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.
14
,
No.
4
,
Augus
t
2025
,
pp.
2964
~
2978
I
S
S
N:
2252
-
8938
,
DO
I
:
10
.
11591/i
jai
.
v
14
.i
4
.
pp
29
64
-
2978
2964
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
ai
.
iaes
c
or
e
.
c
om
E
n
h
an
c
in
g
t
r
ad
itio
n
al
m
ac
h
i
n
e
le
a
r
n
i
n
g
m
e
t
h
o
d
s
u
si
n
g
c
on
c
a
t
e
n
at
io
n
t
w
o
t
r
an
s
f
e
r
le
ar
n
in
g
f
or
c
la
ssi
f
ic
at
io
n
d
e
s
e
r
t
r
e
gi
on
s
Raf
al
Naz
ar
Youn
e
s
A
l
-
T
ah
an
1
,
Rub
a
T
alal
I
b
r
ah
im
2
1
D
e
pa
r
tm
e
nt
of
C
omput
e
r
S
c
ie
nc
e
,
T
he
G
e
ne
r
a
l
D
ir
e
c
to
r
a
te
of
E
duc
a
ti
on i
n N
in
e
ve
h G
ove
r
nor
a
te
,
M
os
ul
,
I
r
a
q
2
D
e
pa
r
tm
e
nt
of
C
omput
e
r
S
c
ie
nc
e
, C
ol
le
g
e
of
C
omput
e
r
S
c
ie
nc
e
a
nd M
a
th
e
ma
ti
c
s
, U
ni
ve
r
s
it
y of
M
o
s
ul
, M
os
ul
, I
r
a
q
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
S
e
p
11,
2024
R
e
vis
e
d
Apr
15,
2025
Ac
c
e
pted
J
un
8,
2025
D
es
er
t
s
c
o
v
er
a
s
i
g
n
i
f
i
can
t
p
o
r
t
i
o
n
of
the
ear
t
h
an
d
p
r
es
en
t
en
v
i
r
o
n
me
n
t
a
l
an
d
eco
n
o
m
i
c
d
i
ff
i
cu
l
t
i
es
o
w
i
n
g
to
t
h
ei
r
h
a
rs
h
c
o
n
d
i
t
i
o
n
s
.
Sat
e
l
l
i
t
e
rem
o
t
e
s
en
s
i
n
g
i
ma
g
es
(SRSI)
h
av
e
ev
o
l
v
ed
i
n
t
o
an
i
mp
o
rt
a
n
t
t
o
o
l
fo
r
mo
n
i
t
o
ri
n
g
an
d
s
t
u
d
y
i
n
g
t
h
e
s
e
reg
i
o
n
s
as
t
ec
h
n
o
l
o
g
y
h
as
a
d
v
a
n
ce
d
.
Mach
i
n
e
l
e
arn
i
n
g
(ML
)
is
cri
t
i
cal
in
e
v
al
u
at
i
n
g
t
h
es
e
i
mag
e
s
an
d
ex
t
rac
t
i
n
g
v
a
l
u
a
b
l
e
i
n
f
o
rmat
i
o
n
fro
m
t
h
em,
res
u
l
t
i
n
g
in
a
b
et
t
er
k
n
o
w
l
ed
g
e
of
h
ard
s
e
t
t
i
n
g
s
a
n
d
i
n
crea
s
i
n
g
effo
r
t
s
t
o
w
ard
s
u
s
t
ai
n
ab
l
e
d
ev
e
l
o
p
me
n
t
in
d
es
er
t
reg
i
o
n
s
.
As
a
res
u
l
t
,
in
t
h
i
s
s
t
u
d
y
,
fo
u
r
ML
ap
p
ro
ach
e
s
w
ere
en
h
an
ce
d
by
h
y
b
ri
d
i
z
i
n
g
t
h
em
w
i
t
h
p
re
-
t
rai
n
i
n
g
met
h
o
d
s
to
ach
i
ev
e
mu
l
t
i
mo
d
e
l
l
ear
n
i
n
g
.
T
w
o
p
re
-
t
rai
n
i
n
g
ap
p
r
o
ach
e
s
(X
cep
t
i
o
n
an
d
D
e
n
es
e
N
et
2
0
1
)
w
ere
u
s
ed
to
ex
t
ract
feat
u
res
,
w
h
i
ch
w
ere
c
o
n
ca
t
en
a
t
ed
a
n
d
fe
d
i
n
t
o
ML
al
g
o
r
i
t
h
ms
:
l
i
g
h
t
g
ra
d
i
e
n
t
b
o
o
s
t
i
n
g
m
o
d
e
l
(L
G
BM),
d
eci
s
i
o
n
t
ree
(
D
T
),
k
-
n
eare
s
t
n
e
i
g
h
b
o
rs
(
K
N
N
),
an
d
n
aï
v
e
Bay
es
(N
B
).
In
ad
d
i
t
i
o
n
,
an
en
s
em
b
l
e
v
o
t
i
n
g
w
as
u
s
e
d
to
i
mp
r
o
v
e
the
o
u
t
c
o
mes
of
ML
al
g
o
r
i
t
h
ms
(
D
T
,
N
B,
an
d
K
N
N
)
an
d
o
v
erc
o
me
t
h
e
i
r
fl
aw
s
.
The
m
o
d
e
l
s
w
er
e
t
es
t
ed
on
t
w
o
d
a
t
as
e
t
s
an
d
h
y
b
r
i
d
L
G
BM
o
u
t
p
erf
o
rmed
o
t
h
er
t
rad
i
t
i
o
n
al
ML
met
h
o
d
s
by
9
9
%
in
accu
racy
,
p
reci
s
i
o
n
,
recal
l
,
an
d
F1
s
co
re,
an
d
by
1
0
0
%
in
area
u
n
d
er
the
c
u
rv
e
(
A
U
C
)
-
r
ece
i
v
er
o
p
era
t
i
n
g
ch
arac
t
eri
s
t
i
c
(
RO
C
)
.
K
e
y
w
o
r
d
s
:
De
c
is
ion
tr
e
e
De
ns
e
Ne
t201
K
-
ne
a
r
e
s
t
ne
ighbor
L
ight
gr
a
dient
boos
ti
ng
model
M
a
c
hine
lea
r
ning
S
a
telli
te
r
e
mot
e
s
e
ns
ing
im
a
ge
s
Xc
e
pti
on
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
:
R
uba
T
a
lal
I
br
a
him
De
pa
r
tm
e
nt
of
C
omput
e
r
S
c
ienc
e
,
C
oll
e
ge
of
C
om
puter
S
c
ienc
e
a
nd
M
a
thema
ti
c
s
,
Unive
r
s
it
y
of
M
os
ul
Al
-
Ha
dba
a
R
oa
d
-
M
o
s
ul,
I
r
a
q
E
mail:
r
uba
tala
l@uom
os
ul.
e
du.
iq
1.
I
NT
RODU
C
T
I
ON
O
ur
p
la
n
e
t
E
a
r
th
is
m
a
de
up
of
2
9%
l
a
n
d
(
c
on
ti
ne
nt
s
a
n
d
i
s
l
a
n
d
s
)
,
wi
th
t
h
e
r
e
m
a
ini
ng
71%
c
o
ntr
ol
le
d
by
wa
ter
[
1]
.
T
h
e
l
a
n
d
is
th
e
n
s
e
pa
r
a
t
e
d
in
to
tw
o
c
a
te
gor
i
e
s
:
l
a
nd
s
s
uit
e
d
f
o
r
h
a
b
it
a
t
io
n
a
n
d
a
gr
ic
ul
tur
e
(
gr
e
e
n
m
e
a
do
w
s
)
a
n
d
b
a
r
r
e
n
de
s
e
r
t
gr
o
u
nd
s
t
ha
t
a
r
e
un
s
uit
a
bl
e
f
or
bo
th
[
2]
.
D
e
s
e
r
t
if
i
c
a
t
io
n
is
a
n
a
tur
a
l
ph
e
n
om
e
n
on
th
a
t
c
a
u
s
e
s
l
a
n
d
d
e
t
e
r
i
or
a
ti
on
ow
in
g
to
wi
nd
a
nd
d
r
if
ti
n
g
s
a
nd.
It
is
one
of
th
e
e
n
vir
on
me
nt
a
l
di
s
a
s
t
e
r
s
t
h
a
t
m
u
s
t
be
pr
e
v
e
n
te
d
or
mi
t
ig
a
t
e
d
by
kn
o
win
g
its
d
yn
a
m
ic
gr
ow
th
a
n
d
e
x
t
e
nt,
as
w
e
ll
as
e
x
a
mi
ni
ng
c
li
ma
ti
c
a
n
d
g
e
o
gr
a
ph
i
c
a
l
to
p
olo
gy
s
u
c
h
as
t
e
mp
e
r
a
t
ur
e
,
p
la
nt
c
ov
e
r
,
r
a
i
nf
a
l
l
r
a
t
e
,
l
a
t
it
ud
e
,
a
nd
lo
ng
it
u
d
e
[
3]
.
In
r
e
c
e
n
t
d
e
c
a
de
s
,
r
e
m
ot
e
s
e
n
s
i
ng
s
a
t
e
l
li
t
e
s
h
a
ve
be
e
n
a
b
l
e
to
r
e
c
or
d
th
e
E
a
r
t
h'
s
t
op
ol
og
y,
r
e
s
ul
ti
ng
in
r
e
mo
t
e
s
e
n
s
in
g
i
ma
ge
s
wi
th
hi
gh
r
e
s
o
lu
ti
on
a
n
d
a
dv
a
n
c
e
d
pr
o
c
e
s
s
in
g,
as
w
e
l
l
as
lo
w
p
r
i
c
e
s
a
nd
r
a
pi
d
a
c
qu
i
s
it
io
n
[
4]
.
R
e
s
e
a
r
c
h
e
r
s
ha
v
e
be
e
n
a
bl
e
to
u
s
e
t
he
s
e
i
m
a
g
e
s
in
m
a
ny
im
po
r
t
a
n
t
r
e
s
e
a
r
c
h
f
ie
ld
s
s
u
c
h
as
e
nv
ir
onm
e
nt
a
l
mo
ni
tor
in
g
[
5]
,
n
a
v
ig
a
t
io
n
a
n
d
m
a
p
pin
g
[6
]
,
G
oo
gl
e
E
a
r
t
h
a
n
d
O
p
e
n
S
tr
e
e
t
M
a
p
[
7]
.
M
a
n
u
a
l
c
l
a
s
s
if
ic
a
ti
o
n
of
na
t
ur
a
l
r
e
g
io
n
s
u
s
in
g
g
e
og
r
a
p
hi
c
inf
or
m
a
t
io
n
s
ys
t
e
m
s
or
r
e
m
ot
e
s
e
n
s
i
ng
is
e
x
c
e
e
d
in
gl
y
c
ha
ll
e
n
gi
ng
o
wi
ng
to
a
la
c
k
of
un
d
e
r
s
t
a
n
di
ng
of
th
e
ir
s
u
r
f
a
c
e
,
g
e
o
gr
a
p
hy,
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
E
nhanc
ing
tr
adit
ional
mac
hine
lear
ning
me
thods
us
ing
c
onc
atenation
…
(
R
afal
N
az
ar
Y
oune
s
A
l
-
T
a
han
)
2965
a
n
d
c
h
a
n
gi
ng
il
lum
in
a
t
io
n,
a
ll
of
w
hi
c
h
ha
v
e
r
e
s
u
lt
e
d
in
a
c
on
s
ta
nt
c
ha
ng
e
in
th
e
f
e
a
tu
r
e
s
of
th
e
s
e
r
e
gio
n
s
,
e
s
p
e
c
i
a
ll
y
d
e
s
e
r
t
r
e
gi
on
s
.
T
he
r
e
f
or
e
,
an
e
l
e
c
tr
o
nic
c
la
s
s
if
i
c
a
t
io
n
h
a
s
b
e
c
om
e
ne
c
e
s
s
a
r
y
be
c
a
u
s
e
it
d
o
e
s
n
ot
r
e
q
uir
e
ti
me,
e
f
f
or
t,
a
n
d
f
i
e
l
d
e
x
pe
r
t
s
,
un
li
ke
m
a
nu
a
l
c
la
s
s
if
i
c
a
ti
on,
w
hi
c
h
a
l
s
o
r
e
q
uir
e
s
hi
gh
c
o
s
t
of
l
o
c
a
li
z
a
ti
on
a
n
d
th
e
w
il
l
in
gn
e
s
s
of
d
a
t
a
a
n
a
l
y
s
i
s
to
u
nd
e
r
t
a
ke
l
e
s
s
r
e
s
e
a
r
c
h.
In
r
e
c
e
nt
y
e
a
r
s
,
m
a
c
h
in
e
l
e
a
r
n
in
g
(
M
L
)
te
c
h
ni
qu
e
s
h
a
v
e
e
mer
g
e
d
to
c
l
a
s
s
if
y
na
tur
a
l
r
e
g
io
n
s
u
s
i
ng
r
e
mo
te
s
e
n
s
in
g
im
a
g
e
s
a
n
d
h
a
v
e
a
c
hi
e
v
e
d
go
od
r
e
s
ult
s
[
8]
.
M
a
ny
r
e
s
e
a
r
c
he
r
s
ha
ve
u
s
e
d
p
r
e
-
tr
a
i
ne
d
M
L
me
th
od
s
to
c
l
a
s
s
if
y
de
s
e
r
t,
gr
e
e
n
,
a
n
d
wa
ter
r
e
g
io
ns
a
nd
h
a
v
e
a
c
h
ie
v
e
d
e
x
c
e
l
le
nt
r
e
s
u
lt
s
,
bu
t
s
o
m
e
of
t
he
m
h
a
ve
f
a
c
e
d
pr
ob
le
m
s
s
uc
h
as
o
v
e
r
f
i
tt
in
g,
w
hi
c
h
h
a
s
a
f
f
e
c
te
d
t
he
a
c
c
ur
a
c
y
[
9]
.
To
r
e
du
c
e
t
h
e
s
pr
e
a
d
of
de
s
e
r
t
r
e
gi
on
s
a
n
d
t
he
p
h
e
n
om
e
n
on
of
de
s
e
r
t
if
i
c
a
t
io
n
th
a
t
h
a
s
in
c
r
e
a
s
e
d
r
e
c
e
nt
ly
a
nd
to
e
nc
our
a
ge
s
u
s
ta
in
a
b
le
a
gr
ic
ul
tu
r
e
t
hi
s
s
t
ud
y
i
nv
e
s
t
ig
a
t
e
d
to
c
la
s
s
if
y
na
tur
a
l
r
e
gi
on
s
(
d
e
s
e
r
t,
gr
e
e
n,
a
n
d
w
a
t
e
r
)
a
nd
de
te
c
ts
d
e
s
e
r
t
r
e
g
io
n
s
in
p
a
r
ti
c
u
la
r
,
in
a
d
di
ti
on
to
i
mpr
ov
in
g
th
e
q
u
a
li
ty
of
t
r
a
di
ti
o
n
a
l
M
L
t
e
c
h
niq
u
e
s
d
ue
to
t
he
ir
s
a
tu
r
a
ti
o
n
w
h
e
n
c
l
a
s
s
if
yi
ng
a
l
a
r
ge
a
mo
un
t
of
d
a
t
a
,
as
w
e
l
l
as
to
a
v
oi
d
ov
e
r
f
it
t
in
g,
wh
ic
h
u
s
u
a
ll
y
o
c
c
ur
s
in
pr
e
-
tr
a
i
n
e
d
l
e
a
r
n
in
g
th
a
t
a
c
qu
ir
e
s
w
e
ig
ht
s
a
n
d
bi
a
s
e
s
th
a
t
r
e
p
r
e
s
e
n
t
th
e
c
h
a
r
a
c
ter
i
s
t
ic
s
of
th
e
d
a
t
a
s
e
t.
S
o,
tr
a
d
it
i
on
a
l
a
n
d
mo
d
e
r
n
M
L
m
e
t
ho
d
s
w
e
r
e
h
ybr
id
iz
e
d
w
it
h
t
wo
pr
e
-
tr
a
i
ne
d
l
e
a
r
ni
ng
me
th
od
s
,
Xc
e
pt
io
n
a
nd
D
e
n
s
e
Ne
t
20
1,
to
e
xt
r
a
c
t
f
e
a
t
ur
e
s
a
n
d
a
c
hi
e
v
e
hi
gh
-
q
ua
li
ty
a
n
d
m
ul
ti
-
mo
de
l
l
e
a
r
n
in
g
whi
l
e
pr
e
vio
us
s
t
ud
i
e
s
h
a
ve
n
ot
a
dd
r
e
s
s
e
d
t
hi
s
.
T
h
e
s
e
f
e
a
tur
e
s
w
e
r
e
t
h
e
n
p
a
s
s
e
d
to
m
or
e
th
a
n
o
ne
M
L
te
c
hni
qu
e
to
be
c
la
s
s
if
ie
d.
A
d
dit
io
na
ll
y,
an
e
n
s
e
mb
le
me
th
od
wa
s
u
s
e
d
b
e
tw
e
e
n
mor
e
M
L
m
e
t
ho
d
s
to
im
pr
o
ve
a
c
c
ur
a
c
y.
T
h
e
pa
pe
r
'
s
c
o
ntr
ib
uti
on
s
c
a
n
be
s
um
ma
r
i
z
e
d
in
t
he
f
ol
lo
wi
ng
poi
nt
s
:
‒
A
p
pl
yin
g
tw
o
d
a
t
a
s
e
t
s
,
f
ir
s
t
s
a
t
e
ll
it
e
r
e
mo
te
s
e
n
s
i
ng
i
ma
g
e
s
(
S
R
S
I
)
w
a
s
ta
k
e
n
f
r
o
m
K
a
g
gl
e
w
e
b
s
i
te
a
n
d,
th
e
s
e
c
on
d
d
a
t
a
s
e
t
w
a
s
c
o
ll
e
c
te
d
f
r
om
m
a
ny
w
e
b
s
it
e
s
li
ke
:
K
a
g
gl
e
,
N
AS
A,
a
n
d
Ni
mb
o.
A
ls
o,
o
nl
y
t
hr
e
e
c
l
a
s
s
e
s
w
e
r
e
t
a
k
e
n
f
r
o
m
t
he
tw
o
d
a
t
a
s
e
t
s
,
wh
i
c
h
a
r
e
(
d
e
s
e
r
t,
gr
e
e
n
a
r
e
a
s
,
a
n
d
w
a
t
e
r
)
.
‒
P
e
r
f
or
m
in
g
ma
ny
pr
e
pr
o
c
e
s
s
in
g
on
t
he
t
wo
d
a
t
a
s
e
t
s
,
s
u
c
h
as
(
r
e
s
i
z
i
ng
im
a
ge
s
,
tr
a
n
s
f
or
ma
ti
on,
c
a
nn
y
d
e
te
c
t
io
n,
b
ou
nd
in
g
b
ox
im
a
g
e
s
,
c
r
o
pp
in
g
im
a
ge
s
,
a
n
d
n
or
m
a
l
iz
a
t
io
n)
.
‒
A
p
pl
yin
g
tw
o
t
r
a
n
s
f
e
r
l
e
a
r
n
in
g
te
c
h
ni
qu
e
s
(
X
c
e
p
ti
o
n
a
n
d
D
e
n
s
e
N
e
t
20
1)
to
e
xtr
a
c
t
f
e
a
t
ur
e
s
a
n
d
a
c
c
omp
li
s
h
m
ult
i
-
m
od
e
l
le
a
r
ni
ng.
‒
A
f
t
e
r
c
o
nc
a
t
e
na
ti
ng
th
e
tr
a
n
s
f
e
r
le
a
r
ni
ng
ou
tc
ome
s
,
t
he
mul
ti
-
f
e
a
t
ur
e
s
w
il
l
be
f
e
d
i
nt
o
tr
a
d
it
io
na
l
a
n
d
m
od
e
r
n
M
L
a
l
gor
it
h
m
s
s
u
c
h
as
(
l
ig
ht
gr
a
di
e
n
t
boo
s
ti
ng
m
od
e
l
(
L
G
B
M
)
,
de
c
is
io
n
tr
e
e
(
DT
)
,
k
-
ne
a
r
e
s
t
n
e
ig
hb
or
s
(
K
NN
)
,
a
nd
n
a
ï
v
e
B
a
y
e
s
(
NB
)
).
a
dd
it
i
on
a
l
ly,
th
e
w
or
k
us
e
d
an
e
n
s
e
mb
le
v
ot
in
g
m
e
t
ho
d
a
m
on
g
s
t
t
hr
e
e
tr
a
di
ti
on
a
l
ML
a
l
go
r
it
hm
s
to
im
p
r
ov
e
a
c
c
ur
a
c
y
a
n
d
p
e
r
f
or
m
a
n
c
e
.
T
he
r
e
m
a
in
d
e
r
of
th
e
s
tu
dy
is
s
tr
u
c
t
ur
e
d
as
f
o
ll
ow
s
:
s
e
c
ti
on
2
w
il
l
a
dd
r
e
s
s
r
e
l
a
t
e
d
w
or
k
s
.
S
e
c
ti
on
s
3
a
n
d
4
w
il
l
of
f
e
r
tr
a
n
s
f
e
r
l
e
a
r
ni
ng
a
nd
M
L
m
e
t
ho
d
s
.
S
e
c
t
io
n
5
wi
ll
d
e
s
c
r
i
be
r
e
s
e
a
r
c
h
m
e
t
ho
do
lo
gy.
S
e
c
ti
on
6
wi
ll
pr
e
s
e
nt
e
v
a
l
ua
ti
on
m
e
t
ho
d
s
.
F
in
a
l
ly,
s
e
c
ti
on
7
w
il
l
of
f
e
r
th
e
r
e
s
u
lt
s
a
n
d
d
i
s
c
u
s
s
i
on
s
,
f
ol
lo
w
e
d
by
s
e
c
ti
on
8,
w
hi
c
h
is
t
he
c
o
n
c
l
u
s
io
n.
2.
R
E
L
A
T
E
D
WO
RK
M
a
n
y
r
e
s
e
a
r
c
h
e
r
s
h
a
ve
f
oc
u
s
e
d
t
h
e
ir
e
f
f
or
t
s
on
u
s
in
g
M
L
to
c
l
a
s
s
i
f
y
na
tur
a
l
r
e
gio
n
s
or
S
R
S
I
i
ma
g
e
s
.
In
2
01
7
P
r
it
t
a
n
d
C
h
e
r
n
[
10
]
pr
o
po
s
e
d
de
e
p
l
e
a
r
n
in
g
s
y
s
t
e
m
u
s
in
g
c
on
vo
lu
ti
on
a
l
n
e
ur
a
l
ne
t
wor
k
s
(
C
NN
)
w
hi
c
h
c
l
a
s
s
if
i
e
d
obj
e
c
t
s
in
hi
gh
-
r
e
s
olu
ti
on
s
a
te
ll
i
t
e
im
a
ge
r
y
f
r
om
th
e
I
A
R
P
A
f
u
nc
ti
on
a
l
ma
p
of
t
h
e
w
or
ld
(
f
M
oW
)
d
a
t
a
s
e
t
in
to
63
c
l
a
s
s
e
s
wi
th
8
3%
a
c
c
ur
a
c
y.
It
i
nt
e
gr
a
te
d
im
a
ge
f
e
a
tu
r
e
s
a
n
d
m
e
t
a
da
ta,
a
c
hi
e
v
in
g
s
e
c
on
d
pl
a
c
e
in
t
he
f
M
oW
T
op
C
od
e
r
c
o
mp
e
t
it
i
on
a
n
d
th
e
s
y
s
te
m
a
c
hi
e
v
e
d
95%
or
h
ig
h
e
r
a
c
c
ur
a
c
y
in
15
c
la
s
s
e
s
a
n
d
pl
a
c
e
d
s
e
c
on
d
in
t
he
f
M
o
W
T
op
C
od
e
r
c
h
a
l
le
ng
e
wit
h
a
s
c
or
e
of
76
5,
6
63
.
In
2
01
8
B
u
s
c
om
be
a
nd
R
it
c
h
ie
[
1
1]
int
r
o
du
c
e
d
a
m
e
t
ho
d
f
or
e
f
f
i
c
i
e
ntl
y
tr
a
in
in
g
d
e
e
p
C
N
N
s
(
D
C
N
N
s
)
u
s
in
g
c
on
di
ti
on
a
l
r
a
n
do
m
f
i
e
l
d
s
(
C
R
F
s
)
wi
th
mi
nim
a
l
ma
nu
a
l
s
up
e
r
vi
s
i
on
to
c
l
a
s
s
i
f
y
n
a
t
ur
a
l
l
a
nd
s
c
a
p
e
s
.
It
d
e
m
on
s
tr
a
t
e
d
t
h
e
a
ppr
o
a
c
h'
s
e
f
f
e
c
ti
ve
n
e
s
s
in
l
a
nd
s
c
a
p
e
-
s
c
a
le
im
a
g
e
c
la
s
s
if
i
c
a
ti
on
a
n
d
h
ig
hl
ig
ht
s
its
p
ot
e
n
ti
a
l
f
or
a
c
c
ur
a
t
e
p
ix
e
l
-
l
e
v
e
l
c
l
a
s
s
if
ic
a
t
io
n
u
s
in
g
tr
a
ns
f
e
r
l
e
a
r
ni
ng,
a
nd
it
p
r
e
s
e
n
te
d
a
w
or
kf
l
ow
f
or
e
f
f
i
c
i
e
n
tl
y
c
r
e
a
t
in
g
l
a
be
le
d
i
m
a
g
e
r
y
a
nd
r
e
tr
a
i
nin
g
D
C
N
N
s
f
or
s
e
m
a
n
ti
c
c
l
a
s
s
if
ic
a
ti
on.
T
h
e
w
or
kf
l
ow,
u
s
in
g
M
ob
il
e
N
e
t
V
2,
a
c
hi
e
ve
d
hi
gh
c
la
s
s
if
i
c
a
ti
on
a
c
c
u
r
a
c
i
e
s
(
9
1
to
9
8%
)
a
c
r
o
s
s
v
a
r
i
ou
s
d
a
ta
s
e
t
s
.
In
2
02
0,
L
e
e
et
al
.
[
1
2]
u
s
e
d
d
e
e
p
le
a
r
nin
g
to
c
las
s
if
y
hu
ma
n
-
i
nd
u
c
e
d
d
e
f
or
e
s
t
a
t
io
n,
it
f
ou
nd
t
ha
t
U
-
N
e
t
o
ut
pe
r
f
or
me
d
S
e
gN
e
t
in
a
c
c
ur
a
c
y
(
74.
8%
v
s
.
63.
3%
)
,
p
a
r
t
ic
ul
a
r
l
y
in
di
s
ti
ng
ui
s
h
in
g
f
or
e
s
t
f
r
o
m
non
-
f
or
e
s
t
a
r
e
a
s
.
By
c
on
s
tr
u
c
ti
n
g
pr
e
c
i
s
e
tr
a
in
ing
da
ta
s
e
t
s
,
13
c
la
s
s
e
s
w
e
r
e
f
or
m
e
d
to
di
s
ti
ng
ui
s
h
f
or
e
s
t
a
n
d
n
on
-
f
or
e
s
t
a
r
e
a
s
.
T
he
s
t
ud
y
h
ig
hl
ig
ht
s
t
h
e
p
ot
e
n
ti
a
l
of
d
e
e
p
-
l
e
a
r
n
in
g
mo
de
l
s
in
a
n
a
ly
z
i
ng
d
e
f
or
e
s
t
a
t
io
n,
w
hi
le
a
c
kn
ow
le
dg
in
g
th
e
ne
e
d
f
or
mor
e
a
d
v
a
n
c
e
d
a
l
go
r
it
hm
s
a
nd
l
a
r
g
e
r
d
a
ta
s
e
t
s
f
or
i
mpr
ov
e
d
a
c
c
ur
a
c
y
a
nd
b
r
o
a
d
e
r
a
pp
li
c
a
ti
on.
Al
s
o,
i
n
20
20,
Ha
q
et
al
.
[
1
3]
d
e
m
on
s
tr
a
t
e
d
t
he
e
f
f
e
c
t
iv
e
n
e
s
s
of
d
e
e
p
le
a
r
ni
ng
-
ba
s
e
d
s
u
pe
r
vi
s
e
d
im
a
ge
c
la
s
s
if
i
c
a
t
io
n
u
s
in
g
u
nm
a
n
ne
d
a
e
r
i
a
l
ve
hi
c
l
e
(
UA
V
)
-
co
ll
e
c
t
e
d
d
a
ta
f
or
f
or
e
s
t
a
r
e
a
c
la
s
s
if
i
c
a
t
io
n.
It
hi
gh
li
ghte
d
th
e
s
i
gn
if
i
c
a
n
t
r
o
l
e
of
U
AV
s
a
nd
de
e
p
l
e
a
r
ni
ng
in
m
a
na
gi
ng
a
n
d
pl
a
nni
ng
f
or
e
s
t
a
r
e
a
s
t
hr
e
a
te
n
e
d
by
d
e
f
or
e
s
t
a
t
io
n.
T
h
e
r
e
s
ul
t
s
s
ho
w
e
d
t
h
a
t
an
a
c
c
u
r
a
c
y
w
a
s
9
3.
28%
a
nd
a
Ka
pp
a
c
o
e
f
f
i
c
i
e
nt
w
a
s
0.
8
98
8.
In
2
02
0
,
R
a
hm
a
n
et
al
.
[
14
]
e
v
a
l
u
a
t
e
d
t
he
p
e
r
f
or
m
a
nc
e
of
M
L
a
l
gor
it
h
m
s
(
r
a
n
do
m
f
or
e
s
t
,
s
up
po
r
t
v
e
c
t
or
m
a
c
h
in
e
(
S
V
M
)
,
a
n
d
s
t
a
c
k
e
d
a
l
gor
it
hm
s
)
on
c
la
s
s
if
yi
ng
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.
14,
No.
4,
Augus
t
2025
:
296
4
-
2978
2966
l
a
nd
u
s
e
a
n
d
l
a
n
d
c
ov
e
r
c
h
a
n
ge
s
u
s
i
ng
l
a
n
d
s
a
t
-
8,
s
e
nt
in
e
l
-
2,
a
n
d
pl
a
ne
t
im
a
g
e
s
in
r
ur
a
l
a
nd
u
r
b
a
n
a
r
e
a
s
.
T
he
s
e
n
ti
ne
l
-
2
im
a
ge
wi
th
S
V
M
pe
r
f
or
me
d
b
e
s
t,
a
c
hi
e
v
in
g
h
ig
h
a
c
c
u
r
a
c
y,
a
id
in
g
in
mo
ni
tor
in
g
f
r
a
g
m
e
nt
e
d
l
a
nd
s
c
a
p
e
s
in
B
a
ng
la
d
e
s
h
a
n
d
b
e
y
on
d
a
n
d
f
o
und
th
a
t
it
s
s
e
nti
n
e
l
-
2
im
a
ge
r
y
out
p
e
r
f
o
r
m
s
l
a
nd
s
a
t
-
8
a
n
d
p
l
a
n
e
t
in
a
c
c
u
r
a
c
y,
wi
th
t
h
e
S
V
M
a
c
hi
e
vi
ng
t
he
h
ig
he
s
t
r
e
s
ul
t
s
w
he
n
u
s
e
d
wi
th
s
e
n
ti
n
e
l
-
2
d
a
ta.
In
bo
th
B
h
ol
a
(
r
ur
a
l
)
a
n
d
Dh
a
ka
(
ur
b
a
n
)
,
S
V
M
pr
ovi
d
e
d
t
h
e
hig
h
e
s
t
o
ve
r
a
l
l
a
c
c
ur
a
c
y
(
0.
96
9
a
n
d
0.
9
83
)
a
nd
Ka
pp
a
v
a
lu
e
s
(
0.
9
48
a
nd
0.
9
68)
.
In
20
22,
G
e
v
a
e
r
t
a
nd
B
e
lg
iu
[
15
]
pr
op
o
s
e
d
la
nd
s
c
a
p
e
me
tr
i
c
s
to
a
s
s
e
s
s
t
he
s
i
mi
lar
it
y
b
e
t
w
e
e
n
t
r
a
ini
ng
a
nd
te
s
ti
ng
im
a
g
e
s
f
or
b
ui
ld
in
g
i
de
nt
if
i
c
a
ti
on
wi
th
f
u
ll
y
c
on
vol
ut
io
na
l
n
e
tw
or
k
s
(
F
C
N
s
)
.
T
h
e
m
od
e
l
tr
a
i
ne
d
on
D
a
r
e
s
S
a
l
a
a
m
im
a
g
e
s
s
ho
we
d
t
h
e
h
ig
he
s
t
g
e
ne
r
a
l
i
z
a
ti
on,
w
h
il
e
t
he
Z
a
n
z
i
b
a
r
-
tr
a
i
n
e
d
mo
de
l
h
a
d
t
he
lo
w
e
s
t.
T
he
c
la
s
s
i
f
i
c
a
ti
on
a
c
c
ur
a
c
ie
s
a
r
e
lo
we
r
t
ha
n
t
ho
s
e
in
th
e
o
pe
n
c
i
ti
e
s
AI
c
ha
ll
e
n
g
e
du
e
to
li
mi
te
d
tr
a
i
ni
ng
da
ta
f
or
e
v
a
lu
a
ti
n
g
g
e
n
e
r
a
l
i
z
a
bil
it
y.
T
h
is
s
tu
dy
f
oc
u
s
e
s
on
i
d
e
n
ti
f
yi
ng
i
ma
g
e
s
im
i
la
r
it
y
me
tr
i
c
s
th
a
t
p
r
e
di
c
t
m
od
e
l
p
e
r
f
or
m
a
n
c
e
r
a
th
e
r
th
a
n
a
c
hi
e
v
in
g
ma
xi
mu
m
a
c
c
ur
a
c
y.
In
2
0
23,
C
h
a
u
dh
a
r
i
et
al
.
[
1
6]
e
xp
lo
r
e
d
dr
ou
gh
t
pr
e
d
ic
ti
on
u
s
in
g
s
a
t
e
ll
i
te
im
a
g
e
s
a
nd
d
e
e
p
l
e
a
r
nin
g
m
od
e
l
s
.
T
he
y
c
om
p
a
r
e
d
E
f
f
ic
ie
nt
N
e
t
wit
h
o
th
e
r
C
NN
va
r
i
a
n
t
s
l
ik
e
A
le
x
Ne
t
a
nd
vi
s
ua
l
g
e
om
e
tr
y
gr
ou
p
ne
tw
or
k
(
VG
G
Ne
t
)
.
It
f
ou
nd
t
h
a
t
E
f
f
ic
ie
nt
N
e
t
o
ut
pe
r
f
or
m
s
th
e
s
e
mo
de
l
s
w
hi
c
h
a
c
hi
e
ve
hi
gh
e
r
a
c
c
ur
a
c
y
in
b
in
a
r
y
d
r
o
ug
ht
c
l
a
s
s
i
f
i
c
a
ti
on,
a
nd
f
o
un
d
t
h
a
t
v
a
r
i
a
n
t
s
of
C
N
N
a
r
e
c
o
mm
on
ly
u
s
e
d
in
i
ma
ge
pr
o
c
e
s
s
in
g.
T
h
i
s
s
tu
dy
e
v
a
l
ua
t
e
d
th
e
ir
e
f
f
e
c
ti
v
e
n
e
s
s
f
or
dr
ou
gh
t
c
l
a
s
s
if
ic
a
ti
on
u
s
in
g
s
a
t
e
l
li
t
e
im
a
ge
s
f
r
o
m
Ko
l
a
r
,
K
a
r
n
a
ta
k
a
a
nd
E
f
f
ic
i
e
nt
N
e
t
ou
tp
e
r
f
or
m
s
tr
a
d
it
io
na
l
C
N
N
mo
de
l
s
li
ke
C
N
N,
Al
e
x
N
e
t,
a
nd
V
GG
N
e
t,
a
c
h
ie
vi
ng
h
ig
he
r
a
c
c
ur
a
c
ie
s
of
0.
9
1
to
0.
94.
D
e
s
p
it
e
C
N
N
'
s
s
u
pe
r
io
r
p
e
r
f
or
ma
nc
e
wi
th
an
a
c
c
ur
a
c
y
of
0.
9
7,
a
ll
m
ode
l
s
n
e
e
d
e
x
te
nd
e
d
tr
a
i
nin
g
p
e
r
io
d
s
.
3.
T
RA
NSF
E
R
L
E
AR
NI
NG
3.
1.
Xc
e
p
t
ion
t
r
an
s
f
e
r
lear
n
in
g
T
he
Xc
e
pti
on
model
is
a
pr
e
-
tr
a
ini
ng
model
on
the
I
mage
Ne
t
da
tas
e
t.
T
he
model
wa
s
r
e
c
e
ntl
y
de
s
igned
as
an
e
xtens
ion
of
the
I
nc
e
pti
onv3
model
.
It
wa
s
invente
d
by
C
holl
e
t
[
17
]
.
It
is
mor
e
r
obus
t
a
nd
ha
s
les
s
ove
r
f
it
ti
ng
dif
f
iculti
e
s
than
c
ur
r
e
nt
popular
pr
e
-
tr
a
ini
ng
models
li
ke
VG
G16
[
18
]
.
Xc
e
pti
on
model
is
ba
s
e
d
on
the
pr
inciple
of
de
p
th
-
wis
e
s
e
p
a
r
a
ble
c
onvolut
ion
ins
tea
d
of
c
las
s
ic
c
onvolut
ion.
T
he
de
pth
-
wis
e
s
e
pa
r
a
ble
c
onvolut
i
on
pa
s
s
e
s
thr
ough
two
s
tage
s
that
a
r
e
a
ppli
e
d
in
r
e
ve
r
s
e
manne
r
.
F
i
r
s
t
s
tage
c
a
ll
e
d
de
pth
wis
e
c
onvolut
ion
whic
h
doe
s
not
a
pply
a
c
onvolut
ional
f
i
lt
e
r
to
a
ll
c
ha
nne
ls
at
the
s
a
me
ti
me,
but
r
a
ther
a
ppli
e
s
it
to
each
input
c
ha
nne
l
f
or
r
e
duc
ing
c
omput
a
ti
ons
a
nd
memor
y
s
pa
c
e
us
e
d.
S
e
c
ond
s
tage
c
a
ll
e
d
P
oint
wis
e
c
onvolut
ion
w
hich
int
e
gr
a
te
the
f
ir
s
t
s
tage
de
pth
wis
e
c
onvolut
io
n
output
ove
r
a
ll
c
ha
nne
ls
by
us
ing
a
1
×
1
c
onvolut
ion
[
19]
.
Ac
c
or
ding
to
F
igu
r
e
1,
the
Xc
e
pti
on
a
r
c
hit
e
c
tur
e
c
ons
is
ts
of
thr
e
e
pr
i
mar
y
pa
r
ts
.
T
he
f
i
r
s
t
c
a
ll
e
d
e
ntr
y
f
low,
whic
h
is
whe
r
e
the
da
ta
is
f
ir
s
t
p
r
oc
e
s
s
e
d.
T
he
da
ta
is
s
ubs
e
que
ntl
y
s
e
nt
via
the
mi
d
dle
f
low
,
whic
h
is
r
e
pe
a
ted
e
ight
ti
mes
,
a
nd
las
tl
y
th
r
ough
th
e
e
xit
f
low
.
F
igur
e
1.
X
c
e
pti
on
a
r
c
hit
e
c
tur
e
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
E
nhanc
ing
tr
adit
ional
mac
hine
lear
ning
me
thods
us
ing
c
onc
atenation
…
(
R
afal
N
az
ar
Y
oune
s
A
l
-
T
a
han
)
2967
F
igur
e
1
s
howe
d
that
thr
e
e
blocks
c
ons
is
t
of
c
onvo
lut
ion
laye
r
s
with
a
number
of
r
e
c
ti
f
ied
li
ne
a
r
unit
(
R
e
L
U
)
a
nd
max
-
pooli
ng
laye
r
s
be
twe
e
n
them.
T
he
e
ntr
y
f
low
block
c
ons
is
ts
of
e
ight
c
onvolut
i
on
laye
r
s
,
while
the
mi
ddle
f
low
block
c
ons
is
ts
of
24,
a
nd
f
in
a
ll
y
the
e
xit
f
low
block
c
ons
is
ts
of
f
our
c
onvolut
io
n
laye
r
s
.
Af
ter
that
,
the
g
lobal
a
ve
r
a
ge
pooli
ng
laye
r
f
oll
ow
e
d
the
c
onvolut
ion
laye
r
s
to
c
onve
r
t
to
the
f
ull
y
c
onne
c
ted
laye
r
whic
h
r
e
duc
e
d
the
number
of
pa
r
a
mete
r
s
.
3.
2.
De
n
s
e
Ne
t
201
t
r
an
s
f
e
r
lear
n
in
g
It
is
a
t
r
a
ns
f
e
r
lea
r
ning
model
tr
a
ined
on
the
I
m
a
ge
Ne
t
da
tas
e
t
a
nd
buil
t
on
the
C
NN
pr
inciples
.
It
wa
s
pr
opos
e
d
by
Hua
ng
et
al
.
[
20]
a
nd
wa
s
uti
li
z
e
d
in
va
r
ious
major
f
ields
of
a
r
t
if
icia
l
int
e
l
li
ge
nc
e
,
including
objec
t
de
tec
ti
on
a
nd
c
las
s
if
ica
ti
on,
due
to
its
c
a
pa
c
it
y
to
r
e
us
e
f
e
a
tur
e
s
a
nd
r
e
duc
e
the
pr
o
blem
of
va
nis
hing
gr
a
dients
,
as
we
ll
as
its
us
a
g
e
of
a
li
m
it
e
d
number
of
f
e
a
tur
e
s
.
De
ns
e
Ne
t201
r
e
li
e
s
on
a
s
im
ple
s
tr
a
tegy,
whic
h
is
to
c
onne
c
t
a
ll
laye
r
s
in
a
f
e
e
d
-
f
or
wa
r
d
wa
y
so
that
each
laye
r
is
f
e
d
f
r
om
a
ll
p
r
e
vio
us
laye
r
s
a
nd
a
ls
o
pa
s
s
e
s
its
f
e
a
tur
e
maps
to
s
ubs
e
que
nt
lay
e
r
s
[
21]
.
De
ns
e
Ne
t201's
ke
y
c
omponents
a
r
e
de
ns
e
blocks
a
nd
tr
a
ns
it
ion
laye
r
s
(
s
e
e
F
igu
r
e
2)
.
F
igur
e
2.
De
ns
e
Ne
t201
a
r
c
hit
e
c
tur
e
T
he
f
unda
menta
l
f
e
a
tur
e
of
the
model
ne
twor
k
is
de
ns
e
block
s
,
whic
h
a
r
e
made
up
of
s
e
ve
r
a
l
bott
lene
c
k
laye
r
s
.
I
n
f
or
mation
f
r
om
each
laye
r
is
c
onne
c
ted
via
the
de
ns
e
c
onne
c
ti
on
mode
ins
ide
t
he
de
ns
e
block,
gua
r
a
ntee
ing
that
the
output
s
ize
r
e
mains
c
ons
is
tent
thr
oughout.
De
ns
e
Ne
t
c
ontr
ols
the
a
mount
of
c
ha
nne
ls
us
ing
bott
lene
c
k
laye
r
s
,
tr
a
ns
it
ion
laye
r
s
,
a
nd
a
gr
owth
r
a
te
[
22
]
.
4.
M
AC
HI
NE
L
E
AR
NI
NG
4.
1.
Naïve
B
aye
s
algorit
h
m
It
is
n
a
me
d
NB
be
c
a
u
s
e
t
h
e
c
om
pu
ta
ti
on
s
of
t
he
p
r
o
ba
bi
li
t
y
f
or
e
a
c
h
c
la
s
s
a
r
e
r
e
d
uc
e
d
to
m
a
k
e
its
c
o
mp
ut
a
ti
on
tr
a
c
ta
bl
e
.
It
is
f
a
m
ou
s
in
m
ult
icl
a
s
s
i
f
i
c
a
ti
on
d
om
a
i
n.
NB
c
l
a
s
s
if
i
e
r
s
r
e
l
y
on
B
a
y
e
s
i
a
n
a
l
gor
it
hm
s
[
2
3]
.
T
h
e
s
e
a
r
e
ba
s
e
d
on
B
a
y
e
s
'
th
e
or
e
m,
an
e
qu
a
t
io
n
th
a
t
de
s
c
r
ib
e
s
th
e
r
e
la
ti
on
s
hi
p
b
e
tw
e
e
n
t
h
e
c
o
ndi
ti
on
a
l
pr
ob
a
b
il
it
i
e
s
of
s
t
a
ti
s
t
ic
a
l
d
a
t
a
.
In
B
a
y
e
s
i
a
n
c
l
a
s
s
if
i
c
a
t
io
n,
we
w
a
n
t
to
k
no
w
th
e
pr
o
b
a
b
il
i
ty
of
a
l
a
b
e
l
gi
v
e
n
s
om
e
o
b
s
e
r
v
a
b
le
c
h
a
r
a
c
t
e
r
i
s
ti
c
s
[
2
4]
.
In
oth
e
r
w
or
d
s
,
it
e
x
pl
a
i
n
s
t
h
e
l
ik
e
l
ih
oo
d
of
an
e
ve
nt
o
c
c
u
r
r
i
ng
giv
e
n
on
p
a
s
t
kn
ow
le
dg
e
of
t
he
o
c
c
ur
r
e
n
c
e
of
a
no
th
e
r
e
v
e
n
t.
To
ma
k
e
t
he
f
or
e
c
a
s
t,
c
o
mp
ut
e
P
(
A
|B
)
,
w
hi
c
h
is
t
h
e
li
ke
li
ho
od
of
A
o
c
c
u
r
r
i
ng
if
B
is
tr
ue
.
F
ur
th
e
r
m
or
e
,
P
(
B
|
A)
r
e
pr
e
s
e
n
t
s
th
e
l
ik
e
li
h
oo
d
of
B
oc
c
ur
r
i
ng
if
A
is
tr
u
e
.
P
(
B
)
a
n
d
P
(
A)
a
r
e
th
e
p
r
o
ba
bi
li
t
y
of
s
e
e
in
g
o
ne
wi
th
ou
t
th
e
o
th
e
r
,
as
il
lu
s
tr
a
t
e
d
in
(
1)
[
2
5]
.
(
|
)
=
(
|
)
(
)
(
)
(
1)
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.
14,
No.
4,
Augus
t
2025
:
296
4
-
2978
2968
4.
2.
De
c
is
ion
t
r
e
e
algori
t
h
m
In
p
ub
li
c
li
f
e
,
wh
il
e
c
o
n
s
id
e
r
in
g
a
s
p
e
c
if
ic
t
opi
c
a
nd
ma
ki
ng
a
d
e
c
i
s
i
on,
on
e
mu
s
t
c
a
r
e
f
ull
y
c
o
n
s
i
de
r
a
ll
of
t
h
e
a
dv
a
nt
a
g
e
s
a
n
d
d
ow
n
s
i
de
s
of
t
h
e
o
pt
io
n
or
th
e
a
l
te
r
n
a
ti
v
e
s
a
v
a
il
a
bl
e
.
S
im
i
la
r
ly,
in
ML
,
DT
do
t
h
e
s
a
m
e
f
u
n
c
ti
on
wit
h
m
or
e
pr
e
c
i
s
io
n,
t
a
k
in
g
in
to
a
c
c
o
un
t
a
ll
r
e
l
e
v
a
n
t
v
a
r
i
a
b
le
s
to
m
a
ke
t
h
e
o
pt
im
a
l
d
e
c
i
s
i
on
[
2
6]
.
DT
a
r
e
a
po
we
r
f
u
l
t
oo
l
ut
il
iz
e
d
in
a
va
r
ie
ty
of
d
om
a
i
n
s
,
i
n
c
lu
di
ng
c
la
s
s
if
i
c
a
t
io
n,
i
ma
g
e
pr
o
c
e
s
s
in
g,
a
n
d
p
a
t
ter
n
r
e
c
o
gn
it
i
on
[
2
7]
.
It
m
a
y
be
u
s
e
d
as
s
ta
ti
s
ti
c
a
l
p
r
o
c
e
s
s
e
s
to
di
s
c
o
ve
r
d
a
ta,
e
x
tr
a
c
t
te
xt,
i
de
nt
if
y
m
is
s
in
g
d
a
t
a
in
a
c
l
a
s
s
,
e
nha
n
c
e
s
e
a
r
c
h
e
ng
in
e
s
,
a
n
d
h
a
s
a
v
a
r
i
e
t
y
of
m
e
d
ic
in
a
l
u
s
e
s
[
28]
.
It
c
o
n
s
i
s
t
s
of
r
o
ot
no
d
e
s
(
t
op
n
od
e
s
)
,
br
a
n
c
he
s
(
l
in
k
s
)
,
a
n
d
l
e
a
f
no
de
s
.
In
a
DT
,
te
s
ti
ng
ta
ke
s
p
l
a
c
e
on
t
h
e
in
ter
ior
no
de
s
,
a
n
d
th
e
ou
tp
ut
is
p
e
r
f
or
m
e
d
on
t
he
l
e
a
f
no
d
e
s
.
E
a
c
h
n
od
e
r
e
pr
e
s
e
nt
s
a
f
e
a
t
ur
e
,
e
a
c
h
b
r
a
n
c
h
is
a
c
c
o
un
ta
bl
e
f
or
t
h
e
de
c
i
s
io
n,
a
nd
e
a
c
h
l
e
a
f
di
s
pl
a
y
s
t
he
r
e
s
u
lt
.
T
h
e
w
a
y
DT
wor
k
is
tha
t,
e
a
c
h
i
nt
e
r
na
l
no
de
div
id
e
s
th
e
d
a
t
a
s
e
t
i
nto
s
u
b
s
e
t
s
d
e
pe
n
di
ng
on
a
f
e
a
t
ur
e
c
r
it
e
r
i
on.
T
h
e
o
bj
e
c
ti
ve
is
to
m
a
ke
t
he
s
u
b
s
e
t
s
as
pu
r
e
as
p
o
s
s
i
bl
e
,
w
hi
c
h
me
a
n
s
th
e
y
s
h
oul
d
o
nl
y
i
nc
lu
de
d
a
t
a
p
oin
t
s
f
r
om
th
e
s
a
m
e
c
a
te
go
r
i
z
a
ti
on
c
la
s
s
.
T
h
e
m
o
s
t
c
r
it
e
r
i
on
f
u
nc
ti
on
s
in
DT
u
s
e
d
a
r
e
:
G
ini
i
nd
e
x
a
nd
e
ntr
op
y
m
e
a
s
ur
e
s
.
W
h
e
n
an
e
l
e
m
e
nt
is
r
a
n
do
ml
y
c
l
a
s
s
if
ie
d
a
c
c
or
di
ng
to
th
e
d
is
tr
i
bu
ti
on
of
l
a
be
le
d
in
t
he
s
e
t,
th
e
G
ini
i
nd
e
x
m
e
a
s
ur
e
s
t
h
e
pr
ob
a
b
il
it
y
of
in
c
or
r
e
c
t
ly
c
l
a
s
s
if
yi
ng
th
a
t
e
le
me
nt.
4.
3.
K
-
n
e
ar
e
s
t
n
e
igh
b
or
s
algorit
h
m
KNN
is
a
wide
ly
us
e
d
s
upe
r
vis
e
d
ML
a
lgor
it
hm,
pa
r
ti
c
ula
r
ly
e
f
f
e
c
ti
ve
f
or
c
las
s
if
ica
ti
on
a
nd
r
e
gr
e
s
s
ion
tas
ks
.
T
he
f
unda
menta
l
pr
inciple
be
hin
d
KNN
is
that
s
im
il
a
r
ins
tanc
e
s
a
r
e
li
ke
ly
to
e
xis
t
c
los
e
to
each
other
withi
n
the
f
e
a
tur
e
s
pa
c
e
,
a
ll
owing
f
or
t
he
c
a
tegor
iza
ti
on
of
ne
w
s
a
mpl
e
s
ba
s
e
d
on
their
pr
oxim
it
y
to
a
lr
e
a
dy
c
las
s
if
ied
da
ta
point
s
(
ne
ighbor
s
)
[
29
]
.
T
he
mos
t
c
omm
on
dis
tanc
e
met
r
ics
us
e
d
in
KNN
include
E
uc
li
de
a
n
dis
tanc
e
,
M
a
nha
tt
a
n
dis
tanc
e
,
M
inkows
ki
dis
tanc
e
,
c
os
ine
s
im
il
a
r
it
y,
a
nd
c
or
r
e
lati
on
[
29
]
.
Among
thes
e
,
E
uc
li
de
a
n
dis
tanc
e
is
pa
r
ti
c
ular
ly
we
ll
-
known
a
nd
is
mathe
matica
ll
y
de
f
ined
as
the
s
tr
a
ight
-
li
ne
dis
tanc
e
be
twe
e
n
two
point
s
in
a
mul
ti
dim
e
ns
i
ona
l
s
pa
c
e.
4.
4.
L
igh
t
gr
ad
ient
b
oos
t
in
g
m
ac
h
in
e
Gr
a
dient
boos
ti
ng
mac
hines
(
GB
M
s
)
a
r
e
a
type
of
e
ns
e
mbl
e
lea
r
ning
method
that
c
ons
tr
uc
t
an
a
ddit
ive
model
f
r
om
s
im
ple
DT
.
T
he
s
e
tr
e
e
s
,
whic
h
a
r
e
not
highl
y
op
ti
mi
z
e
d
indi
vidually
,
a
r
e
then
c
ombi
ne
d
by
opti
mi
z
ing
a
los
s
f
unc
ti
on
,
lea
ding
to
s
tr
ong
e
r
pr
e
dictive
pe
r
f
or
manc
e
[
30]
.
L
ight
GB
M
of
f
e
r
s
f
a
s
ter
tr
a
ini
ng
s
pe
e
ds
a
nd
g
r
e
a
ter
e
f
f
icie
nc
y
than
many
other
a
lgor
it
h
ms
.
T
his
is
pr
im
a
r
il
y
due
to
its
hi
s
togr
a
m
-
ba
s
e
d
a
ppr
oa
c
h,
whic
h
buc
ke
ts
c
onti
nuous
f
e
a
tur
e
va
lues
int
o
dis
c
r
e
te
bins
,
the
r
e
by
a
c
c
e
ler
a
ti
ng
the
tr
a
ini
ng
pr
oc
e
s
s
.
L
ight
GB
M
u
s
e
s
a
lea
f
-
wi
s
e
a
lgor
it
hm
to
gr
ow
tr
e
e
s
ve
r
ti
c
a
ll
y,
s
e
lec
ti
ng
the
lea
f
that
mos
t
r
e
duc
e
s
the
los
s
f
or
s
pli
tt
ing.
To
opti
mi
z
e
t
r
a
ini
ng,
L
ight
GB
M
e
mpl
oys
a
t
e
c
hnique
c
a
ll
e
d
gr
a
dient
-
ba
s
e
d
one
-
s
id
e
s
a
mpl
ing
(
GO
S
S
)
,
whic
h
f
oc
us
e
s
on
da
ta
ins
tanc
e
s
with
lar
ge
r
g
r
a
dients
,
a
s
s
umi
ng
that
ins
tan
c
e
s
with
s
maller
gr
a
dients
a
r
e
a
lr
e
a
dy
we
ll
-
tr
a
ined
a
nd
can
be
ignor
e
d.
5.
RE
S
E
AR
CH
M
E
T
HO
DOL
OG
Y
T
he
pr
opos
e
d
methodology
c
ons
is
ted
of
many
s
teps
:
pr
e
-
pr
oc
e
s
s
ing
a
nd
f
e
a
tur
e
,
f
oll
owe
d
by
c
las
s
if
ica
ti
on
pr
oc
e
s
s
us
ing
both
tr
a
dit
ional
a
nd
s
tate
of
a
r
t
ML
tec
hniques
li
ke
:
DT
,
NB
,
KNN,
L
G
B
M
,
a
nd
e
ns
e
mbl
e
voti
ng.
F
igur
e
3
s
hows
the
wor
kf
low
of
the
pr
opos
e
d
methodology
whic
h
a
ppli
e
d
on
W
in
dows
10
a
nd
4
-
c
or
e
C
P
U
with
a
p
r
oc
e
s
s
ing
s
pe
e
d
of
2.
00
G
Hz
.
M
e
mor
y
c
a
pa
c
it
y
of
16.
0
GB
.
5.
1.
De
s
c
r
ip
t
ion
of
d
a
t
as
e
t
T
he
pr
opos
e
d
models
we
r
e
tes
ted
on
two
S
R
S
I
da
tas
e
ts
f
or
ge
ne
r
a
li
z
a
ti
on
pur
pos
e
s
.
T
he
f
ir
s
t
da
tas
e
t
wa
s
take
n
f
r
om
the
Ka
ggle
we
bs
it
e
a
nd
c
ons
i
s
ts
of
(4
,
131)
im
a
ge
s
c
las
s
if
ied
int
o
only
thr
e
e
c
las
s
e
s
(
de
s
e
r
t,
gr
e
e
n_a
r
e
a
,
a
nd
wa
ter
)
.
T
he
s
e
c
ond
da
tas
e
t
is
s
im
il
a
r
to
the
f
ir
s
t
da
tas
e
t,
but
it
include
s
e
xtr
a
da
ta
f
r
om
s
e
ve
r
a
l
s
our
c
e
s
s
uc
h
as
the
Ka
ggle
we
bs
it
e
[
31]
,
NA
S
A
[
32]
a
nd
Nimbo
[
33]
to
ba
lanc
e
th
r
e
e
c
las
s
e
s
.
T
he
tot
a
l
number
of
im
a
g
es
wa
s
(6
,
900)
im
a
ge
s
o
ve
r
thr
e
e
c
las
s
e
s
.
T
he
da
ta
wa
s
s
pli
t
int
o
80%
tr
a
ini
ng
a
nd
20%
tes
ti
ng.
5.
2.
I
m
age
p
r
e
-
p
r
oc
e
s
s
in
g
P
r
e
pr
oc
e
s
s
ing
is
an
e
s
s
e
nti
a
l
s
tep
f
or
f
indi
ng
r
e
lev
a
nt
f
e
a
tur
e
s
in
S
R
S
I
a
nd
e
ns
ur
ing
that
the
da
ta
is
r
e
a
dy
f
or
c
e
r
tain
kind
of
a
na
lys
is
.
F
igur
e
4
de
picts
many
pr
oc
e
dur
e
s
that
we
r
e
pe
r
f
or
med
to
dig
it
a
l
im
a
ge
s
:
‒
R
e
s
i
z
e
im
a
g
e
s
:
I
t
is
a
v
it
a
l
s
t
e
p
in
e
n
s
ur
in
g
th
a
t
a
l
l
i
m
a
g
e
s
a
r
e
u
nif
or
m
a
nd
of
e
q
u
a
l
s
i
z
e
.
F
ur
t
h
e
r
m
or
e
,
lo
we
r
in
g
t
he
nu
mb
e
r
of
pix
e
l
s
in
im
a
ge
s
wil
l
mi
ni
mi
z
e
th
e
n
um
be
r
of
pr
oc
e
s
s
or
s
a
n
d
m
e
m
or
y
r
e
qu
ir
e
d.
In
th
i
s
w
or
k,
th
e
r
e
s
i
z
e
f
un
c
ti
on
in
P
yt
ho
n
w
a
s
u
s
e
d
to
un
if
or
ml
y
s
c
a
l
e
t
h
e
im
a
g
e
s
to
1
50
wi
dt
h
*
15
0
h
e
i
gh
t
s
.
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
E
nhanc
ing
tr
adit
ional
mac
hine
lear
ning
me
thods
us
ing
c
onc
atenation
…
(
R
afal
N
az
ar
Y
oune
s
A
l
-
T
a
han
)
2969
‒
I
mage
tr
a
ns
f
or
m:
Ope
nC
V
is
a
popular
P
ython
li
br
a
r
y
f
or
digi
tal
im
a
ge
pr
oc
e
s
s
ing.
T
his
li
br
a
r
y
include
s
(
c
vtC
olor
f
unc
ti
on)
,
whic
h
c
onve
r
ts
im
a
ge
s
f
r
om
B
GR
c
olor
s
pa
c
e
to
R
GB
f
or
c
lar
it
y
a
nd
s
im
ple
dis
play
us
ing
the
matplot
li
b
li
br
a
r
y
.
‒
C
a
nny
de
tec
ti
on:
T
he
c
r
it
ica
l
e
dge
s
of
the
S
R
S
I
we
r
e
highl
igh
ted
a
nd
p
r
e
c
is
e
ly
a
na
lyze
d
us
ing
the
c
a
nny
e
dge
a
ppr
oa
c
h
[
34]
.
‒
B
ounding
box:
I
t
is
an
e
s
s
e
nti
a
l
a
nnotation
a
ppr
oa
c
h
f
or
digi
tal
im
a
ge
s
.
An
a
bs
tr
a
c
t
r
e
c
tangle
s
e
r
ve
s
as
both
an
it
e
m
dis
c
ove
r
y
tool
a
nd
a
r
e
f
e
r
e
nc
e
point
f
or
im
a
ge
s
.
‒
C
r
opping
im
a
ge
:
T
he
tec
hnique
of
e
li
mi
na
ti
ng
unn
e
c
e
s
s
a
r
y
white
r
e
gions
a
nd
e
dge
s
f
r
om
S
R
S
I
in
or
de
r
to
identif
y
the
e
dge
s
with
the
mos
t
r
e
leva
nt
e
leme
n
ts
.
‒
Nor
maliza
ti
on:
I
t
is
a
c
omm
on
im
a
ge
pr
oc
e
s
s
ing
t
e
c
hnique
that
c
ha
nge
s
the
int
e
ns
it
y
r
a
nge
of
pixels
to
be
twe
e
n
0
a
nd
1.
It
is
a
c
omm
on
f
unc
ti
on
to
c
onve
r
t
an
input
im
a
ge
int
o
a
r
a
nge
of
pixel
va
lues
that
a
r
e
mor
e
plea
s
a
nt
to
the
hu
man
e
ye
.
F
igur
e
3.
W
or
kf
low
of
pr
opos
e
d
methodology
F
igur
e
4.
P
r
e
-
pr
oc
e
s
s
ing
im
a
ge
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.
14,
No.
4,
Augus
t
2025
:
296
4
-
2978
2970
5.
3.
F
e
a
t
u
r
e
e
xt
r
ac
t
ion
P
r
e
-
tr
a
ined
models
(
Xc
e
pti
on
a
nd
De
ne
s
e
Ne
t201)
we
r
e
us
e
d
to
e
xtr
a
c
t
f
e
a
tur
e
s
f
r
om
S
R
S
I
da
tas
e
ts
by
pa
s
s
ing
them
thr
ough
numer
ous
c
onvolut
ional
laye
r
s
of
two
p
r
e
-
tr
a
ined
models
,
r
e
s
ult
ing
in
2D
matr
ice
s
with
f
r
oz
e
n
we
ight
s
a
nd
r
e
movi
ng
the
output
laye
r
.
T
he
c
ha
r
a
c
ter
is
ti
c
s
of
Xc
e
pti
on
tr
a
ns
f
e
r
lea
r
ning
li
ke
de
pth
wis
e
c
onvolut
ion
a
nd
point
wis
e
c
onvolut
io
n
we
r
e
a
ppli
e
d
to
de
c
r
e
a
s
e
c
omput
ing
p
r
oc
e
s
s
e
s
in
f
e
a
tur
e
e
xtr
a
c
ti
on.
Als
o,
De
ne
s
e
Ne
t201
c
onne
c
ts
a
ll
lay
e
r
s
s
uc
h
that
each
laye
r
r
e
c
e
ives
input
f
r
om
a
ll
pr
e
vious
laye
r
s
while
a
ls
o
pa
s
s
ing
on
f
e
a
tur
e
maps
to
s
ubs
e
que
nt
laye
r
s
.
T
he
r
e
f
or
e
,
it
r
e
duc
e
s
the
pr
oblem
of
va
nis
hing
gr
a
dients
.
T
he
c
onc
a
tena
ti
on
pr
oc
e
s
s
wi
l
l
c
ombi
ne
the
output
s
of
the
two
models
in
o
r
de
r
to
t
r
a
in
thes
e
matr
ice
s
f
or
c
las
s
if
ica
ti
on
us
ing
tr
a
dit
ional
a
nd
s
tate
of
a
r
t
M
L
a
ppr
oa
c
he
s
.
S
o,
c
onc
a
tena
ti
o
n
pr
oc
e
s
s
a
c
hieve
d
mul
ti
model
lea
r
ning
a
nd
high
-
qua
li
ty
r
e
pr
e
s
e
ntation.
F
igu
r
e
5
s
hows
f
e
a
tur
e
s
e
xt
r
a
c
ted
by
two
tr
a
ns
f
e
r
lea
r
ning.
F
igur
e
5.
F
e
a
tur
e
e
xtr
a
c
ti
on
a
f
te
r
c
onc
a
tena
te
two
t
r
a
ns
f
e
r
lea
r
ning
5.
4.
Clas
s
if
icat
ion
p
r
oc
e
s
s
In
thi
s
wor
k
,
t
r
a
dit
ional
a
nd
s
tate
of
a
r
t
ML
methods
(
DT
,
NB
,
KNN
a
nd
L
GB
M
)
we
r
e
us
e
d
to
c
las
s
if
y
S
R
S
I
da
tas
e
t
a
nd
a
c
hieve
be
s
t
r
e
s
ult
s
.
F
ur
ther
mor
e
,
an
e
ns
e
mbl
e
voti
ng
tec
hnique
wa
s
a
ppli
e
d
a
mong
thr
e
e
tr
a
dit
ional
ML
(
DT
,
NB
,
KNN)
to
in
c
r
e
a
s
e
the
pe
r
f
or
manc
e
of
model
.
T
he
s
e
f
our
ML
methods
a
r
e
mor
e
s
e
ns
it
ive
to
r
e
leva
nt
da
ta,
can
s
c
a
le
with
lar
ge
da
ta,
is
non
-
pa
r
a
metr
ic,
mea
ning
it
can
a
da
pt
to
the
da
ta
a
nd
doe
s
not
a
s
s
ume
a
f
ixed
model
f
o
r
m,
c
a
n
int
e
r
pr
e
t
a
nd
e
xtr
a
c
t
the
im
po
r
tanc
e
of
a
f
e
a
tur
e
,
a
nd
can
ha
ndle
both
c
a
tegor
ica
l
a
nd
numer
ica
l
da
ta.
On
the
other
ha
nd,
the
hype
r
pa
r
a
mete
r
s
of
metho
ds
we
r
e
opti
mi
z
e
d
by
us
ing
gr
id
s
e
a
r
c
h
a
lgor
it
hm
be
c
a
us
e
it
include
s
c
r
os
s
-
v
a
li
da
ti
on,
whic
h
divi
de
s
the
da
ta
int
o
k
-
f
olds
f
or
tr
a
ini
ng
a
nd
other
s
f
or
e
va
luation,
a
nd
t
he
n
r
e
pe
a
ts
the
pr
oc
e
s
s
to
e
ns
ur
e
that
each
pa
r
t
of
the
da
ta
is
us
e
d
at
lea
s
t
onc
e
in
tr
a
ini
ng
a
nd
e
va
luation
.
In
thi
s
s
tudy,
va
lue
of
k
-
f
olds
wa
s
10.
T
a
ble
1
dis
plays
the
im
por
tant
hype
r
pa
r
a
mete
r
va
lues
.
T
uning
hype
r
pa
r
a
mete
r
s
a
r
e
a
wa
y
to
a
void
p
r
obl
e
ms
r
e
late
d
to
ove
r
f
it
ti
ng
a
nd
unde
r
f
it
ti
ng
.
W
he
n
hype
r
pa
r
a
mete
r
va
lues
a
r
e
low,
the
model
is
una
b
le
to
dis
ti
nguis
h
be
twe
e
n
the
da
ta
dur
ing
the
t
r
a
in
ing
a
nd
tes
ti
ng
s
tage
s
,
r
e
s
ult
ing
in
unde
r
f
it
ti
ng
.
I
nc
r
e
a
s
ing
hype
r
pa
r
a
mete
r
va
lues
lea
ds
to
an
ove
r
f
it
t
ing
a
nd
c
ompl
ica
ted
model.
S
o,
in
thi
s
wor
k
a
g
r
id
s
e
a
r
c
h
method
is
e
mpl
oye
d
to
pr
ovide
ba
lanc
e
d
r
e
s
ult
s
w
it
h
f
ine
-
tuni
ng
e
s
s
e
nti
a
l
hype
r
pa
r
a
mete
r
s
.
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
E
nhanc
ing
tr
adit
ional
mac
hine
lear
ning
me
thods
us
ing
c
onc
atenation
…
(
R
afal
N
az
ar
Y
oune
s
A
l
-
T
a
han
)
2971
T
a
ble
1.
Hype
r
pa
r
a
mete
r
s
opti
mi
z
a
ti
on
M
ode
ls
H
ype
r
pa
r
a
me
te
r
s
L
G
B
M
num_l
e
a
ve
s
=
30, n
_
e
s
ti
ma
to
r
s
=
100,
m
a
x_de
pt
h=
7
DT
M
a
x_l
e
a
f
_node
s
=
20,
ma
x_de
pt
h
=
10
NV
V
a
r
_s
moot
hi
ng=
le
-
9
KNN
N
n_ne
ig
hbor
s
=
5, l
e
a
f
_s
iz
e
=
30
E
ns
e
mbl
e
vot
in
g
V
ot
in
g=
’
s
of
t
’
, n
_j
obs
=
-
1
6.
P
E
RF
ORM
AN
CE
E
VA
L
UA
T
I
ON
Af
ter
de
ve
lopi
ng
the
models
,
thei
r
pe
r
f
or
manc
e
wa
s
e
va
luate
d
us
ing
a
va
r
iety
of
mea
s
ur
e
s
,
including
a
c
c
ur
a
c
y,
r
e
c
a
ll
,
p
r
e
c
is
ion,
f1
-
s
c
or
e
a
n
d
r
e
c
e
iver
ope
r
a
ti
ng
c
ha
r
a
c
ter
is
ti
c
(
R
OC
)
-
a
r
e
a
u
nde
r
the
c
ur
ve
(
AUC
)
.
Ac
c
ur
a
c
y
mea
s
ur
e
s
c
las
s
if
ica
ti
on
tas
k
pe
r
f
or
manc
e
by
c
ounti
ng
the
number
of
c
or
r
e
c
tl
y
e
va
luate
d
ins
tanc
e
s
a
c
r
os
s
a
ll
da
ta
s
a
mpl
e
s
.
R
e
c
a
ll
is
a
us
e
f
ul
qua
nti
ty
mea
s
ur
e
f
or
de
tec
ti
ng
mod
e
l
e
r
r
o
r
s
.
W
hil
e
pr
e
c
is
ion
is
a
qua
li
ty
metr
ic
that
r
e
f
e
r
s
to
th
e
pe
r
c
e
ntage
of
c
or
r
e
c
tl
y
identif
ied
pos
it
ive
ins
tan
c
e
s
.
T
he
f
1
-
s
c
or
e
is
a
metr
ic
int
e
nde
d
to
s
tr
ike
a
c
o
mpr
omi
s
e
be
twe
e
n
pr
e
c
is
ion
a
nd
r
e
c
a
ll
.
F
inally
,
AUC
-
R
OC
is
a
c
las
s
if
ica
ti
on
mea
s
ur
e
that
de
ter
m
ines
how
we
ll
a
c
las
s
if
ier
dis
ti
nguis
he
s
be
twe
e
n
c
las
s
e
s
at
va
r
ious
thr
e
s
holds
.
It
de
mons
tr
a
tes
the
t
r
a
de
-
of
f
be
twe
e
n
s
pe
c
if
icity
a
nd
s
e
ns
it
ivi
ty
in
tes
ti
ng
t
ha
t
yield
numer
ica
l
f
indi
ngs
ins
tea
d
of
a
binar
y
pos
it
ive
or
ne
ga
ti
ve
c
onc
lus
ion.
T
he
AUC
-
R
OC
(
de
c
is
ion
thr
e
s
holds
)
gives
the
be
s
t
c
ut
-
of
f
f
or
both
s
e
ns
it
ivi
ty
a
nd
s
pe
c
if
icity.
T
he
R
OC
c
ur
ve
f
o
r
each
c
las
s
is
dis
playe
d
both
the
tr
ue
pos
it
ive
r
a
te
a
nd
f
a
ls
e
pos
it
ive
r
a
te.
W
he
n
the
AUC
va
lue
f
or
each
c
las
s
is
1.
0,
it
im
pli
e
s
pe
r
f
e
c
t
dis
c
r
im
ination,
whe
r
e
a
s
0.
5
s
hows
no
dis
c
r
im
inat
ion
i.
e
(
r
a
ndom
gue
s
s
ing)
.
T
he
s
e
metr
ics
a
r
e
e
xpr
e
s
s
e
d
as
f
oll
ows
in
(2
)
to
(
5
)
[
35
]
:
=
+
(
2)
=
+
(
3)
1
−
=
2
∗
∗
+
(
4)
=
+
+
+
+
(
5)
7.
RE
S
UL
T
S
AND
DI
S
CU
S
S
I
ON
7.
1.
F
irs
t
d
at
as
e
t
(
s
at
e
ll
it
e
im
age
s
)
F
igur
e
6
s
hows
the
AUC
-
R
O
C
of
M
L
models
(
f
ir
s
t
da
tas
e
t)
.
It
wa
s
noted
in
F
igur
e
6
(
a
)
,
that
R
OC
-
AUC
metr
ic
of
the
L
GB
M
a
lgor
it
hm
ha
s
a
c
hieve
d
the
highes
t
pe
r
c
e
ntage
,
whic
h
is
100%
due
to
L
GB
M
d
e
c
r
e
a
s
e
s
the
c
o
s
t
of
los
s
by
s
pli
tt
ing
the
tr
e
e
int
o
lea
ve
s
r
a
ther
than
at
the
de
pth
leve
l
e
mpl
oye
d
in
pr
ior
boos
ti
ng
methods
.
M
or
e
ove
r
,
it
f
ol
lows
pa
r
a
ll
e
l
lea
r
ning
us
ing
lar
ge
da
ta
to
s
pe
e
d
up
the
da
ta
tr
a
ini
ng
pr
oc
e
s
s
.
Unlike
the
NB
a
lgor
it
hm
in
F
igu
r
e
6(
b)
,
whic
h
a
c
hieve
d
the
lowe
s
t
pe
r
c
e
ntage
62%
a
mong
the
mentioned
methods
be
c
a
us
e
it
r
e
li
e
s
on
the
a
s
s
umpt
ion
that
the
f
e
a
tur
e
s
a
r
e
c
las
s
if
ying
da
ta
s
e
ts
with
c
ompl
e
x
hier
a
r
c
hica
l
s
tr
uc
tu
r
e
s
.
As
f
or
F
igu
r
e
s
6
(
c
)
to
6
(
e
)
,
they
a
c
hieve
d
be
s
t
indepe
nde
nt,
a
nd
thus
the
model’
s
pr
e
dictions
may
be
inac
c
ur
a
te,
in
a
ddit
ion
to
its
be
ing
uns
uit
a
ble
f
or
r
e
s
ult
s
in
the
AUC
-
R
OC
a
nd
c
las
s
if
y
s
a
mpl
e
c
las
s
e
s
.
F
igur
e
s
7
dis
play
the
c
onf
us
ion
matr
ix
of
S
R
S
I
c
las
s
if
ica
ti
on
to
s
how
mor
e
a
bout
the
r
e
s
ult
s
a
nd
how
they
c
ha
nge
a
c
r
os
s
thr
e
e
c
las
s
e
s
in
f
ive
mod
e
ls
.
T
he
c
onf
us
ion
matr
ix
s
hows
how
dif
f
icult
it
is
f
or
the
f
ive
models
to
c
hoos
e
be
twe
e
n
th
r
e
e
di
f
f
e
r
e
nt
c
las
s
e
s
(
de
s
e
r
t,
gr
e
e
n_a
r
e
a
,
a
nd
wa
ter
)
.
It
is
a
numer
i
c
a
l
table
that
il
lus
tr
a
ted
whe
r
e
ther
e
is
c
onf
us
ion
on
a
c
las
s
if
ier
.
It
is
de
s
igned
to
li
nk
p
r
e
dictions
to
the
or
igi
na
l
c
las
s
e
s
to
whic
h
the
da
ta
be
longs
.
It
is
us
e
d
in
s
upe
r
vis
e
d
lea
r
ning
f
or
c
a
lcula
ti
ng
a
va
r
iety
of
metr
ics
in
a
d
dit
ion
to
a
c
c
ur
a
c
y.
A
c
onf
us
ion
matr
ix
c
r
e
a
ted
f
or
the
s
a
me
tes
t
s
e
t
of
a
da
tas
e
t
but
us
ing
va
r
ious
c
las
s
if
ier
s
may
a
ls
o
a
s
s
is
t
a
na
lyze
their
r
e
lative
s
tr
e
ngths
a
nd
we
a
kn
e
s
s
e
s
a
nd
dr
a
w
r
e
c
omm
e
nda
ti
ons
a
bout
how
to
c
ombi
ne
them
f
or
be
s
t
pe
r
f
o
r
manc
e
.
Als
o,
in
F
igur
e
7(
a
)
,
I
t
wa
s
noti
c
e
d
that
diagona
l
e
leme
nts
r
e
pr
e
s
e
nt
the
c
or
r
e
c
t
pr
e
dictions
,
s
o
the
L
GB
M
model
c
las
s
if
ied
the
highe
r
va
lues
,
indi
c
a
ti
ng
that
the
model
is
be
tt
e
r
a
t
pr
e
dicting
th
is
s
pe
c
if
ic
c
las
s
.
W
hil
e
a
f
e
w
of
f
-
diagona
l
va
lues
we
r
e
obs
e
r
ve
d,
s
howing
that
the
model
L
GB
M
s
uc
c
e
e
de
d
in
not
mi
xing
be
twe
e
n
the
c
las
s
s
a
mpl
e
s
.
I
n
F
igur
e
7(
b)
,
the
NB
model
c
las
s
if
ied
the
lowe
s
t
diagon
a
l
va
lues
,
indi
c
a
ti
ng
that
the
model
is
poor
a
t
pr
e
dicting
thes
e
s
pe
c
if
ic
c
las
s
e
s
.
How
e
ve
r
,
it
s
howe
d
high
o
f
f
-
diag
ona
l
va
lues
,
i
mpl
ying
that
the
model
c
onf
us
e
d
c
las
s
s
a
mpl
e
s
or
f
a
il
e
d
t
o
c
a
ptur
e
the
ne
c
e
s
s
a
r
y
dis
ti
nc
ti
ons
be
twe
e
n
thes
e
c
las
s
e
s
.
As
f
or
F
igu
r
e
s
7(
c
)
to
7
(
e
)
,
they
a
c
hieve
d
moder
a
t
e
a
nd
good
r
e
s
ult
s
a
nd
c
las
s
if
y
s
a
mpl
e
c
las
s
e
s
.
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.
14,
No.
4,
Augus
t
2025
:
296
4
-
2978
2972
I
n
T
a
ble
2
,
it
wa
s
noti
c
e
d
that
a
c
c
ur
a
c
y,
p
r
e
c
is
ion,
r
e
c
a
ll
a
nd
f
1
-
s
c
or
e
metr
ics
of
the
L
GB
M
model
a
r
e
be
tt
e
r
than
the
other
models
,
whic
h
mea
ns
tha
t
the
L
GB
M
s
uc
c
e
e
d
s
in
pr
e
dicting
mor
e
pos
it
ive
s
a
mpl
e
s
(
S
R
S
I
)
than
the
other
model.
How
e
ve
r
,
the
NB
model
wa
s
not
lucky
in
pr
e
dicting
c
or
r
e
c
t
s
a
m
ples
a
nd
a
c
hieve
d
poor
r
e
s
ult
s
c
ompar
e
d
to
the
r
e
s
t
of
th
e
other
models
be
c
a
us
e
it
is
inade
qua
te
f
or
c
a
te
gor
izing
da
tas
e
ts
with
c
ompl
ica
ted
hier
a
r
c
hica
l
s
tr
uc
tur
e
s
.
As
a
c
ons
e
que
nc
e
,
the
outcome
s
of
tr
a
dit
ional
models
(
DT
,
KN
N
,
a
nd
NB
)
ha
ve
be
e
n
im
pr
ove
d
by
e
m
ployi
ng
e
ns
e
mbl
e
voti
ng,
whic
h
c
ombi
ne
s
the
s
tr
e
ngths
of
e
a
c
h
model
while
r
e
duc
ing
the
inf
luenc
e
of
f
a
ult
s
in
other
models
.
I
n
other
wor
ds
,
it
r
e
li
e
s
on
mer
ging
the
f
indi
ngs
of
s
e
ve
r
a
l
models
to
obtain
high
pe
r
f
or
ma
nc
e
a
nd
a
c
c
ur
a
c
y.
(
a
)
(
b)
(
c
)
(
d)
(
e
)
F
igur
e
6.
AUC
-
R
OC
of
ML
models
(
f
ir
s
t
da
tas
e
t)
f
or
,
(
a
)
L
GB
M
,
(
b
)
NB
,
(
c
)
KNN
,
(
d)
DT
,
a
nd
(
e
)
e
ns
e
mbl
e
voti
ng
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
E
nhanc
ing
tr
adit
ional
mac
hine
lear
ning
me
thods
us
ing
c
onc
atenation
…
(
R
afal
N
az
ar
Y
oune
s
A
l
-
T
a
han
)
2973
(
a
)
(
b)
(
c
)
(
d)
(
e
)
F
igur
e
7.
C
onf
us
ion
matr
ix
of
ML
models
(
f
i
r
s
t
da
tas
e
t)
f
or
,
(
a
)
L
GB
M
,
(
b)
NB
,
(
c
)
KNN
,
(
d
)
DT
,
a
n
d
(
e
)
e
ns
e
mbl
e
voti
ng
T
a
ble
2.
E
va
luation
metr
ics
of
ML
models
(
F
ir
s
t
d
a
tas
e
t)
E
va
lu
a
ti
on
me
tr
ic
s
L
G
B
M
DT
NB
KNN
E
ns
e
mbl
e
vot
in
g
P
r
e
c
is
io
n
0.99
0.93
0.57
0.96
0.96
R
e
c
a
ll
0.99
0.93
0.50
0.95
0.96
F1
-
s
c
or
e
0.99
0.93
0.48
0.96
0.96
A
c
c
ur
a
c
y
0.99
0.93
0.50
0.95
0.96
7.
2.
S
e
c
on
d
d
at
as
e
t
(
d
at
a
c
oll
e
c
t
e
d
f
r
om
m
u
lt
i
p
le
s
ou
r
c
e
s
)
F
igur
e
8
s
hows
the
AU
C
-
R
OC
f
or
M
L
models
(
s
e
c
ond
da
tas
e
t)
.
Ac
c
or
ding
to
F
igur
e
8(
a
)
,
L
GB
M
outper
f
or
med
othe
r
models
in
AU
C
-
R
OC
with
a
s
c
or
e
of
100
%
.
I
n
F
igur
e
8
(
b)
,
NB
ha
d
th
e
lowe
s
t
AUC
-
R
OC
va
lue
in
the
s
e
c
ond
da
tas
e
t,
a
bout
70
%
,
bu
t
it
outper
f
or
med
the
f
ir
s
t
da
tas
e
t
due
to
the
lar
ge
da
ta
Evaluation Warning : The document was created with Spire.PDF for Python.