I
AE
S
I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
Ar
t
if
icial
I
n
t
e
ll
ig
e
n
c
e
(
I
J
-
AI
)
Vol.
1
4
,
No.
1
,
F
e
br
ua
r
y
202
5
,
pp.
667
~
679
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
1
4
.i
1
.
pp
66
7
-
679
667
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
A
n
ove
l
m
o
d
e
l
to
d
e
t
e
c
t
a
n
d
c
at
e
gor
iz
e
o
b
j
e
c
t
s
f
r
om
i
m
age
s
by
u
si
n
g
a
h
y
b
r
id
m
ac
h
in
e
le
ar
n
i
n
g
m
od
e
l
Nilam
b
ar
S
e
t
h
i
1
,
Ve
t
u
k
u
r
i
Ve
n
k
a
t
a
S
iva
Ram
a
Raj
u
2
,
Ve
n
k
a
t
a
S
r
in
ivas
L
ok
avar
ap
u
3
,
Ravi
B
ab
u
De
var
e
d
d
i
4
,
S
h
iva
S
h
an
k
ar
Re
d
d
y
3
,
S
il
p
a
Nr
u
s
im
h
ad
r
i
5
1
D
e
pa
r
tm
e
nt
of
C
omput
e
r
S
c
ie
nc
e
a
nd
E
ngi
ne
e
r
in
g,
G
a
ndhi
I
ns
ti
tu
te
of
E
ngi
ne
e
r
in
g a
nd T
e
c
hnol
ogy
U
ni
ve
r
s
it
y,
G
unupur
,
I
nd
ia
2
K
one
r
u
L
a
ks
hma
ia
h
E
duc
a
ti
on
F
ounda
ti
on,
V
a
dde
s
w
a
r
a
m
,
I
n
di
a
3
S
a
gi
R
a
ma
K
r
is
hna
m
R
a
ju
E
ngi
ne
e
r
in
g C
ol
le
ge
,
B
hi
ma
va
r
a
m,
I
ndi
a
4
G
I
T
A
M
S
c
hool
of
T
e
c
hnol
ogy,
G
I
T
A
M
U
ni
ve
r
s
it
y,
H
yde
r
a
ba
d,
I
ndi
a
5
S
hr
i
V
is
hnu
E
ngi
ne
e
r
in
g
C
ol
le
ge
f
or
W
ome
n
(
A
)
,
B
hi
ma
va
r
a
m,
I
ndi
a
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
M
a
r
3
,
2024
R
e
vis
e
d
Oc
t
19
,
2024
Ac
c
e
pted
Oc
t
23
,
2024
As
h
u
ma
n
s
,
we
can
eas
i
l
y
reco
g
n
i
ze
an
d
d
i
s
t
i
n
g
u
i
s
h
d
i
ffere
n
t
feat
u
res
of
o
b
j
ect
s
in
i
ma
g
es
due
to
our
b
rai
n
’
s
ab
i
l
i
t
y
to
u
n
c
o
n
s
ci
o
u
s
l
y
l
ear
n
fro
m
a
s
et
of
i
mag
e
s
.
The
o
b
j
ec
t
i
v
es
of
t
h
i
s
effo
r
t
are
to
d
e
v
el
o
p
a
mo
d
e
l
t
h
a
t
is
cap
a
b
l
e
of
i
d
e
n
t
i
fy
i
n
g
an
d
cat
e
g
o
r
i
zi
n
g
o
b
j
ect
s
t
h
a
t
are
p
re
s
en
t
w
i
t
h
i
n
i
ma
g
es
.
We
i
mp
o
rt
e
d
the
d
at
a
s
et
fr
o
m
K
eras
a
n
d
l
o
a
d
ed
it
u
s
i
n
g
d
a
t
a
l
o
a
d
ers
to
ac
h
i
e
v
e
t
h
i
s
.
We
t
h
en
u
t
i
l
i
z
ed
v
ar
i
o
u
s
d
eep
l
ear
n
i
n
g
a
l
g
o
ri
t
h
ms
,
s
u
c
h
as
v
i
s
u
al
g
eo
me
t
ry
g
ro
u
p
(
V
G
G
)
-
16
an
d
a
s
i
mp
l
e
n
e
t
-
ran
d
o
m
fo
r
es
t
h
y
b
ri
d
mo
d
e
l
,
to
cl
as
s
i
f
y
the
o
b
j
e
ct
s
.
A
ft
er
cl
a
s
s
i
fi
ca
t
i
o
n
,
t
h
e
accu
rac
y
o
b
t
ai
n
e
d
by
V
G
G
1
6
a
n
d
the
h
y
b
ri
d
mo
d
el
w
as
8
4
.
7
%
a
n
d
8
9
.
6
%
,
res
p
ec
t
i
v
el
y
.
T
h
erefo
re,
the
p
r
o
p
o
s
e
d
mo
d
e
l
s
u
cces
s
fu
l
l
y
d
e
t
ect
s
o
b
j
ec
t
s
in
i
mag
e
s
u
s
i
n
g
a
s
i
mp
l
e
n
et
as
a
feat
u
re
ex
t
rac
t
o
r
a
n
d
a
ran
d
o
m
f
o
res
t
fo
r
o
b
j
ect
c
l
as
s
i
f
i
cat
i
o
n
,
ach
i
ev
i
n
g
b
e
t
t
er
accu
racy
t
h
an
V
G
G
1
6
.
K
e
y
w
o
r
d
s
:
Ar
ti
f
icia
l
int
e
ll
igenc
e
C
omput
e
r
vis
ion
De
e
p
lea
r
ning
M
a
c
hine
lea
r
ning
Obje
c
t
de
tec
ti
on
Vis
ua
l
ge
ometr
y
gr
oup
Th
i
s
is
an
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
the
CC
BY
-
SA
l
i
ce
n
s
e.
C
or
r
e
s
pon
din
g
A
u
th
or
:
S
hiva
S
ha
nka
r
R
e
ddy
As
s
is
tant
P
r
of
e
s
s
or
,
S
a
gi
R
a
maK
r
is
hna
m
R
a
ju
E
n
ginee
r
ing
C
oll
e
ge
B
him
a
va
r
a
m,
Andhr
a
P
r
a
de
s
h,
I
ndia
E
mail:
s
hiva
.
c
s
e
s
r
kr
@gmail.
c
om
1.
I
NT
RODU
C
T
I
ON
C
omput
e
r
s
ha
ve
a
unique
wa
y
of
int
e
r
pr
e
ti
ng
v
is
ua
ls
.
W
hil
e
human
br
a
ins
unde
r
s
tand
im
a
ge
s
,
c
omput
e
r
s
br
e
a
k
them
int
o
numer
ica
l
va
lues
a
nd
u
s
e
a
lgor
it
hms
to
identif
y
s
igni
f
ica
nt
e
leme
nts
in
s
ti
l
l
photos
,
videos
,
gr
a
phics
,
a
nd
li
ve
f
e
e
ds
.
T
his
pr
oc
e
s
s
is
known
as
c
omput
e
r
vis
ion
.
I
mage
p
r
oc
e
s
s
ing
a
lgor
it
hms
c
omput
e
r
s
can
int
e
r
pr
e
t
a
nd
c
ompr
e
he
nd
vis
ua
ls
f
r
om
a
s
ingl
e
or
s
e
r
ies
of
pictur
e
s
.
F
igur
e
1
s
hows
th
e
va
r
ious
objec
t
de
tec
ti
on
in
a
n
im
a
ge
.
By
a
na
lyzing
a
nd
f
i
lt
e
r
ing
mi
ll
ions
of
us
e
r
-
uploade
d
photos
,
mac
hine
vi
s
ion
can
a
c
c
ur
a
tely
de
tec
t
pe
ople
a
nd
a
utom
obil
e
s
on
the
r
o
a
d
[
1]
.
Nume
r
ous
c
ompanie
s
a
r
e
inves
ti
ng
i
n
im
a
ge
r
e
c
ognit
ion
tec
hnology
to
a
na
lyze
vis
ua
l
da
ta
as
c
omput
e
r
vis
ion
a
dva
nc
e
s
.
T
his
tec
hnique
is
uti
l
ize
d
in
va
r
ious
f
ields
,
s
uc
h
as
medic
a
l
p
ictur
e
a
na
lys
is
,
a
utonom
ous
objec
t
identif
ica
ti
on
in
a
utom
obil
e
s
,
a
nd
f
a
c
ial
r
e
c
ognit
ion
in
s
e
c
ur
it
y
s
ys
tems
[
2]
.
Us
ing
mac
hine
vis
ion,
a
r
ti
f
icia
l
int
e
ll
igenc
e
(
AI
)
,
a
nd
tr
a
ined
a
lg
or
it
hms
,
a
pr
og
r
a
m
or
s
ys
tem
can
identif
y
objec
ts
,
pe
ople
,
loca
ti
ons
,
a
nd
a
c
ti
vit
ies
in
photogr
a
phs
[
3]
.
I
m
a
ge
r
e
c
ognit
ion
ha
s
be
c
ome
incr
e
a
s
ingl
y
popular
due
to
a
dva
nc
e
ments
in
mac
hine
lea
r
ning
a
nd
pr
oc
e
s
s
in
g
c
a
pa
c
it
y
[
4]
.
T
he
a
im
is
to
loca
te
a
nd
r
e
c
ognize
objec
ts
withi
n
an
im
a
ge
.
C
las
s
if
ica
ti
on
invol
ve
s
de
ter
mi
ning
the
objec
t
type
withi
n
the
bounding
box
[
5
]
.
Obje
c
t
identif
ica
ti
on
can
be
c
ha
ll
e
nging
due
to
oc
c
lus
ion,
c
lut
ter
,
it
e
m
a
ppe
a
r
a
nc
e
,
s
ize
,
a
nd
or
ienta
ti
on
dif
f
e
r
e
nc
e
s
.
O
bjec
t
r
e
c
ognit
ion
methods
us
e
im
a
ge
p
r
oc
e
s
s
ing,
mac
hine
lea
r
ning,
a
nd
c
omput
e
r
vis
ion
tec
hniques
[
6]
,
[
7]
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
1
4
,
No.
1
,
F
e
br
ua
r
y
202
5
:
66
7
-
679
668
Obje
c
t
de
tec
ti
on
ha
s
numer
ous
indus
tr
ial
a
nd
f
ield
a
ppli
c
a
ti
ons
.
M
a
ny
s
ur
ve
il
lanc
e
s
ys
tems
us
e
objec
t
de
tec
ti
on
to
moni
tor
the
r
e
a
l
-
time
moveme
nt
of
pe
ople,
ve
hicle
s
,
a
nd
other
objec
ts
.
It
is
a
ls
o
c
r
uc
ial
to
a
utonom
ous
dr
ivi
ng
s
ys
tems
,
whic
h
mus
t
r
e
c
ognize
a
nd
tr
a
c
k
other
c
a
r
s
,
pe
de
s
tr
ians
,
a
nd
obs
tac
les
to
e
ns
ur
e
s
a
f
e
na
vigation
[
8]
.
F
ig
ur
e
1
.
Obje
c
t
de
tec
ti
on
in
an
i
mage
Obje
c
t
de
tec
ti
on
is
a
ls
o
us
e
d
in
medic
a
l
im
a
ging
to
loca
te
tum
our
s
,
les
ions
,
a
nd
other
pa
thol
ogica
l
s
igns
,
a
idi
ng
diagnos
is
a
nd
ther
a
py
planning
.
In
a
ddit
ion
to
thes
e
a
ppli
c
a
ti
ons
,
ob
jec
t
de
tec
ti
on
is
ut
il
ize
d
in
a
gr
icultur
e
to
identi
f
y
c
r
ops
,
moni
tor
plant
he
a
lt
h
,
a
nd
tr
a
c
k
c
r
op
gr
owth
.
It
is
a
ls
o
us
e
d
in
s
por
ts
a
na
lyt
ics
to
moni
tor
playe
r
a
nd
ba
ll
moveme
nt
thr
oughout
ga
mes
,
a
s
s
is
ti
ng
c
oa
c
h
e
s
a
nd
a
na
ly
s
ts
in
c
r
e
a
ti
ng
be
tt
e
r
tea
m
s
tr
a
tegie
s
.
Ar
t
a
nd
c
ult
ur
a
l
he
r
it
a
ge
a
ppli
c
a
ti
ons
e
m
ploy
objec
t
de
tec
ti
on
to
identif
y
a
nd
e
va
luate
objec
t
s
withi
n
pa
int
ings
,
s
c
ulpt
ur
e
s
,
a
nd
o
ther
a
r
tef
a
c
ts
[
9]
.
T
hi
s
tec
hnology
e
na
bles
r
e
s
e
a
r
c
he
r
s
a
nd
his
tor
ians
be
tt
e
r
to
unde
r
s
tand
thes
e
objec
ts
’
his
tor
ica
l
s
igni
f
ica
nc
e
a
nd
c
ult
ur
a
l
i
mpor
tanc
e
.
2.
RE
L
AT
E
D
WORK
S
ult
a
na
et
al
.
[
1]
s
tudi
e
d
objec
t
identi
f
ica
ti
on
mode
ls
li
ke
r
e
gion
-
ba
s
e
d
c
onvolut
ional
ne
ur
a
l
ne
twor
ks
(
R
C
NN
)
a
nd
R
e
f
ineD
e
t,
whic
h
can
be
c
las
s
if
ied
i
nto
one
-
s
tage
a
nd
two
-
s
tage
methods
.
S
e
ve
r
a
l
ne
w
models
ha
ve
given
r
e
s
e
a
r
c
he
r
s
hope
f
or
quicke
r
a
nd
mor
e
a
c
c
ur
a
te
r
e
a
l
-
time
it
e
m
de
tec
ti
on.
Ha
que
et
al
.
[
2
]
pr
opos
e
d
a
vis
ua
l
ge
ometr
y
gr
oup
(
VG
G)
-
r
e
s
idual
ne
twor
k
(
R
e
s
Ne
t
)
ne
twor
k
objec
t
identif
ica
ti
on
model
that
r
e
duc
e
d
the
VGG
ne
twor
k
tr
a
ini
ng
e
r
r
or
a
nd
i
mpr
ove
d
ti
ny
it
e
m
de
tec
ti
on.
T
he
model
c
ombi
ne
d
the
VGG
a
n
d
R
e
s
Ne
t
ne
twor
k
to
de
c
r
e
a
s
e
laye
r
s
ize
a
nd
tr
a
ini
ng
e
r
r
o
r
.
T
he
upgr
a
de
d
c
onvolut
ional
a
r
c
hit
e
c
tur
e
of
th
e
model
pe
r
f
or
med
we
ll
,
with
an
objec
t
de
tec
ti
on
a
c
c
ur
a
c
y
of
85.
8%
.
R
a
dovic
et
al
.
[
3
]
s
ugge
s
ted
a
c
onvolut
ional
ne
ur
a
l
ne
twor
ks
(
C
NN
)
-
ba
s
e
d
objec
t
identif
ica
ti
on
model
f
or
a
e
r
ial
im
a
ge
s
.
T
he
you
only
look
onc
e
(
YO
L
O
)
platf
or
m
us
e
s
a
C
NN
to
f
oll
ow,
r
e
c
ognize
,
a
nd
c
las
s
if
y
u
n
manne
d
a
e
r
ial
v
e
hicle
s
(
UAV
)
video
s
tr
e
a
ms
in
r
e
a
l
ti
me.
De
s
pit
e
ha
ving
84%
a
c
c
ur
a
c
y
in
de
tec
ti
on
a
nd
c
las
s
if
ica
ti
on,
the
YO
L
O
platf
or
m
is
mor
e
e
f
f
icie
nt
tha
n
typi
c
a
l
mac
hine
lea
r
ning
methods
.
It
is
pe
r
f
e
c
t
f
o
r
UAV
a
utonom
ous
f
lyi
ng.
Ga
lvez
et
al
.
[
7]
s
ugge
s
ted
a
C
NN
model
f
or
objec
t
de
tec
ti
on
.
T
his
model
us
e
d
C
NN
to
r
e
c
ognize
e
nvir
onmenta
l
it
e
ms
.
S
ingl
e
s
hot
de
tec
tor
(
S
S
D
)
with
M
obil
e
Ne
t
of
f
e
r
s
g
r
e
a
t
de
tec
ti
on
s
pe
e
d
but
poor
a
c
c
ur
a
c
y,
whe
r
e
a
s
f
a
s
ter
-
R
C
NN
with
I
nc
e
pti
onV2
ha
s
low
s
pe
e
d
but
higher
a
c
c
ur
a
c
y.
SSD
with
M
obil
e
Ne
t
is
idea
l
f
or
de
tec
ti
ng
objec
ts
quickly,
pa
r
ti
c
ular
ly
in
r
e
a
l
-
time
a
ppli
c
a
ti
ons
.
F
or
h
igh
-
a
c
c
ur
a
c
y
de
tec
ti
on,
one
s
hould
us
e
f
a
s
ter
-
R
C
NN
with
I
nc
e
pti
onV2.
Ka
ng
et
al
.
[
8]
pr
opos
e
d
us
ing
tubele
ts
with
C
N
N
f
or
v
ideo
objec
t
de
tec
ti
on
(
VO
D)
.
T
his
model
int
r
oduc
e
d
a
de
e
p
lea
r
ning
(
DL
)
f
r
a
mew
or
k
f
or
video
objec
t
r
e
c
ognit
ion
us
ing
tempor
a
l
inf
or
mat
ion.
T
his
f
r
a
mew
or
k
won
the
V
O
D
c
ha
ll
e
nge
with
given
d
a
ta
a
nd
pe
r
f
or
med
we
ll
on
I
mage
Ne
t
objec
t
identif
ica
ti
on
f
r
om
video.
Yu
et
al
.
[
9]
p
r
e
s
e
nted
an
a
dva
nc
e
d
o
bjec
t
de
tec
ti
on
ne
twor
k
unit
box.
T
his
s
tudy
e
xa
mi
ne
d
the
unique
bounding
box
p
r
e
diction
los
s
,
int
e
r
s
e
c
ti
on
ove
r
union
(
I
oU
)
.
I
ns
tea
d
of
f
our
indepe
nde
nt
va
r
i
a
bles
,
the
I
oU
los
s
laye
r
r
e
gr
e
s
s
e
s
an
objec
t
c
a
ndidate
’
s
bounding
box,
s
pe
e
ding
c
omput
a
ti
on
a
nd
im
pr
ov
in
g
objec
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
nov
e
l
mode
l
to
de
tec
t
and
c
ategor
iz
e
objec
ts
fr
o
m
image
s
by
us
ing
a
hy
br
id
mac
hine
…
(
N
il
ambar
Se
thi
)
669
loca
li
z
a
ti
on.
T
he
y
a
ls
o
pr
e
s
e
nted
the
UnitB
ox,
a
s
ophis
ti
c
a
ted
objec
t
identif
ica
ti
on
ne
twor
k
f
or
f
a
c
e
r
e
c
ognit
ion
that
pe
r
f
or
ms
wor
ld
-
c
las
s
ba
s
e
d
on
I
oU
los
s
.
In
th
e
ir
r
e
s
e
a
r
c
h
on
im
a
ge
c
las
s
if
ica
ti
on,
T
a
mm
ina
[
10
]
uti
li
z
e
d
tr
a
ns
f
e
r
lea
r
ning
a
long
with
VGG1
6
a
nd
de
e
p
C
NN
.
T
he
pr
e
-
tr
a
ined
model
inf
o
r
mation
wa
s
us
e
d
in
thi
s
s
tudy
to
a
c
hieve
Ga
o
et
al
.
[
11]
int
r
oduc
e
d
de
e
p
s
pa
ti
a
l
f
e
a
tur
e
tr
a
ns
f
or
mation
(
DSF
T
)
f
or
or
iente
d
a
e
r
i
a
l
objec
t
de
tec
ti
on
.
T
he
y
us
e
d
D
S
F
T
ne
twor
ks
to
identif
y
o
bjec
ts
in
a
e
r
ial
photos
.
T
he
y
f
ound
that
the
a
e
r
ial
de
tec
tor
r
e
quir
e
s
a
lot
of
r
otating
da
ta
s
ince
or
dinar
y
C
NN
s
c
a
nnot
r
e
pr
e
s
e
nt
di
r
e
c
ti
on
c
ha
nge
.
T
he
y
a
ls
o
p
r
e
s
e
nted
the
polar
iza
ti
on
f
unc
ti
on
to
c
ons
tr
uc
t
e
s
s
e
nti
a
l
f
e
a
tur
e
s
e
lec
ti
on
a
nd
de
tec
ti
on
f
e
a
tur
e
s
.
M
ohs
in
et
al
.
[
12
]
s
u
r
ve
y
e
d
de
e
p
ne
ur
a
l
ne
two
r
ks
a
nd
a
lgo
r
i
thm
s
f
or
de
f
e
c
t
de
tec
t
ion
.
T
he
y
wa
nt
e
d
to
unde
r
s
tand
th
e
f
unc
ti
o
nin
g
of
t
he
late
s
t
i
ndus
t
r
ia
l
T
e
c
hn
olog
ies
f
o
r
de
tec
ti
n
g
f
a
u
lt
s
in
qu
a
li
ty
c
on
tr
o
l
a
nd
h
ow
they
can
be
a
ppl
ied
in
mo
r
e
c
o
mpl
e
x
wa
ys
.
T
he
y
a
ls
o
dis
c
us
s
e
d
im
p
r
ov
ing
inf
e
r
e
nc
e
pe
r
f
o
r
manc
e
,
mo
de
l
c
or
r
e
c
tnes
s
f
o
r
ti
ny
e
r
r
or
s
,
ne
two
r
k
de
s
ign
f
o
r
de
p
loym
e
nt
on
lo
w
-
r
e
s
ou
r
c
e
de
vice
s
,
a
nd
de
e
p
mode
l
tr
a
ini
n
g
with
les
s
pr
oc
e
s
s
ing
powe
r
.
S
ha
n
ka
r
et
al
.
[
13
]
wor
ke
d
on
im
a
ge
s
to
r
e
t
r
ieve
t
he
da
ta
by
us
i
ng
a
ge
ne
ti
c
a
lgo
r
it
h
m
.
T
he
i
r
m
ode
l
r
e
d
uc
e
s
the
ti
me
in
ter
va
l
a
n
d
inc
r
e
a
s
e
s
pr
e
c
is
io
n
us
in
g
th
e
c
on
tent
-
ba
s
e
d
im
a
ge
r
e
t
r
iev
a
l
(
C
B
I
R
)
mo
de
l
.
Al
ha
r
thi
et
al
.
[
1
4]
pr
e
s
e
n
ted
3
-
di
mens
iona
l
c
on
vol
uti
ona
l
ne
two
r
ks
(
C
3D
)
-
ba
s
e
d
on
mas
s
ive
c
r
owd
a
b
nor
mal
be
ha
vio
r
r
e
c
ogn
it
i
on
.
T
he
y
e
va
lua
ted
wa
ys
to
im
pr
ove
a
be
r
r
a
n
t
be
ha
v
iou
r
de
tec
ti
o
n
u
s
ing
the
Ha
jj
c
r
owd
da
tas
e
t
.
T
he
y
e
x
plo
r
e
d
DL
a
lg
or
it
h
ms
to
c
a
teg
or
i
z
e
Ha
jj
da
tas
e
t
videos
as
a
n
omal
ous
.
T
he
C
3D
model
a
c
h
ieve
d
e
xc
e
ll
e
n
t
F1
-
s
c
or
e
,
pr
e
c
is
ion
,
a
nd
a
c
c
ur
a
c
y
on
t
he
tes
t
s
e
t,
wi
th
s
c
or
e
s
of
0
.
9
0,
0
.
89
,
a
nd
0
.
8
8.
Ka
hler
et
al
.
[
15
]
p
r
op
os
e
d
an
AI
-
ba
s
e
d
a
pp
r
oa
c
h
to
de
tec
t
s
u
r
f
a
c
e
de
f
e
c
ts
in
a
v
iati
on
c
ompo
n
e
nts
f
or
r
e
c
ti
f
ic
a
ti
o
n.
T
he
y
p
r
e
s
e
n
ted
a
de
f
e
c
t
mon
it
o
r
in
g
s
ys
tem
f
o
r
lan
ding
g
e
a
r
c
o
mpone
nts
us
ing
AI
to
id
e
nti
f
y
the
e
ndpoi
nt
of
t
he
p
r
oc
e
s
s
.
S
on
g
et
al
.
[
16
]
us
e
d
an
o
b
jec
ti
ve
a
n
d
s
yn
the
ti
c
da
ta
-
t
r
a
ine
d
Y
OL
Ov4
objec
t
de
tec
ti
on
ne
two
r
k
to
de
tec
t
c
or
r
os
io
n,
mo
nit
or
gr
indi
ng
,
a
nd
of
f
e
r
f
e
e
d
ba
c
k
w
he
n
t
he
e
nd
poi
nt
is
a
c
hie
ve
d
.
To
m
a
nipu
late
the
vide
o
da
taba
s
e
to
de
tec
t
d
upl
ica
te
v
ideos
us
in
g
f
r
a
mes
a
nd
the
ir
i
de
nt
if
i
c
a
ti
on
[
1
7]
.
Ahme
d
et
al
.
[
18]
p
r
e
s
e
nted
an
int
e
r
ne
t
of
thi
ngs
(
I
oT
)
-
e
na
bled
DL
f
r
a
mew
or
k
f
or
mul
ti
ple
objec
t
r
e
c
ognit
ion
in
a
e
r
ial
r
e
mot
e
-
s
e
ns
ing
photos
.
T
he
s
t
udy
us
e
d
the
py
r
a
mi
d
s
c
e
ne
pa
r
s
ing
ne
twor
k
(
P
S
P
Ne
t
)
DL
-
ba
s
e
d
s
e
gmenta
ti
on
model
to
s
e
pa
r
a
te
objec
ts
us
ing
AI
a
nd
a
c
hieve
d
high
a
c
c
ur
a
c
y
r
a
tes
us
ing
VG
G16,
R
e
s
Ne
t
-
0,
a
nd
M
obil
e
Ne
t
models
.
L
or
a
ks
a
et
al
.
[
19]
pr
opos
e
d
an
SSD
-
VG
G16
-
ba
s
e
d
model
to
identi
fy
c
a
nc
e
r
-
c
a
u
s
ing
os
teo
s
a
r
c
oma
lung
nodules
.
T
he
s
tudy
a
im
e
d
to
buil
d
an
AI
s
ys
tem
that
can
r
e
c
ognize
lung
nodules
us
ing
a
s
tr
a
ight
f
or
wa
r
d
model
with
hig
h
a
c
c
ur
a
c
y
r
a
tes
.
Ma
et
al
.
[
20]
s
ugge
s
ted
an
e
nd
-
to
-
e
nd
s
c
a
le
-
a
wa
r
e
s
pli
t
-
mer
ge
-
e
nha
n
c
e
ment
ne
twor
k
f
or
r
e
mot
e
s
e
ns
ing
objec
t
r
e
c
ognit
ion
us
ing
va
r
ious
modul
e
s
th
at
im
pr
ove
objec
t
s
a
li
e
nc
y,
objec
t
s
e
gmenta
ti
on,
a
nd
f
e
a
tur
e
a
li
gnment.
W
a
ng
et
al
.
[
21]
pr
opos
e
d
a
hybr
id
f
e
a
tur
e
-
a
li
gne
d
ne
twor
k
(
HFANe
t)
to
r
e
c
ognize
s
a
li
e
nt
objec
ts
in
opti
c
a
l
r
e
mot
e
s
e
ns
ing
im
a
ge
s
.
T
he
s
tudy
us
e
d
an
a
djac
e
nt
f
e
a
tur
e
-
a
li
gne
d
modul
e
(
AFA
M
)
to
in
tegr
a
te
ne
a
r
by
f
e
a
tur
e
s
us
ing
unpa
r
a
m
e
ter
ize
d
a
li
gnment
a
nd
de
vis
e
d
a
ne
w
int
e
r
a
c
ti
ve
guidanc
e
los
s
(
I
GL
os
s
)
to
blend
e
dge
de
tec
ti
on
a
nd
s
a
li
e
nc
y.
To
r
e
c
ognize
objec
ts
li
ke
f
lowe
r
s
pe
c
ies
us
ing
DL
models
,
the
a
uthor
s
a
c
hieve
d
e
xc
e
ll
e
nt
r
e
s
u
lt
s
[
22]
.
L
oga
na
than
et
al
.
[
23]
s
ugge
s
ted
us
ing
upg
r
a
de
d
gl
oba
l
pooli
ng
(
GP
)
-
R
C
NN
to
identi
f
y
a
nd
r
e
c
ognize
mul
ti
ple
objec
ts
.
T
he
s
tudy
s
howe
d
that
GP
-
R
C
NN
outper
f
or
med
R
C
NN
by
21%
in
de
tec
ti
ng
ti
ny
a
nd
lar
ge
o
bjec
ts
.
S
ha
nka
r
et
al
.
[
24
]
c
las
s
if
ied
the
voice
us
ing
va
r
io
us
mac
hine
lea
r
ning
models
.
T
he
ir
wor
k
obtaine
d
good
r
e
s
ult
s
c
ompar
e
d
wi
th
the
e
xis
ti
ng
models
.
Ya
s
a
s
wini
et
al
.
[
25
]
identif
ied
the
model
to
a
v
oid
r
oa
d
a
c
c
idents
us
ing
DL
models
.
Abdu
a
nd
Noor
[
26]
pr
e
s
e
nted
a
VG
G16
-
ba
s
e
d
DL
model
f
or
domes
ti
c
tr
a
s
h
c
las
s
if
ica
ti
on
to
c
las
s
if
y
tr
a
s
h
it
e
ms
,
a
c
hieving
ove
r
96%
pr
e
c
is
ion
a
ppr
opr
iate
ly.
S
hya
m
et
al
.
[
27]
s
ugge
s
ted
an
uns
upe
r
vis
e
d
domain
a
da
ptation
tec
hnique
tha
t
leve
r
a
ge
s
a
s
im
ulator
-
c
r
e
a
ted
s
ynthetic
da
tas
e
t
to
e
ns
ur
e
pe
r
f
or
manc
e
c
ons
is
tenc
y
of
mul
ti
-
objec
t
-
tr
a
c
king
(
M
OT
)
a
lgor
it
hms
ove
r
va
r
ious
manua
ll
y
a
nnot
a
ted
r
e
a
l
s
it
ua
ti
ons
.
Li
et
al
.
[
28]
pr
opos
e
d
a
s
e
lf
-
knowle
dge
dis
ti
ll
a
ti
on
-
ba
s
e
d
f
e
w
-
s
hot
objec
t
identif
ica
ti
on
a
ppr
oa
c
h
to
r
e
duc
e
ove
r
f
it
t
ing
in
f
e
w
-
s
hot
objec
t
r
e
c
ognit
ion.
T
he
y
de
mons
tr
a
ted
their
method
’
s
a
ppli
c
a
bi
li
ty
a
nd
pr
a
c
ti
c
a
li
ty
us
ing
be
nc
hmar
k
da
tas
e
ts
.
P
r
a
va
ll
ika
a
nd
B
a
s
ka
r
[
29]
s
tudi
e
d
us
ing
VG
G16
a
nd
s
uppo
r
t
ve
c
tor
mac
hine
(
S
VM
)
a
lgo
r
it
hms
f
or
br
a
in
tum
our
magn
e
ti
c
r
e
s
ona
nc
e
im
a
ging
(
M
R
I
)
i
mage
c
las
s
if
ica
ti
on.
VG
G16
outper
f
or
med
S
VM
in
a
c
c
ur
a
c
y
a
nd
pr
e
c
is
ion,
im
p
r
oving
the
ove
r
a
ll
pe
r
f
o
r
manc
e
.
S
a
jj
a
d
a
nd
He
mav
a
thi
[
30
]
e
xplor
e
d
objec
t
-
de
tec
ti
ng
methods
a
nd
how
to
buil
d
models
f
o
r
a
c
c
ur
a
te
r
e
s
ult
s
us
ing
DL
tec
hniques
.
Ya
ng
et
al
.
[
31]
pr
opos
e
d
a
method
f
o
r
identi
f
yin
g
hidden
objec
ts
in
pa
s
s
ive
mi
ll
im
e
ter
-
wa
ve
im
a
g
e
s
.
T
his
method
is
a
one
-
s
tage
a
nc
hor
-
f
r
e
e
de
tec
t
or
that
is
f
a
s
t
a
nd
a
c
c
ur
a
te.
It
ut
il
ize
s
c
onvolut
ion
tec
hni
que
s
a
nd
tr
a
ns
f
or
mer
a
na
lys
is
.
A
f
ter
c
onduc
ti
ng
r
e
s
e
a
r
c
h,
it
wa
s
f
ound
that
thi
s
a
ppr
oa
c
h
is
s
upe
r
ior
to
the
mos
t
a
dva
nc
e
d
tec
hniques
in
s
pe
e
d
a
nd
a
c
c
ur
a
c
y.
3.
M
E
T
HO
DOL
OG
Y
M
a
c
hines
mus
t
be
a
ble
to
r
e
c
ognize
objec
ts
in
im
a
ge
s
to
unde
r
s
tand
a
nd
in
ter
pr
e
t
the
vis
ua
l
wo
r
ld,
jus
t
li
ke
humans
do.
C
omput
e
r
vis
ion
is
a
f
ield
th
a
t
r
e
li
e
s
on
mac
hines
r
e
c
ognizing
a
nd
identif
ying
thi
ngs
in
pictur
e
s
a
nd
videos
.
Obje
c
t
de
tec
ti
on
is
us
e
d
in
va
r
ious
dis
c
ipl
ines
,
including
r
oboti
c
s
,
to
he
l
p
r
obots
unde
r
s
tand
a
nd
int
e
r
p
r
e
t
vis
ua
l
inf
or
mation
.
S
igni
f
ica
ntl
y
im
pr
ove
pr
oduc
ti
vit
y
,
s
a
f
e
ty
,
a
nd
e
f
f
icie
nc
y
a
c
r
os
s
many
s
e
c
tor
s
.
To
f
ind
objec
ts
in
an
input
pictur
e
,
t
he
de
e
p
c
onvolut
ional
ne
ur
a
l
ne
twor
k
(
DC
NN
)
V
GG
16
can
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
1
4
,
No.
1
,
F
e
br
ua
r
y
202
5
:
66
7
-
679
670
be
us
e
d.
T
he
goa
l
is
to
de
tec
t
vis
ua
l
objec
ts
a
nd
c
r
e
a
te
bounding
boxe
s
a
c
c
ur
a
tely.
To
t
r
a
in
the
mode
l,
tagge
d
photos
a
nd
objec
t
a
nnotations
s
hould
be
us
e
d.
T
he
ult
im
a
te
objec
ti
ve
is
to
a
c
hieve
high
objec
t
d
e
tec
ti
on
a
c
c
ur
a
c
y
with
m
ini
mal
f
a
ls
e
pos
it
ives
a
nd
ne
ga
ti
ve
s
.
A
ppli
c
a
ti
ons
s
uc
h
as
a
utonom
ous
d
r
ivi
ng,
pictur
e
s
e
gmenta
ti
on,
a
nd
objec
t
identi
f
ica
ti
on
a
r
e
mainly
us
e
d
.
3.
1.
Ob
j
e
c
t
ives
T
he
main
goa
l
is
to
r
e
c
ognize
objec
ts
in
photos
a
nd
s
or
t
them
int
o
s
pe
c
if
ic
c
a
tegor
ies
.
T
he
p
r
im
a
r
y
objec
ti
ve
s
a
r
e
to
de
tec
t
objec
ts
f
r
o
m
a
gr
e
a
ter
dis
tanc
e
a
nd
in
the
s
hor
tes
t
pos
s
ibl
e
ti
me.
C
ompar
e
d
to
e
a
r
li
e
r
models
,
the
a
im
is
to
c
a
tegor
ize
objec
ts
int
o
mor
e
c
a
tegor
ies
.
T
he
objec
ti
ve
is
to
identi
f
y
ob
jec
ts
in
lo
w
-
qua
li
ty
im
a
ge
s
without
c
onve
r
ti
ng
them
int
o
high
-
qua
li
ty
photos
.
3.
2.
Dat
as
e
t
T
he
C
a
na
dian
I
ns
ti
tut
e
F
o
r
Adva
nc
e
d
R
e
s
e
a
r
c
h
(
C
I
F
AR
-
10)
pr
ovides
im
a
ge
s
c
omm
only
us
e
d
to
t
r
a
in
mac
hine
lea
r
ning
a
nd
c
omput
e
r
vis
ion
a
lgor
it
hms
.
T
his
da
tas
e
t
is
one
of
the
mos
t
popular
c
hoice
s
f
or
thi
s
pur
pos
e
.
M
a
c
hine
lea
r
ning
a
lgo
r
it
hms
that
identif
y
objec
ts
in
im
a
ge
s
lea
r
n
by
e
xa
mpl
e
,
a
nd
C
I
F
AR
-
10
pictur
e
s
a
r
e
be
ne
f
icia
l
f
or
th
is
type
of
tr
a
in
ing.
T
he
da
tas
e
t
is
de
s
igned
to
he
lp
c
omput
e
r
s
dis
ti
nguis
h
be
twe
e
n
dif
f
e
r
e
nt
objec
ts
in
im
a
ge
s
,
a
nd
it
is
an
e
xc
e
ll
e
nt
r
e
s
our
c
e
f
or
r
e
s
e
a
r
c
he
r
s
who
wa
nt
to
tes
t
mul
ti
ple
a
lgor
it
h
ms
.
T
he
pictur
e
s
in
C
I
F
AR
-
10
a
r
e
r
e
latively
low
-
r
e
s
olut
ion
(
32
×
32)
,
whic
h
make
s
them
pa
r
ti
c
ular
ly
us
e
f
ul
f
o
r
tes
ti
ng
dif
f
e
r
e
nt
C
NN
.
3.
2.
1.
Dat
as
e
t
d
e
s
c
r
ip
t
ion
C
I
F
AR
-
10
c
ontains
60,
000
32
×
32
c
olor
pictur
e
s
in
10
c
a
tegor
ies
.
F
igur
e
2
s
hows
that
a
ir
plane
s
,
c
a
r
s
,
bir
ds
,
c
a
ts
,
de
e
r
,
dogs
,
f
r
ogs
,
hor
s
e
s
,
s
hips
,
a
nd
tr
uc
ks
each
ha
ve
6,
000
photos
.
T
he
da
tas
e
t
c
ontain
s
50,
000
tr
a
ini
ng
a
nd
10,
000
tes
t
photos
.
F
ive
ba
tche
s
of
10,
000
tr
a
ini
ng
pictur
e
s
e
xis
t.
One
s
e
t
of
1,
000
tes
t
pho
togr
a
phs
pe
r
c
a
tegor
y
is
us
e
d.
T
he
t
r
a
ini
ng
photos
a
r
e
r
a
nd
oml
y
s
or
ted,
with
s
ome
ba
tche
s
ha
ving
mo
r
e
of
o
ne
type.
E
ve
r
y
tr
a
ini
ng
ba
tch
ha
s
5,
000
photos
f
r
om
each
c
a
tegor
y.
C
I
F
AR
-
10
c
ontains
60,
000
32
×
32
c
olor
pi
c
tur
e
s
in
10
c
a
tegor
ies
.
F
igur
e
2
s
hows
that
a
ir
plane
s
,
c
a
r
s
,
bir
ds
,
c
a
ts
,
de
e
r
,
dogs
,
f
r
ogs
,
ho
r
s
e
s
,
s
hips
,
a
nd
tr
u
c
ks
each
ha
ve
6,
000
photos
.
T
he
da
tas
e
t
c
ontains
50,
000
t
r
a
i
ning
a
nd
10,
000
tes
t
photos
.
F
ive
ba
tche
s
of
10,
000
tr
a
ini
ng
pictur
e
s
e
xis
t.
One
s
e
t
of
1
,
000
tes
t
photog
r
a
phs
pe
r
c
a
tegor
y
is
us
e
d.
T
he
tr
a
ini
ng
photos
a
r
e
r
a
ndoml
y
s
or
ted,
with
s
ome
ba
tche
s
ha
ving
mor
e
of
one
type.
E
ve
r
y
tr
a
ini
ng
ba
tch
ha
s
5,
000
photos
f
r
o
m
each
c
a
tegor
y.
F
ig
ur
e
2
.
S
a
mpl
e
da
tas
e
t
f
or
objec
t
de
tec
ti
on
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
nov
e
l
mode
l
to
de
tec
t
and
c
ategor
iz
e
objec
ts
fr
o
m
image
s
by
us
ing
a
hy
br
id
mac
hine
…
(
N
il
ambar
Se
thi
)
671
3.
3
.
M
od
e
ls
u
s
e
d
To
r
e
a
c
h
the
objec
ti
ve
of
thi
s
wo
r
k,
we
us
e
d
VG
G1
6
,
YO
L
O
-
v3,
YO
L
O
-
v4
,
a
nd
the
pr
opos
e
d
model
,
whic
h
c
ons
is
ts
of
s
im
plene
t
-
r
a
ndom
f
or
e
s
t
models
.
F
or
thes
e
models
,
we
ha
ve
e
va
luate
d
va
r
ious
pe
r
f
or
manc
e
metr
ics
.
T
hos
e
a
r
e
a
c
c
ur
a
c
y,
pr
e
c
is
ion,
r
e
c
a
ll
,
a
nd
F1
s
c
or
e
.
T
his
s
e
c
ti
on
de
s
c
r
ibes
e
ach
model
c
lea
r
l
y
to
he
lp
the
r
e
a
de
r
unde
r
s
tand
thi
s
wo
r
k.
3.
3.
1.
VGG16
VG
G16
is
a
DC
NN
a
r
c
hit
e
c
tur
e
de
ve
loped
by
the
Oxf
or
d
vis
ua
l
ge
ometr
y
gr
oup
.
T
he
ne
two
r
k
wa
s
tr
a
ined
us
ing
I
mage
Ne
t,
a
c
oll
e
c
ti
on
of
mi
ll
ions
of
tagge
d
pictur
e
s
or
ga
nize
d
int
o
1
,
000
c
a
tegor
ies
.
VG
G16
c
ons
is
ts
of
3
f
ull
y
c
onne
c
ted
a
nd
13
c
onvolut
ional
laye
r
s
,
a
nd
it
uti
li
z
e
s
max
-
pooli
ng
laye
r
s
with
2
×
2
f
il
ter
s
a
nd
a
s
tr
ide
of
2
pixels
.
T
he
c
onvolut
ional
laye
r
s
us
e
s
mall
3
×
3
f
il
ter
s
with
1.
T
he
ne
twor
k
ha
s
mor
e
f
il
ter
s
,
r
a
nging
f
r
o
m
64
in
the
f
ir
s
t
laye
r
to
512
in
the
f
inal
laye
r
.
You
can
f
ind
the
model’
s
a
r
c
hit
e
c
tur
e
in
F
i
gur
e
3
.
F
ig
ur
e
3
.
VGG
a
r
c
hit
e
c
tur
e
Algor
it
hm
1:
VG
G16
S
tep
1:
D
e
f
ine
the
tr
a
ns
f
or
mat
ions
f
or
tr
a
ini
ng
a
nd
tes
ti
ng
da
ta
S
tep
2:
De
f
ine
a
f
unc
ti
on
f
or
loading
t
r
a
ini
ng
da
ta
a.
`
tr
a
in_t
r
a
ns
f
or
m
`
is
us
e
d
to
de
f
ine
the
tr
a
ns
f
or
mati
on.
b.
L
oa
d
the
C
I
F
AR
10
da
tas
e
t
f
or
t
r
a
ini
ng
us
ing
`
tor
c
h
-
vis
ion
Da
tas
e
t
s
.
’
c.
C
I
F
AR
10`
,
a
nd
then
us
e
the
`
Da
taL
oa
de
r
`
c
las
s
f
r
om
`
tor
c
h
.
uti
ls
.
da
ta`
to
load
the
da
ta.
d.
R
e
tur
n
the
da
ta
loade
r
.
S
tep
3:
C
r
e
a
te
a
f
unc
ti
on
f
or
loading
tes
ti
ng
da
ta,
us
ing
`
tes
t_t
r
a
ns
f
or
m`
to
de
f
ine
the
tr
a
ns
f
or
mation.
a.
L
oa
d
the
C
I
F
AR
10
da
tas
e
t
f
or
tes
ti
ng
us
ing
`
tor
c
hv
is
ion.
da
tas
e
ts
.
C
I
F
AR
10`
,
then
us
e
the
`
Da
taL
oa
de
r
`
c
las
s
to
load
the
da
ta.
b.
R
e
tur
n
the
da
ta
loade
r
.
S
tep
4:
Us
e
the
f
unc
ti
ons
c
r
e
a
ted
in
s
tep
2
a
nd
s
tep
3
to
ge
t
the
tr
a
in
a
nd
tes
t
da
ta
loade
r
s
,
r
e
s
pe
c
ti
ve
ly.
S
tep
5:
D
e
f
ine
a
dictionar
y
c
las
s
.
S
tep
6:
C
r
e
a
te
the
B
ott
lene
c
k
c
las
s
,
a.
I
nit
ialize
the
bott
lene
c
k
block
laye
r
s
us
ing
the
`
ini
t
`
f
unc
ti
on
.
b.
De
f
ine
the
f
o
r
wa
r
d
f
unc
t
ion
to
pa
s
s
input
thr
ough
t
he
bott
lene
c
k.
S
tep
7:
C
r
e
a
te
the
T
r
a
ns
it
ion
c
las
s
a.
I
nit
ialize
the
t
r
a
ns
it
ion
block
laye
r
s
us
ing
the
`
in
it
`
f
unc
ti
on
b.
S
e
t
the
f
o
r
wa
r
d
f
unc
ti
on
to
pa
s
s
input
via
the
tr
a
ns
i
ti
on
block.
S
tep
8:
C
r
e
a
te
the
VG
G16
ne
twor
k
by
ini
ti
a
li
z
ing
the
VG
G16
ne
twor
k
laye
r
s
us
ing
the
`
ini
t`
f
unc
ti
on.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
1
4
,
No.
1
,
F
e
br
ua
r
y
202
5
:
66
7
-
679
672
a.
S
e
t
the
inner
_c
ha
nne
ls
a
nd
c
onv1
laye
r
s
,
a
nd
c
r
e
a
te
de
ns
e
a
nd
tr
a
ns
it
ion
blocks
us
ing
`
_make
_de
ns
e
_laye
r
s
.
b.
Add
the
f
inal
ba
tch
nor
m
a
nd
r
e
c
ti
f
ied
li
ne
a
r
unit
(
R
e
L
U
)
a
c
ti
va
ti
on
laye
r
,
a
nd
c
las
s
if
y
us
ing
a
ve
r
a
ge
pooli
ng
a
nd
f
ull
y
l
inked
laye
r
s
.
c.
De
f
ine
the
VG
G16
ne
twor
k
input
f
o
r
wa
r
d
f
unc
ti
on
with
the
a
ppr
op
r
iate
pa
r
a
mete
r
s
.
S
tep
9:
C
r
e
a
te
a
VG
G16
ne
twor
k
with
the
a
ppr
opr
iate
pa
r
a
mete
r
s
.
F
or
the
e
xplana
ti
on
f
or
A
lgor
it
hm
1,
we
ne
e
d
to
i
mpor
t
s
e
ve
r
a
l
l
ibr
a
r
ies
,
including
nu
mpy,
pa
nda
s
,
os
,
ti
me,
matplot
li
b
,
a
nd
P
yT
or
c
h
.
We
mus
t
a
ls
o
im
por
t
r
e
late
d
modul
e
s
f
o
r
da
ta
p
r
oc
e
s
s
ing
,
model
de
ve
lopm
e
nt,
a
nd
t
r
a
ini
ng,
s
uc
h
as
tor
c
h
a
nd
to
r
c
h
vis
ion.
Ne
xt,
we
pe
r
f
or
m
da
ta
pr
e
p
r
oc
e
s
s
ing
to
pr
e
pa
r
e
the
da
ta
f
or
the
model.
T
h
is
invol
ve
s
us
ing
tr
a
in_t
r
a
ns
f
or
m
a
nd
tes
t_t
r
a
ns
f
or
m
to
pe
r
f
or
m
da
ta
a
ugm
e
ntation
methods
s
uc
h
as
r
a
ndom
hor
izonta
l
f
li
p,
r
otation,
a
nd
nor
maliza
ti
on
.
We
then
us
e
ge
t_t
r
a
ini
ng_da
taloa
de
r
a
nd
ge
t_t
e
s
ti
ng_da
taloa
de
r
to
ge
ne
r
a
te
da
ta
loade
r
s
f
or
loading
t
r
a
ini
ng
a
nd
tes
ti
ng
da
ta
ba
tche
s
.
T
he
s
e
f
unc
ti
ons
pr
e
pr
oc
e
s
s
the
da
ta
us
ing
tr
a
in_t
r
a
ns
f
or
m
a
nd
tes
t_
tr
a
ns
f
or
m
t
r
a
ns
f
or
mer
s
.
Onc
e
the
da
ta
is
pr
e
pa
r
e
d,
we
move
on
to
da
ta
vis
ua
li
z
a
ti
on
us
ing
M
a
tpl
otl
ib.
He
r
e
,
we
plo
t
the
f
ir
s
t
25
tr
a
ini
ng
da
ta
pictu
r
e
s
,
dis
playing
im
a
ge
s
with
their
c
las
s
na
mes
in
a
5
×
5
g
r
id.
T
he
ne
xt
s
tep
is
model
a
r
c
hit
e
c
tur
e
,
whe
r
e
we
i
mpl
e
ment
De
ns
e
Ne
t
us
ing
P
yT
or
c
h
modul
e
s
.
T
he
de
s
ign
include
s
a
bott
lene
c
k
laye
r
,
a
tr
a
ns
it
ion
laye
r
,
a
nd
a
de
ns
e
block.
T
he
bott
lene
c
k
laye
r
c
ons
is
ts
of
thr
e
e
laye
r
s
:
ba
tch
nor
maliza
ti
on,
r
e
c
ti
f
ied
li
ne
a
r
unit
(
R
e
L
U
)
a
c
ti
va
ti
on,
a
nd
c
onvolut
ional
.
Af
ter
a
pplyi
ng
thes
e
laye
r
s
to
the
input
f
e
a
tur
e
map,
the
output
f
e
a
tur
e
map
is
c
onc
a
tena
ted
with
it
a
long
th
e
c
ha
nne
l
a
xis
to
f
or
m
the
f
inal
output
.
T
he
t
r
a
ns
it
ion
l
a
ye
r
c
ons
is
ts
of
ba
tch
nor
maliza
ti
on
a
nd
c
onvolut
ional
l
a
ye
r
s
.
Af
ter
a
pplyi
ng
thes
e
laye
r
s
to
the
input
f
e
a
t
ur
e
map,
the
output
f
e
a
tur
e
map
is
s
e
nt
thr
ough
an
a
ve
r
a
ge
pooli
ng
laye
r
to
downs
a
mpl
e
its
s
pa
ti
a
l
dim
e
ns
ions
by
2.
T
he
bott
lene
c
k
laye
r
s
a
nd
the
input
f
e
a
tur
e
map
f
o
r
m
the
de
ns
e
block
a
long
the
c
ha
nne
l
a
xis
bott
lene
c
k
laye
r
s
output
f
e
a
tur
e
maps
f
r
om
c
onc
a
tena
ted
f
e
a
tur
e
maps
.
C
onc
a
tena
ti
ng
the
output
f
e
a
tur
e
maps
f
r
om
each
bott
lene
c
k
laye
r
with
the
input
f
e
a
tur
e
map
a
long
t
he
c
ha
nne
l
a
xis
pr
oduc
e
s
the
f
inal
output
.
We
a
ls
o
us
e
the
VG
G16
model,
whic
h
us
e
s
bott
le
ne
c
k,
tr
a
ns
it
ion,
a
nd
de
ns
e
block
laye
r
s
.
A
f
ter
one
c
onvolut
ional
laye
r
,
de
ns
e
blocks
a
nd
t
r
a
ns
it
ion
laye
r
s
make
up
the
model.
T
he
f
inal
de
ns
e
block
output
unde
r
goe
s
ba
tch
nor
maliza
ti
on
a
nd
R
e
L
U
a
c
ti
va
ti
on
be
f
or
e
a
global
a
ve
r
a
ge
pooli
ng
laye
r
.
F
inal
out
put
c
las
s
pr
oba
bil
it
ies
a
r
e
pr
oduc
e
d
by
pa
s
s
ing
the
global
a
ve
r
a
ge
pooli
ng
laye
r
output
via
a
f
ull
y
c
onne
c
t
e
d
laye
r
.
F
inally,
we
tr
a
in
the
model
us
ing
the
s
tocha
s
ti
c
gr
a
dient
de
s
c
e
nt
(
S
GD
)
op
ti
mi
z
e
r
a
nd
lea
r
ning
r
a
te
s
c
he
duler
.
T
he
tr
a
ini
ng
pr
oc
e
dur
e
invol
ve
s
it
e
r
a
ti
ng
ove
r
ba
tche
s
of
tr
a
ini
ng
da
ta
,
c
a
lcula
ti
ng
the
f
or
w
a
r
d
pa
s
s
,
c
r
os
s
-
e
ntr
opy
los
s
f
unc
ti
on
,
g
r
a
dients
,
a
nd
opti
m
ize
r
model
we
ight
upda
tes
.
Af
t
er
a
pr
e
de
ter
m
ined
n
umber
of
tr
a
ini
ng
e
poc
hs
,
the
model’
s
pe
r
f
or
manc
e
is
a
s
s
e
s
s
e
d
us
ing
tes
ti
ng
da
ta.
Af
ter
t
r
a
ini
ng,
the
model
’
s
a
c
c
ur
a
c
y
a
nd
los
s
of
tes
ti
ng
da
ta
a
r
e
dis
playe
d.
3.
3.
2.
Hyb
r
id
m
od
e
l
(
s
im
p
le
N
et
-
r
an
d
om
f
or
e
s
t
)
T
he
s
im
pleN
e
t
-
r
a
ndom
f
or
e
s
t
hybr
id
model
is
an
mac
hine
lea
r
ning
a
ppr
oa
c
h
that
im
pr
ove
s
the
c
a
tegor
iza
ti
on
of
pictur
e
da
tas
e
ts
.
T
he
model
c
o
mbi
ne
s
two
tec
hniques
:
S
im
pleN
e
t,
a
C
NN
that
e
xtr
a
c
ts
r
e
leva
nt
f
e
a
tur
e
s
f
r
o
m
r
a
w
p
ictur
e
da
ta,
a
nd
a
r
a
nd
om
f
or
e
s
t
c
las
s
if
ier
,
whic
h
ha
ndl
es
high
-
dim
e
ns
ional
input
da
ta.
T
h
is
hybr
id
model
us
e
s
the
s
tr
e
ngths
of
b
oth
tec
hniques
to
a
c
hieve
be
tt
e
r
pe
r
f
or
manc
e
in
pictur
e
c
las
s
if
ica
ti
on
tas
ks
.
S
im
pleN
e
t
e
f
f
e
c
ti
ve
ly
e
xtr
a
c
ts
the
f
e
a
tur
e
s
f
r
om
r
a
w
pictur
e
da
ta.
At
the
s
a
me
ti
me,
the
r
a
ndom
f
o
r
e
s
t
c
las
s
if
ier
ge
ne
r
a
li
z
e
s
we
ll
to
ne
w
c
a
s
e
s
a
nd
ha
ndles
high
-
dim
e
ns
ional
input
d
a
ta
—
the
a
r
c
hit
e
c
tur
e
of
thi
s
hybr
id
model
is
p
r
e
s
e
nted
in
F
i
gur
e
4
.
F
ig
ur
e
4
.
Ar
c
hit
e
c
tur
e
of
hyb
r
id
model
(
s
im
ple
N
et
-
R
F
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
nov
e
l
mode
l
to
de
tec
t
and
c
ategor
iz
e
objec
ts
fr
o
m
image
s
by
us
ing
a
hy
br
id
mac
hine
…
(
N
il
ambar
Se
thi
)
673
A
l
go
r
i
t
hm
2:
Hy
b
r
i
d
m
od
e
l
(
s
im
pl
e
n
e
t
-
r
a
nd
om
f
or
e
s
t
)
S
t
e
p
1:
I
m
p
o
r
t
n
u
m
p
y
,
m
a
tp
l
o
t
l
ib
.
p
y
p
lo
t
,
R
a
nd
o
m
F
o
r
e
s
t
C
la
s
s
i
f
i
e
r
,
c
la
s
s
i
f
ic
a
t
io
n
_
r
e
p
o
r
t
,
c
o
n
f
us
io
n
_
ma
t
r
i
x
,
c
i
f
a
r
1
0
,
M
o
d
e
l
,
I
n
p
u
t
,
C
on
v
2
D
,
De
ns
e
,
F
l
a
tt
e
n
,
n
p
_ut
i
l
s
,
E
a
r
l
y
S
to
p
p
in
g
,
a
n
d
M
o
de
l
C
h
e
c
k
p
o
in
t
.
S
t
e
p
2:
D
i
vi
de
t
he
C
I
F
AR
-
10
d
a
t
a
s
e
t
i
nt
o
t
r
a
in
i
ng
a
nd
tes
ti
n
g
s
e
ts
us
in
g
c
if
a
r
1
0
.
l
oa
d_
da
t
a
(
)
.
Di
vi
de
e
a
c
h
p
i
xe
l
va
lu
e
by
2
55
to
no
r
ma
li
z
e
d
a
t
a
.
O
ne
-
h
ot
e
nc
od
e
l
a
b
e
ls
wi
th
np
_u
t
il
s
.
to
_c
a
te
go
r
ica
l
(
)
.
S
t
e
p
3:
Us
e
Ke
r
a
s
’
f
u
nc
ti
on
a
l
AP
I
to
d
e
f
i
ne
S
i
mp
le
Ne
t
—
mo
de
l
c
o
mp
il
a
t
io
n
us
in
g
c
a
t
e
g
or
ic
a
l
_c
r
os
s
e
n
t
r
o
py
l
os
s
,
A
da
m
o
p
ti
mi
z
e
r
,
a
nd
a
c
c
u
r
a
c
y
me
t
r
ic
.
S
a
v
e
t
he
be
s
t
mo
de
l
t
h
r
o
ug
ho
u
t
t
r
a
in
i
ng
us
i
ng
e
a
r
l
y
s
t
op
p
in
g
a
n
d
m
od
e
l
c
he
c
kp
o
in
t
c
a
ll
ba
c
ks
.
S
t
e
p
4:
T
r
a
in
S
im
pl
e
Ne
t
us
i
ng
t
he
mo
de
l
f
i
t
(),
b
a
t
c
h
s
iz
e
12
8
,
20
e
p
oc
hs
,
a
n
d
10
%
va
li
da
ti
on
da
ta
.
L
oa
d
t
he
s
t
or
e
d
be
s
t
S
i
mp
leN
e
t
mo
de
l
us
i
ng
mo
de
l
.
l
oa
d
_w
e
i
gh
ts
(
)
.
S
t
e
p
5:
Us
e
be
s
t_
mo
de
l
.
p
r
e
d
ic
t(
)
to
e
xt
r
a
c
t
tr
a
i
ni
ng
a
nd
t
e
s
t
S
i
mp
le
Ne
t
f
e
a
t
u
r
e
s
.
R
e
s
ha
pe
r
e
t
r
ie
ve
d
f
e
a
t
ur
e
s
to
a
2D
a
r
r
a
y
us
i
ng
r
e
s
ha
pe
,
S
t
e
p
6:
T
r
a
in
a
R
a
ndom
F
or
e
s
ts
model
us
ing
R
a
ndomF
or
e
s
tC
las
s
if
ier
(
)
,
100
e
s
ti
mator
s
,
e
ntr
opy
c
r
it
e
r
ia,
a
nd
42
r
a
ndom
s
tate
s
.
T
e
s
t
the
R
a
ndom
F
o
r
e
s
ts
model
us
ing
c
las
s
if
ica
ti
on_r
e
por
t(
)
a
nd
c
onf
us
ion_m
a
tr
ix(
)
.
S
t
e
p
7:
C
ha
r
t
l
os
s
a
n
d
a
c
c
u
r
a
c
y
us
i
ng
ma
tp
l
ot
li
b
.
P
y
pl
ot
.
S
t
e
p
8:
Us
e
mat
p
lo
tl
ib
.
P
yt
ho
n
us
e
s
th
e
R
a
nd
o
m
F
o
r
e
s
ts
mo
de
l
to
s
h
ow
th
e
a
c
c
u
r
a
te
a
n
d
pr
e
d
ic
te
d
l
a
be
ls
f
o
r
th
e
f
i
r
s
t
10
te
s
t
ph
ot
os
.
−
Algor
it
hm
e
xplana
ti
on
T
he
e
xplana
ti
on
f
or
A
lgo
r
it
hm
2
is
that
t
he
pr
oc
e
s
s
invol
ve
s
us
ing
the
C
I
F
AR
-
10
da
ta
s
e
t
to
de
ve
lop
a
DL
model
known
as
s
im
pleN
e
t.
T
his
model
is
then
us
e
d
to
e
xtr
a
c
t
f
e
a
tur
e
s
f
r
om
the
da
tas
e
t,
a
f
ter
whic
h
the
im
a
ge
s
in
the
da
tas
e
t
a
r
e
c
las
s
if
ied
us
ing
a
r
a
ndom
f
or
e
s
t
c
las
s
if
ier
ba
s
e
d
on
the
r
e
tr
ieve
d
f
e
a
tur
e
s
.
T
he
NumP
y,
matplot
li
b,
a
nd
Ke
r
a
s
modul
e
s
a
r
e
im
por
ted
f
or
t
he
C
I
F
AR
-
10
da
ta
s
e
t,
model
a
r
c
hit
e
c
tur
e
c
omponents
,
a
nd
c
a
ll
ba
c
k
methods
to
be
gin
the
pr
oc
e
s
s
.
T
he
s
c
iki
t
-
lea
r
n’
s
r
a
ndom
f
or
e
s
t
c
las
s
if
ier
,
c
las
s
if
ica
ti
on
r
e
por
t,
a
nd
c
onf
us
ion
matr
ix
(
C
F
M
)
f
unc
ti
ons
a
r
e
a
ls
o
im
por
te
d.
T
he
C
I
F
AR
-
10
da
tas
e
t
is
loade
d
a
nd
nor
malize
d
,
a
nd
the
labe
ls
a
r
e
one
-
hot
e
nc
ode
d
us
ing
np_uti
ls
.
to_ca
te
gor
ica
l.
T
he
s
im
pleN
e
t
model
c
om
pr
is
e
s
s
ix
c
onv
olut
ional
laye
r
s
a
nd
a
de
ns
e
output
laye
r
with
ten
uni
ts
pe
r
C
I
F
AR
-
10
c
las
s
.
T
he
model
us
e
s
c
a
tegor
ica
l
c
r
os
s
-
e
ntr
opy
as
the
los
s
f
unc
ti
on,
Ada
m
opti
mi
z
e
r
,
a
nd
a
c
c
ur
a
c
y
metr
ic.
T
he
s
im
p
leN
e
t
mo
de
l
is
t
r
a
ined
on
C
I
F
AR
-
10
f
o
r
20
e
poc
hs
,
wit
h
a
ba
tc
h
s
iz
e
of
128
a
nd
a
va
li
da
ti
o
n
s
pli
t
of
0
.
1
.
If
the
va
li
da
ti
on
los
s
doe
s
not
im
p
r
o
ve
a
f
te
r
f
ive
e
poc
hs
,
tr
a
ini
n
g
is
s
toppe
d,
a
nd
the
be
s
t
model
is
s
a
ve
d
us
ing
th
e
M
o
de
lC
he
c
kpoi
nt
c
a
l
lbac
k
me
tho
d.
A
f
te
r
load
ing
th
e
be
s
t
si
mp
leN
e
t
m
ode
l
,
th
e
r
e
s
ha
p
e
me
thod
is
us
e
d
to
f
la
tt
e
n
the
t
r
a
in
ing
a
n
d
tes
t
s
e
ts
int
o
one
-
d
im
e
ns
io
na
l
a
r
r
a
ys
f
or
f
e
a
t
ur
e
e
x
tr
a
c
ti
o
n.
T
he
s
e
f
e
a
t
ur
e
s
then
tr
a
i
n
a
r
a
n
dom
f
or
e
s
t
c
las
s
i
f
ie
r
wit
h
100
t
r
e
e
s
a
nd
e
ntr
o
py
f
or
n
ode
s
pl
it
t
ing
.
T
he
c
las
s
i
f
ica
t
ion
r
e
po
r
t
a
nd
C
F
M
f
unc
ti
ons
a
r
e
us
e
d
to
o
utp
ut
t
he
tes
t
s
e
t
pic
tu
r
e
labe
l
s
th
e
t
r
a
ine
d
r
a
ndom
f
o
r
e
s
t
c
las
s
if
ier
pr
e
d
icte
d.
T
he
a
ppr
oa
c
h
a
ls
o
dis
pla
ys
s
i
mpl
e
N
e
t
mo
de
l
los
s
a
nd
a
c
c
ur
a
c
y
c
ur
ve
s
du
r
i
ng
t
r
a
in
ing
.
F
i
na
ll
y,
the
a
lg
or
it
h
m
r
uns
ove
r
the
f
i
r
s
t
10
tes
t
p
hotos
,
in
dica
t
ing
t
he
a
c
tua
l
a
nd
pr
e
dic
t
e
d
la
be
ls
a
nd
s
how
ing
t
he
im
a
g
e
us
ing
ma
tpl
otl
ib
.
We
i
n
i
t
i
a
l
l
y
i
m
p
o
r
t
e
d
t
h
e
d
a
t
a
s
e
t
f
r
o
m
K
e
r
a
s
a
n
d
l
oa
d
e
d
it
u
s
in
g
d
a
t
a
l
o
a
de
r
s
.
D
a
t
a
p
r
e
p
a
r
a
t
i
on
i
n
v
o
l
ve
d
u
s
i
ng
t
r
a
i
n_
t
r
a
ns
f
o
r
m
a
n
d
t
e
s
t
_t
r
a
ns
f
o
r
m
a
f
te
r
d
a
ta
c
o
l
l
e
c
t
i
o
n
.
T
h
e
s
e
t
r
a
ns
f
o
r
m
e
r
s
e
mp
l
o
y
r
a
n
d
o
m
ho
r
i
z
on
t
a
l
f
l
i
ps
,
r
a
n
d
o
m
r
o
ta
t
i
o
n
,
a
nd
n
o
r
m
a
li
z
a
t
i
o
n
to
p
r
e
p
a
r
e
da
t
a
f
o
r
t
h
e
m
o
d
e
l
.
F
i
na
l
l
y
,
we
us
e
d
ge
t
_
t
r
a
i
n
i
ng
_
da
t
a
loa
d
e
r
a
nd
g
e
t
_
te
s
t
i
n
g
_d
a
ta
l
oa
d
e
r
to
c
o
ns
t
r
u
c
t
t
h
e
t
r
a
i
n
i
n
g
a
nd
t
e
s
t
i
ng
d
a
t
a
l
o
a
de
r
s
.
A
f
t
e
r
s
p
l
i
t
t
in
g
t
h
e
d
a
t
a
,
we
u
s
e
d
DL
t
e
c
hn
i
q
ue
s
l
ik
e
VG
G
16
a
n
d
a
h
y
b
r
id
m
o
d
e
l
(
s
i
mp
l
e
Ne
t
-
r
a
nd
o
m
f
o
r
e
s
t
)
to
r
e
c
o
gn
i
z
e
o
b
je
c
ts
in
i
mag
e
s
.
We
e
v
a
lu
a
te
d
t
h
e
ou
t
c
om
e
s
u
s
i
n
g
a
c
c
u
r
a
c
y
me
t
r
i
c
s
,
w
h
ic
h
s
ho
w
e
d
t
h
a
t
th
e
h
yb
r
i
d
m
o
de
l
h
a
d
an
a
c
c
u
r
a
c
y
of
8
9
.
6
%
.
4.
RE
S
UL
T
S
T
h
e
da
tas
e
t
is
p
r
e
p
r
o
c
e
s
s
e
d
a
f
te
r
c
o
l
lec
ti
o
n
.
Da
ta
p
r
e
p
a
r
a
t
i
on
in
vo
lv
e
s
us
in
g
t
r
a
in
_
tr
a
ns
f
or
m
a
nd
t
e
s
t_
t
r
a
ns
f
o
r
m
to
t
r
a
i
n
a
nd
tes
t
t
he
m
ode
l
by
lo
a
d
ing
ba
tch
e
s
of
p
r
e
pr
oc
e
s
s
e
d
da
ta
.
T
he
b
e
s
t
s
im
pl
e
Ne
t
m
o
de
l
is
l
oa
de
d
,
a
nd
th
e
r
e
s
ha
p
e
met
ho
d
is
u
s
e
d
to
f
l
a
t
te
n
t
he
t
r
a
in
in
g
a
nd
t
e
s
t
s
e
ts
i
nt
o
o
ne
-
di
me
ns
i
on
a
l
a
r
r
a
ys
to
e
x
t
r
a
c
t
f
e
a
t
ur
e
s
.
T
he
s
e
f
e
a
tu
r
e
s
a
r
e
th
e
n
u
ti
li
z
e
d
to
t
r
a
i
n
a
r
a
nd
om
f
o
r
e
s
t
c
l
a
s
s
i
f
ie
r
w
it
h
10
0
t
r
e
e
s
a
nd
e
n
t
r
o
py
f
o
r
n
ode
s
pl
it
t
in
g
.
T
he
c
l
a
s
s
i
f
ica
t
io
n
r
e
p
o
r
t
a
n
d
C
F
M
f
un
c
t
io
ns
ou
tp
ut
th
e
t
e
s
t
s
e
t
pi
c
t
u
r
e
lab
e
ls
p
r
e
di
c
te
d
by
th
e
t
r
a
i
ne
d
r
a
nd
om
f
o
r
e
s
t
c
las
s
if
ie
r
.
We
dis
c
o
ve
r
e
d
t
ha
t
V
GG
16
ha
s
an
a
c
c
u
r
a
c
y
of
8
4
.
7
%
,
w
h
il
e
t
he
hy
b
r
id
m
od
e
l
(
s
im
pl
e
Ne
t
-
r
a
n
do
m
f
o
r
e
s
t
)
h
a
s
an
a
c
c
u
r
a
c
y
of
89
.6
%.
T
h
e
r
e
f
o
r
e
,
th
e
h
yb
r
id
mo
de
l
is
89
.6
%
m
o
r
e
a
c
c
ur
a
te
.
4.
1.
Conf
u
s
ion
m
a
t
r
ix
f
or
VGG16
T
he
C
F
M
f
or
the
pr
opos
e
d
s
ys
tem
VG
G16
is
s
ho
wn
in
F
igur
e
5
.
T
he
va
li
da
ti
on
c
ur
ve
s
f
or
VG
G16
a
r
e
s
hown
in
F
igur
e
s
6
a
nd
7.
Us
ing
thi
s
C
F
M
make
s
it
e
a
s
y
to
e
va
luate
the
pe
r
f
or
manc
e
of
met
r
ics
to
pr
e
dict
a
c
c
ur
a
te
r
e
s
ult
s
.
F
o
r
that
pu
r
pos
e
,
we
ha
ve
c
a
lcula
t
e
d
the
C
F
M
.
Ac
c
ur
a
c
y
a
nd
l
os
s
f
o
r
VG
G16
a
r
e
a
ls
o
s
hown
in
F
igur
e
s
6
a
nd
7.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
1
4
,
No.
1
,
F
e
br
ua
r
y
202
5
:
66
7
-
679
674
F
ig
ur
e
5
.
C
F
M
f
o
r
VG
G16
F
igur
e
6.
Va
li
da
ti
on
c
u
r
ve
of
a
c
c
ur
a
c
y
f
or
VG
G16
F
igur
e
7.
Va
li
da
ti
on
c
u
r
ve
of
los
s
f
or
VG
G16
4.
2
.
Conf
u
s
ion
m
a
t
r
ix
f
or
h
yb
r
id
m
od
e
l
(
s
im
p
le
N
et
-
r
an
d
om
f
or
e
s
t
)
T
he
C
F
M
f
or
the
hybr
id
model
is
s
hown
in
F
igur
e
8.
T
he
va
li
da
ti
on
c
ur
ve
f
or
the
hybr
id
model
is
s
hown
in
F
igur
e
9.
T
he
pe
r
f
or
manc
e
metr
ics
us
e
d
in
thi
s
wor
k
a
r
e
s
hown
in
T
a
ble
1.
Us
ing
thi
s
C
F
M
make
s
it
e
a
s
y
to
e
va
luate
the
pe
r
f
or
manc
e
of
metr
ics
to
pr
e
dict
a
c
c
ur
a
te
r
e
s
ult
s
.
F
o
r
that
pu
r
pos
e
,
we
ha
ve
c
a
lcula
ted
the
C
F
M
.
Ac
c
ur
a
c
y
a
nd
los
s
f
or
hybr
id
models
a
r
e
a
ls
o
s
hown
in
F
igu
r
e
9.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
Ar
ti
f
I
ntell
I
S
S
N:
2252
-
8938
A
nov
e
l
mode
l
to
de
tec
t
and
c
ategor
iz
e
objec
ts
fr
o
m
image
s
by
us
ing
a
hy
br
id
mac
hine
…
(
N
il
ambar
Se
thi
)
675
F
ig
ur
e
8
.
C
F
M
f
o
r
hybr
id
model
F
ig
ur
e
9
.
V
a
li
da
ti
on
c
u
r
ve
s
f
or
the
hybr
id
model
T
a
ble
1
obs
e
r
ve
d
that
a
c
c
ur
a
c
y
wa
s
89.
6%
,
pr
e
c
is
i
on
wa
s
76%
,
r
e
c
a
ll
wa
s
78.
5%
,
F1
-
s
c
or
e
wa
s
81%
,
a
nd
M
a
tt
he
ws
c
or
r
e
lation
c
oe
f
f
icie
nt
(
M
C
C
)
got
0.
82%
f
or
the
p
r
opos
e
d
model
by
c
ompar
ing
it
with
e
xis
ti
ng
models
.
F
r
o
m
T
a
ble
1,
the
ob
taine
d
va
lues
f
o
r
a
c
c
ur
a
c
y
a
r
e
s
hown
in
F
igur
e
10
,
the
pe
r
f
or
manc
e
me
tr
ics
li
ke
pr
e
c
is
ion,
r
e
c
a
ll
,
a
nd
F
1
-
s
c
or
e
gr
a
ph
a
r
e
s
hown
in
F
igur
e
11,
a
nd
the
gr
a
ph
f
or
M
C
C
is
s
hown
in
F
igur
e
12.
T
he
outcome
s
a
r
e
e
va
luate
d
ba
s
e
d
on
a
c
c
ur
a
c
y,
pr
e
c
is
ion,
a
nd
r
e
c
a
ll
whe
n
tes
ti
ng
da
ta
to
make
p
r
e
dictions
.
T
he
be
s
t
model
is
r
e
c
omm
e
nde
d
a
f
te
r
c
ompa
r
ing
the
pr
e
dictions
made
by
dif
f
e
r
e
nt
a
lgor
it
hms
.
A
mong
the
s
ugge
s
ted
tec
hniques
,
VG
G16
a
c
hieve
d
a
c
las
s
if
ica
ti
on
a
c
c
ur
a
c
y
of
84.
7
%
,
while
the
hyb
r
i
d
model
(
s
im
pleN
e
t
-
r
a
ndom
f
or
e
s
t
)
a
tt
a
ined
an
a
c
c
ur
a
c
y
of
89.
6%
.
B
a
s
e
d
on
the
r
e
s
ult
s
,
the
hybr
id
model
pe
r
f
or
med
be
tt
e
r
be
twe
e
n
the
two
methods
.
T
a
ble
1
.
P
e
r
f
o
r
manc
e
of
the
model
S
.
No
M
ode
l
A
c
c
ur
a
c
y
P
r
e
c
is
io
n
R
e
c
a
ll
F1
-
s
c
or
e
M
C
C
1
V
G
G
16
84.7
72.3
73.5
75
0.68
2
Y
O
L
O
-
v3
82.8
71
69.8
73.4
0.71
3
Y
O
L
O
-
v4
85.7
74.2
75.6
79
0.74
4
P
r
opos
e
d
mode
l
(
S
im
pl
e
N
et
-
r
a
ndom f
or
e
s
t
)
89.6
76
78.5
81
0.82
F
ig
ur
e
10
.
Gr
a
ph
f
or
a
c
c
ur
a
c
y
with
va
r
ious
models
F
ig
ur
e
11
.
Gr
a
ph
f
or
va
r
ious
pe
r
f
or
manc
e
metr
ics
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2252
-
8938
I
nt
J
Ar
ti
f
I
ntell
,
Vol.
1
4
,
No.
1
,
F
e
br
ua
r
y
202
5
:
66
7
-
679
676
F
ig
ur
e
12
.
Gr
a
ph
f
or
M
C
C
with
va
r
ious
models
T
he
input
photos
a
r
e
de
picte
d
in
F
igu
r
e
13(
a
)
,
whil
e
the
c
r
oppe
d
one
s
a
r
e
s
hown
in
F
igur
e
13(
b)
.
We
e
xc
luded
blur
r
y
pho
tos
by
t
r
im
mi
ng
them.
T
he
c
r
oppe
d
pictur
e
s
a
r
e
dis
playe
d
in
F
igur
e
13.
To
ide
nti
f
y
a
nd
c
ount
it
e
ms
,
we
de
ve
loped
c
us
tom
YO
L
Ov4
f
unc
ti
ons
.
A
high
c
onf
idenc
e
s
c
or
e
im
pli
e
s
a
high
pr
oba
b
il
it
y
that
the
de
tec
ted
it
e
m
is
a
membe
r
of
the
c
las
s
that
YO
L
Ov4
pr
e
dicte
d.
Util
izing
a
thr
e
s
hold
a
ll
owe
d
us
to
e
xc
lude
de
tec
ti
ons
with
a
low
c
onf
idenc
e
leve
l.
(
a
)
(
b)
F
igur
e
13.
Output
s
c
r
e
e
ns
f
or
the
i
mage
s
a
s
given
i
n
(
a
)
input
im
a
ge
s
a
nd
f
r
om
that
(
b)
ou
tput
im
a
ge
s
Evaluation Warning : The document was created with Spire.PDF for Python.