I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
1
5
,
No
.
4
,
A
u
g
u
s
t 2
0
2
6
,
p
p
.
3
6
6
0
~
3
6
7
1
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
4
.
p
p
3
6
6
0
-
3
6
7
1
3660
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
M
ulti
-
sta
g
e hybri
d Y
O
LO
-
driv
en a
nd Mo
bileNe
tV2
-
CNN
v
a
ria
nts f
o
r robu
st f
ish
freshn
ess
c
la
ss
ificatio
n
Ra
s
ee
da
H
a
m
za
h
1
,
Ro
s
niza
Ro
s
la
n
1
,
Am
ni M
u
nira
K
hid
ir
1
,
L
a
la
Septe
m
Riza
2
1
C
e
n
t
e
r
o
f
C
o
m
p
u
t
i
n
g
S
c
i
e
n
c
e
s,
F
a
c
u
l
t
y
o
f
C
o
mp
u
t
e
r
a
n
d
M
a
t
h
e
ma
t
i
c
a
l
S
c
i
e
n
c
e
s,
U
n
i
v
e
r
si
t
i
Te
k
n
o
l
o
g
i
M
A
R
A
,
M
e
l
a
k
a
,
M
a
l
a
y
s
i
a
2
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
Ed
u
c
a
t
i
o
n
,
F
a
c
u
l
t
y
o
f
M
a
t
h
e
m
a
t
i
c
s
a
n
d
S
c
i
e
n
c
e
E
d
u
c
a
t
i
o
n
,
U
n
i
v
e
r
s
i
t
a
s Pe
n
d
i
d
i
k
a
n
I
n
d
o
n
e
s
i
a
,
B
a
n
d
u
n
g
,
I
n
d
o
n
e
si
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Sep
2
5
,
2
0
2
5
R
ev
is
ed
J
u
n
9
,
2
0
2
6
Acc
ep
ted
J
u
l 9
,
2
0
2
6
Th
is
stu
d
y
p
re
se
n
ts
a
m
u
lt
i
-
sta
g
e
h
y
b
r
id
re
p
re
se
n
tati
o
n
lea
rn
in
g
fr
a
m
e
wo
rk
fo
r
ro
b
u
st
fish
fre
sh
n
e
ss
c
las
sifica
ti
o
n
to
a
d
d
re
ss
th
e
c
rit
ica
l
c
h
a
ll
e
n
g
e
o
f
re
li
a
b
le
q
u
a
li
ty
a
ss
e
ss
m
e
n
t
in
u
n
p
re
se
rv
e
d
fo
o
d
s
u
p
p
ly
c
h
a
in
s.
Th
e
p
ro
p
o
se
d
p
ip
e
li
n
e
o
p
e
ra
tes
in
two
sta
g
e
s:
sta
g
e
-
1
e
m
p
lo
y
s
y
o
u
o
n
ly
l
o
o
k
o
n
c
e
(
YO
LO
)
v
8
n
a
s
fe
a
tu
re
-
g
a
ted
d
e
t
e
c
to
r
to
v
a
li
d
a
te
in
p
u
ts
a
n
d
e
li
m
in
a
te
n
o
n
-
fish
ima
g
e
s,
wh
il
e
sta
ge
-
2
lev
e
ra
g
e
s
tran
sfe
r
-
lea
rn
e
d
M
o
b
il
e
Ne
tV
2
v
a
rian
ts
e
n
h
a
n
c
e
d
with
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
(
CNN
)
lay
e
rs,
fin
e
-
tu
n
in
g
,
a
d
a
p
ti
v
e
lea
rn
i
n
g
ra
te
sc
h
e
d
u
lers
,
a
n
d
e
x
p
a
n
d
e
d
fu
l
ly
c
o
n
n
e
c
ted
lay
e
rs
fo
r
h
iera
rc
h
ica
l
c
las
sifica
ti
o
n
in
to
t
h
r
e
e
fre
sh
n
e
ss
c
l
a
ss
e
s
i.
e
.
,
h
ig
h
ly
fr
e
sh
,
fre
sh
,
a
n
d
n
o
t
fre
sh
.
T
h
e
fra
m
e
wo
rk
h
a
s
b
e
e
n
train
e
d
a
n
d
e
v
a
lu
a
ted
o
n
tw
o
c
u
ra
ted
d
a
tas
e
ts
c
o
m
p
risin
g
4
,
5
0
0
fis
h
a
n
d
n
o
n
-
f
ish
ima
g
e
s
a
n
d
1
1
,
1
1
1
fre
sh
n
e
ss
-
lab
e
led
in
c
lu
d
in
g
a
u
g
m
e
n
te
d
d
a
ta
se
t
to
in
c
re
a
se
d
iv
e
rsity
.
Th
e
e
x
p
e
rime
n
tal
re
su
lt
s
sh
o
we
d
sig
n
ifi
c
a
n
t
imp
r
o
v
e
m
e
n
ts.
Th
e
b
e
st
p
e
rfo
rm
a
n
c
e
m
o
d
e
l
tran
sfe
r
lea
rn
in
g
(
TL
)
-
M
o
b
il
e
Ne
t
V2
+
CNN
+
fin
e
-
tu
n
i
n
g
a
c
h
iev
e
d
9
8
.
0
7
%
train
in
g
a
c
c
u
ra
c
y
a
n
d
6
7
.
7
2
%
v
a
li
d
a
ti
o
n
a
c
c
u
ra
c
y
o
n
8
0
:2
0
s
p
li
t,
a
n
d
train
in
g
a
c
c
u
ra
c
y
o
f
9
7
.
5
7
%
,
v
a
li
d
a
ti
o
n
a
c
c
u
ra
c
y
o
f
9
7
.
2
1
%
t
h
ro
u
g
h
1
0
-
fo
ld
c
ro
ss
-
v
a
li
d
a
ti
o
n
.
Th
e
c
o
m
p
a
ra
ti
v
e
b
e
n
c
h
m
a
rk
in
g
c
o
n
firme
d
t
h
a
t
d
u
a
l
-
sta
g
e
d
e
sig
n
o
u
t
p
e
rfo
rm
e
d
b
a
se
li
n
e
M
o
b
il
e
Ne
tV2
a
n
d
YO
LOv
5
s
m
o
d
e
ls
a
c
ro
ss
p
re
c
isio
n
,
re
c
a
ll
,
a
n
d
F
1
-
sc
o
re
.
T
h
e
fin
d
i
n
g
s
h
i
g
h
li
g
h
ted
si
g
n
ifi
c
a
n
t
v
a
l
u
e
o
f
in
teg
ra
ti
n
g
d
e
tec
ti
o
n
-
d
ri
v
e
n
v
a
li
d
a
ti
o
n
wit
h
tran
sfe
r
-
lea
rn
in
g
c
la
ss
ifi
c
a
ti
o
n
,
a
n
d
p
r
o
p
o
se
n
e
w
b
e
n
c
h
m
a
rk
fo
r
in
telli
g
e
n
t
fre
sh
n
e
ss
m
o
n
it
o
r
in
g
.
F
o
r
fu
tu
re
wo
rk
,
th
is
st
u
d
y
a
ims
to
e
x
p
lo
r
e
a
tt
e
n
ti
o
n
-
b
a
se
d
m
o
d
e
ls,
d
a
ta
i
n
teg
ra
ti
o
n
,
a
n
d
sp
e
c
ies
d
i
v
e
rsity
.
K
ey
w
o
r
d
s
:
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
v
ar
ian
ts
Dee
p
lear
n
in
g
Fis
h
f
r
esh
n
ess
T
r
an
s
f
er
lear
n
i
n
g
-
Mo
b
ileNetV2
Yo
u
o
n
ly
lo
o
k
o
n
ce
-
d
r
iv
e
n
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
:
R
o
s
n
iza
R
o
s
lan
C
en
ter
o
f
C
o
m
p
u
tin
g
Scien
ce
s
,
Facu
lty
o
f
C
o
m
p
u
ter
an
d
Ma
th
em
atica
l Scie
n
ce
s
Un
iv
er
s
iti T
ek
n
o
lo
g
i M
AR
A
7
7
3
0
0
Me
r
lim
au
,
Me
lak
a,
Ma
lay
s
ia
E
m
ail:
r
o
s
n
iza@t
m
s
k
.
u
itm
.
ed
u
.
m
y
1.
I
NT
RO
D
UCT
I
O
N
T
h
e
g
lo
b
al
f
is
h
in
d
u
s
tr
y
is
v
e
r
y
im
p
o
r
ta
n
t
f
o
r
f
o
o
d
s
ec
u
r
ity
an
d
n
u
tr
itio
n
,
esp
ec
ially
in
d
ev
elo
p
in
g
r
eg
io
n
s
,
as
f
is
h
is
th
e
p
r
im
ar
y
s
o
u
r
ce
o
f
p
r
o
tein
an
d
ess
en
tial
m
icr
o
n
u
t
r
ien
ts
s
u
ch
as
io
d
in
e,
s
elen
iu
m
,
an
d
v
itam
in
B
1
2
[
1
]
.
Ho
wev
e
r
,
f
is
h
is
o
n
e
o
f
th
e
m
o
s
t
p
er
is
h
ab
le
f
o
o
d
p
r
o
d
u
cts,
with
f
r
esh
n
ess
r
ap
id
ly
d
eter
io
r
atin
g
d
u
e
t
o
en
z
y
m
at
ic,
m
icr
o
b
ial,
a
n
d
p
h
y
s
ico
ch
em
ical
ch
an
g
es
d
u
r
in
g
s
to
r
a
g
e
an
d
d
is
tr
ib
u
tio
n
.
W
h
en
f
is
h
is
s
to
r
ed
im
p
r
o
p
er
ly
af
ter
h
ar
v
est,
it
ca
n
tu
r
n
in
to
h
is
tid
in
e.
T
h
is
p
r
o
ce
s
s
ca
n
ca
u
s
e
s
co
m
b
r
o
id
p
o
is
o
n
in
g
,
g
astro
in
test
in
al
illn
ess
,
r
esp
ir
ato
r
y
co
m
p
licatio
n
s
,
an
d
p
o
ten
tially
f
atal
o
u
tco
m
es
[
2
]
–
[
4
]
.
B
ey
o
n
d
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mu
lti
-
s
ta
g
e
h
yb
r
id
YOLO
-
d
r
iv
en
a
n
d
Mo
b
ileN
etV
2
-
C
N
N
va
r
ia
n
ts
fo
r
r
o
b
u
s
t
…
(
R
a
s
ee
d
a
H
a
mza
h
)
3661
h
ea
lth
r
is
k
s
,
s
p
o
ilag
e
cr
ea
tes
r
ip
p
le
ef
f
ec
ts
ac
r
o
s
s
s
u
p
p
ly
c
h
ain
s
,
u
n
d
er
m
i
n
in
g
c
o
n
s
u
m
er
t
r
u
s
t,
co
n
tr
ib
u
tes
to
f
o
o
d
waste,
an
d
th
r
ea
ten
to
ec
o
n
o
m
ic
e
x
p
o
r
tin
g
n
atio
n
s
[
4
]
.
T
h
e
tr
ad
itio
n
al
m
eth
o
d
s
s
u
ch
as
th
e
q
u
ality
in
d
ex
m
eth
o
d
ar
e
wid
ely
u
s
ed
to
s
co
r
e
s
en
s
o
r
y
attr
ib
u
tes
in
clu
d
in
g
co
lo
r
,
o
d
o
r
,
an
d
ey
e
o
r
g
ill
clar
ity
,
b
u
t
r
em
ain
o
p
e
r
ato
r
-
d
ep
en
d
en
t
an
d
in
co
n
s
is
ten
t
ac
r
o
s
s
ev
alu
ato
r
s
[
5
]
.
Hu
m
an
ev
alu
at
o
r
s
ar
e
lik
ely
to
f
atig
u
e
an
d
b
i
as,
wh
ile
r
elate
d
co
s
ts
to
m
ain
tain
s
k
ill
team
s
ar
e
ex
p
en
s
iv
e
[
6
]
,
[
7
]
.
T
h
e
s
en
s
o
r
y
ev
alu
atio
n
b
ased
o
n
p
h
en
o
ty
p
ic
m
ar
k
er
s
s
u
ch
as
ey
e
clar
ity
,
g
ill
co
lo
r
atio
n
,
an
d
b
o
d
y
s
tr
u
ctu
r
e
r
em
ain
p
r
ac
ticed
b
u
t
c
o
n
s
tr
ain
ed
with
s
u
b
jectiv
ity
,
in
co
n
s
is
ten
cy
,
a
n
d
d
e
p
en
d
e
n
ce
o
n
h
u
m
an
ex
p
e
r
ts
[
5
]
,
[
6
]
,
[
8
]
.
T
h
e
in
s
tr
u
m
e
n
tal
m
eth
o
d
s
s
u
ch
as
to
r
r
y
m
eter
a
n
d
s
p
ec
tr
o
s
co
p
y
-
b
ased
p
r
o
v
id
e
m
o
r
e
o
b
jectiv
ity
b
u
t
r
eq
u
ir
e
h
ig
h
ac
q
u
is
itio
n
c
o
s
ts
,
im
p
r
ac
tical
d
esig
n
,
a
n
d
lack
o
f
s
ca
lab
ilit
y
[
4
]
,
[
8
]
.
T
h
e
m
an
u
al
m
et
h
o
d
s
a
r
e
tim
e
-
c
o
n
s
u
m
in
g
,
a
n
d
s
o
m
e
ex
is
tin
g
d
i
g
ital
s
y
s
tem
s
,
s
u
ch
as
f
is
h
ey
e
-
b
ased
class
if
ier
s
,
h
av
e
ac
h
iev
e
d
m
o
d
est
ac
cu
r
a
cy
with
7
7
.
3
%,
f
u
r
t
h
er
lim
itin
g
ad
o
p
ti
o
n
[
9
]
,
[
1
0
]
.
T
h
e
em
e
r
g
in
g
tech
n
o
lo
g
ies
s
u
ch
as
b
io
s
en
s
o
r
s
an
d
elec
tr
o
n
ic
n
o
s
es
ca
n
d
etec
t
ch
e
m
ical
m
ar
k
er
s
o
f
s
p
o
ilag
e,
b
u
t
r
e
m
ain
at
ex
p
er
im
en
tal
s
tag
e
with
h
ig
h
co
s
ts
an
d
tech
n
ical
b
ar
r
ier
s
to
c
o
m
m
er
cializ
atio
n
.
T
h
e
r
e
c
en
t
a
d
v
an
c
e
m
en
t
s
i
n
d
e
e
p
l
e
ar
n
in
g
,
c
o
n
v
o
l
u
t
io
n
a
l
n
e
u
r
a
l
n
e
t
wo
r
k
s
(
C
N
N
s
)
a
n
d
l
ig
h
t
w
e
ig
h
t
o
b
j
e
c
t
d
e
t
e
c
to
r
s
s
u
c
h
a
s
y
o
u
o
n
ly
l
o
o
k
o
n
c
e
(
Y
O
L
O
)
v
8
h
a
v
e
tr
a
n
s
f
o
r
m
ed
a
u
to
m
a
t
ed
v
i
s
u
a
l
i
n
s
p
e
c
t
io
n
b
y
en
ab
l
i
n
g
r
ea
l
-
t
i
m
e
f
e
a
tu
r
e
e
x
t
r
a
c
t
i
o
n
an
d
o
b
je
c
t
l
o
c
a
li
z
a
t
i
o
n
[
9
]
–
[
1
6
]
.
F
u
r
th
er
m
o
r
e
,
t
r
an
s
f
er
l
e
ar
n
in
g
w
i
t
h
b
a
c
k
b
o
n
e
s
l
i
k
e
Mo
b
i
le
N
e
t
V
2
a
l
lo
w
s
p
r
e
-
tr
a
in
e
d
m
o
d
e
l
s
t
o
a
d
a
p
t
e
f
f
e
c
t
i
v
e
l
y
to
s
p
e
c
ia
l
i
z
ed
a
g
r
i
cu
l
t
u
r
a
l
d
o
m
a
in
s
w
i
t
h
li
m
i
t
e
d
a
n
n
o
ta
t
e
d
d
a
t
a
[
1
7
]
,
[
1
8
]
.
De
s
p
i
t
e
t
h
e
s
e
ad
v
an
c
e
s
,
th
e
e
x
i
s
t
i
n
g
v
i
s
i
o
n
-
b
a
s
ed
ap
p
r
o
a
ch
e
s
s
h
o
w
s
o
m
e
l
i
m
i
t
a
ti
o
n
s
o
f
s
in
g
l
e
-
f
e
a
tu
r
e
C
N
N
c
l
a
s
s
i
f
i
er
s
l
a
ck
r
e
l
i
ab
i
l
i
t
y
a
c
r
o
s
s
d
i
v
er
s
e
s
p
ec
i
e
s
[
9
]
–
[
1
1
]
,
r
o
b
o
t
i
c
v
i
s
u
a
l
s
y
s
t
e
m
s
d
em
a
n
d
s
t
r
i
c
t
l
y
c
o
n
t
r
o
l
l
e
d
s
t
u
d
io
e
n
v
i
r
o
n
m
e
n
t
s
[
4
]
,
an
d
s
t
a
n
d
a
lo
n
e
l
ig
h
t
we
i
g
h
t
C
N
N
s
l
a
c
k
i
n
p
u
t
v
a
l
id
a
t
i
o
n
m
ec
h
an
i
s
m
s
t
o
f
i
l
te
r
o
u
t
b
a
c
k
g
r
o
u
n
d
n
o
i
s
e
o
r
n
o
n
-
f
i
s
h
d
i
s
t
r
a
c
to
r
s
[
1
9
]
–
[
2
3
]
.
T
h
u
s
,
t
h
e
s
e
s
h
o
r
t
co
m
in
g
s
h
i
g
h
l
i
g
h
t
i
n
t
eg
r
a
t
ed
d
u
a
l
-
s
t
a
g
e
f
r
a
m
ew
o
r
k
t
h
a
t
u
n
i
f
i
e
s
r
e
a
l
-
t
im
e
d
e
te
c
t
i
o
n
-
d
r
iv
e
n
in
p
u
t
g
a
t
i
n
g
w
i
t
h
f
i
n
e
-
g
r
a
i
n
ed
,
m
u
l
t
i
-
r
eg
i
o
n
c
l
a
s
s
i
f
i
c
a
t
io
n
u
n
d
e
r
a
s
c
a
l
ab
l
e
p
i
p
e
l
in
e
[
1
5
]
–
[
2
1
]
,
[
2
4
]
–
[
2
7
]
.
T
h
er
ef
o
r
e,
th
is
s
tu
d
y
in
tr
o
d
u
ce
s
Fu
s
io
n
Net
-
YOL
O
m
o
d
el,
d
u
al
-
s
tag
e
h
y
b
r
i
d
d
ee
p
lear
n
in
g
f
r
am
ewo
r
k
f
o
r
f
is
h
f
r
esh
n
es
s
class
if
icatio
n
.
I
n
th
e
f
ir
s
t
s
tag
e,
YOL
Ov
8
n
as
in
tellig
e
n
t
in
p
u
t
v
alid
ato
r
,
d
etec
tio
n
o
f
f
is
h
v
e
r
s
u
s
n
o
n
-
f
is
h
im
ag
es
a
n
d
to
r
ed
u
ce
f
a
ls
e
er
r
o
r
.
I
n
th
e
s
ec
o
n
d
s
tag
e,
tr
an
s
f
er
lear
n
i
n
g
(
TL
)
-
Mo
b
ileNetV2
v
ar
ian
ts
with
C
NN
lay
er
s
,
f
in
e
-
tu
n
in
g
s
tr
ateg
ies,
ad
ap
tiv
e
lear
n
i
n
g
r
ate
s
ch
ed
u
ler
s
,
an
d
ad
d
itio
n
al
f
u
lly
co
n
n
ec
ted
la
y
er
s
to
p
er
f
o
r
m
h
ier
ar
ch
ical
class
if
icatio
n
o
f
f
r
esh
n
ess
lev
els.
T
h
is
d
u
al
-
s
tag
e
in
teg
r
atio
n
estab
lis
h
es
a
m
eth
o
d
o
lo
g
ical
b
etwe
en
d
etec
tio
n
an
d
class
if
icatio
n
as
to
im
p
r
o
v
e
r
o
b
u
s
tn
ess
,
en
h
an
ce
d
is
cr
im
in
ativ
e
ca
p
ab
i
lity
,
an
d
s
ca
lab
ilit
y
ac
r
o
s
s
d
at
asets
.
T
h
e
m
ain
co
n
tr
ib
u
tio
n
s
o
f
th
is
s
tu
d
y
ar
e
as
f
o
ll
o
ws
:
−
Du
al
-
s
tag
e
h
y
b
r
i
d
f
r
am
ewo
r
k
:
to
p
r
o
p
o
s
e
n
o
v
el
ar
ch
itectu
r
e
th
at
in
teg
r
ates
YOL
Ov
8
n
as
f
ea
tu
r
e
-
g
ated
d
etec
tio
n
s
tag
e
with
TL
-
Mo
b
ileNetV2
v
ar
ian
ts
f
o
r
f
is
h
f
r
esh
n
ess
clas
s
if
icatio
n
.
T
h
is
f
r
am
ewo
r
k
to
en
s
u
r
e
th
at
o
n
ly
v
alid
ate
d
f
is
h
im
ag
es
ar
e
f
u
r
th
er
p
r
o
ce
s
s
ed
as
to
r
ed
u
ce
e
r
r
o
r
p
r
o
p
ag
atio
n
an
d
im
p
r
o
v
e
o
v
er
all
ap
p
licatio
n
r
eliab
ilit
y
.
−
Mu
lti
-
r
eg
io
n
f
ea
tu
r
e
in
te
g
r
at
io
n
:
to
en
h
an
ce
t
h
e
v
ar
i
o
u
s
d
atasets
r
o
b
u
s
tn
ess
th
r
o
u
g
h
th
e
ca
p
tu
r
e
lo
ca
lized
an
ato
m
ical
f
ea
tu
r
es
i.e
.
,
b
o
d
y
,
ey
es,
an
d
g
ills
v
ia
d
ata
au
g
m
en
tatio
n
tr
an
s
f
o
r
m
atio
n
.
T
h
is
ap
p
r
o
ac
h
allo
ws
th
e
f
r
am
ewo
r
k
to
g
en
er
alize
a
cr
o
s
s
v
ar
io
u
s
s
p
ec
ies,
im
ag
in
g
co
n
d
itio
n
s
,
an
d
v
ar
y
i
n
g
ca
p
tu
r
e
en
v
ir
o
n
m
en
ts
.
−
C
o
m
p
r
eh
en
s
iv
e
b
e
n
ch
m
ar
k
:
to
v
alid
ate
lay
er
ed
ev
al
u
atio
n
t
o
r
ep
r
esen
t
th
e
s
ig
n
if
ican
t
p
r
o
p
o
s
ed
v
ar
ian
ts
m
o
d
el.
At
s
tag
e
-
1
,
YOL
Ov
8
n
co
n
s
is
ten
tly
o
u
t
p
er
f
o
r
m
e
d
YOL
Ov
5
s
in
d
etec
tio
n
ac
cu
r
ac
y
an
d
f
alse
-
p
o
s
itiv
e
r
ed
u
ctio
n
.
At
s
tag
e
-
2
,
th
e
en
h
an
ce
d
T
L
-
Mo
b
ileNetV2
m
o
d
els
o
u
t
p
er
f
o
r
m
b
aselin
e
Mo
b
ileNetV2
ac
r
o
s
s
ev
er
y
m
e
tr
ics.
−
Ad
v
an
ce
d
TL
e
n
h
an
ce
m
e
n
ts
:
to
ex
ten
d
th
e
Mo
b
ileNetV2
b
ac
k
b
o
n
e
with
ad
d
itio
n
al
C
NN
lay
er
s
,
f
in
e
-
tu
n
in
g
,
ad
ap
tiv
e
lear
n
in
g
r
ate
s
ch
ed
u
ler
s
,
an
d
f
u
lly
co
n
n
ec
ted
lay
er
s
th
at
in
cr
ea
s
ed
v
alid
atio
n
ac
cu
r
ac
y
f
r
o
m
6
3
.
6
7
%
(
b
aselin
e
Mo
b
ileNetV2
)
t
o
6
7
.
7
2
%
(
T
L
-
Mo
b
ileNetV2
+
C
NN
+
f
in
e
-
tu
n
in
g
)
,
an
d
s
tab
le
g
en
e
r
aliza
tio
n
at
6
7
.
0
1
% with
ad
ap
tiv
e
lea
r
n
in
g
r
ate
s
ch
ed
u
lin
g
.
−
T
r
an
s
f
er
ab
le
m
eth
o
d
o
lo
g
ical
f
r
am
ewo
r
k
:
to
g
en
er
ate
d
u
al
-
s
tag
e
m
eth
o
d
o
lo
g
y
f
o
r
h
ier
ar
ch
ica
l
class
if
icatio
n
task
s
.
T
h
e
s
ca
lab
ilit
y
an
d
ad
a
p
tab
ilit
y
as
to
e
x
ten
d
u
n
p
r
eser
v
ed
f
o
o
d
q
u
alit
y
m
o
n
ito
r
in
g
d
o
m
ain
s
an
d
in
tellig
en
t im
ag
e
class
if
icatio
n
.
T
h
e
s
tr
u
ctu
r
e
o
f
th
is
p
ap
er
is
o
r
g
an
ized
as
f
o
llo
ws.
Sectio
n
1
in
tr
o
d
u
ce
s
r
esear
ch
s
tu
d
y
b
a
ck
g
r
o
u
n
d
,
m
o
tiv
atio
n
,
liter
atu
r
e
r
e
v
iew,
p
r
o
p
o
s
ed
s
o
lu
tio
n
,
an
d
co
n
tr
i
b
u
tio
n
s
.
Sectio
n
2
d
escr
ib
es
th
e
p
r
o
p
o
s
ed
m
eth
o
d
s
i.e
.
,
im
ag
e
d
ataset
p
r
e
p
ar
atio
n
,
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
,
s
tag
e
-
1
f
ea
tu
r
e
-
g
ated
d
etec
tio
n
m
o
d
el
s
i.e
.
,
YOL
Ov
5
s
v
s
.
YOL
Ov
8
n
,
an
d
s
tag
e
-
2
Mo
b
ileNetV2
-
b
ased
v
ar
ian
ts
.
Sectio
n
3
p
r
esen
ts
th
e
ex
p
er
i
m
en
tal
r
esu
lts
an
d
co
m
p
ar
ativ
e
b
en
c
h
m
ar
k
in
g
a
g
ain
s
t
ex
is
tin
g
m
et
h
o
d
s
.
Fin
ally
,
se
ctio
n
4
c
o
n
clu
d
es
th
e
r
esear
ch
s
tu
d
y
,
co
n
tr
ib
u
tio
n
s
to
th
e
s
tate
-
of
-
t
h
e
-
ar
t,
an
d
r
ec
o
m
m
e
n
d
atio
n
s
f
o
r
f
u
tu
r
e
r
esear
ch
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
4
,
Au
g
u
s
t 2
0
2
6
:
3
6
6
0
-
3
6
7
1
3662
2.
M
E
T
H
O
D
T
h
e
aim
o
f
t
h
is
s
tu
d
y
is
to
im
p
r
o
v
e
m
u
lti
-
s
tag
e
h
y
b
r
id
r
e
p
r
esen
tatio
n
lear
n
in
g
f
r
am
ewo
r
k
f
o
r
r
o
b
u
s
t
f
is
h
f
r
esh
n
ess
class
if
icatio
n
[
1
9
]
,
[
2
0
]
.
T
h
is
s
tu
d
y
in
teg
r
ates
YOL
Ov
5
s
v
s
YOL
Ov
8
n
f
ea
t
u
r
e
-
g
ated
d
etec
tio
n
at
th
e
f
ir
s
t
s
tag
e
with
Mo
b
ile
NetV2
v
ar
ian
ts
at
th
e
s
ec
o
n
d
s
tag
e
to
ac
h
iev
e
f
in
e
-
g
r
ain
ed
f
r
esh
n
ess
p
r
ed
ictio
n
[
1
4
]
,
[
1
8
]
,
[
2
3
]
.
T
h
is
d
u
al
-
s
tag
e
r
ep
r
esen
tatio
n
aim
s
t
o
r
e
d
u
c
e
er
r
o
r
p
r
o
p
ag
atio
n
,
im
p
r
o
v
e
m
o
d
el
s
tab
ilit
y
,
an
d
en
h
an
ce
g
en
er
aliza
tio
n
ac
r
o
s
s
v
ar
io
u
s
d
atasets
,
an
d
im
ag
in
g
co
n
d
itio
n
s
.
T
h
e
o
v
er
all
p
r
o
p
o
s
ed
m
o
d
el
is
s
h
o
wn
in
Fig
u
r
e
1
an
d
o
u
tlin
es
as
f
o
llo
ws:
i)
im
ag
e
d
ataset
p
r
ep
ar
atio
n
,
ii)
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
,
iii)
s
tag
e
-
1
f
ea
tu
r
e
-
g
ated
d
etec
tio
n
v
ia
Y
OL
Ov
5
s
v
er
s
u
s
YOL
Ov
8
n
,
a
n
d
iv
)
s
tag
e
-
2
r
o
b
u
s
t
f
is
h
f
r
esh
n
ess
class
if
icatio
n
with
Mo
b
ileNetV2
v
ar
ian
ts
.
Fig
u
r
e
1
.
Pro
p
o
s
ed
m
u
lti
-
s
tag
e
h
y
b
r
i
d
p
ip
elin
e
2
.
1
.
I
ma
g
e
da
t
a
s
et
prepa
ra
t
i
o
n
T
h
e
im
ag
e
d
ataset
p
r
ep
ar
ati
o
n
f
o
r
th
is
s
tu
d
y
h
as
b
ee
n
co
n
d
u
cte
d
to
e
x
p
er
im
e
n
t
th
e
d
u
al
-
s
tag
e
f
r
am
ewo
r
k
th
at
i
n
teg
r
ates
s
tag
e
-
1
f
ea
tu
r
e
-
g
ated
d
etec
tio
n
v
ia
YOL
Ov
8
f
o
r
f
is
h
v
e
r
s
u
s
n
o
n
-
f
is
h
v
alid
atio
n
,
an
d
s
tag
e
-
2
f
r
esh
n
ess
class
if
ic
atio
n
f
o
r
f
is
h
ca
teg
o
r
izatio
n
in
to
f
r
esh
n
ess
lev
els
[
4
]
,
[
1
4
]
.
T
h
is
s
tu
d
y
to
en
s
u
r
e
th
at
o
n
ly
v
alid
f
is
h
im
ag
es
p
r
o
ce
ed
to
f
r
esh
n
ess
class
if
icat
io
n
.
T
h
is
ap
p
r
o
ac
h
r
ed
u
ce
s
f
alse
p
o
s
itiv
es,
m
itig
ate
er
r
o
r
p
r
o
p
a
g
atio
n
,
an
d
s
tr
en
g
t
h
en
in
g
r
o
b
u
s
tn
ess
[
9
]
–
[
1
3
]
.
2
.
1
.
1
.
Sta
g
e
-
1
f
e
a
t
ure
-
g
a
t
ed
det
ec
t
io
n
Fo
r
s
tag
e
-
1
,
th
e
to
tal
im
ag
e
d
ataset
o
f
4
,
6
9
2
im
ag
es
h
as
b
e
en
co
llected
ac
r
o
s
s
two
ca
teg
o
r
ies:
f
is
h
an
d
n
o
n
-
f
is
h
[
1
4
]
,
[
1
6
]
.
T
h
e
f
i
s
h
class
co
n
tain
ed
1
,
5
6
0
im
ag
es,
wh
ich
also
u
s
ed
as
in
p
u
t
f
o
r
s
tag
e
-
2
f
r
esh
n
ess
class
if
icatio
n
.
T
h
ese
im
ag
es
h
av
e
b
ee
n
o
b
tain
ed
f
r
o
m
th
r
ee
r
ep
o
s
ito
r
ies
with
m
u
ltip
le
an
ato
m
ical
r
an
g
es
o
f
b
o
d
y
,
ey
e,
an
d
g
ill
r
eg
io
n
s
[
4
]
,
[
1
3
]
.
T
h
e
n
o
n
-
f
is
h
cl
ass
co
n
tain
ed
3
,
1
3
2
im
ag
es,
in
clu
d
in
g
m
a
r
in
e
b
ac
k
g
r
o
u
n
d
s
,
m
ar
k
et
en
v
ir
o
n
m
en
ts
,
an
d
ir
r
elev
an
t
o
b
jects
as
to
b
u
ild
in
v
ar
ian
ce
ag
ain
s
t
en
v
ir
o
n
m
en
tal
n
o
is
e.
T
h
e
d
etails o
f
th
e
im
a
g
e
d
atas
et
o
f
s
tag
e
-
1
a
r
e
p
r
esen
ted
in
T
ab
le
1
.
T
ab
le
1
.
Stag
e
-
1
im
ag
e
d
ataset
d
escr
ip
tio
n
f
o
r
f
ea
tu
r
e
-
g
ated
d
etec
tio
n
D
a
t
a
s
e
t
I
mag
e
d
a
t
a
se
t
s
o
u
r
c
e
To
t
a
l
n
u
m
b
e
r
o
f
i
ma
g
e
s
F
i
sh
d
a
t
a
s
e
t
F
i
sh
f
r
e
sh
n
e
ss
c
l
a
ss
i
f
i
c
a
t
i
o
n
d
a
t
a
s
e
t
f
r
o
m
K
a
g
g
l
e
,
f
i
s
h
e
y
e
f
r
e
s
h
n
e
ss
d
a
t
a
set
f
r
o
m
M
e
n
d
e
l
e
y
,
a
n
d
p
r
i
m
a
r
y
d
a
t
a
s
e
t
o
f
f
i
s
h
i
ma
g
e
s
t
h
a
t
c
a
p
t
u
r
e
d
a
n
d
c
o
l
l
e
c
t
e
d
a
t
l
o
c
a
l
f
i
sh
m
a
r
k
e
t
s
i
n
M
a
l
a
y
si
a
1
,
5
6
0
N
o
n
-
f
i
s
h
d
a
t
a
s
e
t
C
O
V
I
D
-
19
c
h
e
st
X
-
r
a
y
i
ma
g
e
d
a
t
a
s
e
t
,
j
e
l
l
y
f
i
s
h
i
ma
g
e
d
a
t
a
s
e
t
,
a
n
i
mal
i
m
a
g
e
d
a
t
a
set
-
c
a
t
s
,
d
o
g
s,
a
n
d
f
o
x
e
s
,
a
n
d
g
a
r
b
a
g
e
i
m
a
g
e
d
a
t
a
se
t
f
r
o
m
K
a
g
g
l
e
3
,
1
3
2
To
t
a
l
4
,
6
9
2
2
.
1
.
2
.
Sta
g
e
-
2
ro
bu
s
t
f
re
s
hn
ess
cla
s
s
if
ica
t
io
n
Fo
r
s
t
ag
e
-
2
,
th
e
to
tal
o
f
1
1
,
7
3
6
n
u
m
b
e
r
o
f
i
m
a
g
e
d
a
tase
t
t
h
a
t
h
as
b
ee
n
c
o
ll
ec
t
e
d
t
o
ex
t
e
n
d
t
h
e
ex
p
e
r
i
m
e
n
t
o
n
r
o
b
u
s
t
f
is
h
f
r
es
h
n
ess
i
n
t
o
t
h
r
ee
cl
ass
es:
h
i
g
h
l
y
f
r
esh
,
f
r
es
h
,
an
d
n
o
t
f
r
esh
.
T
h
i
s
d
at
ase
t
i
n
te
g
r
at
e
d
m
u
lti
p
l
e
s
o
u
r
c
es
t
o
en
s
u
r
e
d
i
v
e
r
s
it
y
in
a
n
at
o
m
ic
al
cu
es
an
d
s
p
o
il
ag
e
i
n
d
i
ca
t
o
r
s
,
i.
e
.
i
)
f
is
h
f
r
es
h
n
ess
class
i
f
i
ca
t
io
n
d
a
tas
et,
ii
)
f
r
esh
n
ess
o
f
t
h
e
f
is
h
ey
es d
at
aset
,
a
n
d
iii
)
p
r
im
a
r
y
d
atas
et
o
f
f
is
h
i
m
a
g
es th
at
h
as b
ee
n
ca
p
t
u
r
ed
a
n
d
c
o
l
le
cte
d
at
lo
c
al
f
is
h
m
a
r
k
ets
in
Ma
la
y
s
i
a.
All
d
a
tase
ts
e
m
p
h
asi
ze
d
i
n
t
o
an
at
o
m
i
ca
l
r
e
g
i
o
n
s
(
b
o
d
y
,
e
y
e
,
a
n
d
g
il
l)
,
to
e
n
s
u
r
e
t
h
a
t
class
if
ica
ti
o
n
r
el
ie
d
o
n
m
u
lti
-
r
e
g
i
o
n
al
f
ea
t
u
r
es
r
at
h
e
r
th
a
n
s
i
n
g
le
at
tr
ib
u
t
e
s
u
c
h
as
f
is
h
e
y
e
a
p
p
e
ar
a
n
ce
.
T
h
e
d
et
ails
o
f
i
m
a
g
e
d
at
aset
d
es
cr
i
p
t
io
n
o
f
s
ta
g
e
-
2
ta
b
u
la
te
d
as
i
n
T
a
b
l
e
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mu
lti
-
s
ta
g
e
h
yb
r
id
YOLO
-
d
r
iv
en
a
n
d
Mo
b
ileN
etV
2
-
C
N
N
va
r
ia
n
ts
fo
r
r
o
b
u
s
t
…
(
R
a
s
ee
d
a
H
a
mza
h
)
3663
T
ab
le
2
.
Stag
e
-
2
im
ag
e
d
ataset
d
escr
ip
tio
n
f
o
r
f
is
h
f
r
esh
n
ess
class
if
icatio
n
I
mag
e
d
a
t
a
se
t
s
o
u
r
c
e
F
i
sh
f
r
e
sh
n
e
ss
c
l
a
ss
To
t
a
l
n
u
m
b
e
r
o
f
i
ma
g
e
s
F
i
sh
f
r
e
sh
n
e
ss
c
l
a
ss
i
f
i
c
a
t
i
o
n
d
a
t
a
s
e
t
f
r
o
m
K
a
g
g
l
e
F
r
e
sh
2
,
5
8
1
N
o
t
f
r
e
s
h
1
,
8
9
5
F
i
sh
e
y
e
f
r
e
s
h
n
e
ss
d
a
t
a
set
f
r
o
m
M
e
n
d
e
l
e
y
H
i
g
h
l
y
f
r
e
s
h
1
,
7
6
4
F
r
e
sh
1
,
3
2
0
N
o
t
f
r
e
s
h
1
,
3
0
6
P
r
i
mary
d
a
t
a
set
H
i
g
h
l
y
f
r
e
s
h
9
4
9
F
r
e
sh
9
5
7
N
o
t
f
r
e
s
h
9
6
4
To
t
a
l
11
,
736
2
.
2
.
Da
t
a
pre
-
pro
ce
s
s
ing
T
h
e
p
r
im
ar
y
o
b
jectiv
e
o
f
th
e
d
ata
p
r
e
-
p
r
o
ce
s
s
in
g
p
h
ase
is
to
n
o
r
m
alize
in
p
u
t
r
e
p
r
esen
tatio
n
s
an
d
t
o
q
u
an
tify
th
e
p
r
o
g
r
ess
iv
e
r
ep
r
e
s
en
tatio
n
o
f
T
L
s
tr
ateg
ies
ac
r
o
s
s
1
1
,
7
3
6
m
u
lti
-
r
e
g
io
n
al
im
a
g
es
[
1
7
]
,
[
1
8
]
.
T
h
e
r
aw
im
ag
e
h
as
b
ee
n
ex
p
e
r
im
en
ted
to
s
tan
d
ar
d
im
ag
e
d
im
en
s
io
n
an
d
f
ea
tu
r
e
en
h
an
ce
m
en
t
as
to
ac
h
iev
e
f
in
e
-
g
r
ain
e
d
f
r
esh
n
ess
lev
els
ac
r
o
s
s
th
r
ee
-
class
f
r
esh
n
ess
(
h
ig
h
ly
f
r
esh
,
f
r
esh
,
an
d
n
o
t
f
r
esh
)
[
1
7
]
,
[
1
8
]
.
T
h
is
p
h
ase
f
u
r
th
e
r
ev
alu
ates
s
ev
en
ex
p
er
im
en
tal
m
o
d
el
v
a
r
ian
ts
f
r
o
m
th
e
b
aselin
e
Mo
b
ileNe
tV2
b
ac
k
b
o
n
e
th
at
in
co
r
p
o
r
ates
cu
s
to
m
C
NN
h
e
ad
,
p
a
r
tial
f
in
e
-
t
u
n
in
g
,
a
d
ap
ti
v
e
lear
n
i
n
g
r
ate
s
ch
ed
u
ler
s
,
a
n
d
e
x
p
an
d
ed
f
u
lly
co
n
n
ec
ted
class
if
ier
lay
er
s
[
1
7
]
,
[
2
7
]
.
Fu
r
t
h
er
m
o
r
e,
tar
g
ete
d
d
ata
au
g
m
e
n
tatio
n
an
d
im
a
g
e
r
esizin
g
h
as
b
ee
n
ap
p
lied
t
o
r
e
d
u
ce
o
v
e
r
f
itti
n
g
,
o
p
tim
ize
g
r
ad
ien
t
d
escen
t
co
n
v
er
g
en
ce
,
an
d
im
p
r
o
v
e
f
ea
tu
r
e
in
v
ar
ian
ce
ac
r
o
s
s
b
o
d
y
,
ey
e,
a
n
d
g
ill a
n
ato
m
ical
r
eg
io
n
s
[
2
1
]
,
[
2
5
]
.
2
.
2
.
1
.
Da
t
a
a
ug
m
ent
a
t
io
n
T
h
e
aim
o
f
th
is
p
h
ase
is
to
in
c
r
ea
s
e
th
e
d
iv
er
s
ity
o
f
th
e
d
ataset,
p
r
ev
en
t
o
v
er
f
itti
n
g
,
an
d
in
cr
ea
s
e
th
e
n
u
m
b
er
o
f
f
is
h
im
ag
e
d
ataset
[
1
2
]
,
[
1
5
]
,
[
2
1
]
.
Fo
r
s
tag
e
-
1
,
f
ea
tu
r
e
-
g
ated
d
etec
tio
n
,
d
at
a
au
g
m
en
tatio
n
o
n
g
eo
m
etr
ic
an
d
p
h
o
to
m
etr
ic
tr
an
s
f
o
r
m
atio
n
s
to
s
tr
en
g
th
e
n
in
v
ar
ia
n
ce
in
d
etec
tio
n
.
Fo
r
s
tag
e
-
2
,
r
o
b
u
s
t
f
r
esh
n
ess
class
if
icatio
n
,
au
g
m
en
tatio
n
em
p
h
asized
v
ar
ia
b
ilit
y
in
illu
m
in
atio
n
a
n
d
c
o
n
tr
ast
to
s
im
u
late
r
ea
l
-
wo
r
ld
ca
p
tu
r
e
co
n
d
itio
n
s
.
T
h
e
au
g
m
e
n
tatio
n
o
p
e
r
atio
n
s
th
at
h
av
e
b
ee
n
em
p
lo
y
ed
f
o
r
s
tag
e
-
1
an
d
s
tag
e
-
2
ar
e
tab
u
lated
as T
ab
le
3
.
T
ab
le
3
.
Data
au
g
m
en
tatio
n
tr
an
s
f
o
r
m
atio
n
s
f
o
r
s
tag
e
-
1
an
d
s
tag
e
-
2
m
o
d
el
A
u
g
m
e
n
t
a
t
i
o
n
T
e
c
h
n
i
q
u
e
S
t
a
g
e
-
1
S
t
a
g
e
-
2
F
l
i
p
p
i
n
g
H
o
r
i
z
o
n
t
a
l
a
n
d
v
e
r
t
i
c
a
l
H
o
r
i
z
o
n
t
a
l
a
n
d
v
e
r
t
i
c
a
l
R
o
t
a
t
i
o
n
9
0
°
(
c
l
o
c
k
w
i
se
,
c
o
u
n
t
e
r
c
l
o
c
k
w
i
s
e
u
p
si
d
e
d
o
w
n
)
;
r
a
n
d
o
m
−
1
5
°
t
o
+
1
5
°
R
a
n
d
o
m
−
2
5
°
t
o
+
2
5
°
C
r
o
p
p
i
n
g
/
z
o
o
m
0
–
2
0
%
z
o
o
m
a
p
p
l
i
e
d
N
o
t
a
p
p
l
i
e
d
G
r
a
y
sc
a
l
e
c
o
n
v
e
r
si
o
n
1
5
%
o
f
i
ma
g
e
s
c
o
n
v
e
r
t
e
d
N
o
t
a
p
p
l
i
e
d
B
r
i
g
h
t
n
e
ss
a
d
j
u
s
t
me
n
t
B
e
t
w
e
e
n
−
1
5
%
a
n
d
+
1
5
%
r
a
n
g
e
R
a
n
d
o
m
i
z
e
d
b
r
i
g
h
t
n
e
ss
sh
i
f
t
s
C
o
n
t
r
a
st
/
sh
a
r
p
n
e
ss
N
o
t
a
p
p
l
i
e
d
R
a
n
d
o
m
i
z
e
d
c
o
n
t
r
a
st
a
n
d
sh
a
r
p
n
e
ss
2
.
2
.
2
.
I
m
a
g
e
re
s
izing
All
im
ag
e
d
atasets
h
av
e
b
ee
n
r
esized
to
s
tan
d
a
r
d
2
2
4
×
2
2
4
p
ix
els
to
n
o
r
m
alize
th
e
in
p
u
t
im
a
g
e
d
im
en
s
io
n
an
d
m
atch
th
e
r
eq
u
ir
ed
in
p
u
t size
o
f
th
e
Mo
b
ileNetV2
-
b
ased
m
o
d
el
[
1
8
]
,
[
2
2
]
.
T
h
e
s
tan
d
ar
d
im
ag
e
d
im
en
s
io
n
al
is
to
en
s
u
r
e
b
atc
h
-
p
r
o
c
ess
in
g
co
m
p
atib
le
a
n
d
u
n
if
o
r
m
m
atr
ix
o
p
er
atio
n
s
ac
r
o
s
s
h
eter
o
g
en
e
o
u
s
im
ag
es
f
r
o
m
m
u
ltip
le
r
e
p
o
s
ito
r
y
s
o
u
r
ce
s
.
Fu
r
th
er
m
o
r
e,
t
h
is
s
p
atial
r
eso
lu
tio
n
o
p
tim
izes
th
e
h
y
b
r
id
m
o
d
e
l
p
er
f
o
r
m
an
ce
b
y
b
ala
n
cin
g
c
o
m
p
u
tatio
n
al
ef
f
icie
n
cy
an
d
f
ea
tu
r
e
p
r
eser
v
atio
n
,
th
er
eb
y
r
etain
f
in
e
-
g
r
ain
e
d
v
is
u
al
d
etails ac
r
o
s
s
th
e
ey
e,
g
ill an
d
b
o
d
y
d
u
r
in
g
th
e
m
o
d
el
tr
ain
in
g
[
2
2
]
,
[
2
5
]
.
2
.
3
.
St
a
g
e
-
1
f
ea
t
ure
-
g
a
t
ed
d
et
ec
t
io
n
T
h
e
aim
o
f
s
tag
e
-
1
m
o
d
el
is
to
d
etec
t
an
d
class
if
y
th
e
f
ea
t
u
r
e
-
g
atin
g
o
f
f
is
h
im
ag
e
r
ep
r
esen
tatio
n
[
1
4
]
,
[
1
6
]
.
T
wo
m
o
d
els
f
r
o
m
th
e
YOL
O
f
am
ily
i.e
.
,
YOL
Ov
5
s
an
d
YOL
Ov
8
n
h
av
e
b
e
en
ex
p
er
im
e
n
ted
an
d
ev
alu
ated
as
to
b
en
ch
m
a
r
k
t
h
e
p
er
f
o
r
m
an
ce
[
2
3
]
.
T
h
is
s
tu
d
y
em
p
l
o
y
ed
YOL
Ov
5
s
,
an
d
YOL
Ov
8
n
as
to
v
alid
ate
f
is
h
im
a
g
e.
B
y
f
ilte
r
in
g
o
u
t
n
o
n
-
f
is
h
i
n
p
u
ts
,
t
h
is
s
tag
e
ac
ted
as
f
ea
tu
r
e
-
g
ati
n
g
m
ec
h
an
is
m
t
h
at
r
ed
u
ce
d
f
alse p
o
s
itiv
es a
n
d
m
i
n
im
ized
er
r
o
r
p
r
o
p
a
g
atio
n
in
to
s
tag
e
-
2.
2
.
3
.
1
.
YO
L
O
v
5
s
T
h
is
s
tu
d
y
ex
p
er
im
en
ted
Y
OL
Ov
5
s
as
th
e
b
aselin
e
o
b
ject
d
etec
tio
n
m
o
d
el,
an
d
a
lig
h
tweig
h
t
m
em
b
er
o
f
th
e
YOL
Ov
5
f
am
ily
[
1
6
]
.
YOL
Ov
5
s
co
m
b
in
es
r
ea
l
-
tim
e
d
etec
tio
n
with
b
o
u
n
d
in
g
b
o
x
lo
ca
lizatio
n
,
en
ab
lin
g
id
en
tif
icatio
n
o
f
f
is
h
o
b
jects
in
cl
u
tter
ed
en
v
ir
o
n
m
en
ts
[
2
3
]
.
T
h
e
s
tag
e
-
1
d
ataset
co
n
tain
ed
4
,
6
9
2
lab
eled
im
ag
es
s
p
lit
b
etwe
en
f
is
h
(
1
,
5
6
0
im
ag
es)
a
n
d
n
o
n
-
f
is
h
(
3
,
1
3
2
im
a
g
es).
Fo
r
YOL
Ov
5
s
tr
ain
in
g
,
b
o
u
n
d
in
g
b
o
x
es
h
av
e
b
ee
n
m
an
u
ally
a
n
n
o
tated
f
o
r
th
e
f
is
h
class
,
an
d
n
o
n
-
f
is
h
im
ag
es
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
4
,
Au
g
u
s
t 2
0
2
6
:
3
6
6
0
-
3
6
7
1
3664
lab
elled
as
b
ac
k
g
r
o
u
n
d
.
T
h
e
ex
p
er
im
en
te
d
k
ey
f
ea
tu
r
es
ar
e
to
en
h
a
n
ce
m
u
ltil
ev
el
f
ea
tu
r
e
f
u
s
io
n
an
d
d
ec
o
u
p
le
d
atten
tio
n
m
ec
h
an
i
s
m
s
,
r
ed
u
ce
f
alse
p
o
s
itiv
es
an
d
im
p
r
o
v
e
d
etec
tio
n
o
f
s
m
all
o
b
jects,
s
im
p
le,
p
ar
am
eter
-
f
r
ee
atten
tio
n
m
o
d
u
le
(
SimA
M
)
an
d
c
u
s
to
m
c
h
an
n
el
atten
tio
n
m
o
d
u
les
s
tr
en
g
th
en
f
o
c
u
s
o
n
r
elev
an
t
f
ea
t
u
r
es
in
co
m
p
lex
s
ce
n
es,
to
ev
alu
ate
a
d
v
an
c
ed
in
ter
s
ec
tio
n
o
v
e
r
u
n
io
n
(
I
o
U)
-
b
ased
lo
s
s
es,
in
clu
d
in
g
f
o
ca
l
m
in
im
u
m
p
o
i
n
t
d
is
tan
ce
I
o
U
an
d
co
m
p
lete
in
ter
s
ec
tio
n
o
v
er
u
n
i
o
n
(
C
I
o
U
)
,
ad
d
r
ess
class
im
b
alan
ce
an
d
im
p
r
o
v
e
co
n
v
er
g
en
ce
;
an
d
b
id
ir
ec
tio
n
al
f
e
atu
r
e
p
y
r
am
id
n
etwo
r
k
(
B
iFP
N
)
with
co
o
r
d
in
ate
atten
tio
n
r
ef
i
n
es
f
ea
tu
r
e
r
e
p
r
esen
tatio
n
,
b
o
o
s
tin
g
d
etec
tio
n
ac
cu
r
a
cy
.
T
h
e
tr
ain
in
g
ex
p
er
im
en
t
h
as
b
ee
n
p
er
f
o
r
m
ed
f
o
r
5
e
p
o
ch
s
wit
h
an
i
n
p
u
t
s
ize
o
f
6
4
0
×
6
4
0
p
ix
els.
T
h
e
ex
p
er
im
en
tal
r
esu
lts
ac
h
iev
ed
an
ac
ce
p
tab
le
7
6
.
7
4
5
%
p
r
ec
is
io
n
an
d
6
2
.
7
6
6
%
r
ec
all
at
ep
o
ch
f
o
u
r
.
Ho
wev
er
,
th
is
s
tu
d
y
s
h
o
wed
two
m
ain
is
s
u
es:
i)
in
cr
ea
s
ed
co
m
p
u
tat
io
n
al
o
v
er
h
ea
d
d
u
e
t
o
b
o
u
n
d
in
g
b
o
x
r
eg
r
ess
io
n
as
to
r
ed
u
ce
s
u
itab
ilit
y
f
o
r
lig
h
tweig
h
t
d
e
p
lo
y
m
e
n
t;
an
d
i
i)
r
ed
u
ce
d
p
r
ec
is
io
n
i
n
clu
tter
ed
b
ac
k
g
r
o
u
n
d
s
,
allo
win
g
s
o
m
e
n
o
n
-
f
is
h
im
ag
es
to
p
ass
th
r
o
u
g
h
as f
alse p
o
s
itiv
es.
2
.
3
.
2
.
YO
L
O
v
8
n
T
h
e
YOL
Ov
8
n
h
as
b
ee
n
ex
p
er
im
en
ted
to
co
m
p
ar
e
with
Y
OL
Ov
5
s
f
o
r
f
is
h
v
s
n
o
n
-
f
is
h
d
etec
tio
n
.
T
h
e
aim
is
to
r
em
o
v
e
th
e
b
o
u
n
d
in
g
b
o
x
an
n
o
tatio
n
s
a
n
d
s
im
p
lify
th
e
p
ip
elin
e
w
h
ile
p
r
o
d
u
cin
g
h
ig
h
d
is
cr
im
in
ativ
e
p
er
f
o
r
m
a
n
ce
[
1
4
]
,
[
2
3
]
.
Ke
y
f
ac
t
o
r
s
o
f
YO
L
Ov
8
n
p
ip
elin
e
a
r
e
to
r
e
p
r
esen
t
b
in
ar
y
p
r
o
b
lem
(
f
is
h
an
d
n
o
n
-
f
is
h
)
,
th
e
n
b
o
u
n
d
in
g
b
o
x
an
n
o
tatio
n
s
is
elim
in
ated
,
h
en
ce
s
im
p
lif
y
th
e
d
etec
tio
n
p
ip
elin
e
wh
ile
p
r
o
v
id
es
h
ig
h
d
is
cr
im
in
ativ
e
p
er
f
o
r
m
an
ce
,
o
p
tim
ize
f
o
r
r
ea
l
-
tim
e
d
etec
tio
n
,
an
d
s
u
itab
le
f
o
r
f
is
h
er
ies
ap
p
licatio
n
s
.
YOL
Ov
8
n
also
c
o
m
b
in
e
Fas
ter
Net
b
lo
ck
s
an
d
ef
f
icien
t
m
u
lti
-
s
ca
le
atten
tio
n
(
EMA
)
atten
tio
n
as
t
o
e
n
h
a
n
c
e
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
a
n
d
d
e
t
e
c
t
i
o
n
a
c
c
u
r
a
c
y
,
e
m
p
l
o
y
s
B
iF
P
N
-
b
as
e
d
m
u
l
t
i
-
s
c
a
le
f
e
a
t
u
r
e
f
u
s
i
o
n
t
o
o
p
t
i
m
i
z
e
f
e
at
u
r
e
p
y
r
a
m
i
d
s
a
m
p
l
i
n
g
,
r
e
d
u
c
e
s
p
a
r
a
m
e
te
r
s
f
o
r
ef
f
i
c
i
e
n
t
r
e
al
-
t
i
m
e
i
n
f
e
r
e
n
c
e
i
n
r
e
s
o
u
r
c
e
-
c
o
n
s
t
r
ai
n
e
d
en
v
ir
o
n
m
en
ts
,
ac
h
iev
es
lo
wer
co
m
p
lex
ity
a
n
d
f
aster
in
f
er
e
n
ce
,
an
d
o
f
f
e
r
s
s
tr
o
n
g
p
er
f
o
r
m
an
ce
f
o
r
s
p
ee
d
a
n
d
ac
cu
r
ac
y
,
ac
h
iev
in
g
h
ig
h
e
r
o
v
er
all
p
er
f
o
r
m
an
ce
m
et
r
ics co
m
p
ar
ed
to
YOL
Ov
5
s
[
1
4
]
,
[
2
3
]
.
2
.
4
.
St
a
g
e
-
2
ro
bu
s
t
f
re
s
hn
es
s
cla
s
s
if
ica
t
io
n wit
h T
L
-
M
o
b
ileNet
V2
v
a
ria
nts
Fo
llo
win
g
s
tag
e
-
1
v
alid
atio
n
with
YOL
Ov
8
n
,
o
n
l
y
v
er
if
ie
d
f
is
h
im
ag
es
h
av
e
b
ee
n
p
r
o
ce
s
s
ed
in
to
s
tag
e
-
2
f
o
r
h
ier
ar
c
h
ical
f
r
esh
n
ess
class
if
icat
io
n
in
to
h
ig
h
ly
f
r
esh
,
f
r
esh
,
an
d
n
o
t
f
r
esh
[
1
5
]
,
[
2
6
]
.
T
h
is
s
tag
e
aim
s
to
ex
p
e
r
im
en
t
th
e
lig
h
tw
eig
h
t
Mo
b
ileNetV2
-
b
ased
ar
c
h
itectu
r
e
as
to
im
p
r
o
v
e
e
f
f
icie
n
cy
,
ac
c
u
r
ac
y
,
an
d
s
u
itab
le
f
o
r
h
ig
h
-
d
im
en
s
io
n
al
v
is
u
al
task
s
u
n
d
er
co
m
p
u
tatio
n
al
co
n
s
tr
ain
ts
[
1
8
]
,
[
2
5
]
.
T
h
e
b
aselin
e
Mo
b
ileNetV2
ad
ap
te
d
b
y
r
e
m
o
v
in
g
th
e
o
r
ig
i
n
al
to
p
lay
e
r
s
an
d
in
teg
r
atin
g
c
u
s
to
m
ized
class
if
icatio
n
h
ea
d
co
m
p
r
is
in
g
a
v
er
ag
e
p
o
o
lin
g
,
d
en
s
e
lay
er
s
,
d
r
o
p
o
u
t,
a
n
d
s
o
f
tm
ax
o
u
t
p
u
t.
T
h
e
o
b
jectiv
e
is
to
m
o
d
if
y
th
e
ex
tr
ac
tio
n
o
f
d
o
m
ai
n
-
s
p
ec
if
ic
f
ea
tu
r
es,
m
ain
l
y
f
r
o
m
ey
e,
g
ill
,
an
d
b
o
d
y
r
eg
i
o
n
s
[
1
5
]
–
[
2
7
]
.
T
h
e
to
tal
o
f
s
ev
en
r
ef
in
em
en
ts
T
L
–
M
o
b
ileNetV2
v
ar
ian
ts
h
a
v
e
b
ee
n
d
e
v
elo
p
ed
th
at
in
co
r
p
o
r
ate
C
NN
ex
ten
s
io
n
s
,
f
in
e
-
tu
n
in
g
,
ad
ap
tiv
e
lear
n
in
g
r
ate
s
ch
ed
u
l
er
s
,
an
d
ex
p
an
d
ed
f
u
lly
co
n
n
e
cted
lay
er
s
as
ex
p
er
im
en
ted
in
T
ab
le
4
.
T
h
e
aim
s
ar
e
to
ad
v
a
n
ce
r
e
p
r
esen
tatio
n
lear
n
in
g
,
m
itig
ate
o
v
er
f
itti
n
g
,
an
d
im
p
r
o
v
e
g
en
er
aliza
tio
n
ac
r
o
s
s
h
eter
o
g
en
e
o
u
s
d
atasets
,
th
u
s
estab
lis
h
in
g
a
r
o
b
u
s
t
f
r
esh
n
ess
class
if
ic
atio
n
p
ip
eli
n
e
th
at
o
u
t
p
er
f
o
r
m
s
co
n
v
en
tio
n
al
s
in
g
le
-
m
o
d
el
ap
p
r
o
ac
h
es.
T
h
e
s
ev
en
ex
p
er
im
e
n
tal
r
ef
in
em
e
n
t m
o
d
els ar
e
s
u
m
m
ar
ize
d
as
s
h
o
wn
in
T
ab
le
4
.
T
ab
le
4
.
T
h
e
s
ev
en
p
r
o
p
o
s
ed
T
L
–
Mo
b
ileNetV2
v
ar
ia
n
ts
m
o
d
el
M
o
d
e
l
D
e
scri
p
t
i
o
n
P
e
r
f
o
r
ma
n
c
e
B
a
se
l
i
n
e
M
o
b
i
l
e
N
e
t
V
2
P
r
e
t
r
a
i
n
e
d
o
n
I
mag
e
N
e
t
,
u
s
e
d
a
s
l
i
g
h
t
w
e
i
g
h
t
f
e
a
t
u
r
e
e
x
t
r
a
c
t
o
r
.
A
c
h
i
e
v
e
d
st
r
o
n
g
t
r
a
i
n
i
n
g
a
c
c
u
r
a
c
y
(
9
4
.
2
6
%)
b
u
t
l
o
w
v
a
l
i
d
a
t
i
o
n
a
c
c
u
r
a
c
y
(
6
3
.
6
7
%),
i
n
d
i
c
a
t
i
n
g
o
v
e
r
f
i
t
t
i
n
g
a
n
d
w
e
a
k
g
e
n
e
r
a
l
i
z
a
t
i
o
n
.
TL
-
M
o
b
i
l
e
N
e
t
V
2
I
n
c
o
r
p
o
r
a
t
e
d
t
r
a
n
sf
e
r
l
e
a
r
n
i
n
g
w
i
t
h
p
r
e
t
r
a
i
n
e
d
w
e
i
g
h
t
s
a
n
d
c
u
st
o
m
c
l
a
ss
i
f
i
c
a
t
i
o
n
h
e
a
d
.
I
mp
r
o
v
e
d
v
a
l
i
d
a
t
i
o
n
a
c
c
u
r
a
c
y
t
o
6
7
.
1
1
%,
sh
o
w
i
n
g
b
e
t
t
e
r
a
d
a
p
t
a
b
i
l
i
t
y
t
o
d
o
mai
n
-
sp
e
c
i
f
i
c
f
e
a
t
u
r
e
s.
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
C
o
m
b
i
n
e
c
o
n
v
o
l
u
t
i
o
n
a
l
l
a
y
e
r
s
o
n
t
o
p
o
f
M
o
b
i
l
e
N
e
t
V
2
t
o
c
a
p
t
u
r
e
f
i
n
e
r
l
o
c
a
l
t
e
x
t
u
r
e
s.
B
o
o
st
e
d
t
r
a
i
n
i
n
g
a
c
c
u
r
a
c
y
(
9
5
.
3
9
%)
a
n
d
i
mp
r
o
v
e
d
st
a
b
i
l
i
t
y
i
n
b
o
d
y
a
n
d
g
i
l
l
d
a
t
a
set
s
.
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
+
f
i
n
e
-
t
u
n
i
n
g
C
o
m
b
i
n
e
d
C
N
N
e
x
t
e
n
s
i
o
n
s
w
i
t
h
sel
e
c
t
i
v
e
f
i
n
e
-
t
u
n
i
n
g
o
f
t
o
p
M
o
b
i
l
e
N
e
t
V
2
l
a
y
e
r
s.
A
c
h
i
e
v
e
d
t
h
e
h
i
g
h
t
r
a
i
n
i
n
g
a
c
c
u
r
a
c
y
(
9
8
.
0
7
%)
a
n
d
h
i
g
h
e
s
t
v
a
l
i
d
a
t
i
o
n
a
c
c
u
r
a
c
y
(
6
7
.
7
2
%),
d
e
m
o
n
st
r
a
t
i
n
g
i
mp
r
o
v
e
d
f
e
a
t
u
r
e
d
i
scr
i
mi
n
a
t
i
o
n
a
n
d
r
o
b
u
st
n
e
ss a
c
r
o
ss
a
u
g
m
e
n
t
e
d
d
a
t
a
set
s.
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
+
l
e
a
r
n
i
n
g
r
a
t
e
sch
e
d
u
l
e
r
I
n
t
r
o
d
u
c
e
d
d
y
n
a
mi
c
s
c
h
e
d
u
l
e
r
s
t
o
st
a
b
i
l
i
z
e
c
o
n
v
e
r
g
e
n
c
e
a
n
d
r
e
d
u
c
e
l
o
ss
o
sc
i
l
l
a
t
i
o
n
s
.
R
e
a
c
h
e
d
v
a
l
i
d
a
t
i
o
n
a
c
c
u
r
a
c
y
o
f
6
6
.
7
8
%,
sh
o
w
i
n
g
b
a
l
a
n
c
e
d
g
e
n
e
r
a
l
i
z
a
t
i
o
n
a
c
r
o
ss
d
a
t
a
se
t
s.
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
+
a
d
d
i
t
i
o
n
a
l
f
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
s
Ex
p
a
n
d
e
d
f
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
s
t
o
i
n
c
r
e
a
s
e
r
e
p
r
e
se
n
t
a
t
i
o
n
a
l
d
e
p
t
h
.
I
mp
r
o
v
e
d
p
e
r
f
o
r
ma
n
c
e
mo
d
e
r
a
t
e
l
y
(
6
5
.
1
0
%
v
a
l
i
d
a
t
i
o
n
a
c
c
u
r
a
c
y
)
,
b
u
t
w
i
t
h
d
i
mi
n
i
sh
i
n
g
r
e
t
u
r
n
s
a
n
d
h
i
g
h
e
r
c
o
m
p
u
t
a
t
i
o
n
a
l
c
o
s
t
.
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
+
f
i
n
e
-
t
u
n
i
n
g
+
a
d
a
p
t
i
v
e
l
e
a
r
n
i
n
g
r
a
t
e
sc
h
e
d
u
l
e
r
I
n
t
e
g
r
a
t
e
d
a
l
l
r
e
f
i
n
e
me
n
t
s
i
n
t
o
t
h
e
mo
st
a
d
v
a
n
c
e
d
c
o
n
f
i
g
u
r
a
t
i
o
n
.
A
c
h
i
e
v
e
d
t
h
e
h
i
g
h
e
st
t
r
a
i
n
i
n
g
a
c
c
u
r
a
c
y
(
9
8
.
6
3
%)
a
n
d
st
r
o
n
g
v
a
l
i
d
a
t
i
o
n
a
c
c
u
r
a
c
y
(
6
7
.
0
1
%),
w
i
t
h
r
o
b
u
s
t
n
e
ss.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mu
lti
-
s
ta
g
e
h
yb
r
id
YOLO
-
d
r
iv
en
a
n
d
Mo
b
ileN
etV
2
-
C
N
N
va
r
ia
n
ts
fo
r
r
o
b
u
s
t
…
(
R
a
s
ee
d
a
H
a
mza
h
)
3665
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
is
s
ec
tio
n
p
r
esen
ts
th
e
ex
p
er
im
en
tal
f
in
d
in
g
s
o
f
th
e
p
r
o
p
o
s
ed
m
u
lti
-
s
tag
e
h
y
b
r
id
r
ep
r
esen
tatio
n
lear
n
in
g
f
r
am
ewo
r
k
f
o
r
r
o
b
u
s
t
f
is
h
f
r
esh
n
ess
class
if
icatio
n
[
4
]
,
[
1
2
]
,
[
1
9
]
.
T
h
e
d
etailed
ex
p
lan
atio
n
o
f
s
tag
e
-
1
f
ea
tu
r
e
-
g
ated
d
etec
tio
n
with
YOL
Ov
5
s
an
d
YOL
Ov
8
n
,
s
tag
e
-
2
r
o
b
u
s
t
f
r
esh
n
ess
c
lass
if
icatio
n
u
s
in
g
Mo
b
ileNetV2
v
ar
ian
ts
as
f
o
ll
o
ws.
T
h
e
a
n
aly
s
is
r
esu
lts
ar
e
p
r
esen
ted
b
elo
w
with
ac
c
u
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e.
T
h
e
co
m
p
ar
ativ
e
ev
alu
atio
n
s
with
p
r
i
o
r
s
tu
d
ie
s
ar
e
also
d
is
cu
s
s
ed
[
1
0
]
,
[
1
3
]
,
[
2
1
]
,
[
2
6
]
.
3
.
1
.
St
a
g
e
-
1
f
ea
t
ure
-
g
a
t
ed
d
et
ec
t
io
n
v
ia
YO
L
O
v
5
s
–
YO
L
O
v
8
n:
f
is
h v
er
s
us
no
n
-
f
is
h
T
h
e
f
ir
s
t
s
tag
e
o
f
th
e
p
ip
elin
e
ev
alu
ated
YOL
Ov
5
s
an
d
Y
OL
Ov
8
n
as
co
m
p
etin
g
ar
c
h
itectu
r
es
f
o
r
f
ea
tu
r
e
-
g
ated
d
etec
tio
n
o
f
f
is
h
v
er
s
u
s
n
o
n
-
f
is
h
class
if
icatio
n
p
r
io
r
to
f
r
esh
n
ess
an
aly
s
is
[
1
4
]
,
[
1
6
]
.
B
ased
o
n
T
ab
le
5
,
it
is
s
h
o
wn
th
at
YO
L
Ov
8
n
c
o
n
s
is
ten
tly
o
u
t
p
er
f
o
r
m
ed
YOL
Ov
5
s
ac
r
o
s
s
all
k
ey
ev
alu
atio
n
m
etr
ics.
T
h
is
f
ea
tu
r
e
-
g
ate
d
r
ep
r
esen
ts
p
r
e
-
f
ilter
th
at
ef
f
ec
tiv
ely
s
u
p
p
r
ess
es
b
ac
k
g
r
o
u
n
d
d
is
tr
ac
to
r
s
,
m
ar
k
et
en
v
ir
o
n
m
en
t
n
o
is
e,
an
d
n
o
n
-
f
i
s
h
o
b
jects,
th
er
eb
y
to
p
r
ev
e
n
t f
alse
-
p
o
s
itiv
e
er
r
o
r
p
r
o
p
a
g
atio
n
in
to
s
tag
e
-
2
[
1
6
]
.
Fu
r
th
er
m
o
r
e
,
YOL
Ov
8
n
d
em
o
n
s
tr
ated
f
aster
in
f
er
en
ce
s
p
e
ed
s
an
d
r
ed
u
ce
d
p
ar
am
eter
c
o
m
p
lex
ity
co
m
p
a
r
ed
to
YOL
Ov
5
s
,
th
er
e
b
y
v
alid
f
o
r
p
r
ac
tical
s
u
itab
ilit
y
f
o
r
r
ea
l
-
tim
e
ed
g
e
d
e
p
lo
y
m
e
n
t
in
f
is
h
er
ies
q
u
ality
m
o
n
ito
r
in
g
[
1
4
]
,
[
2
3
]
.
T
ab
le
5
.
YOL
OV5
s
an
d
YOL
Ov
8
n
Ep
o
c
h
Y
O
LO
v
5
s
Y
O
LO
v
8
n
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
Tr
a
i
n
i
n
g
a
c
c
u
r
a
c
y
Tr
a
i
n
i
n
g
l
o
ss
V
a
l
i
d
a
t
i
o
n
l
o
ss
0
0
.
1
8
0
3
6
0
.
2
5
8
9
7
0
.
9
8
3
7
0
.
1
9
0
6
6
0
.
0
7
3
9
8
1
0
.
4
8
6
5
0
.
4
7
8
1
9
0
.
9
9
4
5
7
0
.
0
4
6
3
6
0
.
0
8
1
5
2
2
0
.
4
9
7
8
9
0
.
4
5
9
4
5
0
.
9
8
9
1
3
0
.
0
4
3
7
3
0
.
0
6
1
5
1
3
0
.
6
4
5
2
9
0
.
5
5
5
6
7
0
.
9
9
4
5
7
0
.
0
3
1
1
4
0
.
0
7
0
2
5
4
0
.
7
6
7
4
5
0
.
6
2
7
6
6
0
.
9
9
4
5
7
0
.
0
1
5
2
2
0
.
0
7
8
8
3
3
.
2
.
St
a
g
e
-
2
ro
bu
s
t
f
re
s
hn
es
s
cla
s
s
if
ica
t
io
n o
n 8
0
:
2
0
s
pli
t
:
ba
s
eline
t
o
hy
brids
T
h
e
a
i
m
o
f
s
t
a
g
e
-
2
i
s
t
o
b
en
ch
m
a
r
k
th
e
e
f
f
e
c
t
i
v
en
e
s
s
o
f
p
r
o
g
r
e
s
s
i
v
e
Mo
b
i
l
e
Ne
t
V
2
e
n
h
an
c
e
m
en
t
s
u
n
d
e
r
an
8
0
:
2
0
t
r
a
in
i
n
g
–
v
a
l
i
d
a
t
i
o
n
s
p
l
i
t,
f
r
o
m
th
e
b
a
s
e
l
i
n
e
b
a
c
k
b
o
n
e
t
o
h
y
b
r
id
v
a
r
i
an
t
s
[
1
5
]
,
[
2
6
]
.
T
ab
l
e
6
s
h
o
w
s
av
e
r
a
g
e
t
r
a
in
in
g
a
n
d
v
a
l
i
d
a
t
io
n
a
cc
u
r
a
c
y
f
o
r
Mo
b
il
e
N
e
t
V
2
v
ar
i
a
n
t
s
m
o
d
e
l
p
e
r
f
o
r
m
an
c
e.
B
a
s
e
d
o
n
T
a
b
l
e
6
,
b
a
s
e
l
i
n
e
M
o
b
i
l
eN
e
t
V
2
a
ch
i
e
v
ed
6
3
.
6
7
%
v
a
l
i
d
a
t
io
n
a
c
cu
r
a
cy
t
r
an
s
f
e
r
l
e
ar
n
in
g
an
d
C
N
N
i
n
t
e
g
r
a
t
i
o
n
p
r
o
v
id
ed
in
c
r
em
e
n
t
a
l
g
a
i
n
s
.
F
i
n
e
-
t
u
n
i
n
g
f
u
r
th
e
r
i
m
p
r
o
v
ed
s
t
a
b
i
l
i
t
y
,
r
e
a
ch
i
n
g
6
7
.
7
2
%
,
w
h
i
l
e
t
h
e
h
y
b
r
i
d
w
i
th
a
d
a
p
ti
v
e
l
e
a
r
n
i
n
g
r
a
t
e
s
ch
e
d
u
l
i
n
g
a
c
h
i
e
v
e
d
t
h
e
b
e
s
t
t
r
ad
e
-
o
f
f
b
e
t
w
e
e
n
tr
a
i
n
in
g
a
c
c
u
r
a
c
y
(
9
8
.
6
3
%)
an
d
v
a
li
d
a
t
i
o
n
a
cc
u
r
a
c
y
(
6
7
.
0
1
%)
.
T
h
e
s
e
r
e
s
u
l
t
s
h
ig
h
l
ig
h
t
th
e
s
t
a
g
e
-
2
m
o
d
e
l
p
a
r
t
i
cu
l
a
r
ly
C
N
N
ex
t
e
n
s
i
o
n
s
a
n
d
a
d
ap
t
i
v
e
s
ch
e
d
u
l
er
s
,
m
it
i
g
a
t
e
o
v
e
r
f
i
t
t
i
n
g
a
n
d
s
t
r
e
n
g
th
e
n
g
e
n
er
a
l
i
z
a
t
io
n
c
o
m
p
a
r
ed
to
t
h
e
b
a
s
e
l
i
n
e.
T
ab
le
6
.
Av
e
r
ag
e
tr
ain
in
g
an
d
v
alid
atio
n
ac
cu
r
ac
y
f
o
r
Mo
b
ileNetV2
v
ar
ian
ts
m
o
d
el
p
er
f
o
r
m
an
ce
M
o
d
e
l
Tr
a
i
n
i
n
g
a
c
c
u
r
a
cy
V
a
l
i
d
a
t
i
o
n
a
c
c
u
r
a
c
y
B
a
se
l
i
n
e
M
o
b
i
l
e
N
e
t
V
2
9
4
.
2
6
6
3
.
6
7
TL
-
M
o
b
i
l
e
N
e
t
V
2
9
3
.
0
3
6
7
.
1
1
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
9
5
.
3
9
6
4
.
7
2
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
+
f
i
n
e
-
t
u
n
i
n
g
9
8
.
0
7
6
7
.
7
2
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
+
l
e
a
r
n
i
n
g
r
a
t
e
sc
h
e
d
u
l
e
r
9
6
.
8
1
6
6
.
7
8
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
+
A
d
d
i
t
i
o
n
a
l
F
C
l
a
y
e
r
s
9
5
.
2
5
6
5
.
1
0
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
+
f
i
n
e
-
t
u
n
i
n
g
+
a
d
a
p
t
i
v
e
l
e
a
r
n
i
n
g
r
a
t
e
sc
h
e
d
u
l
e
r
9
8
.
6
3
6
7
.
0
1
3
.
3
.
St
a
g
e
-
2
ev
a
lua
t
i
o
n m
et
r
ics f
o
r
M
o
bil
eNe
t
V2
v
a
ria
nts a
cr
o
s
s
da
t
a
s
et
T
h
e
a
i
m
o
f
th
i
s
s
e
c
t
i
o
n
i
s
t
o
e
v
a
lu
a
t
e
an
d
c
o
m
p
ar
e
t
h
e
p
er
f
o
r
m
a
n
c
e
o
f
T
L
-
Mo
b
i
l
e
N
e
t
V
2
v
a
r
i
an
t
s
a
cr
o
s
s
v
ar
i
o
u
s
d
a
t
as
e
t
s
.
T
h
e
s
e
e
v
a
lu
a
t
i
o
n
s
h
a
v
e
b
e
en
co
n
d
u
c
te
d
to
d
e
t
e
r
m
i
n
e
th
e
ef
f
ec
t
o
f
a
r
c
h
i
t
ec
t
u
r
a
l
r
ef
i
n
e
m
en
t
s
,
o
p
t
i
m
iz
a
t
i
o
n
s
t
r
a
t
eg
i
e
s
,
an
d
t
r
a
in
i
n
g
d
u
r
a
t
io
n
o
n
t
h
e
r
o
b
u
s
t
n
e
s
s
a
n
d
g
e
n
er
a
l
i
z
a
t
io
n
o
f
t
h
e
f
r
e
s
h
n
es
s
c
l
a
s
s
i
f
i
c
a
t
i
o
n
s
y
s
t
e
m
[
1
5
]
,
[
2
5
]
,
[
2
6
]
.
T
h
i
s
s
e
c
t
i
o
n
e
x
p
e
r
i
m
en
t
e
d
f
o
r
e
ac
h
v
a
r
i
an
t
f
o
r
t
r
a
in
i
n
g
an
d
v
a
l
id
a
t
i
o
n
o
n
h
e
t
er
o
g
e
n
e
o
u
s
d
a
ta
s
e
t
s
c
o
n
s
i
s
t
i
n
g
o
f
b
o
d
y
,
ey
e
,
a
n
d
g
i
l
l
i
m
ag
e
s
a
c
r
o
s
s
t
h
r
e
e
f
r
e
s
h
n
e
s
s
l
e
v
e
l
s
(
h
i
g
h
l
y
f
r
e
s
h
,
f
r
e
s
h
,
an
d
n
o
t
f
r
e
s
h
)
,
a
n
d
a
c
r
o
s
s
5
,
1
0
,
an
d
1
5
e
p
o
c
h
s
,
f
o
cu
s
i
n
g
o
n
t
r
a
i
n
in
g
ac
c
u
r
a
c
y
,
v
al
i
d
a
t
io
n
a
cc
u
r
a
c
y
,
a
n
d
g
e
n
er
a
l
i
z
a
t
io
n
t
r
e
n
d
s
b
a
s
e
d
o
n
s
e
v
en
r
ef
i
n
e
m
en
t
s
v
ar
i
a
n
t
s
[
1
8
]
,
[
2
7
]
a
s
s
h
o
w
n
in
T
a
b
l
e
s
7
t
o
14
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
4
,
Au
g
u
s
t 2
0
2
6
:
3
6
6
0
-
3
6
7
1
3666
B
ased
o
n
th
e
T
a
b
le
1
4
,
t
h
e
b
aselin
e
Mo
b
ileNetV2
ac
h
ie
v
ed
m
o
d
est
ac
c
u
r
ac
y
(
9
2
.
2
3
%
tr
ain
in
g
,
6
2
.
5
4
%
v
alid
ati
o
n
)
,
wh
ile
tr
an
s
f
er
lear
n
in
g
i
m
p
r
o
v
ed
v
al
id
atio
n
to
6
5
.
9
0
%.
T
h
e
C
NN
lay
er
s
ex
ten
s
io
n
p
r
o
d
u
ce
d
s
m
all
s
ig
n
if
ican
t,
wh
ile
in
co
r
p
o
r
atin
g
f
in
e
-
tu
n
i
n
g
s
ig
n
if
ican
tly
p
r
o
d
u
ce
d
v
alid
atio
n
ac
cu
r
ac
y
to
6
7
.
8
8
%
f
r
o
m
T
L
-
M
o
b
il
eNe
tV2
.
T
h
e
lear
n
in
g
r
ate
s
ch
ed
u
ler
m
ain
tain
ed
b
alan
ce
d
p
er
f
o
r
m
a
n
ce
(
9
3
.
3
9
%
tr
ain
in
g
,
6
5
.
5
2
%
v
a
lid
atio
n
)
,
wh
er
ea
s
ad
d
itio
n
al
f
u
lly
co
n
n
ec
ted
lay
er
s
led
to
o
v
er
f
itti
n
g
with
r
ed
u
ce
d
v
alid
atio
n
(
6
1
.
8
9
%).
T
h
e
b
est
p
er
f
o
r
m
a
n
ce
m
o
d
e
l,
T
L
-
Mo
b
ileNetV2
with
C
NN,
an
d
f
in
e
-
tu
n
in
g
,
r
esu
lted
9
4
.
9
7
%
±
0
.
0
4
tr
ai
n
in
g
a
n
d
6
7
.
8
8
%
±
2
.
1
0
v
a
lid
atio
n
ac
cu
r
a
cy
.
T
h
u
s
,
f
in
e
-
tu
n
in
g
in
teg
r
atio
n
en
h
an
ce
r
o
b
u
s
tn
ess
m
o
r
e
ef
f
e
ctiv
ely
th
an
in
cr
ea
s
ed
ar
c
h
itectu
r
al
d
ep
th
.
T
h
is
m
o
d
el
d
em
o
n
s
tr
ates
b
o
th
s
tr
o
n
g
ac
cu
r
ac
y
an
d
s
tab
le
co
n
v
er
g
e
n
ce
.
T
ab
le
7
.
Stag
e
-
2
–
tr
ain
in
g
,
an
d
v
alid
atio
n
ac
cu
r
ac
y
f
o
r
b
aselin
e
Mo
b
ileNetV2
m
o
d
el
p
er
f
o
r
m
an
ce
D
a
t
a
s
e
t
5
e
p
o
c
h
s
1
0
e
p
o
c
h
s
1
5
e
p
o
c
h
s
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
F
i
sh
f
r
e
sh
n
e
ss
c
l
a
ss
i
f
i
c
a
t
i
o
n
d
a
t
a
s
e
t
1
0
0
1
0
0
1
0
0
1
0
0
1
0
0
1
0
0
Th
e
f
r
e
sh
n
e
ss
o
f
t
h
e
f
i
s
h
e
y
e
s
d
a
t
a
se
t
6
5
.
0
2
4
1
.
5
1
7
0
.
8
8
4
1
.
1
6
7
2
.
9
3
4
1
.
8
5
P
r
i
mary
d
a
t
a
set
–
b
o
d
y
8
8
.
4
5
5
4
.
7
9
9
2
.
3
9
6
7
.
0
2
9
3
.
8
3
6
7
.
5
5
P
r
i
mary
d
a
t
a
set
–
e
y
e
s
9
2
.
4
9
5
2
.
8
8
9
6
.
8
9
4
9
.
2
1
9
8
.
8
3
5
3
.
4
0
P
r
i
mary
d
a
t
a
set
–
g
i
l
l
s
9
0
.
7
4
5
6
.
8
4
9
7
.
7
8
6
0
.
0
0
9
8
.
1
7
5
7
.
3
7
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
b
o
d
y
9
0
.
2
8
7
2
.
3
1
9
2
.
4
6
7
4
.
6
4
9
3
.
9
9
7
4
.
6
4
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
e
y
e
s
9
4
.
7
0
5
0
.
9
3
9
6
.
6
8
5
6
.
3
5
9
8
.
2
9
5
6
.
7
2
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
g
i
l
l
s
9
3
.
6
6
5
6
.
3
1
9
6
.
9
0
5
7
.
7
3
9
8
.
0
6
5
7
.
8
0
A
v
e
r
a
g
e
8
9
.
4
2
6
0
.
7
0
9
3
.
0
0
6
3
.
2
6
9
4
.
2
6
6
3
.
6
7
T
ab
le
8
.
Stag
e
-
2
–
tr
ain
in
g
,
an
d
v
alid
atio
n
ac
cu
r
ac
y
f
o
r
T
L
–
Mo
b
ileNetV2
m
o
d
el
p
e
r
f
o
r
m
a
n
ce
D
a
t
a
s
e
t
5
e
p
o
c
h
s
1
0
e
p
o
c
h
s
1
5
e
p
o
c
h
s
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
F
i
sh
f
r
e
sh
n
e
ss
c
l
a
ss
i
f
i
c
a
t
i
o
n
d
a
t
a
s
e
t
9
9
.
7
8
1
0
0
1
0
0
1
0
0
1
0
0
1
0
0
Th
e
f
r
e
sh
n
e
s
s
o
f
t
h
e
f
i
s
h
e
y
e
s
d
a
t
a
se
t
5
9
.
7
8
4
1
.
7
3
6
6
.
3
0
4
3
.
7
9
7
0
.
2
0
4
3
.
3
3
P
r
i
mary
d
a
t
a
set
–
b
o
d
y
8
9
.
3
7
6
4
.
8
9
9
3
.
1
8
6
4
.
8
9
9
5
.
1
4
7
0
.
2
1
P
r
i
mary
d
a
t
a
set
–
e
y
e
s
9
2
.
3
6
5
4
.
4
5
9
8
.
0
6
5
4
.
9
7
9
9
.
2
2
5
8
.
1
8
P
r
i
mary
d
a
t
a
set
–
g
i
l
l
s
9
3
.
6
1
6
1
.
0
5
9
8
.
8
3
6
4
.
7
4
9
9
.
4
8
6
7
.
3
7
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
b
o
d
y
8
9
.
7
8
7
5
.
4
7
9
2
.
3
3
7
5
.
5
5
9
2
.
9
9
7
6
.
9
0
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
e
y
e
s
9
3
.
1
8
5
6
.
9
4
9
6
.
5
7
5
9
.
9
1
9
7
.
2
6
5
5
.
7
5
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
g
i
l
l
s
9
3
.
1
9
6
1
.
0
9
9
6
.
8
7
6
5
.
2
0
9
7
.
1
1
6
5
.
1
2
A
v
e
r
a
g
e
8
8
.
8
8
6
4
.
4
5
9
2
.
7
7
6
6
.
1
3
9
3
.
9
3
6
7
.
1
1
T
ab
le
9
.
Stag
e
-
2
–
tr
ain
in
g
,
an
d
v
alid
atio
n
ac
cu
r
ac
y
f
o
r
T
L
–
Mo
b
ileNetV2
+
C
NN
m
o
d
el
p
er
f
o
r
m
a
n
ce
D
a
t
a
s
e
t
5
e
p
o
c
h
s
1
0
e
p
o
c
h
s
1
5
e
p
o
c
h
s
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
F
i
sh
f
r
e
sh
n
e
ss
c
l
a
ss
i
f
i
c
a
t
i
o
n
d
a
t
a
s
e
t
1
0
0
1
0
0
1
0
0
9
9
.
8
9
9
9
.
6
9
9
9
.
4
4
Th
e
f
r
e
sh
n
e
ss
o
f
t
h
e
f
i
s
h
e
y
e
s
d
a
t
a
se
t
5
1
.
6
4
4
2
.
4
2
6
1
.
2
0
4
1
.
2
8
7
9
.
0
5
4
1
.
9
6
P
r
i
mary
d
a
t
a
set
–
b
o
d
y
8
7
.
9
0
7
0
.
7
4
9
1
.
3
8
6
5
.
9
6
9
4
.
0
9
6
8
.
6
2
P
r
i
mary
d
a
t
a
set
–
e
y
e
s
9
5
.
0
8
6
0
.
2
1
9
9
.
6
1
5
2
.
8
8
9
9
.
8
7
5
3
.
9
3
P
r
i
mary
d
a
t
a
set
–
g
i
l
l
s
9
4
.
5
2
5
7
.
8
9
9
8
.
1
7
6
5
.
2
6
9
8
.
7
0
6
9
.
4
7
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
b
o
d
y
9
0
.
6
2
6
8
.
1
0
9
2
.
6
0
6
7
.
7
2
9
4
.
7
8
7
0
.
5
8
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
e
y
e
s
9
5
.
2
4
5
6
.
5
7
9
7
.
2
6
5
4
.
7
9
9
8
.
3
9
5
3
.
4
5
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
g
i
l
l
s
9
3
.
0
0
5
8
.
0
3
9
7
.
8
9
5
6
.
8
3
9
8
.
5
8
6
0
.
2
7
A
v
e
r
a
g
e
8
8
.
5
0
6
4
.
2
5
9
2
.
2
6
6
3
.
0
8
9
5
.
3
9
6
4
.
7
2
T
ab
le
1
0
.
Stag
e
-
2
–
tr
ain
i
n
g
,
a
n
d
v
alid
atio
n
ac
c
u
r
ac
y
f
o
r
T
L
–
Mo
b
ileNetV2
+
C
NN
+
f
in
e
-
tu
n
in
g
m
o
d
el
p
er
f
o
r
m
a
n
ce
D
a
t
a
s
e
t
5
e
p
o
c
h
s
1
0
e
p
o
c
h
s
1
5
e
p
o
c
h
s
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
F
i
sh
f
r
e
sh
n
e
ss
c
l
a
ss
i
f
i
c
a
t
i
o
n
d
a
t
a
s
e
t
9
9
.
9
2
9
9
.
8
9
1
0
0
1
0
0
9
9
.
9
7
1
0
0
Th
e
f
r
e
sh
n
e
ss
o
f
t
h
e
f
i
s
h
e
y
e
s
d
a
t
a
se
t
7
3
.
3
8
4
0
.
2
5
7
0
.
0
5
4
1
.
5
1
9
3
.
8
2
4
1
.
6
2
P
r
i
mary
d
a
t
a
set
–
b
o
d
y
9
4
.
6
2
6
7
.
5
5
9
4
.
3
6
6
5
.
9
6
7
0
.
7
4
9
5
.
1
4
P
r
i
mary
d
a
t
a
set
–
e
y
e
s
9
9
.
2
2
6
0
.
7
3
9
9
.
8
7
7
0
.
1
6
9
7
.
6
7
6
0
.
7
3
P
r
i
mary
d
a
t
a
set
–
g
i
l
l
s
9
9
.
2
2
5
9
.
4
7
9
9
.
8
1
6
2
.
1
1
1
0
0
6
5
.
7
9
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
b
o
d
y
9
5
.
4
9
7
2
.
4
6
9
7
.
5
2
6
7
.
1
2
9
8
.
5
3
7
4
.
8
7
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
e
y
e
s
9
9
.
3
0
6
0
.
7
3
9
9
.
2
0
5
8
.
5
7
9
9
.
7
8
6
1
.
2
5
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
g
i
l
l
s
9
8
.
3
6
6
7
.
2
1
9
8
.
7
5
6
9
.
5
3
9
9
.
6
6
6
6
.
7
7
A
v
e
r
a
g
e
9
4
.
9
4
6
6
.
0
0
9
4
.
9
5
6
6
.
8
7
9
5
.
0
2
7
0
.
7
7
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mu
lti
-
s
ta
g
e
h
yb
r
id
YOLO
-
d
r
iv
en
a
n
d
Mo
b
ileN
etV
2
-
C
N
N
va
r
ia
n
ts
fo
r
r
o
b
u
s
t
…
(
R
a
s
ee
d
a
H
a
mza
h
)
3667
T
ab
le
1
1
.
Stag
e
-
2
–
tr
ain
i
n
g
,
a
n
d
v
alid
atio
n
ac
c
u
r
ac
y
f
o
r
T
L
–
Mo
b
ileNetV2
+
C
NN
+
lear
n
in
g
r
ate
s
ch
ed
u
ler
m
o
d
el
p
er
f
o
r
m
a
n
ce
D
a
t
a
s
e
t
5
e
p
o
c
h
s
1
0
e
p
o
c
h
s
1
5
e
p
o
c
h
s
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
F
i
sh
f
r
e
sh
n
e
ss
c
l
a
ss
i
f
i
c
a
t
i
o
n
d
a
t
a
s
e
t
9
9
.
9
7
9
9
.
8
9
1
0
0
1
0
0
1
0
0
1
0
0
Th
e
f
r
e
sh
n
e
ss
o
f
t
h
e
f
i
s
h
e
y
e
s
d
a
t
a
se
t
5
1
.
3
8
4
0
.
9
4
7
5
.
1
8
4
2
.
4
2
8
3
.
6
3
4
1
.
5
1
P
r
i
mary
d
a
t
a
set
–
b
o
d
y
8
5
.
8
3
6
2
.
7
7
9
3
.
1
8
6
7
.
0
2
9
5
.
6
7
7
0
.
7
4
P
r
i
mary
d
a
t
a
set
–
e
y
e
s
9
4
.
8
2
6
1
.
2
6
9
9
.
2
2
5
4
.
4
5
1
0
0
6
0
.
2
1
P
r
i
mary
d
a
t
a
set
–
g
i
l
l
s
9
5
.
5
7
5
7
.
8
9
9
7
.
3
9
6
1
.
0
5
9
9
.
6
1
6
3
.
1
6
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
b
o
d
y
8
9
.
9
5
7
3
.
9
7
9
4
.
6
4
7
2
.
4
6
9
6
.
7
3
7
7
.
0
5
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
e
y
e
s
9
6
.
4
6
5
7
.
5
4
9
9
.
8
1
6
1
.
4
0
9
9
.
5
0
5
5
.
8
3
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
g
i
l
l
s
9
4
.
3
7
6
0
.
6
4
9
9
.
1
0
6
4
.
5
3
9
9
.
3
3
6
5
.
8
0
A
v
e
r
a
g
e
8
8
.
5
4
6
4
.
3
6
9
4
.
8
2
6
5
.
4
2
9
6
.
8
1
6
6
.
7
9
T
ab
le
1
2
.
Stag
e
-
2
–
tr
ain
i
n
g
,
a
n
d
v
alid
atio
n
ac
c
u
r
ac
y
f
o
r
T
L
-
Mo
b
ileNetV2
+
C
NN
+
ad
d
itio
n
al
f
u
lly
co
n
n
ec
ted
la
y
er
s
m
o
d
el
p
er
f
o
r
m
an
ce
D
a
t
a
s
e
t
5
e
p
o
c
h
s
1
0
e
p
o
c
h
s
1
5
e
p
o
c
h
s
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
F
i
sh
f
r
e
sh
n
e
ss
c
l
a
ss
i
f
i
c
a
t
i
o
n
d
a
t
a
s
e
t
9
9
.
8
9
9
9
.
6
6
9
9
.
8
6
9
9
.
6
6
1
0
0
9
9
.
8
9
Th
e
f
r
e
sh
n
e
ss
o
f
t
h
e
f
i
s
h
e
y
e
s
d
a
t
a
se
t
5
0
.
9
0
4
1
.
1
6
6
9
.
3
1
4
1
.
6
2
8
1
.
2
1
4
5
.
7
2
P
r
i
mary
d
a
t
a
set
–
b
o
d
y
5
6
.
9
6
5
6
.
3
8
8
5
.
1
7
5
7
.
9
8
9
0
.
9
4
6
1
.
7
0
P
r
i
mary
d
a
t
a
set
–
e
y
e
s
6
6
.
5
8
3
8
.
2
2
8
3
.
4
2
5
1
.
8
3
9
7
.
0
2
5
7
.
0
7
P
r
i
mary
d
a
t
a
set
–
g
i
l
l
s
5
3
.
3
2
4
2
.
1
1
8
4
.
3
5
4
8
.
9
5
9
7
.
6
5
6
4
.
2
1
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
b
o
d
y
8
7
.
7
8
7
4
.
2
7
9
4
.
7
2
7
0
.
3
5
9
6
.
8
1
7
4
.
5
7
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
e
y
e
s
9
2
.
6
4
5
7
.
0
2
9
8
.
1
8
5
9
.
2
4
9
9
.
2
2
5
2
.
1
9
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
g
i
l
l
s
9
3
.
4
9
6
0
.
8
7
9
7
.
9
5
6
5
.
3
5
9
9
.
1
2
6
5
.
4
7
A
v
e
r
a
g
e
7
5
.
2
0
5
8
.
7
1
8
9
.
1
2
6
1
.
8
7
9
5
.
2
5
6
5
.
1
0
T
ab
le
1
3
.
Stag
e
-
2
–
tr
ain
i
n
g
,
a
n
d
v
alid
atio
n
ac
c
u
r
ac
y
f
o
r
T
L
-
Mo
b
ileNetV2
+
C
NN
+
f
in
e
-
tu
n
in
g
+
ad
ap
tiv
e
lear
n
in
g
r
ate
s
ch
e
d
u
ler
m
o
d
el
p
er
f
o
r
m
an
ce
D
a
t
a
s
e
t
5
e
p
o
c
h
s
1
0
e
p
o
c
h
s
1
5
e
p
o
c
h
s
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
F
i
sh
f
r
e
sh
n
e
ss
c
l
a
ss
i
f
i
c
a
t
i
o
n
d
a
t
a
s
e
t
9
9
.
9
4
1
0
0
1
0
0
9
9
.
8
9
1
0
0
1
0
0
Th
e
f
r
e
sh
n
e
ss
o
f
t
h
e
f
i
s
h
e
y
e
s
d
a
t
a
se
t
7
3
.
8
7
4
4
.
8
1
8
7
.
5
3
3
9
.
0
0
9
4
.
7
3
3
9
.
4
5
P
r
i
mary
d
a
t
a
set
–
b
o
d
y
9
5
.
6
7
6
7
.
0
2
9
4
.
8
8
6
5
.
4
3
9
5
.
1
4
7
6
.
0
6
P
r
i
mary
d
a
t
a
set
–
e
y
e
s
9
9
.
0
9
5
7
.
5
9
1
0
0
6
0
.
2
1
9
9
.
7
4
5
1
.
3
1
P
r
i
mary
d
a
t
a
set
–
g
i
l
l
s
9
9
.
8
7
5
6
.
3
2
1
0
0
6
2
.
6
3
1
0
0
6
7
.
3
7
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
b
o
d
y
9
6
.
8
4
6
8
.
6
2
9
7
.
5
8
6
6
.
3
7
9
9
.
6
8
6
9
.
2
2
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
e
y
e
s
9
9
.
0
2
6
2
.
5
1
9
9
.
8
1
5
9
.
6
9
9
9
.
9
3
5
8
.
6
5
P
r
i
mary
d
a
t
a
set
w
i
t
h
d
a
t
a
a
u
g
me
n
t
a
t
i
o
n
–
g
i
l
l
s
9
9
.
2
7
6
0
.
4
6
9
9
.
7
6
6
8
.
9
8
9
9
.
8
1
7
3
.
9
9
A
v
e
r
a
g
e
9
5
.
4
5
6
4
.
6
7
9
7
.
4
4
6
5
.
2
1
9
8
.
6
3
6
7
.
0
1
T
ab
le
1
4
.
Stag
e
-
2
f
o
r
T
L
-
M
o
b
ileNetV2
v
ar
ian
ts
m
o
d
el
–
av
e
r
ag
e
±
s
tan
d
ar
d
d
ev
iatio
n
(
SD)
M
o
d
e
l
Tr
a
i
n
i
n
g
±
S
D
V
a
l
i
d
a
t
i
o
n
±
S
D
Ep
o
c
h
s
B
a
se
l
i
n
e
M
o
b
i
l
e
N
e
t
V
2
9
2
.
2
3
±
2
.
0
6
6
2
.
5
4
±
1
.
3
2
5
,
1
0
,
1
5
TL
-
M
o
b
i
l
e
N
e
t
V
2
9
1
.
8
6
±
2
.
1
5
6
5
.
9
0
±
1
.
1
0
5
,
1
0
,
1
5
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
9
2
.
0
5
±
2
.
8
2
6
4
.
0
2
±
0
.
6
7
5
,
1
0
,
1
5
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
+
f
i
n
e
-
t
u
n
i
n
g
9
4
.
9
7
±
0
.
0
4
6
7
.
8
8
±
2
.
1
0
5
,
1
0
,
1
5
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
+
l
e
a
r
n
i
n
g
r
a
t
e
sc
h
e
d
u
l
e
r
9
3
.
3
9
±
3
.
4
8
6
5
.
5
2
±
1
.
2
2
5
,
1
0
,
1
5
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
+
a
d
d
i
t
i
o
n
a
l
F
C
l
a
y
e
r
s
8
6
.
5
2
±
8
.
3
6
6
1
.
8
9
±
2
.
6
2
5
,
1
0
,
1
5
TL
-
M
o
b
i
l
e
N
e
t
V
2
+
C
N
N
+
f
i
n
e
-
t
u
n
i
n
g
+
a
d
a
p
t
i
v
e
l
e
a
r
n
i
n
g
r
a
t
e
sc
h
e
d
u
l
e
r
9
7
.
1
7
±
1
.
2
9
6
5
.
3
0
±
1
.
0
0
5
,
1
0
,
1
5
Ov
er
all,
th
e
Mo
b
ileNetV2
v
a
r
ian
ts
s
h
o
w
s
tr
o
n
g
g
en
er
aliza
tio
n
p
atter
n
ac
r
o
s
s
5
,
1
0
,
a
n
d
1
5
ep
o
c
h
s
.
T
h
e
b
aselin
e
Mo
b
ileNetV2
an
d
T
L
-
M
o
b
ileNetV2
s
h
o
w
s
m
all
v
alid
atio
n
r
ates
th
at
in
d
i
ca
te
tr
an
s
f
er
lear
n
i
n
g
alo
n
e
im
p
r
o
v
es
ad
a
p
tatio
n
b
u
t
d
o
es
n
o
t
f
u
lly
ad
d
r
ess
o
v
er
f
itti
n
g
.
T
h
e
a
d
d
itio
n
al
C
NN
(
T
L
-
Mo
b
ileNetV2
+
C
NN)
r
aise
s
tr
ain
in
g
ac
cu
r
ac
y
with
r
elativ
ely
s
m
all
f
lu
ctu
atio
n
s
in
v
alid
atio
n
ac
cu
r
ac
y
,
wh
ile
f
in
e
-
tu
n
e
d
v
ar
ian
t
(
T
L
-
Mo
b
ileNetV2
+
C
NN
+
f
in
e
-
tu
n
in
g
)
ac
h
iev
e
s
th
e
h
ig
h
est
av
e
r
ag
e
v
alid
atio
n
ac
cu
r
ac
y
wit
h
s
lig
h
tly
lar
g
er
v
ar
iab
ilit
y
ac
r
o
s
s
ep
o
ch
s
.
T
h
e
lear
n
in
g
r
ate
s
ch
ed
u
ler
m
o
d
el
m
ain
tain
s
b
alan
ce
d
tr
ad
e
-
o
f
f
b
etwe
en
f
it
an
d
g
en
er
aliza
tio
n
as
co
n
f
ig
u
r
atio
n
with
ad
d
it
io
n
al
f
u
lly
c
o
n
n
ec
te
d
lay
er
s
clea
r
ly
o
v
er
f
its
,
as
r
ef
lecte
d
b
y
h
ig
h
tr
ain
in
g
b
u
t
wea
k
er
v
alid
atio
n
ac
cu
r
ac
y
.
T
h
e
r
ef
o
r
e
,
th
e
m
o
s
t
ad
v
an
ce
d
m
o
d
el
(TL
-
Mo
b
ileNetV2
+
C
NN
+
f
in
e
-
tu
n
in
g
)
d
eliv
er
s
h
ig
h
av
e
r
ag
e
ac
cu
r
ac
y
with
c
o
m
p
ar
ati
v
ely
s
m
all
ch
an
g
es
ac
r
o
s
s
ep
o
ch
s
,
in
d
icatin
g
m
o
r
e
s
tab
le
o
p
tim
izatio
n
b
e
h
av
i
o
r
an
d
r
o
b
u
s
t
g
en
er
aliza
tio
n
o
n
h
eter
o
g
en
eo
u
s
f
r
esh
n
ess
d
atasets
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
4
,
Au
g
u
s
t 2
0
2
6
:
3
6
6
0
-
3
6
7
1
3668
3
.
4
.
St
a
g
e
-
2
ro
bu
s
t
nes
s
f
re
s
h c
la
s
s
if
ica
t
io
n v
ia
1
0
-
f
o
ld cr
o
s
s
-
v
a
lid
a
t
io
n
T
h
e
aim
o
f
th
is
ex
p
er
im
en
t
is
to
ev
alu
ate
th
e
r
o
b
u
s
tn
ess
an
d
s
tab
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
TL
-
Mo
b
ileNetV2
+
C
NN
+
f
in
e
-
tu
n
in
g
v
ar
ian
t
ac
r
o
s
s
co
m
b
in
atio
n
d
atasets
.
T
h
e
s
in
g
le
8
0
:2
0
s
p
lit
ex
p
er
im
en
ts
m
ay
p
r
o
d
u
ce
b
ia
s
r
esu
lts
d
ep
en
d
in
g
o
n
th
e
tr
a
in
–
test
s
tr
u
ctu
r
e
d
ataset
1
0
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
p
r
o
v
id
es
a
m
o
r
e
co
m
p
r
e
h
e
n
s
iv
e
m
ea
s
u
r
e
o
f
g
en
er
aliz
atio
n
b
y
a
v
er
ag
in
g
o
u
tc
o
m
es
ac
r
o
s
s
m
u
ltip
le
f
o
ld
s
[
1
0
]
,
[
1
5
]
,
[
2
1
]
.
T
ab
les 1
5
an
d
1
6
s
h
o
w
t
h
e
1
0
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
r
esu
lts
.
T
ab
le
1
5
.
Stag
e
-
2
–
tr
ain
i
n
g
,
a
n
d
v
alid
atio
n
ac
c
u
r
ac
y
f
o
r
1
0
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
F
o
l
d
Tr
a
i
n
i
n
g
a
c
c
u
r
a
c
y
V
a
l
i
d
a
t
i
o
n
a
c
c
u
r
a
c
y
1
0
.
9
7
3
6
0
.
9
6
7
6
2
6
2
0
.
9
7
1
9
0
.
9
6
6
6
9
7
3
0
.
9
7
5
7
0
.
9
6
4
8
9
7
4
0
.
9
7
2
7
0
.
9
6
7
5
9
7
5
0
.
9
7
5
0
.
9
5
5
8
9
6
6
0
.
9
7
1
9
0
.
9
7
2
0
9
7
7
0
.
9
7
2
4
0
.
9
5
7
6
9
6
8
0
.
9
7
3
2
0
.
9
4
8
6
9
5
9
0
.
9
7
5
5
0
.
9
5
1
3
9
5
10
0
.
9
7
4
9
0
.
9
5
7
6
9
6
A
v
e
r
a
g
e
0
.
9
7
3
6
9
0
.
9
6
1
0
2
9
T
ab
le
1
6
.
Stag
e
-
2
–
a
v
er
ag
e
p
r
e
cisi
o
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
f
o
r
10
-
f
o
ld
cr
o
s
s
v
alid
atio
n
ac
r
o
s
s
class
v
ar
ian
ts
C
l
a
s
s
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
s
c
o
r
e
H
i
g
h
l
y
f
r
e
s
h
0
.
9
0
0
.
9
2
0
.
9
1
F
r
e
sh
0
.
9
7
0
.
9
5
0
.
9
6
N
o
t
f
r
e
s
h
0
.
9
9
0
.
9
9
0
.
9
9
B
ased
o
n
T
a
b
le
1
5
,
t
h
e
m
o
d
el
m
ain
tain
ed
co
n
s
is
ten
tly
h
ig
h
tr
ai
n
in
g
an
d
v
alid
atio
n
p
e
r
f
o
r
m
a
n
ce
ac
r
o
s
s
all
f
o
ld
s
,
with
av
er
a
g
e
tr
ain
in
g
ac
cu
r
ac
y
o
f
9
7
.
3
7
%
an
d
v
alid
atio
n
ac
cu
r
ac
y
o
f
9
6
.
1
0
%.
T
h
e
s
m
all
v
ar
ian
ce
b
etwe
en
f
o
ld
s
d
em
o
n
s
tr
ates
th
at
f
r
am
ewo
r
k
g
en
e
r
alize
s
well
to
u
n
s
ee
n
d
ata,
m
itig
ate
o
v
er
f
itti
n
g
,
an
d
p
r
o
v
e
s
tab
ilit
y
ac
r
o
s
s
v
ar
io
u
s
d
atasets
.
Fo
r
class
-
wi
s
e
ev
alu
atio
n
f
u
r
th
er
s
tr
en
g
t
h
en
s
th
ese
f
in
d
in
g
s
.
T
ab
le
1
6
r
ep
o
r
ts
th
e
a
v
er
ag
e
d
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
f
o
r
ea
c
h
f
r
esh
n
ess
ca
teg
o
r
y
.
T
h
e
n
o
t
f
r
esh
class
ac
h
ie
v
ed
n
ea
r
-
p
e
r
f
ec
t
p
er
f
o
r
m
an
ce
(
p
r
ec
is
io
n
=
0
.
9
9
,
r
ec
a
ll
=
0
.
9
9
,
F1
-
s
co
r
e
=
0
.
9
9
)
,
o
u
tp
er
f
o
r
m
an
ce
p
r
o
p
o
s
ed
m
o
d
el
p
er
f
o
r
m
an
ce
with
h
ig
h
r
eliab
ilit
y
.
T
h
e
f
r
esh
class
al
s
o
d
em
o
n
s
tr
ated
s
tr
o
n
g
o
u
tco
m
e
s
(
p
r
ec
is
io
n
=
0
.
9
7
,
r
ec
all
=
0
.
9
5
,
F1
-
s
co
r
e
=
0
.
9
6
)
,
p
r
o
p
o
s
ed
th
e
m
i
d
-
r
an
g
e
f
r
esh
n
ess
lev
els
ar
e
co
n
s
is
ten
tly
ca
p
tu
r
ed
.
Fo
r
th
e
h
ig
h
l
y
f
r
esh
class
,
al
th
o
u
g
h
p
r
ec
is
io
n
(
0
.
9
0
)
an
d
r
ec
all
(
0
.
9
2
)
s
lig
h
tly
lo
wer
t
h
an
o
th
er
class
es,
th
e
av
er
a
g
e
F1
-
s
c
o
r
e
o
f
0
.
9
1
r
em
ai
n
s
co
m
p
etiti
v
e,
h
i
g
h
lig
h
ted
th
e
ch
allen
g
es
o
f
d
is
tin
g
u
is
h
in
g
f
i
n
e
-
g
r
ain
e
d
f
r
esh
n
ess
cu
es
at
th
e
u
p
p
e
r
q
u
ality
b
o
u
n
d
ar
y
.
Ov
er
all,
th
ese
r
esu
lts
in
d
icate
th
at
th
e
d
u
al
-
s
tag
e
a
r
ch
itectu
r
e
ac
h
ie
v
es
b
o
th
ac
c
u
r
ac
y
a
n
d
r
o
b
u
s
tn
ess
,
o
u
t
p
er
f
o
r
m
th
e
tr
ad
itio
n
al
b
aselin
e
Mo
b
ileNetV2
m
o
d
el
r
ep
o
r
ted
in
p
r
i
o
r
s
tu
d
ies.
4.
CO
NCLU
SI
O
N
T
h
is
p
ap
er
p
r
esen
ted
m
u
lti
-
s
tag
e
h
y
b
r
id
r
ep
r
esen
tatio
n
lear
n
in
g
f
r
am
ewo
r
k
th
at
in
teg
r
ates
YOL
Ov
8
-
d
r
iv
e
n
f
ea
tu
r
e
-
g
ate
d
d
etec
tio
n
with
Mo
b
ileNetV2
-
C
NN
tr
an
s
f
er
lear
n
in
g
f
o
r
f
is
h
f
r
esh
n
ess
class
if
icatio
n
.
T
h
e
two
s
tag
es
p
ip
elin
e
v
alid
ates
in
p
u
t
im
ag
es
b
ef
o
r
e
p
r
ed
icts
f
r
es
h
n
ess
ac
r
o
s
s
th
r
ee
ca
teg
o
r
ies:
h
ig
h
ly
f
r
esh
,
f
r
esh
,
an
d
n
o
t
f
r
esh
.
T
h
e
en
h
an
ce
d
m
o
d
el
ac
r
o
s
s
Mo
b
ileNetV2
,
C
NN
ex
ten
s
io
n
s
,
f
in
e
-
tu
n
in
g
,
an
d
a
d
ap
tiv
e
l
ea
r
n
in
g
r
ate
s
ch
ed
u
lin
g
b
ac
k
b
o
n
e
ac
h
iev
ed
h
ig
h
est
tr
ain
in
g
ac
cu
r
ac
y
o
f
9
8
.
0
7
%
an
d
v
ali
d
atio
n
ac
cu
r
a
cy
o
f
6
7
.
7
2
%
u
s
in
g
8
0
:2
0
s
p
l
it,
wh
ile
1
0
-
f
o
ld
cr
o
s
s
-
v
alid
at
io
n
p
r
esen
ted
p
ea
k
tr
ain
in
g
ac
cu
r
ac
y
o
f
9
7
.
5
7
%,
an
d
v
alid
atio
n
ac
cu
r
ac
y
o
f
9
7
.
2
1
%.
T
h
e
f
in
d
in
g
s
h
ig
h
lig
h
t
th
e
v
alu
e
o
f
d
u
al
f
ea
tu
r
e
-
g
ated
d
etec
tio
n
with
v
ar
io
u
s
T
L
-
Mo
b
ileNetV2
class
if
icatio
n
in
th
e
s
tate
o
f
th
e
ar
t
f
o
r
f
is
h
f
r
esh
n
ess
m
o
n
ito
r
in
g
.
Ho
we
v
er
,
th
er
e
h
av
e
s
o
m
e
lim
itatio
n
s
s
u
ch
as
m
i
s
clas
s
if
icatio
n
s
o
c
ca
s
io
n
ally
o
cc
u
r
r
ed
b
etwe
en
ad
jace
n
t
f
r
esh
n
ess
ca
teg
o
r
ies
wh
en
m
in
im
al
v
is
u
al
d
if
f
er
en
ce
s
an
d
lim
ited
to
f
is
h
s
p
ec
ies
d
ataset.
B
ey
o
n
d
f
is
h
a
p
p
licatio
n
,
th
e
p
r
o
p
o
s
ed
d
u
al
-
s
tag
e
f
r
am
ewo
r
k
o
f
f
er
s
tr
an
s
f
er
a
b
le
A
I
m
eth
o
d
o
lo
g
y
f
o
r
b
r
o
ad
e
r
p
e
r
is
h
ab
le
f
o
o
d
q
u
ality
m
o
n
ito
r
i
n
g
.
T
h
e
YOL
O
-
b
ased
f
ea
tu
r
e
-
g
ated
d
etec
tio
n
s
tag
e
ca
n
b
e
r
ep
u
r
p
o
s
ed
t
o
v
alid
ate
m
ea
t
cu
ts
,
p
o
u
ltry
s
u
r
f
ac
es,
o
r
f
r
u
it
an
d
v
e
g
etab
le
d
ef
o
r
m
ities
,
wh
ile
th
e
Mo
b
ileNetV2
-
C
NN
class
if
ie
r
ca
n
b
e
f
in
e
-
tu
n
ed
t
o
lea
r
n
d
o
m
ain
f
r
esh
n
ess
cu
es
s
u
ch
as
b
r
u
is
in
g
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
Mu
lti
-
s
ta
g
e
h
yb
r
id
YOLO
-
d
r
iv
en
a
n
d
Mo
b
ileN
etV
2
-
C
N
N
va
r
ia
n
ts
fo
r
r
o
b
u
s
t
…
(
R
a
s
ee
d
a
H
a
mza
h
)
3669
d
is
co
lo
r
atio
n
,
a
n
d
d
e
h
y
d
r
atio
n
.
T
h
e
a
r
ch
itectu
r
e
m
ay
s
tr
en
g
th
en
th
e
e
x
tr
ac
tio
n
o
f
s
u
b
tle
s
p
atial
tex
tu
r
es
an
d
p
r
o
g
r
ess
iv
e
d
e
g
r
ad
atio
n
in
d
ica
to
r
s
.
ACK
NO
WL
E
DG
M
E
N
T
S
T
h
e
au
th
o
r
s
wo
u
ld
lik
e
t
o
th
an
k
th
e
Un
iv
er
s
iti T
ek
n
o
lo
g
i M
AR
A
f
o
r
r
esear
ch
f
u
n
d
in
g
an
d
s
u
p
p
o
r
t.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
T
h
is
wo
r
k
s
u
p
p
o
r
ted
b
y
th
e
Un
iv
er
s
iti
T
ek
n
o
l
o
g
i
MA
R
A
th
r
o
u
g
h
th
e
p
r
o
ject
“
L
ig
h
tw
eig
h
t
d
ee
p
n
eu
r
al
n
etwo
r
k
p
r
u
n
in
g
m
o
d
el
lin
g
b
ased
o
n
a
d
ap
tiv
e
ap
p
r
o
x
im
ate
g
en
er
al
m
atr
ix
m
u
ltip
licatio
n
an
d
iter
ativ
e
s
tr
u
ctu
r
ed
f
ilter
f
u
s
io
n
”
u
n
d
er
Gr
an
t 1
0
0
-
T
NC
PI/I
NT
1
6
/6
/2
(
0
6
0
/2
0
2
4
)
.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
id
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
R
aseed
a
Ham
za
h
✓
✓
✓
✓
✓
✓
✓
✓
R
o
s
n
iza
R
o
s
lan
✓
✓
✓
✓
✓
✓
Am
n
i M
u
n
ir
a
Kh
id
i
r
✓
✓
✓
✓
✓
✓
✓
✓
✓
L
ala
Sep
tem
R
iza
✓
✓
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
ST
A
T
E
M
E
NT
Au
th
o
r
s
s
tate
n
o
co
n
f
lict o
f
in
t
er
est
DATA AV
AI
L
AB
I
L
I
T
Y
T
h
e
d
ata
th
at
s
u
p
p
o
r
t th
e
f
in
d
i
n
g
s
o
f
th
is
s
tu
d
y
a
r
e
as f
o
llo
w
s
:
–
Fis
h
f
r
esh
n
ess
clas
s
if
icatio
n
in
Kag
g
le
at
h
ttp
s
://www.
k
ag
g
le.
co
m
/d
atasets
/m
u
h
am
m
a
d
ab
u
r
ay
a
n
/f
is
h
-
f
r
esh
n
ess
-
class
if
icatio
n
.
–
T
h
e
f
r
esh
n
ess
o
f
th
e
f
is
h
ey
es
d
ataset
in
Me
n
d
eley
d
ata
at
h
ttp
s
://d
ata.
m
en
d
eley
.
co
m
/d
at
asets
/x
zy
x
7
p
b
r
3
w/1
.
–
Prim
ar
y
d
ataset
o
f
f
is
h
im
a
g
es
th
at
ca
p
tu
r
ed
an
d
c
o
llected
at
l
o
ca
l
f
is
h
m
ar
k
ets
in
Ma
lay
s
ia
a
t
h
ttp
s
://d
r
iv
e.
g
o
o
g
le.
co
m
/d
r
iv
e
/f
o
ld
er
s
/1
-
o
6
wq
m
0
2
e
k
L
m
zDL
E
UAOz
SF
T
y
k
L
Si6
o
2
5
?u
s
p
=
d
r
iv
e_
lin
k
–
C
h
est
X
-
r
ay
(
C
OVI
D
-
1
9
&
Pn
eu
m
o
n
ia)
im
ag
e
d
ataset
in
Kag
g
le
at
h
ttp
s
://www.
k
ag
g
le.
co
m
/d
atasets
/p
r
ash
an
t2
6
8
/ch
est
-
x
r
a
y
-
co
v
id
1
9
-
p
n
eu
m
o
n
ia
.
–
J
elly
f
is
h
im
ag
e
d
ataset
in
Kag
g
le
at
h
ttp
s
://www.
k
ag
g
le.
co
m
/d
atasets
/an
s
h
tan
war
/jelly
f
is
h
-
ty
p
es
–
An
im
al
im
ag
e
d
ataset
-
ca
ts
,
d
o
g
s
,
an
d
f
o
x
es
in
Kag
g
le
at
h
ttp
s
://www.
k
ag
g
le.
co
m
/d
atasets
/s
n
m
ah
s
a/an
im
al
-
im
ag
e
-
d
ataset
-
ca
ts
-
dogs
-
an
d
-
f
o
x
es
.
–
Gar
b
ag
e
im
ag
e
d
ataset
in
Kag
g
le
at
h
ttp
s
://www.
k
ag
g
le.
co
m
/d
atasets
/f
ar
za
d
n
ek
o
u
ei/tr
as
h
-
ty
p
e
-
im
a
g
e
-
d
ataset
.
RE
F
E
R
E
NC
E
S
[
1
]
B
.
D
.
A
b
e
r
a
a
n
d
M
.
A
.
A
d
i
m
a
s,
“
H
e
a
l
t
h
b
e
n
e
f
i
t
s
a
n
d
h
e
a
l
t
h
r
i
s
k
s
o
f
c
o
n
t
a
m
i
n
a
t
e
d
f
i
sh
c
o
n
s
u
m
p
t
i
o
n
:
c
u
r
r
e
n
t
r
e
sea
r
c
h
o
u
t
p
u
t
s
,
r
e
sea
r
c
h
a
p
p
r
o
a
c
h
e
s
,
a
n
d
p
e
r
s
p
e
c
t
i
v
e
s
,
”
H
e
l
i
y
o
n
,
2
0
2
4
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
h
e
l
i
y
o
n
.
2
0
2
4
.
e
3
3
9
0
5
.
[
2
]
M
.
F
.
A
.
R
a
z
a
k
e
t
a
l
.
,
“
S
c
o
mb
r
o
i
d
p
o
i
s
o
n
i
n
g
a
m
o
n
g
p
r
i
s
o
n
e
r
s
o
f
a
p
r
i
s
o
n
i
n
J
o
h
o
r
,
M
a
l
a
y
si
a
,
”
M
a
l
a
y
s
i
a
n
J
o
u
r
n
a
l
o
f
P
u
b
l
i
c
H
e
a
l
t
h
M
e
d
i
c
i
n
e
,
v
o
l
.
2
4
,
n
o
.
1
,
p
p
.
2
3
2
–
2
3
8
,
2
0
2
4
.
[
3
]
M
.
A
.
I
b
r
a
h
i
m
,
S
.
M
.
I
b
r
a
h
i
m
,
a
n
d
F
.
H
.
A
b
o
z
a
i
d
,
“
G
u
i
d
e
l
i
n
e
s
o
f
f
i
sh
s
u
p
p
l
y
c
h
a
i
n
f
r
o
m
p
o
st
h
a
r
v
e
st
i
n
g
t
o
m
a
r
k
e
t
i
n
g
,
”
Eg
y
p
t
i
a
n
J
o
u
rn
a
l
o
f
Aq
u
a
t
i
c
Bi
o
l
o
g
y
&
F
i
sh
e
ri
e
s
,
v
o
l
.
2
8
,
n
o
.
5
,
2
0
2
4
,
d
o
i
:
1
0
.
2
1
6
0
8
/
e
j
a
b
f
.
2
0
2
4
.
3
7
8
8
7
5
.
[
4
]
Y
.
E.
T
.
Y
a
s
i
n
,
I
.
A
.
O
z
k
a
n
,
a
n
d
M
.
K
ö
k
l
ü
,
“
D
e
t
e
c
t
i
o
n
o
f
f
i
s
h
f
r
e
s
h
n
e
ss
u
si
n
g
a
r
t
i
f
i
c
i
a
l
i
n
t
e
l
l
i
g
e
n
c
e
m
e
t
h
o
d
s
,
”
Eu
r
o
p
e
a
n
Fo
o
d
Re
se
a
rc
h
a
n
d
T
e
c
h
n
o
l
o
g
y
,
v
o
l
.
2
4
9
,
n
o
.
8
,
p
p
.
1
9
7
9
–
1
9
9
0
,
2
0
2
3
,
d
o
i
:
1
0
.
1
0
0
7
/
s
0
0
2
1
7
-
0
2
3
-
0
4
2
7
1
-
4.
[
5
]
E.
Es
t
e
v
e
s
a
n
d
J.
A
n
í
b
a
l
,
“
S
e
n
so
r
y
e
v
a
l
u
a
t
i
o
n
o
f
se
a
f
o
o
d
f
r
e
s
h
n
e
ss
u
si
n
g
t
h
e
q
u
a
l
i
t
y
i
n
d
e
x
me
t
h
o
d
:
a
met
a
-
a
n
a
l
y
s
i
s,
”
I
n
t
e
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
F
o
o
d
Mi
c
r
o
b
i
o
l
o
g
y
,
v
o
l
.
3
3
4
,
2
0
2
0
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
j
f
o
o
d
mi
c
r
o
.
2
0
2
0
.
1
0
8
9
3
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.