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.
1
,
F
e
br
ua
r
y
2025
,
pp.
222
~
230
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
1
.
pp
22
2
-
230
222
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
C
om
p
a
r
is
on
of
f
as
t
e
r
r
e
gi
on
-
b
ase
d
c
on
vol
u
t
io
n
al
n
e
t
w
or
k
f
or
al
gor
ith
m
s
f
o
r
gr
a
p
e
l
e
ave
s
c
la
ssi
f
ic
at
io
n
M
oe
c
h
am
m
ad
S
ar
os
a
1
,
P
u
t
e
r
i
Nurul
M
a’
r
i
f
ah
1
,
M
il
a
Ku
s
u
m
awar
d
an
i
1
,
Di
m
as
F
irm
a
n
d
a
Al
Ri
z
a
2
1
D
e
pa
r
tm
e
nt
of
E
le
c
tr
ic
a
l
E
ngi
ne
e
r
in
g, S
ta
te
P
ol
yt
e
c
hni
c
of
M
a
la
ng, M
a
la
ng, I
ndone
s
ia
2
D
e
pa
r
tm
e
nt
of
B
io
s
ys
t
e
ms
E
ngi
ne
e
r
in
g
,
F
a
c
ul
ty
of
A
gr
ic
ul
tu
r
a
l
T
e
c
hnol
ogy, Unive
r
s
it
y of
B
r
a
w
ij
a
ya
,
Ma
la
ng
,
I
ndone
s
i
a
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
De
c
15
,
2023
R
e
vis
e
d
J
ul
2
,
2024
Ac
c
e
pted
J
ul
26
,
2024
T
h
e
s
h
a
p
es
o
f
l
ea
v
es
d
i
s
t
i
n
g
u
i
s
h
t
h
e
In
d
o
n
es
i
an
g
ra
p
e
v
ari
an
t
s
.
T
h
e
g
rap
e
l
eav
e
s
mi
g
h
t
l
o
o
k
t
h
e
s
ame
at
f
i
rs
t
g
l
an
ce,
b
u
t
t
h
ere
ar
e
d
i
ffere
n
ces
i
n
l
eaf
s
h
a
p
es
an
d
ch
arac
t
eri
s
t
i
cs
w
h
en
o
b
s
er
v
ed
cl
o
s
el
y
.
T
h
i
s
res
earch
u
s
e
s
a
d
eep
l
earn
i
n
g
met
h
o
d
co
m
b
i
n
ed
w
i
t
h
t
h
e
f
as
t
er
re
g
i
o
n
-
b
as
e
d
co
n
v
o
l
u
t
i
o
n
a
l
n
e
u
ral
n
et
w
o
r
k
(
R
-
CN
N
)
al
g
o
ri
t
h
m
w
i
t
h
t
h
e
In
cep
t
i
o
n
n
et
w
o
rk
arch
i
t
ect
u
re,
Res
N
e
t
V
2
,
Res
N
et
-
1
5
2
,
Res
N
e
t
-
1
0
1
,
an
d
Res
N
e
t
-
5
0
,
an
d
u
s
es
CO
C
O
w
ei
g
h
t
s
t
rai
n
ed
t
o
cl
as
s
i
f
y
fi
v
e
g
ra
p
e
v
ari
e
t
i
e
s
t
h
r
o
u
g
h
l
e
af
i
mag
es
.
T
h
e
s
t
u
d
y
co
l
l
ect
e
d
5
0
0
i
mag
e
s
t
o
b
e
u
s
ed
as
a
n
i
n
d
e
p
e
n
d
e
n
t
d
at
a
s
et
.
T
h
e
re
s
u
l
t
s
s
h
o
w
t
h
a
t
n
et
w
o
rk
i
m
p
ro
v
emen
t
s
can
effect
i
v
e
l
y
i
mp
r
o
v
e
o
p
era
t
i
n
g
effi
ci
e
n
c
y
.
T
h
ere
are
al
s
o
l
i
m
i
t
a
t
i
o
n
s
t
o
t
ra
i
n
i
n
g
s
co
r
es
b
eca
u
s
e
t
h
e
F1
s
co
re
v
al
u
e
t
e
n
d
s
t
o
s
t
a
b
i
l
i
ze
o
r
d
ecrea
s
e
at
a
cert
ai
n
p
o
i
n
t
.
In
t
h
e
In
cep
t
i
o
n
Res
N
et
V
2
arch
i
t
ect
u
re,
w
i
t
h
t
h
e
h
i
g
h
e
s
t
av
era
g
e
F1
s
co
re
o
f
9
2
%
,
t
h
e
av
erag
e
co
m
p
u
t
i
n
g
t
i
me
fo
r
t
rai
n
i
n
g
an
d
t
es
t
i
n
g
i
s
l
o
n
g
er
t
h
an
o
t
h
er
n
e
t
w
o
r
k
arch
i
t
ect
u
res
.
T
h
i
s
s
u
g
g
e
s
t
s
t
h
a
t
t
h
e
al
g
o
ri
t
h
m
ca
n
cl
as
s
i
f
y
t
y
p
e
s
o
f
g
ra
p
es
b
as
e
d
o
n
t
h
e
i
r
l
eav
e
s
.
K
e
y
w
o
r
d
s
:
Gr
a
pe
plant
Gr
a
pe
va
r
ieties
I
nc
e
pti
on
R
e
s
Ne
t
L
e
a
f
de
tec
ti
on
R
e
s
Ne
t
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
:
M
oe
c
ha
mm
a
d
S
a
r
os
a
De
pa
r
tm
e
nt
of
E
lec
tr
ica
l
E
nginee
r
ing,
S
tate
P
olyt
e
c
hnic
of
M
a
lang
S
t.
S
oe
ka
r
no
-
Ha
tt
a
09
,
M
a
lang
65141,
I
ndone
s
ia
E
mail:
ms
a
r
os
a
@poli
ne
ma.
a
c
.
id
1.
I
NT
RODU
C
T
I
ON
Gr
a
pe
s
be
long
to
the
V
it
ac
e
ae
f
a
mi
ly
[
1]
,
whic
h
a
r
e
known
to
ha
ve
a
number
of
he
a
lt
h
be
ne
f
it
s
[
2]
,
[
3
]
.
I
t
is
im
pe
r
a
ti
ve
to
unde
r
s
tand
the
gr
a
pe
v
a
r
iant
to
de
ter
mi
ne
the
be
s
t
c
ult
ivation
tec
hnique,
pos
s
ibl
e
qua
li
ty,
a
nd
c
omm
e
r
c
ial
va
lue
[
4]
,
[
5
]
.
Gr
a
pe
gr
owe
r
s
a
r
e
wor
king
on
e
ns
ur
ing
a
pr
e
c
is
e
identif
ica
ti
on
o
f
gr
a
pe
s
va
r
ieties
,
a
s
we
ll
a
s
de
te
r
mi
ning
how
to
gr
ow
c
utt
i
ngs
ba
s
e
d
on
e
a
c
h
va
r
iety
a
nd
c
a
lcula
ti
ng
it
s
s
uppl
y
pr
ice
.
T
he
va
r
ieties
of
gr
a
pe
s
c
a
n
be
dis
ti
nguis
he
d
on
th
e
ba
s
is
of
their
lea
f
s
ha
pe
s
[
6]
.
Gr
a
pe
g
r
owe
r
s
a
r
e
tr
ying
to
f
ind
the
be
s
t
wa
y
to
f
ind
a
p
r
e
c
is
e
identif
ica
ti
on
of
gr
a
pe
va
r
ieties
,
a
s
we
ll
a
s
to
de
ter
mi
ne
the
be
s
t
c
ult
ivation
tec
hnique
f
or
dif
f
e
r
e
nt
types
of
gr
a
pe
va
r
iants
with
the
be
s
t
c
omm
e
r
c
ial
va
lues
.
Va
r
ious
s
tud
ie
s
ha
ve
de
s
igned
dif
f
e
r
e
nt
methods
of
c
las
s
if
ying
lea
ve
s
of
va
r
ious
ty
pe
s
of
plants
ove
r
the
las
t
f
e
w
ye
a
r
s
.
S
ome
of
thes
e
methods
a
r
e
mas
k
a
lgo
r
it
hm
r
e
gion
-
ba
s
e
d
c
onvolut
ional
ne
ur
a
l
ne
twor
k
(
R
-
C
NN
)
a
nd
VG
G16
us
e
d
to
dis
ti
ng
uis
h
lea
f
s
ha
pe
s
[
7]
,
the
c
onvolut
ional
ne
ur
a
l
ne
twor
k
(
C
N
N)
tec
hnique
[
8]
,
C
NN
to
a
na
lyze
lea
f
dis
e
a
s
e
[
9]
,
[
10]
,
a
nd
the
s
tanda
r
d
R
e
s
Ne
t
-
50
C
NN
model’
s
a
tt
e
nti
on
r
e
s
idual
lea
r
ning
s
tr
a
tegy
(
AR
e
s
Ne
t
-
50)
[
11]
.
De
e
p
lea
r
ning
tec
hniques
,
whic
h
he
lp
c
las
s
if
y
objec
ts
mor
e
a
c
c
ur
a
tely,
a
r
e
us
e
d
by
mos
t
s
tudi
e
s
to
c
las
s
if
y
plant
lea
ve
s
.
Us
ing
de
e
p
lea
r
ning
methods
f
or
im
a
ge
r
e
c
ognit
ion
a
nd
a
ls
o
c
las
s
if
ica
ti
on,
ha
s
wide
ly
s
pr
e
a
d
in
r
e
s
e
a
r
c
h
[
12
]
–
[
14
]
.
Anothe
r
c
las
s
if
ica
ti
on
method,
C
NN
is
one
of
the
mos
t
c
omm
on
a
nd
wi
de
ly
us
e
d
de
e
p
lea
r
ning
models
that
ha
ve
be
e
n
pr
ove
n
to
h
a
ve
good
pe
r
f
or
manc
e
due
to
thei
r
e
xc
e
ll
e
nt
c
a
pa
bil
it
y
of
lea
r
ning
pr
ope
r
ti
e
s
of
a
n
objec
t
us
ing
a
lar
ge
num
be
r
of
ne
twor
k
a
r
c
hit
e
c
tur
e
s
[
15]
,
[
16]
.
M
e
a
nwhil
e
,
a
ne
w
method
s
ugge
s
ted
by
L
iu
e
t
al
.
[
17]
,
F
a
s
ter
R
-
C
NN
,
is
c
ur
r
e
ntl
y
be
ing
de
ve
loped
.
G
r
a
pe
lea
f
va
r
iants
a
r
e
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
ompar
is
on
of
fas
ter
r
e
gion
-
bas
e
d
c
onv
olut
ional
ne
ur
al
ne
tw
or
k
for
algor
it
hms
…
(
M
oe
c
hamm
ad
Sar
os
a
)
223
identif
ied
us
ing
a
de
e
p
lea
r
ning
method
wi
th
th
e
F
a
s
ter
R
-
C
NN
a
lgor
it
hm
c
omb
ined
with
the
I
nc
e
pti
on
R
e
s
Ne
t
V2,
R
e
s
Ne
t
-
152,
R
e
s
Ne
t
-
101,
a
nd
R
e
s
Ne
t
-
50
ne
tw
or
k
a
r
c
hit
e
c
tur
e
a
nd
us
e
p
r
e
-
tr
a
ined
C
OC
O
we
ight
s
.
T
his
r
e
s
e
a
r
c
h
e
mpl
oys
f
ive
types
of
g
r
a
pe
lea
ve
s
,
n
a
mely
a
c
a
de
mi
c
,
jupi
ter
,
loca
l,
taldun,
a
nd
t
r
a
ns
f
igur
a
ti
on
.
2.
M
E
T
HO
D
2
.
1.
Re
s
e
ar
c
h
d
at
a
c
oll
e
c
t
io
n
I
n
the
da
ta
c
oll
e
c
ti
on
pr
oc
e
s
s
,
im
a
ge
s
o
f
g
r
a
pe
lea
ve
s
a
r
e
take
n
a
nd
us
e
d
to
ge
ne
r
a
te
a
da
tas
e
t.
T
he
da
tas
e
t
c
ontains
500
im
a
ge
s
,
with
100
f
or
e
a
c
h
gr
a
pe
lea
f
va
r
iety.
T
his
r
e
s
e
a
r
c
h
f
oc
us
e
s
on
f
ive
gr
a
pe
va
r
ieties
,
whic
h
a
r
e
a
c
a
de
mi
c
s
,
jupi
ter
s
,
loca
l,
taldun,
a
nd
tr
a
ns
f
igur
a
ti
on
.
F
igur
e
s
1
(
a
)
to
1
(
e
)
s
hows
the
gr
a
p
e
lea
f
in
one
of
our
r
e
s
e
a
r
c
he
r
s
'
ga
r
de
ns
.
T
he
r
e
s
olut
ion
of
the
im
a
ge
s
is
a
djus
ted
to
the
c
ondit
ions
s
e
t
o
ut
in
the
pr
e
c
ondit
ioni
ng
we
ight
,
na
mely
the
C
OC
O
pr
e
t
r
a
ined
method.
T
he
a
djus
tm
e
nt
is
made
a
f
ter
c
oll
e
c
ti
ng
da
ta
on
gr
a
pe
plant
lea
f
im
a
ge
s
.
C
OC
O
ha
s
a
pr
e
-
tr
a
ined
we
ight
of
640×
640
.
F
igur
e
1
,
whic
h
r
e
pr
e
s
e
nts
e
a
c
h
c
las
s
,
s
hows
e
xa
mpl
e
s
of
the
im
a
ge
s
that
a
r
e
a
ppli
e
d
to
t
his
da
tas
e
t.
C
las
s
he
r
e
r
e
f
e
r
s
to
the
type
of
gr
a
pe
v
a
r
iety.
(
a
)
(
b)
(
c
)
(
d)
(
e
)
F
igur
e
1.
Da
tas
e
ts
(
a
)
a
c
a
de
mi
c
,
(
b)
jupi
ter
,
(
c
)
loc
a
l,
(
d
)
taldun
,
a
nd
(
e
)
tr
a
ns
f
igur
a
ti
on
2
.
2
.
Ann
o
t
at
ion
an
d
lab
e
li
n
g
T
he
p
r
oc
e
s
s
of
mar
king
the
gr
a
pe
lea
f
in
a
pictur
e
i
s
c
a
ll
e
d
a
nnotation
a
nd
labe
li
ng.
T
he
ini
ti
a
l
s
tep
is
to
dr
a
w
a
box
on
the
lea
f
s
ur
f
a
c
e
of
the
gr
a
pe
pl
a
nt
a
nd
mar
k
e
a
c
h
of
the
boxe
s
with
labe
ls
"
Ac
a
de
mi
c
"
,
"
J
upit
e
r
"
,
"
L
oc
a
l"
,
"
T
a
ldun"
,
a
nd
"
T
r
a
ns
f
igu
r
a
ti
o
n"
.
A
f
il
e
wi
th
a
n
e
xtens
ion
'.
xml
'
is
ge
ne
r
a
ted
b
y
im
a
ge
a
nnotation
a
nd
labe
li
ng
in
the
P
ASC
AL
VO
C
f
or
mat.
T
he
a
nnotation
a
nd
la
be
li
ng
pha
s
e
pr
oduc
e
s
a
ba
s
e
li
ne
tr
uth
whic
h
is
then
us
e
d
to
c
a
lcula
te
the
r
e
g
r
e
s
s
ion
los
s
of
bounding
box
de
tec
ti
on
point
s
on
objec
ts
de
tec
ted
dur
ing
tr
a
ini
ng.
T
he
pr
oc
e
s
s
e
mpl
oys
im
a
ge
s
of
t
wa
r
e
.
An
e
xa
mpl
e
of
the
type
of
a
nnotation
a
nd
labe
li
ng
is
s
hown
in
F
igur
e
2
.
F
igur
e
2
.
L
a
be
li
ng
pr
oc
e
s
s
us
ing
s
of
twa
r
e
c
a
ll
e
d
L
a
be
lI
mg
S
ubs
e
que
ntl
y,
the
da
ta
is
divi
de
d
int
o
s
e
ve
r
a
l
c
a
teg
or
ies
,
s
uc
h
a
s
tes
ti
ng
a
nd
t
r
a
ini
ng
da
ta
,
onc
e
a
ll
the
im
a
ge
s
ha
ve
be
e
n
a
nnotate
d
a
nd
labe
led.
T
a
ble
1
s
hows
the
pa
r
ti
ti
on
s
c
he
me
f
or
the
da
tas
e
ts
.
I
n
tr
a
i
ning,
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.
1
,
F
e
br
ua
r
y
2025
:
22
2
-
230
224
r
e
s
ult
of
im
a
ge
a
nnotation
a
nd
labe
li
ng
may
not
be
dir
e
c
tl
y
us
e
d.
T
o
be
us
e
d
in
tr
a
ini
ng
a
nd
modeling
a
c
ti
vit
ies
,
the
r
e
s
ult
ing
a
nnotation
a
nd
labe
l
f
il
e
s
that
a
r
e
ge
n
e
r
a
ted
with
a
n
e
xtens
ion
f
or
mat
'.
x
ml
'
ha
ve
to
be
c
onve
r
ted
int
o
'.
c
s
v'
f
il
e
f
o
r
mats
.
T
a
ble
2
i
ll
us
tr
a
tes
the
c
onve
r
s
ion
r
e
s
ult
s
f
or
the
f
il
e
to
'.
c
s
v'.
T
a
ble
1.
T
a
ble
o
f
d
a
tas
e
t
s
pli
ts
E
xp
s
c
he
ma
ti
c
-
T
ot
a
l
t
r
a
in
in
g
d
a
ta
T
ot
a
l
t
e
s
t
d
a
ta
1
400
50
2
450
50
T
a
bl
e
2
.
C
onve
r
s
ion
r
e
s
ult
s
of
XM
L
to
C
S
V
f
il
e
s
F
il
e
na
me
W
id
e
T
a
ll
C
la
s
s
X
mi
n
Y
mi
n
X
ma
x
Y
ma
x
A
c
a
de
mi
c
5.j
pg
640
640
a
c
a
de
mi
c
55
142
543
622
J
upi
te
r
6.j
pg
640
640
ju
pi
te
r
87
6
615
596
L
oc
a
l4
.j
pg
640
640
lo
c
a
l
6
9
590
512
T
a
ld
un3.jpg
640
640
ta
ld
un
22
7
507
537
T
r
a
ns
f
ig
ur
a
ti
on2.jpg
640
640
tr
a
ns
f
ig
ur
a
ti
on
114
85
635
545
2
.
3
.
M
od
e
li
n
g
F
a
s
ter
R
-
C
N
N
is
a
n
im
pr
ove
d
C
NN
method
that
is
de
ve
loped
f
r
om
R
-
C
NN
a
nd
F
a
s
t
R
-
C
NN
[
18]
–
[
20]
.
W
ha
t
s
e
ts
thi
s
method
a
pa
r
t
f
r
om
it
s
pr
e
de
c
e
s
s
or
s
is
the
upgr
a
de
f
r
om
the
s
e
lec
ti
ve
s
e
a
r
c
h
f
e
a
tur
e
to
a
r
e
gion
pr
opos
a
l
ne
twor
k
(
R
P
N)
[
21]
.
T
his
s
t
udy
e
mpl
oys
the
F
a
s
ter
R
-
C
NN
a
lgor
it
hm
with
I
nc
e
pti
on
R
e
s
Ne
t
V2,
R
e
s
Ne
t
-
152,
R
e
s
Ne
t
-
101,
a
nd
R
e
s
N
e
t
-
50
a
r
c
hit
e
c
tur
e
.
A
modeling
e
xe
r
c
is
e
wa
s
c
a
r
r
ied
out
be
f
or
e
the
e
xpe
r
im
e
nt
us
ing
Google
C
oll
a
bor
a
to
r
y
too
ls
.
F
igur
e
3
s
hows
a
model
of
the
F
a
s
ter
R
-
C
NN
a
lgor
it
hm.
F
igur
e
3
.
M
ode
l
of
the
F
a
s
ter
R
-
C
NN
a
lgor
it
hm
il
l
us
tr
a
ti
on
T
he
F
a
s
ter
R
-
C
NN
a
r
c
hit
e
c
tur
a
l
model
be
gins
wit
h
input
ti
ng
a
n
im
a
ge
a
nd
the
lea
f
f
e
a
tur
e
s
whic
h
then
a
r
e
r
e
c
ove
r
e
d
us
ing
ba
c
kbone
C
N
N.
T
he
C
NN
known
a
s
a
tec
hnology
that
u
s
e
s
a
s
e
t
of
c
onvolu
ti
ona
ls
to
r
e
tr
ieve
f
e
a
tur
e
s
in
or
de
r
[
22
]
to
c
r
e
a
te
laye
r
s
of
c
a
r
e
on
the
f
inal
r
e
s
ult
a
t
e
a
c
h
pha
s
e
of
tr
a
ini
ng
[
23]
.
T
his
method
c
a
n
identif
y
the
it
e
ms
r
e
pr
e
s
e
nted
in
a
n
im
a
ge
[
24]
.
T
he
r
e
s
e
a
r
c
h
a
ls
o
tes
ted
the
or
ga
nizing
c
a
pa
c
it
y
of
C
OC
O,
c
omput
e
r
vis
ion,
I
mage
Ne
t,
a
nd
na
tur
a
l
la
ngua
ge
pr
oc
e
s
s
ing
(
NL
P
)
to
us
e
lar
ge
im
a
ge
c
las
s
if
ica
ti
on
da
ta.
T
he
C
NN
ba
c
kbone
s
us
e
d
in
thi
s
s
tudy
a
r
e
I
nc
e
pti
on
R
e
s
Ne
t
V2,
R
e
s
Ne
t
-
152,
R
e
s
Ne
t
-
101,
a
nd
R
e
s
Ne
t
-
50.
R
e
s
N
e
t
wa
s
one
od
th
e
be
s
t
de
e
p
n
e
ur
a
l
ne
twor
ks
in
the
2015
c
las
s
if
ica
ti
on
c
ompetit
ion
know
a
s
I
M
AG
E
NE
T
lar
ge
s
c
a
le
vis
ua
l
r
e
c
ognit
ion
c
ompeti
ti
on
(
I
L
S
VR
C
2015)
.
R
e
s
Ne
t
-
18,
R
e
s
Ne
t
-
50,
101,
1
10,
152
,
164
,
a
nd
R
e
s
Ne
t
-
1202
a
r
e
a
ll
identica
l
va
r
iants
wi
th
a
va
r
iable
number
of
laye
r
s
[
25]
.
T
he
las
t
numb
e
r
in
the
R
e
s
Ne
t
a
r
c
hit
e
c
tur
e
na
me
indi
c
a
tes
the
tot
a
l
laye
r
s
of
the
a
r
c
hit
e
c
tur
e
f
or
f
e
a
tur
e
e
xt
r
a
c
ti
on
in
a
n
im
a
ge
.
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
ompar
is
on
of
fas
ter
r
e
gion
-
bas
e
d
c
onv
olut
ional
ne
ur
al
ne
tw
or
k
for
algor
it
hms
…
(
M
oe
c
hamm
ad
Sar
os
a
)
225
B
e
ginni
ng
while
the
I
nc
e
pti
on
c
onvolut
ion
laye
r
in
R
e
s
Ne
t
V2
c
ombi
ne
s
the
I
nc
e
pti
on
a
ls
o
R
e
s
Ne
t
laye
r
s
,
R
e
s
Ne
t
V2
is
a
blend
of
both
laye
r
s
.
I
nc
e
pti
on
R
e
s
Ne
t
V2
is
a
mi
xtur
e
of
the
I
nc
e
pti
on
a
nd
R
e
s
N
e
t
laye
r
s
,
whe
r
e
a
s
the
I
nc
e
pti
on
c
onvolut
ion
laye
r
R
e
s
Ne
t
V2
is
a
c
ombi
na
ti
on
of
the
I
nc
e
pti
on
a
nd
R
e
s
N
e
t
laye
r
s
[
26]
.
T
he
R
P
N
,
whic
h
r
e
quir
e
s
f
utur
e
map
output
f
r
om
a
ba
c
kbone
ne
twor
k,
is
the
ne
xt
s
e
c
ti
on.
T
o
dis
play
the
"
Anc
hor
s
"
s
e
t
f
o
r
e
a
c
h
loca
ti
on
in
the
Output
f
e
a
tur
e
ma
p,
R
P
N
is
pe
r
f
or
med
by
putt
ing
it
on
a
n
input
im
a
ge
.
T
he
a
nc
hor
s
s
how
dif
f
e
r
e
nt
s
ize
s
a
nd
r
a
ti
o
s
of
objec
ts
s
hown
by
the
im
a
ge
s
.
F
or
P
ASC
AL
VO
C
,
the
a
nc
hor
s
ha
ve
thr
e
e
s
c
a
le
box
s
ize
s
(
128²,
256²,
51
2²)
a
nd
thr
e
e
a
s
pe
c
t
r
a
ti
os
(
1:1
,
1:2
,
a
nd
2:1)
,
so
t
he
r
e
a
r
e
nine
pos
s
ibl
e
a
nc
hor
s
plac
e
d
on
the
im
a
ge
inpu
t
t
o
the
output
f
e
a
tur
e
map
[
27
]
.
T
he
output
o
f
thi
s
pr
oc
e
s
s
de
ter
mi
ne
s
the
pr
oba
bil
it
y
that
a
ny
of
the
nine
a
nc
hor
point
s
on
the
ba
c
kbone
f
e
a
tur
e
map
c
ontains
o
bjec
ts
a
t
that
point
[
28]
.
T
he
thi
r
d
pa
r
t
is
the
r
e
gion
of
int
e
r
e
s
t
(
R
OI
)
poll
in
g
laye
r
,
whic
h
us
e
s
a
maximum
poll
ing
ope
r
a
ti
on
to
c
oll
e
c
t
f
e
a
tur
e
s
f
r
om
the
f
e
a
tur
e
map
a
nd
to
c
ha
nge
their
s
ize
to
a
f
ixed
s
ize
.
A
s
ingl
e
-
dim
e
ns
ional
f
e
a
tu
r
e
ve
c
tor
c
ontaining.
T
he
R
OI
input
f
or
the
laye
r
that
f
ull
y
c
onne
c
ted
will
be
the
poll
ing
laye
r
’
s
output
o
r
ga
nize
d
a
s
a
s
ingl
e
-
dim
e
ns
ional
f
e
a
tur
e
ve
c
tor
.
Af
ter
going
thr
ough
the
laye
r
s
that
f
ull
y
c
onne
c
ted,
the
f
e
a
tur
e
s
a
r
e
a
ls
o
f
e
d
int
o
the
r
e
g
r
e
s
s
ion
a
nd
c
las
s
if
ica
ti
on
br
a
nc
he
s
i
n
the
f
inal
s
e
c
ti
on,
whic
h
p
r
e
dicts
the
objec
t’
s
c
or
r
e
c
t
matc
h
.
T
his
wa
y,
it
is
pos
s
ibl
e
,
f
or
e
xa
mpl
e
,
to
ge
ne
r
a
te
a
n
im
a
ge
of
objec
ts
with
the
de
s
ignation
of
boundi
ng
boxe
s
a
nd
a
pos
s
ibl
e
c
las
s
if
ica
ti
on
r
e
s
ult
[
28]
.
2
.
4
.
M
e
as
u
r
e
m
e
n
t
T
r
ue
pos
it
ive
(
T
P
)
,
f
a
ls
e
pos
it
ive
(
F
P
)
,
a
nd
f
a
ls
e
ne
ga
ti
ve
(
F
N)
va
lues
we
r
e
obtaine
d
f
r
om
the
mea
s
ur
e
ment
method
us
e
d
to
tes
t
the
gr
a
pe
lea
f
c
las
s
if
ica
ti
on
s
ys
tem.
FP
de
nor
e
s
that
bounding
box
i
de
nti
f
ied
objec
ts
but
f
a
il
e
d
to
identif
y
gr
a
pe
lea
ve
s
,
while
FN
indi
c
a
tes
that
the
bounding
box
did
not
c
ontain
a
ny
objec
ts
in
the
pr
ovided
f
igur
e
.
TP
de
notes
that
the
boundin
g
box
de
tec
ted
gr
a
pe
plan
lea
ve
s
s
uc
c
e
s
s
f
ull
y
[
29]
,
[
30]
.
F
1
s
c
or
e
s
,
r
e
c
a
ll
,
a
nd
pr
e
c
is
ion
a
r
e
c
omput
e
d
us
ing
t
he
s
e
pa
r
a
mete
r
s
.
R
e
c
a
ll
is
the
de
gr
e
e
of
de
tec
ti
on
s
uc
c
e
s
s
,
whe
r
e
a
s
pr
e
c
is
ion
is
the
a
c
c
ur
a
c
y
of
the
de
tec
ti
on
r
e
s
ult
s
.
T
he
F
1
s
c
or
e
wa
s
f
ound
to
ha
ve
a
ba
lanc
e
be
twe
e
n
r
e
c
a
ll
a
nd
pr
e
c
is
ion.
T
he
f
oll
owing
f
or
mul
a
wa
s
us
e
d
to
de
ter
mi
ne
the
F
1
s
c
or
e
s
,
r
e
c
a
ll
,
a
nd
pr
e
c
is
ion
[
29]
.
P
r
e
cis
io
n
=
TP
TP
+
FP
(
1)
Re
ca
l
l
=
TP
TP
+
FN
(
2)
F1
S
co
r
e
=
2
×
P
r
e
c
i
s
i
on
×
R
e
c
a
ll
P
r
e
c
i
s
i
on
+
R
e
c
a
ll
(
3)
3.
RE
S
UL
T
S
AN
D
DI
S
CU
S
S
I
ON
3.
1.
Dat
a
t
r
ain
in
g
p
r
oc
e
s
s
c
om
p
u
t
in
g
t
i
m
e
T
he
c
omput
ing
ti
me
is
the
ti
me
ne
e
de
d
by
a
c
omp
uter
to
p
r
oc
e
s
s
a
n
a
lgor
it
hm
a
nd
tr
a
in
da
ta.
T
a
ble
3
s
hows
the
tr
a
ini
ng
ti
me
r
e
quir
e
d
f
or
e
a
c
h
e
xpe
r
im
e
nt.
F
ur
ther
mor
e
,
the
I
nc
e
pti
on
R
e
s
Ne
t
V2
a
r
c
hit
e
c
t
ur
e
take
s
the
longes
t
to
c
ompl
e
te
the
tr
a
ini
ng
,
while
the
R
e
s
Ne
t
-
50
a
r
c
hit
e
c
tur
e
ha
s
the
f
a
s
tes
t
tr
a
ini
ng
ti
me.
T
a
ble
3
.
T
r
a
in
ing
pr
oc
e
s
s
c
omput
ing
ti
me
N
e
twor
k
a
r
c
hi
te
c
tu
r
e
E
xpe
r
im
e
nt
t
o
-
C
omput
e
T
im
e
(
mi
nut
e
)
3,000
4,000
5,000
I
nc
e
pt
io
n R
e
s
ne
t
V
2
1
69
69
73
2
81
85
91
R
e
s
N
e
t
-
152
1
25
31
43
2
35
35
61
R
e
s
N
e
t
-
101
1
25
31
41
2
31
31
49
R
e
s
N
e
t
-
50
1
19
29
39
2
29
29
47
3.
2
.
Dat
a
t
e
s
t
in
g
p
r
oc
e
s
s
c
om
p
u
t
in
g
t
im
e
E
a
c
h
ne
twor
k
a
r
c
hit
e
c
tur
e
f
or
the
tes
ti
ng
p
r
oc
e
s
s
r
e
quir
e
s
va
r
ying
c
omput
a
t
ional
ti
me
de
pe
nding
on
the
c
ompl
e
xit
y
of
the
model.
T
e
s
ti
ng
t
im
e
on
the
I
n
c
e
pti
on
R
e
s
Ne
t
V2
a
r
c
hit
e
c
tu
r
e
tends
to
be
longer
c
ompar
e
d
to
other
a
r
c
hit
e
c
tur
e
s
,
with
a
n
a
ve
r
a
ge
tes
ti
ng
ti
m
e
r
e
a
c
hing
mor
e
than
100
s
e
c
onds
.
T
a
ble
4
s
hows
that
the
c
omput
a
ti
ona
l
c
os
t
of
the
I
nc
e
pti
on
R
e
s
Ne
t
V
2
a
r
c
hit
e
c
tur
e
is
much
higher
,
making
it
a
n
i
mpor
tant
c
ons
ider
a
ti
on
in
f
ield
a
ppli
c
a
ti
ons
,
e
ve
n
though
it
ha
s
be
tt
e
r
pe
r
f
or
manc
e
in
ter
ms
o
f
de
tec
ti
on
a
c
c
ur
a
c
y.
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.
1
,
F
e
br
ua
r
y
2025
:
22
2
-
230
226
T
a
ble
4
.
T
e
s
ti
ng
p
r
oc
e
s
s
c
omput
ing
ti
me
N
e
twor
k
a
r
c
hi
te
c
tu
r
e
E
xpe
r
im
e
nt
t
o
-
C
omput
e
T
im
e
(
s
e
c
ond)
3,000
4,000
5,000
I
nc
e
pt
io
n R
e
s
ne
t
V
2
1
107
103
101
2
101
103
101
R
e
s
N
e
t
-
152
1
78
78
80
2
80
73
73
R
e
s
N
e
t
-
101
1
55
58
57
2
58
55
65
R
e
s
N
e
t
-
50
1
48
45
45
2
44
43
42
3.
3
.
T
ot
al
los
s
whi
le
m
od
e
li
n
g
A
ke
y
r
e
s
ult
o
f
the
modeling
pr
oc
e
s
s
f
or
g
r
a
pe
lea
f
s
ha
pe
s
is
the
los
s
f
unc
t
ion.
W
e
a
kne
s
s
e
s
a
r
e
f
ound
a
t
e
ve
r
y
s
tage
of
the
modeling
pr
oc
e
s
s
.
T
a
ble
5
s
hows
whe
r
e
the
dif
f
e
r
e
nc
e
be
twe
e
n
the
pe
r
f
o
r
manc
e
s
of
the
modeling
pr
oc
e
s
s
oc
c
ur
s
.
R
e
ga
r
ding
the
a
ve
r
a
ge
lo
s
s
of
the
dif
f
e
r
e
nt
a
r
c
hit
e
c
tur
e
s
,
I
nc
e
pti
on
R
e
s
Ne
t
V2
s
hows
the
lowe
s
t
a
ve
r
a
ge
los
s
,
while
the
a
ve
r
a
ge
los
s
of
R
e
s
Ne
t
-
50
is
the
highes
t.
T
a
ble
5
.
L
os
s
of
F
a
s
ter
R
-
C
NN
in
the
modeling
p
r
oc
e
s
s
S
te
p
I
nc
e
pt
io
n R
e
s
ne
t
V
2
R
e
s
N
e
t
-
152
R
e
s
N
e
t
-
101
R
e
s
N
e
t
-
50
e
xpe
r
im
e
nt
t
o
-
e
xpe
r
im
e
nt
t
o
-
e
xpe
r
im
e
nt
t
o
-
e
xpe
r
im
e
nt
t
o
-
1
2
1
2
1
2
1
2
3
,
000
0.0872
0.0723
0.1065
0.1016
0.1375
0.1302
0.1507
0.1464
4
,
000
0.1052
0.0859
0.0825
0.0767
0.1232
0.1114
0.1361
0.1346
5
,
000
0.1096
0.1055
0.1334
0.1287
0.1398
0.1328
0.1146
0.1096
A
ve
r
a
ge
0.1007
0.0879
0.1074
0.1023
0.1335
0.1248
0.1338
0.1302
A
ve
r
a
ge
0.0943
0.1049
0.1291
0.1320
3.
4
.
T
r
u
e
p
os
it
ive
,
f
a
ls
e
p
os
it
ive,
a
n
d
f
als
e
n
e
ga
t
ive
t
e
s
t
r
e
s
u
lt
s
T
he
T
P
,
F
P
,
a
nd
F
N
tes
t
r
e
s
ult
s
s
how
the
s
ys
tem's
a
bil
it
y
to
identif
y
a
nd
c
las
s
if
y
g
r
a
pe
lea
f
objec
ts
.
T
he
tes
t
r
e
s
ult
s
in
T
a
ble
6
on
the
I
nc
e
pti
on
R
e
s
Ne
t
V2
a
r
c
hit
e
c
tur
e
s
how
a
higher
s
uc
c
e
s
s
r
a
te
in
de
tec
ti
ng
gr
a
pe
lea
ve
s
with
higher
T
P
va
lues
a
nd
lowe
r
F
P
a
nd
F
N
c
ompar
e
d
to
other
a
r
c
hit
e
c
tur
e
s
.
T
his
indi
c
a
tes
that
thi
s
model
is
a
ble
to
r
e
c
ognize
objec
ts
mo
r
e
a
c
c
ur
a
tely,
a
lt
hough
it
r
e
quir
e
s
mor
e
c
omput
a
ti
on
ti
me
,
making
it
a
be
tt
e
r
c
hoice
f
o
r
a
ppli
c
a
ti
ons
with
high
p
r
e
c
is
ion
r
e
quir
e
ments
.
T
a
ble
6
.
T
P
,
F
P
,
a
nd
F
N
t
e
s
t
r
e
s
ult
s
N
e
twor
k
a
r
c
hi
te
c
tu
r
e
E
xpe
r
im
e
nt
t
o
-
S
te
p
TP
FP
FN
I
nc
e
pt
io
n R
e
s
N
e
t
V
2
1
3,000
44
5
1
4,000
42
7
1
5,000
42
6
2
2
3,000
44
5
1
4,000
44
5
1
5,000
43
6
1
R
e
s
N
e
t
-
152
1
3,000
39
9
2
4,000
41
7
2
5,000
40
8
2
2
3,000
40
8
2
4,000
42
6
2
5,000
42
7
1
R
e
s
N
e
t
-
101
1
3,000
36
11
3
4,000
39
9
2
5,000
39
10
1
2
3,000
39
8
3
4,000
41
8
1
5,000
41
8
1
R
e
s
N
e
t
-
50
1
3,000
34
13
3
4,000
37
11
2
5,000
39
9
2
2
3,000
36
11
3
4,000
38
10
2
5,000
41
7
2
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
ompar
is
on
of
fas
ter
r
e
gion
-
bas
e
d
c
onv
olut
ional
ne
ur
al
ne
tw
or
k
for
algor
it
hms
…
(
M
oe
c
hamm
ad
Sar
os
a
)
227
3.
5
.
P
r
e
c
is
ion
,
r
e
c
all
,
a
n
d
F
1
s
c
or
e
t
e
s
t
r
e
s
u
lt
s
T
he
s
c
or
e
f
o
r
a
c
c
ur
a
c
y,
r
e
c
a
ll
,
a
nd
F
1
is
c
a
lcula
ted
us
ing
the
r
e
s
ult
s
of
T
P
,
F
P
,
a
nd
F
N
in
T
a
ble
6.
F
o
r
the
F
a
s
ter
R
-
C
NN
a
r
c
hit
e
c
tur
a
l
ne
twor
k
models
,
T
a
ble
7
c
ontains
a
c
c
ur
a
c
y,
r
e
c
a
ll
,
a
nd
F
1
s
c
or
e
s
.
Us
ing
the
F
a
s
ter
R
-
C
NN
a
lgor
it
hm,
the
c
omput
a
ti
ons
de
m
ons
tr
a
te
the
ince
pti
on
of
R
e
s
Ne
t
V2,
R
e
s
Ne
t
-
152,
R
e
s
Ne
t
-
101,
a
nd
R
e
s
Ne
t
-
50
ne
twor
k
a
r
c
hit
e
c
tur
a
l
models
a
c
hi
e
ve
a
nd
a
ve
r
a
ge
F
1
s
c
or
e
of
93%
,
90%
,
88
%
,
a
nd
86%
,
r
e
s
pe
c
ti
ve
ly.
T
his
indi
c
a
tes
the
F
a
s
ter
R
-
C
NN
a
lgo
r
it
hm,
I
nc
e
pti
on
R
e
s
Ne
t
V2
c
a
n
de
tec
t
a
nd
c
las
s
if
y
objec
ts
mor
e
e
f
f
e
c
ti
ve
ly
.
T
a
ble
7
.
P
r
e
c
is
ion,
r
e
c
a
ll
,
a
nd
F
1
s
c
or
e
tes
t
r
e
s
ult
s
N
e
twor
k
a
r
c
hi
te
c
tu
r
e
E
xpe
r
im
e
nt
t
o
-
S
te
p
P
r
e
c
is
io
n
(%)
R
e
c
a
ll
(%)
F
1
s
c
or
e
(%)
I
nc
e
pt
io
n R
e
s
N
e
t
V
2
1
3,000
90
98
94
4,000
86
98
91
5,000
88
95
91
A
ve
r
a
ge
92
2
3,000
90
98
94
4,000
90
98
94
5,000
88
98
92
A
ve
r
a
ge
93
A
ve
r
a
ge
92
R
e
s
N
e
t
-
152
1
3,000
81
95
88
4,000
85
95
90
5,000
83
95
89
A
ve
r
a
ge
89
2
3,000
83
95
89
4,000
88
95
91
5,000
86
98
91
A
ve
r
a
ge
9
0
A
ve
r
a
ge
90
R
e
s
N
e
t
-
101
1
3,000
77
92
84
4,000
81
95
88
5,000
80
98
88
A
ve
r
a
ge
86
2
3,000
83
93
88
4,000
84
98
90
5,000
84
98
90
A
ve
r
a
ge
89
A
ve
r
a
ge
88
R
e
s
N
e
t
-
50
1
3,000
72
92
81
4,000
77
95
85
5,000
81
95
88
A
ve
r
a
ge
85
2
3,000
77
92
84
4,000
79
95
86
5,000
85
95
90
A
ve
r
a
ge
87
A
ve
r
a
ge
86
3.
6
.
E
f
f
e
c
t
s
o
n
F
1
s
c
or
e
r
e
s
u
lt
s
of
t
h
e
e
n
t
ire
n
u
m
b
e
r
of
t
r
ain
in
g
s
t
e
p
s
F
igur
e
s
4
(
a
)
a
nd
4(
b
)
s
hows
the
e
f
f
e
c
t
of
the
to
tal
number
of
tr
a
in
ing
s
teps
a
s
a
f
unc
ti
on
of
F
1
s
c
or
e
r
e
s
ult
s
in
both
ini
ti
a
l
a
nd
s
ubs
e
que
nt
tr
ials
.
T
he
r
e
is
a
tr
a
ini
ng
s
tep
li
mi
t
of
3
,
000
s
teps
on
the
I
nc
e
pti
on
R
e
s
Ne
t
V2
ne
twor
k
a
r
c
hit
e
c
tur
e
.
F
o
r
R
e
s
Ne
t
-
152
a
nd
R
e
s
Ne
t
-
101
ne
twor
k
a
r
c
hit
e
c
tur
e
s
,
the
tr
a
ini
ng
s
tep
li
mi
t
is
a
t
4
,
000
s
teps
.
F
or
R
e
s
Ne
t
50,
the
li
mi
t
on
tr
a
in
ing
s
teps
is
not
ye
t
known
be
c
a
us
e
the
F
1
s
c
or
e
c
ons
tantly
r
is
e
s
a
s
a
ddit
ional
tr
a
ini
ng
s
teps
a
r
e
a
dde
d.
T
he
ov
e
r
a
ll
li
mi
t
of
the
t
r
a
ini
ng
s
tep
is
s
e
t
s
ince
th
e
F
1
s
c
or
e
tends
to
be
s
table
or
de
c
r
e
a
s
ing
a
t
a
given
point
in
the
e
x
e
r
c
is
e
.
3.
7.
T
h
e
e
f
f
e
c
t
s
of
ave
r
age
los
s
on
ave
r
age
F
1
s
c
or
e
F
igur
e
5
s
hows
the
mea
n
im
pa
c
t
of
a
c
ompl
e
te
los
s
on
a
n
F
1
s
c
or
e
.
T
he
model
is
c
ons
ider
e
d
to
pe
r
f
or
m
be
tt
e
r
in
e
xa
mi
ning
th
e
c
ha
r
a
c
ter
is
ti
c
s
of
a
n
obje
c
t
be
c
a
us
e
the
ge
ne
r
a
l
mea
n
los
s
dur
ing
tr
a
ini
ng
is
lowe
r
.
He
nc
e
,
we
c
a
n
s
e
e
a
n
incr
e
a
s
e
in
the
a
ve
r
a
ge
F
1
s
c
or
e
.
A
f
a
il
ur
e
to
de
tec
t
gr
a
pe
lea
f
type
is
ge
ne
r
a
ll
y
due
to
a
n
objec
t
be
ing
too
s
hor
t,
unde
f
inable
,
a
mbi
guous
,
or
due
to
the
li
ght
a
f
f
e
c
ti
ng
the
im
a
ge
take
n.
3.
8
.
De
t
e
c
t
ion
r
e
s
u
lt
s
Along
with
the
r
e
s
ult
s
of
the
a
lgor
it
hm
c
ompar
is
on
that
ha
s
be
e
n
c
a
r
r
ied
out,
we
made
the
c
hoice
to
us
e
the
F
a
s
ter
R
-
C
NN
R
e
s
N
e
t
-
101
a
lgor
it
hm
to
tes
t
the
de
tec
ti
on
a
c
c
ur
a
c
y
of
gr
a
pe
lea
f
im
a
g
e
s
by
c
o
ns
ider
ing
the
e
r
r
or
r
e
s
ult
s
dur
ing
modeling
,
tes
ti
ng
c
omput
in
g
ti
me,
a
nd
the
a
ve
r
a
ge
F
1
s
c
or
e
.
F
igur
e
s
6
(
a
)
to
6(
e
)
s
hows
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.
1
,
F
e
br
ua
r
y
2025
:
22
2
-
230
228
a
n
e
xa
mpl
e
o
f
the
de
tec
ti
on
a
c
c
ur
a
c
y
of
thi
s
r
e
s
e
a
r
c
h
tes
t.
I
n
a
ddit
ion
,
the
im
a
ge
de
tec
ti
on
a
c
c
ur
a
c
y
ha
s
r
e
a
c
he
d
100%
,
whic
h
s
hows
that
the
s
ys
tem
is
a
ble
to
a
na
lyze
the
model
a
c
c
ur
a
tely
a
nd
de
tec
t
a
nd
c
a
tegor
ize
it
c
or
r
e
c
tl
y.
De
tec
ti
on
e
r
r
or
due
to
the
da
tas
e
t
not
be
i
ng
va
r
ied
a
nd
the
li
ghti
ng
whe
n
taking
the
pictur
e
.
(
a
)
(
b)
F
igur
e
4.
E
f
f
e
c
ts
of
tr
a
ini
ng
on
the
F
1
s
c
or
e
(
a
)
f
ir
s
t
tr
ial
a
nd
(
b
)
s
e
c
ond
tr
ial
F
igur
e
5.
T
he
e
f
f
e
c
t
o
f
a
ve
r
a
ge
los
s
on
a
ve
r
a
ge
F
1
s
c
or
e
(
a
)
(
b)
(
c
)
(
d)
(
e
)
F
igur
e
6.
De
tec
ti
on
r
e
s
ult
s
of
(
a
)
a
c
a
de
mi
c
,
(
b)
jup
it
e
r
,
(
c
)
loca
l
,
(
d)
taldun,
a
nd
(
e
)
tr
a
ns
f
igur
a
ti
on
4.
CONC
L
USI
ON
T
his
s
tudy
de
s
c
r
ibes
tec
hniques
f
or
c
las
s
if
ying
a
nd
making
us
e
of
the
F
a
s
ter
R
-
C
NN
a
lgor
it
hm
to
identif
y
g
r
a
pe
lea
ve
s
a
ls
o
with
the
R
e
s
Ne
t
V2,
R
e
s
Ne
t
-
152,
R
e
s
Ne
t
-
101,
a
nd
R
e
s
Ne
t
-
50
ne
twor
k
a
r
c
h
it
e
c
tur
e
s
,
a
ll
of
whic
h
s
e
r
ve
a
s
t
r
a
ined
ne
two
r
ks
f
o
r
f
e
a
tur
e
e
xtr
a
c
ti
on.
T
he
e
xpe
r
im
e
nt
r
e
s
ult
s
s
how
that
the
a
ve
r
a
ge
F
1
s
c
or
e
s
a
r
e
93%
,
90%
,
88%
,
a
nd
86%
r
e
s
pe
c
ti
ve
ly,
with
the
be
s
t
F
1
s
c
or
e
on
the
I
nc
e
pti
on
R
e
s
Ne
t
V2
ne
twor
k
a
r
c
hit
e
c
tur
e
(
a
n
a
ve
r
a
ge
los
s
of
0.
943)
.
H
owe
ve
r
,
the
ti
me
ne
e
de
d
f
or
t
r
a
ini
ng
a
nd
tes
ti
ng
is
f
a
r
mor
e
int
e
ns
ive
than
the
other
ne
twor
k
a
r
c
hit
e
c
tur
e
s
.
W
hil
e
the
a
r
c
hit
e
c
tur
e
with
the
f
a
s
tes
t
c
omput
in
g
ti
me
f
or
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
ompar
is
on
of
fas
ter
r
e
gion
-
bas
e
d
c
onv
olut
ional
ne
ur
al
ne
tw
or
k
for
algor
it
hms
…
(
M
oe
c
hamm
ad
Sar
os
a
)
229
tr
a
ini
ng
or
tes
ti
n
g
is
the
R
e
s
Ne
t
-
50
ne
twor
k
a
r
c
hit
e
c
tur
e
,
the
F
1
s
c
or
e
of
thi
s
a
r
c
hit
e
c
tur
e
is
the
lowe
s
t
c
ompar
e
d
to
other
s
.
F
ur
the
r
mor
e
,
the
tr
a
ini
ng
s
te
p
li
mi
t
f
o
r
the
I
nc
e
pti
on
R
e
s
Ne
t
V2
is
a
t
s
tep
3
,
000,
while
R
e
s
Ne
t
152
a
nd
R
e
s
Ne
t
-
101
ne
twor
k
a
r
c
hit
e
c
tur
e
s
ha
ve
4
,
000
s
tep
t
r
a
ini
ng
s
tep
l
im
it
.
M
e
a
nwhile,
the
tr
a
ini
ng
s
tep
li
mi
t
f
or
the
R
e
s
Ne
t
-
50
a
r
c
hit
e
c
tur
e
,
given
t
ha
t
the
F
1
s
c
or
e
c
onti
nue
s
to
incr
e
a
s
e
a
s
the
e
xe
r
c
is
e
s
tep
incr
e
a
s
e
s
,
is
unknown.
T
he
r
e
s
e
a
r
c
h
e
s
tablis
he
s
a
g
e
ne
r
a
l
li
mi
t
on
the
number
of
tr
a
ini
ng
s
teps
a
s
the
F
1
s
c
or
e
tends
to
s
tabili
z
e
or
de
c
li
ne
a
t
s
ome
point
du
r
ing
a
n
e
xe
r
c
is
e
.
I
t
c
a
n
be
c
onc
luded
that
the
F
a
s
ter
R
C
NN
-
ba
s
e
d
de
tec
ti
on
a
nd
c
las
s
if
ica
ti
on
s
ys
tem
f
o
r
gr
a
pe
lea
f
c
a
n
a
na
lyze
objec
t
pr
ope
r
ti
e
s
mor
e
e
f
f
e
c
ti
ve
ly
if
tot
a
l
los
s
e
s
dur
ing
tr
a
ini
n
g
a
r
e
lowe
r
e
d.
AC
KNOWL
E
DGE
M
E
NT
S
T
he
r
e
s
e
a
r
c
he
r
s
would
li
ke
to
e
xpr
e
s
s
gr
a
ti
tud
e
towa
r
ds
the
Dir
e
c
tor
a
te
Ge
ne
r
a
l
of
R
e
s
e
a
r
c
h
E
nha
nc
e
ment
a
nd
De
ve
lopm
e
nt,
M
ini
s
tr
y
of
E
du
c
a
ti
on,
C
ult
ur
e
,
R
e
s
e
a
r
c
h,
a
nd
T
e
c
hnology,
f
or
t
he
gr
a
nt
pr
ovided
with
De
c
r
e
e
Numbe
r
173/
S
P
K/D.
D4/
P
P
K.
01.
AP
T
V/VI
/2023
a
nd
Agr
e
e
ment/C
ontr
a
c
t
Numbe
r
12882/P
L
2.
1/HK/2023
.
RE
F
E
RE
NC
E
S
[
1]
M
. S
a
r
os
a
, P
. N
. M
a
a
’
r
if
a
h, M
. K
us
uma
w
a
r
da
ni
, a
nd D
. F
. A
l
R
iz
a
, “
V
it
is
V
in
e
r
a L
.
le
a
f
de
te
c
ti
on us
in
g
fa
s
te
r
R
-
C
N
N
,”
B
I
O
W
e
b
C
onf
e
r
e
nc
e
, vol
. 117, pp. 1
–
9, 2024, doi:
10.1051/bi
oc
onf
/2
02411701021.
[
2]
A
.
R
a
he
mi
, J
.
C
.
D
.
P
e
te
r
s
on,
a
nd
K
. T
.
L
und,
“
G
r
a
pe
S
pe
c
ie
s
,
”
in
G
r
ape
R
oot
s
to
c
k
s
and
R
e
la
te
d
Spe
c
ie
s
,
S
pr
in
ge
r
,
C
ha
m,
2
022,
pp. 5
–
21
,
doi
:
10.1007/978
-
3
-
030
-
99407
-
5_2.
[
3]
A
.
M
.
W
a
lk
e
r
,
C
.
H
e
in
it
z
,
S
.
R
ia
z
,
a
nd
J
.
U
r
e
t
s
ky,
“
G
r
a
pe
ta
xonomy
a
nd
ge
r
mp
la
s
m,”
in
T
he
G
r
ape
G
e
nom
e
,
S
pr
in
ge
r
,
C
ha
m,
2019, pp. 25
–
38
,
doi
:
10.1007/978
-
3
-
030
-
18601
-
2_2.
[
4]
G
.
H
a
s
a
na
li
ye
va
e
t
al
.
,
“
E
f
f
e
c
ts
of
pr
oduc
ti
on
r
e
gi
on,
pr
oduc
ti
on
s
ys
te
ms
a
nd
gr
a
pe
ty
pe
/v
a
r
ie
ty
on
nut
r
it
io
na
l
qua
li
ty
pa
r
a
me
te
r
s
of
t
a
bl
e
gr
a
pe
s
;
r
e
s
ul
ts
f
r
om a
U
K
r
e
ta
il
s
ur
ve
y,”
F
oods
, vol
. 9,
no. 12, 2020, doi:
10.3390/f
oods
9121874.
[
5]
B
.
S
ut
e
r
,
A
.
D
.
I
r
vi
ne
,
M
.
G
ow
dy,
Z
.
D
a
i,
a
nd
C
.
V.
L
e
e
u
w
e
n,
“
A
da
pt
in
g
w
in
e
gr
a
pe
r
ip
e
ni
ng
to
gl
oba
l
c
h
a
nge
r
e
qui
r
e
s
a
m
ul
ti
-
tr
a
it
a
ppr
oa
c
h,”
F
r
ont
ie
r
s
i
n P
la
nt
Sc
ie
nc
e
, vol
. 12, pp. 1
–
17, 2
021, doi:
10.3389/f
pl
s
.2021.624867.
[
6]
Y
.
K
ha
n
e
t
al
.
,
“
A
nt
io
xi
da
nt
pot
e
nt
ia
l
in
th
e
le
a
ve
s
of
gr
a
pe
va
r
ie
ti
e
s
(
V
it
is
v
in
if
e
r
a
L
.)
g
r
ow
n
in
di
f
f
e
r
e
nt
s
oi
l
c
ompos
it
io
ns
,”
A
r
abi
an J
our
nal
of
C
he
m
is
t
r
y
, vol
. 14, no. 11, 2021, doi
:
10.10
16/
j.
a
r
a
bj
c
.2021.103412.
[
7]
K
.
Y
a
ng,
W
.
Z
hong,
a
nd
F
.
L
i,
“
L
e
a
f
s
e
gme
nt
a
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on
w
it
h
a
c
ompl
ic
a
te
d
ba
c
kgr
ound
us
in
g
de
e
p
le
a
r
ni
ng,”
A
gr
onomy
, vol
. 10, no. 11, 2020, doi
:
10.3390/a
gr
onomy10111721.
[
8]
N
.
A
.
O
th
ma
n,
N
.
S
.
D
a
ma
nhur
i,
N
.
M
.
A
li
,
B
.
C
.
C
.
M
e
ng,
a
nd
A
.
A
.
A
.
S
a
ma
t,
“
P
la
nt
le
a
f
c
la
s
s
if
ic
a
ti
on
us
in
g
c
onvolut
i
ona
l
ne
ur
a
l
ne
twor
k
,”
2022
8t
h
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
C
ont
r
ol
,
D
e
c
is
io
n
and
I
nf
or
m
at
io
n
T
e
c
hnol
ogi
e
s
(
C
o
D
I
T
)
,
I
s
ta
nbul
,
T
ur
ke
y
,
pp. 1043
–
1048, 2022, doi:
10.1109/C
oD
I
T
55151.2022.9804121.
[
9]
J
. H
a
ng, D. Z
ha
ng, P
. C
he
n, J
. Z
ha
ng, a
nd B
. W
a
ng, “
C
la
s
s
if
ic
a
ti
on of
pl
a
nt
l
e
a
f
d
is
e
a
s
e
s
ba
s
e
d on im
pr
ove
d c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k,”
Se
ns
or
s
, vol
. 19, no. 19, pp. 1
–
14, 2019, doi:
10.3390/s
19194161.
[
10]
Y
.
T
oda
a
nd
F
.
O
kur
a
,
“
H
ow
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
s
di
a
gnos
e
pl
a
nt
di
s
e
a
s
e
,
”
P
la
nt
P
he
nom
ic
s
,
vol
.
2019,
2019,
doi
:
10.3413
3/
2019/9237136.
[
11]
A
.
P
a
nde
y
a
nd
K
.
J
a
in
,
“
P
la
nt
le
a
f
di
s
e
a
s
e
c
l
a
s
s
if
ic
a
ti
on
us
in
g
de
e
p
a
tt
e
nt
io
n
r
e
s
id
ua
l
ne
twor
k
opt
im
iz
e
d
by
oppos
it
io
n
-
ba
s
e
d
s
ymbi
ot
ic
or
ga
ni
s
ms
s
e
a
r
c
h
a
lg
or
it
hm,”
N
e
ur
al
C
om
put
in
g
and
A
ppl
ic
at
io
ns
,
vol
.
34,
pp.
21049
–
21066,
2022,
doi
:
10.1007/
s
00
521
-
022
-
07587
-
6.
[
12]
P
.
N
.
M
a
’
r
if
a
h,
M
.
S
a
r
os
a
,
a
nd
E
.
R
oha
di
,
“
G
a
r
ba
g
e
c
la
s
s
if
i
c
a
ti
on
us
in
g
F
a
s
te
r
R
-
C
N
N
,”
2023
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
E
le
c
tr
ic
al
and I
nf
or
m
at
io
n T
e
c
hnol
ogy
(
I
E
I
T
)
,
M
a
la
ng, I
ndone
s
ia
,
pp. 196
–
201, 2023, doi:
10.1109/I
E
I
T
59852.2023.10335519.
[
13]
Q
.
L
v,
S
.
Z
ha
ng,
a
nd
Y
.
W
a
ng,
“
D
e
e
p
le
a
r
ni
ng
mode
l
of
im
a
ge
c
la
s
s
if
ic
a
ti
on
us
in
g
ma
c
hi
ne
le
a
r
ni
n
g,”
A
dv
anc
e
s
in
M
ul
ti
m
e
di
a
,
vol
. 2022, 2022, doi:
10.1155/2022/
3351256.
[
14]
Z
.
L
i
e
t
al
.
,
“
A
hi
gh
-
pr
e
c
is
io
n
d
e
te
c
ti
on
me
th
od
of
hydr
oponic
le
tt
uc
e
s
e
e
dl
in
gs
s
ta
tu
s
b
a
s
e
d
on
im
pr
ove
d
F
a
s
t
e
r
R
C
N
N
,”
C
o
m
pu
te
r
s
and E
le
c
tr
oni
c
s
i
n A
gr
ic
ul
tu
r
e
, vol
. 182, 2021, d
oi
:
10.1016/j
.c
ompa
g.2021.106054.
[
15]
M
. M
. T
a
ye
, “
T
he
or
e
ti
c
a
l
unde
r
s
ta
ndi
ng of
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k : c
onc
e
pt
s
, a
r
c
hi
te
c
tu
r
e
s
, a
ppl
ic
a
ti
ons
, f
ut
ur
e
di
r
e
c
ti
on
s
,”
C
om
put
at
io
n
, vol
. 11, no. 52, 2023.
[
16]
X
.
W
a
ng,
Y
.
Z
ha
o,
a
nd
F
.
P
our
pa
na
h,
“
R
e
c
e
nt
a
dv
a
nc
e
s
in
de
e
p
le
a
r
ni
ng,”
I
nt
e
r
nat
io
nal
J
our
nal
of
M
ac
hi
ne
L
e
ar
ni
ng
and
C
y
be
r
ne
ti
c
s
, vol
. 11, no. 4, pp. 747
–
750, 2020, doi:
10.1007/s
1
3042
-
020
-
01096
-
5.
[
17]
S
.
L
iu
,
H
.
B
a
n,
Y
.
S
ong,
M
.
Z
ha
ng,
a
nd
F
.
Y
a
ng,
“
M
e
th
od
f
or
de
te
c
ti
ng
C
hi
ne
s
e
te
xt
s
in
na
tu
r
a
l
s
c
e
ne
s
ba
s
e
d
on
im
pr
ove
d
f
a
s
te
r
R
-
C
N
N
,”
I
nt
e
r
nat
io
nal
J
our
nal
of
P
at
te
r
n
R
e
c
ogni
ti
on
and
A
r
ti
fi
c
ia
l
I
nt
e
ll
ig
e
nc
e
,
vol
.
34,
no.
2,
2020,
doi
:
10.1142/S
021800142053002X
.
[
18]
L
.
J
ia
ng,
J
.
C
he
n,
H
.
T
odo,
Z
.
T
a
ng,
S
.
L
iu
,
a
nd
Y
.
L
i,
“
A
pp
li
c
a
ti
on
of
a
F
a
s
t
R
C
N
N
ba
s
e
d
on
upp
e
r
a
nd
lo
w
e
r
la
y
e
r
s
in
f
a
c
e
r
e
c
ogni
t
io
n,”
C
om
put
at
io
nal
I
nt
e
ll
ig
e
nc
e
and N
e
ur
o
s
c
ie
nc
e
, vo
l.
2021, 2021, doi:
10.1155/2021/
9945934.
[
19]
P
.
N
.
M
.
M
a
’
r
if
a
h,
M
.
S
a
r
os
a
,
a
nd
E
.
R
oha
di
,
“
C
ompa
r
is
on
of
F
a
s
te
r
R
-
C
N
N
R
e
s
N
e
t
-
50
a
nd
R
e
s
N
e
t
-
101
me
th
ods
f
o
r
r
e
c
yc
l
in
g
w
a
s
te
de
te
c
t
io
n,”
I
nt
e
r
nat
io
nal
J
ou
r
nal
of
C
om
put
e
r
A
ppl
ic
at
io
ns
T
e
c
hnol
ogy
and R
e
s
e
ar
c
h
,
vol
. 12, no.
12
, pp. 26
–
32, 2023,
doi
:
10.7753/i
jc
a
tr
1212.1006.
[
20]
W
.
L
i
u,
S
.
L
i
a
o
,
W
.
H
u
,
X
.
L
i
a
n
g,
a
n
d
X
.
C
h
e
n,
“
L
e
a
r
ni
ng
e
f
f
i
c
i
e
n
t
s
in
gl
e
-
s
ta
g
e
pe
d
e
s
tr
ia
n
d
e
t
e
c
t
or
s
by
a
s
y
mpt
ot
ic
lo
c
a
li
z
a
t
io
n
f
it
ti
n
g
,”
P
r
oc
e
e
d
in
g
s
of
th
e
E
u
r
op
e
a
n
C
on
fe
r
e
n
c
e
on
C
o
m
p
ut
e
r
V
i
s
i
on
(
E
C
C
V
)
,
p
p.
6
43
–
6
59
,
20
18
,
do
i:
1
0.
10
07
/9
78
-
3
-
0
30
-
0
12
64
-
9_
38
.
[
21]
W
.
Z
ou,
Z
.
Z
ha
ng,
Y
.
P
e
ng,
C
.
X
ia
ng,
S
.
T
ia
n,
a
nd
L
.
Z
ha
ng,
“
S
C
-
R
P
N
:
A
s
tr
ong
c
or
r
e
la
ti
on
le
a
r
ni
ng
f
r
a
me
w
or
k
f
or
r
e
gi
on
pr
opos
a
l
,”
I
E
E
E
T
r
ans
. I
m
age
P
r
oc
e
s
s
.
, vol
. 30, pp. 4084
–
4098
, 2021, doi:
10.1109/T
I
P
.2021.3069547.
[
22]
L
.
F
a
n,
T
.
Z
ha
ng,
a
nd
W
.
D
u,
“
O
pt
ic
a
l
-
f
lo
w
-
ba
s
e
d
f
r
a
me
w
or
k
to
boos
t
vi
de
o
obj
e
c
t
de
te
c
ti
on
pe
r
f
or
ma
nc
e
w
it
h
ob
je
c
t
e
nha
nc
e
me
nt
,”
E
x
pe
r
t
Sy
s
te
m
s
w
it
h A
ppl
ic
at
io
ns
, vol
. 170, 202
0, 2021, doi:
10.1016/j
.e
s
w
a
.2020.114544.
[
23]
A
.
A
ji
t,
K
.
A
c
ha
r
ya
,
a
nd
A
.
S
a
ma
nt
a
,
“
A
r
e
vi
e
w
o
f
c
onvolu
ti
o
na
l
ne
ur
a
l
ne
two
r
ks
,”
2020
I
nt
e
r
nat
io
nal
C
onf
e
r
e
nc
e
on
E
m
e
r
gi
ng
T
r
e
nds
in
I
nf
or
m
at
io
n
T
e
c
hnol
ogy
and
E
ngi
ne
e
r
in
g
(
ic
-
E
T
I
T
E
)
,
V
e
ll
or
e
,
I
ndi
a
,
pp.
1
–
5,
2020,
doi
:
10.110
9/
ic
-
E
T
I
T
E
47903.2020.049.
[
24]
J
.
L
i
e
t
al
.
,
“
A
ut
oma
ti
c
de
te
c
ti
on
a
nd
c
la
s
s
if
ic
a
ti
on
s
y
s
te
m
of
dome
s
ti
c
w
a
s
te
vi
a
mul
ti
mode
l
c
a
s
c
a
de
d
c
onvolut
io
na
l
ne
ur
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.
1
,
F
e
br
ua
r
y
2025
:
22
2
-
230
230
ne
tw
or
k,”
I
E
E
E
T
r
ans
ac
ti
ons
on I
ndus
tr
ia
l
I
nf
or
m
at
ic
s
, vol
. 18
, no. 1, pp. 163
–
173, 2022, doi:
10.1109/T
I
I
.2021.3085669.
[
25]
L
.
I
c
hi
m
a
nd
D
.
P
ope
s
c
u,
“
M
e
la
noma
de
te
c
ti
on
u
s
in
g
a
n
obj
e
c
ti
ve
s
ys
te
m
ba
s
e
d
on
mul
ti
pl
e
c
onne
c
t
e
d
ne
ur
a
l
ne
twor
ks
,
”
I
E
E
E
A
c
c
e
s
s
, vol
. 8, pp. 179189
–
179202, 2020, doi:
10.1109/AC
C
E
S
S
.2020.3028248.
[
26]
J
.
W
a
ng,
X
.
H
e
,
S
.
F
a
mi
ng,
G
.
L
u,
H
.
C
ong,
a
nd
Q
.
J
ia
n
g,
“
A
r
e
a
l
-
ti
me
br
id
ge
c
r
a
c
k
de
te
c
ti
on
me
th
od
ba
s
e
d
on
a
n
im
pr
ove
d
I
nc
e
pt
io
n R
e
s
N
e
t
V
2
s
tr
uc
tu
r
e
,”
I
E
E
E
A
c
c
e
s
s
, vol
. 9, pp. 93209
–
93223, 2021, doi:
10.1109/AC
C
E
S
S
.2021.3093210.
[
27]
Y
.
P
.
C
he
n,
Y
.
L
i,
a
nd
G
.
W
a
ng,
“
A
n
e
nh
a
nc
e
d
r
e
gi
on
pr
opos
a
l
ne
twor
k
f
or
o
bj
e
c
t
d
e
te
c
ti
on
u
s
in
g
de
e
p
le
a
r
ni
ng
m
e
th
od,”
P
L
oS
O
ne
, vol
. 13, no. 9, pp. 1
–
26, 2018, doi:
10.1371/j
our
na
l.
pone
.0203897.
[
28]
W
.
G
u
e
t
al
.
,
“
H
ig
h
a
c
c
ur
a
c
y
th
yr
oi
d
tu
mor
im
a
ge
r
e
c
ogni
ti
on
ba
s
e
d
on
hybr
id
mul
ti
pl
e
mode
ls
opt
im
iz
a
ti
on,”
I
E
E
E
A
c
c
e
s
s
,
vol
.
8,
pp. 128426
–
128439, 2020, doi:
10.1109/AC
C
E
S
S
.2020.300
8401.
[
29]
N
.
A
.
P
r
a
s
e
ty
o,
P
r
a
now
o,
a
nd
A
.
J
.
S
a
nt
o
s
o,
“
A
ut
oma
ti
c
d
e
te
c
ti
on
a
nd
c
a
lc
ul
a
ti
on
of
pa
lm
oi
l
f
r
e
s
h
f
r
ui
t
bunc
he
s
us
in
g
f
a
s
te
r
R
-
C
N
N
,”
I
nt
e
r
nat
io
nal
J
ou
r
nal
of
A
ppl
ie
d
Sc
ie
nc
e
and
E
ngi
ne
e
r
in
g
,
vol
.
17,
no.
2,
pp.
121
–
134,
2020,
doi
:
10.6703/I
J
A
S
E
.202005_17(
2)
.121.
[
30]
M
.
S
a
r
os
a
,
N
.
M
una
,
a
nd
E
.
R
oha
di
,
“
P
e
r
f
or
ma
nc
e
of
f
a
s
te
r
R
-
C
N
N
to
de
te
c
t
pl
a
s
ti
c
w
a
s
te
,”
I
nt
e
r
nat
io
nal
J
ou
r
nal
of
A
dv
anc
e
d
T
r
e
nds
i
n C
om
put
e
r
Sc
ie
nc
e
and E
ngi
ne
e
r
in
g
, vol
. 9, no.
5, pp.
7756
–
7762, 2020, doi:
10.30534/i
ja
tc
s
e
/2
020/
120952020.
B
I
OG
RA
P
HI
E
S
OF
AU
T
HO
RS
M
o
echa
m
m
a
d
Sa
r
o
s
a
recei
v
ed
t
h
e
d
i
p
l
o
ma
o
f
en
g
i
n
eer
i
n
g
t
ech
n
o
l
o
g
y
fro
m
U
n
i
v
er
s
i
t
e
d
e
N
an
c
y
I,
Fran
ce
i
n
1
9
8
9
.
H
e
o
b
t
ai
n
ed
h
i
s
mas
t
er’
s
an
d
d
o
ct
o
ral
d
e
g
rees
fro
m
Ban
d
u
n
g
In
s
t
i
t
u
t
e
o
f
T
ec
h
n
o
l
o
g
y
,
In
d
o
n
es
i
a
i
n
2
0
0
2
an
d
2
0
0
7
res
p
ect
i
v
e
l
y
.
H
e
h
as
b
ee
n
t
h
e
reci
p
i
en
t
o
f
s
e
v
eral
re
s
earch
g
ran
t
s
f
u
n
d
ed
b
y
t
h
e
Mi
n
i
s
t
r
y
o
f
Res
earc
h
,
T
ech
n
o
l
o
g
y
a
n
d
H
i
g
h
er
E
d
u
cat
i
o
n
o
f
t
h
e
Rep
u
b
l
i
c
o
f
In
d
o
n
es
i
a.
H
i
s
cu
rren
t
res
ear
ch
i
n
t
eres
t
s
l
i
e
i
n
i
n
f
o
rmat
i
o
n
an
d
co
mmu
n
i
ca
t
i
o
n
t
ech
n
o
l
o
g
y
,
art
i
fi
c
i
al
i
n
t
el
l
i
g
en
ce,
m
o
b
i
l
e
co
mp
u
t
i
n
g
,
an
d
Io
T
.
H
e
can
b
e
co
n
t
act
e
d
at
emai
l
:
ms
ar
o
s
a@
p
o
l
i
n
ema.
ac.
i
d
.
P
uteri
N
urul
M
a
’ri
fa
h
co
mp
l
et
e
d
D
i
p
l
o
ma
s
t
u
d
i
es
i
n
T
el
eco
mm
u
n
i
ca
t
i
o
n
s
an
d
Mas
t
er
o
f
E
l
ect
r
i
cal
E
n
g
i
n
eeri
n
g
co
n
cen
t
rat
i
o
n
i
n
T
e
l
eco
mmu
n
i
cat
i
o
n
Sci
e
n
ce
an
d
In
f
o
rmat
i
o
n
T
ech
n
o
l
o
g
y
at
t
h
e
St
a
t
e
Po
l
y
t
ec
h
n
i
c
o
f
Mal
an
g
.
In
2
0
2
0
i
t
p
as
s
ed
t
h
e
3
3
r
d
PIMN
A
S
i
n
t
h
e
fi
el
d
o
f
Co
mmu
n
i
t
y
Serv
i
ce
Res
earc
h
,
an
d
i
n
2
0
2
3
recei
v
ed
a
Mas
t
er's
T
h
es
i
s
Res
earc
h
g
ra
n
t
fro
m
t
h
e
M
i
n
i
s
t
ry
o
f
E
d
u
cat
i
o
n
,
Cu
l
t
u
re,
Res
earc
h
,
an
d
T
ec
h
n
o
l
o
g
y
.
W
h
i
l
e
s
t
u
d
y
i
n
g
ma
s
t
er
s
,
h
e
h
as
b
eco
me
a
t
eac
h
i
n
g
a
s
s
i
s
t
an
t
h
el
p
i
n
g
t
each
d
i
g
i
t
a
l
i
m
ag
e
p
ro
ce
s
s
i
n
g
a
n
d
ar
t
i
f
i
ci
a
l
i
n
t
el
l
i
g
e
n
ce
w
o
r
k
s
h
o
p
co
u
rs
e
s
,
aft
er
g
ra
d
u
a
t
i
n
g
fr
o
m
t
h
e
mas
t
er'
s
d
eg
r
ee
i
n
v
o
l
v
ed
i
n
res
earch
a
t
t
h
e
In
t
e
g
rat
e
d
A
p
p
l
i
e
d
T
ec
h
n
o
l
o
g
y
Res
earc
h
Cen
t
er
(PRI
T
T
I)
St
a
t
e
Po
l
y
t
ech
n
i
c
o
f
Mal
a
n
g
.
S
h
e
can
b
e
co
n
t
ac
t
ed
at
ema
i
l
:
p
n
u
r
u
l
mar
i
fah
@
g
ma
i
l
.
co
m.
M
i
l
a
Kus
um
a
w
a
rda
n
i
recei
v
ed
t
h
e
M.
T
.
d
eg
ree
i
n
E
l
ect
r
o
E
n
g
i
n
eeri
n
g
fro
m
Braw
i
j
ay
a
U
n
i
v
ers
i
t
y
o
f
Mal
an
g
,
In
d
o
n
es
i
a,
i
n
2
0
1
0
.
Sh
e
i
s
a
l
ect
u
rer
i
n
D
i
g
i
t
al
T
el
ec
o
mmu
n
i
ca
t
i
o
n
s
N
et
w
o
r
k
s
St
u
d
y
Pro
g
ram
o
f
E
l
ect
ri
cal
E
n
g
i
n
eer
i
n
g
D
ep
ar
t
men
t
o
f
St
at
e
Po
l
y
t
ec
h
n
i
cs
o
f
Ma
l
an
g
.
Sh
e
can
b
e
co
n
t
ac
t
e
d
at
emai
l
:
mi
l
a.
k
u
s
u
maw
ar
d
an
i
@
p
o
l
i
n
ema.
ac.
i
d
.
D
i
m
a
s
F
i
rm
a
nda
A
l
R
i
z
a
g
o
t
a
D
o
ct
o
ral
d
eg
r
ee
fro
m
K
y
o
t
o
U
n
i
v
ers
i
t
y
,
J
a
p
an
i
n
t
h
e
fi
el
d
o
f
Bi
o
-
s
en
s
i
n
g
E
n
g
i
n
eeri
n
g
i
n
2
0
1
9
.
T
h
e
d
o
ct
o
r
al
s
t
u
d
y
w
a
s
co
mp
l
et
e
d
w
i
t
h
t
h
e
L
PD
P
p
res
i
d
e
n
t
i
al
s
c
h
o
l
ars
h
i
p
.
He
h
a
s
p
u
b
l
i
s
h
ed
m
o
re
t
h
a
n
7
0
s
ci
en
t
i
f
i
c
p
a
p
ers
u
n
t
i
l
2
0
2
2
i
n
cl
u
d
i
n
g
d
o
ze
n
s
o
f
t
h
em
i
n
Q
1
rep
u
t
a
b
l
e
i
n
t
er
n
at
i
o
n
al
j
o
u
rn
a
l
s
.
Re
cen
t
l
y
,
h
e
recei
v
ed
Y
o
u
n
g
Re
s
earch
e
r's
A
cad
em
i
c
E
n
co
u
rag
eme
n
t
A
w
ar
d
2
0
2
1
fro
m
T
h
e
J
a
p
an
es
e
So
ci
e
t
y
o
f
A
g
ri
c
u
l
t
u
ra
l
Mach
i
n
er
y
an
d
Fo
o
d
E
n
g
i
n
eer
s
(J
S
A
M).
Cu
rren
t
l
y
,
h
e
i
s
t
h
e
H
ead
o
f
t
h
e
Mech
at
r
o
n
i
cs
L
ab
o
rat
o
ry
o
f
A
g
r
o
-
i
n
d
u
s
t
ry
T
o
o
l
s
a
n
d
Mach
i
n
er
y
,
D
e
p
art
me
n
t
o
f
A
g
r
i
cu
l
t
u
r
al
E
n
g
i
n
eer
i
n
g
,
Facu
l
t
y
o
f
A
g
ri
c
u
l
t
u
ra
l
T
ech
n
o
l
o
g
y
,
U
n
i
v
ers
i
t
a
s
Braw
i
j
ay
a.
H
e
ca
n
b
e
co
n
t
ac
t
e
d
at
emai
l
:
d
i
mas
f
i
rman
d
a@
u
b
.
ac.
i
d
.
Evaluation Warning : The document was created with Spire.PDF for Python.