I
nte
rna
t
io
na
l J
o
urna
l o
f
I
nfo
rm
a
t
ics a
nd
Co
m
m
un
ica
t
io
n T
ec
hn
o
lo
g
y
(
I
J
-
I
CT
)
Vo
l.
1
5
,
No
.
4
,
D
e
c
em
b
er
20
2
6
,
p
p
.
1
6
4
8
~1
6
6
3
I
SS
N:
2252
-
8
7
7
6
,
DOI
:
1
0
.
1
1
5
9
1
/iji
ct
.
v
1
5
i
4
.
pp
1
6
4
8
-
1
6
6
3
1648
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ict.
ia
esco
r
e.
co
m
From da
ta to
deci
sio
ns: ensemble
d
eep learning
fram
ewo
rk f
o
r
ma
ize seed va
riet
y
clas
sifica
tion wi
th
e
x
pla
ina
ble
AI
E
m
m
a
nu
el
Asa
nte
1
,
O
bed Appia
h
2
,
E
ze
k
iel
M
e
ns
a
h
M
a
rt
ey
1
,
E
m
ma
nu
el
E
f
f
a
h
3
,
E
ric
O
po
k
u
4
,
E
rnest
K
wa
m
e
Am
po
m
a
h
1
1
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
i
n
g
a
n
d
D
a
t
a
A
n
a
l
y
t
i
c
s
,
U
n
i
v
e
r
s
i
t
y
o
f
M
i
n
e
s
a
n
d
T
e
c
h
n
o
l
o
g
y
,
T
a
r
k
w
a
,
G
h
a
n
a
2
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
U
n
i
v
e
r
s
i
t
y
o
f
E
n
e
r
g
y
a
n
d
N
a
t
u
r
a
l
R
e
s
o
u
r
c
e
s
,
S
u
n
y
a
n
i
,
G
h
a
n
a
3
D
e
p
a
r
t
me
n
t
o
f
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
U
n
i
v
e
r
si
t
y
o
f
M
i
n
e
s
a
n
d
Te
c
h
n
o
l
o
g
y
,
T
a
r
k
w
a
,
G
h
a
n
a
4
B
r
e
d
e
se
n
C
e
n
t
e
r
,
U
n
i
v
e
r
s
i
t
y
o
f
T
e
n
n
e
ssee
,
K
n
o
x
v
i
l
l
e
,
U
S
A
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
an
1
0
,
2
0
2
6
R
ev
is
ed
Au
g
4
,
2
0
2
6
Acc
ep
ted
Au
g
2
8
,
2
0
2
6
Ac
c
u
ra
te
a
n
d
n
o
n
-
d
e
stru
c
ti
v
e
i
d
e
n
ti
fica
ti
o
n
o
f
m
a
ize
se
e
d
v
a
rieties
is
e
ss
e
n
ti
a
l
fo
r
m
a
in
tain
in
g
se
e
d
p
u
rit
y
,
imp
ro
v
in
g
c
ro
p
p
ro
d
u
c
ti
v
it
y
,
a
n
d
su
p
p
o
rti
n
g
a
g
ricu
lt
u
ra
l
q
u
a
li
t
y
a
ss
u
ra
n
c
e
.
Alth
o
u
g
h
re
c
e
n
t
st
u
d
ies
re
p
o
r
t
h
ig
h
a
c
c
u
ra
c
ies
u
sin
g
a
d
v
a
n
c
e
d
i
m
a
g
in
g
a
n
d
d
e
e
p
lea
rn
i
n
g
tec
h
n
i
q
u
e
s,
m
a
n
y
re
ly
o
n
h
y
p
e
rsp
e
c
t
ra
l
se
n
so
rs
o
r
c
o
n
tr
o
ll
e
d
a
c
q
u
isit
i
o
n
e
n
v
ir
o
n
m
e
n
t
s,
o
r
o
ffe
r
li
m
it
e
d
i
n
terp
re
tab
il
it
y
,
t
h
e
re
b
y
r
e
strictin
g
re
a
l
-
wo
rl
d
a
d
o
p
ti
o
n
.
T
h
is
stu
d
y
in
v
e
stig
a
tes
m
a
ize
se
e
d
v
a
riety
c
las
sifica
ti
o
n
u
sin
g
lo
w
-
c
o
st
RG
B
ima
g
in
g
a
n
d
a
n
e
n
se
m
b
le
d
e
e
p
lea
rn
i
n
g
fra
m
e
wo
rk
,
fo
c
u
sin
g
o
n
th
r
e
e
wid
e
ly
c
u
lt
iv
a
ted
G
h
a
n
a
ian
v
a
rieties
:
B
h
ih
il
ifa,
S
a
n
z
a
l
S
ima
,
a
n
d
Wan
g
Da
taa
.
I
n
c
o
ll
a
b
o
ra
ti
o
n
wit
h
He
rit
a
g
e
S
e
e
d
s
G
h
a
n
a
,
a
n
e
x
p
e
rt
-
v
e
rifi
e
d
d
a
tas
e
t
o
f
1
7
,
7
2
4
m
a
ize
se
e
d
ima
g
e
s
w
a
s
c
u
ra
ted
u
n
d
e
r
re
a
li
stic
a
c
q
u
isit
i
o
n
c
o
n
d
i
ti
o
n
s
with
a
n
A
p
p
l
e
iP
h
o
n
e
1
2
.
A
b
a
se
li
n
e
c
o
n
v
o
lu
t
io
n
a
l
n
e
u
ra
l
n
e
two
rk
(CNN
)
wa
s
d
e
v
e
lo
p
e
d
fo
r
b
e
n
c
h
m
a
rk
i
n
g
,
f
o
ll
o
we
d
b
y
tra
n
sfe
r
lea
rn
in
g
wit
h
De
n
se
Ne
t1
6
9
,
M
o
b
i
leN
e
tV2
,
a
n
d
Xc
e
p
ti
o
n
a
rc
h
it
e
c
tu
re
s.
A
n
e
n
se
m
b
le
fra
m
e
wo
rk
in
te
g
ra
ti
n
g
De
n
se
Ne
t
1
6
9
a
n
d
Xc
e
p
ti
o
n
wa
s
p
ro
p
o
se
d
t
o
i
m
p
r
o
v
e
ro
b
u
stn
e
ss
a
n
d
g
e
n
e
ra
li
z
a
ti
o
n
.
Ac
ro
ss
fiv
e
-
fo
ld
c
r
o
ss
-
v
a
li
d
a
ti
o
n
,
th
e
e
n
se
m
b
le
a
c
h
iev
e
d
a
m
e
a
n
a
c
c
u
ra
c
y
o
f
9
8
.
8
9
±0
.
1
7
%
,
p
re
c
isio
n
o
f
9
8
.
8
6
±
0
.
1
8
%
,
re
c
a
ll
o
f
9
8
.
8
0
±0
.
1
7
%
,
a
n
d
F
1
-
sc
o
re
o
f
9
8
.
8
3
±
0
.
1
7
%
.
T
o
e
n
h
a
n
c
e
tran
sp
a
re
n
c
y
,
G
ra
d
-
CAM
a
n
d
sa
li
e
n
c
y
m
a
p
s
we
re
e
m
p
lo
y
e
d
t
o
v
isu
a
li
z
e
d
isc
rimin
a
ti
v
e
fe
a
tu
re
s
in
flu
e
n
c
i
n
g
m
o
d
e
l
d
e
c
isi
o
n
s
.
Re
su
lt
s
d
e
m
o
n
stra
te
th
a
t
a
c
c
u
ra
te
a
n
d
in
terp
re
tab
le
c
las
sifica
ti
o
n
o
f
m
a
ize
se
e
d
v
a
rieties
is
a
c
h
iev
a
b
le
u
sin
g
st
a
n
d
a
rd
RG
B
ima
g
e
ry
with
o
u
t
s
p
e
c
ialize
d
se
n
sin
g
e
q
u
i
p
m
e
n
t.
Ke
ywo
r
d
s
:
E
n
s
em
b
le
d
ee
p
lear
n
in
g
E
x
p
lain
ab
le
ag
r
icu
ltu
r
al
AI
L
o
w
-
r
eso
u
r
ce
r
g
b
im
ag
in
g
Ma
ize
s
ee
d
v
ar
iety
class
if
icatio
n
Pre
cisi
o
n
f
ar
m
in
g
Seed
v
ar
iety
id
en
tific
atio
n
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
E
m
m
an
u
e
l
Asan
te
Dep
ar
tm
en
t o
f
C
o
m
p
u
tin
g
an
d
Data
An
aly
tics
,
Sch
o
o
l o
f
R
ai
lway
an
d
I
n
f
r
astru
ctu
r
al
Dev
el
o
p
m
en
t
Un
iv
er
s
ity
o
f
Min
es a
n
d
T
ec
h
n
o
lo
g
y
T
ar
k
wa,
Gh
an
a
E
m
ail:
ea
s
an
te@
u
m
at.
ed
u
.
g
h
1.
I
NT
RO
D
UCT
I
O
N
Ma
iz
e
(
Ze
a
m
a
ys
L
.
)
is
o
n
e
o
f
t
h
e
m
o
s
t
im
p
o
r
ta
n
t
c
er
ea
l
cr
o
p
s
g
l
o
b
all
y
,
s
er
v
i
n
g
as
a
s
tap
le
f
o
o
d
,
liv
est
o
c
k
f
ee
d
,
a
n
d
i
n
d
u
s
t
r
ial
r
aw
m
ate
r
i
al
[
1
]
,
[
2
]
.
Gl
o
b
al
m
aiz
e
p
r
o
d
u
c
ti
o
n
n
o
w
a
cc
o
u
n
ts
f
o
r
a
p
p
r
o
x
i
m
at
el
y
3
9
%
o
f
to
tal
c
er
ea
l
o
u
t
p
u
t
,
a
g
r
o
w
th
d
r
i
v
en
b
y
ad
v
an
ce
s
i
n
h
y
b
r
id
b
r
e
e
d
i
n
g
a
n
d
s
ee
d
i
m
p
r
o
v
e
m
e
n
t
tec
h
n
o
lo
g
i
es
[
3
]
,
[
4
]
.
H
o
we
v
e
r
,
th
ese
a
d
v
a
n
ce
s
h
av
e
als
o
le
d
t
o
a
r
a
p
i
d
p
r
o
li
f
er
ati
o
n
o
f
v
ar
i
eti
es
wi
th
h
ig
h
l
y
s
im
il
ar
m
o
r
p
h
o
lo
g
i
ca
l
c
h
ar
ac
t
er
is
tics
,
m
a
k
i
n
g
r
eli
ab
le
i
d
e
n
ti
f
icat
io
n
o
f
s
e
ed
v
a
r
ie
ties
i
n
c
r
e
a
s
in
g
l
y
c
h
all
e
n
g
in
g
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
F
r
o
m
d
a
ta
to
d
ec
is
io
n
s
:
en
s
emb
le
d
ee
p
lea
r
n
in
g
fr
a
mewo
r
k
fo
r
ma
iz
e
s
ee
d
va
r
iety
…
(
E
mma
n
u
el
A
s
a
n
te
)
1649
Acc
u
r
ate
id
en
tific
atio
n
o
f
m
ai
ze
s
ee
d
v
a
r
ieties
is
cr
itical
f
o
r
m
ain
tain
in
g
v
a
r
ietal
p
u
r
ity
,
o
p
t
im
izin
g
y
ield
,
an
d
en
s
u
r
i
n
g
ec
o
n
o
m
ic
r
etu
r
n
s
f
o
r
f
ar
m
e
r
s
.
Seed
v
ar
iety
m
ix
in
g
d
u
r
in
g
g
r
o
wth
,
h
ar
v
esti
n
g
,
tr
an
s
p
o
r
tatio
n
,
an
d
s
to
r
ag
e
p
o
s
es
s
er
io
u
s
ch
allen
g
es
to
s
ee
d
p
u
r
ity
an
d
q
u
ality
co
n
tr
o
l
[
4
]
,
[
5
]
.
Sin
ce
d
if
f
er
en
t
m
aize
v
ar
ieties
ar
e
ad
ap
ted
to
s
p
ec
if
ic
s
o
il
ty
p
es,
clim
ates,
an
d
cu
ltiv
atio
n
p
r
ac
tices,
en
s
u
r
in
g
v
ar
ietal
p
u
r
ity
b
ef
o
r
e
p
lan
tin
g
is
ess
en
tial f
o
r
b
o
th
a
g
r
icu
lt
u
r
al
p
r
o
d
u
ctiv
ity
an
d
ec
o
n
o
m
ic
s
u
s
tain
ab
ilit
y
[
5
]
,
[
6
]
.
T
r
ad
itio
n
al
id
en
tific
atio
n
m
eth
o
d
s
,
in
clu
d
in
g
DNA
m
o
lecu
lar
m
a
r
k
er
an
aly
s
is
,
p
r
o
tein
elec
tr
o
p
h
o
r
esis
,
g
as
ch
r
o
m
at
o
g
r
ap
h
y
,
an
d
ex
p
er
t
v
is
u
al
in
s
p
ec
tio
n
,
ar
e
h
ig
h
ly
ac
c
u
r
ate
b
u
t
lab
o
r
-
in
ten
s
iv
e,
co
s
tly
,
an
d
d
ep
e
n
d
en
t
o
n
s
p
ec
ialized
lab
o
r
ato
r
y
e
q
u
ip
m
e
n
t
[
3
]
,
[
7
]
.
T
h
ese
co
n
s
tr
ain
ts
m
a
k
e
th
em
im
p
r
ac
tical
f
o
r
r
o
u
tin
e,
la
r
g
e
-
s
ca
le
s
ee
d
in
s
p
ec
tio
n
,
cr
ea
tin
g
a
p
r
ess
in
g
d
em
an
d
f
o
r
r
a
p
id
,
n
o
n
-
d
estru
ctiv
e,
an
d
co
s
t
-
ef
f
ec
tiv
e
alter
n
ativ
es.
I
n
r
esp
o
n
s
e,
r
esear
ch
e
r
s
h
a
v
e
in
cr
ea
s
in
g
ly
ex
p
lo
r
ed
i
m
ag
e
-
b
ased
n
o
n
-
d
estru
ctiv
e
tech
n
iq
u
es.
Ma
ch
in
e
v
is
io
n
,
n
ea
r
-
i
n
f
r
ar
e
d
s
p
ec
tr
o
s
co
p
y
(
NI
R
S),
an
d
h
y
p
er
s
p
ec
tr
al
im
ag
i
n
g
(
HSI
)
h
av
e
d
em
o
n
s
tr
ated
s
tr
o
n
g
class
if
icatio
n
p
er
f
o
r
m
a
n
ce
[
5
]
,
[
8
]
–
[
1
1
]
.
Ma
ch
in
e
v
is
io
n
ex
tr
ac
ts
m
o
r
p
h
o
lo
g
ical
f
e
atu
r
es
s
u
ch
as
s
ize,
s
h
ap
e,
co
lo
r
,
a
n
d
tex
tu
r
e
f
r
o
m
R
GB
im
ag
es
[
1
2
]
,
wh
ile
NI
R
S
ca
p
tu
r
es
c
h
em
ical
-
b
o
n
d
ab
s
o
r
p
tio
n
in
t
h
e
7
8
0
–
2
5
0
0
n
m
r
an
g
e
[
4
]
.
HSI
in
te
g
r
ates
b
o
th
s
p
atial
an
d
s
p
ec
tr
al
in
f
o
r
m
atio
n
,
p
r
o
v
id
in
g
r
ich
r
ep
r
esen
tatio
n
s
f
o
r
d
is
tin
g
u
is
h
in
g
v
is
u
ally
s
im
ilar
v
ar
ieties
[
4
]
,
[
5
]
.
Desp
ite
th
e
ir
ef
f
ec
tiv
e
n
ess
,
h
y
p
er
s
p
ec
tr
al
an
d
s
p
ec
tr
o
s
co
p
ic
s
y
s
tem
s
ar
e
ex
p
en
s
iv
e,
co
m
p
u
tatio
n
ally
d
e
m
an
d
in
g
,
a
n
d
c
o
m
p
lex
to
d
e
p
lo
y
[
1
3
]
,
[
1
4
]
,
lim
itin
g
ad
o
p
tio
n
i
n
lo
w
-
r
eso
u
r
ce
ag
r
icu
ltu
r
al
en
v
i
r
o
n
m
en
ts
.
Stan
d
ar
d
R
GB
im
a
g
in
g
,
r
ea
d
ily
av
ailab
le
v
ia
s
m
ar
tp
h
o
n
es
an
d
lo
w
-
co
s
t c
am
er
as,
o
f
f
er
s
a
f
ar
m
o
r
e
ac
ce
s
s
ib
le
alter
n
ativ
e.
Dee
p
lear
n
in
g
h
as
em
e
r
g
ed
as
a
p
o
we
r
f
u
l
t
o
o
l
f
o
r
i
m
ag
e
-
b
ased
s
ee
d
class
if
icati
o
n
,
wit
h
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
s
(
C
NNs)
an
d
ad
v
a
n
ce
d
ar
ch
it
ec
tu
r
es
r
ep
o
r
tin
g
h
ig
h
ac
c
u
r
ac
ies
ac
r
o
s
s
m
u
ltip
le
s
tu
d
ies
[
5
]
-
[
7
]
,
[
1
5
]
,
[
1
6
]
.
Ho
wev
er
,
m
an
y
ex
is
tin
g
m
aize
s
ee
d
class
if
icatio
n
s
tu
d
ies
s
u
f
f
er
f
r
o
m
o
n
e
o
r
m
o
r
e
o
f
th
e
f
o
llo
win
g
lim
itatio
n
s
:
(
i)
r
elian
ce
o
n
ex
p
e
n
s
iv
e
im
ag
in
g
m
o
d
alities
th
at
r
estrict
d
ep
lo
y
m
en
t
o
n
lo
w
-
r
eso
u
r
ce
o
r
m
o
b
ile
d
ev
ices;
(
ii)
lack
o
f
ev
alu
atio
n
o
n
r
e
g
io
n
-
s
p
ec
if
ic
v
ar
ieties,
p
ar
ticu
lar
ly
th
o
s
e
cu
ltiv
ated
in
s
u
b
-
Sah
ar
an
Af
r
ica;
an
d
(
iii)
i
n
s
u
f
f
icien
t
m
o
d
el
in
ter
p
r
etab
il
ity
,
wh
ich
is
in
cr
ea
s
in
g
ly
im
p
o
r
tan
t
f
o
r
tr
u
s
t
an
d
ad
o
p
tio
n
in
ag
r
ic
u
ltu
r
al
d
ec
is
i
o
n
-
m
ak
i
n
g
s
y
s
tem
s
.
T
h
is
s
tu
d
y
ad
d
r
ess
es
th
ese
g
ap
s
b
y
i
n
v
esti
g
atin
g
ac
cu
r
a
te,
in
ter
p
r
etab
le,
an
d
r
eso
u
r
c
e
-
ef
f
icien
t
m
aize
s
ee
d
v
ar
iety
class
if
icatio
n
u
s
in
g
lo
w
-
c
o
s
t
R
GB
i
m
ag
in
g
.
W
e
f
o
c
u
s
o
n
th
r
ee
wid
ely
cu
ltiv
ated
Gh
an
aian
v
ar
ieties,
B
h
ih
ilifa
,
San
za
l
Sima
,
an
d
W
an
g
Data
a,
ca
p
tu
r
ed
with
a
s
tan
d
ar
d
s
m
a
r
tp
h
o
n
e
ca
m
er
a
.
I
n
co
llab
o
r
atio
n
with
Her
itag
e
S
ee
d
s
Gh
an
a,
we
c
u
r
ated
a
lar
g
e,
ex
p
er
t
-
v
er
i
f
ied
d
ataset
o
f
1
7
,
7
2
4
m
aize
s
ee
d
im
ag
es
u
n
d
er
r
e
alis
tic
ac
q
u
is
i
tio
n
co
n
d
itio
n
s
.
T
o
o
u
r
k
n
o
w
led
g
e,
th
is
is
th
e
f
ir
s
t
p
u
b
l
icly
av
ailab
le,
ex
p
e
r
t
-
lab
eled
R
GB
d
ataset
o
f
Gh
an
aian
m
aize
s
ee
d
v
ar
ieties,
ad
d
r
ess
in
g
a
cr
itical
g
a
p
in
r
eg
io
n
-
s
p
ec
if
ic
ag
r
icu
ltu
r
al
AI
r
esear
ch
.
W
e
p
r
o
p
o
s
e
an
e
n
s
em
b
le
d
ee
p
lear
n
i
n
g
f
r
am
ewo
r
k
in
teg
r
atin
g
c
o
m
p
lem
e
n
tar
y
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
f
r
o
m
f
in
e
-
t
u
n
e
d
Den
s
eNe
t1
6
9
[
1
7
]
an
d
Xce
p
tio
n
[
1
8
]
a
r
ch
itectu
r
es.
T
o
p
r
o
m
o
te
tr
an
s
p
ar
en
c
y
an
d
tr
u
s
t
in
AI
-
d
r
iv
en
ag
r
ic
u
ltu
r
al
s
y
s
tem
s
,
we
in
co
r
p
o
r
at
e
ex
p
lain
ab
le
AI
tech
n
iq
u
es,
s
p
ec
if
ically
Gr
ad
-
C
AM
[
1
9
]
an
d
s
alien
cy
m
ap
s
[
2
0
]
,
to
v
is
u
alize
th
e
d
is
cr
im
in
ativ
e
s
ee
d
f
ea
tu
r
es
in
f
lu
en
cin
g
m
o
d
el
p
r
ed
ictio
n
s
.
T
h
e
m
ain
co
n
tr
ib
u
tio
n
s
o
f
th
is
wo
r
k
ar
e
as f
o
llo
ws:
−
A
lar
g
e,
r
eg
io
n
-
s
p
ec
if
ic
m
aiz
e
s
ee
d
im
ag
e
d
ataset
f
o
r
th
r
e
e
Gh
an
aian
v
ar
ieties,
c
o
llected
u
s
in
g
lo
w
-
r
eso
u
r
ce
R
GB
im
ag
in
g
an
d
ex
p
er
t la
b
elin
g
,
a
n
d
m
a
d
e
p
u
b
lic
ly
av
ailab
le.
−
A
co
m
p
u
tatio
n
ally
ef
f
icien
t
e
n
s
em
b
le
d
ee
p
lear
n
in
g
m
o
d
el
th
at
ac
h
iev
es
9
8
.
8
9
±
0
.
1
7
%
c
r
o
s
s
-
v
alid
ated
ac
cu
r
ac
y
with
o
u
t sp
ec
ialized
s
en
s
in
g
h
ar
d
wa
r
e.
−
An
in
ter
p
r
etab
le
class
if
icatio
n
f
r
a
m
ewo
r
k
th
at
p
r
o
v
id
es
v
is
u
al
ex
p
lan
atio
n
s
o
f
m
o
d
el
d
ec
is
io
n
s
v
ia
Gr
ad
-
C
AM
an
d
s
alien
cy
m
ap
s
,
s
u
p
p
o
r
tin
g
tr
u
s
t a
n
d
ad
o
p
tio
n
in
ag
r
icu
ltu
r
al
p
r
ac
tice.
B
y
ex
p
licitly
ad
d
r
ess
in
g
ac
ce
s
s
ib
ilit
y
,
in
ter
p
r
etab
ilit
y
,
a
n
d
r
ea
l
-
wo
r
ld
ap
p
licab
ilit
y
,
th
is
s
tu
d
y
ad
v
an
ce
s
m
aize
s
ee
d
v
ar
iety
class
if
icatio
n
b
ey
o
n
d
lab
o
r
ato
r
y
-
s
ca
le
s
o
lu
tio
n
s
with
d
ir
ec
t
r
elev
an
ce
f
o
r
s
ee
d
q
u
ality
ass
u
r
an
ce
in
Gh
a
n
a
a
n
d
n
eig
h
b
o
r
in
g
W
est
Af
r
ican
c
o
u
n
tr
ies.
T
h
e
r
e
m
ain
d
er
o
f
th
i
s
p
ap
er
is
o
r
g
a
n
ized
as
f
o
llo
ws:
Secti
o
n
2
d
escr
ib
es
th
e
r
esear
ch
m
eth
o
d
,
d
ata
s
et,
an
d
ex
p
er
im
en
tal
s
etu
p
.
Sectio
n
3
p
r
esen
ts
r
esu
lts
an
d
d
is
cu
s
s
io
n
.
Sectio
n
4
co
n
cl
u
d
es th
e
p
a
p
er
.
2.
M
E
T
H
O
D
T
h
is
s
ec
tio
n
d
escr
ib
es
th
e
c
o
m
p
lete
m
eth
o
d
o
lo
g
ical
p
ip
eli
n
e
f
o
r
m
aize
s
ee
d
v
ar
iety
cla
s
s
if
icatio
n
,
co
v
er
in
g
d
ata
ac
q
u
is
itio
n
,
p
r
ep
r
o
ce
s
s
in
g
,
m
o
d
el
d
e
v
elo
p
m
en
t,
en
s
em
b
le
c
o
n
s
tr
u
ctio
n
,
an
d
e
v
alu
atio
n
as
s
h
o
wn
in
Fig
u
r
e
1
.
T
h
e
s
tu
d
y
em
p
lo
y
ed
a
m
u
lti
-
p
h
ase
d
esi
g
n
:
(
i)
d
at
aset
cu
r
atio
n
u
n
d
e
r
r
ea
lis
tic
co
n
d
itio
n
s
u
s
in
g
a
1
2
MP
p
h
o
n
e
ca
m
er
a;
(
ii)
d
ev
elo
p
m
e
n
t
o
f
a
b
aselin
e
C
NN
f
o
r
b
en
ch
m
ar
k
i
n
g
;
(
iii
)
tr
an
s
f
er
lear
n
in
g
with
th
r
ee
p
r
e
-
tr
ain
e
d
ar
ch
it
ec
tu
r
es;
(
iv
)
en
s
em
b
le
co
n
s
t
r
u
ctio
n
v
ia
s
tack
in
g
;
an
d
(
v
)
ev
alu
atio
n
o
f
th
e
p
r
o
p
o
s
ed
en
s
em
b
le
u
s
in
g
f
iv
e
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
an
d
ex
p
l
ain
ab
ilit
y
v
is
u
aliza
tio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
1
5
,
No
.
4
,
D
e
c
em
b
er
20
2
6
:
1
6
4
8
-
1
6
6
3
1650
Fig
u
r
e
1
.
E
n
d
-
to
-
en
d
p
ip
elin
e
f
o
r
Gh
an
aia
n
m
aize
s
ee
d
v
a
r
ie
ty
class
if
icatio
n
,
f
r
o
m
ex
p
er
t
-
v
er
if
ied
R
GB
im
ag
e
ac
q
u
is
itio
n
th
r
o
u
g
h
p
r
e
p
r
o
ce
s
s
in
g
,
tr
an
s
f
er
lear
n
in
g
,
en
s
em
b
le
co
n
s
tr
u
ctio
n
,
f
iv
e
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
,
an
d
ex
p
lain
ab
ilit
y
an
aly
s
is
v
ia
Gr
ad
-
C
AM
an
d
s
alien
cy
m
a
p
s
2
.
1
.
E
x
perim
ent
a
l set
up
a
nd
env
iro
nm
ent
All
ex
p
er
im
en
ts
wer
e
c
o
n
d
u
cted
o
n
a
Dell
XPS
1
3
(
2
0
2
4
)
l
ap
to
p
e
q
u
ip
p
e
d
with
an
I
n
tel
C
o
r
e
Ultr
a
7
1
5
5
H
p
r
o
ce
s
s
o
r
(
u
p
to
4
.
8
GHz
,
1
6
co
r
es),
1
6
GB
L
PDDR
5
x
R
AM
(
7
4
6
7
MT
/s
)
,
an
d
I
n
tel
Ar
c
in
teg
r
ate
d
g
r
ap
h
ics.
T
h
e
s
o
f
twar
e
e
n
v
ir
o
n
m
en
t
c
o
n
s
is
ted
o
f
Py
t
h
o
n
3
.
1
0
.
1
2
,
T
en
s
o
r
Flo
w
2
.
1
3
.
0
with
th
e
in
teg
r
ated
Ker
as
API
(
tf
.
k
e
r
as),
Nu
m
Py
1
.
2
4
.
3
,
s
cik
it
-
lear
n
1
.
3
.
0
,
Ma
t
p
lo
tlib
3
.
7
.
2
,
an
d
O
p
en
C
V
4
.
8
.
0
.
Mo
d
el
tr
ain
in
g
was
p
er
f
o
r
m
e
d
o
n
a
C
PU
d
u
e
to
th
e
a
b
s
en
ce
o
f
a
d
is
cr
ete
GPU
in
th
e
XPS
1
3
co
n
f
ig
u
r
atio
n
,
th
er
e
b
y
v
alid
atin
g
th
e
f
e
asib
ilit
y
o
f
th
e
p
r
o
p
o
s
ed
ap
p
r
o
ac
h
o
n
ac
ce
s
s
ib
le
h
ar
d
war
e
.
2
.
2
.
Da
t
a
s
et
2
.
2
.
1
.
I
m
a
g
e
a
cquis
it
io
n
T
h
e
d
ataset
co
m
p
r
is
es
im
ag
es
o
f
th
r
e
e
m
aize
s
ee
d
v
ar
i
eties,
W
an
g
Data
a,
San
za
l
Sima
,
an
d
B
h
ih
ilifa
,
s
o
u
r
ce
d
f
r
o
m
Her
it
ag
e
Seed
s
Gh
an
a.
T
h
ese
c
u
ltiv
ar
s
ar
e
f
r
eq
u
e
n
tly
g
r
o
wn
in
n
o
r
th
er
n
Gh
an
a
.
Seed
s
wer
e
m
an
u
ally
ca
teg
o
r
ized
an
d
lab
eled
b
y
d
o
m
ain
ex
p
er
ts
at
Her
itag
e
Seed
s
Gh
an
a
p
r
i
o
r
t
o
im
ag
e
ca
p
tu
r
e.
I
m
ag
es
wer
e
ac
q
u
ir
ed
u
s
in
g
an
Ap
p
le
iPh
o
n
e
1
2
s
m
ar
tp
h
o
n
e
(
1
2
M
P
wid
e
-
an
g
le
s
en
s
o
r
,
f
/1
.
6
ap
er
tu
r
e,
2
6
m
m
eq
u
iv
alen
t f
o
ca
l
len
g
th
,
o
p
tical
im
ag
e
s
tab
ilizatio
n
,
Ap
p
le
A1
4
B
io
n
ic
p
r
o
ce
s
s
o
r
)
,
p
r
o
d
u
cin
g
o
r
ig
in
al
J
PEG
im
ag
es
at
4
0
3
2
×3
0
2
4
p
ix
els.
A
u
n
if
o
r
m
b
lu
e
b
ac
k
g
r
o
u
n
d
was
u
s
ed
d
u
r
in
g
ca
p
tu
r
e
to
en
h
an
ce
s
ee
d
v
is
ib
ilit
y
a
n
d
r
e
d
u
ce
b
ac
k
g
r
o
u
n
d
n
o
is
e
u
n
d
er
n
atu
r
al
d
ay
lig
h
t
c
o
n
d
itio
n
s
.
All
im
ag
es
wer
e
s
u
b
s
eq
u
en
tly
r
esized
to
2
2
4
×
2
2
4
p
ix
els
to
en
s
u
r
e
u
n
if
o
r
m
ity
ac
r
o
s
s
th
e
d
ataset.
T
h
e
ex
p
er
t
-
v
er
if
ied
la
b
elin
g
p
r
o
ce
s
s
was
ca
r
r
ied
o
u
t
o
v
er
two
wee
k
s
to
en
s
u
r
e
th
e
v
alid
ity
a
n
d
ac
c
u
r
a
cy
o
f
all
an
n
o
tatio
n
s
.
T
h
e
im
a
g
e
ac
q
u
is
itio
n
s
etu
p
is
illu
s
tr
ated
in
Fig
u
r
e
2
,
an
d
t
h
e
cu
r
atio
n
p
ip
elin
e
is
s
h
o
wn
i
n
Fig
u
r
e
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
F
r
o
m
d
a
ta
to
d
ec
is
io
n
s
:
en
s
emb
le
d
ee
p
lea
r
n
in
g
fr
a
mewo
r
k
fo
r
ma
iz
e
s
ee
d
va
r
iety
…
(
E
mma
n
u
el
A
s
a
n
te
)
1651
Fig
u
r
e
2
.
I
m
ag
e
ac
q
u
is
itio
n
s
etu
p
Fig
u
r
e
3.
Data
s
et
cu
r
atio
n
p
ip
elin
e
2
.
2
.
2.
Da
t
a
s
et
des
cr
iptio
n a
nd
s
t
a
t
is
t
ics
T
h
e
in
itial
d
ataset
co
m
p
r
is
ed
3
,
2
0
2
im
ag
es
d
is
tr
ib
u
te
d
ac
r
o
s
s
th
r
ee
v
ar
ieties:
1
,
3
2
8
B
h
ih
ilifa
,
8
5
0
San
za
l
Sima
,
an
d
1
,
0
2
4
W
an
g
Data
a.
T
ab
le
1
s
u
m
m
ar
izes
th
e
d
is
tr
ib
u
tio
n
b
ef
o
r
e
an
d
af
ter
au
g
m
en
tatio
n
.
Fo
r
th
e
b
aselin
e
C
NN
an
d
in
d
iv
i
d
u
al
tr
an
s
f
e
r
lear
n
in
g
m
o
d
els,
an
8
0
/2
0
s
tr
atif
ied
s
p
lit
was
u
s
ed
to
s
ep
ar
ate
tr
ain
in
g
an
d
test
d
ata,
with
an
ad
d
itio
n
al
2
0
% o
f
th
e
tr
ain
in
g
s
et
r
eser
v
ed
f
o
r
v
alid
atio
n
.
T
h
e
class
d
is
tr
ib
u
tio
n
ac
r
o
s
s
th
ese
s
p
lits
i
s
s
h
o
wn
in
Fig
u
r
e
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
1
5
,
No
.
4
,
D
e
c
em
b
er
20
2
6
:
1
6
4
8
-
1
6
6
3
1652
Fo
r
f
in
al
en
s
em
b
le
ev
alu
ati
o
n
,
f
i
ve
-
f
o
l
d
cr
o
s
s
-
v
alid
atio
n
was
co
n
d
u
cte
d
u
s
in
g
Stra
tifie
d
KFo
ld
(
n
_
s
p
lits
=5
,
s
h
u
f
f
le=
T
r
u
e
,
r
an
d
o
m
_
s
tate=
4
2
)
to
en
s
u
r
e
co
n
s
is
ten
t
clas
s
p
r
o
p
o
r
tio
n
s
ac
r
o
s
s
all
f
o
ld
s
.
I
n
ea
ch
f
o
ld
,
th
e
en
s
em
b
le
was
tr
ain
e
d
f
r
o
m
s
cr
atch
o
n
th
e
tr
ain
in
g
p
o
r
tio
n
o
f
th
at
f
o
ld
a
n
d
e
v
al
u
a
ted
o
n
th
e
h
eld
-
o
u
t
test
p
o
r
tio
n
.
T
h
is
p
r
o
ce
d
u
r
e
en
s
u
r
es
th
at
n
o
test
d
ata
was
s
ee
n
d
u
r
in
g
tr
ain
in
g
in
an
y
f
o
ld
,
p
r
o
v
id
i
n
g
an
u
n
b
iased
esti
m
ate
o
f
g
en
er
ali
za
tio
n
p
er
f
o
r
m
a
n
ce
.
Ma
cr
o
-
a
v
er
ag
ed
p
r
ec
is
io
n
,
r
ec
all,
F1
-
s
co
r
e,
an
d
ac
c
u
r
ac
y
wer
e
co
m
p
u
te
d
p
er
f
o
ld
,
with
f
in
al
p
er
f
o
r
m
an
ce
r
ep
o
r
ted
as
m
ea
n
±
s
tan
d
ar
d
d
e
v
iatio
n
ac
r
o
s
s
all
f
iv
e
f
o
ld
s
.
T
ab
le
1
.
Data
s
et
d
escr
ip
tio
n
b
ef
o
r
e
an
d
af
ter
a
u
g
m
e
n
tatio
n
M
a
i
z
e
V
a
r
i
e
t
y
I
n
i
t
i
a
l
D
a
t
a
s
e
t
S
i
z
e
A
u
g
m
e
n
t
e
d
D
a
t
a
s
e
t
S
i
z
e
B
h
i
h
i
l
i
f
a
1
,
3
2
8
6
,
4
8
0
S
a
n
z
a
l
S
i
m
a
8
5
0
5
,
1
0
0
W
a
n
g
D
a
t
a
a
1
,
0
2
4
6
,
1
4
4
Fig
u
r
e
4
.
C
lass
d
is
tr
ib
u
tio
n
o
f
B
h
ih
ilifa
,
San
za
l Sim
a,
an
d
W
an
g
Data
a
im
ag
es a
cr
o
s
s
tr
ain
i
n
g
,
v
alid
atio
n
,
an
d
test
s
p
lits
,
co
n
f
ir
m
in
g
c
o
n
s
is
ten
t c
las
s
p
r
o
p
o
r
tio
n
s
f
o
llo
win
g
s
tr
atif
ied
s
am
p
lin
g
2
.
2
.
3
.
Da
t
a
a
ug
m
ent
a
t
io
n
T
o
en
h
an
ce
d
ataset
d
iv
er
s
it
y
an
d
r
e
d
u
ce
o
v
er
f
itti
n
g
,
s
tan
d
ar
d
au
g
m
e
n
tatio
n
tr
an
s
f
o
r
m
atio
n
s
wer
e
ap
p
lied
u
s
in
g
Ke
r
as’s
I
m
ag
eDa
taGe
n
er
ato
r
:
r
o
tatio
n
_
r
a
n
g
e=
2
0
°,
wid
th
_
s
h
if
t_
r
an
g
e=
0
.
2
,
h
eig
h
t_
s
h
if
t_
r
a
n
g
e=
0
.
2
,
s
h
ea
r
_
r
an
g
e=
0
.
2
,
an
d
c
h
an
n
el_
s
h
if
t_
r
an
g
e=
2
0
.
T
h
ese
t
r
an
s
f
o
r
m
atio
n
s
s
im
u
late
r
ea
lis
tic
v
ar
iatio
n
s
in
s
ee
d
o
r
ien
tatio
n
,
s
ca
le,
an
d
illu
m
in
atio
n
en
co
u
n
ter
ed
d
u
r
in
g
p
r
ac
ti
ca
l
s
ee
d
h
an
d
lin
g
.
Af
ter
au
g
m
en
tatio
n
,
th
e
d
ataset
ex
p
an
d
ed
as
in
d
icate
d
in
T
ab
le
1
,
p
r
o
v
id
in
g
s
u
f
f
icie
n
t
v
ar
iab
ilit
y
f
o
r
r
o
b
u
s
t
m
o
d
el
tr
ain
in
g
.
2
.
2
.
4
.
Da
t
a
p
re
pro
ce
s
s
ing
All
im
ag
e
f
ile
p
ath
s
a
n
d
la
b
els
wer
e
co
n
v
er
te
d
to
Nu
m
Py
ar
r
a
y
s
to
s
tr
ea
m
lin
e
m
an
ip
u
latio
n
.
An
I
m
ag
eDa
taGe
n
er
ato
r
p
ip
e
lin
e
n
o
r
m
alize
d
p
ix
el
v
al
u
es
b
y
s
ca
lin
g
1
/2
5
5
t
o
th
e
[
0
,
1
]
r
an
g
e,
s
tab
ilizin
g
g
r
ad
ien
t
u
p
d
ate
s
d
u
r
in
g
b
ac
k
p
r
o
p
a
g
atio
n
.
C
lass
lab
els
we
r
e
o
n
e
-
h
o
t
en
co
d
ed
to
b
e
c
o
m
p
atib
le
with
th
e
ca
teg
o
r
ical
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
f
u
n
ctio
n
.
I
m
a
g
es
wer
e
r
esized
to
2
2
4
×2
2
4
×3
(
R
GB
)
p
ix
els
as
r
eq
u
ir
ed
b
y
th
e
p
r
e
-
tr
ain
ed
ar
ch
itectu
r
es.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
F
r
o
m
d
a
ta
to
d
ec
is
io
n
s
:
en
s
emb
le
d
ee
p
lea
r
n
in
g
fr
a
mewo
r
k
fo
r
ma
iz
e
s
ee
d
va
r
iety
…
(
E
mma
n
u
el
A
s
a
n
te
)
1653
2
.
3
.
E
v
a
lua
t
i
o
n m
et
rics
T
h
e
m
o
d
els
in
th
is
wo
r
k
ar
e
ev
alu
ated
u
s
in
g
f
o
u
r
w
id
ely
ad
o
p
ted
p
er
f
o
r
m
an
ce
in
d
icato
r
s
:
Acc
u
r
ac
y
(
1
)
,
Pre
cisi
o
n
(
2
)
,
R
ec
all
(
3
)
,
an
d
F1
-
s
co
r
e
(
4
)
.
T
h
ese
m
etr
ics
co
llectiv
ely
m
ea
s
u
r
e
th
e
co
r
r
ec
tn
ess
,
s
en
s
itiv
ity
,
an
d
b
alan
ce
o
f
t
h
e
class
if
icatio
n
p
er
f
o
r
m
an
c
e.
T
h
e
ev
al
u
atio
n
was
co
n
d
u
cted
u
s
in
g
5
-
f
o
l
d
cr
o
s
s
-
v
alid
atio
n
,
an
d
th
e
r
ep
o
r
ted
r
esu
lts
r
ep
r
esen
t
th
e
m
ea
n
±
s
tan
d
ar
d
d
ev
iatio
n
ac
r
o
s
s
f
o
ld
s
to
en
s
u
r
e
s
tatis
t
ical
r
eliab
ilit
y
an
d
r
o
b
u
s
tn
ess
o
f
th
e
en
s
em
b
le
m
o
d
el.
=
+
+
+
+
(
1
)
=
+
(
2
)
=
+
(
3
)
1
−
=
2
∗
∗
+
(
4
)
W
h
er
e
T
P,
T
N,
FP
,
an
d
FN
r
ep
r
esen
t
tr
u
e
p
o
s
itiv
e,
tr
u
e
n
e
g
ativ
e,
f
alse
p
o
s
itiv
e,
an
d
f
alse
n
eg
ativ
e,
r
esp
ec
tiv
ely
.
2
.
4
.
M
o
del dev
elo
pm
ent
2
.
4
.
1
.
B
a
s
eline
m
o
del
A
b
aselin
e
C
N
N,
as
s
h
o
wn
in
Fig
u
r
e
5
,
was
d
ev
elo
p
ed
to
p
r
o
v
id
e
a
p
er
f
o
r
m
a
n
ce
b
en
c
h
m
ar
k
f
o
r
co
m
p
ar
is
o
n
with
th
e
tr
a
n
s
f
er
lear
n
in
g
a
n
d
en
s
em
b
le
ap
p
r
o
ac
h
es.
T
h
e
ar
ch
itectu
r
e
co
n
s
is
ts
o
f
two
co
n
v
o
l
u
tio
n
al
b
lo
ck
s
,
ea
ch
c
o
n
tain
in
g
a
C
o
n
v
2
D
lay
er
(
3
×3
k
er
n
els,
R
eL
U
ac
tiv
atio
n
)
,
a
Ma
x
Po
o
lin
g
2
D
lay
er
f
o
r
s
p
atial
d
o
wn
-
s
am
p
lin
g
,
an
d
a
D
r
o
p
o
u
t
lay
er
to
m
itig
ate
o
v
er
f
itti
n
g
.
T
h
e
f
ir
s
t
b
lo
ck
u
s
es
3
2
f
ilter
s
an
d
th
e
s
ec
o
n
d
u
s
es
6
4
.
Af
te
r
th
e
co
n
v
o
l
u
tio
n
al
b
lo
c
k
s
,
a
g
lo
b
al
av
er
ag
e
p
o
o
lin
g
(
GAP)
lay
er
r
ed
u
ce
s
th
e
s
p
atial
d
im
en
s
io
n
s
,
f
o
llo
wed
b
y
a
Den
s
e
lay
er
(
1
2
8
u
n
its
,
R
eL
U)
,
a
Dr
o
p
o
u
t la
y
er
,
an
d
a
s
o
f
tm
ax
o
u
tp
u
t la
y
er
with
3
u
n
its
co
r
r
esp
o
n
d
in
g
t
o
th
e
3
m
aize
v
ar
ieties.
T
h
e
m
o
d
el
was
co
m
p
iled
with
t
h
e
Ad
am
o
p
tim
izer
(
lear
n
in
g
r
ate=
0
.
0
0
1
)
an
d
ca
teg
o
r
ical
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
.
T
r
ain
in
g
u
s
ed
b
atc
h
s
ize
5
,
5
0
ep
o
ch
s
,
an
d
ea
r
l
y
s
to
p
p
in
g
with
p
atien
c
e=
5
.
T
h
e
b
aselin
e
m
o
d
el
s
er
v
e
d
th
r
e
e
p
u
r
p
o
s
es:
(
i)
estab
lis
h
in
g
a
p
er
f
o
r
m
an
ce
l
o
wer
b
o
u
n
d
;
(
ii)
v
alid
atin
g
d
atas
et
q
u
ality
an
d
lear
n
ab
ilit
y
;
an
d
(
iii)
g
u
id
in
g
s
u
b
s
eq
u
en
t
ar
ch
itectu
r
al
an
d
h
y
p
er
p
ar
am
eter
d
ec
is
io
n
s
.
Fig
u
r
e
5
.
B
aselin
e
C
NN
ar
ch
itectu
r
e
f
o
r
Gh
an
aian
m
aize
s
ee
d
v
ar
iety
class
if
icatio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
1
5
,
No
.
4
,
D
e
c
em
b
er
20
2
6
:
1
6
4
8
-
1
6
6
3
1654
2
.
4
.
2
.
P
re
-
t
r
a
ined
m
o
dels
Acc
ess
to
lar
g
e,
well
-
an
n
o
tated
ag
r
icu
ltu
r
al
d
atasets
is
lim
ited
in
m
an
y
Af
r
ican
co
n
tex
t
s
,
m
ak
in
g
tr
an
s
f
er
lear
n
in
g
a
h
ig
h
ly
s
u
itab
le
s
tr
ateg
y
.
T
r
an
s
f
e
r
lear
n
in
g
en
a
b
les
m
o
d
els
p
r
e
-
tr
ai
n
ed
o
n
I
m
ag
eNe
t
[
2
1
]
to
ad
ap
t
th
eir
lear
n
e
d
f
ea
tu
r
e
r
e
p
r
esen
tatio
n
s
to
s
m
aller
,
d
o
m
ain
-
s
p
ec
if
ic
d
atasets
s
u
ch
as
m
aize
s
ee
d
im
ag
es,
s
u
b
s
tan
tially
r
ed
u
cin
g
t
r
ain
in
g
tim
e,
p
r
ev
e
n
tin
g
o
v
er
f
itti
n
g
,
a
n
d
im
p
r
o
v
in
g
g
en
e
r
aliza
tio
n
[
2
2
]
.
T
h
r
ee
p
r
e
-
tr
ain
ed
ar
ch
itectu
r
e
s
wer
e
s
elec
ted
f
o
r
th
ei
r
co
m
p
lem
en
tar
y
s
tr
en
g
th
s
in
v
is
u
al
f
ea
tu
r
e
ex
tr
ac
tio
n
:
(
i)
Mo
b
ileNetV2
[
2
3
]
,
ch
o
s
en
f
o
r
its
s
u
itab
ilit
y
f
o
r
m
o
b
ile
an
d
em
b
ed
d
e
d
s
y
s
tem
s
v
ia
in
v
er
ted
r
esid
u
als
an
d
lin
ea
r
b
o
ttlen
ec
k
s
;
(
ii)
Den
s
eNe
t1
6
9
[
1
7
]
,
s
e
lecte
d
f
o
r
its
d
en
s
e
co
n
n
ec
tiv
ity
th
at
en
co
u
r
ag
es
f
ea
tu
r
e
r
e
u
s
e
an
d
ef
f
icien
t
g
r
ad
ien
t
f
lo
w;
an
d
(
iii
)
Xce
p
tio
n
[
1
8
]
,
ch
o
s
en
f
o
r
its
d
ep
th
wis
e
s
ep
ar
ab
le
co
n
v
o
l
u
tio
n
s
th
at
im
p
r
o
v
e
r
e
p
r
esen
tatio
n
al
ca
p
ac
ity
wh
ile
r
ed
u
cin
g
c
o
m
p
u
tatio
n
al
co
s
t.
All
th
r
ee
h
av
e
b
ee
n
ex
ten
s
iv
ely
v
alid
ated
o
n
s
ee
d
,
cr
o
p
,
a
n
d
p
la
n
t c
lass
if
icatio
n
task
s
[
1
]
,
[
8
]
,
[
2
2
]
.
Fin
e
-
tu
n
in
g
was
p
er
f
o
r
m
ed
b
y
f
r
ee
zi
n
g
t
h
e
co
n
v
o
l
u
tio
n
al
b
ase
o
f
ea
ch
m
o
d
el
to
p
r
eser
v
e
lear
n
e
d
lo
w
-
lev
el
f
ea
tu
r
es,
an
d
ap
p
e
n
d
in
g
task
-
s
p
ec
if
ic
lay
er
s
:
a
GAP
lay
er
,
a
Den
s
e
lay
er
(
1
2
8
u
n
its
,
R
eL
U)
,
a
Dr
o
p
o
u
t
lay
er
(
r
ate=
0
.
2
)
,
a
n
d
a
th
r
ee
-
class
s
o
f
tm
ax
o
u
t
p
u
t
lay
er
.
T
r
ain
in
g
h
y
p
er
p
ar
a
m
eter
s
wer
e
co
n
s
is
ten
t
ac
r
o
s
s
all
m
o
d
els:
b
atch
s
ize=
5
,
in
itial
lear
n
in
g
r
ate=
0
.
0
0
1
,
m
ax
im
u
m
5
0
ep
o
c
h
s
,
ea
r
ly
s
to
p
p
in
g
(
p
atien
ce
=5
)
,
a
n
d
R
ed
u
ce
L
R
On
Plateau
s
ch
ed
u
lin
g
(
f
ac
to
r
=0
.
2
,
m
in
im
u
m
lear
n
in
g
r
ate=
1
×1
0
⁻⁴)
to
ad
a
p
t
th
e
lear
n
in
g
r
ate
w
h
en
v
alid
atio
n
l
o
s
s
p
latea
u
ed
.
T
h
e
f
in
e
-
tu
n
i
n
g
p
r
o
ce
s
s
is
illu
s
tr
ated
in
Fig
u
r
e
6
.
Fig
u
r
e
6
.
Fin
e
-
tu
n
in
g
p
ip
elin
e
f
o
r
th
e
p
r
e
-
tr
ain
e
d
tr
an
s
f
er
lear
n
in
g
a
r
ch
itectu
r
es
(
Den
s
eNe
t1
6
9
,
Mo
b
ileNetV2
,
an
d
Xce
p
tio
n
)
2
.
5
.
P
r
o
po
s
ed
ens
em
ble a
pp
ro
a
ch
An
en
s
em
b
le
lear
n
in
g
s
tr
ateg
y
was
im
p
lem
en
ted
to
co
m
b
i
n
e
th
e
p
r
ed
ictiv
e
s
tr
en
g
t
h
s
o
f
m
u
ltip
le
f
in
e
-
tu
n
e
d
m
o
d
els.
T
h
e
s
tack
in
g
tech
n
iq
u
e
a
g
g
r
e
g
ated
th
e
p
r
o
b
ab
ilis
tic
o
u
tp
u
ts
f
r
o
m
th
e
t
wo
b
est
-
p
er
f
o
r
m
in
g
in
d
iv
id
u
al
m
o
d
els,
Den
s
eNe
t1
6
9
an
d
Xce
p
tio
n
,
in
to
a
m
eta
-
m
o
d
el
th
at
g
en
er
ates
th
e
f
in
al
p
r
ed
ictio
n
as
s
h
o
wn
in
Fig
u
r
e
7
.
T
h
is
ap
p
r
o
ac
h
e
n
h
an
ce
s
ac
cu
r
ac
y
a
n
d
r
o
b
u
s
tn
ess
b
y
lev
e
r
ag
in
g
th
e
co
m
p
lem
en
tar
y
lear
n
in
g
p
atter
n
s
o
f
th
e
two
ar
ch
itectu
r
es.
E
ac
h
f
in
e
-
tu
n
ed
m
o
d
el
p
r
o
d
u
c
ed
a
p
r
o
b
ab
ilit
y
d
is
tr
ib
u
tio
n
o
v
er
th
e
th
r
ee
clas
s
es.
T
h
ese
o
u
tp
u
ts
wer
e
co
n
ca
ten
ated
to
f
o
r
m
a
s
in
g
le
f
ea
tu
r
e
v
ec
to
r
,
wh
ic
h
was th
en
p
ass
ed
to
a
m
eta
-
class
if
ier
co
n
s
is
tin
g
o
f
a
Den
s
e
lay
er
(
R
eL
U
ac
tiv
atio
n
)
,
a
Dr
o
p
o
u
t la
y
er
(
r
ate=
0
.
2
)
,
an
d
a
s
o
f
tm
ax
o
u
tp
u
t la
y
e
r
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
F
r
o
m
d
a
ta
to
d
ec
is
io
n
s
:
en
s
emb
le
d
ee
p
lea
r
n
in
g
fr
a
mewo
r
k
fo
r
ma
iz
e
s
ee
d
va
r
iety
…
(
E
mma
n
u
el
A
s
a
n
te
)
1655
Fig
u
r
e
7
.
Pro
p
o
s
ed
e
n
s
em
b
le
ar
ch
ite
ctu
r
e
co
m
b
in
in
g
Den
s
eNe
t1
6
9
an
d
Xce
p
tio
n
v
ia
f
ea
tu
r
e
co
n
ca
ten
atio
n
an
d
a
s
tack
ed
m
eta
-
class
if
ier
T
h
e
en
s
em
b
le
p
r
o
ce
s
s
is
f
o
r
m
alize
d
as
(
5
)
-
(
1
0
)
:
=
(
(
I
)
)
(
5
)
169
=
(
169
(
I
)
)
(
6
)
=
c
on
c
a
t
(
,
169
)
(
7
)
=
R
e
L
U
(
,
)
(
8
)
=
(
,
=
0
.
2
)
(
9
)
=
s
oft
ma
x
(
,
)
(
1
0
)
W
h
er
e
,
I
: I
n
p
u
t
ten
s
o
r
wi
th
s
h
ap
e
(
2
2
4
,
2
2
4
,
3
)
.
B
Xception
: T
h
e
b
ase
m
o
d
el
Xce
p
tio
n
with
o
u
t th
e
t
o
p
lay
er
s
.
B
DenseNet169
: T
h
e
b
ase
m
o
d
el
D
en
s
eNe
t1
6
9
with
o
u
t th
e
to
p
la
y
er
s
.
: T
h
e
GAP
o
p
er
atio
n
.
F
Xception
: O
u
tp
u
t f
ea
tu
r
e
m
ap
f
r
o
m
th
e
Xce
p
tio
n
b
ase
m
o
d
el
a
f
ter
GAP
.
F
DenseNet169
: O
u
tp
u
t f
ea
tu
r
e
m
a
p
f
r
o
m
th
e
Den
s
eNe
t1
6
9
b
ase
m
o
d
el
af
ter
GAP
.
F
concat
: Co
n
ca
ten
ated
f
ea
tu
r
e
v
ec
to
r
f
r
o
m
all
th
r
ee
b
ase
m
o
d
e
ls
.
W
d
:
W
eig
h
t
m
atr
ix
o
f
th
e
d
e
n
s
e
lay
er
with
R
eL
U
ac
tiv
at
io
n
.
,
b
d
:
B
ias
v
ec
to
r
o
f
th
e
d
e
n
s
e
lay
er
with
R
eL
U
ac
tiv
atio
n
.
D:
Ou
tp
u
t o
f
th
e
d
en
s
e
lay
er
with
R
eL
U
ac
tiv
atio
n
.
,
D
drop
: O
u
tp
u
t a
f
ter
a
p
p
ly
in
g
d
r
o
p
o
u
t
to
D.
W
o
: We
ig
h
t m
atr
ix
o
f
t
h
e
f
in
al
d
en
s
e
lay
er
(
o
u
tp
u
t la
y
er
)
wit
h
s
o
f
tm
ax
ac
tiv
atio
n
.
b
o
: Bi
as
v
ec
to
r
o
f
th
e
f
in
al
d
en
s
e
lay
er
(
o
u
tp
u
t la
y
er
)
with
s
o
f
tm
ax
ac
tiv
atio
n
.
P: Fin
al
class
p
r
o
b
ab
ilit
ies af
t
er
ap
p
ly
in
g
th
e
s
o
f
tm
a
x
f
u
n
cti
o
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
7
7
6
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
,
Vo
l.
1
5
,
No
.
4
,
D
e
c
em
b
er
20
2
6
:
1
6
4
8
-
1
6
6
3
1656
2
.
5
.
1
.
E
ns
em
ble
ca
nd
ida
t
e
s
e
lect
io
n
(
Abla
t
io
n)
An
ab
latio
n
s
tu
d
y
was
co
n
d
u
cted
to
id
en
tify
th
e
o
p
tim
al
en
s
em
b
le
co
n
f
ig
u
r
atio
n
.
All
p
a
ir
wis
e
an
d
tr
ip
le
co
m
b
in
atio
n
s
o
f
t
h
e
th
r
e
e
f
in
e
-
tu
n
e
d
m
o
d
els
wer
e
ev
al
u
ated
o
n
th
e
test
s
et.
As
s
u
m
m
ar
ized
in
T
a
b
le
2
,
th
e
Xce
p
tio
n
+
Den
s
eNe
t1
6
9
(
X+
D)
co
m
b
in
atio
n
ac
h
ie
v
ed
th
e
b
est
p
er
f
o
r
m
an
ce
a
cr
o
s
s
all
m
etr
ics
(
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
e
ca
ll,
F1
-
s
co
r
e
=
0
.
9
7
)
with
a
m
o
d
er
ate
tr
ain
ab
le
p
ar
a
m
eter
f
o
o
tp
r
in
t
o
f
3
.
1
8
MB,
ju
s
tify
in
g
its
s
elec
tio
n
f
o
r
th
e
p
r
o
p
o
s
ed
en
s
em
b
le.
T
ab
le
2
.
Ab
latio
n
s
tu
d
y
: p
e
r
f
o
r
m
an
ce
co
m
p
ar
is
o
n
o
f
en
s
em
b
le
co
n
f
ig
u
r
atio
n
s
o
n
test
d
ata
M
o
d
e
l
s
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F
1
-
sc
o
r
e
T
r
a
i
n
a
b
l
e
P
a
r
a
ms (
M
B
)
D
+
M
0
.
9
5
0
.
9
5
0
.
9
5
0
.
9
5
2
.
5
2
X
+
M
0
.
9
5
0
.
9
5
0
.
9
5
0
.
9
5
2
.
8
5
X
+
D
0
.
9
7
0
.
9
7
0
.
9
7
0
.
9
7
3
.
1
8
X
+
D
+
M
0
.
9
6
0
.
9
6
0
.
9
6
0
.
9
6
4
.
2
7
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
is
s
ec
tio
n
p
r
esen
ts
an
d
in
ter
p
r
ets
th
e
ex
p
er
im
e
n
tal
r
esu
lts
ac
r
o
s
s
f
o
u
r
s
u
b
-
s
tu
d
ies:
(
3
.
1
)
b
aselin
e
C
NN
p
er
f
o
r
m
a
n
ce
,
(
3
.
2
)
c
o
m
p
ar
ativ
e
e
v
alu
atio
n
o
f
in
d
iv
id
u
al
p
r
e
-
tr
ain
e
d
m
o
d
els
an
d
th
e
en
s
em
b
le,
(
3
.
3
)
class
-
wis
e
p
r
ed
ictio
n
an
aly
s
is
an
d
s
tatis
tical
v
ali
d
atio
n
,
an
d
(
3
.
4
)
ex
p
lain
ab
ilit
y
v
is
u
aliza
tio
n
.
E
ac
h
s
u
b
-
s
ec
tio
n
c
o
n
tex
tu
aliz
es f
in
d
in
g
s
r
elativ
e
to
t
h
e
s
tu
d
y
’
s
o
b
jectiv
es a
n
d
p
r
io
r
liter
atu
r
e.
3
.
1
.
B
a
s
eline
re
s
ults
Fig
u
r
e
8
p
r
esen
ts
th
e
b
aselin
e
C
NN’
s
tr
ain
in
g
cu
r
v
es
Fig
u
r
e
8
(
a)
an
d
co
n
f
u
s
io
n
m
atr
i
x
o
n
th
e
test
s
et
Fig
u
r
e
8
(
b
)
.
T
h
e
m
o
d
el
wa
s
tr
ain
ed
f
o
r
5
0
e
p
o
ch
s
an
d
ac
h
iev
ed
a
tr
ain
in
g
ac
c
u
r
ac
y
o
f
ap
p
r
o
x
im
ately
0
.
9
0
an
d
a
v
alid
atio
n
ac
cu
r
ac
y
o
f
a
p
p
r
o
x
im
ately
0
.
8
5
.
B
o
th
tr
ai
n
in
g
an
d
v
alid
atio
n
lo
s
s
es
d
ec
r
ea
s
ed
co
n
s
is
ten
tly
,
with
tr
ain
in
g
lo
s
s
co
n
v
er
g
in
g
to
ap
p
r
o
x
im
ately
0
.
2
5
.
T
h
e
clo
s
e
alig
n
m
en
t
o
f
th
e
ac
cu
r
a
cy
an
d
lo
s
s
cu
r
v
es
in
d
icate
s
ad
eq
u
ate
g
e
n
er
aliza
tio
n
with
o
u
t sig
n
s
o
f
s
ev
er
e
o
v
e
r
f
itti
n
g
.
T
h
e
b
aselin
e
C
NN
ac
h
i
ev
e
d
an
ac
cu
r
ac
y
o
f
8
5
%
o
n
th
e
test
s
et.
T
ab
le
3
s
h
o
ws
th
at
B
h
ih
ilifa
was
class
if
ied
n
ea
r
-
p
er
f
ec
tl
y
(
p
r
e
cisi
o
n
=1
.
0
0
,
r
ec
all=0
.
9
8
,
F1
=0
.
9
9
)
,
wh
er
ea
s
San
za
l
Sim
a
an
d
W
an
g
Data
a
ac
h
iev
ed
s
u
b
s
tan
tially
lo
wer
F1
-
s
co
r
es
o
f
0
.
7
2
a
n
d
0
.
7
7
,
r
esp
ec
tiv
ely
.
T
h
e
co
n
f
u
s
io
n
m
atr
ix
s
h
o
ws
th
at
6
5
3
San
za
l
Sima
im
ag
es
wer
e
co
r
r
ec
tly
class
if
ied
,
3
6
7
wer
e
m
is
class
if
ied
as
W
an
g
Data
a,
1
,
0
7
9
W
an
g
Data
a
im
ag
es
wer
e
co
r
r
ec
tly
id
e
n
tifie
d
,
an
d
1
5
0
wer
e
m
is
class
if
ie
d
as
San
za
l
Sima
.
T
h
is
p
atter
n
in
d
icate
s
th
at
t
h
e
two
v
ar
ieties s
h
ar
e
o
v
er
lap
p
i
n
g
m
o
r
p
h
o
lo
g
ical
ch
ar
ac
ter
is
tics
th
at
ch
allen
g
e
s
h
allo
w
f
ea
tu
r
e
ex
tr
ac
to
r
s
.
W
h
ile
th
e
b
aselin
e
d
em
o
n
s
tr
ated
th
at
m
ea
n
in
g
f
u
l
p
atter
n
s
ex
is
t
in
th
e
d
ataset,
its
lim
ited
d
is
cr
im
in
ativ
e
ca
p
ac
it
y
m
o
tiv
ated
th
e
u
s
e
o
f
d
ee
p
er
ar
ch
itectu
r
es.
(
a)
(b
)
Fig
u
r
e
8
.
B
aselin
e
tr
ain
in
g
c
u
r
v
e
(
a)
a
n
d
co
n
f
u
s
io
n
m
atr
ix
an
d
(
b
)
o
n
test
s
et
3
.
2
.
Co
m
pa
ra
t
iv
e
s
t
ud
y
o
f
d
ee
p lea
rning
a
pp
ro
a
ches
T
ab
le
3
a
n
d
Fig
u
r
e
9
s
u
m
m
a
r
ize
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
f
o
u
r
d
ee
p
lear
n
in
g
m
o
d
els,
De
n
s
eNe
t1
6
9
,
Mo
b
ileNetV2
,
Xce
p
tio
n
,
an
d
th
e
p
r
o
p
o
s
ed
en
s
em
b
le,
o
n
th
e
test
s
et.
Fig
u
r
e
9
s
h
o
ws
th
e
tr
ain
in
g
cu
r
v
es
f
o
r
ea
ch
m
o
d
el,
an
d
Fig
u
r
e
1
0
s
h
o
ws th
eir
co
r
r
esp
o
n
d
in
g
c
o
n
f
u
s
io
n
m
atr
ices.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J I
n
f
&
C
o
m
m
u
n
T
ec
h
n
o
l
I
SS
N:
2252
-
8
7
7
6
F
r
o
m
d
a
ta
to
d
ec
is
io
n
s
:
en
s
emb
le
d
ee
p
lea
r
n
in
g
fr
a
mewo
r
k
fo
r
ma
iz
e
s
ee
d
va
r
iety
…
(
E
mma
n
u
el
A
s
a
n
te
)
1657
Am
o
n
g
th
e
in
d
i
v
id
u
al
p
r
e
-
tr
ai
n
ed
ar
c
h
itectu
r
es,
Den
s
eNe
t1
6
9
ac
h
ie
v
ed
th
e
h
ig
h
est
p
er
f
o
r
m
an
c
e
at
9
4
%
ac
r
o
s
s
all
m
et
r
ics,
f
o
llo
wed
b
y
Xce
p
tio
n
at
9
3
%
an
d
Mo
b
ileNetV2
at
9
2
%.
All
th
r
ee
s
ig
n
if
ican
tl
y
o
u
tp
er
f
o
r
m
ed
t
h
e
b
aselin
e
C
NN
(
8
5
%),
co
n
f
ir
m
in
g
th
e
b
en
ef
it
o
f
I
m
ag
eNe
t
-
p
r
etr
ain
ed
f
ea
tu
r
e
r
ep
r
esen
tatio
n
s
f
o
r
d
o
m
ain
-
s
p
ec
if
ic
class
if
icatio
n
task
s
.
De
n
s
eNe
t1
6
9
ac
h
iev
ed
th
e
m
o
s
t
s
tab
le
an
d
co
n
s
is
ten
t
tr
ain
in
g
b
eh
av
i
o
r
,
with
tr
ain
in
g
an
d
v
alid
atio
n
ac
cu
r
ac
ies
co
n
v
e
r
g
in
g
n
ea
r
0
.
9
4
an
d
s
tead
ily
d
ec
r
ea
s
in
g
lo
s
s
es
(
Fig
u
r
e
9
(
a
)
)
.
I
ts
co
n
f
u
s
i
o
n
m
atr
ix
(
Fig
u
r
e
10
(
a
)
)
s
h
o
ws
1
,
2
9
0
co
r
r
ec
tl
y
class
if
ied
B
h
ih
ilifa
,
9
1
8
San
za
l
Sima
,
an
d
1
,
1
2
8
W
an
g
Data
a
s
am
p
les,
r
ef
lectin
g
s
tr
o
n
g
d
is
cr
im
in
at
iv
e
ab
ilit
y
d
u
e
to
its
d
en
s
e
co
n
n
ec
tiv
ity
an
d
f
ea
tu
r
e
r
eu
s
e
m
ec
h
an
is
m
.
Mo
b
ileNetV2
r
ea
ch
e
d
a
f
in
al
ac
cu
r
ac
y
o
f
0
.
9
2
af
ter
2
5
ep
o
ch
s
(
Fig
u
r
e
9
(
b
)
)
,
with
m
o
r
e
m
is
class
if
ic
atio
n
s
b
etwe
en
San
za
l
Sima
an
d
W
an
g
Data
a
(
Fig
u
r
e
1
0
(
b
)
)
,
co
n
s
is
ten
t
with
its
lig
h
tweig
h
t
d
esig
n
,
w
h
ich
p
r
io
r
itizes
co
m
p
u
tatio
n
al
e
f
f
i
cien
cy
o
v
er
r
ep
r
esen
tatio
n
al
d
ep
th
.
Xce
p
tio
n
attain
ed
0
.
9
3
af
ter
2
0
ep
o
ch
s
(
Fig
u
r
e
9
(
c)
)
,
ac
h
iev
i
n
g
im
p
r
o
v
ed
class
s
ep
ar
at
io
n
o
v
e
r
Mo
b
ileNetV2
th
r
o
u
g
h
its
d
ep
th
wis
e
s
ep
ar
ab
le
co
n
v
o
l
u
tio
n
s
(
Fig
u
r
e
1
0
(
c
)
).
T
h
e
p
r
o
p
o
s
ed
en
s
em
b
le
f
u
r
t
h
er
im
p
r
o
v
ed
p
e
r
f
o
r
m
an
ce
t
o
0
.
9
7
ac
r
o
s
s
all
m
etr
ics
(
F
ig
u
r
e
9
(
d
)
,
Fig
u
r
e
1
0
(
d
)
)
.
T
h
is
im
p
r
o
v
e
m
en
t
r
ef
lects
th
e
p
r
i
n
cip
le
o
f
en
s
em
b
le
d
iv
er
s
ity
:
Den
s
eNe
t1
6
9
’
s
d
e
n
s
e
f
ea
tu
r
e
p
r
o
p
a
g
atio
n
a
n
d
Xce
p
tio
n
’
s
e
f
f
icien
t
s
p
atial
s
ep
ar
atio
n
ca
p
tu
r
e
co
m
p
lem
en
tar
y
d
is
cr
im
in
ativ
e
p
atter
n
s
,
a
n
d
th
eir
co
m
b
in
atio
n
r
e
d
u
ce
s
in
d
iv
id
u
al
m
o
d
el
b
iases
,
p
ar
ticu
lar
ly
f
o
r
th
e
San
za
l
Sima
–
W
an
g
Data
a
co
n
f
u
s
io
n
th
at
p
er
s
is
ted
in
s
in
g
le
m
o
d
els.
T
h
e
en
s
em
b
le
ac
h
iev
ed
1
,
2
9
1
co
r
r
ec
t
B
h
ih
ilifa
,
9
5
3
San
z
al
Sima
,
an
d
1
,
1
9
5
W
an
g
Data
a
class
if
icatio
n
s
,
w
ith
v
er
y
f
ew
er
r
o
r
s
ac
r
o
s
s
class
es
.
T
ab
le
3
.
C
o
m
p
a
r
is
o
n
o
f
DL
m
o
d
els u
s
in
g
weig
h
ted
-
av
er
a
g
e
Acc
u
r
ac
y
,
Pre
cisi
o
n
,
R
ec
all,
a
n
d
F1
-
Sco
r
e
o
n
test
d
ata
No
M
o
d
e
l
A
c
c
u
r
a
c
y
P
r
e
c
i
s
i
o
n
R
e
c
a
l
l
F1
-
sc
o
r
e
1
M
o
b
i
l
e
N
e
t
v
2
92
92
92
92
2
X
c
e
p
t
i
o
n
93
93
93
93
3
D
e
n
seN
e
t
1
6
9
94
94
94
94
4
En
se
mb
l
e
97
97
97
97
(
a)
(
b
)
(
c)
(
d
)
Fig
u
r
e
9
.
T
r
ain
in
g
i
n
s
ig
h
ts
o
f
(
a)
Mo
b
ileNetv
2
,
(
b
)
Xce
p
tio
n
,
(
c)
Den
s
eNe
t1
6
9
,
an
d
(
d
)
e
n
s
em
b
le
m
o
d
el
3
.
3
.
Cla
s
s
-
wis
e
predict
io
ns
a
nd
s
t
a
t
is
t
ica
l v
a
lid
a
t
io
n
P
e
r
-
c
l
as
s
c
la
s
s
i
f
i
ca
t
i
o
n
r
e
p
o
r
t
f
o
r
a
l
l
m
o
d
e
l
s
,
a
s
s
h
o
w
n
i
n
T
a
b
l
e
4
,
i
n
d
i
c
a
t
e
s
t
h
a
t
t
h
e
p
r
o
g
r
e
s
s
i
v
e
i
m
p
r
o
v
e
m
e
n
t
a
c
r
o
s
s
a
r
c
h
i
t
e
ctu
r
e
s
is
c
o
n
s
i
s
t
e
n
t
a
n
d
i
n
t
e
r
p
r
e
t
a
b
l
e
.
T
h
e
b
a
s
el
i
n
e
C
N
N
ac
h
i
e
v
e
d
n
e
a
r
-
p
e
r
f
e
c
t
Evaluation Warning : The document was created with Spire.PDF for Python.