I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
14
, N
o.
3
,
J
une
2025
, pp.
2246
~
2257
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
14
.i
3
.pp
2246
-
2257
2246
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
R
ob
u
st
t
w
o
-
st
age
ob
j
e
c
t
d
e
t
e
c
t
i
on
u
si
n
g Y
O
L
O
v5
f
or
e
n
h
an
c
i
n
g
t
om
at
o l
e
af
d
i
se
ase
d
e
t
e
c
t
i
on
E
n
d
an
g S
u
r
yaw
at
i
1
, S
yi
f
a A
u
li
yah
H
as
an
ah
2
, R
a
d
e
n
S
an
d
r
a Y
u
w
an
a
1
, Ji
m
m
y A
b
d
e
l
K
ad
ar
1
,
H
il
m
an
F
e
r
d
in
an
d
u
s
P
ar
d
e
d
e
1
1
R
e
s
e
a
r
c
h
C
e
nt
e
r
f
or
A
r
t
i
f
i
c
i
a
l
I
nt
e
l
l
i
ge
nc
e
a
nd C
ybe
r
S
e
c
ur
i
t
y, N
a
t
i
ona
l
R
e
s
e
a
r
c
h a
nd I
nnova
t
i
on A
ge
nc
y, B
a
ndung, I
ndone
s
i
a
2
D
e
pa
r
t
m
e
nt
of
S
t
a
t
i
s
t
i
c
s
,
F
a
c
ul
t
y of
M
a
t
he
m
a
t
i
c
s
a
nd
N
a
t
ur
a
l
S
c
i
e
nc
e
s
, P
a
dj
a
d
j
a
r
a
n U
ni
ve
r
s
i
t
y, S
um
e
da
ng, I
ndone
s
i
a
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
A
ug
13
,
2024
R
e
vi
s
e
d
F
e
b
10
,
2025
A
c
c
e
pt
e
d
M
a
r
15
,
2025
Deep
learning
facilitates
human
activities
across
various
sectors,
inc
luding
agricult
ure.
Early
disease
detection
in
plants,
such
as
tomato
plant
t
hat
are
susceptible
to
diseases,
is
critical
because
it
h
elps
farmers
reduce
loss
es
and
control
the
disease
spread
more
effectively
.
However,
the
abilit
y
of
m
achine
to
recognize
diseased
leaf
objects
is
also
influenced
by
the
quality
o
f
data.
Data
collected
directly
from
the
field
typically
yields
lower
accuracy
due
to
challenges
faced
in
ma
chine
interpret
ation.
To
address
this
challen
ge
,
we
propose
a
two
-
stage
detection
architecture
for
identifying
infected
t
omato
plant
classes,
leveraging
YOLOv5
to
detect
objects
within
the
images
obtained
from
the
field.
We
use
Inception
-
V3
for
classifying
objec
ts
into
known
classes.
Additionally,
w
e
employ
a
combination
of
two
d
ataset:
PlantDoc
s
which
repre
sent
field
data,
and
PlantVill
age
datase
t
which
serve
s
as
a
cleaner
dataset.
Our
experimental
results
indicate
that
the
use
of
YOLOv5
in
handling
data
under
actual
field
conditions
can
enhance
model
performance, a
lthough the accuracy value
is moderate (62.50 %).
K
e
y
w
o
r
d
s
:
C
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
D
e
e
p l
e
a
r
ni
ng
P
la
nt
di
s
e
a
s
e
d
e
te
c
ti
on
T
w
o
-
s
ta
ge
obj
e
c
t
de
te
c
ti
on
Y
O
L
O
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
E
nda
ng S
ur
ya
w
a
ti
R
e
s
e
a
r
c
h
C
e
nt
e
r
f
or
A
r
ti
f
ic
ia
l
I
nt
e
ll
ig
e
nc
e
a
nd
C
ybe
r
S
e
c
ur
it
y, N
a
ti
ona
l
R
e
s
e
a
r
c
h a
nd I
nnova
ti
on A
ge
nc
y
S
a
ngkur
ia
ng
S
t.
, K
S
T
S
a
m
a
un S
a
m
a
di
kun, B
a
ndung, I
ndone
s
ia
E
m
a
il
:
e
nda
029@
br
in
.go.i
d
1.
I
N
T
R
O
D
U
C
T
I
O
N
E
a
r
ly
di
s
e
a
s
e
de
te
c
ti
on i
n pl
a
nt
s
e
na
bl
e
s
f
a
r
m
e
r
s
t
o m
in
im
iz
e
l
os
s
e
s
a
nd mor
e
e
f
f
e
c
ti
ve
ly
c
ont
r
ol
t
he
s
pr
e
a
d
of
di
s
e
a
s
e
[
1]
.
C
e
r
ta
in
pl
a
nt
s
,
pa
r
ti
c
ul
a
r
ly
to
m
a
to
e
s
,
a
r
e
vul
ne
r
a
bl
e
to
a
va
r
ie
ty
of
d
is
e
a
s
e
s
th
a
t
c
a
n
r
e
duc
e
c
r
op
pr
oduc
ti
vi
ty
a
nd
f
r
u
it
qua
li
ty
.
B
a
c
te
r
ia
l
s
pot
,
la
te
b
li
ght
,
le
a
f
m
ol
d,
s
e
pt
or
ia
le
a
f
s
pot
,
a
nd
s
pi
de
r
m
it
e
s
a
r
e
a
m
ong
th
e
di
s
e
a
s
e
s
th
a
t
a
f
f
e
c
t
to
m
a
to
pl
a
nt
s
.
C
ons
e
que
nt
ly
,
e
a
r
ly
de
te
c
ti
on
of
di
s
e
a
s
e
s
in
to
m
a
to
pl
a
nt
s
i
s
c
r
uc
ia
l
to
m
in
im
iz
in
g l
os
s
e
s
[
2]
.
E
xi
s
ti
ng
r
e
s
e
a
r
c
h
in
di
c
a
te
s
s
ig
ni
f
ic
a
nt
a
dva
nc
e
m
e
nt
s
in
de
v
e
lo
pi
ng
s
ys
te
m
s
f
or
id
e
nt
if
yi
ng
a
n
d
c
la
s
s
if
yi
ng
pl
a
nt
di
s
e
a
s
e
s
us
in
g
m
a
c
hi
ne
le
a
r
ni
ng
m
e
th
ods
th
a
t
ut
il
iz
e
im
a
ge
s
of
in
f
e
c
te
d
le
a
ve
s
.
I
ni
ti
a
ll
y,
th
e
s
e
m
e
th
ods
r
e
li
e
d
on
m
a
nua
l
f
e
a
tu
r
e
e
xt
r
a
c
ti
on,
de
m
a
ndi
ng
e
xpe
r
t
knowle
dge
a
nd
li
m
it
in
g
th
e
qua
li
ty
a
nd
r
e
le
va
nc
e
of
f
e
a
tu
r
e
s
.
A
lg
or
it
hm
s
s
uc
h
a
s
s
uppor
t
ve
c
to
r
m
a
c
hi
ne
s
,
de
c
is
io
n
tr
e
e
s
,
k
-
ne
a
r
e
s
t
ne
ig
hbor
s
,
na
ïv
e
B
a
ye
s
,
a
nd
r
a
ndom
f
or
e
s
t
s
ha
ve
de
m
ons
tr
a
te
d
th
e
pot
e
nt
ia
l
o
f
tr
a
di
ti
ona
l
m
a
c
hi
ne
le
a
r
ni
ng
in
a
gr
ic
ul
tu
r
a
l
a
ppl
ic
a
ti
ons
[
3]
,
[
4]
.
T
he
a
dve
nt
of
de
e
p
le
a
r
ni
ng
ha
s
r
e
vol
ut
io
ni
z
e
d
tr
a
di
ti
ona
l
m
a
c
hi
ne
le
a
r
ni
ng
th
r
ough
a
ut
om
a
ti
c
f
e
a
tu
r
e
e
xt
r
a
c
ti
on
[
5]
,
gr
e
a
tl
y
im
pr
ovi
ng
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y.
D
e
e
p
le
a
r
ni
ng,
f
ir
s
t
in
tr
oduc
e
d
in
1943
[
6]
,
c
ont
in
ue
s
to
e
vol
ve
a
nd
i
s
w
id
e
ly
a
ppl
ie
d
a
c
r
os
s
va
r
io
us
dom
a
in
s
,
in
c
lu
di
ng
te
xt
r
e
c
ogni
ti
on
[
7]
,
[
8]
,
s
pe
e
c
h
r
e
c
ogni
ti
on
[
9]
,
[
10]
,
a
nd
im
a
ge
r
e
c
ogni
ti
on
[
11
]
,
[
12]
.
O
ne
of
th
e
de
e
p
le
a
r
ni
ng
a
r
c
hi
te
c
tu
r
e
s
c
om
m
onl
y
us
e
d
f
or
im
a
ge
c
la
s
s
if
ic
a
ti
on
is
th
e
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
(
C
N
N
)
.
C
N
N
le
ve
r
a
ge
s
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
R
obus
t
tw
o
-
s
ta
ge
obj
e
c
t
de
te
c
ti
on us
in
g Y
O
L
O
v
5 f
o
r
e
nhanc
in
g
t
om
at
o l
e
af
…
(
E
ndang Sur
y
aw
at
i
)
2247
m
ove
m
e
nt
of
c
onvolut
io
n
ke
r
ne
ls
to
c
la
s
s
if
y
obj
e
c
ts
ba
s
e
d
on
vi
s
ua
l
f
e
a
tu
r
e
s
s
u
c
h
a
s
c
ol
or
,
te
xt
ur
e
,
a
nd
le
a
f
e
dge
s
.
T
hi
s
a
ppr
oa
c
h
de
li
ve
r
e
d
s
upe
r
io
r
pe
r
f
or
m
a
nc
e
f
or
v
a
r
io
us
im
a
ge
da
ta
ta
s
ks
w
hi
le
pr
ogr
e
s
s
iv
e
ly
s
upe
r
s
e
di
ng
tr
a
di
ti
ona
l
m
a
c
hi
ne
le
a
r
ni
ng
m
e
th
ods
[
11]
.
H
ow
e
ve
r
,
f
or
s
m
a
ll
da
ta
s
e
t
us
e
c
a
s
e
s
,
th
e
tr
a
di
ti
ona
l
m
a
c
hi
ne
l
e
a
r
ni
ng s
ti
ll
out
pe
r
f
or
m
s
[
13]
.
I
n
a
gr
ic
ul
tu
r
e
a
nd
pl
a
nt
a
ti
on
a
ppl
ic
a
ti
ons
,
e
xi
s
ti
ng
r
e
s
e
a
r
c
h
in
di
c
a
te
s
th
a
t
C
N
N
s
a
r
e
c
a
p
a
bl
e
of
im
pr
ovi
ng
c
la
s
s
if
ic
a
ti
on
a
c
c
ur
a
c
y
th
r
ough
va
r
io
us
popula
r
a
r
c
hi
te
c
tu
r
e
s
.
D
if
f
e
r
e
nt
C
N
N
a
r
c
hi
te
c
tu
r
e
s
ha
ve
be
e
n w
id
e
ly
us
e
d t
o i
de
nt
if
y pl
a
nt
s
or
c
la
s
s
if
y pl
a
nt
di
s
e
a
s
e
s
. A
le
xN
e
t,
G
oogl
e
N
e
t,
a
nd V
G
G
-
16 e
a
c
h pos
s
e
s
s
di
s
ti
nc
t
c
ha
r
a
c
te
r
is
ti
c
s
f
or
id
e
nt
if
yi
ng
a
nd
c
la
s
s
if
yi
ng
di
s
e
a
s
e
d
le
a
ve
s
[
14]
.
T
h
e
I
nc
e
pt
io
n
a
r
c
hi
te
c
tu
r
e
ha
s
be
e
n
a
ppl
ie
d
f
or
c
la
s
s
if
yi
ng
f
r
ui
t
pl
a
nt
s
[
15]
,
lu
ng
c
a
nc
e
r
[
16]
,
a
nd
pl
a
nt
di
s
e
a
s
e
s
[
17]
–
[
20]
.
C
N
N
a
r
c
hi
te
c
tu
r
e
s
c
a
te
gor
iz
e
d
a
s
s
ki
p
c
onne
c
ti
on
a
r
c
hi
te
c
tu
r
e
s
,
ha
v
e
a
ls
o
be
e
n
us
e
d
f
or
c
la
s
s
if
yi
ng
pl
a
nt
di
s
e
a
s
e
s
s
uc
h
a
s
R
e
s
N
e
t
[
19]
,
[
21]
–
[
23]
,
a
nd
D
e
ns
e
N
e
t
[
19]
,
[
21]
,
[
24]
,
a
nd
a
ls
o
D
e
ns
e
N
e
t
f
or
de
te
c
ti
ng
pl
a
nt
nut
r
ie
nt
de
f
ic
ie
nc
ie
s
[
25]
.
O
th
e
r
C
N
N
a
r
c
hi
te
c
tu
r
e
s
,
de
ve
lo
p
e
d
f
or
im
pr
ove
d
pe
r
f
or
m
a
nc
e
a
nd
e
f
f
ic
ie
nc
y,
in
c
lu
de
C
om
N
e
t
[
26]
,
E
f
f
ic
ie
nt
N
e
t
[
19]
,
[
21]
,
[
24
]
,
M
obi
le
N
e
t
[
20]
–
[
2
2]
,
[
27]
,
a
nd
I
nc
e
pt
io
nR
e
s
N
e
t
[
21]
,
[
22]
.
I
n
it
s
de
ve
lo
pm
e
nt
,
s
om
e
r
e
s
e
a
r
c
h
e
r
s
ha
ve
pr
opos
e
d
m
ode
ls
c
a
te
go
r
iz
e
d
unde
r
th
e
de
te
c
to
r
f
a
m
il
y,
na
m
e
ly
one
-
s
ta
ge
a
nd
two
-
s
ta
ge
obj
e
c
t
de
te
c
ti
on.
Y
O
L
O
is
r
e
c
ogni
z
e
d
a
s
a
popula
r
one
-
s
ta
ge
obj
e
c
t
de
te
c
ti
on
m
ode
l,
w
hi
le
th
e
r
e
gi
on
-
ba
s
e
d
C
N
N
f
a
m
il
y
f
a
ll
s
in
to
two
-
s
ta
ge
obj
e
c
t
de
te
c
ti
on.
W
u
e
t
al
.
[
28]
a
ppl
ie
s
two
le
a
r
ni
ng
m
ode
ls
, Y
O
L
O
v5 a
nd E
f
f
ic
ie
nt
N
e
tV2, to c
la
s
s
if
y t
om
a
to
l
e
a
f
d
is
e
a
s
e
s
.
N
e
ve
r
th
e
le
s
s
,
m
a
ny
s
tu
di
e
s
on
pl
a
nt
di
s
e
a
s
e
c
la
s
s
if
ic
a
ti
on
r
e
ly
he
a
vi
ly
on
c
l
e
a
n
da
t
a
s
e
t
s
,
w
hi
c
h
e
na
bl
e
m
ode
ls
to
a
c
hi
e
ve
hi
gh
a
c
c
ur
a
c
y.
H
ow
e
ve
r
,
m
o
s
t
da
ta
s
e
ts
f
ound
in
r
e
a
l
-
w
or
ld
e
nvi
r
onm
e
nt
s
a
r
e
c
a
pt
ur
e
d
unde
r
unc
ont
r
ol
le
d,
r
e
a
l
-
w
or
ld
c
ondi
ti
ons
,
unl
ik
e
la
b
or
a
to
r
y
da
ta
s
e
ts
.
W
e
r
e
f
e
r
to
s
uc
h
da
ta
s
e
ts
a
s
"
di
r
ty
da
ta
s
e
ts
,"
r
e
pr
e
s
e
nt
in
g
r
e
a
l
-
w
or
ld
c
ondi
ti
ons
.
O
f
te
n,
m
o
de
ls
s
tr
uggl
e
to
pe
r
f
or
m
w
e
ll
w
he
n
te
s
te
d
on
di
r
ty
da
ta
s
e
ts
.
T
hi
s
s
it
ua
ti
on
pr
e
s
e
nt
s
a
c
ha
ll
e
ngi
ng
t
a
s
k
f
or
m
a
c
hi
ne
s
,
w
hi
c
h
m
us
t
r
e
c
ogni
z
e
a
nd
c
la
s
s
if
y
obj
e
c
ts
f
r
om
r
e
a
l
-
c
ondi
ti
on da
ta
i
nt
o pr
e
de
f
in
e
d c
a
te
gor
ie
s
.
B
a
s
e
d on thi
s
ba
c
kgr
ound, our
r
e
s
e
a
r
c
h f
oc
us
e
s
on i
m
pr
ovi
ng
m
ode
l
pe
r
f
or
m
a
nc
e
, pa
r
ti
c
ul
a
r
ly
w
he
n
te
s
te
d
w
it
h
r
e
a
l
-
w
or
ld
(
di
r
ty
)
da
ta
s
e
ts
,
to
de
ve
lo
p
a
r
obus
t
m
o
de
l
f
or
c
la
s
s
if
yi
ng
to
m
a
to
pl
a
nt
le
a
f
di
s
e
a
s
e
s
.
T
he
r
e
a
r
e
s
e
ve
r
a
l
c
r
uc
ia
l
f
a
c
to
r
s
to
c
ons
id
e
r
to
a
ddr
e
s
s
th
is
r
e
s
e
a
r
c
h
que
s
ti
on.
F
ir
s
t,
w
e
e
m
pl
oy
a
n
obj
e
c
t
de
te
c
to
r
a
nd
a
c
la
s
s
if
ie
r
to
pr
opos
e
a
two
-
s
ta
ge
obj
e
c
t
de
te
c
ti
o
n
a
r
c
hi
te
c
tu
r
e
.
S
e
c
ond,
w
e
le
ve
r
a
ge
Y
O
L
O
v5
to
be
in
te
gr
a
te
d
in
to
th
e
a
r
c
hi
te
c
tu
r
e
a
s
a
n
obj
e
c
t
de
te
c
to
r
,
pe
r
f
or
m
in
g
pr
e
-
p
r
oc
e
s
s
in
g
ta
s
k
s
be
f
or
e
th
e
da
ta
e
nt
e
r
s
th
e
c
la
s
s
if
ie
r
.
F
or
th
e
pr
e
li
m
in
a
r
y
r
e
s
e
a
r
c
h
of
our
pr
opos
e
d
a
r
c
hi
te
c
tu
r
e
,
w
e
c
ons
id
e
r
ut
il
iz
in
g
Y
O
L
O
v5,
w
hi
c
h
of
f
e
r
s
ba
la
nc
e
d
pe
r
f
or
m
a
nc
e
,
s
p
e
e
d,
a
li
g
ht
w
e
ig
ht
m
ode
l,
a
nd
a
da
pt
a
bi
li
ty
f
or
f
ut
ur
e
r
e
qui
r
e
m
e
nt
s
w
hi
le
a
ls
o
a
c
c
ount
in
g
f
or
th
e
c
ons
tr
a
in
ts
of
our
c
ur
r
e
nt
ha
r
dw
a
r
e
[
27]
,
[
29
]
.
W
e
ut
il
iz
e
I
nc
e
pt
io
n
-
V
3
to
c
la
s
s
if
y
de
te
c
te
d
obj
e
c
t
s
f
r
om
Y
O
L
O
v5
in
to
k
now
n
to
m
a
to
di
s
e
a
s
e
c
la
s
s
e
s
. T
he
ju
s
ti
f
ic
a
ti
on
f
or
c
hoos
in
g
I
nc
e
pt
io
n
-
V
3
a
s
our
ba
s
e
li
ne
c
la
s
s
if
ie
r
is
th
a
t
it
is
qui
te
e
f
f
ic
ie
nt
in
te
r
m
s
of
c
om
put
a
ti
ona
l
c
os
t,
ha
s
a
s
im
pl
e
de
s
ig
n
m
ode
l,
a
nd
is
s
tr
a
ig
ht
f
or
w
a
r
d
to
s
tu
dy
[
3
0]
.
M
a
ny
s
tu
di
e
s
us
e
th
is
m
ode
l
a
s
a
ba
s
e
li
ne
a
nd
a
c
hi
e
v
e
good
pe
r
f
or
m
a
nc
e
.
T
hi
r
d,
w
e
ut
il
iz
e
th
e
P
la
nt
D
oc
s
d
a
ta
s
e
t
to
r
e
pr
e
s
e
nt
th
e
c
ha
ll
e
nge
s
of
r
e
a
l
-
w
or
ld
c
ondi
ti
ons
(
di
r
ty
da
ta
s
e
ts
)
.
M
e
a
nw
hi
le
,
th
e
P
la
nt
V
il
la
ge
da
ta
s
e
t
is
u
s
e
d
to
v
a
li
da
te
th
e
f
in
di
ngs
of
m
a
ny
s
tu
di
e
s
th
a
t
r
e
ly
on
c
le
a
n
da
ta
s
e
ts
.
P
la
nt
D
oc
s
a
nd
P
la
nt
V
il
la
ge
w
il
l
be
a
lt
e
r
na
te
ly
us
e
d
a
s
tr
a
in
in
g
a
nd
te
s
ti
ng
da
ta
.
H
ow
e
ve
r
,
w
e
a
s
s
um
e
th
a
t
th
e
r
ol
e
of
Y
O
L
O
v5
i
n
pr
e
-
pr
oc
e
s
s
in
g
ta
s
ks
w
il
l
be
m
or
e
e
f
f
e
c
ti
ve
w
he
n
th
e
m
ode
l
is
tr
a
in
e
d
a
nd
t
e
s
te
d
us
in
g
th
e
P
la
nt
D
oc
s
da
ta
s
e
t.
F
our
th
,
w
e
a
im
to
a
s
s
e
s
s
w
h
e
th
e
r
Y
O
L
O
v5
a
s
a
pr
e
-
pr
oc
e
s
s
or
c
a
n
im
pr
ove
c
la
s
s
if
ie
r
pe
r
f
or
m
a
nc
e
.
T
he
c
la
s
s
if
ie
r
w
il
l
be
e
va
lu
a
t
e
d
w
it
h
a
nd
w
it
hout
Y
O
L
O
v5 pr
e
-
pr
oc
e
s
s
in
g.
2.
M
E
T
H
O
D
2.1. I
n
c
e
p
t
io
n
-
V3
I
nc
e
pt
io
n
-
V
3
is
a
de
e
p
le
a
r
ni
ng
a
r
c
hi
te
c
tu
r
e
th
a
t
ha
s
a
c
hi
e
ve
d
a
n
a
c
c
ur
a
c
y
of
m
or
e
th
a
n
78.1%
on
c
la
s
s
if
ic
a
ti
on
ta
s
ks
in
vol
vi
ng
1000
c
la
s
s
e
s
on
th
e
I
m
a
ge
N
e
t
d
a
ta
s
e
t
[
31]
.
T
hi
s
l
e
ve
l
of
a
c
c
ur
a
c
y
r
e
nde
r
s
it
s
ui
ta
bl
e
f
or
va
r
io
us
im
a
ge
r
e
c
ogni
ti
on
ta
s
ks
.
S
e
ve
r
a
l
s
tu
di
e
s
ha
ve
be
e
n
c
onduc
te
d
to
c
la
s
s
if
y
28
f
lo
w
e
r
s
pe
c
ie
s
u
s
in
g t
he
I
nc
e
pt
io
n
-
V
3 a
r
c
hi
te
c
tu
r
e
a
nd t
r
a
ns
f
e
r
l
e
a
r
ni
n
g t
o e
nha
nc
e
a
c
c
ur
a
c
y by r
e
tr
a
in
in
g t
he
f
lo
w
e
r
c
a
te
gor
y
c
ol
le
c
ti
on.
B
a
s
e
d
on
th
e
e
xpe
r
im
e
nt
r
e
s
ul
ts
,
f
r
om
th
e
two
da
ta
s
e
ts
us
e
d,
th
e
O
xf
or
d
-
17
a
nd
O
xf
or
d
-
102
f
lo
w
e
r
da
ta
s
e
ts
,
th
e
r
e
s
ul
ti
ng
a
c
c
ur
a
c
y
is
95%
.
T
hi
s
in
di
c
a
te
s
th
a
t
I
nc
e
pt
io
n
-
V
3
pe
r
f
or
m
s
w
e
ll
in
im
a
ge
c
la
s
s
if
ic
a
ti
on t
a
s
ks
, e
v
e
n w
it
h da
ta
s
e
ts
c
ont
a
in
in
g nume
r
ous
c
la
s
s
c
a
te
gor
ie
s
[
32]
.
A
ddi
ti
ona
ll
y,
I
nc
e
pt
io
n
is
de
s
ig
ne
d
to
de
li
ve
r
hi
gh
pe
r
f
or
m
a
nc
e
r
e
s
ul
ts
w
it
h
a
lo
w
e
r
c
om
put
a
ti
ona
l
lo
a
d
c
om
pa
r
e
d
to
ot
he
r
a
r
c
hi
te
c
tu
r
e
s
.
T
hi
s
i
s
a
c
hi
e
va
bl
e
du
e
t
o
th
e
f
e
w
e
r
pa
r
a
m
e
te
r
s
in
I
nc
e
pt
io
n
c
om
pa
r
e
d
to
ot
he
r
a
r
c
hi
te
c
tu
r
e
s
. I
nc
e
pt
io
n
-
V
3 i
s
a
n a
dva
nc
e
m
e
nt
of
t
he
e
a
r
li
e
r
a
r
c
hi
te
c
tu
r
e
, I
nc
e
pt
io
n
-
V
1, i
nt
r
oduc
e
d i
n
2014
a
s
G
oog
L
e
N
e
t
[
33]
.
S
e
ve
r
a
l
m
odi
f
ic
a
ti
ons
ha
ve
be
e
n
im
pl
e
m
e
nt
e
d
in
th
is
a
r
c
hi
te
c
tu
r
e
c
om
pa
r
e
d
to
it
s
pr
e
de
c
e
s
s
or
,
in
c
lu
di
ng
f
a
c
to
r
iz
a
ti
on
in
to
s
m
a
ll
e
r
c
onvolut
io
ns
,
s
pa
ti
a
l
f
a
c
to
r
iz
a
ti
on
in
to
a
s
ym
m
e
tr
ic
c
onvolut
io
ns
,
ut
il
iz
a
ti
on
of
a
uxi
li
a
r
y
c
la
s
s
if
ie
r
s
,
a
nd
e
f
f
ic
ie
nt
g
r
id
r
e
duc
ti
on.
F
ig
u
r
e
1
s
how
th
e
I
nc
e
pt
io
n
-
V
3
a
r
c
hi
te
c
tu
r
e
ge
ne
r
a
ll
y.
O
ve
r
a
ll
,
th
e
I
nc
e
pt
io
n
-
V
3
a
r
c
hi
te
c
tu
r
e
c
om
pr
is
e
s
th
ir
te
e
n
m
odul
e
s
:
one
s
te
m
m
odul
e
,
te
n
in
c
e
pt
io
n
m
odul
e
s
,
two
r
e
duc
ti
on
m
odul
e
s
,
a
nd
one
a
uxi
li
a
r
y
c
la
s
s
if
ie
r
m
odul
e
.
T
hi
s
c
om
bi
na
ti
on
of
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
3
,
J
une
20
25
:
2246
-
2257
2248
m
odul
e
s
a
ll
ow
s
I
nc
e
pt
io
n
-
V
3
to
pr
oc
e
s
s
im
a
g
e
s
e
f
f
ic
ie
nt
ly
,
c
a
pt
ur
in
g
a
w
id
e
r
a
nge
of
f
e
a
tu
r
e
s
a
t
m
ul
ti
pl
e
s
c
a
le
s
w
hi
le
m
a
in
ta
in
in
g a
ba
la
nc
e
be
twe
e
n
c
om
put
a
ti
ona
l
c
os
t
a
nd pe
r
f
or
m
a
nc
e
.
F
ig
ur
e
1. T
he
i
nc
e
pt
io
n
-
V
3 a
r
c
hi
te
c
tu
r
e
2.2. YOL
O
v5
T
he
a
r
c
hi
te
c
tu
r
e
of
C
N
N
i
s
e
xc
e
ll
e
nt
f
or
c
la
s
s
if
ic
a
ti
on.
N
e
ve
r
t
he
le
s
s
,
obj
e
c
t
de
te
c
ti
on
c
a
n
b
e
a
good
s
ol
ut
io
n
in
c
e
r
ta
in
c
a
s
e
s
to
e
ns
ur
e
th
a
t
th
e
c
la
s
s
if
ie
d
im
a
ge
s
do
not
c
ont
a
in
noi
s
e
or
ot
he
r
im
a
ge
s
out
s
id
e
th
e
in
te
nde
d obje
c
t.
Y
O
L
O
v5 i
s
a
m
e
th
od s
ui
ta
bl
e
f
or
obj
e
c
t
de
te
c
ti
on. YO
L
O
v5 i
s
t
he
e
vol
ut
io
n of
th
e
f
a
m
il
y o
f
Y
O
L
O
.
W
id
e
ly
us
e
d,
th
is
obj
e
c
t
de
te
c
ti
on
m
e
th
od
b
a
la
nc
e
s
s
pe
e
d
a
nd
de
te
c
ti
on
pe
r
f
or
m
a
nc
e
,
a
nd
a
ls
o
of
f
e
r
s
a
s
m
a
ll
e
r
m
ode
l
w
e
ig
ht
[
27]
,
e
na
bl
in
g
e
f
f
e
c
ti
ve
m
ul
t
i
-
s
c
a
le
obj
e
c
t
de
te
c
ti
on
[
29]
.
Y
O
L
O
v5
a
ls
o
be
c
om
e
s
a
s
ui
ta
bl
e
a
nd e
a
s
y m
e
th
od t
o be
m
odi
f
ie
d
f
or
e
nha
nc
e
m
e
nt
s
i
n f
ur
th
e
r
de
ve
lo
pm
e
nt
ne
e
ds
[
27]
,
[
29]
. Y
O
L
O
v5
c
a
n
de
te
c
t
a
nd
c
la
s
s
if
y
m
ul
ti
pl
e
obj
e
c
ts
in
c
lu
di
ng
hum
a
ns
,
a
ni
m
a
ls
,
a
nd
ve
hi
c
le
s
,
in
a
n
im
a
ge
or
vi
de
o.
F
ig
ur
e
2 i
s
a
de
pi
c
ti
on of
Y
O
L
O
v5 a
r
c
hi
te
c
tu
r
e
.
F
ig
ur
e
2. T
he
Y
O
L
O
v5 ne
twor
k a
r
c
hi
te
c
tu
r
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
R
obus
t
tw
o
-
s
ta
ge
obj
e
c
t
de
te
c
ti
on us
in
g Y
O
L
O
v
5 f
o
r
e
nhanc
in
g
t
om
at
o l
e
af
…
(
E
ndang Sur
y
aw
at
i
)
2249
Y
O
L
O
v5
is
a
va
il
a
bl
e
in
f
iv
e
di
f
f
e
r
e
nt
s
iz
e
s
,
ba
s
e
d
on
th
e
num
be
r
of
la
ye
r
s
a
nd
pa
r
a
m
e
te
r
s
it
pos
s
e
s
s
e
s
.
T
hi
s
a
r
c
hi
te
c
tu
r
e
c
om
pr
is
e
s
th
r
e
e
m
a
in
pa
r
ts
:
th
e
ba
c
kbone
,
ne
c
k,
a
nd
he
a
d.
T
he
ba
c
kbon
e
is
r
e
s
pons
ib
le
f
or
f
or
m
in
g
f
e
a
tu
r
e
s
i
n t
he
i
m
a
ge
, l
e
ve
r
a
gi
ng t
he
C
S
P
D
a
r
kne
t5
3 a
r
c
hi
te
c
tu
r
e
, w
hi
c
h i
s
a
m
odi
f
ie
d
ve
r
s
io
n of
D
a
r
kne
t.
T
he
c
r
os
s
-
s
ta
ge
pa
r
ti
a
l
(
C
S
P
)
s
tr
uc
tu
r
e
he
lp
s
ove
r
c
om
e
gr
a
di
e
nt
pr
obl
e
m
s
by s
pl
it
ti
ng
t
he
f
lo
w
of
gr
a
di
e
nt
s
[
34]
,
r
e
duc
in
g
th
e
num
be
r
of
pa
r
a
m
e
te
r
s
,
a
nd
c
om
put
in
g
th
e
lo
a
d.
I
n
ot
he
r
w
or
ds
,
th
e
B
ot
tl
e
ne
c
kC
S
P
c
a
n
h
a
ndl
e
th
e
f
e
a
tu
r
e
m
a
p
e
xt
r
a
c
ti
on
a
nd
r
e
duc
e
gr
a
di
e
nt
in
f
or
m
a
ti
on
dupl
ic
a
ti
on
in
th
e
C
N
N
opt
im
iz
a
ti
on
pr
oc
e
s
s
.
M
e
a
nw
hi
le
,
th
e
s
pa
ti
a
l
pyr
a
m
id
po
ol
in
g
(
S
P
P
)
m
odul
e
e
nha
nc
e
s
th
e
de
te
c
ti
on
of
ta
r
ge
ts
a
t
di
f
f
e
r
e
nt
s
c
a
le
s
by
a
ggr
e
ga
ti
ng
f
e
a
tu
r
e
s
f
r
om
m
ul
ti
pl
e
la
ye
r
s
.
T
he
ne
c
k
is
th
e
pa
r
t
th
a
t
c
onne
c
ts
th
e
ba
c
kbone
w
it
h
th
e
he
a
d,
r
e
s
pon
s
ib
le
f
or
m
e
r
gi
ng
f
e
a
tu
r
e
s
f
r
o
m
di
f
f
e
r
e
nt
s
c
a
le
s
.
T
he
he
a
d
is
r
e
s
pon
s
ib
le
f
or
de
te
c
ti
ng
obj
e
c
ts
.
S
im
il
a
r
to
ot
he
r
Y
O
L
O
a
r
c
hi
te
c
tu
r
e
s
,
it
us
e
s
Y
O
L
O
la
ye
r
s
to
bui
ld
th
is
pa
r
t.
T
he
out
put
o
f
th
is
pa
r
t
in
c
lu
de
s
bounding boxes
a
nd c
la
s
s
pr
oba
bi
li
ti
e
s
.
2.3. YOL
O
v5 as
p
r
e
-
p
r
oc
e
s
s
in
g m
e
t
h
od
f
o
r
t
w
o
-
s
t
age
s
ob
j
e
c
t
d
e
t
e
c
t
io
n
ar
c
h
it
e
c
t
u
r
e
A
s
w
e
ha
ve
e
xpl
a
in
e
d
in
th
e
in
tr
oduc
ti
on
s
e
c
ti
on,
w
e
ut
il
iz
e
Y
O
L
O
v5
to
s
uppor
t
th
e
pr
e
-
pr
oc
e
s
s
in
g
s
ta
ge
,
in
c
lu
di
ng
lo
c
a
li
z
in
g
a
nd
de
te
c
ti
ng
obj
e
c
ts
w
it
hi
n
a
n
im
a
ge
.
T
hi
s
s
ta
g
e
is
th
e
f
ir
s
t
s
te
p
in
th
e
two
-
s
ta
ge
obj
e
c
t
de
te
c
ti
on
pr
oc
e
s
s
.
T
h
e
r
e
s
e
a
r
c
h
c
om
m
e
nc
e
d
w
it
h
th
e
pr
e
pr
oc
e
s
s
in
g
s
ta
g
e
,
w
he
r
e
w
e
a
ppl
ie
d
Y
O
L
O
v5
to
t
w
o da
ta
s
e
ts
f
or
obj
e
c
t
de
te
c
ti
on. W
e
a
ppl
y t
hi
s
pr
oc
e
s
s
t
o t
he
s
e
le
c
te
d da
t
a
s
e
t
s
, f
oc
us
in
g t
he
i
m
a
ge
s
on t
he
im
por
ta
nt
a
r
e
a
s
f
or
e
a
s
ie
r
a
nd mor
e
a
c
c
ur
a
te
c
la
s
s
if
ic
a
ti
on.
T
he
f
ir
s
t
s
te
p
in
th
e
obj
e
c
t
de
te
c
ti
on
pr
oc
e
s
s
in
Y
O
L
O
v5
in
vol
ve
s
e
xt
r
a
c
ti
ng
f
e
a
tu
r
e
s
f
r
om
e
a
c
h
da
ta
s
e
t,
us
in
g
a
r
e
s
ol
ut
io
n
of
768
×
768
×
3
f
r
om
th
e
or
ig
in
a
l
ba
c
kbone
.
T
hi
s
pa
r
t
s
pl
it
s
e
a
c
h
im
a
ge
in
to
f
e
a
tu
r
e
m
a
ps
,
e
a
c
h
r
e
pr
e
s
e
nt
in
g
th
e
im
a
ge
a
t
di
f
f
e
r
e
nt
le
ve
ls
of
a
bs
tr
a
c
ti
on.
T
he
ne
c
k
pa
r
t
th
e
n
c
onc
a
te
na
te
s
th
e
s
e
f
e
a
tu
r
e
m
a
ps
to
a
ggr
e
ga
te
in
f
or
m
a
ti
on
f
r
om
va
r
io
us
s
c
a
le
s
.
A
f
te
r
c
onc
a
te
na
ti
on,
a
c
onvolut
io
n
pr
oc
e
s
s
w
it
h
32
ke
r
ne
ls
tr
a
n
s
f
or
m
s
th
e
c
on
c
a
te
na
te
d
f
e
a
tu
r
e
m
a
p
s
in
to
a
3
20
×
320
×
32
f
e
a
tu
r
e
m
a
p.
T
he
h
e
a
d
pa
r
t
th
e
n
lo
c
a
li
z
e
s
a
nd
de
te
c
ts
obj
e
c
t
s
f
r
om
th
e
s
e
va
r
io
us
s
c
a
le
s
of
f
e
a
t
ur
e
m
a
ps
,
ul
ti
m
a
te
ly
pr
oduc
in
g
c
la
s
s
if
ic
a
ti
on
r
e
s
ul
ts
a
nd
obj
e
c
t
c
oor
di
na
te
s
.
F
ig
ur
e
3
il
lu
s
tr
a
te
s
th
e
a
r
c
hi
te
c
tu
r
e
of
a
two
-
s
ta
ge
obj
e
c
t
de
te
c
ti
on
s
ys
te
m
,
w
he
r
e
t
he
Y
O
L
O
v5 obje
c
t
de
te
c
ti
on pr
oc
e
s
s
f
or
m
s
a
n i
nt
e
gr
a
l
pa
r
t
of
t
he
e
nt
ir
e
s
ys
te
m
.
F
ig
ur
e
3. T
he
pr
opos
e
d a
r
c
hi
te
c
tu
r
e
of
t
w
o
-
s
ta
ge
s
obj
e
c
t
de
te
c
ti
on
D
ur
in
g
obj
e
c
t
de
te
c
ti
on,
w
e
tr
a
in
th
e
Y
O
L
O
v5
m
ode
l
u
s
in
g
t
he
or
ig
in
a
l
da
ta
s
e
t
to
ge
ne
r
a
te
a
ne
w
da
ta
s
e
t.
T
hi
s
ne
w
d
a
ta
s
e
t
c
on
s
is
ts
of
c
la
s
s
if
ie
d
le
a
f
obj
e
c
ts
r
e
s
ul
ti
ng
f
r
om
th
e
de
te
c
ti
on
pr
oc
e
s
s
,
s
pe
c
if
ic
a
ll
y
to
m
a
to
le
a
f
im
a
ge
s
c
la
s
s
if
ie
d
a
s
e
it
he
r
di
s
e
a
s
e
d
or
he
a
lt
hy.
T
he
or
ig
in
a
l
da
ta
s
e
t
c
om
pr
is
e
s
to
m
a
to
le
a
f
im
a
ge
s
,
bot
h
di
s
e
a
s
e
d
a
nd
he
a
lt
hy,
obt
a
in
e
d
f
r
om
th
e
P
la
nt
D
oc
s
a
nd
P
la
nt
V
il
la
ge
da
ta
s
e
t
s
.
A
f
te
r
obj
e
c
t
de
te
c
ti
on,
w
e
pr
oc
e
e
d
to
th
e
c
la
s
s
if
ic
a
ti
on
s
ta
ge
.
I
n
th
is
s
ta
g
e
,
w
e
tr
a
in
th
e
I
nc
e
pt
io
n
-
V
3
m
ode
l
to
c
la
s
s
if
y
to
m
a
to
l
e
a
f
di
s
e
a
s
e
s
us
in
g t
he
n
e
w
da
ta
s
e
t.
I
nt
e
r
e
s
ti
ngl
y,
th
is
s
tu
dy
s
how
s
ho
w
Y
O
L
O
v5,
w
hi
c
h
is
a
n
obj
e
c
t
de
te
c
to
r
,
c
ont
r
ib
ut
e
s
to
th
e
pr
e
-
pr
oc
e
s
s
in
g
s
ta
g
e
to
s
uppor
t
th
e
I
nc
e
pt
io
n
-
V
3
c
la
s
s
if
ie
r
in
id
e
nt
if
yi
ng
di
s
e
a
s
e
d
to
m
a
to
le
a
ve
s
.
U
s
in
g
a
da
ta
s
e
t
th
a
t
a
c
c
ur
a
te
ly
r
e
f
le
c
t
s
r
e
a
l
-
w
or
ld
c
ondi
ti
ons
,
s
uc
h
a
s
t
he
P
la
nt
D
oc
s
d
a
ta
s
e
t
in
F
ig
ur
e
4,
s
ig
ni
f
ic
a
nt
ly
e
nha
nc
e
s
it
s
e
f
f
e
c
ti
ve
ne
s
s
.
T
hi
s
f
ig
ur
e
il
lu
s
tr
a
te
s
a
s
a
m
pl
e
im
a
ge
f
r
om
P
la
nt
D
oc
s
ta
ke
n
in
f
ie
ld
c
ondi
ti
ons
.
T
hr
ough
th
e
obj
e
c
t
de
te
c
ti
on
pr
oc
e
s
s
,
Y
O
L
O
v5
lo
c
a
li
z
e
s
a
nd
de
te
c
ts
th
r
e
e
di
s
e
a
s
e
d
le
a
f
obj
e
c
ts
,
th
e
n
c
r
ops
th
e
s
e
le
a
ve
s
f
r
om
th
e
or
ig
in
a
l
im
a
ge
,
r
e
s
ul
ti
ng
in
th
r
e
e
s
e
pa
r
a
t
e
di
s
e
a
s
e
d
le
a
f
im
a
ge
s
,
a
s
s
how
n
on
th
e
r
ig
ht
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
3
,
J
une
20
25
:
2246
-
2257
2250
s
id
e
of
th
e
a
r
r
ow
in
F
ig
ur
e
4.
T
he
I
nc
e
pt
io
n
-
V
3
c
la
s
s
if
ie
r
f
in
d
s
it
e
a
s
ie
r
to
r
e
c
ogni
z
e
a
nd
c
la
s
s
if
y
th
e
le
a
ve
s
us
in
g t
he
s
e
t
hr
e
e
i
ndi
vi
dua
l
le
a
f
i
m
a
ge
s
, a
s
oppo
s
e
d t
o us
in
g t
h
e
or
ig
in
a
l
im
a
ge
on t
he
l
e
f
t
s
id
e
.
F
ig
ur
e
4. O
bj
e
c
t
de
te
c
ti
on by YO
L
O
v5 on P
la
nt
D
oc
s
s
a
m
pl
e
i
m
a
ge
W
e
e
xc
lu
s
iv
e
ly
pr
e
s
e
nt
t
he
P
la
nt
D
oc
s
s
a
m
pl
e
i
m
a
ge
f
or
t
he
obj
e
c
t
de
te
c
ti
on pr
oc
e
s
s
, a
s
i
t
s
how
c
a
s
e
s
th
e
a
bi
li
ty
to
de
te
c
t
a
nd
c
r
op
m
ul
ti
pl
e
le
a
f
obj
e
c
ts
w
it
hi
n
a
n
im
a
ge
in
to
s
e
pa
r
a
te
obj
e
c
t
im
a
ge
s
.
H
ow
e
ve
r
,
w
e
a
ppl
ie
d
th
e
s
a
m
e
pr
e
-
pr
oc
e
s
s
in
g
us
in
g
Y
O
L
O
v5
to
th
e
P
l
a
nt
V
il
la
ge
da
ta
s
e
t
in
our
s
tu
dy,
de
s
pi
te
it
s
c
la
s
s
if
ic
a
ti
on
a
s
a
c
l
e
a
n
da
ta
s
e
t.
P
la
nt
V
il
la
ge
im
a
ge
s
a
r
e
w
e
ll
-
or
ga
ni
z
e
d
to
m
a
to
le
a
f
im
a
ge
s
a
r
r
a
nge
d
in
la
bor
a
to
r
y s
e
tt
in
gs
w
it
h unif
or
m
c
ol
or
ba
c
kgr
ounds
.
2.4. Dat
as
e
t
p
r
e
p
ar
at
io
n
P
r
e
pa
r
in
g
th
e
da
ta
be
f
o
r
e
us
in
g
th
e
da
ta
s
e
ts
to
tr
a
in
th
e
m
ode
l
is
a
not
he
r
s
te
p
in
th
e
p
r
e
-
pr
oc
e
s
s
in
g
s
ta
ge
.
D
a
ta
pr
e
pa
r
a
ti
on
is
e
s
s
e
nt
ia
l
f
or
a
c
hi
e
vi
ng
e
x
c
e
ll
e
nt
m
ode
l
pe
r
f
or
m
a
nc
e
.
I
n
our
s
tu
dy,
Y
O
L
O
v5'
s
obj
e
c
t
de
te
c
ti
on
pr
oc
e
s
s
pr
oduc
e
s
th
e
pr
e
p
a
r
e
d
da
ta
,
w
hi
c
h
w
e
r
e
f
e
r
to
a
s
th
e
ne
w
da
ta
s
e
t,
a
s
il
lu
s
tr
a
te
d
in
F
ig
ur
e
4.
A
s
pr
e
vi
ous
ly
e
xpl
a
in
e
d,
w
e
us
e
two
di
f
f
e
r
e
nt
da
ta
s
e
ts
:
th
e
P
la
nt
V
il
la
ge
da
ta
s
e
t
a
nd
th
e
P
la
nt
D
o
c
s
da
ta
s
e
t.
T
he
P
la
nt
V
il
la
ge
da
t
a
s
e
t
c
ons
i
s
ts
of
38
c
la
s
s
c
a
te
gor
ie
s
ba
s
e
d
on
di
s
e
a
s
e
ty
pe
s
f
or
va
r
io
us
pl
a
nt
s
pe
c
ie
s
,
to
ta
li
ng
54,303
im
a
g
e
s
a
c
r
os
s
a
ll
c
l
a
s
s
e
s
.
T
he
P
la
nt
D
oc
s
da
t
a
s
e
t
c
ons
is
t
s
of
2,598
im
a
ge
s
f
r
om
13
pl
a
nt
s
pe
c
ie
s
,
w
it
h
a
to
ta
l
of
17
c
la
s
s
c
a
te
gor
ie
s
ba
s
e
d
on
di
s
e
a
s
e
ty
pe
s
.
P
la
nt
D
oc
s
a
nd
P
la
nt
V
il
la
ge
a
r
e
publ
ic
da
ta
s
e
ts
th
a
t
a
r
e
w
id
e
ly
us
e
d
f
or
de
ve
lo
pi
ng
a
nd
te
s
ti
ng
pl
a
nt
di
s
e
a
s
e
de
te
c
ti
on
m
ode
ls
a
nd
c
a
n
be
a
c
c
e
s
s
e
d
f
r
e
e
ly
.
E
a
c
h
of
th
e
m
ha
s
th
e
ir
ow
n
uni
que
c
h
a
r
a
c
t
e
r
is
ti
c
s
.
I
m
a
ge
s
s
a
m
pl
e
s
f
or
th
e
P
la
nt
V
il
la
ge
da
ta
s
e
t
a
r
e
s
how
n i
n F
ig
ur
e
5, a
nd F
ig
ur
e
6 i
ll
us
tr
a
te
s
a
m
pl
e
of
t
he
i
m
a
ge
s
f
or
t
he
P
la
nt
D
oc
s
da
ta
s
e
t.
F
ig
ur
e
5. P
la
nt
V
il
la
ge
s
a
m
pl
e
i
m
a
ge
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
R
obus
t
tw
o
-
s
ta
ge
obj
e
c
t
de
te
c
ti
on us
in
g Y
O
L
O
v
5 f
o
r
e
nhanc
in
g
t
om
at
o l
e
af
…
(
E
ndang Sur
y
aw
at
i
)
2251
F
ig
ur
e
6. P
la
nt
D
oc
s
s
a
m
pl
e
i
m
a
ge
B
ot
h
f
ig
ur
e
s
hi
ghl
ig
ht
th
e
di
f
f
e
r
e
nc
e
s
in
im
a
ge
c
ondi
ti
o
ns
be
twe
e
n
th
e
two
da
ta
s
e
ts
.
T
he
P
la
nt
V
il
la
ge
da
ta
s
e
t
e
xhi
bi
ts
a
r
e
la
ti
ve
ly
c
le
a
n c
ondi
ti
on due
t
o
th
e
pr
oc
e
s
s
of
c
a
pt
ur
in
g i
m
a
ge
s
t
a
ke
s
pl
a
c
e
i
n
a
c
ont
r
ol
le
d
e
nvi
r
onm
e
nt
s
e
tt
in
g.
M
e
a
nw
hi
l
e
,
th
e
P
la
nt
D
oc
s
da
ta
s
e
t
c
ont
a
in
s
im
a
g
e
s
th
a
t
m
a
y
c
ont
a
in
m
ul
ti
pl
e
le
a
ve
s
,
e
a
c
h
w
it
h
va
r
yi
ng
ba
c
kgr
ounds
a
nd
li
ght
in
g
c
ondi
ti
ons
.
F
or
our
s
tu
dy,
w
e
us
e
onl
y
to
m
a
to
pl
a
nt
s
f
r
om
e
a
c
h
da
ta
s
e
t,
f
oc
u
s
in
g
on
a
s
ubs
e
t
of
6
di
s
e
a
s
e
c
la
s
s
e
s
,
na
m
e
ly
b
a
c
te
r
ia
l
s
pot
,
l
a
te
bl
ig
ht
,
le
a
f
m
ol
d, s
e
pt
or
ia
l
e
a
f
s
pot
, m
os
a
ic
vi
r
us
, ye
ll
ow
l
e
a
f
c
ur
l
vi
r
us
, a
nd 1 he
a
lt
hy c
la
s
s
.
T
a
bl
e
1
s
how
s
th
e
da
ta
di
s
tr
ib
ut
io
n
of
th
e
or
ig
in
a
l
da
ta
s
e
ts
us
e
d
a
s
in
put
f
or
th
e
obj
e
c
t
de
te
c
ti
on
pr
oc
e
s
s
.
T
he
r
e
a
r
e
12,357
im
a
ge
s
f
r
om
th
e
P
la
nt
V
il
la
ge
da
ta
s
e
t
a
nd
648
im
a
ge
s
f
r
om
th
e
P
la
nt
D
oc
s
da
ta
s
e
t.
S
in
c
e
Y
O
L
O
v5
c
a
n
de
te
c
t
m
ul
ti
pl
e
c
la
s
s
e
s
in
a
s
in
gl
e
im
a
ge
,
th
e
num
be
r
of
in
s
ta
nc
e
s
in
e
a
c
h
c
la
s
s
in
th
e
P
la
nt
D
oc
s
da
ta
s
e
t
h
a
s
c
ha
nge
d,
a
s
s
ho
w
n
in
T
a
bl
e
2.
T
hi
s
c
ha
nge
onl
y
oc
c
ur
s
in
th
e
P
la
nt
D
oc
s
da
ta
s
e
t
be
c
a
us
e
t
h
e
P
la
nt
V
il
la
ge
da
ta
s
e
t
c
ons
i
s
ts
of
s
in
gl
e
-
le
a
f
i
m
a
ge
s
.
T
a
bl
e
1.
T
he
da
ta
di
s
tr
ib
ut
io
n f
or
t
w
o or
ig
in
a
l
da
ta
s
e
ts
D
i
s
e
a
s
e
c
l
a
s
s
P
l
a
nt
V
i
l
l
a
ge
P
l
a
nt
D
oc
s
B
a
c
t
e
r
i
a
l
s
pot
1914
107
L
a
t
e
bl
i
ght
1689
111
L
e
a
f
m
ol
d
857
91
S
e
pt
or
i
a
l
e
a
f
s
pot
1582
148
M
os
a
i
c
vi
r
us
307
54
Y
e
l
l
ow
l
e
a
f
c
ur
l
vi
r
us
4671
75
H
e
a
l
t
hy
1337
62
T
a
bl
e
2.
T
he
da
ta
di
s
tr
ib
ut
io
n f
or
t
he
P
la
nt
D
oc
s
da
ta
s
e
t
D
i
s
e
a
s
e
c
l
a
s
s
O
r
i
gi
na
l
da
t
a
s
e
t
N
e
w
da
t
a
s
e
t
B
a
c
t
e
r
i
a
l
s
pot
107
265
L
a
t
e
bl
i
ght
111
141
L
e
a
f
m
ol
d
91
368
S
e
pt
or
i
a
l
e
a
f
s
pot
148
195
M
os
a
i
c
vi
r
us
54
482
Y
e
l
l
ow
l
e
a
f
c
ur
l
vi
r
us
75
1095
H
e
a
l
t
hy
62
582
2.5. E
xp
e
r
im
e
n
t
al
s
e
t
u
p
W
e
di
vi
de
e
a
c
h
da
ta
s
e
t
in
to
80%
tr
a
in
in
g
da
ta
a
nd
20%
te
s
ti
ng
da
ta
,
r
e
s
pe
c
ti
ve
ly
.
T
h
e
da
ta
s
e
ts
w
e
us
e
d
c
ont
a
in
im
a
ge
s
w
it
h
va
r
io
us
pi
xe
l
s
iz
e
s
.
T
he
r
e
f
or
e
,
s
om
e
pi
xe
l
tr
a
ns
f
or
m
a
ti
ons
o
r
a
dj
us
tm
e
nt
s
a
r
e
r
e
qui
r
e
d
to
a
da
pt
to
th
e
m
ode
l
a
r
c
hi
te
c
tu
r
e
.
W
e
s
ta
nda
r
di
z
e
d
a
ll
da
ta
s
iz
e
s
to
128
×
128
to
e
ns
ur
e
uni
f
or
m
it
y.
W
e
s
c
a
le
th
e
pi
xe
l
va
lu
e
s
f
r
om
0
-
255
to
0
-
1,
s
pe
e
di
ng
up
th
e
m
ode
l'
s
tr
a
in
in
g
a
nd
hom
oge
ni
z
in
g
th
e
va
lu
e
s
in
th
e
da
ta
.
W
e
a
ls
o
us
e
da
ta
a
ugm
e
nt
a
ti
on
to
di
ve
r
s
if
y
th
e
a
v
a
il
a
bl
e
da
ta
,
a
ll
ow
in
g
th
e
m
ode
l
to
le
a
r
n
f
r
o
m
a
ddi
ti
ona
l
da
ta
dur
in
g
th
e
tr
a
in
in
g
p
r
oc
e
s
s
.
T
hi
s
c
a
n
le
a
d
to
be
tt
e
r
r
e
s
ul
ts
by
c
a
pt
ur
in
g
ta
r
ge
te
d
c
ha
r
a
c
te
r
is
ti
c
s
. A
ugm
e
nt
a
ti
on t
e
c
hni
que
s
u
s
e
d i
nc
lu
de
s
hi
f
t,
r
ot
a
ti
on, s
he
a
r
, z
oom
, a
nd f
li
p.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
3
,
J
une
20
25
:
2246
-
2257
2252
T
he
im
a
ge
s
pr
e
pa
r
e
d
in
th
e
pr
e
-
pr
oc
e
s
s
in
g
s
ta
g
e
a
r
e
th
e
n
in
put
in
to
th
e
I
nc
e
pt
io
n
-
V
3
c
la
s
s
if
ie
r
.
W
e
us
e
th
e
s
e
im
a
ge
s
to
tr
a
in
th
e
m
ode
l
f
or
opt
im
a
l
pe
r
f
o
r
m
a
nc
e
in
c
la
s
s
if
yi
ng
di
s
e
a
s
e
s
in
to
m
a
to
pl
a
nt
s
.
A
s
out
li
ne
d
in
our
pr
opos
a
l,
th
is
s
tu
dy
e
m
pha
s
iz
e
s
le
ve
r
a
gi
n
g
Y
O
L
O
v5
to
de
te
c
t
obj
e
c
t
s
w
it
hi
n
im
a
ge
s
,
e
nha
nc
in
g
th
e
pr
e
-
pr
oc
e
s
s
in
g
s
ta
ge
in
a
two
-
s
ta
ge
obj
e
c
t
de
te
c
ti
on
a
r
c
hi
te
c
tu
r
e
.
O
ur
hope
is
to
im
pr
ove
I
nc
e
pt
io
n
-
V
3 m
ode
l
pe
r
f
or
m
a
nc
e
i
n de
te
c
ti
ng t
om
a
to
l
e
a
f
di
s
e
a
s
e
s
by i
nc
or
por
a
ti
ng Y
O
L
O
v5.
T
o
te
s
t
our
pr
opos
a
l,
w
e
tr
a
in
a
nd
te
s
t
th
e
I
nc
e
pt
io
n
-
V
3
m
ode
l
us
in
g
a
c
om
bi
na
ti
on
of
two
da
ta
s
e
ts
,
a
ll
ow
in
g
th
e
m
ode
l
to
le
a
r
n
f
r
om
di
ve
r
s
e
da
ta
ty
pe
s
.
W
e
a
lt
e
r
na
te
ly
us
e
th
e
s
e
two
da
ta
s
e
ts
a
s
tr
a
in
in
g
a
nd
te
s
ti
ng
da
ta
.
A
ddi
ti
ona
ll
y,
w
e
tr
a
in
th
e
I
nc
e
pt
io
n
-
V
3
m
ode
l
w
it
hout
us
in
g
Y
O
L
O
v5
in
th
e
pr
e
-
pr
oc
e
s
s
in
g
s
ta
ge
,
w
hi
c
h
w
e
de
f
in
e
d
a
s
our
b
a
s
e
li
ne
a
r
c
hi
te
c
tu
r
e
.
I
n
th
e
b
a
s
e
li
ne
a
r
c
hi
te
c
tu
r
e
,
w
e
di
r
e
c
tl
y
tr
a
in
a
nd
te
s
t
th
e
I
nc
e
pt
io
n
-
V
3
m
ode
l
us
in
g
th
e
two
or
ig
in
a
l
da
ta
s
e
ts
,
by
pa
s
s
in
g
th
e
Y
O
L
O
v5
pr
e
-
pr
oc
e
s
s
in
g
s
ta
ge
.
T
o
s
uppor
t
th
e
tr
a
in
in
g
a
nd
te
s
ti
ng
of
th
e
m
ode
l,
w
e
us
e
th
e
f
ol
lo
w
in
g
hype
r
pa
r
a
m
e
te
r
s
e
tt
in
gs
:
A
da
m
opt
im
iz
e
r
w
it
h a
l
e
a
r
ni
ng r
a
te
of
1×
10⁻
⁴, a
ba
tc
h
s
iz
e
of
32, a
nd
30 e
poc
hs
pe
r
e
xpe
r
im
e
nt
.
3.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
W
e
di
vi
de
d
th
e
m
ode
l
pe
r
f
or
m
a
nc
e
r
e
s
ul
ts
in
to
two
s
ub
s
e
c
ti
o
ns
:
th
e
f
ir
s
t
s
e
c
ti
on,
w
he
n
th
e
m
ode
l
us
e
s
P
la
nt
V
il
la
ge
a
s
tr
a
in
in
g
da
ta
,
a
nd
th
e
s
e
c
ond
s
e
c
ti
on,
w
h
e
n
th
e
m
ode
l
us
e
s
P
la
nt
D
oc
s
a
s
tr
a
in
in
g
da
ta
,
in
c
lu
di
ng
c
om
pa
r
is
ons
be
twe
e
n
th
e
be
s
t
pr
opos
e
d
a
nd
it
s
ba
s
e
li
ne
.
T
hi
s
c
om
pa
r
is
on
il
lu
s
tr
a
te
s
th
e
e
f
f
e
c
t
of
us
in
g
Y
O
L
O
v5
in
th
e
p
r
e
-
pr
oc
e
s
s
in
g
s
ta
ge
on
m
ode
l
pe
r
f
o
r
m
a
nc
e
.
T
o
c
la
r
if
y
th
e
te
r
m
in
ol
ogy,
"
pr
opos
e
d"
r
e
f
e
r
s
t
o t
he
p
r
e
-
pr
oc
e
s
s
in
g m
e
th
od t
ha
t
us
e
s
Y
O
L
O
v5, while
"
ba
s
e
li
ne
"
r
e
f
e
r
s
t
o t
he
s
ta
nda
r
d p
r
e
-
pr
oc
e
s
s
in
g
m
e
th
od
th
a
t
doe
s
not
u
s
e
Y
O
L
O
v5.
W
e
a
ls
o
u
s
e
th
e
te
r
m
"
P
V
"
to
r
e
f
e
r
to
th
e
P
la
nt
V
il
la
ge
da
ta
s
e
t
a
nd
"
P
D
"
to
r
e
f
e
r
t
o t
he
P
la
nt
D
oc
s
da
ta
s
e
t.
3.1.
P
e
r
f
or
m
an
c
e
m
od
e
l
b
as
e
d
on
P
la
n
t
V
il
la
ge
as
t
h
e
t
r
ai
n
i
n
g d
at
a
T
a
bl
e
3
de
m
ons
tr
a
te
s
th
a
t
th
e
I
nc
e
pt
io
n
-
V
3
m
ode
l,
tr
a
in
e
d
a
n
d
te
s
te
d
on
th
e
P
la
nt
V
il
la
g
e
da
ta
s
e
t,
a
c
hi
e
ve
d
a
n
a
c
c
ur
a
c
y
va
lu
e
of
98.28%
f
or
bot
h
our
ba
s
e
li
ne
a
nd
th
e
pr
opos
e
d
m
ode
l.
C
onve
r
s
e
ly
,
te
s
ti
ng
th
e
m
ode
l
w
it
h
th
e
P
la
nt
D
oc
d
a
ta
s
e
t
r
e
s
ul
ts
in
a
d
e
c
r
e
a
s
e
in
it
s
pe
r
f
or
m
a
nc
e
.
H
ow
e
ve
r
,
a
s
m
e
nt
io
ne
d
in
th
e
in
tr
oduc
ti
on,
th
is
out
c
om
e
is
not
s
ur
pr
is
in
g,
gi
ve
n
th
a
t
m
a
ny
c
la
s
s
if
ic
a
ti
ons
a
c
hi
e
ve
hi
gh
a
c
c
ur
a
c
y
w
he
n
us
in
g
c
le
a
n
d
a
ta
s
e
t
s
,
pa
r
ti
c
ul
a
r
ly
f
or
to
m
a
to
pl
a
nt
di
s
e
a
s
e
s
[
14]
.
W
e
a
ls
o
obs
e
r
ve
th
a
t
in
th
is
c
a
s
e
,
th
e
us
e
of
Y
O
L
O
v5 doe
s
not
s
ig
ni
f
ic
a
nt
ly
i
nf
lu
e
nc
e
pe
r
f
or
m
a
nc
e
i
m
pr
ove
m
e
nt
.
T
a
bl
e
3. M
od
e
l
pe
r
f
or
m
a
nc
e
w
he
n t
r
a
in
e
d by
P
la
nt
V
il
la
ge
da
ta
s
e
t
T
e
s
t
i
ng da
t
a
A
c
c
ur
a
c
y (
%
)
A
r
c
hi
t
e
c
t
ur
e
P
l
a
nt
V
i
l
l
a
ge
98.32
ba
s
e
l
i
ne
P
l
a
nt
V
i
l
l
a
ge
98.32
P
r
opos
e
d
P
l
a
nt
D
oc
s
21.54
ba
s
a
e
l
i
ne
P
l
a
nt
D
oc
s
15.28
P
r
opos
e
d
3.2.
P
e
r
f
or
m
an
c
e
m
od
e
l
b
as
e
d
on
t
h
e
P
la
n
t
D
oc
s
as
t
h
e
t
r
ai
n
in
g d
at
a
B
a
s
e
d
on
th
e
a
c
c
ur
a
c
y
c
ur
ve
s
pr
e
s
e
nt
e
d
in
F
ig
ur
e
s
7
a
nd
8,
w
e
c
a
n
obs
e
r
ve
th
a
t
th
e
m
ode
l
ove
r
f
i
ts
w
it
h
a
hi
gh
a
c
c
ur
a
c
y
dur
in
g
tr
a
in
in
g,
but
th
e
va
li
da
ti
on
pr
oc
e
s
s
r
e
ve
a
ls
a
de
c
li
ne
in
th
e
m
ode
l'
s
pe
r
f
or
m
a
nc
e
.
I
t
oc
c
ur
s
w
he
n
th
e
m
ode
l
le
a
r
ns
th
e
tr
a
in
in
g
da
ta
to
o
pr
e
c
is
e
ly
,
in
c
lu
di
ng
noi
s
e
,
w
hi
c
h
ne
ga
ti
ve
ly
im
pa
c
ts
it
s
pe
r
f
or
m
a
nc
e
on
te
s
ti
ng
da
ta
.
I
n
F
ig
ur
e
8,
w
e
c
a
n s
e
e
our
m
ode
l
pe
r
f
or
m
a
nc
e
th
r
ough
s
om
e
of
th
e
c
ur
ve
s
w
it
h
di
f
f
e
r
e
nt
le
ve
ls
of
f
lu
c
tu
a
ti
on.
W
e
obs
e
r
ve
th
a
t
th
e
pr
opo
s
e
d
c
u
r
ve
w
it
h
a
hi
ghe
r
f
lu
c
tu
a
ti
on
in
di
c
a
te
s
th
a
t
th
e
m
ode
l
ha
s
m
or
e
di
f
f
ic
ul
ty
in
ge
ne
r
a
li
z
in
g
th
e
le
a
r
ne
d
f
e
a
tu
r
e
s
f
r
om
th
e
di
r
ty
P
la
nt
D
oc
s
da
ta
s
e
t
to
th
e
c
le
a
n
e
r
P
la
nt
V
il
la
ge
da
ta
s
e
t.
O
ve
r
a
ll
, t
he
us
e
of
t
he
P
la
nt
D
oc
da
ta
s
e
t
te
nds
t
o de
c
r
e
a
s
e
m
ode
l
pe
r
f
or
m
a
nc
e
, bot
h w
it
h
th
e
ba
s
e
li
ne
a
nd t
he
pr
opo
s
e
d m
ode
l.
N
e
ve
r
th
e
le
s
s
,
th
e
r
e
is
s
om
e
th
in
g
in
tr
ig
ui
ng
to
no
te
in
th
e
r
e
s
ul
ts
pr
e
s
e
nt
e
d
in
T
a
bl
e
4.
W
he
n
th
e
m
ode
l
is
tr
a
in
e
d
on
th
e
P
la
nt
D
oc
s
da
ta
s
e
t
a
nd
te
s
te
d
on
th
e
P
la
nt
V
il
la
ge
da
ta
s
e
t,
it
s
how
s
a
s
li
ght
in
c
r
e
a
s
e
in
a
c
c
ur
a
c
y
of
13.9%
.
S
im
il
a
r
ly
,
tr
a
in
in
g
a
nd
te
s
ti
ng
th
e
m
ode
l
on
th
e
P
la
nt
D
oc
da
ta
s
e
t
r
e
s
ul
ts
in
a
gr
e
a
te
r
a
c
c
ur
a
c
y
in
c
r
e
a
s
e
of
27.08%
,
w
hi
c
h
c
a
us
e
s
th
e
m
ode
l'
s
pe
r
f
or
m
a
nc
e
to
a
c
hi
e
ve
a
n
a
c
c
ur
a
c
y
of
62.50%
.
T
hi
s
de
m
ons
tr
a
te
s
th
a
t
u
s
in
g
Y
O
L
O
v5
c
a
n
a
s
s
is
t
in
im
pr
ovi
ng
a
c
c
u
r
a
c
y,
e
ve
n
th
ough
th
e
r
e
s
ul
ts
obt
a
in
e
d
a
r
e
not
ve
r
y hi
gh.
W
he
n
tr
a
in
e
d
a
nd
t
e
s
te
d
on
th
e
f
ie
ld
da
ta
s
e
t
(
P
la
nt
D
oc
s
da
ta
s
e
t
)
,
our
pr
opos
e
d
a
r
c
hi
te
c
tu
r
e
a
c
hi
e
ve
d
62.50%
a
c
c
ur
a
c
y.
W
hi
le
th
is
va
lu
e
is
n'
t
ve
r
y
hi
gh,
it
hi
ghl
ig
ht
s
Y
O
L
O
v5'
s
s
tr
e
ngt
h
in
obj
e
c
t
d
e
te
c
ti
on
dur
in
g
pr
e
-
pr
oc
e
s
s
in
g,
w
hi
c
h
i
s
ke
y
f
or
id
e
nt
if
yi
ng
to
m
a
to
pl
a
nt
le
a
f
di
s
e
a
s
e
s
.
T
hi
s
unde
r
s
c
or
e
s
th
e
im
por
ta
nc
e
of
us
in
g
r
e
a
l
-
c
ondi
ti
on
da
ta
s
e
ts
to
bui
ld
r
obus
t
m
ode
ls
.
H
ow
e
ve
r
,
it
is
ne
c
e
s
s
a
r
y
to
be
a
tt
e
nt
iv
e
to
pr
e
pa
r
in
g
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
R
obus
t
tw
o
-
s
ta
ge
obj
e
c
t
de
te
c
ti
on us
in
g Y
O
L
O
v
5 f
o
r
e
nhanc
in
g
t
om
at
o l
e
af
…
(
E
ndang Sur
y
aw
at
i
)
2253
da
ta
s
e
t
w
e
ll
,
w
hi
c
h
ne
e
ds
s
ys
t
e
m
a
ti
c
a
na
ly
s
is
of
f
ie
ld
-
s
pe
c
if
ic
da
ta
va
r
ia
ti
ons
a
nd
th
e
ir
in
f
lu
e
nc
e
on
th
e
m
ode
l'
s
e
r
r
or
r
a
te
s
.
F
ig
ur
e
7. T
r
a
in
in
g a
c
c
ur
a
c
y
F
ig
ur
e
8. V
a
li
da
ti
on
a
c
c
ur
a
c
y
T
a
bl
e
4. M
od
e
l
pe
r
f
or
m
a
nc
e
w
he
n t
r
a
in
e
d by
P
la
nt
D
oc
s
d
a
ta
s
e
t
T
e
s
t
i
ng da
t
a
A
c
c
ur
a
c
y (
%
)
A
r
c
hi
t
e
c
t
ur
e
P
l
a
nt
V
i
l
l
a
ge
12.59
ba
s
e
l
i
ne
P
l
a
nt
V
i
l
l
a
ge
26.49
pr
opos
e
d
P
l
a
nt
D
oc
s
35.42
ba
s
a
e
l
i
ne
P
l
a
nt
D
oc
s
62.50
P
r
opos
e
d
M
a
ny
r
e
s
e
a
r
c
he
r
s
in
th
e
pr
e
vi
ous
s
tu
dy
f
oc
us
on
a
c
hi
e
vi
ng
hi
gh
a
c
c
ur
a
c
y
us
in
g
c
le
a
n
da
ta
s
e
ts
w
it
h
va
r
io
us
C
N
N
a
r
c
hi
te
c
tu
r
e
s
,
but
our
r
e
s
ul
t
s
s
ugge
s
t
th
a
t
popula
r
C
N
N
s
s
tr
uggl
e
w
it
h
r
e
a
l
-
c
ondi
ti
on
da
ta
s
e
t
s
.
T
o
ve
r
if
y
th
is
,
w
e
te
s
te
d
s
e
ve
r
a
l
C
N
N
s
on
th
e
P
la
nt
D
oc
s
da
ta
s
e
t,
s
how
in
g
a
dr
op
in
pe
r
f
or
m
a
nc
e
,
a
s
de
ta
il
e
d
in
T
a
bl
e
5.
T
he
P
D
da
ta
s
e
t
e
nc
om
pa
s
s
e
s
im
a
ge
s
of
di
s
e
a
s
e
s
a
nd
unw
a
nt
e
d
obj
e
c
ts
,
boa
s
t
s
a
w
id
e
r
a
nge
of
im
a
ge
s
iz
e
s
,
a
nd
m
a
y
in
c
lu
de
m
ul
ti
pl
e
le
a
ve
s
w
it
h
a
va
r
ie
ty
of
ba
c
kgr
ounds
a
nd
li
ght
in
g
c
ondi
ti
on
s
.
T
h
e
c
onvolut
io
na
l
la
ye
r
s
f
in
d
it
c
ha
ll
e
ngi
ng
to
e
xt
r
a
c
t
f
e
a
tu
r
e
s
f
r
om
th
e
P
D
da
ta
s
e
t,
w
hi
c
h
f
r
e
que
nt
ly
c
ont
a
in
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
, V
ol
.
14
, N
o.
3
,
J
une
20
25
:
2246
-
2257
2254
ir
r
e
le
va
nt
ba
c
kgr
ound
or
noi
s
e
.
T
he
r
e
f
or
e
,
th
e
m
ode
l
ha
s
di
f
f
ic
ul
ty
di
s
ti
ngui
s
hi
ng
di
s
e
a
s
e
s
pot
s
or
noi
s
e
.
I
t
is
ne
c
e
s
s
a
r
y
to
a
ppl
y
a
r
obus
t
pr
e
-
pr
oc
e
s
s
in
g
te
c
hni
que
to
he
lp
lo
c
a
li
z
e
th
e
de
s
ir
e
d
di
s
e
a
s
e
s
pot
s
a
nd
s
e
p
a
r
a
te
th
e
m
f
r
om
th
e
noi
s
e
obj
e
c
ts
.
T
o
im
pr
ove
th
e
qua
li
ty
of
th
e
noi
s
e
a
nd
in
c
r
e
a
s
e
th
e
a
m
ount
of
a
r
ti
f
ic
ia
l
da
ta
,
w
e
ne
e
d t
o a
ppl
y t
he
a
ugm
e
nt
a
ti
on t
e
c
hni
que
. T
hi
s
w
il
l
a
ll
ow
u
s
t
o a
da
pt
t
he
m
ode
l
to
doma
in
s
w
it
h di
f
f
e
r
e
n
t
le
ve
ls
of
va
r
ia
bi
li
ty
[
35]
.
F
ur
th
e
r
,
w
e
c
a
n
be
ne
f
it
f
r
om
Y
O
L
O
v5'
s
c
a
pa
bi
li
ty
in
va
r
io
us
a
ugm
e
nt
a
ti
on
te
c
hni
que
s
due
t
o i
ts
e
a
s
e
of
m
odi
f
ic
a
ti
on.
T
a
bl
e
5. C
om
pa
r
is
on of
pe
r
f
or
m
a
nc
e
be
twe
e
n our
pr
opos
e
d a
nd
ot
he
r
C
N
N
a
r
c
hi
te
c
tu
r
e
s
A
r
c
hi
t
e
c
t
ur
e
A
c
c
ur
a
c
y (
%
)
R
e
s
ne
t
-
50
41.07
D
e
ns
e
N
e
t
-
121
43.57
M
obi
l
e
N
e
t
-
V3
33.13
E
f
f
i
c
i
e
nt
N
e
t
-
V2
43.48
I
nc
e
pt
i
onR
e
s
N
e
t
-
V2
32.71
I
nc
e
pt
i
on
-
V3
35.42
O
ur
pr
opos
e
d
62.50
T
o
de
ve
lo
p
a
r
obus
t
c
la
s
s
if
ic
a
ti
on
m
ode
l,
w
e
ne
e
d
to
tr
a
in
a
nd
te
s
t
th
e
m
ode
l
us
in
g
a
f
ie
ld
da
ta
s
e
t
w
it
h
hi
gh
va
r
ia
bi
li
ty
th
a
t
r
e
pr
e
s
e
nt
s
th
e
r
e
a
l
e
nvi
r
onm
e
nt
,
in
c
lu
di
ng
va
r
io
us
im
a
ge
ba
c
kgr
ounds
a
nd
noi
s
e
[
36]
.
W
e
ut
il
iz
e
d
th
e
h
ig
hl
y
va
r
ia
bl
e
P
la
nt
D
oc
s
da
ta
s
e
t
f
or
our
pr
opos
e
d
a
r
c
hi
te
c
tu
r
e
s
[
37]
,
but
T
a
bl
e
2
r
e
ve
a
ls
a
n
im
ba
la
nc
e
in
th
e
num
be
r
of
s
a
m
pl
e
s
f
or
e
a
c
h
di
s
e
a
s
e
ty
pe
.
C
la
s
s
im
ba
la
nc
e
a
r
is
e
s
w
he
n
one
di
s
e
a
s
e
c
la
s
s
dom
in
a
te
s
th
e
da
ta
s
e
t,
le
a
vi
ng
ot
he
r
di
s
e
a
s
e
s
unde
r
r
e
pr
e
s
e
nt
e
d.
T
hi
s
im
ba
la
nc
e
m
a
k
e
s
th
e
m
ode
l
di
f
f
ic
ul
t
to
ge
ne
r
a
li
z
e
im
por
ta
nt
f
e
a
tu
r
e
s
to
ne
w
da
ta
,
a
s
it
le
a
r
ns
ove
r
ly
s
pe
c
if
ic
pa
tt
e
r
ns
a
nd
ig
nor
e
s
m
or
e
ge
ne
r
a
l
one
s
.
F
or
f
ut
ur
e
w
or
k,
it
is
ne
c
e
s
s
a
r
y
to
in
c
r
e
a
s
e
th
e
a
m
ount
of
da
ta
t
o
e
ns
ur
e
a
ba
la
n
c
e
d
num
be
r
f
or
e
a
c
h
c
la
s
s
w
hi
le
a
l
s
o
e
ns
ur
in
g
ba
la
nc
e
d
va
r
ia
bi
li
ty
[
38]
,
[
39]
.
W
e
a
ls
o
c
ons
id
e
r
e
d
c
om
bi
ni
ng
th
e
di
r
ty
a
nd
c
le
a
n
da
ta
s
e
ts
t
o a
im
a
t
a
ba
la
nc
e
d va
r
ia
bi
li
ty
o
f
t
he
da
ta
s
e
t.
T
he
us
e
of
Y
O
L
O
s
ti
ll
pr
ovi
de
s
c
onf
id
e
nc
e
a
s
a
r
obus
t
pr
e
-
pr
oc
e
s
s
or
.
Y
O
L
O
v5
s
ig
ni
f
ic
a
nt
ly
a
id
s
th
e
m
ode
l
in
f
oc
us
i
ng
on
th
e
e
xt
r
a
c
te
d
f
e
a
tu
r
e
s
.
H
ow
e
ve
r
,
th
e
us
e
of
ba
c
kgr
ound
r
e
m
ova
l
te
c
hni
que
s
ne
e
d
s
to
be
in
vol
ve
d
to
im
pr
ove
da
ta
qua
li
ty
.
W
it
hout
r
obus
t
pr
e
-
pr
oc
e
s
s
in
g,
th
e
m
ode
l
ha
s
di
f
f
ic
ul
ty
e
xt
r
a
c
ti
ng
f
oc
u
s
e
d
f
e
a
tu
r
e
s
,
r
e
s
ul
ti
ng
in
de
c
r
e
a
s
e
d a
c
c
ur
a
c
y
[
36]
.
F
or
im
pr
ovi
ng
th
e
m
ode
l'
s
a
bi
li
ty
to
ge
n
e
r
a
li
z
e
f
e
a
tu
r
e
s
be
tw
e
e
n
dom
a
in
s
w
it
h
di
f
f
e
r
e
nt
va
r
ia
bi
li
ty
,
it
is
ne
c
e
s
s
a
r
y t
o s
e
le
c
t
th
e
r
ig
ht
doma
in
a
da
pt
a
ti
on t
e
c
hni
que
a
nd r
e
gul
a
r
iz
a
ti
on t
e
c
hni
que
. F
ur
th
e
r
m
or
e
, w
e
m
us
t
e
nha
nc
e
th
e
hype
r
pa
r
a
m
e
te
r
v
a
lu
e
s
e
tt
in
gs
f
or
m
ode
l
tr
a
in
in
g,
in
c
lu
di
ng
le
a
r
ni
ng
r
a
te
,
opt
im
iz
e
r
,
a
nd
ba
t
c
h
s
iz
e
, w
hi
le
a
l
s
o i
m
pl
e
m
e
nt
in
g s
ui
ta
bl
e
r
e
gul
a
r
iz
a
ti
on t
e
c
hni
que
s
.
4.
C
O
N
C
L
U
S
I
O
N
I
n
th
is
s
tu
dy,
w
e
pr
opos
e
a
two
-
s
ta
ge
de
te
c
ti
on
a
r
c
hi
te
c
tu
r
e
f
or
id
e
nt
if
yi
ng
c
la
s
s
e
s
of
in
f
e
c
te
d
to
m
a
to
pl
a
nt
s
.
I
n
th
e
pr
e
-
pr
oc
e
s
s
in
g
s
ta
ge
,
w
e
ut
il
iz
e
Y
O
L
O
v5
to
de
te
c
t
obj
e
c
ts
w
it
hi
n
th
e
im
a
ge
s
.
T
h
e
de
te
c
te
d
obj
e
c
ts
f
r
om
th
e
P
la
nt
D
oc
s
a
nd
P
la
nt
V
il
la
ge
da
ta
s
e
t
a
r
e
th
e
n
c
la
s
s
if
ie
d
in
to
known
c
la
s
s
e
s
us
in
g
I
nc
e
pt
io
n
-
V
3 m
ode
l.
O
ur
e
va
lu
a
ti
on of
t
w
o da
ta
s
e
ts
c
onf
ir
m
s
t
ha
t
our
pr
opos
e
d a
r
c
hi
te
c
tu
r
e
i
s
m
or
e
e
f
f
e
c
ti
ve
f
or
di
s
e
a
s
e
d
to
m
a
to
pl
a
nt
de
te
c
ti
on,
s
pe
c
if
ic
a
ll
y
w
he
n
th
e
c
l
a
s
s
if
ie
r
m
ode
l
is
tr
a
in
e
d
a
nd
te
s
te
d
us
in
g
th
e
P
la
nt
D
oc
s
da
ta
s
e
t.
I
n
th
i
s
c
a
s
e
,
Y
O
L
O
v5
s
uppor
t
our
a
r
c
hi
te
c
t
ur
e
f
or
de
te
c
ti
ng
r
e
gi
ons
of
in
te
r
e
s
t
(
R
O
I
)
a
nd
di
s
ti
ngui
s
hi
ng
im
por
ta
nt
f
e
a
tu
r
e
s
f
r
om
th
e
noi
s
e
w
hi
c
h
a
r
e
pr
e
s
e
nt
in
th
e
P
la
nt
D
oc
s
d
a
ta
s
e
t.
A
lt
hough
th
e
e
xpe
r
im
e
nt
a
l
r
e
s
ul
ts
s
how
a
m
ode
r
a
te
a
c
c
ur
a
c
y
va
lu
e
(
62.50
%
)
,
th
is
r
e
s
e
a
r
c
h
ha
s
th
e
pot
e
nt
ia
l
f
or
f
ut
ur
e
im
pr
ove
m
e
nt
.
W
e
ne
e
d
to
pr
e
pa
r
e
th
e
da
ta
s
e
t
w
it
h
ba
la
nc
e
d
v
a
r
ia
bi
li
ty
to
a
c
hi
e
ve
d
a
m
or
e
r
obus
t
m
ode
l
f
o
r
de
te
c
ti
ng
di
s
e
a
s
e
d
to
m
a
to
le
a
ve
s
.
O
ur
hope
is
to
de
ve
lo
p
a
m
or
e
a
c
c
ur
a
te
m
ode
l
by
us
in
g
a
ba
la
nc
e
d
da
ta
s
e
t,
a
s
ophi
s
ti
c
a
te
d
pr
e
-
pr
oc
e
s
s
or
li
ke
Y
O
L
O
v5,
th
e
a
ppr
opr
ia
te
r
e
gul
a
r
iz
a
ti
on
te
c
hni
que
s
,
a
nd
th
e
s
ui
ta
bl
e
dom
a
in
a
da
pt
a
ti
on t
e
c
hni
que
.
F
U
N
D
I
N
G
I
N
F
O
R
M
A
T
I
O
N
T
hi
s
r
e
s
e
a
r
c
h
w
a
s
f
unde
d
by
R
um
a
h
P
r
ogr
a
m
A
r
ti
f
ic
ia
l
I
nt
e
ll
ig
e
nc
e
,
B
ig
D
a
ta
,
da
n
T
e
knol
ogi
K
om
put
a
s
i
unt
uk
B
io
di
ve
r
s
it
a
s
da
n
C
it
r
a
S
a
te
li
t
-
E
le
c
tr
oni
c
s
a
nd
I
nf
or
m
a
ti
c
s
R
e
s
e
a
r
c
h
O
r
ga
ni
z
a
ti
on
(
O
R
E
I
)
-
N
a
ti
ona
l
R
e
s
e
a
r
c
h
a
nd
I
nnova
ti
on
A
ge
nc
y
(
B
R
I
N
)
,
C
o
nt
r
a
c
t
N
um
be
r
:
83/
I
I
I
.6.4/
H
K
/2
023,
M
a
r
c
h
3,
2023,
in
c
ol
la
bor
a
ti
on
w
it
h
D
e
pa
r
tm
e
nt
of
S
ta
ti
s
ti
c
s
,
F
a
c
ul
ty
of
M
a
th
e
m
a
ti
c
s
a
nd
N
a
tu
r
a
l
S
c
ie
nc
e
s
a
t
P
a
dj
a
dj
a
r
a
n U
ni
ve
r
s
it
y.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
R
obus
t
tw
o
-
s
ta
ge
obj
e
c
t
de
te
c
ti
on us
in
g Y
O
L
O
v
5 f
o
r
e
nhanc
in
g
t
om
at
o l
e
af
…
(
E
ndang Sur
y
aw
at
i
)
2255
A
U
T
H
O
R
C
O
N
T
R
I
B
U
T
I
O
N
S
S
T
A
T
E
M
E
N
T
T
hi
s
jo
ur
na
l
us
e
s
th
e
C
ont
r
ib
ut
or
R
ol
e
s
T
a
xonomy
(
C
R
e
di
T
)
to
r
e
c
ogni
z
e
in
di
vi
dua
l
a
ut
hor
c
ont
r
ib
ut
io
ns
, r
e
duc
e
a
ut
hor
s
hi
p di
s
put
e
s
,
a
nd f
a
c
il
it
a
te
c
ol
la
bo
r
a
ti
on.
N
am
e
o
f
A
u
t
h
or
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
E
nda
ng
S
ur
ya
w
a
ti
✓
✓
✓
✓
✓
✓
✓
✓
✓
S
yi
f
a
A
ul
iy
a
h H
a
s
a
na
h
✓
✓
✓
✓
✓
✓
✓
✓
✓
R
a
de
n S
a
ndr
a
Y
uw
a
na
✓
✓
✓
✓
✓
✓
✓
J
im
m
y A
bde
l
K
a
da
r
✓
✓
✓
✓
✓
✓
✓
H
il
m
a
n F
e
r
di
na
ndus
P
a
r
de
de
✓
✓
✓
✓
✓
✓
✓
✓
C
:
C
onc
e
pt
ua
l
i
z
a
t
i
on
M
:
M
e
t
hodol
ogy
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
da
t
i
on
Fo
:
Fo
r
m
a
l
a
na
l
ys
i
s
I
:
I
nve
s
t
i
ga
t
i
on
R
:
R
e
s
our
c
e
s
D
:
D
a
t
a
C
ur
a
t
i
on
O
:
W
r
i
t
i
ng
-
O
r
i
gi
na
l
D
r
a
f
t
E
:
W
r
i
t
i
ng
-
R
e
vi
e
w
&
E
di
t
i
ng
Vi
:
Vi
s
ua
l
i
z
a
t
i
on
Su
:
Su
pe
r
vi
s
i
on
P
:
P
r
oj
e
c
t
a
dm
i
ni
s
t
r
a
t
i
on
Fu
:
Fu
ndi
ng a
c
qui
s
i
t
i
on
C
O
N
F
L
I
C
T
O
F
I
N
T
E
R
E
S
T
S
T
A
T
E
M
E
N
T
A
ut
hor
s
s
ta
te
no c
onf
li
c
t
of
i
nt
e
r
e
s
t.
D
A
T
A
A
V
A
I
L
A
B
I
L
I
T
Y
T
he
da
ta
s
e
ts
us
e
d
in
th
is
s
tu
dy
a
r
e
publ
ic
ly
a
c
c
e
s
s
ib
le
a
n
d
w
e
r
e
obt
a
in
e
d
f
r
om
ope
n
-
s
our
c
e
r
e
pos
it
or
ie
s
,
s
pe
c
if
ic
a
ll
y:
-
T
he
da
ta
th
a
t
s
uppor
t
th
e
f
in
di
ngs
of
th
is
s
tu
dy,
in
c
onne
c
ti
on
w
it
h
th
e
c
le
a
n
d
a
ta
s
e
t
a
r
e
ope
nl
y
a
v
a
il
a
bl
e
in
[
P
la
nt
V
il
la
ge
da
ta
s
e
t]
a
t
ht
tp
s
:/
/g
it
hub.c
om
/s
pM
oha
nt
y/
P
la
nt
V
il
la
ge
-
D
a
ta
s
e
t
-
T
he
da
ta
th
a
t
s
uppor
t
th
e
f
in
di
ngs
of
th
is
s
tu
dy
th
a
t
r
e
la
te
d
t
o
th
e
di
r
ty
da
ta
s
e
t
a
r
e
ope
nl
y
a
va
il
a
bl
e
in
[
P
la
nt
D
oc
]
a
t
ht
tp
s
:/
/g
it
hub.c
om
/p
r
a
ti
kka
ya
l/
P
la
nt
D
oc
-
D
a
ta
s
e
t
R
E
F
E
R
E
N
C
E
S
[
1]
A
.
A
bba
s
e
t
al
.
,
“
D
r
one
s
i
n
pl
a
nt
di
s
e
a
s
e
a
s
s
e
s
s
m
e
nt
,
e
f
f
i
c
i
e
nt
m
oni
t
or
i
ng,
a
nd
de
t
e
c
t
i
on:
a
w
a
y
f
or
w
a
r
d
t
o
s
m
a
r
t
a
gr
i
c
ul
t
ur
e
,”
A
gr
onom
y
, vol
. 13, no. 6, pp. 1
–
26, 2023, doi
:
10.3390/
a
gr
onom
y13061524.
[
2]
M
.
B
ha
nda
r
i
,
T
.
B
.
S
ha
hi
,
A
.
N
e
upa
ne
,
a
nd
K
.
B
.
W
a
l
s
h,
“
B
ot
a
ni
c
X
-
a
i
:
i
de
nt
i
f
i
c
a
t
i
on
of
t
om
a
t
o
l
e
a
f
di
s
e
a
s
e
s
us
i
ng
a
n
e
xpl
a
na
t
i
on
-
dr
i
ve
n de
e
p
-
l
e
a
r
ni
ng m
ode
l
,”
J
our
nal
of
I
m
agi
ng
, vol
. 9, no. 2, 2023, doi
:
10.3390/
j
i
m
a
gi
ng9020053.
[
3]
B
.
S
.
N
a
w
a
l
e
a
nd
H
.
D
.
G
a
da
de
,
“
A
s
ys
t
e
m
a
t
i
c
r
e
vi
e
w
:
de
t
e
c
t
i
ng
pl
a
nt
di
s
e
a
s
e
s
us
i
ng
m
a
c
hi
ne
l
e
a
r
ni
ng
t
e
c
hni
que
s
,”
i
n
2023
11t
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
E
m
e
r
gi
ng
T
r
e
nd
s
i
n
E
ngi
ne
e
r
i
ng
&
T
e
c
hnol
ogy
-
S
i
gnal
and
I
nf
or
m
at
i
on
P
r
oc
e
s
s
i
ng
(
I
C
E
T
E
T
-
S
I
P
)
,
I
E
E
E
, 2023, pp. 1
–
5
, doi
:
10.1109/
I
C
E
T
E
T
-
S
I
P
58143.2023.10151590.
[
4]
T
.
S
.
X
i
a
n
a
nd
R
.
N
ga
di
r
a
n,
“
P
l
a
nt
di
s
e
a
s
e
s
c
l
a
s
s
i
f
i
c
a
t
i
on
us
i
ng
m
a
c
hi
ne
l
e
a
r
ni
ng,”
J
our
nal
of
P
hy
s
i
c
s
:
C
onf
e
r
e
n
c
e
Se
r
i
e
s
,
vol
. 1962, no. 1, 2021, doi
:
10.1088/
1742
-
6596/
1962/
1/
012024.
[
5]
M
.
W
u,
J
.
Z
hou,
Y
.
P
e
ng,
S
.
W
a
ng,
a
nd
Y
.
Z
ha
ng,
“
D
e
e
p
l
e
a
r
ni
ng
f
o
r
i
m
a
ge
c
l
a
s
s
i
f
i
c
a
t
i
on:
a
r
e
vi
e
w
,”
i
n
P
r
oc
e
e
di
ngs
of
202
3
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
M
e
di
c
al
I
m
agi
ng
and
C
om
put
e
r
-
A
i
de
d
D
i
agnos
i
s
(
M
I
C
A
D
2023
)
,
2024,
pp.
352
–
362
,
doi
:
10.1007/
978
-
981
-
97
-
1335
-
6_31.
[
6]
W
.
S
.
M
c
C
ul
l
oc
h
a
nd
W
.
P
i
t
t
s
,
“
A
l
ogi
c
a
l
c
a
l
c
ul
us
of
t
he
i
de
a
s
i
m
m
a
ne
nt
i
n
ne
r
vous
a
c
t
i
vi
t
y,”
T
he
B
ul
l
e
t
i
n
of
M
at
he
m
at
i
c
al
B
i
ophy
s
i
c
s
, vol
. 5, no. 4, pp.
115
–
133, 1943, doi
:
10.1007/
B
F
02478259.
[
7]
M
.
E
l
t
a
y,
A
.
Z
i
dour
i
,
a
nd
I
.
A
hm
a
d,
“
E
xpl
o
r
i
ng
de
e
p
l
e
a
r
ni
ng
a
ppr
oa
c
he
s
t
o
r
e
c
ogni
z
e
ha
ndw
r
i
t
t
e
n
a
r
a
bi
c
t
e
xt
s
,”
I
E
E
E
A
c
c
e
s
s
,
vol
. 8, pp. 89882
–
89898, 2020, doi
:
10.1109/
A
C
C
E
S
S
.2020.2994248.
[
8]
Z
.
Z
ha
ng
e
t
al
.
,
“
D
e
ns
e
r
e
s
i
dua
l
ne
t
w
or
k:
e
nha
nc
i
ng
gl
oba
l
de
ns
e
f
e
a
t
ur
e
f
l
ow
f
or
c
ha
r
a
c
t
e
r
r
e
c
ogni
t
i
on,”
N
e
ur
al
N
e
t
w
or
k
s
,
vol
. 139, pp. 77
–
85, 2021, doi
:
10.1016/
j
.ne
une
t
.2021.02.005.
[
9]
K
. N
oda
, Y
. Y
a
m
a
guc
hi
,
K
. N
a
ka
da
i
, H
.
G
. O
kuno, a
nd T
. O
ga
t
a
,
“
A
udi
o
-
vi
s
u
a
l
s
pe
e
c
h r
e
c
ogni
t
i
on u
s
i
ng de
e
p l
e
a
r
ni
ng,”
A
ppl
i
e
d
I
nt
e
l
l
i
ge
nc
e
, vol
. 42, no. 4, pp. 722
–
737, 2015, doi
:
10.1007/
s
10489
-
014
-
0629
-
7.
[
10]
A
.
B
.
N
a
s
s
i
f
,
I
.
S
ha
hi
n,
I
.
A
t
t
i
l
i
,
M
.
A
z
z
e
h,
a
nd
K
.
S
ha
a
l
a
n,
“
S
pe
e
c
h
r
e
c
ogni
t
i
on
us
i
ng
de
e
p
ne
ur
a
l
ne
t
w
or
ks
:
a
s
ys
t
e
m
a
t
i
c
r
e
vi
e
w
,”
I
E
E
E
A
c
c
e
s
s
, vol
. 7, pp. 19143
–
19165, 2019, doi
:
10.1109/
A
C
C
E
S
S
.2019.2896880.
[
11]
X
.
Z
ha
o,
L
.
W
a
ng,
Y
.
Z
ha
ng,
X
.
H
a
n,
M
.
D
e
ve
c
i
,
a
nd
M
.
P
a
r
m
a
r
,
“
A
r
e
vi
e
w
of
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
ks
i
n
c
om
put
e
r
vi
s
i
on,”
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
R
e
v
i
e
w
, vol
. 57, no. 4, 2024, doi
:
10.1007/
s
10462
-
024
-
10721
-
6.
[
12]
Y
.
L
i
,
“
R
e
s
e
a
r
c
h
a
nd
a
ppl
i
c
a
t
i
on
of
de
e
p
l
e
a
r
ni
ng
i
n
i
m
a
ge
r
e
c
ogni
t
i
on,”
i
n
2
022
I
E
E
E
2nd
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
P
ow
e
r
,
E
l
e
c
t
r
oni
c
s
and C
om
put
e
r
A
ppl
i
c
at
i
ons
(
I
C
P
E
C
A
)
,
I
E
E
E
, 2022, pp. 994
–
999
, d
oi
:
10.1109/
I
C
P
E
C
A
53709.2022.9718847.
[
13]
P
.
W
a
ng,
E
.
F
a
n,
a
nd
P
.
W
a
ng,
“
C
om
pa
r
a
t
i
ve
a
na
l
ys
i
s
of
i
m
a
ge
c
l
a
s
s
i
f
i
c
a
t
i
on
a
l
gor
i
t
hm
s
ba
s
e
d
on
t
r
a
di
t
i
ona
l
m
a
c
hi
ne
l
e
a
r
ni
ng
a
nd de
e
p l
e
a
r
ni
ng,”
P
at
t
e
r
n R
e
c
ogni
t
i
on L
e
t
t
e
r
s
, vol
. 141, pp. 61
–
67, 2021, doi
:
10.1016/
j
.pa
t
r
e
c
.2020.07.042.
[
14]
E
. S
ur
ya
w
a
t
i
, R
. S
us
t
i
ka
, R
.
S
. Y
uw
a
n
a
, A
. S
ube
kt
i
,
a
nd H
. F
. P
a
r
de
d
e
, “
D
e
e
p
s
t
r
uc
t
ur
e
d c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k f
or
t
om
a
t
o
di
s
e
a
s
e
s
de
t
e
c
t
i
on,”
i
n
2018 I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on A
dv
anc
e
d
C
om
put
e
r
Sc
i
e
nc
e
and I
nf
or
m
at
i
on Sy
s
t
e
m
s
(
I
C
A
C
SI
S)
, I
E
E
E
,
2018, pp. 385
–
390
, doi
:
10.1109/
I
C
A
C
S
I
S
.2018.8618169.
[
15]
P
.
S
um
a
r
i
,
A
.
M
.
K
a
s
s
i
m
,
S
.
Q
.
O
ng,
G
.
N
a
i
r
,
A
.
D
.
R
a
gh
e
e
d,
a
nd
N
.
F
.
A
m
i
nuddi
n,
“
C
l
a
s
s
i
f
i
c
a
t
i
on
of
j
a
c
kf
r
ui
t
a
nd
c
e
m
pe
da
k
us
i
ng
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k
a
nd
t
r
a
ns
f
e
r
l
e
a
r
ni
ng,”
I
A
E
S
I
nt
e
r
nat
i
ona
l
J
our
nal
of
A
r
t
i
f
i
c
i
al
I
nt
e
l
l
i
ge
nc
e
,
vol
.
11,
no.
4,
pp. 1353
–
1361, 2022, doi
:
10.11591/
i
j
a
i
.v11.i
4.pp1353
-
1361.
Evaluation Warning : The document was created with Spire.PDF for Python.