
tuni
ng
is
ne
e
de
d
to
f
ur
ther
i
mpr
ove
pe
r
f
or
manc
e
,
be
c
a
us
e
the
a
c
c
ur
a
c
y
gr
a
ph
s
hows
f
luctua
ti
ons
be
twe
e
n
tr
a
ini
ng
a
c
c
ur
a
c
y
a
n
d
va
li
da
ti
on
a
c
c
ur
a
c
y
.
T
his
lea
ds
to
the
model
b
e
ing
les
s
e
f
f
e
c
ti
ve
a
nd
incons
is
tent.
T
a
ble
6.
Ac
c
ur
a
c
y
of
models
us
ing
Ada
m
opti
mi
z
e
r
M
ode
l
L
e
a
r
ni
ng
r
a
te
B
a
tc
h
s
iz
e
E
poc
h
A
c
c
ur
a
c
y
(%)
V
G
G
N
e
t
0.001
32
10
96
R
e
s
N
e
t
0.001
32
10
60
G
oogL
e
N
e
t
0.001
32
10
100
T
a
ble
7
dis
plays
the
a
c
c
ur
a
c
y
va
lues
f
r
om
the
e
xpe
r
im
e
nts
with
the
thr
e
e
a
r
c
hit
e
c
tur
e
s
.
I
t
c
a
n
be
c
onc
luded
that
the
GoogL
e
Ne
t
a
r
c
hit
e
c
tur
e
a
c
hiev
e
d
the
highes
t
a
c
c
ur
a
c
y.
How
e
ve
r
,
hype
r
pa
r
a
mete
r
tuni
ng
is
ne
e
de
d
to
f
u
r
ther
im
pr
ove
pe
r
f
or
manc
e
.
T
a
ble
7
s
hows
the
a
c
c
ur
a
c
y
va
lues
f
r
om
thr
e
e
types
of
a
r
c
hit
e
c
tur
e
s
tes
ted
us
ing
the
Ada
m
opti
mi
z
e
r
,
whic
h
is
c
ons
ider
e
d
the
be
s
t
opti
mi
z
e
r
a
nd
GoogL
e
N
e
t
whic
h
is
c
ons
ider
e
d
the
be
s
t
a
r
c
hit
e
c
tur
e
a
s
the
a
c
c
ur
a
c
y
r
e
maine
d
a
t
100
%
,
T
he
pr
opos
e
d
method
in
th
is
s
tudy
tends
to
ha
ve
a
much
higher
a
c
c
ur
a
c
y
p
r
opor
ti
o
n
than
other
a
r
c
hit
e
c
tur
e
s
.
Ac
c
or
ding
to
ou
r
s
tud
y,
lowe
r
a
c
c
ur
a
c
y
doe
s
not
ne
c
e
s
s
a
r
il
y
indi
c
a
te
poor
model
pe
r
f
or
manc
e
in
c
las
s
if
ica
ti
on.
T
he
p
r
opos
e
d
opti
mi
z
a
ti
on
tec
hnique
c
a
n
potentially
im
p
r
ove
a
c
c
ur
a
c
y
with
t
h
e
a
va
il
a
ble
da
tas
e
t.
T
his
s
tudy
tes
ted
the
pe
r
f
o
r
m
a
nc
e
of
a
c
ompr
e
he
ns
ive
C
NN
model
with
the
opt
im
ize
d
m
ode
l.
How
e
ve
r
,
mor
e
thor
ough
r
e
s
e
a
r
c
h
may
be
n
e
e
de
d
to
va
li
da
te
it
s
a
c
c
ur
a
c
y,
e
s
pe
c
ially
in
r
e
lation
to
the
li
mi
tations
of
the
da
tas
e
t.
T
a
ble
7.
Ac
c
ur
a
c
y
of
tunni
ng
pa
r
a
mete
r
GoogL
e
Ne
t
a
r
c
hit
e
c
tur
e
M
ode
l
L
e
a
r
ni
ng
r
a
te
B
a
tc
h
s
iz
e
E
poc
h
A
c
c
ur
a
c
y
(%)
1s
t
e
xpe
r
im
e
nt
0.001
32
50
100
2nd
e
xpe
r
im
e
nt
0.001
32
50
100
4.
CONC
L
USI
ON
Our
f
indi
ngs
pr
ovide
c
onc
lus
ive
e
videnc
e
that
e
xpe
r
im
e
nts
c
onduc
ted
on
ba
ti
k
mot
if
c
las
s
if
ica
ti
on
us
ing
C
NN
r
e
ve
a
l
that
the
high
c
ompl
e
xit
y
o
f
ba
ti
k
mot
if
s
s
igni
f
ica
ntl
y
a
f
f
e
c
ts
model
pe
r
f
or
manc
e
,
a
s
the
ha
ndli
ng
of
e
a
c
h
c
las
s
a
f
f
e
c
ts
the
ove
r
a
ll
r
e
s
ult
s
.
I
nit
ial
e
xpe
r
im
e
nts
with
the
o
r
igi
na
l
da
tas
e
t
s
howe
d
s
ubopti
mal
model
pe
r
f
o
r
manc
e
,
c
ha
r
a
c
ter
ize
d
b
y
a
c
c
ur
a
c
y
a
nd
va
li
da
ti
on
c
u
r
ve
s
indi
c
a
ti
ng
ov
e
r
f
it
ti
ng,
a
c
hieving
only
75
%
a
c
c
ur
a
c
y
with
a
lea
r
ning
r
a
t
e
of
0.
001
,
a
ba
tch
s
ize
of
32,
a
nd
50
e
poc
hs
.
How
e
ve
r
,
by
e
mpl
oying
da
ta
s
e
gmenta
ti
on,
da
ta
a
ugment
a
ti
on,
s
e
lec
ti
ng
the
be
s
t
opti
mi
z
e
r
,
uti
li
z
ing
a
n
opti
mal
a
r
c
hit
e
c
tur
e
,
a
nd
tuni
ng
hype
r
pa
r
a
mete
r
s
,
model
pe
r
f
or
manc
e
im
p
r
ove
d
s
igni
f
ica
ntl
y.
T
he
be
s
t
m
ode
l
wa
s
obtaine
d
by
tr
a
ini
ng
on
da
ta
that
unde
r
we
nt
s
pe
c
if
ic
pr
e
pr
oc
e
s
s
ing
f
or
e
a
c
h
c
las
s
,
us
ing
the
Ada
m
opti
mi
z
e
r
with
hype
r
pa
r
a
mete
r
tuni
ng
s
e
t
to
a
lea
r
ning
r
a
te
of
0.
001,
a
ba
tch
s
ize
of
32,
a
nd
50
e
poc
hs
.
I
n
the
hype
r
pa
r
a
mete
r
tuni
ng
e
xpe
r
im
e
nt
with
the
VG
GN
e
t
a
r
c
hit
e
c
tur
e
,
it
wa
s
s
hown
that
ther
e
is
a
n
im
p
r
ove
ment
in
the
pr
e
diction
of
the
k
ume
li
c
las
s
,
a
c
hieving
a
n
a
c
c
ur
a
c
y
of
100%
.
Our
s
tudy
s
hows
that
opti
mi
z
a
ti
on
tec
hniques
on
C
NN
model
pe
r
f
or
manc
e
c
a
n
be
be
tt
e
r
a
nd
mor
e
a
c
c
ur
a
te
than
be
f
or
e
.
F
u
tur
e
s
tudi
e
s
c
a
n
de
ve
lop
other
f
e
a
tur
e
s
in
ba
ti
k
a
nd
e
xpl
or
e
f
e
a
s
ibl
e
methods
to
p
r
oduc
e
the
be
s
t
model
pe
r
f
or
manc
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.