I
A
E
S
I
n
t
e
r
n
at
io
n
al
Jou
r
n
al
of
A
r
t
if
ic
ia
l
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
V
ol
.
14
, N
o.
1
,
F
e
bua
r
y
2025
, pp.
658
~
666
I
S
S
N
:
2252
-
8938
,
D
O
I
:
10.11591/
ij
a
i.
v
14
.i
1
.pp
658
-
666
658
Jou
r
n
al
h
om
e
page
:
ht
tp
:
//
ij
ai
.
ia
e
s
c
or
e
.c
om
E
n
h
a
n
c
e
m
e
n
t
of
Y
O
L
O
v
5
f
or
au
t
om
at
i
c
w
e
e
d
d
e
t
e
c
t
i
on
t
h
r
ou
gh
b
ac
k
b
o
n
e
op
t
i
m
i
z
at
i
on
M
oh
am
m
e
d
H
ab
ib
1
, S
al
m
a S
e
k
h
r
a
1
,
A
d
il
T
an
n
ou
c
h
e
2
,
Y
ou
s
s
e
f
O
u
n
e
j
j
ar
1
1
L
a
bor
a
t
or
y
of
S
pe
c
t
r
om
e
t
r
y, M
a
t
e
r
i
a
l
s
a
nd A
r
c
ha
e
om
a
t
e
r
i
a
l
s
“
L
A
S
M
A
R
”
,
F
a
c
ul
t
y of
S
c
i
e
nc
e
s
,
M
oul
a
y I
s
m
a
i
l
U
ni
ve
r
s
i
t
y
,
M
e
kne
s
,
M
or
oc
c
o
2
L
a
bor
a
t
or
y of
E
ngi
ne
e
r
i
ng a
nd A
ppl
i
e
d T
e
c
hnol
ogy “
L
I
T
A
”
, H
i
gh
S
c
hool
of
T
e
c
hnol
ogy,
U
ni
ve
r
s
i
t
é
S
ul
t
a
n M
oul
a
y S
l
i
m
a
ne
,
B
e
ni
M
e
l
l
a
l
,
M
or
oc
c
o
A
r
t
ic
le
I
n
f
o
A
B
S
T
R
A
C
T
A
r
ti
c
le
h
is
to
r
y
:
R
e
c
e
iv
e
d
M
a
r
15
,
2024
R
e
vi
s
e
d
J
ul
6
,
2024
A
c
c
e
pt
e
d
O
c
t
18
,
2024
In
the
context
of
our
research
project,
which
involves
developing
a
r
obotic
system
capable
of
eliminating
weeds
using
deep
learning
techni
cs,
the
selection
of
powerful
object
detection
model
is
essential.
Object
de
tectors
typically
consist
of
three
components:
backbone
,
neck,
and
prediction
head.
In
this
study,
we
propose
an
enhancement
to
the
you
only
look
once
ve
rsion
5
(
YOLO
v5
)
network
by
using
the
most
popular
convolut
ional
neural
ne
tworks
(
CNN
)
networks
(such
as
DarkNet
and
MobileNet)
as
back
bones.
The
objective
of
this
study
is
to
identify
the
best
backbone
that
can
i
mprove
YOLO
v5
'
s
performance
while
preserving
its
other
layers
(neck
and
head)
.
I
n
terms
of
detecting
and
ultra
-
localizing
pea
crops.
Additionally,
we
compared
their
results
with
those
of
the
most
commonly
used
object
detectors.
Our
findings
indicate
that
the
fastest
models
among
the
ne
tworks
studied
were
MobileNet,
YOLO
-
tiny,
and
YOLO
v5
,
with
speeds
r
anging
from
5
to
14
milliseconds
per
image.
Among
these
models,
Mobil
eNetv1
demonstrated
the
highest
accuracy,
achieving
average
precision
(AP)
s
core
of
89.3%
for
intersectio
n
over
union
(IoU)
threshold
of
0.5.
However,
the
accuracy
of
this
model
dec
reased
when
we
increased
the
threshold,
sug
gesting
that
it
does
not
provide
perfect
crop
delineation.
On
the
other
hand,
while
YOLO
v5
had a lower AP score than MobileNe
tv1 at an IoU threshold
of 0.5,
it exhibited gr
eater sta
bility when f
aced with
variations
in this thresho
ld.
K
e
y
w
o
r
d
s
:
C
om
put
e
r
vi
s
io
n
C
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
ba
c
kbone
D
e
e
p l
e
a
r
ni
ng
O
bj
e
c
t
de
te
c
to
r
s
S
m
a
r
t
f
a
r
m
in
g
W
e
e
d de
te
c
ti
on
This is an
open
acce
ss artic
le unde
r the
CC BY
-
SA
license.
C
or
r
e
s
pon
di
n
g A
u
th
or
:
M
oha
m
m
e
d H
a
bi
b
L
a
bor
a
to
r
y of
S
pe
c
tr
om
e
t
r
y, M
a
te
r
ia
ls
a
nd A
r
c
ha
e
om
a
te
r
ia
l
s
“
L
A
S
M
A
R
”
, F
a
c
ul
ty
of
S
c
ie
nc
e
s
M
oul
a
y
I
s
m
a
il
U
ni
ve
r
s
it
y
M
e
kne
s
,
M
or
oc
c
o
E
m
a
il
:
ha
bi
bm
oha
m
e
dt
s
e
i
@
gm
a
il
.c
om
1.
I
N
T
R
O
D
U
C
T
I
O
N
W
it
h
th
e
gr
ow
in
g
gl
oba
l
popula
ti
on
a
nd
th
e
ne
e
d
to
p
r
oduc
e
m
or
e
f
ood
s
us
ta
in
a
bl
y,
f
a
r
m
e
r
s
a
nd
r
e
s
e
a
r
c
he
r
s
a
r
e
tu
r
ni
ng
to
a
dva
nc
e
d
te
c
hnol
ogi
e
s
,
in
c
lu
di
ng
m
a
c
hi
ne
le
a
r
ni
ng
to
opt
im
iz
e
a
gr
ic
ul
tu
r
a
l
pr
oc
e
s
s
e
s
.
T
hi
s
te
c
hnol
ogy
he
lp
s
s
ol
v
e
va
r
io
us
a
gr
ic
ul
tu
r
a
l
is
s
ue
s
,
s
u
c
h
a
s
w
e
e
d
c
ont
r
ol
.
M
a
c
hi
ne
le
a
r
ni
ng
m
e
th
ods
c
a
n
a
id
in
th
e
pr
e
c
is
e
id
e
nt
if
ic
a
ti
on
a
nd
m
a
na
ge
m
e
nt
of
w
e
e
ds
,
th
e
r
e
by
r
e
duc
in
g
th
e
ne
e
d
f
or
c
he
m
ic
a
l
pr
odu
c
ts
a
nd i
m
pr
ovi
ng c
r
op yie
ld
s
.
I
n
th
is
c
ont
e
xt
,
our
r
e
s
e
a
r
c
h
te
a
m
h
a
s
a
lr
e
a
dy
m
a
d
e
s
ig
ni
f
ic
a
nt
pr
ogr
e
s
s
in
s
ol
vi
ng
pr
obl
e
m
s
r
e
la
te
d
to
w
e
e
d
c
ont
r
ol
. I
n t
he
ir
w
or
k,
T
a
nnouc
he
e
t
al
.
[
1]
ut
il
iz
e
d a
m
e
th
od ba
s
e
d on
a
s
e
t
of
di
s
c
r
im
in
a
nt
c
la
s
s
if
ie
r
s
c
ons
tr
uc
te
d
us
in
g
H
a
a
r
-
li
ke
f
e
a
tu
r
e
s
.
T
he
y
a
c
hi
e
v
e
d
a
not
a
bl
e
m
il
e
s
to
ne
by
de
ve
lo
pi
ng
a
n
e
f
f
ic
ie
nt
s
h
a
pe
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
E
nhanc
e
m
e
nt
of
Y
O
L
O
v
5 f
or
aut
om
at
ic
w
e
e
d d
e
te
c
ti
on t
hr
ough
bac
k
bone
…
(
M
ohamm
e
d H
abi
b)
659
de
s
c
r
ip
to
r
s
pe
c
if
ic
a
ll
y
d
e
s
ig
ne
d
to
di
s
ti
ngui
s
h
be
twe
e
n
two
ty
pe
s
of
w
e
e
d
s
p
e
c
ie
s
:
m
onoc
ot
s
a
nd
di
c
ot
s
[
2]
.
B
ot
h s
tu
di
e
s
e
m
pl
oye
d m
a
c
hi
ne
l
e
a
r
ni
ng t
e
c
hni
que
s
w
it
h t
he
a
i
m
of
opt
im
iz
in
g he
r
bi
c
id
e
us
a
ge
.
I
n t
he
c
ont
e
xt
of
s
it
e
-
s
p
e
c
if
ic
w
e
e
d m
a
na
g
e
m
e
nt
(
S
S
W
M
)
, w
hi
c
h a
im
s
t
o
t
a
r
ge
t
a
nd pr
e
c
i
s
e
ly
c
ont
r
ol
w
e
e
ds
,
w
e
ha
ve
de
ve
lo
pe
d
r
e
s
e
a
r
c
h
to
m
e
e
t
th
is
ne
e
d. O
ur
s
tu
d
ie
s
f
oc
us
e
d
on
us
in
g
de
e
p
le
a
r
ni
ng
m
ode
ls
f
or
de
te
c
ti
ng
a
nd
id
e
nt
if
yi
ng
c
r
ops
a
m
ong
w
e
e
ds
.
I
n
our
in
it
ia
l
s
tu
dy,
w
e
tr
a
in
e
d
th
e
F
a
s
te
r
r
e
gi
on
-
ba
s
e
d
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
k
(
R
C
N
N
)
r
e
s
id
ua
l
ne
twor
ks
(
R
e
s
N
e
t
)
-
50
m
ode
l
to
de
te
c
t
a
nd
lo
c
a
te
c
r
ops
w
it
h
hi
gh pr
e
c
is
io
n
[
3]
. I
n
a
s
ubs
e
que
nt
pr
oj
e
c
t,
w
e
e
va
lu
a
te
d t
he
pe
r
f
or
m
a
nc
e
of
t
he
you
onl
y l
ook onc
e
ve
r
s
io
n
5
(
Y
O
L
O
v5
)
obj
e
c
t
de
te
c
to
r
by
te
s
ti
ng
di
f
f
e
r
e
nt
s
iz
e
s
a
nd
le
a
r
ni
n
g
m
e
th
ods
,
in
c
lu
di
ng
tr
a
in
in
g
f
r
om
s
c
r
a
tc
h
a
nd
tr
a
ns
f
e
r
le
a
r
ni
ng.
T
hi
s
a
ll
ow
e
d
us
to
a
c
hi
e
ve
hi
gh
a
c
c
ur
a
c
y
a
nd
f
a
s
t
lo
c
a
li
z
a
ti
on
s
pe
e
d
s
f
or
c
r
ops
.
F
ur
th
e
r
m
or
e
,
w
e
e
xpl
or
e
d
va
r
io
us
s
e
gm
e
nt
a
ti
on
ne
twor
ks
to
e
f
f
e
c
ti
ve
ly
di
s
ti
ngui
s
h
c
r
ops
f
r
om
w
e
e
ds
[
4]
.
T
hi
s
a
n
a
ly
s
is
e
na
bl
e
d
u
s
to
a
c
c
ur
a
te
ly
de
te
r
m
in
e
th
e
pos
it
io
ns
of
w
e
e
ds
b
y
e
li
m
in
a
ti
ng
th
e
c
r
op
r
e
gi
ons
id
e
nt
if
ie
d
by
Y
O
L
O
v5
.
I
n
a
not
he
r
s
tu
dy,
w
e
pr
opos
e
d
a
c
la
s
s
if
ic
a
ti
on
m
ode
l
ba
s
e
d
on
R
e
s
N
e
t
a
nd
M
obi
le
N
e
t
f
or
id
e
nt
if
yi
ng
a
nd
di
f
f
e
r
e
nt
ia
ti
ng
be
twe
e
n
c
r
ops
a
nd
w
e
e
d
s
,
a
nd
c
om
p
a
r
e
d
it
s
pe
r
f
or
m
a
nc
e
w
it
h
popula
r
c
onvolut
io
na
l
ne
ur
a
l
ne
twor
ks
(
C
N
N
s
)
[
5]
.
I
n
th
e
s
a
m
e
c
ont
e
xt
,
W
a
ng
e
t
al
.
[
6]
s
um
m
a
r
iz
e
d
th
e
pr
ogr
e
s
s
m
a
de
in
w
e
e
d
de
t
e
c
ti
on
us
in
g
c
om
put
e
r
vi
s
io
n
a
nd
im
a
ge
pr
oc
e
s
s
in
g
te
c
hni
que
s
.
A
ll
th
e
s
e
te
c
hni
que
s
h
a
ve
f
oc
us
e
d
on
di
r
e
c
t
de
te
c
ti
on
of
w
e
e
d
f
e
a
tu
r
e
s
,
a
de
li
c
a
te
ta
s
k
in
vi
e
w
of
th
e
w
id
e
va
r
ie
ty
o
f
w
e
e
d
s
pe
c
ie
s
.
O
ur
w
e
e
d
de
te
c
ti
on
m
e
th
od
in
vol
ve
s
a
two
-
s
te
p
a
ppr
oa
c
h:
ve
ge
ta
ti
on/
gr
ound
di
s
c
r
im
in
a
ti
on
(
s
e
gm
e
nt
a
ti
on)
f
ol
lo
w
e
d
by
c
r
op/
w
e
e
d
di
s
c
r
im
in
a
ti
on
(
obj
e
c
t
de
te
c
ti
on)
, w
hi
c
h a
ll
ow
s
us
t
o e
xt
r
a
c
t
th
e
w
e
e
d
s
a
lo
ne
.
O
bj
e
c
t
de
te
c
ti
on
c
ont
in
ue
s
to
be
a
s
ig
ni
f
ic
a
nt
a
r
e
a
of
r
e
s
e
a
r
c
h
in
de
e
p
le
a
r
ni
ng,
w
it
h
va
r
io
us
a
ppl
ic
a
ti
ons
s
how
c
a
s
in
g
th
e
c
a
p
a
bi
li
ti
e
s
of
obj
e
c
t
de
t
e
c
to
r
s
in
s
o
lv
in
g
c
om
put
e
r
vi
s
io
n
pr
obl
e
m
s
.
T
h
e
s
e
in
c
lu
de
f
a
c
ia
l
r
e
c
ogni
ti
on
[
7]
, pe
de
s
tr
ia
n de
te
c
ti
on
[
8]
, vi
de
o
a
na
ly
s
is
[
9
]
,
a
nd
lo
go de
te
c
ti
on
[
10]
. C
ur
r
e
nt
ly
, t
he
r
e
a
r
e
two ma
jo
r
c
a
te
gor
ie
s
of
obj
e
c
t
de
te
c
to
r
s
:
two
-
s
ta
g
e
de
te
c
to
r
s
, s
uc
h a
s
R
C
N
N
a
nd i
ts
va
r
ia
nt
s
, a
nd
s
in
gl
e
-
s
ta
ge
de
te
c
to
r
s
, w
hi
c
h a
r
e
ge
n
e
r
a
ll
y m
or
e
pow
e
r
f
ul
t
ha
n t
he
f
or
m
e
r
, l
ik
e
t
he
Y
O
L
O
m
ode
l
a
nd i
ts
ve
r
s
io
ns
.
T
he
r
e
m
a
r
ka
bl
e
s
uc
c
e
s
s
of
Y
O
L
O
'
s
a
r
c
hi
te
c
tu
r
e
ha
s
le
d
to
nu
m
e
r
ous
r
e
s
e
a
r
c
h
e
f
f
or
ts
a
im
e
d
a
t
i
ts
im
pr
ove
m
e
nt
.
I
n
[
11
]
,
[
12]
,
th
e
a
ut
hor
s
opt
im
iz
e
d
th
e
num
be
r
a
nd
s
iz
e
of
bounding
boxe
s
ge
ne
r
a
te
d
by
Y
O
L
O
v3
a
nd
Y
O
L
O
v5
dur
in
g
tr
a
in
in
g
us
in
g
K
-
m
e
a
ns
a
nd
K
-
m
e
a
ns
+
+
c
lu
s
t
e
r
in
g
te
c
hni
que
s
.
A
f
e
a
tu
r
e
f
us
io
n
m
e
th
od
c
a
ll
e
d
P
B
-
F
P
N
w
a
s
pr
opos
e
d,
bui
ld
in
g
upon
pa
th
a
g
gr
e
ga
ti
on
ne
twor
k
(
P
A
N
e
t)
a
nd
bi
di
r
e
c
ti
ona
l
f
e
a
tu
r
e
pyr
a
m
id
ne
twor
k
(
B
iF
P
N
)
te
c
hni
que
s
[
13]
.
A
not
he
r
s
tu
dy
f
oc
us
e
d
on
opt
im
iz
in
g
th
e
Y
O
L
O
v5
m
ode
l
s
pe
c
if
ic
a
ll
y
f
or
pl
a
nt
di
s
e
a
s
e
id
e
nt
if
ic
a
ti
on
[
14]
.
A
ddi
ti
ona
ll
y,
r
e
s
e
a
r
c
he
r
s
pr
opos
e
d
a
di
la
ti
on
te
c
hni
que
f
or
th
e
s
pa
ti
a
l
pyr
a
m
id
m
odul
e
(
S
P
P
)
to
in
c
or
po
r
a
te
m
ul
ti
-
s
c
a
le
in
f
o
r
m
a
ti
on
a
nd
a
ddr
e
s
s
s
c
a
le
va
r
ia
ti
on
is
s
ue
s
in
Y
O
L
O
v3
[
15]
.
I
n
th
e
dom
a
in
of
ki
w
i
a
gr
ic
ul
tu
r
e
f
a
ul
t
de
te
c
ti
on,
th
e
a
ut
hor
s
of
[
16]
in
tr
oduc
e
d
s
e
ve
r
a
l
e
nha
nc
e
m
e
nt
s
to
Y
O
L
O
v5
. T
he
s
e
im
pr
ove
m
e
nt
s
in
c
lu
de
d
th
e
a
ddi
ti
on
of
a
s
m
a
ll
ta
r
ge
t
de
te
c
ti
on
la
ye
r
a
nd
a
S
E
L
a
ye
r
,
a
s
w
e
ll
a
s
m
odi
f
ic
a
ti
ons
to
th
e
lo
s
s
f
unc
ti
on
f
r
o
m
di
s
ta
nc
e
-
in
te
r
s
e
c
ti
on
ove
r
uni
on
(
D
i
oU
)
to
c
om
pl
e
te
-
in
te
r
s
e
c
ti
on ove
r
uni
on
(
C
i
oU
)
.
W
h
e
n
it
c
o
m
e
s
t
o
e
n
ha
nc
in
g
th
e
a
c
c
ur
a
c
y
of
ob
je
c
t
de
t
e
c
ti
on
n
e
twor
k
s
,
a
r
c
h
it
e
c
tu
r
e
pl
a
y
s
a
vi
t
a
l
r
ol
e
a
m
o
ng
va
r
io
us
p
a
r
a
m
e
te
r
s
.
T
yp
ic
a
l
ly
,
ob
je
c
t
de
te
c
t
or
s
c
on
s
i
s
t
of
t
hr
e
e
m
a
i
n
c
om
p
on
e
n
ts
:
t
he
ba
c
k
bo
ne
,
t
he
ne
c
k
,
a
nd
th
e
pr
e
di
c
ti
on
he
a
d
.
I
n
th
i
s
r
e
s
e
a
r
c
h
,
w
e
pr
o
po
s
e
a
n
e
nh
a
n
c
e
m
e
n
t
to
th
e
Y
O
L
O
v
5
n
e
t
w
or
k
b
y
u
s
i
ng
th
e
m
o
s
t
pop
ul
a
r
C
N
N
n
e
t
w
or
ks
(
s
u
c
h
a
s
D
a
r
kN
e
t
,
a
nd
M
obi
le
N
e
t)
a
s
b
a
c
kbo
ne
s
.
T
h
e
o
bj
e
c
ti
v
e
of
t
hi
s
s
t
udy
i
s
t
o
i
de
nt
if
y
th
e
b
e
s
t
b
a
c
kb
on
e
th
a
t
c
a
n
im
pr
ov
e
Y
O
L
O
v5
'
s
p
e
r
f
or
m
a
nc
e
w
h
il
e
pr
e
s
e
r
vi
ng
it
s
o
th
e
r
la
ye
r
s
(
ne
c
k
a
nd
he
a
d)
.
W
e
f
o
c
u
s
o
n
e
v
a
l
ua
ti
n
g
th
e
ir
p
e
r
f
o
r
m
a
n
c
e
i
n
t
h
e
a
c
c
ur
a
te
d
e
t
e
c
t
io
n
a
nd
l
oc
a
li
z
a
ti
on
of
p
e
a
c
r
o
p
s
.
A
ddi
ti
o
na
ll
y
,
w
e
w
i
ll
c
o
m
p
a
r
e
t
he
r
e
s
ul
t
s
obt
a
i
ne
d
by
th
e
s
e
n
e
tw
or
k
s
w
i
th
th
os
e
o
f
o
th
e
r
c
om
m
o
nl
y
u
s
e
d
obj
e
c
t
de
te
c
t
or
s
.
2.
M
E
T
H
O
D
I
n
th
is
s
e
c
ti
on
w
e
w
il
l
s
how
you
th
e
m
e
th
ods
a
nd
m
a
te
r
ia
ls
w
e
us
e
d
to
c
a
r
r
y
out
th
is
s
tu
dy.
A
m
ode
r
n
obj
e
c
t
de
te
c
to
r
c
on
s
is
ts
of
3
m
a
in
pa
r
ts
, a
ba
c
kbon
e
,
a
n
e
c
k
,
a
n
d
a
he
a
d
(
F
ig
ur
e
1)
.
T
h
e
ba
c
kbon
e
c
om
pone
nt
of
a
n
obj
e
c
t
de
te
c
ti
on
m
ode
l
c
ons
is
ts
of
a
s
e
t
of
la
ye
r
s
r
e
s
pons
ib
le
f
or
e
xt
r
a
c
ti
ng
f
e
a
tu
r
e
s
f
r
om
in
put
im
a
ge
s
.
T
he
s
e
la
ye
r
s
a
im
to
c
a
pt
ur
e
de
ta
il
e
d
in
f
or
m
a
ti
on
f
r
om
th
e
im
a
ge
s
. T
he
ne
c
k
c
om
pone
nt
,
on
th
e
ot
he
r
h
a
nd,
is
ty
pi
c
a
ll
y us
e
d t
o ga
th
e
r
a
nd me
r
ge
t
he
out
put
f
e
a
tu
r
e
s
obt
a
in
e
d f
r
om
t
he
ba
c
kbone
. I
t
th
e
n s
e
nds
f
e
a
tu
r
e
m
a
ps
of
di
f
f
e
r
e
nt
s
iz
e
s
b
a
c
k t
o
t
he
de
t
e
c
ti
on he
a
d
s
.
T
hi
s
pr
oc
e
s
s
e
n
a
b
le
s
t
he
m
ode
l
to
de
t
e
c
t
obj
e
c
ts
of
v
a
r
yi
ng s
iz
e
s
w
it
hi
n
th
e
im
a
ge
.
F
in
a
ll
y,
th
e
de
te
c
ti
on
he
a
d
pe
r
f
or
m
s
c
a
lc
ul
a
ti
ons
f
or
bounding
box
r
e
gr
e
s
s
io
n
a
nd
pr
oba
bi
li
ty
e
s
ti
m
a
ti
on.
I
n
th
is
s
tu
dy
w
e
pr
opos
e
to
m
a
ke
a
m
odi
f
ic
a
ti
on
to
th
e
a
r
c
hi
te
c
tu
r
e
of
Y
O
L
O
v5
by
r
e
pl
a
c
in
g
it
s
ba
c
kbone
w
it
h t
he
popula
r
C
N
N
s
:
D
a
r
kN
e
t,
M
obi
le
N
e
t,
R
e
s
N
e
t
a
nd
vi
s
ua
l
ge
om
e
tr
y gr
oup (
V
G
G
)
,
a
s
s
how
n
in
F
ig
ur
e
1, a
nd t
o e
va
lu
a
te
t
he
pe
r
f
or
m
a
nc
e
of
e
a
c
h m
ode
l.
2.1.
Y
O
L
O
v3
Y
O
L
O
,
or
"
you
onl
y
lo
ok
onc
e
,"
is
a
w
e
ll
-
known
one
-
s
ta
ge
obj
e
c
t
de
te
c
to
r
th
a
t
pr
e
di
c
ts
th
e
lo
c
a
ti
on
a
nd
c
la
s
s
of
obj
e
c
ts
in
a
s
in
gl
e
pa
s
s
,
m
a
ki
ng
it
ge
n
e
r
a
ll
y
f
a
s
te
r
t
ha
n
de
te
c
to
r
s
us
in
g
r
e
gi
on
pr
opos
a
l
ne
twor
ks
(
R
P
N
)
[
17]
.
Y
O
L
O
v3,
one
of
th
e
m
os
t
popula
r
ve
r
s
io
ns
,
e
nha
nc
e
s
th
e
Y
O
L
O
a
r
c
hi
te
c
tu
r
e
,
r
e
s
ul
ti
ng
in
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
14
, N
o.
1
,
F
e
br
ua
r
y
2025
:
658
-
666
660
im
pr
ove
d
a
c
c
ur
a
c
y
a
nd
s
pe
e
d.
Y
O
L
O
v3'
s
a
r
c
hi
te
c
tu
r
e
c
ons
is
t
s
of
th
r
e
e
m
a
in
c
om
pone
nt
s
:
th
e
ba
c
kbone
,
ne
c
k,
a
nd
he
a
d.
I
t
us
e
s
D
a
r
kne
t5
3
a
s
th
e
ba
c
kbone
,
c
om
pr
is
in
g
53
c
onvolut
io
na
l
la
ye
r
s
in
s
pi
r
e
d
by
D
a
r
kne
t1
9,
in
c
lu
di
ng B
ot
tl
e
ne
c
k m
odul
e
s
w
it
h s
hor
tc
ut
c
onn
e
c
ti
ons
f
or
e
f
f
ic
ie
nt
f
e
a
tu
r
e
e
xt
r
a
c
ti
on.
F
ig
ur
e
1.
T
he
s
tr
uc
tu
r
e
of
a
m
ode
r
n obje
c
t
de
te
c
to
r
ne
twor
k
T
he
n
e
c
k
in
c
or
por
a
te
s
th
e
f
e
a
tu
r
e
pyr
a
m
id
ne
twor
k
(
F
P
N
)
m
e
th
od
to
f
us
e
lo
w
a
nd
hi
gh
-
le
ve
l
f
e
a
tu
r
e
s
,
ups
a
m
pl
in
g t
he
ba
c
kbone
'
s
out
put
a
nd c
om
bi
ni
ng i
t
w
it
h l
ow
-
le
ve
l
out
put
s
. T
he
he
a
d us
e
s
k
-
m
e
a
ns
c
lu
s
te
r
in
g
to
de
te
r
m
in
e
bounding
box
c
oor
di
na
te
s
,
di
s
tr
ib
ut
in
g
ni
ne
c
lu
s
t
e
r
s
a
m
ong
th
r
e
e
out
put
s
c
a
le
s
[
18]
,
[
19]
.
I
n
a
s
e
pa
r
a
te
s
tu
dy,
r
e
s
e
a
r
c
he
r
s
pr
opos
e
d
Y
O
L
O
v3
-
ti
ny,
a
c
om
pa
c
t
ve
r
s
io
n
us
in
g
D
a
r
kne
t1
9
a
s
th
e
ba
c
kbone
a
nd
a
s
im
pl
if
ie
d ne
c
k w
it
h a
s
in
gl
e
c
onc
a
te
na
ti
on.
Y
O
L
O
v3
-
ti
ny e
m
pl
oys
t
w
o f
e
a
tu
r
e
m
a
ps
of
di
f
f
e
r
e
nt
s
c
a
le
s
f
or
de
te
c
ti
on
pr
e
di
c
ti
ons
,
w
hi
c
h
a
r
e
th
e
n
pr
oc
e
s
s
e
d
by
a
s
im
il
a
r
s
e
ns
in
g
he
a
d
a
s
in
Y
O
L
O
v3
[
20]
.
T
hi
s
c
om
pa
c
t
ve
r
s
io
n
ba
la
nc
e
s
s
pe
e
d
a
nd
a
c
c
ur
a
c
y,
m
a
ki
ng
it
s
ui
ta
bl
e
f
or
e
m
be
dde
d
s
ys
te
m
s
w
it
h
li
m
it
e
d
c
om
put
a
ti
ona
l
r
e
s
our
c
e
s
.
2.2.
Y
O
L
O
v4
Y
O
L
O
v4
is
a
r
e
c
e
nt
it
e
r
a
ti
on
of
Y
O
L
O
d
e
te
c
to
r
s
,
f
e
a
t
ur
in
g
a
m
ode
r
n
a
r
c
hi
te
c
tu
r
e
w
it
h
C
S
P
D
a
r
kN
e
t5
3
a
s
it
s
ba
c
kbone
.
B
ui
lt
upon
th
e
D
a
r
kN
e
t
53
a
r
c
hi
te
c
tu
r
e
us
e
d
in
Y
O
L
O
v3,
C
S
P
D
a
r
kN
e
t5
3
in
c
lu
de
s
C
onv
m
odul
e
s
a
nd
a
ne
w
B
ot
tl
e
N
e
c
k
C
S
P
m
odul
e
in
s
pi
r
e
d
by
th
e
C
S
P
N
e
t
te
c
hni
que
[
21]
.
A
f
te
r
num
e
r
ous
e
xpe
r
im
e
nt
s
,
th
e
a
ut
hor
s
de
te
r
m
in
e
d
th
a
t
C
S
P
D
a
r
k
N
e
t5
3
is
th
e
opt
im
a
l
m
ode
l
c
om
pa
r
e
d
to
ot
he
r
ne
twor
ks
te
s
te
d.
Y
O
L
O
v4'
s
ne
c
k
in
c
or
por
a
te
s
a
n
a
ddi
ti
ona
l
S
P
P
m
odul
e
to
ge
ne
r
a
te
r
e
pr
e
s
e
nt
a
ti
ons
f
r
om
im
a
ge
s
of
a
r
bi
tr
a
r
y
s
i
z
e
s
[
22]
.
I
t
a
ls
o
u
s
e
s
P
A
N
e
t
in
th
e
n
e
c
k,
i
m
pr
ovi
ng
obj
e
c
t
lo
c
a
li
z
a
ti
on a
c
c
ur
a
c
y
th
r
ough
e
nha
nc
e
d
f
e
a
tu
r
e
f
us
io
n
f
r
om
bot
to
m
to
to
p
a
nd
to
p
to
bot
to
m
.
T
he
he
a
d
of
Y
O
L
O
v4
is
s
im
il
a
r
to
th
a
t
o
f
Y
O
L
O
v3.
A
not
he
r
s
tu
dy
in
tr
oduc
e
d
Y
O
L
O
v4
-
ti
ny,
a
s
m
a
ll
e
r
ve
r
s
io
n
w
it
h
a
r
e
duc
e
d
C
S
P
D
a
r
kN
e
t5
3
ba
c
kbone
f
e
a
tu
r
in
g
3
B
ot
tl
e
N
e
c
kC
S
P
m
odul
e
s
in
s
te
a
d
of
27.
T
he
e
xt
r
a
c
te
d
f
e
a
tu
r
e
s
a
r
e
pr
oc
e
s
s
e
d
by
a
r
e
duc
e
d
ne
c
k
ba
s
e
d
on
th
e
F
P
N
te
c
hni
que
,
s
im
il
a
r
to
Y
O
L
O
v3
-
ti
ny,
be
f
or
e
be
in
g
s
e
nt
to
th
e
Y
O
L
O
v3
he
a
d
f
or
pr
e
di
c
ti
ng
lo
c
a
ti
on a
nd c
la
s
s
[
23]
.
2.3.
Y
O
L
O
v5
Y
O
L
O
v5
,
a
r
e
c
e
nt
Y
O
L
O
v
e
r
s
io
n,
is
known
f
or
it
s
im
pr
e
s
s
iv
e
a
c
c
ur
a
c
y
a
nd
s
pe
e
d.
I
ts
a
r
c
hi
te
c
tu
r
e
c
lo
s
e
ly
r
e
s
e
m
bl
e
s
Y
O
L
O
v4,
us
in
g
th
e
C
S
P
D
a
r
kN
e
t5
3
ba
c
kbone
w
it
h
m
odi
f
ic
a
ti
ons
.
Y
O
L
O
v5
in
c
lu
de
s
uni
que
m
odul
e
s
li
ke
C
3,
c
om
pr
is
in
g
th
r
e
e
C
onv
m
odul
e
s
a
nd
a
B
ot
tl
e
N
e
c
kC
S
P
,
w
hi
c
h
m
it
ig
a
te
s
gr
a
di
e
nt
in
f
or
m
a
ti
on
dupl
ic
a
ti
on.
I
t
a
ls
o
f
e
a
tu
r
e
s
th
e
S
P
P
F
m
odul
e
f
o
r
im
p
r
ove
d
f
e
a
tu
r
e
e
xpr
e
s
s
io
n,
in
s
pi
r
e
d
by
S
P
P
ne
twor
ks
bu
t
f
a
s
te
r
.
Y
O
L
O
v5
'
s
ne
c
k
us
e
s
P
A
N
e
t
f
or
f
e
a
tu
r
e
f
us
io
n,
a
nd
th
e
out
put
s
a
r
e
s
e
nt
to
th
e
de
te
c
ti
on
he
a
d,
s
im
il
a
r
to
Y
O
L
O
v3, f
o
r
pr
e
di
c
ti
ng obje
c
t
lo
c
a
ti
on a
nd c
la
s
s
[
13]
.
2.4.
T
h
e
M
ob
il
e
N
e
t
n
e
t
w
or
k
s
M
obi
le
N
e
t
is
r
e
now
ne
d
f
or
it
s
s
pe
e
d
a
nd
li
ght
w
e
ig
ht
a
r
c
hi
te
c
t
ur
e
,
id
e
a
l
f
or
m
obi
le
a
ppl
ic
a
ti
ons
.
I
t
ut
il
iz
e
s
de
pt
hw
is
e
s
e
p
a
r
a
bl
e
c
onvolut
io
ns
(
D
W
C
onv)
f
or
e
f
f
ic
ie
nt
f
e
a
tu
r
e
e
xt
r
a
c
ti
on,
c
ons
is
ti
ng
of
a
de
pt
hw
is
e
c
onvolut
io
n
la
ye
r
f
ol
lo
w
e
d
by
ba
tc
h
nor
m
a
li
z
a
ti
on
(
B
a
tc
hN
or
m
)
a
nd
R
e
L
U
a
c
ti
va
ti
on.
M
obi
le
N
e
t
v
1
us
e
s
D
W
C
onv
-
s
e
p
bl
oc
k
s
,
w
hi
c
h
in
c
lu
de
D
W
C
onv
m
odul
e
s
f
ol
lo
w
e
d
by
C
onv
m
odul
e
s
[
24]
.
M
obi
le
N
e
t
v
2
in
tr
oduc
e
s
B
ot
tl
e
N
e
c
k
-
M
ob
bl
oc
ks
,
f
e
a
tu
r
in
g
c
onvolut
io
n
la
ye
r
s
w
it
h
a
n
e
xpa
ns
io
n
la
ye
r
a
t
th
e
in
put
a
nd
a
pr
oj
e
c
ti
on l
a
ye
r
a
t
th
e
out
put
, i
nc
or
por
a
ti
ng s
hor
tc
ut
c
onne
c
ti
ons
a
nd t
he
D
W
C
onv modul
e
a
t
th
e
ir
c
or
e
[
25]
.
2.5.
T
h
e
R
e
s
N
e
t
m
od
e
ls
R
e
s
N
e
t
w
a
s
in
tr
oduc
e
d
to
ta
c
kl
e
th
e
va
ni
s
hi
ng
gr
a
di
e
nt
pr
obl
e
m
in
tr
a
in
in
g
de
e
p
ne
ur
a
l
ne
twor
ks
.
R
e
s
N
e
t
m
ode
ls
a
r
e
s
im
il
a
r
t
o V
G
G
ne
twor
ks
but
i
nc
lu
de
s
hor
tc
ut
c
onne
c
ti
ons
t
h
a
t
li
nk t
he
i
nput
a
nd output
of
e
a
c
h
m
odul
e
,
e
ns
ur
in
g
s
m
oot
h
gr
a
di
e
nt
f
lo
w
.
T
he
s
e
c
onne
c
ti
o
ns
he
lp
m
it
ig
a
te
th
e
va
ni
s
hi
ng
gr
a
di
e
nt
is
s
ue
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
E
nhanc
e
m
e
nt
of
Y
O
L
O
v
5 f
or
aut
om
at
ic
w
e
e
d d
e
te
c
ti
on t
hr
ough
bac
k
bone
…
(
M
ohamm
e
d H
abi
b)
661
D
e
pe
ndi
ng
on
th
e
m
ode
l
s
i
z
e
,
R
e
s
N
e
t
c
a
n
c
ont
a
in
di
f
f
e
r
e
nt
n
um
be
r
s
of
R
e
s
C
onv
m
odul
e
s
,
s
uc
h
a
s
34,
50,
or
101
[
26]
.
2.6.
T
h
e
V
G
G
m
od
e
ls
V
G
G
ne
twor
ks
a
r
e
r
e
c
ogni
z
e
d
a
s
f
unda
m
e
nt
a
l
c
onvolu
ti
ona
l
ne
twor
ks
known
f
or
th
e
ir
s
tr
a
ig
ht
f
or
w
a
r
d
a
r
c
hi
te
c
tu
r
e
.
T
he
y
c
ons
is
t
of
a
s
e
que
nc
e
of
c
onv
ol
ut
io
na
l
la
ye
r
s
(
C
onv)
s
ta
c
ke
d
on
to
p
of
e
a
c
h
ot
he
r
.
T
he
va
r
io
us
ve
r
s
io
ns
of
V
G
G
a
r
e
di
s
ti
ngui
s
he
d
by
th
e
n
um
be
r
of
c
onvolut
io
na
l
la
ye
r
s
e
m
pl
oye
d,
s
uc
h
a
s
V
G
G
16
a
nd
V
G
G
19
[
27]
.
2.7.
G
e
n
e
r
al
iz
e
d
i
n
t
e
r
s
e
c
t
io
n
ove
r
u
n
io
n
l
os
s
W
he
n
tr
a
in
in
g
obj
e
c
t
de
te
c
to
r
s
,
th
e
lo
s
s
i
s
de
te
r
m
in
e
d
by
two
f
unc
ti
ons
,
th
e
c
la
s
s
if
ic
a
ti
on
lo
s
s
a
nd
th
e
bounding
box
r
e
gr
e
s
s
io
n
lo
s
s
.
T
he
m
os
t
c
om
m
onl
y
us
e
d
lo
s
s
f
unc
ti
on
f
or
bounding
box
r
e
gr
e
s
s
io
n
is
th
e
in
te
r
s
e
c
ti
on ove
r
uni
on (
I
oU
)
a
nd i
ts
de
r
iv
a
ti
ons
. T
he
I
oU
i
s
de
f
in
e
d by
(
1)
.
(
,
)
=
∩
∪
(
1)
w
he
r
e
A
i
s
t
he
pr
e
di
c
ti
on a
nd B
i
s
t
he
t
r
ue
bounding box.
T
he
lo
s
s
f
unc
ti
on
i
s
e
xpr
e
s
s
e
d
a
s
a
di
s
ta
nc
e
,
L
os
s
I
oU
=
1
−
I
oU
.
T
he
r
e
a
r
e
ot
he
r
f
unc
ti
ons
d
e
r
iv
e
d
f
r
om
I
oU
th
a
t
e
na
bl
e
th
e
e
v
a
lu
a
ti
on
of
th
e
di
s
s
im
il
a
r
it
y
be
twe
e
n
th
e
pr
e
di
c
te
d
a
nd
tr
ue
bounding
boxe
s
,
e
v
e
n
in
c
a
s
e
s
w
he
r
e
th
e
y
do
not
ove
r
la
p.
O
ne
s
uc
h
f
unc
ti
on
is
th
e
ge
ne
r
a
li
z
e
d
in
te
r
s
e
c
ti
on
ove
r
uni
on
(
G
I
oU
)
,
w
hi
c
h
a
ddr
e
s
s
e
s
th
e
is
s
u
e
of
di
s
jo
in
t
bounding
boxe
s
(
A
a
nd
B
)
.
G
I
oU
pos
s
e
s
s
e
s
th
e
s
a
m
e
s
c
a
le
in
va
r
ia
nc
e
pr
ope
r
ti
e
s
a
s
I
oU
a
nd i
s
de
f
in
e
d by the
(
2)
a
nd (
3)
.
(
,
)
=
(
,
)
−
|
|
−
|
∪
|
|
|
(
2)
=
1
−
(
,
)
=
1
−
(
,
)
+
|
|
−
|
∪
|
|
|
(
3)
w
it
h C
i
s
t
he
s
m
a
ll
box tha
t
e
n
c
lo
s
e
s
t
he
boxe
s
A
a
nd B
[
13]
.
I
n
th
is
s
tu
dy,
w
e
w
il
l
ut
il
iz
e
th
e
G
I
oU
f
unc
ti
on
to
c
a
lc
ul
a
te
th
e
bounding
box
r
e
gr
e
s
s
io
n
lo
s
s
a
t
e
a
c
h
it
e
r
a
ti
on.
F
or
c
la
s
s
if
ic
a
ti
on,
w
e
us
e
d
th
e
bi
na
r
y
c
r
os
s
e
nt
r
opy
w
it
h
lo
gi
ts
lo
s
s
f
unc
ti
on
f
r
om
th
e
Y
O
L
O
v5
m
ode
l
tr
a
in
in
g.
T
he
B
C
E
W
it
hL
ogi
ts
lo
s
s
c
om
bi
ne
s
th
e
f
unc
ti
ona
li
ti
e
s
of
two
S
ig
m
oi
d
f
unc
ti
ons
a
nd
th
e
B
C
E
L
os
s
by
us
in
g t
he
l
og
-
s
um
-
e
xp me
th
od. T
he
B
C
E
L
os
s
W
it
hL
ogi
ts
i
s
de
f
in
e
d a
s
(
4)
.
ℓ
(
,
)
=
=
{
1
,
…
,
}
,
=
−
[
⋅
(
)
+
(
1
−
)
⋅
(
1
−
(
)
)
]
(
4)
N
is
th
e
ba
tc
h
s
iz
e
.
T
he
r
e
s
ul
t
is
a
m
or
e
num
e
r
ic
a
ll
y
s
ta
bl
e
f
un
c
ti
on
c
om
pa
r
e
d
to
us
in
g
S
ig
m
oi
d
f
ol
lo
w
e
d
by
B
C
E
L
os
s
s
e
p
a
r
a
te
ly
[
28]
.
2.8.
D
at
as
e
t
ac
q
u
is
it
io
n
D
a
ta
S
e
t
a
c
qui
s
it
io
n
s
ti
ll
r
e
m
a
in
s
one
of
t
he
gr
e
a
t
c
ha
ll
e
nge
s
of
D
e
e
pL
e
a
r
ni
ng, a
gr
e
a
t
D
a
t
a
S
e
t
m
e
a
ns
a
good
le
a
r
ni
ng.
I
n
our
s
tu
dy,
w
e
ut
il
iz
e
d
a
pr
e
vi
ous
ly
c
ol
le
c
te
d
a
nd
pr
e
pa
r
e
d
D
a
ta
S
e
t
f
r
om
our
e
a
r
li
e
r
w
or
k
[
3]
f
or
m
ode
l
t
r
a
in
in
g.
T
he
im
a
ge
s
w
e
ha
ve
c
ol
le
c
te
d
c
ont
a
in
t
he
pe
a
c
r
op
a
nd
th
e
w
e
e
ds
w
it
h
th
e
s
e
di
f
f
e
r
e
nt
s
pe
c
ie
s
.
F
ig
ur
e
2
il
lu
s
tr
a
te
s
th
e
te
c
hni
que
s
e
m
pl
oye
d
f
or
im
a
g
e
a
c
qui
s
it
io
n.
T
o
f
ur
th
e
r
e
nha
nc
e
th
e
D
a
ta
S
e
t,
w
e
e
m
pl
oye
d i
m
a
ge
pr
oc
e
s
s
in
g t
ool
s
f
or
da
ta
a
ugm
e
nt
a
ti
on.
F
ig
ur
e
2.
T
h
e
m
a
t
e
r
i
a
l
s
u
s
e
d f
or
t
he
a
c
q
ui
s
i
ti
o
n o
f
d
a
t
a
s
e
t
a
di
gi
t
a
l
c
a
m
e
r
a
w
a
s
f
i
xe
d
in
40
c
m
f
r
o
m
t
h
e
gr
ou
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
14
, N
o.
1
,
F
e
br
ua
r
y
2025
:
658
-
666
662
F
or
im
a
ge
l
a
be
li
ng, we
ut
il
iz
e
d t
he
Y
O
L
O
m
ode
ls
'
s
pe
c
if
ic
a
nn
ot
a
ti
on f
or
m
a
t.
E
a
c
h
bounding box
is
c
ha
r
a
c
te
r
iz
e
d by
t
he
c
oor
di
na
te
s
of
i
ts
c
e
nt
e
r
poi
nt
(
x, y)
, i
ts
l
e
n
gt
h, w
id
th
, a
nd t
he
c
or
r
e
s
ponding obj
e
c
t
c
la
s
s
.
T
he
s
e
c
oor
di
na
te
s
a
r
e
nor
m
a
li
z
e
d
a
nd
s
a
ve
d
in
a
te
xt
f
il
e
.
T
he
D
a
ta
S
e
t
w
a
s
di
vi
de
d
in
to
th
r
e
e
s
ub
s
e
ts
:
7
,
360 im
a
ge
s
f
or
t
r
a
in
in
g, 1
,
840 im
a
ge
s
f
or
va
li
da
ti
on, a
nd 54 im
a
ge
s
f
or
t
e
s
ti
ng.
3.
R
E
S
U
L
T
S
A
N
D
D
I
S
C
U
S
S
I
O
N
I
n
th
is
s
e
c
ti
on,
w
e
w
il
l
pr
e
s
e
nt
th
e
r
e
s
ul
ts
of
tr
a
in
in
g
a
nd
va
l
id
a
ti
on
of
obj
e
c
t
de
te
c
to
r
s
us
in
g
th
e
pr
e
vi
ous
ly
di
s
c
us
s
e
d
ba
c
kbon
e
a
r
c
hi
te
c
tu
r
e
s
.
W
e
w
il
l
e
va
lu
a
t
e
th
e
ir
pe
r
f
or
m
a
nc
e
in
c
r
op
id
e
nt
if
ic
a
ti
on
a
nd
ul
tr
a
-
lo
c
a
li
z
a
ti
on. T
he
m
ode
ls
w
e
w
il
l
e
xa
m
in
e
a
r
e
de
r
iv
e
d
f
r
o
m
t
he
ne
ur
a
l
ne
twor
ks
a
lr
e
a
dy s
tu
di
e
d.
T
a
bl
e
1
pr
ovi
de
s
a
n
ove
r
vi
e
w
of
th
e
c
ons
tr
uc
te
d
m
ode
ls
,
in
c
lu
di
ng
t
he
ir
s
iz
e
a
nd
c
om
put
a
ti
ona
l
r
e
qui
r
e
m
e
nt
s
.
I
t
s
um
m
a
r
iz
e
s
t
he
c
h
a
r
a
c
te
r
is
ti
c
s
of
e
a
c
h m
ode
l.
T
a
bl
e
1
.
T
he
pr
ope
r
ti
e
s
of
t
he
t
r
a
in
e
d m
ode
ls
M
ode
l
N
a
m
e
B
a
c
kbone
N
e
c
k
L
a
ye
r
s
P
a
r
a
m
e
t
e
r
s
G
f
l
ops
Y
O
L
O
v5
C
S
P
D
a
r
kN
e
t
53 (
m
od)
P
A
N
e
t
214
7235389
16.6
Y
O
L
O
v4
C
S
P
D
a
r
kN
e
t
53
P
A
N
e
t
383
40020445
102.1
Y
O
L
O
v3
D
a
r
kN
e
t
53
FPN
262
61949149
156.6
Y
O
L
O
v4
-
T
i
ny
C
S
P
D
a
r
kN
e
t
19
FPN
-
t
i
ny
91
3105526
6.5
Y
O
L
O
v3
-
T
i
ny
D
a
r
kN
e
t
19
FPN
-
t
i
ny
49
8852366
13.3
V
G
G
16
V
G
G
16
FPN
-
t
i
ny
68
25996222
272.1
V
G
G
19
V
G
G
19
FPN
-
t
i
ny
77
31307198
340.1
M
obne
t
v1
M
obi
l
e
ne
t
v1
FPN
-
t
i
ny
105
16973630
13.3
M
obne
t
v2
M
obi
l
e
ne
t
v2
FPN
-
t
i
ny
200
16774670
18.9
R
e
s
ne
t
50
R
e
s
N
e
t
50
FPN
-
t
i
ny
284
61283390
89.5
T
he
G
f
lo
ps
r
e
f
e
r
s
to
th
e
num
be
r
of
f
lo
a
ti
ng
-
poi
nt
ope
r
a
ti
ons
pe
r
s
e
c
ond,
e
xpr
e
s
s
e
d
in
G
ig
a
.
A
s
s
ho
w
n
in
th
e
T
a
bl
e
1,
Y
O
L
O
v4,
Y
O
L
O
v3,
a
nd
V
G
G
m
ode
ls
ha
ve
hi
gh
c
om
put
a
ti
ona
l
r
e
qui
r
e
m
e
nt
s
,
in
di
c
a
ti
ng
th
a
t
th
e
ir
e
xe
c
ut
io
n
a
nd
tr
a
in
in
g
ti
m
e
s
w
il
l
be
r
e
la
ti
ve
ly
lo
ng.
O
n
th
e
ot
he
r
ha
nd,
Y
O
L
O
-
ti
ny
a
nd
M
obi
le
N
e
t
m
ode
ls
ha
ve
lo
w
e
r
c
om
put
a
ti
ona
l
de
m
a
nds
,
a
ll
ow
in
g
th
e
m
to
be
e
x
e
c
ut
e
d
e
f
f
ic
ie
nt
ly
e
ve
n
on
lo
w
-
pe
r
f
or
m
a
nc
e
pr
oc
e
s
s
or
s
.
3.1.
T
r
ai
n
in
g p
r
e
p
ar
at
io
n
W
e
bui
lt
th
e
s
e
m
ode
ls
w
it
h
th
e
f
unc
ti
ona
li
ty
of
th
e
P
yt
or
c
h
li
br
a
r
y
w
it
h
w
hi
c
h
th
e
Y
O
L
O
v
5
m
ode
l
w
a
s
w
r
it
te
n.
W
e
tr
a
in
e
d,
v
a
li
da
te
d
a
nd
te
s
te
d
a
ll
th
e
c
r
e
a
te
d
m
ode
ls
unde
r
th
e
s
a
m
e
c
ondi
ti
ons
in
te
r
m
s
of
da
ta
s
e
t,
pr
e
-
pr
oc
e
s
s
in
g
te
c
hni
que
s
,
a
nd
us
in
g
th
e
T
e
s
la
T
4
gr
a
phi
c
s
pr
oc
e
s
s
or
.
T
he
hype
r
pa
r
a
m
e
te
r
s
f
or
m
ode
l
tr
a
in
in
g a
r
e
l
is
te
d i
n T
a
bl
e
2.
T
a
bl
e
2
.
T
he
hype
r
-
pa
r
a
m
e
te
r
s
of
t
he
t
r
a
in
in
g of
t
he
m
ode
ls
E
poc
hs
W
a
r
m
up e
poc
hs
B
a
t
c
h S
i
z
e
I
m
g. S
i
z
e
l
r
0
O
pt
i
m
i
z
e
r
M
om
e
nt
um
L
os
s
f
unc
t
i
on
50
3
16
320
0.01
S
G
D
0.937
B
C
E
W
i
t
hL
ogi
t
s
L
os
s
,
G
I
oU
3.2.
V
al
id
at
io
n
r
e
s
u
lt
s
A
f
te
r
tr
a
in
in
g
our
m
ode
ls
,
w
e
va
li
da
te
d
th
e
ir
pe
r
f
or
m
a
nc
e
by
i
nt
r
oduc
in
g
th
e
va
li
da
ti
on
im
a
ge
s
a
nd
c
om
pa
r
in
g t
he
pr
e
di
c
te
d bounding
boxe
s
w
it
h t
he
t
r
ue
bounding
boxe
s
.
W
e
v
a
r
ie
d t
he
s
iz
e
of
t
h
e
i
nput
i
m
a
ge
s
f
r
om
320 to 640 t
o obs
e
r
ve
i
ts
i
m
pa
c
t
on t
he
r
e
s
ul
ts
.
P
r
e
c
is
io
n a
nd r
e
c
a
ll
a
r
e
t
he
m
os
t
c
om
m
onl
y us
e
d m
e
tr
ic
s
f
or
e
va
lu
a
ti
ng t
he
pe
r
f
or
m
a
nc
e
of
a
C
N
N
dur
in
g va
li
da
ti
on. T
he
y a
r
e
de
f
in
e
d a
s
(
5)
a
nd (
6)
:
=
+
(
5)
=
+
(
6)
T
P
r
e
pr
e
s
e
nt
in
g
tr
ue
pos
it
iv
e
,
T
N
f
o
r
t
r
ue
ne
ga
ti
ve
,
F
P
f
or
f
a
ls
e
pos
it
iv
e
,
a
nd
F
N
f
o
r
f
a
ls
e
ne
ga
ti
ve
.
I
n
our
c
a
s
e
,
th
e
pos
it
iv
e
c
la
s
s
c
or
r
e
s
ponds
t
o p
e
a
s
, w
hi
le
t
h
e
ne
ga
ti
ve
c
la
s
s
r
e
pr
e
s
e
nt
s
t
he
b
a
c
kgr
ound
[
5]
.
P
r
e
c
is
io
n
r
e
c
a
ll
(
PR
)
c
ur
ve
s
a
r
e
us
e
d
to
r
e
pr
e
s
e
nt
pr
e
c
is
io
n
v
e
r
s
us
r
e
c
a
ll
,
w
h
e
r
e
pr
e
c
is
io
n
m
e
a
s
ur
e
s
r
e
s
ul
t
a
c
c
ur
a
c
y
a
nd
r
e
c
a
ll
m
e
a
s
ur
e
s
th
e
r
e
le
va
nc
e
of
r
e
s
ul
ts
.
T
he
x
-
a
xi
s
of
P
R
c
ur
ve
s
s
how
s
r
e
c
a
ll
,
a
nd
th
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
E
nhanc
e
m
e
nt
of
Y
O
L
O
v
5 f
or
aut
om
at
ic
w
e
e
d d
e
te
c
ti
on t
hr
ough
bac
k
bone
…
(
M
ohamm
e
d H
abi
b)
663
y
-
a
xi
s
s
how
s
pr
e
c
is
io
n.
F
ig
ur
e
3
di
s
pl
a
ys
P
R
c
ur
ve
s
f
or
m
ode
l
va
li
da
ti
on
a
c
r
os
s
va
r
io
us
I
oU
th
r
e
s
hol
ds
.
P
R
c
ur
ve
s
a
r
e
c
om
m
on
in
bi
na
r
y
c
la
s
s
if
ic
a
ti
on
ta
s
ks
,
li
ke
di
s
ti
ngui
s
hi
ng
be
twe
e
n
P
e
a
a
nd
ba
c
kgr
ound,
to
e
va
lu
a
te
m
ode
l
pe
r
f
or
m
a
nc
e
[
29
]
. A
h
ig
he
r
a
r
e
a
unde
r
t
he
c
ur
ve
(
A
U
C
)
in
di
c
a
te
s
be
tt
e
r
a
c
c
ur
a
c
y a
nd r
e
c
a
ll
, w
it
h hi
gh
pr
e
c
is
io
n s
ig
ni
f
yi
ng f
e
w
e
r
f
a
ls
e
pos
it
iv
e
s
a
nd high r
e
c
a
ll
i
ndi
c
a
ti
ng f
e
w
e
r
f
a
ls
e
ne
ga
ti
ve
s
.
F
r
om
F
ig
ur
e
s
3
(
a
)
to
3
(
d)
,
it
is
e
vi
de
nt
th
a
t
th
e
V
G
G
,
R
e
s
N
e
t5
0,
a
nd
Y
O
L
O
v3
m
ode
ls
e
xhi
bi
te
d
s
upe
r
io
r
A
U
C
va
lu
e
s
c
om
pa
r
e
d
to
th
e
ot
he
r
m
ode
ls
a
c
r
os
s
a
ll
I
oU
th
r
e
s
hol
ds
.
C
onve
r
s
e
ly
,
th
e
M
obi
le
N
e
ts
,
Y
O
L
O
v5
, a
nd
Y
O
L
O
v4
-
ti
ny mode
ls
a
c
hi
e
v
e
d t
he
m
in
im
um
A
U
C
va
lu
e
s
.
I
t
is
w
or
th
not
in
g t
ha
t
th
e
Y
O
L
O
v3
m
ode
l
r
e
m
a
in
e
d
r
e
la
ti
ve
ly
s
ta
bl
e
in
te
r
m
s
of
A
U
C
a
s
th
e
I
o
U
th
r
e
s
hol
d
va
r
ie
d.
H
ow
e
ve
r
,
th
e
M
obi
le
N
e
ts
m
ode
ls
de
m
ons
tr
a
te
d
s
ig
ni
f
ic
a
nt
c
h
a
nge
s
,
pa
r
ti
c
ul
a
r
ly
in
pr
e
c
i
s
io
n,
w
it
h
va
r
yi
ng
I
oU
th
r
e
s
hol
d
s
.
R
e
ga
r
di
ng
r
e
c
a
ll
,
th
e
r
e
w
e
r
e
no
s
ig
ni
f
ic
a
nt
di
f
f
e
r
e
nc
e
s
obs
e
r
ve
d
a
m
ong
th
e
m
ode
ls
or
a
c
r
os
s
th
e
di
f
f
e
r
e
nt
I
oU
th
r
e
s
hol
ds
.
(
a
)
(
b)
(
c
)
(
d)
F
ig
ur
e
3. T
he
P
R
_c
ur
ve
s
of
t
he
va
li
da
ti
on of
t
he
s
tu
di
e
d m
ode
l
s
w
it
h a
n I
oU
t
hr
e
s
hol
d
:
(
a
)
0.5, (
b)
0.75,
(
c
)
0.85
,
a
nd (
d)
0.9
I
n
or
de
r
to
c
la
r
if
y
th
e
r
e
s
ul
t
s
of
m
ode
l
v
a
li
da
ti
on,
w
e
us
e
d
one
of
th
e
m
e
tr
ic
s
th
a
t
be
tt
e
r
r
e
f
le
c
ts
th
e
A
U
C
i
s
t
he
a
ve
r
ge
pr
e
c
is
io
n (
AP
)
.
I
t
is
de
f
in
e
d by
(
7)
[
3]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
14
, N
o.
1
,
F
e
br
ua
r
y
2025
:
658
-
666
664
=
∫
(
)
1
0
(
7)
T
he
r
e
s
ul
ts
of
A
P
va
li
da
ti
on
of
th
e
m
ode
ls
s
tu
di
e
d
in
f
unc
ti
on
of
I
oU
th
r
e
s
hol
d
a
s
w
e
ll
a
s
th
e
s
pe
e
d
of
in
f
e
r
e
nc
e
(
de
te
c
ti
on)
f
or
a
s
in
gl
e
i
m
a
ge
a
r
e
s
how
n i
n T
a
bl
e
3.
T
he
T
a
bl
e
3
f
ur
th
e
r
s
uppor
ts
th
e
f
in
di
ngs
de
pi
c
te
d
in
th
e
P
R
c
ur
ve
s
.
T
he
V
G
G
,
R
e
s
N
e
t,
a
nd
Y
O
L
O
m
ode
ls
e
xhi
bi
t
th
e
hi
ghe
s
t
pr
e
c
is
io
n
v
a
lu
e
s
a
nd
de
m
on
s
tr
a
te
s
t
a
bi
li
ty
in
pr
e
c
is
io
n
a
s
th
e
I
oU
th
r
e
s
hol
d
v
a
r
ie
s
.
H
ow
e
ve
r
,
th
e
s
e
m
ode
l
s
te
nd
to
be
s
lo
w
e
r
in
te
r
m
s
of
e
xe
c
ut
io
n
ti
m
e
.
O
n
th
e
ot
he
r
ha
nd,
Y
O
L
O
v
5
s
ta
nds
out
a
s
a
f
a
s
t
m
ode
l
w
hi
le
m
a
in
ta
in
in
g
good
p
r
e
c
is
io
n.
T
he
Y
O
L
O
-
ti
ny,
M
obi
le
N
e
ts
,
a
nd
Y
O
L
O
v
5
m
ode
ls
a
r
e
id
e
nt
if
ie
d a
s
t
he
f
a
s
te
s
t
m
ode
l
s
, pr
io
r
it
iz
in
g c
om
put
a
ti
ona
l
e
f
f
ic
ie
nc
y.
T
a
bl
e
3
.
T
he
va
li
da
ti
on r
e
s
ul
ts
of
t
he
m
ode
l
s
s
tu
di
e
d a
s
a
f
unc
ti
on of
I
oU
t
hr
e
s
hol
d
T
he
m
ode
l
AP
50
AP
75
AP
85
AP
90
S
pe
e
d (
m
s
)
Y
O
L
O
v5
0.881
0.874
0.852
0.798
10
Y
O
L
O
v4
0.906
0.890
0.859
0.818
29
Y
O
L
O
v3
0.915
0.912
0.898
0.871
28.4
Y
O
L
O
v4
-
T
i
ny
0.881
0.851
0.769
0.654
4.7
Y
O
L
O
v3
-
T
i
ny
0.891
0.877
0.806
0.663
5.9
V
G
G
16
0.946
0.935
0.898
0.830
32.5
V
G
G
19
0.945
0.937
0.900
0.822
40.4
M
obi
l
e
N
e
t
v1
0.893
0.862
0.752
0.593
10.1
M
obi
l
e
N
e
t
v2
0.865
0.842
0.788
0.657
13.9
R
e
s
N
e
t
50
0.923
0.917
0.896
0.841
27.7
F
a
s
t
e
r
R
C
N
N
R
e
s
N
e
t
50
0.957
0.956
-
-
136
SSD
-
M
obi
l
e
N
e
t
v2
0.956
0.899
-
-
17
4.
D
I
S
C
U
S
S
I
O
N
T
hi
s
s
e
c
ti
on
e
va
lu
a
te
s
th
e
r
e
s
ul
ts
a
nd
li
m
it
a
ti
ons
of
our
m
e
th
ods
,
f
oc
us
in
g
on
id
e
nt
if
yi
ng
th
e
be
s
t
ba
c
kbone
to
e
nha
nc
e
Y
O
L
O
v5
'
s
pe
r
f
or
m
a
nc
e
w
hi
le
pr
e
s
e
r
vi
ng
it
s
ot
he
r
pa
r
a
m
e
te
r
s
(
ne
c
k
a
nd
he
a
d
)
.
W
e
di
s
c
ove
r
e
d
th
a
t
V
G
G
,
R
e
s
N
e
t5
0,
a
nd
Y
O
L
O
v3
m
ode
l
s
,
w
hi
c
h
de
m
ons
tr
a
te
d
hi
gh
pr
e
c
is
io
n
a
nd
r
e
c
a
ll
,
a
r
e
e
f
f
e
c
ti
ve
f
or
a
c
hi
e
vi
ng
a
c
c
ur
a
te
r
e
s
ul
ts
a
nd
c
a
pt
ur
in
g
m
os
t
pos
it
iv
e
in
s
ta
nc
e
s
.
H
ow
e
ve
r
,
c
on
s
id
e
r
in
g
th
e
c
ons
tr
a
in
ts
of
lo
w
-
c
a
pa
c
it
y
c
om
put
in
g
s
ys
te
m
s
,
pr
io
r
it
iz
in
g
f
a
s
te
r
m
ode
ls
li
ke
M
obi
le
N
e
t,
Y
O
L
O
-
ti
ny,
a
nd
Y
O
L
O
v5
is
r
e
a
s
ona
bl
e
due
to
th
e
ir
good
ba
la
nc
e
of
s
pe
e
d
a
nd
a
c
c
ur
a
c
y,
m
a
ki
ng
th
e
m
s
ui
ta
bl
e
f
or
de
pl
oym
e
n
t
in
r
e
s
our
c
e
-
li
m
it
e
d
e
nvi
r
onm
e
nt
s
.
F
or
a
n
I
oU
th
r
e
s
hol
d
of
0
.5,
M
obi
le
N
e
tv
1
a
c
hi
e
ve
d
th
e
hi
ghe
s
t
A
U
C
,
r
e
f
le
c
ti
ng
s
tr
ong
pe
r
f
or
m
a
nc
e
in
pr
e
c
i
s
io
n
a
nd
r
e
c
a
ll
.
Y
e
t,
in
c
r
e
a
s
in
g
th
e
I
oU
th
r
e
s
hol
d
c
a
us
e
s
a
dr
op
in
pr
e
c
is
io
n,
a
s
il
lu
s
tr
a
te
d
in
F
ig
ur
e
s
4
(
a
)
a
nd
4
(
b)
,
w
he
r
e
M
obi
le
N
e
tv
1’
s
bounding
box
f
a
il
s
to
a
c
c
ur
a
te
ly
e
nc
lo
s
e
th
e
obj
e
c
t
c
om
pa
r
e
d t
o
Y
O
L
O
v5
[
29]
.
(
a
)
(
b)
F
ig
ur
e
4.
C
om
pa
r
is
on be
twe
e
n t
he
pr
e
di
c
ti
on of
(
a
)
M
obi
le
ne
tv
1 a
nd
(
b)
Y
O
L
O
v
5
m
ode
ls
Y
O
L
O
v5
s
ta
nds
out
a
s
th
e
m
o
s
t
s
ta
bl
e
a
m
ong
f
a
s
t
m
ode
l
s
,
m
a
i
nt
a
in
in
g
c
ons
is
te
nt
a
ve
r
a
ge
pr
e
c
is
io
n
(
A
P
)
a
c
r
os
s
v
a
r
yi
ng
I
oU
th
r
e
s
hol
ds
a
nd
a
c
c
ur
a
te
ly
s
ur
r
ounding
th
e
pe
a
c
r
op.
I
n
c
ont
r
a
s
t,
M
obi
le
N
e
tv
1
e
xc
e
ls
a
t
a
n I
oU
t
hr
e
s
hol
d of
0.5 but
s
tr
uggl
e
s
w
it
h pr
e
c
is
e
l
oc
a
li
z
a
ti
on a
nd s
uf
f
e
r
s
f
r
om
oc
c
lu
s
io
n e
r
r
or
s
, w
hi
c
h c
a
n
le
a
d
to
pa
r
ts
of
th
e
c
r
op
b
e
in
g
e
xc
l
ude
d
f
r
om
th
e
bounding
b
ox
a
nd
m
is
c
l
a
s
s
if
ie
d
a
s
w
e
e
ds
in
s
ubs
e
que
nt
s
e
gm
e
nt
a
ti
on.
T
h
e
r
e
f
or
e
,
w
he
n
s
e
le
c
ti
ng
a
m
ode
l,
it
is
im
por
ta
nt
to
c
ons
id
e
r
bot
h
de
te
c
ti
on
c
a
p
a
bi
li
ti
e
s
a
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
A
r
ti
f
I
nt
e
ll
I
S
S
N
:
2252
-
8938
E
nhanc
e
m
e
nt
of
Y
O
L
O
v
5 f
or
aut
om
at
ic
w
e
e
d d
e
te
c
ti
on t
hr
ough
bac
k
bone
…
(
M
ohamm
e
d H
abi
b)
665
s
ta
bi
li
ty
a
c
r
os
s
di
f
f
e
r
e
nt
I
oU
th
r
e
s
hol
ds
.
S
ol
ut
io
ns
to
im
pr
ove
pe
r
f
or
m
a
nc
e
in
c
lu
de
e
xpa
ndi
ng
th
e
bounding
box to e
ns
ur
e
c
om
pl
e
te
c
ove
r
a
g
e
or
m
odi
f
yi
ng t
he
m
ode
l
a
r
c
hi
te
c
tu
r
e
t
o e
nha
nc
e
r
e
s
ul
ts
.
5.
C
O
N
C
L
U
S
I
O
N
T
hi
s
s
tu
dy
s
uc
c
e
s
s
f
ul
ly
ut
il
iz
e
d
e
s
ta
bl
is
he
d
C
N
N
s
to
de
ve
lo
p
a
n
obj
e
c
t
de
te
c
to
r
c
a
pa
bl
e
of
a
c
c
ur
a
te
ly
lo
c
a
li
z
in
g c
r
ops
a
m
id
s
t
w
e
e
ds
i
n r
e
a
l
-
ti
m
e
i
m
a
ge
s
. I
ni
t
ia
l
hypot
he
s
e
s
w
e
r
e
va
li
da
te
d, a
nd f
in
di
ngs
unde
r
s
c
or
e
th
e
s
upe
r
io
r
it
y
of
m
ode
ls
ba
s
e
d
on V
G
G
,
R
e
s
N
e
t,
Y
O
L
O
v3,
a
n
d
Y
O
L
O
v4
in
te
r
m
s
of
a
c
c
ur
a
c
y,
de
s
pi
te
th
e
ir
hi
gh c
om
put
a
ti
ona
l
de
m
a
nds
.
H
o
w
e
ve
r
, f
or
our
a
ppl
ic
a
ti
on a
im
in
g t
o i
nt
e
gr
a
te
t
hi
s
de
te
c
to
r
i
nt
o a
n e
m
b
e
dde
d
s
ys
te
m
c
a
pa
bl
e
of
s
im
ul
a
ti
ng
m
a
nua
l
w
e
e
d
r
e
m
ov
a
l,
s
pe
e
d
is
c
r
uc
ia
l.
T
h
e
f
a
s
t
e
s
t
m
ode
l
s
li
ke
M
obi
le
N
e
t,
Y
O
L
O
-
ti
ny,
a
nd
Y
O
L
O
v5
pe
r
f
or
m
e
d
w
e
ll
,
w
it
h
s
pe
e
d
s
r
a
n
gi
ng
f
r
om
5
to
14
m
il
li
s
e
c
onds
pe
r
im
a
g
e
.
M
obi
le
N
e
tv
1 s
how
e
d t
he
be
s
t
pe
r
f
or
m
a
nc
e
, a
c
hi
e
vi
ng a
n A
P
of
89.3%
f
or
a
n
I
oU
t
h
r
e
s
hol
d of
0.5,
t
hough
it
s
pe
r
f
or
m
a
nc
e
de
c
r
e
a
s
e
s
w
it
h
hi
ghe
r
I
oU
th
r
e
s
hol
ds
.
L
ooki
ng
a
he
a
d,
our
f
oc
us
w
il
l
be
on
opt
im
iz
in
g
m
ode
l
a
r
c
hi
te
c
tu
r
e
s
,
pa
r
ti
c
ul
a
r
ly
e
nha
nc
in
g
M
obi
le
N
e
tv
1
w
it
h
a
dva
nc
e
d
te
c
hni
que
s
to
a
c
hi
e
ve
e
ve
n
be
tt
e
r
r
e
s
ul
t
s
in
te
r
m
s
of
a
c
c
ur
a
c
y a
nd s
p
e
e
d.
R
E
F
E
R
E
N
C
E
S
[
1]
A
.
T
a
nnouc
he
,
K
.
S
ba
i
,
M
.
R
a
hm
oune
,
R
.
A
gounoune
,
a
nd
A
.
R
a
hm
a
ni
,
“
R
e
a
l
t
i
m
e
w
e
e
d
de
t
e
c
t
i
on
us
i
ng
a
boos
t
e
d
c
a
s
c
a
de
of
s
i
m
pl
e
f
e
a
t
ur
e
s
,”
I
nt
e
r
nat
i
onal
J
our
nal
of
E
l
e
c
t
r
i
c
al
and
C
om
put
e
r
E
ngi
ne
e
r
i
ng
,
vol
.
6,
no.
6,
pp.
2755
–
2765,
2016,
doi
:
10.11591/
i
j
e
c
e
.v6i
6.11878.
[
2]
A
.
T
a
nnouc
he
e
t
al
.
,
“
A
f
a
s
t
a
nd
e
f
f
i
c
i
e
nt
s
ha
pe
de
s
c
r
i
pt
or
f
or
a
n
a
dva
nc
e
d
w
e
e
d
t
ype
c
l
a
s
s
i
f
i
c
a
t
i
on
a
ppr
oa
c
h,
”
I
nt
e
r
nat
i
onal
J
our
na
l
of
E
l
e
c
t
r
i
c
al
and C
om
put
e
r
E
ngi
ne
e
r
i
ng
, vol
. 6, no. 3, pp. 1168
–
1175, 2016, do
i
:
10.11591/
i
j
e
c
e
.v6i
3.9978.
[
3]
H
.
M
oha
m
m
e
d,
A
.
T
a
nnouc
h
e
,
a
nd
Y
.
O
une
j
j
a
r
,
“
W
e
e
d
de
t
e
c
t
i
on
i
n
p
e
a
c
ul
t
i
va
t
i
on
w
i
t
h
t
he
f
a
s
t
e
r
R
C
N
N
R
e
s
N
e
t
50
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
k,”
R
e
v
ue
d’
I
nt
e
l
l
i
ge
nc
e
A
r
t
i
f
i
c
i
e
l
l
e
, vol
. 36, no. 1, pp. 13
–
18, 2022, doi
:
10.18280/
r
i
a
.360102.
[
4]
Y
.
O
une
j
j
a
r
,
A
.
T
a
nnouc
he
,
S
.
S
e
khr
a
,
a
nd
M
.
H
a
bi
b,
“
O
pt
i
m
i
s
a
t
i
on
of
w
e
e
d
m
a
na
ge
m
e
nt
by
i
m
a
ge
s
e
gm
e
nt
a
t
i
on
i
n
pr
e
c
i
s
i
o
n
a
gr
i
c
ul
t
ur
e
,”
I
nt
e
r
nat
i
onal
J
our
nal
of
C
om
put
at
i
onal
V
i
s
i
on and R
obot
i
c
s
, vol
. 1, no. 1, 2024, doi
:
10.1504/
i
j
c
vr
.2024.10062366.
[
5]
M
.
H
a
bi
b,
S
.
S
e
khr
a
,
A
.
T
a
nnouc
he
,
a
nd
Y
.
O
une
j
j
a
r
,
“
T
he
i
de
nt
i
f
i
c
a
t
i
on
o
f
w
e
e
ds
a
nd
c
r
ops
us
i
ng
t
he
popul
a
r
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
ks
,”
D
i
gi
t
al
T
e
c
hnol
ogi
e
s
and A
ppl
i
c
at
i
ons
, pp. 484
–
493, 2023, doi
:
10.1007/
978
-
3
-
031
-
29857
-
8_49.
[
6]
A
.
W
a
ng,
W
.
Z
ha
ng,
a
nd
X
.
W
e
i
,
“
A
r
e
vi
e
w
on
w
e
e
d
d
e
t
e
c
t
i
on
u
s
i
ng
gr
ound
-
ba
s
e
d
m
a
c
hi
ne
vi
s
i
on
a
nd
i
m
a
ge
pr
oc
e
s
s
i
n
g
t
e
c
hni
que
s
,”
C
o
m
put
e
r
s
and E
l
e
c
t
r
oni
c
s
i
n A
gr
i
c
ul
t
ur
e
, vol
. 158, pp. 226
–
240,
2019, doi
:
10.1016/
j
.c
om
pa
g.2019.02.005.
[
7]
W
. L
i
u, Y
.
W
e
n,
Z
. Y
u, M
.
L
i
,
B
. R
a
j
, a
nd
L
. S
ong,
“
S
phe
r
e
F
a
c
e
:
D
e
e
p hype
r
s
phe
r
e
e
m
be
ddi
ng f
or
f
a
c
e
r
e
c
ogni
t
i
on,”
2017
I
E
E
E
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
i
s
i
on
and
P
at
t
e
r
n
R
e
c
ogni
t
i
on
(
C
V
P
R
)
,
H
o
nol
ul
u,
H
I
,
U
S
A
,
2017,
pp.
6738
-
6746,
doi
:
10.1109/
C
V
P
R
.2017.713.
[
8]
J
.
L
i
,
X
.
L
i
a
ng,
S
.
S
he
n,
T
.
X
u,
J
.
F
e
ng,
a
nd
S
.
Y
a
n,
“
S
c
a
l
e
-
a
w
a
r
e
f
a
s
t
R
-
C
N
N
f
or
pe
de
s
t
r
i
a
n
de
t
e
c
t
i
on,”
I
E
E
E
T
r
ans
ac
t
i
ons
on
M
ul
t
i
m
e
di
a
, vol
. 20, no. 4, pp. 985
–
996, 2018, doi
:
10.1109/
T
M
M
.2017.2759508.
[
9]
A
.
K
a
r
pa
t
hy,
G
.
T
ode
r
i
c
i
,
S
.
S
he
t
t
y,
T
.
L
e
ung,
R
.
S
ukt
ha
nka
r
,
a
nd
F
.
F
.
L
i
,
“
L
a
r
ge
-
s
c
a
l
e
vi
de
o
c
l
a
s
s
i
f
i
c
a
t
i
on
w
i
t
h
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
ks
,”
P
r
oc
e
e
di
ngs
of
t
he
I
E
E
E
C
om
put
e
r
Soc
i
e
t
y
C
onf
e
r
e
nc
e
on C
om
put
e
r
V
i
s
i
on and P
at
t
e
r
n R
e
c
ogni
t
i
on
, pp.
1725
–
1732, 2014, doi
:
10.1109/
C
V
P
R
.2014.223.
[
10]
H
.
S
u,
X
.
Z
hu,
a
nd
S
.
G
ong,
“
D
e
e
p
l
e
a
r
ni
ng
l
ogo
d
e
t
e
c
t
i
on
w
i
t
h
da
t
a
e
xpa
n
s
i
on by
s
ynt
he
s
i
s
i
ng
c
ont
e
xt
,”
P
r
oc
e
e
di
ng
s
-
2017
I
E
E
E
W
i
nt
e
r
C
onf
e
r
e
nc
e
on A
ppl
i
c
at
i
ons
of
C
om
put
e
r
V
i
s
i
on, W
A
C
V
2017
, pp. 530
–
539, 2017, doi
:
10.1109/
W
A
C
V
.2017.65.
[
11]
M
.
L
i
u,
X
.
W
a
ng,
A
.
Z
hou,
X
.
F
u,
Y
.
M
a
,
a
nd
C
.
P
i
a
o,
“
U
a
v
-
Y
O
L
O
:
S
m
a
l
l
obj
e
c
t
de
t
e
c
t
i
on
on
unm
a
nne
d
a
e
r
i
a
l
ve
hi
c
l
e
pe
r
s
pe
c
t
i
ve
,”
Se
n
s
or
s
, vol
. 20, no. 8, 2020, doi
:
10.3390/
s
20082238.
[
12]
Z
.
W
a
ng,
L
.
W
u,
T
.
L
i
,
a
nd
P
.
S
hi
,
“
A
s
m
oke
de
t
e
c
t
i
on
m
ode
l
ba
s
e
d
on
i
m
pr
o
ve
d
Y
O
L
O
v5
,”
M
at
he
m
at
i
c
s
,
vol
.
10,
no.
7,
2022,
doi
:
10.3390/
m
a
t
h10071190.
[
13]
H
.
L
i
u,
F
.
S
un,
J
.
G
u,
a
nd L
.
D
e
ng, “
S
F
-
Y
O
L
O
v5
:
a
l
i
ght
w
e
i
ght
s
m
a
l
l
obj
e
c
t
de
t
e
c
t
i
on
a
l
gor
i
t
hm
ba
s
e
d
on
i
m
pr
ove
d
f
e
a
t
ur
e
f
us
i
o
n
m
ode
,”
Se
ns
or
s
, vol
. 22, no. 15, 2022, doi
:
10.3390/
s
22155817.
[
14]
H
.
W
a
ng,
S
.
S
ha
ng,
D
.
W
a
ng,
X
.
H
e
,
K
.
F
e
ng,
a
nd
H
.
Z
hu,
“
P
l
a
nt
di
s
e
a
s
e
de
t
e
c
t
i
on
a
nd
c
l
a
s
s
i
f
i
c
a
t
i
on
m
e
t
hod
ba
s
e
d
on
t
he
opt
i
m
i
z
e
d
l
i
ght
w
e
i
ght
Y
O
L
O
v5
m
ode
l
,”
A
gr
i
c
ul
t
ur
e
, vol
. 12, no. 7, 2022, doi
:
10.3390/
a
gr
i
c
ul
t
ur
e
12070931.
[
15]
X
.
Z
ha
ng,
Y
.
G
a
o,
H
.
W
a
ng,
a
nd
Q
.
W
a
ng,
“
I
m
pr
ove
Y
O
L
O
v3
us
i
ng
di
l
a
t
e
d
s
pa
t
i
a
l
pyr
a
m
i
d
m
odul
e
f
or
m
ul
t
i
-
s
c
a
l
e
obj
e
c
t
de
t
e
c
t
i
on,”
I
nt
e
r
nat
i
onal
J
our
nal
of
A
dv
anc
e
d R
obot
i
c
Sy
s
t
e
m
s
, vol
. 17, no. 4,
2020, doi
:
10.1177/
1729881420936062.
[
16]
J
.
Y
a
o,
J
.
Q
i
,
J
.
Z
ha
ng,
H
.
S
ha
o,
J
.
Y
a
ng,
a
nd
X
.
L
i
,
“
A
r
e
a
l
-
t
i
m
e
de
t
e
c
t
i
on
a
l
gor
i
t
hm
f
or
ki
w
i
f
r
ui
t
de
f
e
c
t
s
ba
s
e
d
on
Y
O
L
O
v5
,
”
E
l
e
c
t
r
oni
c
s
, vol
.
10, no. 14, 2021, doi
:
10.3390/
e
l
e
c
t
r
oni
c
s
10141711.
[
17]
S
.
R
e
n,
K
.
H
e
,
R
.
G
i
r
s
hi
c
k,
a
nd
J
.
S
un,
“
F
a
s
t
e
r
R
-
C
N
N
:
t
ow
a
r
ds
r
e
a
l
-
t
i
m
e
obj
e
c
t
de
t
e
c
t
i
on
w
i
t
h
r
e
gi
on
pr
opos
a
l
ne
t
w
or
ks
,”
I
E
E
E
T
r
ans
ac
t
i
ons
on
P
at
t
e
r
n
A
nal
y
s
i
s
and
M
ac
hi
ne
I
nt
e
l
l
i
ge
nc
e
,
vol
.
39,
no.
6,
pp.
1137
–
1149,
2017,
do
i
:
10.1109/
T
P
A
M
I
.2016.2577031.
[
18]
T.
-
Y
.
L
i
n,
P
.
D
ol
l
á
r
,
R
.
G
i
r
s
hi
c
k,
K
.
H
e
,
B
.
H
a
r
i
ha
r
a
n,
a
nd
S
.
B
e
l
ongi
e
,
“
F
e
a
t
ur
e
pyr
a
m
i
d
ne
t
w
or
ks
f
or
obj
e
c
t
de
t
e
c
t
i
on,”
P
r
oc
e
e
di
ngs
-
2019
I
E
E
E
I
nt
l
C
onf
on
P
ar
al
l
e
l
and
D
i
s
t
r
i
but
e
d
P
r
oc
e
s
s
i
ng
w
i
t
h
A
ppl
i
c
at
i
ons
,
B
i
g
D
at
a
and
C
l
oud
C
om
put
i
ng,
Sus
t
ai
nabl
e
C
om
put
i
ng
and
C
om
m
uni
c
at
i
ons
,
Soc
i
al
C
om
put
i
ng
and
N
e
t
w
o
r
k
i
ng,
I
SP
A
/
B
D
C
l
oud/
Sus
t
ai
nC
om
/
Soc
i
al
C
om
2019
,
pp. 1500
–
1504, D
e
c
. 2016, doi
:
10.1109/
I
S
P
A
-
B
D
C
l
oud
-
S
us
t
a
i
nC
om
-
S
oc
i
a
l
C
om
48970.2019.00217.
[
19]
J
. R
e
dm
on a
nd A
. F
a
r
ha
di
, “
Y
O
L
O
v3:
a
n i
nc
r
e
m
e
nt
a
l
i
m
pr
ove
m
e
nt
,”
ar
X
i
v
-
C
o
m
put
e
r
Sc
i
e
nc
e
,
pp. 1
-
6,
2018.
[
20]
P
.
A
da
r
s
h,
P
.
R
a
t
hi
,
a
nd
M
.
K
um
a
r
,
“
Y
O
L
O
v3
-
T
i
ny:
ob
j
e
c
t
de
t
e
c
t
i
on
a
nd
r
e
c
ogni
t
i
on
us
i
ng
one
s
t
a
ge
i
m
pr
ove
d
m
ode
l
,”
i
n
2020
6t
h
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on
A
dv
anc
e
d
C
om
put
i
ng
and
C
o
m
m
uni
c
at
i
on
Sy
s
t
e
m
s
,
I
C
A
C
C
S
2020
,
2020,
pp.
687
–
694,
doi
:
10.1109/
I
C
A
C
C
S
48705.2020.9074315.
[
21]
C
.
Y
.
W
a
ng,
H
.
Y
.
M
a
r
k
L
i
a
o,
Y
.
H
.
W
u,
P
.
Y
.
C
h
e
n,
J
.
W
.
H
s
i
e
h,
a
nd
I
.
H
.
Y
e
h,
“
C
S
P
N
e
t
:
A
ne
w
ba
c
kbone
t
h
a
t
c
a
n
e
nha
nc
e
l
e
a
r
ni
ng
c
a
pa
bi
l
i
t
y
of
C
N
N
,”
I
E
E
E
C
om
put
e
r
Soc
i
e
t
y
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
i
s
i
on
and
P
at
t
e
r
n
R
e
c
ogni
t
i
on
W
o
r
k
s
hop
s
,
pp
.
1571
–
1580, 2020, doi
:
10.1109/
C
V
P
R
W
50498.2020.00203.
[
22]
K
.
H
e
,
X
.
Z
ha
ng,
S
.
R
e
n,
a
nd
J
.
S
un,
“
S
pa
t
i
a
l
pyr
a
m
i
d
pool
i
ng
i
n
de
e
p
c
onvol
ut
i
ona
l
ne
t
w
or
ks
f
or
vi
s
ua
l
r
e
c
ogni
t
i
on,”
C
om
put
e
r
V
i
s
i
on
–
E
C
C
V
2014
, pp. 346
–
361, 2014, doi
:
10.1007/
978
-
3
-
319
-
10578
-
9_23.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
A
r
ti
f
I
nt
e
ll
,
V
ol
.
14
, N
o.
1
,
F
e
br
ua
r
y
2025
:
658
-
666
666
[
23]
Z
.
J
i
a
ng,
L
.
Z
ha
o,
S
.
L
i
,
a
nd
Y
.
J
i
a
,
“
R
e
a
l
-
t
i
m
e
obj
e
c
t
d
e
t
e
c
t
i
on
m
e
t
hod
ba
s
e
d
on
i
m
pr
ove
d
Y
O
L
O
v4
-
t
i
ny,”
ar
X
i
v
-
C
om
put
e
r
Sc
i
e
nc
e
,
pp. 1
-
11.
[
24]
A
.
G
.
H
ow
a
r
d
e
t
al
.
,
“
M
obi
l
e
N
e
t
s
:
e
f
f
i
c
i
e
nt
c
onvol
ut
i
ona
l
ne
ur
a
l
ne
t
w
or
ks
f
or
m
obi
l
e
vi
s
i
on
a
ppl
i
c
a
t
i
ons
,”
ar
X
i
v
-
C
om
put
e
r
S
c
i
e
nc
e
,
pp. 1
-
9,
2017.
[
25]
M
.
S
a
ndl
e
r
,
A
.
H
ow
a
r
d,
M
.
Z
hu,
A
.
Z
hm
ogi
nov,
a
nd
L
.
C
.
C
he
n,
“
M
obi
l
e
N
e
t
V
2:
i
nve
r
t
e
d
r
e
s
i
dua
l
s
a
nd
l
i
ne
a
r
bot
t
l
e
ne
c
ks
,
”
P
r
oc
e
e
di
ngs
of
t
h
e
I
E
E
E
C
om
put
e
r
Soc
i
e
t
y
C
onf
e
r
e
nc
e
on
C
om
put
e
r
V
i
s
i
on
a
nd
P
at
t
e
r
n
R
e
c
ogni
t
i
on
,
pp.
4510
–
4520,
2018,
doi
:
10.1109/
C
V
P
R
.2018.00474.
[
26]
K
.
H
e
,
X
.
Z
ha
ng,
S
.
R
e
n,
a
nd
J
.
S
un,
“
D
e
e
p
r
e
s
i
dua
l
l
e
a
r
ni
ng
f
or
i
m
a
ge
r
e
c
ogni
t
i
on,”
P
r
oc
e
e
di
ngs
of
t
he
I
E
E
E
C
om
put
e
r
Soc
i
e
t
y
C
onf
e
r
e
nc
e
on C
om
put
e
r
V
i
s
i
on and P
at
t
e
r
n R
e
c
ogni
t
i
on
, vol
. 2016
-
D
e
c
e
m
be
r
, pp. 770
–
778, 2016, doi
:
10.1109/
C
V
P
R
.2016.90.
[
27]
K
.
S
i
m
onya
n
a
nd
A
.
Z
i
s
s
e
r
m
a
n,
“
V
e
r
y
de
e
p
c
onvol
ut
i
ona
l
ne
t
w
or
ks
f
or
l
a
r
ge
-
s
c
a
l
e
i
m
a
ge
r
e
c
ogni
t
i
on,”
3
r
d
I
nt
e
r
nat
i
onal
C
onf
e
r
e
nc
e
on L
e
ar
ni
ng R
e
pr
e
s
e
nt
at
i
ons
, I
C
L
R
2015
-
C
onf
e
r
e
nc
e
T
r
ac
k
P
r
oc
e
e
di
ngs
, 2015.
[
28]
“
B
C
E
w
i
t
h
l
ogi
t
s
l
os
s
,”
P
y
T
or
c
h,
2023.
A
c
c
e
s
s
e
d:
M
a
r
.
04,
2023.
[
O
nl
i
ne
]
.
A
va
i
l
a
bl
e
:
ht
t
ps
:
/
/
pyt
or
c
h.or
g/
doc
s
/
m
a
s
t
e
r
/
ge
ne
r
a
t
e
d/
t
or
c
h.nn.B
C
E
W
i
t
hL
ogi
t
s
L
o
s
s
.ht
m
l
[
29]
“
P
r
e
c
i
s
i
on
-
r
e
c
a
l
l
,”
Sc
i
k
i
t
L
e
ar
n
,
2023.
A
c
c
e
s
s
e
d:
M
a
r
.
04,
2023.
[
O
nl
i
ne
]
.
A
va
i
l
a
bl
e
:
ht
t
ps
:
/
/
s
c
i
ki
t
-
l
e
a
r
n.or
g/
s
t
a
bl
e
/
a
ut
o_e
xa
m
pl
e
s
/
m
ode
l
_s
e
l
e
c
t
i
on/
pl
ot
_pr
e
c
i
s
i
on_r
e
c
a
l
l
.ht
m
l
B
I
O
G
R
A
P
H
I
E
S
O
F
A
U
T
H
O
R
S
Mohammed
Habib
is
currently
a
Ph
.
D
.
student
at
the
Spectrometry,
Materials
and
Archeomaterials
Laboratory
(LASMAR)
Moulay
Ismail
Universit
y,
Faculty
of
Scienc
es
,
Meknes,
Morocco
.
His
research
interests
are
focused
in
computer
vi
sion,
deep
learning,
and
robotics
and
their
application
in
agriculture.
He
can
be
contacted
at
email:
moh.habib@edu.umi.ac.ma
.
Salma
Sekhra
is
currently
a
Ph
.
D
.
student
at
the
Spectrometry,
Materials
and
Archeomaterials
Laboratory
(LASMAR)
,
Moulay
Ismail
Universit
y,
Faculty
of
Scienc
es
,
Meknes,
Morocco
.
Her
research
interests
are
focused
in
computer
vi
sio
n,
deep
learning
and
their
applicati
on in ag
ri
-
food field.
She
can be cont
acted at em
ail:
sekhrasalma3@gmail.com
.
Pr.
Adil
Tannouche
is
a
professor
at
the
Higher
School
of
Technolog
y,
Béni
Mellal,
Sultan
Moulay
Slimane
University,
Morocco.
He
received
his
P
h
.
D
.
in
electronics
and
embedded
s
ystems
from
Moulay
Ismail
University,
Mèknes
,
Mor
occo.
He
is
member
of
Laboratory
of
Engineering
and
Applied
Technologies
and
his
researc
h
area
includes
machine
vision,
artificial
intelligence,
and
applications
in
the
field
of
pr
ecision
agriculture
and
agroindus
try.
He can be contacted at email:
tannouche
@
gmail.com
.
Pr.
Youssef
Ounejjar
was
born
in
Meknes,
Morocco,
in
1971.
H
e
received
the
B.Ing.
and
M.S.
degrees
in
electrical
engineering
from
the
Ecole
Nation
ale
d'
Ingénieurs
de
Sfax,
Sfax,
Tunisia,
in
1996
and
1998,
respectively,
and
the
Ph.D.
degree
in
electrical
engineering
from
the
École
de
Technologie
Supérieure,
Montréal,
QC,
Canada,
in
2
011.
His
current
research
interests
are
the
multilevel
power
converte
rs.
He
is
Associate
Profess
or
at
the
Moulay
Ismail
University, Meknes
, Morocco. He can be contacted at em
ail: ounejjar@
gmail.com
.
Evaluation Warning : The document was created with Spire.PDF for Python.