I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l.
15
,
No
.
3
,
J
u
n
e
20
25
,
p
p
.
3319
~
3
3
3
1
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
15
i
3
.
pp
3
3
1
9
-
3
3
3
1
3319
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Im
pro
v
ed YO
L
O
v
1
0
mo
del f
o
r det
ecting sur
face de
f
ects on
so
la
r photo
v
o
ltaic pa
nels
P
ha
t
T
.
Ng
uy
en
1
,
L
o
c
D.
H
o
2
,
Duy
C.
H
uy
nh
2
1
D
e
p
a
r
t
me
n
t
o
f
C
i
r
c
u
i
t
t
h
e
o
r
y
a
n
d
me
a
su
r
e
m
e
n
t
,
F
a
c
u
l
t
y
o
f
R
a
d
i
o
-
El
e
c
t
r
o
n
i
c
E
n
g
i
n
e
e
r
i
n
g
,
L
e
Q
u
y
D
o
n
Te
c
h
n
i
c
a
l
U
n
i
v
e
r
s
i
t
y
,
H
a
N
o
i
,
V
i
e
t
n
a
m
2
H
U
TEC
H
I
n
st
i
t
u
t
e
o
f
E
n
g
i
n
e
e
r
i
n
g
,
H
U
TEC
H
U
n
i
v
e
r
s
i
t
y
,
H
o
C
h
i
M
i
n
h
C
i
t
y
,
V
i
e
t
n
a
m
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Sep
2
0
,
2
0
2
4
R
ev
is
ed
Dec
1
8
,
2
0
2
4
Acc
ep
ted
J
an
1
4
,
2
0
2
5
S
u
rfa
c
e
d
e
fe
c
ts
g
re
a
tl
y
a
ffe
c
t
th
e
p
e
rf
o
rm
a
n
c
e
a
n
d
se
rv
ic
e
li
fe
o
f
p
h
o
to
v
o
lt
a
ic
(P
V)
m
o
d
u
les
.
De
tec
ti
n
g
t
h
e
se
d
e
fe
c
ts
is
imp
o
rtan
t
t
o
imp
ro
v
e
th
e
m
a
n
a
g
e
m
e
n
t,
re
p
a
ir
a
n
d
m
a
in
ten
a
n
c
e
o
f
P
V
p
a
n
e
ls.
Wi
th
t
h
e
d
e
v
e
lo
p
m
e
n
t
o
f
a
rti
f
icia
l
in
tel
li
g
e
n
c
e
,
c
o
m
p
u
ter
v
isi
o
n
b
ri
n
g
s
h
i
g
h
e
r
a
c
c
u
ra
c
y
a
n
d
lo
we
r
lab
o
r
c
o
sts
th
a
n
trad
i
ti
o
n
a
l
in
s
p
e
c
ti
o
n
m
e
t
h
o
d
s
.
Th
i
s
p
a
p
e
r
in
tr
o
d
u
c
e
s
a
n
imp
ro
v
e
d
P
V
y
o
u
o
n
ly
lo
o
k
o
n
c
e
v
1
0
(
YO
LOv
1
0
)
m
o
d
e
l
f
o
r
d
e
tec
ti
n
g
s
u
rfa
c
e
d
e
fe
c
ts
o
f
P
V
m
o
d
u
les
.
T
h
e
imp
ro
v
e
m
e
n
t
in
c
lu
d
e
s
a
d
d
i
n
g
a
n
e
x
p
o
n
e
n
ti
a
l
m
o
v
i
n
g
a
v
e
ra
g
e
(EM
A)
a
tt
e
n
ti
o
n
m
e
c
h
a
n
ism
to
t
h
e
n
e
c
k
,
u
si
n
g
a
c
y
c
le
g
e
n
e
ra
ti
v
e
a
d
v
e
rsa
rial
n
e
two
rk
(G
AN
)
to
e
n
h
a
n
c
e
th
e
d
a
ta,
a
n
d
re
p
lac
in
g
t
h
e
YO
L
Ov
1
0
h
e
a
d
with
a
YO
LOv
9
h
e
a
d
t
o
re
tain
non
-
m
a
x
imu
m
su
p
p
re
ss
io
n
(NM
S
).
Ex
p
e
rime
n
ts
sh
o
w
t
h
a
t
th
e
p
ro
p
o
se
d
m
o
d
e
l
o
u
t
p
e
rfo
rm
s
sta
te
-
of
-
t
h
e
-
a
rt
m
e
th
o
d
s
su
c
h
a
s
YO
LOv
1
0
s,
n
,
x
,
b
,
l,
a
n
d
e
,
a
c
h
iev
i
n
g
su
p
e
ri
o
r
d
e
t
e
c
ti
o
n
a
c
c
u
ra
c
y
.
De
sp
it
e
t
h
e
in
c
re
a
se
d
c
o
m
p
u
tati
o
n
a
l
c
o
st,
th
e
p
r
o
p
o
se
d
m
e
th
o
d
imp
r
o
v
e
d
m
AP
@0
.
5
a
n
d
m
AP@0
.
5
:
0
.
9
5
b
y
5
.
1
%
a
n
d
6
.
5
%
o
v
e
r
th
e
o
ri
g
in
a
l
YO
LOv
1
0
s
.
K
ey
w
o
r
d
s
:
Atten
tio
n
m
ec
h
an
is
m
Dee
p
lear
n
in
g
Gen
er
ativ
e
ad
v
e
r
s
ar
ial
n
etwo
r
k
Ph
o
to
v
o
ltaic
PV d
ef
ec
t d
etec
tio
n
YOL
Ov
1
0
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Ph
at
T
.
Ng
u
y
e
n
Dep
ar
tm
en
t o
f
C
ir
cu
it
T
h
eo
r
y
an
d
Me
asu
r
em
en
t
,
Facu
lty
o
f
R
ad
io
-
E
lectr
o
n
ic
E
n
g
in
ee
r
in
g
,
L
e
Qu
y
Do
n
T
ec
h
n
ical
Un
iv
er
s
ity
2
3
6
Ho
a
n
g
Qu
o
c
Viet
R
o
ad
,
Ha
No
i,
Viet
n
am
E
m
ail: n
g
u
y
e
n
p
h
at@
lq
d
tu
.
e
d
u
.
v
n
1.
I
NT
RO
D
UCT
I
O
N
No
wad
ay
s
,
en
v
ir
o
n
m
e
n
tal
is
s
u
es
an
d
clim
ate
c
h
an
g
e
ar
e
g
ain
in
g
m
u
ch
atten
tio
n
f
r
o
m
i
n
d
iv
id
u
als,
o
r
g
an
izatio
n
s
an
d
g
o
v
e
r
n
m
en
ts
ar
o
u
n
d
th
e
wo
r
ld
.
On
e
o
f
t
h
e
m
o
s
t
im
p
o
r
tan
t
asp
ec
ts
to
p
r
o
tect
th
e
h
u
m
a
n
en
v
ir
o
n
m
en
t
is
th
e
n
ee
d
to
r
e
d
u
ce
g
r
e
en
h
o
u
s
e
g
as
e
m
is
s
io
n
s
to
d
ea
l
with
th
e
p
r
o
b
lem
o
f
g
l
o
b
al
war
m
i
n
g
.
T
h
e
k
ey
,
th
e
m
ain
d
r
iv
in
g
f
o
r
ce
o
f
th
e
ab
o
v
e
task
is
th
e
n
ee
d
t
o
p
r
o
m
o
te
th
e
d
ev
elo
p
m
en
t
o
f
r
en
ewa
b
le
en
er
g
y
,
r
ep
r
esen
ted
b
y
p
h
o
to
v
o
ltaic
p
o
wer
g
en
er
atio
n
[
1
]
.
Nu
m
er
o
u
s
s
tatis
tical
f
in
d
in
g
s
h
av
e
co
n
f
ir
m
ed
th
e
s
ig
n
if
ican
ce
o
f
p
h
o
to
v
o
ltaic
(
PV)
s
y
s
tem
s
an
d
g
r
id
-
co
n
n
ec
t
ed
PV
p
lan
ts
w
o
r
ld
wid
e.
B
y
t
h
e
en
d
o
f
2
0
2
4
,
th
e
cu
m
u
lativ
e
in
s
talled
ca
p
a
city
o
f
s
o
lar
p
o
wer
s
y
s
tem
s
g
lo
b
al
ly
is
ex
p
ec
ted
to
r
ea
c
h
a
p
p
r
o
x
im
ately
1
.
9
T
W
,
u
p
s
h
ar
p
ly
f
r
o
m
1
.
1
7
7
T
W
at
th
e
en
d
o
f
2
0
2
2
.
T
h
is
in
cr
ea
s
e
r
ef
lects
th
e
ex
p
lo
s
iv
e
g
r
o
w
th
o
f
s
o
lar
en
er
g
y
,
esp
ec
ially
in
lar
g
e
m
ar
k
ets
s
u
ch
as
C
h
in
a
an
d
th
e
Un
ited
St
ates
[
2
]
.
I
n
ad
d
itio
n
,
th
e
in
s
tal
latio
n
o
f
PV
p
lan
ts
h
as
d
r
iv
en
th
e
r
a
p
id
in
cr
ea
s
e
in
s
o
lar
ce
ll
d
ep
lo
y
m
e
n
t
g
lo
b
ally
.
Fo
r
ex
am
p
le,
in
th
e
Un
ited
States
,
th
e
cu
m
u
lativ
e
in
s
talled
ca
p
ac
ity
o
f
s
o
lar
p
o
wer
s
y
s
tem
s
is
ex
p
ec
ted
to
r
ea
ch
ap
p
r
o
x
im
ately
1
6
0
GW
b
y
t
h
e
en
d
o
f
2
0
2
4
[
2
]
.
Me
an
w
h
ile,
in
C
h
in
a,
th
e
cu
m
u
lativ
e
in
s
talled
ca
p
ac
ity
o
f
s
o
lar
PV
s
y
s
tem
s
is
f
o
r
ec
ast
to
r
ea
ch
ap
p
r
o
x
im
ately
6
6
2
GW
[
2
]
.
H
o
wev
er
,
th
e
p
o
wer
ef
f
icien
cy
o
f
p
h
o
t
o
v
o
ltaic
ce
lls
is
f
r
eq
u
en
tly
co
n
s
tr
ain
ed
b
y
d
ef
ec
ts
th
at
im
p
ac
t
b
o
th
th
eir
p
er
f
o
r
m
an
ce
an
d
life
s
p
an
.
T
h
ese
d
ef
ec
ts
ca
n
ca
u
s
e
en
er
g
y
l
o
s
s
b
y
f
o
r
m
in
g
n
ew
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
3
,
J
u
n
e
20
25
:
3
3
1
9
-
3
3
3
1
3320
r
ec
o
m
b
in
atio
n
p
at
h
way
s
,
co
n
v
er
tin
g
lig
h
t
in
to
h
ea
t
in
s
tead
o
f
elec
tr
icity
,
o
r
ev
e
n
d
r
ain
in
g
th
e
e
n
er
g
y
s
to
r
ed
in
th
e
b
atter
y
,
r
ed
u
cin
g
th
e
ef
f
icien
cy
o
f
PV m
o
d
u
les
[
3
]
.
Mo
r
eo
v
er
,
em
er
g
i
n
g
g
e
n
er
atio
n
s
o
f
s
o
lar
ce
lls
,
in
clu
d
in
g
co
p
p
er
-
in
d
iu
m
-
g
alliu
m
-
d
is
u
lf
id
e
(
C
I
GS)
ce
lls
an
d
Per
o
v
s
k
ite
ce
lls
(
PS
C
s
)
,
ar
e
f
ac
in
g
ch
allen
g
es
in
i
m
p
r
o
v
i
n
g
en
er
g
y
co
n
v
er
s
io
n
ef
f
icien
cy
,
r
ed
u
cin
g
m
an
u
f
ac
tu
r
in
g
c
o
s
ts
,
an
d
lim
i
tin
g
en
v
ir
o
n
m
e
n
tal
im
p
ac
ts
d
u
e
to
th
e
u
s
e
o
f
to
x
ic
m
ater
ial
s
[
4
]
,
[
5
]
.
R
ec
en
tly
,
th
e
r
esear
ch
an
d
m
an
u
f
ac
tu
r
i
n
g
in
d
u
s
tr
y
h
as
p
aid
s
p
ec
ial
atten
tio
n
to
PS
C
s
tech
n
o
lo
g
y
d
u
e
to
its
s
im
p
le
f
ab
r
icatio
n
p
r
o
ce
s
s
an
d
h
ig
h
c
o
n
v
er
s
io
n
ef
f
icien
c
y
.
Acc
o
r
d
i
n
g
to
r
ep
o
r
ts
f
r
o
m
KR
I
C
T
an
d
MI
T
,
PS
C
s
h
av
e
ac
h
iev
ed
an
ef
f
icien
c
y
o
f
2
5
.
2
%
[
5
]
.
Ho
wev
er
,
PS
C
s
s
till
s
u
f
f
er
f
r
o
m
m
ajo
r
p
r
o
b
lem
s
o
f
l
o
w
s
tab
ilit
y
,
wh
ich
ar
e
af
f
ec
te
d
b
y
f
ac
to
r
s
s
u
c
h
as
wate
r
in
f
iltra
tio
n
,
te
m
p
er
atu
r
e,
an
d
h
u
m
id
ity
co
n
d
itio
n
s
,
ca
u
s
in
g
ch
e
m
ical
an
d
s
tr
u
ctu
r
al
ch
an
g
es.
T
h
is
ca
n
l
ea
d
to
in
cr
ea
s
ed
co
s
ts
an
d
s
h
o
r
ten
ed
life
s
p
an
.
T
h
u
s
,
ad
v
an
cin
g
f
au
lt
d
etec
tio
n
tech
n
iq
u
es
f
o
r
p
r
e
d
ictiv
e
m
ai
n
ten
an
ce
an
d
co
n
d
itio
n
m
o
n
it
o
r
in
g
o
f
PV
m
o
d
u
les
is
h
ig
h
l
y
im
p
o
r
ta
n
t
[
5
]
,
[
6
]
.
T
h
er
e
a
r
e
th
r
ee
m
et
h
o
d
s
f
o
r
d
etec
tin
g
d
e
f
ec
ts
in
PV
s
y
s
tem
s
,
in
clu
d
in
g
im
a
g
e
p
r
o
ce
s
s
in
g
-
b
ased
m
eth
o
d
s
[
7
]
–
[
1
0
]
,
elec
tr
ical
d
etec
tio
n
m
eth
o
d
s
[
1
1
]
–
[
1
4
]
,
a
n
d
m
ac
h
in
e
lear
n
in
g
-
b
ased
m
eth
o
d
s
[
1
5
]
–
[
2
4
]
[
2
5
]
–
[
3
4
]
[
3
5
]
.
Alth
o
u
g
h
im
ag
e
p
r
o
ce
s
s
in
g
an
d
elec
tr
ical
d
etec
tio
n
m
e
th
o
d
s
h
av
e
ac
h
iev
ed
q
u
ite
g
o
o
d
r
esu
lts
,
th
ey
lack
th
e
ab
ilit
y
to
ad
a
p
t
to
en
v
i
r
o
n
m
en
tal
ch
a
n
g
es,
a
r
e
h
ig
h
ly
d
ep
en
d
e
n
t
o
n
im
a
g
e
q
u
ality
,
an
d
r
e
q
u
ir
e
h
u
m
an
in
ter
v
en
tio
n
to
ad
ju
s
t
th
e
alg
o
r
ith
m
.
R
ec
en
tly
,
m
ac
h
i
n
e
le
ar
n
in
g
h
as
em
e
r
g
ed
as
a
p
o
t
en
tial
an
d
ef
f
ec
tiv
e
s
o
lu
tio
n
,
p
lay
in
g
an
im
p
o
r
tan
t
r
o
le
in
th
e
d
ev
elo
p
m
en
t
o
f
th
e
PV
in
d
u
s
tr
y
.
Dee
p
lear
n
in
g
tech
n
iq
u
es
h
av
e
th
e
p
o
ten
tial
to
g
r
ea
tly
e
n
h
an
ce
d
etec
tio
n
ef
f
icien
c
y
,
o
p
tim
ize
t
h
e
in
s
p
ec
tio
n
p
r
o
ce
s
s
at
PV
p
o
wer
p
lan
ts
,
a
n
d
ai
d
in
th
e
o
p
er
atio
n
an
d
m
ain
ten
a
n
ce
p
r
o
ce
s
s
es.
R
ec
en
t
s
tu
d
ies
o
n
PV
d
e
f
ec
t
d
etec
tio
n
u
s
in
g
m
ac
h
in
e
lear
n
in
g
tech
n
iq
u
es
ar
e
lis
ted
in
T
ab
le
1
.
B
ec
au
s
e
p
h
o
to
v
o
ltaic
s
y
s
tem
s
'
ef
f
ec
tiv
en
ess
an
d
p
er
f
o
r
m
an
ce
ar
e
in
f
lu
en
ce
d
b
y
v
ar
io
u
s
f
ac
to
r
s
,
m
an
y
s
p
ec
if
ic
p
r
o
b
lem
s
ar
e
ch
allen
g
in
g
to
r
eso
lv
e.
Ho
wev
er
,
m
ac
h
in
e
l
ea
r
n
in
g
tech
n
i
q
u
es
ca
n
s
o
lv
e
th
ese
ch
allen
g
es v
e
r
y
well,
m
ak
in
g
th
em
p
o
p
u
lar
in
d
ef
ec
t d
etec
tio
n
m
eth
o
d
s
.
T
ab
le
1
.
Ma
ch
i
n
e
lear
n
in
g
m
o
d
els f
o
r
d
etec
tin
g
f
au
lts
in
p
h
o
to
v
o
ltaic
s
y
s
tem
s
an
d
th
eir
o
u
t
co
m
es
R
e
f
.
D
e
t
e
c
t
i
o
n
t
a
r
g
e
t
M
a
c
h
i
n
e
l
e
a
r
n
i
n
g
-
met
h
o
d
R
e
s
u
l
t
s
[
1
5
]
D
e
f
e
c
t
s
i
n
P
V
c
e
l
l
S
u
p
p
o
r
t
v
e
c
t
o
r
ma
c
h
i
n
e
(
S
V
M
)
,
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
(
C
N
N
)
A
c
c
u
r
a
c
y
(
C
N
N
)
:
8
8
.
4
2
%
;
a
c
c
u
r
a
c
y
(
S
V
M
)
:
8
2
.
44%
[
1
6
]
Ti
n
y
c
r
a
c
k
s
a
n
d
d
a
r
k
s
p
o
t
s
,
R
a
n
d
o
mN
e
t
5
0
A
c
c
u
r
a
c
y
(
R
a
n
d
o
mN
e
t
5
0
)
:
8
8
.
2
3
%,
[
1
7
]
D
e
f
e
c
t
s
i
n
P
V
C
N
N
Th
e
B
E
R
i
s
7
.
7
3
%
f
o
r
t
h
e
b
i
n
a
r
y
c
l
a
s
si
f
i
c
a
t
i
o
n
p
r
o
b
l
e
m.
[
1
8
]
C
r
a
c
k
s,
f
i
n
g
e
r
f
a
i
l
u
r
e
s
S
V
M
,
r
a
n
d
o
m
f
o
r
e
st
(
R
F
)
A
c
c
u
r
a
c
y
(
S
V
M
)
:
9
9
.
7
%;
A
c
c
u
r
a
c
y
(
R
F
)
:
9
6
.
7%
[
1
9
]
D
e
t
e
c
t
d
e
f
e
c
t
i
v
e
p
a
n
e
l
s
D
e
e
p
La
b
V
3
+
,
F
P
N
a
n
d
U
-
N
e
t
A
c
c
u
r
a
c
y
(
U
-
N
e
t
)
:
0
.
9
4
;
A
c
c
u
r
a
c
y
(
F
P
N
)
:
0
.
9
2
;
A
c
c
u
r
a
c
y
(
D
e
e
p
L
a
b
V
3
+
)
:
0
.
8
7
[
2
0
]
D
e
f
e
c
t
s
i
n
P
V
G
A
N
,
V
G
G
1
6
A
c
c
u
r
a
c
y
(
G
A
N
)
:
0
.
9
4
5
;
A
c
c
u
r
a
c
y
(
V
G
G
1
6
)
:
0
.
9
6
[
2
1
]
P
h
o
t
o
v
o
l
t
a
i
c
c
e
l
l
d
e
f
e
c
t
s
Li
g
h
t
C
N
N
93
.
0
2
%
mA
P
[
2
2
]
F
a
i
l
u
r
e
s
i
n
P
V
m
o
d
u
l
e
s
S
V
M
94
.
4
%
mA
P
[
2
3
]
C
r
a
c
k
s,
o
x
y
g
e
n
-
r
e
l
a
t
e
d
d
e
f
e
c
t
s
,
d
e
f
e
c
t
s
w
i
t
h
i
n
c
e
l
l
s
,
a
n
d
s
o
l
d
e
r
d
i
s
c
o
n
n
e
c
t
i
o
n
s.
R
e
sN
e
t
m
o
d
e
l
s
;
Y
O
LO
F
1
sc
o
r
e
s
o
f
0
.
8
3
(
R
e
sN
e
t
1
8
)
a
n
d
0
.
7
8
(
Y
O
LO
)
[
2
4
]
D
a
mag
e
d
g
a
t
e
,
c
o
n
c
e
a
l
e
d
c
r
a
c
k
,
s
u
r
f
a
c
e
s
c
r
a
t
c
h
,
a
n
d
h
o
t
_
sp
o
t
.
Li
n
e
a
r
a
n
d
q
u
a
d
r
a
t
i
c
d
i
s
c
r
i
mi
n
a
n
t
a
n
a
l
y
si
s
(
LD
A
a
n
d
QDA
)
Th
e
Q
D
A
a
l
g
o
r
i
t
h
m
p
e
r
f
o
r
ms
b
e
t
t
e
r
t
h
a
n
LD
A
i
n
t
h
e
S
N
R
i
n
d
e
x
,
e
n
a
b
l
i
n
g
i
t
t
o
e
f
f
i
c
i
e
n
t
l
y
d
e
t
e
c
t
v
a
r
i
o
u
s
d
e
f
e
c
t
s
i
n
P
V
c
e
l
l
s
[
2
5
]
C
r
a
c
k
e
d
,
a
n
d
h
e
a
v
i
l
y
b
u
s
b
a
r
-
c
o
r
r
o
d
e
d
S
V
M
,
R
F
,
a
n
d
A
N
N
A
c
c
u
r
a
c
y
(
S
V
M
)
:
9
8
.
7
7
%
;
A
c
c
u
r
a
c
y
(
R
F
)
:
9
6
.
6
%;
A
c
c
u
r
a
c
y
(
A
N
N
)
:
9
8
.
1
3
%
;
[
2
6
]
C
r
a
c
k
e
d
,
a
n
d
c
o
r
r
o
d
e
d
S
V
M
,
R
F
,
a
n
d
C
N
N
A
c
c
u
r
a
c
y
(
S
V
M
)
:
9
9
.
4
3
%
;
A
c
c
u
r
a
c
y
(
R
F
)
:
9
7
.
4
6
%
;
A
c
c
u
r
a
c
y
(
C
N
N
)
:
9
9
.
7
1
%
;
[
2
7
]
W
i
t
h
o
u
t
d
e
f
e
c
t
s
,
f
i
n
g
e
r
i
n
t
e
r
r
u
p
t
i
o
n
,
m
i
c
r
o
-
c
r
a
c
k
,
f
r
a
c
t
u
r
e
P
r
o
p
o
se
d
C
N
N
8
3
%
mA
P
[
2
8
]
C
r
a
c
k
s,
i
n
t
e
r
c
o
n
n
e
c
t
f
a
i
l
u
r
e
s
o
f
c
e
l
l
s,
i
n
t
e
r
r
u
p
t
i
o
n
s
i
n
c
o
n
t
a
c
t
,
a
n
d
c
o
r
r
o
s
i
o
n
o
f
c
o
n
t
a
c
t
s.
D
e
e
p
l
a
b
v
3
9
5
.
4
%
mA
P
[
2
9
]
C
r
a
c
k
s (l
i
n
e
a
r
a
n
d
st
e
l
l
a
t
e
)
,
b
r
o
k
e
n
g
r
i
d
s,
b
l
a
c
k
c
o
r
e
s,
u
n
a
l
i
g
n
e
d
,
t
h
i
c
k
l
i
n
e
s
R
e
sN
e
t
1
5
2
–
X
c
e
p
t
i
o
n
9
6
.
1
7
%
mA
P
[
3
0
]
P
V
f
a
u
l
t
d
e
t
e
c
t
i
o
n
Y
O
LO
v
2
a
n
d
Y
O
LO
v
3
Y
O
LO
v
2
a
c
h
i
e
v
e
s
a
n
F
1
sc
o
r
e
o
f
8
9
%,
w
h
i
l
e
Y
O
LO
v
3
r
e
a
c
h
e
s
9
1
%
[
3
1
]
F
a
u
l
t
y
a
r
e
a
o
r
h
o
t
_
sp
o
t
o
f
t
h
e
P
V
mo
d
u
l
e
.
Y
O
LO
v
3
3
4
%
mA
P
[
3
2
]
P
V
f
a
u
l
t
d
e
t
e
c
t
i
o
n
Y
O
LO
v
4
,
Y
O
LO
v
4
-
t
i
n
y
Y
O
LO
v
4
h
a
s a
mA
P
o
f
9
8
.
8
%,
w
h
e
r
e
a
s YO
LO
v
4
-
t
i
n
y
h
a
s
a
mA
P
o
f
9
1
.
0
%.
[
3
3
]
P
V
mo
d
u
l
e
s
w
i
t
h
c
r
a
c
k
s
a
n
d
f
r
a
g
me
n
t
s
Y
O
LO
v
5
9
2
.
3
%
mA
P
[
3
4
]
H
o
t
_
sp
o
t
f
a
u
l
t
d
e
t
e
c
t
i
o
n
Y
O
LO
v
5
9
8
.
1
%
mA
P
[
3
5
]
P
V
p
a
n
e
l
d
e
f
e
c
t
d
e
t
e
c
t
i
o
n
Y
O
LO
v
5
9
7
.
8
%
mA
P
[
3
6
]
P
V
f
a
u
l
t
d
e
t
e
c
t
i
o
n
Y
O
LO
v
8
9
4
%
mA
P
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
I
mp
r
o
ve
d
YOLOv1
0
mo
d
el
fo
r
d
etec
tin
g
s
u
r
fa
ce
d
efec
ts
o
n
s
o
la
r
…
(
P
h
a
t T.
N
g
u
ye
n
)
3321
T
o
en
h
an
ce
th
e
q
u
ality
an
d
ef
f
icien
cy
o
f
d
etec
tin
g
d
e
f
ec
ts
in
p
h
o
to
v
o
ltaic
s
y
s
tem
s
an
d
s
u
p
p
o
r
t
th
e
s
u
s
tain
ab
le
ad
v
an
ce
m
e
n
t
o
f
t
h
e
p
h
o
to
v
o
ltaic
s
ec
to
r
a
n
d
n
e
w
en
er
g
y
a
p
p
licatio
n
s
,
a
u
to
m
atin
g
th
e
in
tellig
en
t
d
ef
ec
t
d
etec
tio
n
p
r
o
ce
s
s
th
r
o
u
g
h
ad
v
a
n
ce
d
c
o
m
p
u
te
r
tech
n
o
lo
g
y
is
an
ess
en
tial
tech
n
ical
s
o
lu
tio
n
.
As
im
ag
e
an
aly
s
is
an
d
d
ee
p
lear
n
in
g
tech
n
o
lo
g
ies
co
n
tin
u
e
to
ad
v
an
ce
,
th
e
in
teg
r
atio
n
o
f
co
m
p
u
te
r
v
is
io
n
tech
n
o
lo
g
y
with
s
u
r
f
ac
e
d
ef
ec
t
d
etec
tio
n
i
s
b
ec
o
m
in
g
in
c
r
ea
s
in
g
ly
p
r
ev
alen
t.
C
u
r
r
en
tly
,
m
ac
h
in
e
lear
n
in
g
tec
h
n
iq
u
es
f
o
r
d
etec
tin
g
d
ef
ec
ts
in
PV
ce
lls
ar
e
p
r
im
a
r
ily
ca
teg
o
r
ized
in
t
o
two
ty
p
es.
T
h
e
f
ir
s
t
ty
p
e
in
clu
d
es
tr
ad
itio
n
a
l
m
ac
h
in
e
lear
n
in
g
m
eth
o
d
s
,
wh
ich
d
ep
e
n
d
o
n
m
an
u
ally
d
esig
n
ed
f
ea
tu
r
e
ex
tr
ac
to
r
s
to
estab
lis
h
co
m
p
lex
r
ec
o
g
n
itio
n
r
elatio
n
s
h
ip
s
b
u
t
o
f
ten
f
ac
e
lim
itatio
n
s
in
ter
m
s
o
f
g
en
e
r
aliza
tio
n
an
d
r
o
b
u
s
tn
ess
.
T
h
e
s
ec
o
n
d
ca
teg
o
r
y
in
clu
d
es
alg
o
r
ith
m
s
s
u
ch
as
YOL
O,
r
eg
io
n
-
C
NN
(
R
-
C
NN)
,
Mo
b
ileNet,
I
n
ce
p
tio
n
V3
,
VGG1
6
,
an
d
R
esNet5
0
,
wh
ich
u
s
e
d
ee
p
lear
n
in
g
tech
n
iq
u
es
to
lear
n
f
r
o
m
a
lar
g
e
n
u
m
b
er
o
f
s
am
p
l
es,
p
r
o
v
id
in
g
b
etter
g
en
er
aliza
tio
n
an
d
r
o
b
u
s
tn
ess
[
1
9
]
–
[
2
1
]
,
[
2
7
]
–
[
3
3
]
[
3
4
]
–
[
3
6
]
.
Am
o
n
g
t
h
e
r
ec
en
tly
wid
ely
u
s
ed
d
ee
p
lear
n
i
n
g
m
o
d
els
with
h
ig
h
d
etec
tio
n
s
p
ee
d
an
d
ac
c
u
r
ac
y
,
th
e
YOL
O
m
o
d
el,
f
ir
s
t
in
tr
o
d
u
ce
d
in
2
0
1
6
[
3
7
]
,
s
tan
d
s
o
u
t.
T
o
o
v
e
r
co
m
e
t
h
e
lo
w
d
etec
tio
n
ac
cu
r
ac
y
o
f
th
e
v
1
m
o
d
el
,
YOL
Ov
2
in
tr
o
d
u
ce
d
an
c
h
o
r
b
o
x
es
an
d
b
atch
n
o
r
m
aliza
tio
n
[
3
7
]
.
B
u
ild
in
g
o
n
YOL
Ov
2
,
YOL
Ov
3
[
3
7
]
f
ea
tu
r
ed
a
n
ew
Dar
k
n
et
ar
c
h
itectu
r
e
with
5
3
lay
er
s
,
ad
v
an
cin
g
b
e
y
o
n
d
f
ea
tu
r
e
p
y
r
am
id
n
etwo
r
k
s
to
im
p
r
o
v
e
m
u
lti
-
s
ca
le
f
u
s
io
n
p
r
e
d
ictio
n
an
d
ac
cu
r
ac
y
f
o
r
d
etec
tin
g
s
m
all
an
d
c
o
m
p
lex
o
b
jects.
YOL
Ov
4
[
3
7
]
o
f
f
er
e
d
a
s
im
p
ler
tar
g
et
d
etec
tio
n
m
o
d
el,
r
ed
u
cin
g
th
e
tr
ain
in
g
th
r
esh
o
ld
f
o
r
th
e
alg
o
r
ith
m
.
YOL
Ov
5
d
ev
el
o
p
ed
f
iv
e
v
ar
ian
ts
(
N/S/M/L/X)
b
ased
o
n
v
ar
y
i
n
g
ch
a
n
n
el
r
atio
s
an
d
m
o
d
el
s
izes
[
3
7
]
.
I
n
late
2
0
2
2
,
YOL
Ov
6
a
n
d
YOL
Ov
7
wer
e
r
elea
s
ed
alm
o
s
t
s
im
u
ltan
eo
u
s
ly
,
with
YOL
Ov
6
in
teg
r
atin
g
th
e
R
ep
VGG
s
tr
u
ctu
r
e
to
o
p
tim
i
ze
GPU
p
er
f
o
r
m
a
n
ce
an
d
s
im
p
lify
tech
n
ical
im
p
lem
en
tatio
n
[
3
7
]
.
I
n
J
an
u
a
r
y
2
0
2
3
,
Ultr
aly
tics
in
tr
o
d
u
ce
d
YOL
Ov
8
,
w
h
ich
m
a
d
e
a
m
a
jo
r
b
r
ea
k
th
r
o
u
g
h
in
co
m
p
u
ter
v
is
io
n
[
3
7
]
.
YOL
O
v
8
d
if
f
e
r
s
f
r
o
m
tr
ad
itio
n
al
an
ch
o
r
b
o
x
-
b
ased
m
eth
o
d
s
b
y
a
d
o
p
tin
g
a
n
an
c
h
o
r
box
-
f
r
ee
m
eth
o
d
b
y
d
ir
ec
tly
p
r
ed
ictin
g
th
e
ce
n
ter
o
f
th
e
o
b
ject.
YOL
Ov
9
,
r
elea
s
ed
in
Feb
r
u
ar
y
2
0
2
4
[
3
7
]
,
in
tr
o
d
u
ce
d
two
m
ajo
r
im
p
r
o
v
em
en
ts
:
th
e
p
r
o
g
r
am
m
ab
le
g
r
ad
ien
t
in
f
o
r
m
atio
n
(
PGI
)
f
r
am
ewo
r
k
an
d
th
e
g
en
er
alize
d
ef
f
icien
t
lay
er
p
o
o
lin
g
n
etwo
r
k
(
GE
L
AN)
.
YOL
Ov
1
0
[
3
7
]
,
r
elea
s
ed
in
J
u
n
e
2
0
2
4
,
d
e
m
o
n
s
tr
ate
d
s
u
p
er
io
r
p
er
f
o
r
m
an
ce
o
v
er
its
p
r
ed
ec
ess
o
r
s
.
Sev
er
al
ca
s
e
s
tu
d
ies
[
3
0
–
3
6
]
o
n
PV
f
au
lt
d
ete
ctio
n
u
s
in
g
v
a
r
ian
ts
o
f
YOL
O
h
av
e
d
em
o
n
s
tr
ated
t
h
e
r
ich
n
ess
o
f
m
o
d
els
an
d
m
et
h
o
d
s
ap
p
lie
d
to
th
is
task
.
No
ta
b
ly
,
h
ig
h
ac
cu
r
ac
y
is
m
ain
tain
ed
ac
r
o
s
s
a
v
ar
iety
o
f
d
atasets
an
d
tar
g
ets,
with
m
etr
ics
s
u
ch
as
m
A
P,
AP,
an
d
F1
.
Var
ian
ts
lik
e
YOL
Ov
3
,
YOL
Ov
4
,
an
d
YO
L
Ov
5
o
f
ten
ac
h
iev
e
d
etec
tio
n
ac
cu
r
ac
y
ex
ce
e
d
in
g
9
8
% in
ce
r
tain
ca
s
es.
T
h
e
r
o
b
u
s
tn
ess
an
d
f
lex
ib
ilit
y
o
f
YOL
O
m
o
d
els
in
d
etec
tin
g
f
au
lts
in
PV
s
y
s
tem
s
ar
e
h
ig
h
lig
h
ted
b
y
th
ese
r
esu
lts
.
T
h
er
ef
o
r
e,
YOL
Ov
1
0
was
im
p
r
o
v
ed
an
d
u
s
ed
as
th
e
m
ain
n
etwo
r
k
in
th
is
s
t
u
d
y
,
wh
ich
en
a
b
les
f
ast
p
r
o
ce
s
s
in
g
wh
ile
en
s
u
r
in
g
h
ig
h
ac
c
u
r
ac
y
.
T
h
e
p
ap
e
r
a
ls
o
in
tr
o
d
u
ce
s
th
e
E
MA
atten
tio
n
m
o
d
u
le,
wh
ich
allo
ws
th
e
m
o
d
el
to
e
x
tr
ac
t
f
ea
tu
r
es
ef
f
icien
tly
with
o
u
t
r
e
d
u
cin
g
th
e
d
etec
tio
n
s
p
ee
d
.
Y
OL
Ov
1
0
is
u
s
ed
to
d
etec
t
v
ar
io
u
s
ty
p
es
o
f
d
e
f
ec
ts
o
n
th
e
s
u
r
f
ac
e
o
f
p
h
o
to
v
o
ltaic
p
an
els,
in
clu
d
in
g
b
r
o
k
e
n
,
h
o
t
_
s
p
o
t,
b
lack
_
b
o
r
d
er
,
s
cr
atc,
an
d
n
o
e
lectr
icity
.
Ad
d
itio
n
ally
,
th
is
p
ap
er
in
clu
d
es
a
co
m
p
ar
is
o
n
o
f
YOL
Ov
1
0
with
its
ea
r
lier
v
er
s
io
n
s
.
T
h
e
p
ap
er
'
s
m
ain
co
n
tr
ib
u
tio
n
s
ca
n
b
e
s
u
m
m
ar
ized
as f
o
llo
ws:
a.
I
m
p
r
o
v
e
th
e
n
ec
k
p
ar
t
o
f
YO
L
Ov
1
0
b
y
in
teg
r
atin
g
t
h
e
E
M
A
atten
tio
n
m
ec
h
a
n
is
m
,
to
en
h
an
ce
th
e
a
b
ilit
y
to
ca
p
tu
r
e
th
e
tar
g
et
f
ea
t
u
r
es.
b.
R
ep
lace
th
e
h
ea
d
p
ar
t
o
f
YOL
Ov
1
0
with
th
e
h
ea
d
p
ar
t o
f
Y
OL
Ov
9
,
b
ec
a
u
s
e
we
f
o
u
n
d
th
a
t
r
em
o
v
i
n
g
non
-
m
ax
im
u
m
s
u
p
p
r
ess
io
n
(
NM
S)
in
YOL
Ov
1
0
d
o
es
n
o
t
g
iv
e
g
o
o
d
r
esu
lts
wh
en
d
ea
lin
g
with
m
an
y
ty
p
es
o
f
tar
g
ets with
d
if
f
er
en
t
f
ea
tu
r
es,
s
u
ch
as o
n
th
e
s
u
r
f
ac
e
o
f
PV p
an
els,
f
iv
e
ty
p
es o
f
d
e
f
ec
ts
.
c.
Pro
p
o
s
e
u
s
in
g
th
e
cy
cle
-
GA
N
n
etwo
r
k
f
o
r
d
ata
au
g
m
en
ta
tio
n
,
alth
o
u
g
h
th
e
tr
ain
in
g
ti
m
e
in
cr
ea
s
es,
th
e
ac
cu
r
ac
y
o
f
th
e
m
o
d
el
h
as
b
ee
n
s
ig
n
if
ican
tly
im
p
r
o
v
e
d
.
Data
au
g
m
en
tatio
n
b
r
i
n
g
s
ad
v
a
n
ta
g
es
in
d
etec
tin
g
d
ef
ec
ts
o
n
th
e
s
u
r
f
ac
e
o
f
PV p
an
els.
T
h
e
o
r
g
an
izatio
n
o
f
th
e
r
est
o
f
th
e
p
ap
er
is
as
f
o
llo
ws:
s
e
ctio
n
2
p
r
esen
ts
th
e
p
r
o
p
o
s
ed
m
eth
o
d
,
s
ec
tio
n
3
in
clu
d
es
th
e
r
esu
lts
an
d
d
is
c
u
s
s
io
n
,
an
d
s
ec
tio
n
4
co
n
clu
d
es
with
th
e
f
in
d
in
g
s
an
d
f
u
tu
r
e
d
e
v
elo
p
m
e
n
t
d
ir
ec
tio
n
s
.
2.
M
E
T
H
O
D
2
.
1
.
O
rig
ina
l
YO
L
O
v
1
0
Dev
elo
p
ed
b
y
r
esear
ch
er
s
at
T
s
in
g
h
u
a
Un
iv
er
s
ity
an
d
b
ased
o
n
Ultr
aly
tics
Py
th
o
n
,
YOL
Ov
1
0
in
tr
o
d
u
ce
s
a
n
ew
ap
p
r
o
ac
h
to
r
ea
l
-
tim
e
o
b
ject
d
etec
tio
n
,
ai
m
in
g
to
ad
d
r
ess
th
e
lim
itatio
n
s
o
f
a
r
ch
itectu
r
e
an
d
p
o
s
t
-
p
r
o
ce
s
s
in
g
f
o
u
n
d
in
p
r
ev
io
u
s
v
e
r
s
io
n
s
o
f
YOL
O.
T
h
e
ar
c
h
itectu
r
e
o
f
YOL
Ov
1
0
is
d
ep
icted
in
Fig
u
r
es
1
a
n
d
2
.
T
h
e
a
r
ch
ite
ctu
r
e
o
f
YOL
Ov
1
0
in
h
e
r
its
th
e
ad
v
a
n
tag
es
o
f
p
r
ev
io
u
s
YOL
O
m
o
d
els
wh
ile
ad
d
in
g
im
p
o
r
tan
t
im
p
r
o
v
em
e
n
ts
.
T
h
e
s
tr
u
ctu
r
e
o
f
th
e
m
o
d
el
in
clu
d
es
th
e
f
o
llo
win
g
m
ain
co
m
p
o
n
en
ts
:
B
ac
k
b
o
n
e,
s
er
v
in
g
as
th
e
f
ea
tu
r
e
ex
tr
ac
to
r
,
YOL
Ov
1
0
u
s
es
an
im
p
r
o
v
ed
v
er
s
io
n
o
f
cr
o
s
s
s
tag
e
p
ar
tial
n
etwo
r
k
(
C
SP
Net
)
to
en
h
a
n
ce
g
r
ad
ien
t
f
lo
w
an
d
r
ed
u
ce
r
e
d
u
n
d
an
t
ca
lc
u
latio
n
s
;
T
h
e
Nec
k
n
etwo
r
k
,
d
esig
n
e
d
to
ag
g
r
eg
ate
f
ea
tu
r
es
f
r
o
m
m
u
ltip
le
lev
els
an
d
tr
an
s
m
it
th
em
to
th
e
h
ea
d
,
ef
f
ec
tiv
ely
m
er
g
es
m
u
lti
-
s
ca
le
f
ea
tu
r
es
u
s
in
g
p
ath
a
g
g
r
e
g
atio
n
n
etwo
r
k
(
PAN
)
lay
e
r
s
.
T
h
e
h
ea
d
p
a
r
t
o
f
th
e
YOL
Ov
1
0
n
e
two
r
k
co
n
s
is
ts
o
f
a
one
-
to
-
m
an
y
h
ea
d
an
d
a
o
n
e
-
to
-
o
n
e
h
ea
d
.
T
h
e
one
-
to
-
m
an
y
h
ea
d
g
en
e
r
ates
m
u
ltip
le
e
x
p
ec
tatio
n
s
f
o
r
ea
ch
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
3
,
J
u
n
e
20
25
:
3
3
1
9
-
3
3
3
1
3322
o
b
ject
d
u
r
in
g
tr
ai
n
in
g
t
o
p
r
o
v
id
e
r
ich
s
u
p
er
v
is
io
n
s
ig
n
als
an
d
im
p
r
o
v
e
lear
n
in
g
ac
cu
r
a
cy
.
T
h
e
one
-
to
-
o
n
e
h
ea
d
o
n
ly
g
e
n
er
ates th
e
b
est e
x
p
ec
tatio
n
s
f
o
r
ea
ch
o
b
ject
,
eli
m
in
atin
g
th
e
n
ee
d
f
o
r
NM
S.
Fig
u
r
e
1
.
Netwo
r
k
s
tr
u
ctu
r
e
o
f
YOL
Ov
1
0
d
iag
r
am
C
o
n
v
S
p
l
i
t
B
o
t
t
l
e
n
ec
k
S
h
o
r
t
c
u
t
?
C
o
n
v
C
o
n
v
C
o
n
c
a
t
C
o
n
v
.
B
o
t
t
l
e
n
ec
k
S
h
o
r
t
c
u
t
?
n
C
2
f
S
h
o
r
t
c
u
t
=
?,
n
Bo
t
t
l
e
n
e
c
k
S
h
o
r
t
c
u
t
=
T
r
u
e
C
o
n
v
C
o
n
v
Bo
t
t
l
en
ec
k
S
h
o
r
t
c
u
t
=
F
a
l
s
e
S
P
P
F
C
o
n
v
M
a
x
p
o
o
l
2
d
M
a
x
p
o
o
l
2
d
M
a
x
p
o
o
l
2
d
C
o
n
c
a
t
C
o
n
v
C
o
n
v
2
d
B
a
t
c
h
N
o
r
m
2
d
C
o
n
v
S
i
L
U
C
o
n
v
C
o
n
v
2
d
BN
S
C
D
o
w
n
C
o
n
v
S
p
l
i
t
C
IB
C
IB
C
o
n
c
a
t
C
o
n
v
C
2
f
C
I
B
1
X
1
M
HS
A
P
S
A
F
F
N
1
X
1
Sp
l
i
t
Fig
u
r
e
2
.
Diag
r
a
m
o
f
YOL
Ov
1
0
s
n
etwo
r
k
d
etail
s
tr
u
ctu
r
e
2
.
2
.
I
m
pro
v
ed
YO
L
O
v
1
0
YOL
Ov
1
0
is
th
e
latest
v
er
s
io
n
in
th
e
YOL
O
s
er
ies,
m
ar
k
in
g
an
im
p
o
r
tan
t
s
tep
f
o
r
wa
r
d
i
n
r
ea
l
-
tim
e
tar
g
et
d
etec
tio
n
p
r
o
b
lem
s
.
T
h
e
YOL
Ov
1
0
m
o
d
el
is
h
ig
h
l
y
ap
p
r
ec
iated
f
o
r
its
b
alan
ce
b
et
wee
n
ac
cu
r
ac
y
an
d
in
f
er
en
ce
tim
e,
wh
ic
h
is
s
u
itab
le
f
o
r
r
ea
l
-
tim
e
o
b
ject
d
e
tectio
n
r
eq
u
ir
em
e
n
ts
.
Ho
wev
er
,
wh
en
ap
p
ly
in
g
YOL
Ov
1
0
to
s
p
ec
if
ic
p
r
o
b
le
m
s
with
s
p
ec
ial
d
ata,
ad
ju
s
tm
en
ts
n
ee
d
to
b
e
m
ad
e
t
o
m
ak
e
it
m
o
r
e
s
u
itab
le.
T
h
er
ef
o
r
e,
th
is
p
ap
e
r
im
p
r
o
v
e
s
YOL
Ov
1
0
to
s
im
u
ltan
eo
u
s
ly
d
etec
t
5
t
y
p
es
o
f
d
ef
ec
ts
o
n
PV
an
d
p
r
esen
ts
th
e
p
r
o
p
o
s
ed
n
etwo
r
k
s
tr
u
ctu
r
e
in
Fig
u
r
e
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
I
mp
r
o
ve
d
YOLOv1
0
mo
d
el
fo
r
d
etec
tin
g
s
u
r
fa
ce
d
efec
ts
o
n
s
o
la
r
…
(
P
h
a
t T.
N
g
u
ye
n
)
3323
C
o
n
v
C
o
n
v
C
2
f
(
T
r
u
e
)
C
o
n
v
C
2
f
(
T
r
u
e
)
S
C
D
o
w
n
C
2
f
(
T
r
u
e
)
S
C
D
o
w
n
C
2
f
C
IB
S
P
P
F
B
a
c
k
b
o
n
e
P
S
A
C
o
n
c
a
t
U
p
s
a
m
p
l
e
C
2
f
(
F
a
l
s
e
)
C
o
n
c
a
t
U
p
s
a
m
p
l
e
C
o
n
c
a
t
C
2
f
C
IB
C
2
f
C
IB
S
C
D
o
w
n
C
o
n
c
a
t
V
9
D
e
t
e
c
t
M
i
d
l
e
T
arge
t
V
9
D
e
t
e
c
t
L
a
r
g
e
T
a
r
g
e
t
N
e
c
k
He
a
d
C
2
f
(
F
a
l
s
e
)
V
9
D
e
t
e
c
t
S
m
al
l
T
ar
ge
t
C
o
n
v
A
d
d
ed
m
o
d
u
l
e
s
R
et
a
i
n
ed
m
o
d
u
l
es
r
e
p
l
a
c
e
m
o
d
u
l
e
s
E
M
A
E
M
A
E
M
A
Fig
u
r
e
3
.
I
m
p
r
o
v
ed
YOL
Ov
1
0
(
PV
-
YOL
Ov
1
0
)
n
etwo
r
k
s
tr
u
ctu
r
e
d
iag
r
a
m
2
.
2
.
1
.
An im
pro
v
e
m
ent
o
f
t
he
neck
T
h
e
n
ec
k
p
ar
t
o
f
th
e
YOL
O
v
1
0
n
etwo
r
k
h
as
b
ee
n
ex
te
n
d
ed
with
a
m
u
lti
-
s
ca
le
atten
tio
n
(
E
MA
)
m
o
d
u
le
,
as
illu
s
tr
ated
in
Fig
u
r
e
3
[
3
8
]
,
[
3
9
]
.
T
o
im
p
r
o
v
e
t
h
e
ex
p
r
ess
iv
en
ess
o
f
th
e
PV
-
YOL
Ov
1
0
n
etwo
r
k
an
d
co
n
s
tr
u
ct
d
ep
e
n
d
en
cies,
wh
ile
m
in
im
izin
g
th
e
lo
s
s
o
f
im
p
o
r
tan
t
in
f
o
r
m
atio
n
b
et
wee
n
co
n
v
o
lu
tio
n
al
ex
tr
ac
tio
n
s
,
th
e
E
MA
m
o
d
u
le
h
as b
ee
n
in
teg
r
ate
d
in
to
th
e
n
e
ck
p
ar
t o
f
th
e
n
etwo
r
k
.
T
h
e
s
tr
u
ctu
r
e
o
f
th
e
E
MA
is
s
h
o
wn
in
Fig
u
r
e
4
.
E
M
A,
a
p
ar
allel
atten
tio
n
m
ec
h
an
is
m
co
m
m
o
n
ly
ap
p
lied
in
co
m
p
u
ter
v
is
io
n
ap
p
licatio
n
s
,
en
h
a
n
ce
s
m
o
d
el
p
er
f
o
r
m
an
ce
an
d
s
p
ee
d
s
u
p
p
r
o
ce
s
s
in
g
.
I
n
co
n
tr
ast
to
c
o
n
v
en
tio
n
al
C
NNs,
E
MA
em
p
lo
y
s
a
p
ar
allel
ar
c
h
itectu
r
e
f
o
r
ef
f
icien
t
in
p
u
t
d
ata
p
r
o
ce
s
s
in
g
.
D
u
e
to
t
h
e
p
ar
allel
co
n
v
o
lu
tio
n
f
ea
tu
r
e,
th
e
m
o
d
el
tr
ain
i
n
g
p
r
o
ce
s
s
b
ec
o
m
es
f
aster
wh
en
w
o
r
k
in
g
with
lar
g
e
d
ata,
wh
ile
im
p
r
o
v
in
g
ac
cu
r
ac
y
b
y
s
im
u
ltan
eo
u
s
ly
p
r
o
ce
s
s
in
g
f
ea
tu
r
es
at
d
if
f
er
en
t
s
ca
les.
T
h
e
k
ey
to
th
e
ef
f
ec
tiv
e
n
ess
o
f
E
MA
lies
in
th
e
s
y
n
er
g
is
tic
co
m
b
in
atio
n
o
f
3
×
3
m
ask
an
d
1
×
1
b
r
an
c
h
.
Str
ateg
ically
d
ep
lo
y
in
g
th
is
co
m
b
in
atio
n
allo
ws
f
o
r
s
y
n
th
esizin
g
m
u
lti
-
s
ca
le
s
p
atial
in
f
o
r
m
atio
n
,
d
eliv
er
i
n
g
a
m
ec
h
an
is
m
f
o
r
f
ast
an
d
ef
f
icien
t
f
ee
d
b
ac
k
.
B
y
f
lex
ib
ly
n
a
v
ig
atin
g
t
h
e
co
m
p
lex
en
v
ir
o
n
m
en
t
o
f
f
ea
tu
r
e
e
x
tr
ac
tio
n
,
d
esig
n
i
n
g
en
s
u
r
es
s
tr
o
n
g
a
d
ap
tab
ilit
y
t
o
v
ar
io
u
s
s
p
atial
s
ca
les
in
th
e
d
ata.
C
o
m
p
r
is
in
g
two
m
ain
p
ar
a
llel
b
r
an
ch
es,
th
e
E
MA
ar
c
h
itectu
r
e
in
clu
d
es
o
n
e
b
r
an
ch
f
o
r
p
er
f
o
r
m
i
n
g
g
lo
b
al
clu
s
ter
in
g
in
o
n
e
d
im
e
n
s
io
n
to
en
c
o
d
e
g
lo
b
al
in
f
o
r
m
atio
n
,
wh
ile
th
e
o
th
e
r
b
r
an
ch
is
f
o
r
ex
tr
ac
tin
g
f
ea
tu
r
es
u
s
in
g
3
×
3
co
n
v
o
lu
tio
n
.
M
o
d
u
latin
g
an
d
n
o
r
m
alizin
g
t
h
e
o
u
tp
u
ts
f
r
o
m
b
o
t
h
b
r
an
ch
es
with
a
s
ig
m
o
i
d
f
u
n
c
tio
n
,
in
teg
r
atin
g
th
ese
o
u
t
p
u
ts
th
r
o
u
g
h
a
m
u
lti
-
d
im
en
s
io
n
al
in
ter
ac
tio
n
m
o
d
u
le
ca
p
tu
r
es
p
ix
el
r
elatio
n
s
h
ip
s
.
E
n
h
an
cin
g
o
r
wea
k
en
i
n
g
t
h
e
o
r
ig
in
al
f
ea
tu
r
es
b
y
t
h
e
s
ig
m
o
id
-
m
o
d
u
lated
f
ea
tu
r
e
m
ap
s
lead
s
to
a
m
o
r
e
r
e
f
in
ed
an
d
o
p
tim
ized
r
ep
r
esen
tatio
n
.
C
o
n
s
eq
u
e
n
tly
,
en
c
o
d
in
g
in
ter
-
ch
an
n
el
in
f
o
r
m
atio
n
b
y
E
MA
ad
ju
s
ts
th
e
s
ig
n
if
ican
ce
o
f
d
if
f
e
r
en
t
ch
an
n
els,
wh
ile
m
ain
tain
in
g
th
e
p
r
ec
is
e
s
p
atial
s
tr
u
ctu
r
e
d
etails with
in
th
ese
ch
an
n
els
[
3
9
]
.
2
.
2
.
2
.
An
im
pro
v
e
m
ent
o
f
t
he
hea
d
YOL
Ov
1
0
in
tr
o
d
u
ce
s
a
n
ew
m
eth
o
d
ca
lled
c
o
n
s
is
ten
t
d
u
al
lab
elin
g
f
o
r
NM
S
-
f
r
ee
t
r
ain
in
g
.
T
h
e
NM
S
-
f
r
ee
m
eth
o
d
allo
ws
f
o
r
a
tr
u
ly
en
d
-
to
-
en
d
m
o
d
el
im
p
l
em
en
tatio
n
,
s
im
p
lifie
s
th
e
in
f
er
en
ce
p
ip
elin
e,
an
d
p
o
ten
tially
im
p
r
o
v
es
th
e
o
v
er
a
ll
p
er
f
o
r
m
an
ce
o
f
th
e
s
y
s
tem
.
Ho
wev
er
,
s
in
ce
th
e
g
o
al
o
f
th
i
s
p
ap
er
is
to
d
etec
t
d
ef
ec
ts
with
s
im
ilar
tar
g
et
ch
ar
ac
ter
is
tics
,
s
u
ch
as
b
lack
_
b
o
r
d
er
,
s
cr
atch
,
an
d
b
r
o
k
en
,
wh
ic
h
ar
e
q
u
ite
s
im
ilar
,
ap
p
ly
in
g
th
e
o
n
e
-
to
-
m
a
n
y
an
d
o
n
e
-
to
-
o
n
e
d
etec
tio
n
h
ea
d
s
as
in
YOL
Ov
1
0
will
lik
ely
m
is
-
d
etec
t
s
u
ch
d
ef
ec
ts
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
3
,
J
u
n
e
20
25
:
3
3
1
9
-
3
3
3
1
3324
Fo
r
p
r
o
b
lem
s
th
at
o
n
ly
d
etec
t
o
n
e
ty
p
e
o
f
tar
g
et,
th
e
h
ea
d
o
f
YOL
Ov
1
0
h
as
b
etter
p
er
f
o
r
m
an
ce
th
an
p
r
e
v
io
u
s
v
er
s
io
n
s
,
b
u
t
f
o
r
th
e
p
r
o
b
lem
o
f
class
if
y
in
g
5
ty
p
es
o
f
d
e
f
ec
ts
as
in
th
is
p
ap
er
,
th
e
h
ea
d
o
f
YOL
Ov
1
0
is
r
ep
lace
d
b
y
t
h
e
h
ea
d
o
f
YOL
Ov
9
[
4
0
]
an
d
s
till
u
s
es
N
MS.
Up
o
n
id
en
tify
in
g
an
d
ca
teg
o
r
izin
g
th
e
f
au
lt
y
ar
ea
s
,
NM
S
i
s
em
p
lo
y
ed
to
f
ilter
th
e
d
etec
tio
n
b
o
x
es,
ad
d
r
ess
in
g
th
e
is
s
u
e
o
f
o
v
er
lap
p
in
g
tar
g
et
b
o
x
es.
Dete
r
m
in
in
g
th
e
I
o
U
v
alu
e
b
etwe
en
th
e
d
etec
tio
n
b
o
x
es
with
th
e
g
r
ea
test
co
n
f
id
en
ce
an
d
o
th
er
d
etec
tio
n
b
o
x
es,
as
well
as
ex
am
in
in
g
b
o
x
es
with
I
o
U
v
alu
es
s
u
r
p
ass
in
g
th
e
d
esig
n
ated
th
r
esh
o
ld
,
is
co
n
d
u
cted
.
T
h
is
p
r
o
ce
d
u
r
e
co
n
tin
u
es u
n
til o
n
ly
o
n
e
co
r
r
esp
o
n
d
in
g
d
etec
tio
n
b
o
x
r
e
m
ain
s
f
o
r
ea
ch
o
b
ject.
Fig
u
r
e
4
.
E
MA
m
o
d
u
le
[
3
8
]
2
.
2
.
3
.
Da
t
a
a
ug
m
ent
a
t
io
n us
ing
c
y
cle
-
G
AN
net
wo
rk
T
o
o
v
er
co
m
e
th
e
p
r
o
b
lem
o
f
lim
ited
s
am
p
le
s
ize
f
o
r
d
ee
p
lear
n
in
g
m
o
d
els,
a
c
y
cle
-
G
AN
-
b
ased
s
am
p
le
en
h
a
n
ce
m
en
t
s
tr
ateg
y
was
p
r
o
p
o
s
ed
,
in
wh
ich
th
e
er
r
o
r
r
eg
io
n
s
ar
e
r
a
n
d
o
m
l
y
g
e
n
er
ated
t
o
g
en
er
ate
p
s
eu
d
o
-
s
am
p
les,
wh
ich
h
elp
s
d
iv
er
s
if
y
th
e
ex
is
tin
g
s
am
p
le
s
et.
I
n
th
is
s
tu
d
y
,
c
y
cle
-
GAN
is
ap
p
lied
to
g
en
er
ate
p
s
eu
d
o
-
s
am
p
les
f
r
o
m
th
e
er
r
o
r
r
eg
io
n
s
b
ased
o
n
th
e
o
r
i
g
in
al
d
ata,
i
n
o
r
d
er
t
o
ex
p
an
d
th
e
tr
ai
n
in
g
d
ataset,
d
u
e
to
th
e
s
m
all
s
ize
o
f
th
e
o
r
i
g
in
al
d
ataset
an
d
th
e
im
b
alan
ce
b
etwe
en
th
e
er
r
o
r
class
es.
T
h
e
o
u
ts
tan
d
in
g
ad
v
a
n
tag
e
o
f
c
y
cl
e
-
GAN
is
th
e
ab
ilit
y
t
o
tr
ain
t
h
e
im
ag
e
tr
an
s
f
o
r
m
atio
n
m
o
d
el
with
o
u
t
p
ar
allel
d
ata
p
air
s
[
4
1
]
.
C
y
cle
-
GAN,
an
u
n
s
u
p
er
v
is
ed
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
,
f
ac
ilit
ates
b
id
ir
ec
tio
n
al
tr
an
s
f
o
r
m
atio
n
b
etwe
en
th
e
s
o
u
r
ce
d
o
m
ain
X
an
d
th
e
tar
g
et
d
o
m
ain
Y,
as
d
ep
icted
in
Fig
u
r
e
5
.
I
t
u
s
es
two
g
en
er
ativ
e
n
etwo
r
k
s
G1
an
d
G2
:
G1
tr
a
n
s
f
o
r
m
s
f
r
o
m
X
t
o
Y,
an
d
G2
t
r
an
s
f
o
r
m
s
b
ac
k
f
r
o
m
Y
to
X.
B
o
th
ar
e
co
n
n
ec
te
d
to
d
is
cr
im
in
ativ
e
n
etwo
r
k
s
D1
an
d
D2
af
ter
u
n
d
er
g
o
in
g
a
d
v
e
r
s
ar
ial
tr
ain
in
g
.
T
h
e
n
etwo
r
k
s
G
an
d
D
en
g
ag
e
in
a
co
m
p
etitiv
e
p
r
o
ce
s
s
,
with
D
ac
tin
g
as
a
b
in
ar
y
class
if
ier
attem
p
tin
g
to
d
if
f
er
e
n
tiate
b
etwe
en
r
ea
l
an
d
f
ak
e
im
ag
es,
wh
ile
G
aim
s
to
d
ec
eiv
e
D
b
y
e
n
h
an
ci
n
g
t
h
e
q
u
al
ity
o
f
th
e
g
en
er
ate
d
im
ag
es.
Pro
v
id
in
g
an
im
ag
e
f
r
o
m
th
e
s
o
u
r
ce
d
o
m
ain
as
i
n
p
u
t
to
th
e
g
e
n
er
ativ
e
n
etwo
r
k
G,
it
p
r
o
d
u
ce
s
a
s
y
n
th
etic
im
ag
e
as
o
u
tp
u
t.
R
ec
eiv
in
g
b
o
th
th
e
s
y
n
th
etic
im
ag
e
an
d
a
r
an
d
o
m
im
ag
e
f
r
o
m
th
e
tar
g
et
d
o
m
ain
,
th
e
d
is
cr
im
in
ativ
e
n
etwo
r
k
D
p
r
o
ce
s
s
es
th
em
with
o
u
t
r
eq
u
ir
in
g
p
ai
r
in
g
.
C
y
cle
-
GAN
was
tr
ain
ed
f
o
r
5
0
0
e
p
o
ch
s
with
a
f
ix
ed
lear
n
in
g
r
ate
o
f
0
.
0
0
0
2
.
Af
te
r
th
is
p
r
o
c
ess
,
th
e
s
y
n
th
etic
im
a
g
es
ar
e
g
en
er
ated
b
y
c
y
cle
-
GAN.
I
n
t
h
is
p
ap
er
,
we
cr
ea
te
d
5
0
0
im
a
g
es a
n
d
th
eir
co
r
r
esp
o
n
d
in
g
a
n
n
o
tatio
n
s
to
en
h
a
n
ce
t
h
e
tr
ain
in
g
d
ataset
f
o
r
th
e
PV
-
YOL
Ov
1
0
m
o
d
el.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
I
mp
r
o
ve
d
YOLOv1
0
mo
d
el
fo
r
d
etec
tin
g
s
u
r
fa
ce
d
efec
ts
o
n
s
o
la
r
…
(
P
h
a
t T.
N
g
u
ye
n
)
3325
2
.
2
.
4
.
P
V
def
ec
t
det
ec
t
io
n us
ing
P
V
-
YO
L
O
v
1
0
m
o
del
T
h
e
p
r
o
ce
s
s
o
f
d
etec
tin
g
s
u
r
f
a
ce
d
ef
ec
ts
o
n
p
h
o
to
v
o
ltaic
p
a
n
els
b
ased
o
n
th
e
PV
-
YOL
Ov
1
0
m
o
d
el,
in
clu
d
in
g
th
e
s
tep
s
o
f
d
ata
s
et
co
n
s
tr
u
ctio
n
,
d
ata
au
g
m
e
n
tatio
n
,
m
o
d
el
tr
ai
n
in
g
a
n
d
d
ef
ec
t
d
etec
tio
n
,
is
illu
s
tr
ated
in
Fig
u
r
e
5
.
First,
th
e
d
ata
will
b
e
co
llected
,
p
r
e
-
p
r
o
ce
s
s
ed
,
lab
elled
,
an
d
th
en
d
iv
id
ed
in
to
tr
ai
n
in
g
an
d
test
in
g
d
ata
s
ets.
T
h
e
c
y
cle
-
GAN
n
etwo
r
k
is
u
s
ed
to
au
g
m
en
t
th
e
tr
ain
in
g
d
ata
s
et.
Nex
t,
th
e
tr
ain
in
g
p
ar
am
eter
s
ar
e
s
et
an
d
th
e
d
ee
p
lear
n
in
g
n
etwo
r
k
is
in
itialized
.
Du
r
i
n
g
t
h
is
p
r
o
ce
s
s
,
th
e
w
eig
h
ts
o
f
th
e
m
o
d
el
ar
e
u
p
d
ate
d
th
r
o
u
g
h
ea
ch
i
ter
atio
n
to
en
s
u
r
e
th
e
co
n
v
er
g
en
ce
o
f
th
e
lo
s
s
f
u
n
ctio
n
,
an
d
f
in
ally
th
e
PV
-
YOL
Ov
1
0
m
o
d
el
is
g
e
n
er
ated
to
d
etec
t
d
e
f
ec
ts
o
n
th
e
s
u
r
f
ac
e
o
f
th
e
p
an
els.
T
h
e
d
ef
e
ct
d
etec
tio
n
p
r
o
ce
s
s
is
co
m
p
letely
au
t
o
m
atic,
a
n
d
wh
en
a
n
ew
d
ata
s
et
is
ad
d
ed
,
th
e
m
o
d
el
co
n
f
ig
u
r
atio
n
p
ar
am
eter
s
ca
n
b
e
u
p
d
ated
ac
c
o
r
d
in
g
ly
an
d
r
e
-
t
r
a
in
ed
,
m
ee
tin
g
th
e
ac
tu
al
r
e
q
u
ir
em
en
ts
.
Fig
u
r
e
5
.
T
h
e
PV d
ef
ec
t d
etec
t
io
n
p
r
o
ce
s
s
b
ased
o
n
PV
-
YOL
Ov
1
0
3.
E
XP
E
R
I
M
E
N
T
A
L
P
RE
P
A
RATI
O
N
A
ND
RE
SU
L
T
S
3
.
1
.
Da
t
a
s
et
intr
o
du
ct
io
n
W
e
u
s
ed
th
e
p
u
b
lic
PV
m
u
lti
-
d
ef
ec
t
d
ataset
[
3
5
]
to
v
alid
at
e
th
e
ef
f
ec
tiv
e
n
ess
o
f
o
u
r
m
o
d
el.
T
h
is
d
ataset
in
clu
d
es
f
iv
e
co
m
m
o
n
d
ef
ec
t
ty
p
es:
b
r
o
k
en
,
h
o
t_
s
p
o
t,
b
lack
_
b
o
r
d
er
,
s
cr
atch
,
an
d
n
o
_
elec
tr
icity
.
Fig
u
r
e
6
illu
s
tr
ates
ex
am
p
les
o
f
ea
ch
o
f
th
ese
d
e
f
ec
t
ty
p
es.
I
m
ag
es
f
r
o
m
th
e
PV
m
u
lti
-
d
ef
ec
t
d
ataset
wer
e
p
r
o
ce
s
s
ed
,
with
a
to
tal
o
f
1
,
1
0
8
PV
p
an
el
s
u
r
f
ac
e
d
e
f
ec
t
im
ag
es,
d
iv
id
ed
in
t
o
7
2
.
8
%
f
o
r
th
e
tr
ain
in
g
s
et
an
d
2
7
.
2
%
f
o
r
th
e
test
in
g
s
et.
Sp
ec
if
ically
,
th
er
e
ar
e
8
0
7
im
ag
es
in
th
e
tr
ain
in
g
s
et
an
d
3
0
1
i
m
ag
es
in
th
e
test
in
g
s
et.
W
e
u
s
ed
Py
th
o
n
to
co
n
v
er
t th
e
lab
els f
r
o
m
XM
L
to
T
XT
f
o
r
m
at
an
d
s
er
v
e
f
o
r
m
o
d
el
tr
ain
in
g
.
T
h
er
e
a
r
e
a
to
tal
o
f
4
2
3
5
d
ef
ec
tiv
e
tar
g
e
ts
o
n
th
e
1
,
1
0
8
PV
p
a
n
el
s
u
r
f
ac
e
im
ag
es.
T
h
e
liter
atu
r
e
[
3
5
]
i
n
d
icate
s
th
at
ac
co
u
n
tin
g
f
o
r
th
e
h
ig
h
est
p
r
o
p
o
r
tio
n
o
f
th
e
f
iv
e
d
e
f
ec
t
ty
p
e
s
,
h
o
t
_
s
p
o
ts
co
n
s
titu
te
4
9
.
0
9
%.
Sm
all
s
cr
atch
es,
r
ep
r
esen
tin
g
3
6
.
6
2
%,
ar
e
f
o
llo
wed
b
y
b
lack
-
b
o
r
d
er
tar
g
ets
an
d
b
r
o
k
en
ce
lls
,
wh
ich
ac
co
u
n
t
f
o
r
6
.
0
2
%
an
d
3
.
9
9
%,
r
esp
ec
tiv
ely
.
Fig
u
r
e
6
.
So
m
e
s
am
p
les o
f
5
t
y
p
es o
f
PV su
r
f
ac
e
d
ef
ec
ts
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
3
,
J
u
n
e
20
25
:
3
3
1
9
-
3
3
3
1
3326
3
.
2
.
I
m
plem
ent
a
t
io
n det
a
ils
T
h
e
h
ar
d
war
e
u
s
ed
f
o
r
th
e
ex
p
er
im
en
ts
in
clu
d
e
d
a
GPU
(
R
T
X4
0
7
0
T
i
Su
p
er
1
6
GB
)
,
C
P
U
(
i5
-
1
2
6
0
0
KF)
,
3
2
GB
R
AM
,
alo
n
g
with
Py
T
o
r
c
h
v
er
s
io
n
1
.
1
3
.
1
as
th
e
d
ee
p
lear
n
in
g
p
latf
o
r
m
an
d
Py
t
h
o
n
v
er
s
io
n
3
.
1
0
.
1
3
.
T
h
e
d
ataset
was
tr
ain
ed
o
v
er
3
0
0
ep
o
ch
s
u
s
in
g
th
e
Ad
am
o
p
tim
izatio
n
alg
o
r
ith
m
with
an
in
itial lea
r
n
in
g
r
ate
o
f
0
.
0
1
.
T
h
e
im
p
o
r
ta
n
t p
ar
am
ete
r
s
in
th
e
tr
ain
in
g
p
r
o
ce
s
s
ar
e
d
etailed
i
n
T
ab
le
2
.
T
ab
le
2
.
E
x
p
er
im
en
tal
e
n
v
ir
o
n
m
en
t setu
p
S
TT
P
a
r
a
me
t
e
r
s
S
e
t
u
p
1
Ep
o
c
h
s
3
0
0
2
W
a
r
mu
p
-
e
p
o
c
h
s
3
3
W
a
r
mu
p
-
mo
m
e
n
t
u
m
0
.
8
4
B
a
t
c
h
S
i
z
e
16
5
I
mg
si
z
e
6
4
0
×
6
4
0
6
l
r
f
0
.
0
1
7
l
r
0
0
.
0
1
8
O
p
t
i
mi
z
e
r
A
d
a
m
9
M
o
me
n
t
u
m
0
.
9
3
7
3
.
3
.
E
v
a
lua
t
io
n ind
ica
t
o
rs
T
o
ev
alu
ate
th
e
p
e
r
f
o
r
m
an
ce
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
PV
-
YOL
Ov
1
0
co
m
p
ar
ed
with
t
h
e
n
ewly
in
tr
o
d
u
ce
d
YOL
O
v
e
r
s
io
n
s
(
Y
OL
Ov
1
0
s
,
x
,
n
,
m
,
l)
,
th
is
p
a
p
er
u
s
es
s
tatis
tical
in
d
ices
to
co
m
p
ar
e
th
e
m
o
d
els.
T
h
r
o
u
g
h
o
u
t
th
e
ex
p
er
im
e
n
t,
th
e
ev
alu
atio
n
s
wer
e
p
er
f
o
r
m
ed
b
y
p
er
s
is
ten
tly
tu
n
in
g
t
h
e
p
ar
am
eter
s
.
T
h
e
tr
ain
in
g
an
d
test
in
g
s
ets
o
f
th
e
d
ataset
wer
e
d
iv
id
ed
in
to
s
m
aller
p
ar
ts
to
o
p
tim
ize
th
e
m
o
d
el
tr
ain
in
g
p
r
o
ce
s
s
s
o
th
at
ea
ch
in
d
e
x
co
u
ld
b
e
te
s
ted
in
d
etail.
As
s
h
o
w
n
in
Fig
u
r
e
7
,
t
h
e
r
ec
all
(
1
)
,
p
r
ed
ictio
n
r
ate
(
2
)
,
a
n
d
mA
P
g
r
ap
h
(
4
)
wer
e
v
er
if
ied
.
T
h
e
f
in
d
in
g
s
in
d
icate
d
th
at
th
e
m
AP
cu
r
v
e
p
r
o
g
r
ess
iv
ely
s
tead
ied
f
o
llo
win
g
a
s
p
ec
if
ic
n
u
m
b
er
o
f
tr
ain
in
g
iter
atio
n
s
,
wh
ile
b
o
th
th
e
r
ec
all
r
ate
an
d
p
r
ec
is
io
n
r
ea
c
h
ed
th
e
i
n
ten
d
ed
o
u
tco
m
es.
T
h
e
mA
P
in
d
ex
r
ep
r
esen
ts
th
e
av
er
ag
e
p
r
ec
is
io
n
p
e
r
class
,
wh
ile
AP
(
3
)
is
th
e
a
r
ea
u
n
d
er
th
e
Pr
(
p
r
ec
is
io
n
-
r
ec
all)
c
u
r
v
e,
an
d
p
r
ec
is
io
n
is
u
s
ed
to
ev
alu
ate
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
tar
g
et
d
e
tectio
n
m
o
d
el.
T
h
e
r
ec
all
r
ate,
o
r
all
-
test
r
ate,
r
ep
r
esen
ts
th
e
p
r
o
p
o
r
tio
n
o
f
co
r
r
ec
tly
p
r
ed
icted
class
es
o
u
t
o
f
th
e
to
tal
n
u
m
b
e
r
o
f
tr
u
e
class
es,
in
co
n
tr
ast to
th
e
f
alse d
etec
tio
n
r
ate.
=
+
(
1
)
=
+
(
2
)
wh
er
e
TP
,
FN
,
an
d
FP
r
e
p
r
esen
t
th
e
n
u
m
b
er
o
f
tr
u
e
p
o
s
itiv
es,
f
alse
n
eg
ativ
es,
a
n
d
f
alse
p
o
s
itiv
es,
r
esp
ec
tiv
ely
.
=
∫
(
)
1
0
(
3
)
wh
er
e
an
d
d
en
o
te
th
e
r
ec
all
an
d
p
r
ec
is
io
n
f
o
r
tar
g
et
class
m
an
d
N
d
en
o
tes
th
e
n
u
m
b
er
o
f
tar
g
ets.
T
h
e
av
er
a
g
e
p
r
ec
is
io
n
m
ea
n
is
ca
lcu
lated
as
(
4
)
:
=
1
∑
=
1
(
4
)
wh
er
e
N
s
ig
n
if
ies
th
e
n
u
m
b
er
o
f
ca
teg
o
r
ies
an
d
AP
s
tan
d
s
f
o
r
th
e
a
v
er
ag
e
p
r
ec
is
io
n
o
f
ea
ch
ca
teg
o
r
y
.
I
n
th
e
s
co
p
e
o
f
o
u
r
d
etec
tio
n
task
,
N
=
5
.
3
.
4
.
M
a
in re
s
u
lt
s
T
o
en
s
u
r
e
an
ac
cu
r
ate
r
ep
r
es
en
tatio
n
o
f
th
e
p
r
o
p
o
s
ed
al
g
o
r
ith
m
’
s
d
etec
tio
n
p
e
r
f
o
r
m
an
ce
o
n
th
e
d
ataset
d
etailed
in
s
ec
tio
n
3
.
1
,
in
clu
d
i
n
g
t
h
e
b
ac
k
g
r
o
u
n
d
r
eg
io
n
as
a
ca
teg
o
r
y
in
th
e
c
o
n
f
u
s
io
n
m
atr
ix
,
alo
n
g
with
th
e
f
iv
e
d
ef
ec
ts
,
was
n
ec
ess
ar
y
f
o
r
ev
alu
atin
g
th
e
m
o
d
el’
s
ef
f
ec
tiv
en
ess
.
I
n
th
e
co
n
f
u
s
io
n
m
atr
ix
,
r
ep
r
esen
tin
g
th
e
tr
u
e
class
es
o
f
th
e
o
b
jects
ar
e
th
e
r
o
ws,
wh
ile
in
d
icatin
g
th
e
p
r
ed
ic
ted
class
e
s
ar
e
th
e
co
lu
m
n
s
.
Sin
ce
en
co
m
p
ass
in
g
all
p
ix
els
in
th
e
im
ag
e
th
at
d
o
n
o
t
co
r
r
esp
o
n
d
to
an
y
d
ef
ec
t
is
th
e
b
ac
k
g
r
o
u
n
d
lay
er
,
it
is
n
o
t
a
class
th
e
PV
-
YOL
Ov
1
0
m
o
d
el
is
tr
ain
ed
to
id
en
tify
.
T
h
u
s
,
wh
en
th
e
m
o
d
el
is
ap
p
lied
to
an
im
ag
e,
n
o
d
ef
ec
t sh
o
u
ld
b
e
class
if
ied
as
b
elo
n
g
in
g
to
th
e
b
ac
k
g
r
o
u
n
d
lay
er
.
C
o
n
s
eq
u
en
tly
,
Fig
u
r
e
8
will sh
o
w
a
v
alu
e
o
f
0
i
n
th
e
b
ac
k
g
r
o
u
n
d
-
b
ac
k
g
r
o
u
n
d
co
lu
m
n
o
f
th
e
PV
-
YOL
Ov
1
0
m
o
d
els’
co
n
f
u
s
io
n
m
at
r
ix
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
I
mp
r
o
ve
d
YOLOv1
0
mo
d
el
fo
r
d
etec
tin
g
s
u
r
fa
ce
d
efec
ts
o
n
s
o
la
r
…
(
P
h
a
t T.
N
g
u
ye
n
)
3327
r
ep
r
esen
tin
g
an
in
v
alid
p
r
ed
ictio
n
o
f
th
e
PV
-
YOL
Ov
1
0
m
o
d
el.
T
o
b
etter
u
n
d
e
r
s
tan
d
th
e
v
alu
es
an
d
v
alid
atio
n
r
esu
lts
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
f
o
r
d
etec
tin
g
f
i
v
e
ty
p
es
o
f
PV
s
u
r
f
ac
e
d
e
f
e
cts,
p
lease
r
ef
er
to
T
ab
le
3
.
I
n
T
ab
le
3
,
th
e
n
o
_
e
lectr
icity
s
u
r
f
ac
e
d
ef
ec
t
h
as
t
h
e
h
ig
h
est
m
AP
v
alu
e
o
f
0
.
9
4
5
,
wh
ile
h
o
t_
s
p
o
t,
b
lack
_
b
o
r
d
er
,
b
r
o
k
en
,
an
d
s
cr
atch
h
av
e
m
AP
v
alu
es
o
f
0
.
9
4
4
;
0
.
9
3
2
;
0
.
7
3
8
;
an
d
0
.
6
4
3
,
r
esp
ec
tiv
ely
.
T
h
is
p
r
im
ar
ily
r
esu
lts
f
r
o
m
th
e
f
a
ct
th
at
th
e
n
o
_
elec
tr
icity
d
ef
e
ct
is
lar
g
e
in
s
ize
an
d
th
e
f
ea
tu
r
e
is
d
if
f
icu
lt
to
co
n
f
u
s
e,
wh
ile
h
o
t_
s
p
o
t
h
as
th
e
lar
g
est
n
u
m
b
er
o
f
tr
ain
in
g
s
am
p
les,
ac
co
u
n
tin
g
f
o
r
n
ea
r
ly
h
alf
o
f
th
e
to
tal
s
am
p
les.
I
n
c
o
n
tr
ast,
s
cr
atch
a
n
d
b
r
o
k
e
n
ar
e
s
m
all
in
s
ize
a
n
d
ar
e
ea
s
ily
co
n
f
u
s
ed
with
ea
c
h
o
th
e
r
,
r
esu
ltin
g
in
r
elativ
ely
lo
w
p
r
ec
is
io
n
,
r
ec
all,
an
d
mA
P
v
alu
es o
n
s
cr
atch
a
n
d
b
r
o
k
en
.
Fig
u
r
e
7
.
PV
-
YOL
Ov
1
0
d
etec
tio
n
in
d
ex
ch
ar
t
Fig
u
r
e
8
.
T
h
e
co
n
f
u
s
io
n
m
atr
ix
o
f
PV
-
YOL
Ov
1
0
m
o
d
el
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
15
,
No
.
3
,
J
u
n
e
20
25
:
3
3
1
9
-
3
3
3
1
3328
T
ab
le
3
.
PV
-
YOL
Ov
1
0
d
etec
t
io
n
r
esu
lts
C
a
t
e
g
o
r
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
m
AP
b
r
o
k
e
n
0
.
6
8
7
0
.
7
0
6
0
.
7
3
8
h
o
t
_
sp
o
t
0
.
9
7
4
0
.
8
0
8
0
.
9
4
4
b
l
a
c
k
_
b
o
r
d
e
r
0
.
8
0
9
0
.
9
0
5
0
.
9
3
2
scrat
c
h
0
.
7
0
7
0
.
6
2
9
0
.
6
4
3
n
o
_
e
l
e
c
t
r
i
c
i
t
y
0
.
9
2
7
0
.
9
0
.
9
4
5
3
.
4
.
1
.
Abla
t
io
n e
x
perim
ent
C
o
n
d
u
ctin
g
ab
latio
n
ex
p
er
im
en
ts
f
o
r
a
th
o
r
o
u
g
h
c
o
m
p
ar
i
s
o
n
,
we
aim
to
v
er
if
y
th
e
ef
f
ec
ts
o
f
d
if
f
er
en
t
m
o
d
u
les
o
n
m
o
d
el
p
er
f
o
r
m
a
n
ce
.
Star
tin
g
with
th
e
YOL
Ov
1
0
s
m
o
d
el
as
th
e
b
ase,
we
g
r
ad
u
ally
ad
d
ad
d
itio
n
al
m
o
d
u
les
to
illu
s
tr
at
e
th
e
im
p
o
r
tan
ce
o
f
ea
ch
m
o
d
u
le
an
d
its
in
f
lu
en
ce
o
n
p
er
f
o
r
m
an
ce
.
Sh
o
win
g
th
e
r
esu
lts
o
f
th
ese
ex
p
er
im
en
t
s
,
T
ab
le
4
d
em
o
n
s
tr
ates th
at
ac
h
iev
in
g
th
e
b
est ex
p
e
r
im
en
tal
o
u
tco
m
es r
eq
u
ir
es
u
s
in
g
all
m
o
d
u
les,
in
d
icatin
g
t
h
at
ev
er
y
m
o
d
u
le
is
v
ital to
o
u
r
s
tr
ateg
y
.
I
m
p
r
o
v
in
g
all
m
o
d
el
m
etr
ics,
th
e
ad
d
itio
n
o
f
th
e
E
MA
m
o
d
u
le
r
esu
lts
in
a
0
.
1
%
in
cr
ea
s
e
in
m
AP
@
0.
5
an
d
a
0
.
4
%
r
is
e
in
m
AP
@
0.
5:0
.
95
,
ef
f
ec
tiv
el
y
s
h
o
wca
s
in
g
its
en
h
an
ce
m
en
t
o
f
th
e
m
o
d
el'
s
f
ea
tu
r
e
e
x
tr
ac
tio
n
ca
p
ab
ilit
ies.
Mo
d
if
y
in
g
th
e
h
ea
d
o
f
YOL
Ov
9
r
esu
lted
in
e
n
h
an
ce
m
e
n
ts
o
f
2
.
5
%
an
d
3
.
1
%
in
m
AP
@
0.
5
an
d
m
AP
@
0.
5:0.
95
,
r
esp
ec
tiv
ely
,
co
n
f
ir
m
in
g
th
at
th
is
m
o
d
u
le
d
ir
ec
ts
th
e
n
etwo
r
k
to
ca
p
tu
r
e
m
o
r
e
cr
u
cial
f
ea
tu
r
e
in
f
o
r
m
atio
n
.
I
n
teg
r
atin
g
b
o
th
th
e
h
ea
d
m
o
d
if
icatio
n
an
d
th
e
E
MA
ad
d
itio
n
led
t
o
in
c
r
ea
s
es
o
f
4
.
4
%
an
d
3
.
6
%
in
m
AP
@
0.
5
an
d
m
AP
@
0.
5:0.
95
.
Ultim
ately
,
em
p
lo
y
in
g
d
at
a
au
g
m
en
tatio
n
with
th
e
C
y
cle
-
GAN
n
etwo
r
k
elev
ated
m
AP@
0
.
5
an
d
m
AP
@
0.
5:0.
95
b
y
5
.
1
%
an
d
6
.
5
%.
T
h
e
d
ata
s
h
o
w
n
in
T
ab
le
4
cle
ar
ly
illu
s
tr
ates
th
e
ess
en
tial r
o
le
an
d
ef
f
ec
tiv
e
n
ess
o
f
ea
ch
m
o
d
u
le,
h
ig
h
lig
h
tin
g
th
e
ad
v
a
n
tag
e
o
f
o
u
r
m
eth
o
d
.
T
ab
le
4
.
Ab
latio
n
ex
p
e
r
im
en
t
r
esu
lts
M
o
d
e
l
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
mA
P
@
0.
5
mA
P
@
0.
5
:
0
.
95
Y
O
LO
v
1
0
s
0
.
8
1
5
0
.
6
8
8
0
.
7
8
9
0
.
5
0
6
Y
O
LO
v
1
0
s+
EM
A
0
.
7
6
8
0
.
7
1
7
0
.
7
9
9
0
.
5
1
Y
O
LO
v
1
0
s+H
e
a
d
o
f
Y
O
LO
v
9
0
.
8
0
1
0
.
7
6
4
0
.
8
1
4
0
.
5
3
7
Y
O
LO
v
1
0
s+H
e
a
d
o
f
Y
O
LO
v
9
+
EM
A
0
.
7
9
4
0
.
7
8
7
0
.
8
3
3
0
.
5
4
2
H
e
a
d
o
f
Y
O
LO
v
9
+
E
M
A
+
G
A
N
(
P
V
-
Y
O
LO
v
1
0
)
0
.
8
2
1
0
.
7
9
0
.
8
4
0
.
5
7
1
3
.
4
.
2
.
Co
m
pa
riso
n wit
h sta
t
e
-
of
-
t
he
-
a
rt
m
et
ho
ds
YO
L
O
v
1
0
T
o
s
h
o
wca
s
e
th
e
s
u
p
er
io
r
ity
o
f
th
e
en
h
an
ce
d
PV
-
YOL
Ov
1
0
m
o
d
el,
th
is
s
tu
d
y
ca
r
r
ied
o
u
t
a
s
er
ies
o
f
ev
alu
atio
n
e
x
p
er
im
e
n
ts
,
co
m
p
ar
in
g
its
p
er
f
o
r
m
an
ce
ag
ai
n
s
t
th
e
n
ewly
in
tr
o
d
u
ce
d
YOL
Ov
1
0
m
o
d
els.
Of
f
er
in
g
a
co
n
cise
co
m
p
a
r
is
o
n
o
f
v
ar
io
u
s
YOL
Ov
1
0
s
m
o
d
el
v
er
s
io
n
s
,
T
ab
le
5
h
ig
h
lig
h
t
s
th
e
ar
ch
itectu
r
al
d
iv
er
s
ity
am
o
n
g
th
e
YOL
Ov
1
0
s
v
ar
ia
n
ts
.
T
h
ese
v
ar
ian
ts
,
em
p
h
asizin
g
d
if
f
er
en
t
ar
ch
it
ec
tu
r
al
ap
p
r
o
ac
h
es,
d
em
o
n
s
tr
ate
d
is
tin
ct
tr
ad
e
-
o
f
f
s
b
etwe
en
m
AP
0.
5
;
m
AP
0.
5
-
0.
9
5
;
Pre
cisi
o
n
,
r
ec
all
a
n
d
GFL
OP.
T
h
e
p
r
o
p
o
s
ed
PVDF
-
YOL
Ov
1
0
m
o
d
el
d
em
o
n
s
tr
ates
th
e
h
ig
h
est
p
er
f
o
r
m
an
ce
in
d
etec
tin
g
f
iv
e
t
y
p
es
o
f
s
u
r
f
ac
e
d
e
f
ec
ts
o
n
PV
with
m
AP
0.
5
;
m
AP
0.
5
-
0.
95
;
Pre
cisi
o
n
,
r
ec
all
c
o
m
p
a
r
ed
t
o
th
e
latest
YOL
Ov
1
0
v
er
s
io
n
s
.
H
o
wev
er
,
t
h
er
e
is
a
co
m
p
u
tatio
n
al
tr
ad
e
-
o
f
f
as
th
e
GFLO
P
v
alu
e
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
el
is
th
e
lar
g
est
an
d
th
e
m
o
s
t
co
m
p
u
tatio
n
ally
ad
v
a
n
tag
eo
u
s
is
th
e
YOL
Ov
1
0
n
m
o
d
el
with
th
e
s
m
allest GFL
OP.
T
ab
le
5
.
R
esu
lts
o
f
p
r
o
p
o
s
ed
m
o
d
els PV
-
YOL
Ov
1
0
an
d
YOL
Ov
1
0
v
er
s
io
n
s
M
o
d
e
l
mA
P
@
0.
5
mA
P
@
0.
5
-
0.
9
5
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
G
F
LO
P
Y
O
LO
v
1
0
s
0
.
7
8
9
0
.
5
0
6
0
.
8
1
5
0
.
6
8
8
2
4
.
5
Y
O
LO
v
1
0
m
0
.
7
7
7
0
.
5
1
0
.
7
5
6
0
.
7
1
9
6
3
.
4
Y
O
LO
v
1
0
n
0
.
7
7
5
0
.
5
1
3
0
.
7
4
8
0
.
7
0
1
8
.
2
Y
O
LO
v
1
0
b
0
.
8
0
3
0
.
5
1
8
0
.
7
9
6
0
.
7
4
98
Y
O
LO
v
1
0
l
0
.
8
1
8
0
.
5
3
3
0
.
8
7
8
0
.
7
2
9
1
2
6
.
4
Y
O
LO
v
1
0
x
0
.
8
2
1
0
.
5
4
2
0
.
8
3
3
0
.
7
6
3
1
6
9
.
8
PV
-
Y
O
LO
v
1
0
0
.
8
4
0
.
5
7
1
0
.
8
2
1
0
.
7
9
1
7
6
.
8
3
.
4
.
3
.
Q
ua
lita
t
iv
e
co
m
pa
riso
n
T
o
f
u
r
t
h
er
clar
if
y
th
e
e
f
f
ec
tiv
e
n
ess
o
f
th
e
p
r
o
p
o
s
ed
m
eth
o
d
,
we
s
elec
ted
s
o
m
e
r
ep
r
esen
tati
v
e
im
ag
es
to
co
m
p
a
r
e
with
o
th
er
s
tate
-
of
-
th
e
-
ar
t
m
eth
o
d
s
.
Fig
u
r
e
8
ill
u
s
tr
ates
th
is
s
p
ec
if
ically
.
As
s
h
o
wn
in
Fig
u
r
e
9
,
o
u
r
m
o
d
el
o
u
tp
er
f
o
r
m
s
th
e
Y
OL
Ov
1
0
n
m
o
d
el.
Fig
u
r
e
9
(
a)
clea
r
ly
s
h
o
ws
th
at
th
e
r
e
ar
e
s
o
m
e
m
is
s
ed
d
ef
ec
t
d
etec
tio
n
s
,
m
ar
k
ed
in
r
ed
at
th
r
ee
d
if
f
er
e
n
t
lo
ca
tio
n
s
with
d
if
f
er
en
t
d
ef
ec
t
ty
p
es.
Me
an
w
h
ile,
in
Fig
u
r
e
9
(
b
)
,
th
e
m
is
s
ed
lo
ca
tio
n
s
in
Fig
u
r
e
9
(
a)
wer
e
ac
cu
r
ately
d
etec
te
d
b
y
t
h
e
p
r
o
p
o
s
ed
m
o
d
el.
T
h
i
s
im
p
r
o
v
em
e
n
t
ca
n
Evaluation Warning : The document was created with Spire.PDF for Python.