I
AE
S In
t
er
na
t
io
na
l J
o
urna
l o
f
Art
if
icia
l In
t
ellig
ence
(
I
J
-
AI
)
Vo
l.
1
5
,
No
.
2
,
A
p
r
il 2
0
2
6
,
p
p
.
1
7
1
9
~
1
7
3
2
I
SS
N:
2
2
5
2
-
8
9
3
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijai.v
15
.i
2
.
p
p
1
7
1
9
-
1
7
3
2
1719
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
Co
mpa
ra
tive dee
p learning
stu
dy
f
o
r downy
mildew
detec
tion in
v
eg
etables
Su
pree
t
ha
Sh
iv
a
ra
j
,
M
a
njul
a
Su
nk
a
da
k
a
t
t
e
H
a
la
da
pp
a
D
e
p
a
r
t
me
n
t
o
f
C
o
mp
u
t
e
r
S
c
i
e
n
c
e
E
n
g
i
n
e
e
r
i
n
g
,
U
n
i
v
e
r
s
i
t
y
o
f
V
i
s
v
e
s
v
a
r
a
y
a
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
(
B
a
n
g
a
l
o
r
e
U
n
i
v
e
r
s
i
t
y
)
,
B
e
n
g
a
l
u
r
u
,
I
n
d
i
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
No
v
1
1
,
2
0
2
4
R
ev
is
ed
J
an
2
2
,
2
0
2
6
Acc
ep
ted
Feb
6
,
2
0
2
6
S
e
v
e
ra
l
v
e
g
e
tab
le
c
ro
p
s
a
re
a
ffe
c
ted
b
y
d
o
wn
y
m
il
d
e
w
,
a
m
a
jo
r
f
o
li
a
r
d
ise
a
se
re
su
lt
in
g
i
n
n
o
tab
le
re
d
u
c
ti
o
n
s
in
y
iel
d
.
F
o
r
su
sta
in
a
b
le
a
g
ricu
lt
u
re
a
n
d
d
ise
a
se
p
re
v
e
n
ti
o
n
,
e
a
rly
a
n
d
p
re
c
ise
d
e
tec
ti
o
n
is
c
ru
c
ial.
To
b
e
a
b
le
t
o
d
e
tec
t
d
o
wn
y
m
il
d
e
w
in
fiv
e
v
a
r
ied
v
e
g
e
tab
les
—
b
it
ter
g
o
u
r
d
,
b
o
t
tl
e
g
o
u
r
d
,
c
a
u
li
flo
we
r,
c
u
c
u
m
b
e
r
(Ra
sh
id
),
a
n
d
c
u
c
u
m
b
e
r
(S
u
lt
a
n
a
)
—
t
h
is
stu
d
y
e
v
a
lu
a
tes
th
re
e
d
e
e
p
lea
rn
i
n
g
a
rc
h
it
e
c
tu
re
s:
VG
G
1
9
,
De
n
se
Ne
t2
0
1
,
a
n
d
M
o
b
i
leN
e
tV2
.
T
h
is
wo
r
k
fo
c
u
se
s
o
n
imb
a
lan
c
e
d
d
a
tas
e
ts
c
o
ll
e
c
ted
fro
m
se
v
e
ra
l
s
o
u
rc
e
s,
in
o
p
p
o
siti
o
n
t
o
p
r
io
r
wo
rk
t
h
a
t
d
e
p
e
n
d
e
d
o
n
b
a
lan
c
e
d
lab
o
ra
to
r
y
d
a
tas
e
ts.
Ac
c
u
ra
c
y
,
p
r
e
c
isio
n
,
re
c
a
ll
,
a
n
d
F
1
-
sc
o
re
m
e
tri
c
s
we
re
u
se
d
to
e
v
a
l
u
a
te
th
e
m
o
d
e
ls
sh
o
rtl
y
a
fter
th
e
y
we
re
train
e
d
u
si
n
g
tran
sfe
r
lea
rn
in
g
,
d
a
ta
a
u
g
m
e
n
tatio
n
,
a
n
d
5
-
f
o
ld
c
ro
ss
-
v
a
li
d
a
ti
o
n
.
M
o
d
e
l
f
o
c
u
s
re
g
io
n
s
we
re
a
ss
e
ss
e
d
b
y
u
sin
g
g
ra
d
ien
t
-
we
ig
h
ted
c
las
s
a
c
ti
v
a
ti
o
n
m
a
p
p
in
g
(G
ra
d
-
CAM)
v
isu
a
li
z
a
ti
o
n
s,
a
n
d
sta
ti
stica
l
re
li
a
b
il
it
y
wa
s
a
ss
e
ss
e
d
b
a
se
d
o
n
p
a
ired
t
-
tes
ts
a
n
d
W
il
c
o
x
o
n
sig
n
e
d
-
ra
n
k
tes
ts.
By
a
c
h
iev
i
n
g
m
e
a
n
a
c
c
u
ra
c
ies
ab
o
v
e
9
8
%
a
n
d
sta
ti
stica
ll
y
si
g
n
ifi
c
a
n
t
re
su
lt
s
(p
<
0
.
0
5
)
o
n
c
u
c
u
m
b
e
r
d
a
tas
e
ts,
De
n
se
Ne
t2
0
1
a
c
c
o
m
p
li
sh
e
d
su
p
e
rio
r
p
e
rfo
rm
a
n
c
e
.
De
sp
it
e
a
tt
a
in
in
g
sli
g
h
t
ly
lo
we
r
a
c
c
u
ra
c
y
(
8
9
.
6
–
1
0
0
%
)
,
M
o
b
il
e
Ne
tV2
o
ffe
re
d
th
e
sm
a
ll
e
st
m
o
d
e
l
siz
e
(1
2
.
9
M
B)
a
n
d
m
in
im
u
m
in
fe
re
n
c
e
ti
m
e
(8
5
m
s).
Th
e
p
ro
p
o
se
d
a
p
p
ro
a
c
h
d
e
m
o
n
stra
ted
a
tran
sp
a
re
n
t,
g
e
n
e
ra
li
z
a
b
le,
a
n
d
c
o
m
p
u
tati
o
n
a
ll
y
e
fficie
n
t
d
e
e
p
le
a
rn
in
g
p
i
p
e
li
n
e
f
o
r
p
re
c
isio
n
a
g
r
icu
lt
u
re
’s
re
a
l
-
ti
m
e
d
o
wn
y
m
il
d
e
w
d
e
tec
ti
o
n
.
K
ey
w
o
r
d
s
:
Den
s
eNe
t2
0
1
Do
wn
y
m
ild
ew
E
x
p
l
a
i
n
a
b
l
e
a
r
ti
f
i
c
i
al
i
n
t
el
l
i
g
e
n
c
e
Gr
ad
-
C
AM
I
m
b
alan
ce
d
d
ataset
Mo
b
ileNetV2
VGG1
9
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
:
Su
p
r
ee
th
a
Sh
iv
ar
aj
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
E
n
g
in
ee
r
in
g
Un
iv
er
s
ity
o
f
Vis
v
esv
ar
ay
a
C
o
lleg
e
o
f
E
n
g
in
ee
r
i
n
g
(
B
an
g
al
o
r
e
Un
iv
er
s
ity
)
B
en
g
alu
r
u
,
5
6
0
0
0
1
,
Ka
r
n
atak
a
,
I
n
d
ia
E
m
ail: su
p
r
ee
th
a.
s
3
1
9
1
@
g
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
Plan
t
d
is
ea
s
es
ar
e
a
m
ajo
r
th
r
e
at
to
ag
r
icu
ltu
r
al
s
u
s
tain
ab
ilit
y
an
d
g
lo
b
al
f
o
o
d
s
ec
u
r
ity
,
an
d
ea
r
ly
an
d
ac
cu
r
ate
d
etec
tio
n
is
v
ital
f
o
r
y
ield
p
r
o
tectio
n
.
Am
o
n
g
th
es
e,
o
n
e
o
f
th
e
m
o
s
t
d
estru
ctiv
e
f
o
liar
d
is
ea
s
es
th
at
af
f
ec
ts
a
v
ar
iety
o
f
cr
o
p
s
,
i
n
clu
d
in
g
cr
u
cif
er
o
u
s
v
eg
etab
les
an
d
cu
cu
r
b
its
,
is
d
o
wn
y
m
ild
ew
.
I
n
h
u
m
id
co
n
d
itio
n
s
,
th
e
in
f
ec
tio
n
s
p
r
e
ad
s
r
ap
id
ly
,
lea
d
in
g
t
o
n
ec
r
o
tic
lesi
o
n
s
,
ch
lo
r
o
s
is
,
an
d
s
ig
n
if
ican
t
y
ield
lo
s
s
if
n
o
t a
d
d
r
ess
ed
[
1
]
–
[
3
]
.
Dis
ea
s
e
d
iag
n
o
s
is
h
as
tr
ad
itio
n
ally
r
elie
d
o
n
v
is
u
al
s
y
m
p
to
m
r
ec
o
g
n
itio
n
,
lab
o
r
ato
r
y
a
s
s
ay
s
,
an
d
m
an
u
al
in
s
p
ec
tio
n
b
y
tr
ain
ed
ag
r
o
n
o
m
is
ts
.
Ho
wev
er
,
th
ese
m
eth
o
d
s
ar
e
s
u
b
jectiv
e,
ti
m
e
-
co
n
s
u
m
in
g
,
an
d
o
f
ten
im
p
r
ac
tical
f
o
r
lar
g
e
-
s
ca
le
o
r
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
s
ettin
g
s
.
R
ec
en
t
d
ev
elo
p
m
en
ts
in
d
ee
p
lear
n
i
n
g
an
d
co
m
p
u
ter
v
is
io
n
[
4
]
–
[
6
]
h
a
v
e
s
ig
n
if
ican
tly
r
ev
o
lu
tio
n
iz
ed
p
lan
t
p
ath
o
lo
g
y
b
y
allo
win
g
th
e
au
t
o
m
atic
d
etec
tio
n
o
f
p
lan
t
d
is
ea
s
es
d
ir
ec
tly
f
r
o
m
im
ag
es.
Sh
allo
w
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etw
o
r
k
s
(
C
NNs)
wer
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
7
1
9
-
1
7
3
2
1720
em
p
lo
y
ed
i
n
ea
r
ly
m
o
d
els
f
o
r
d
is
ea
s
e
d
etec
tio
n
[
7
]
–
[
9
]
,
b
u
t
th
ese
tech
n
iq
u
es
s
u
f
f
er
ed
f
r
o
m
o
v
er
f
itti
n
g
an
d
lo
w
g
en
er
aliza
tio
n
o
n
r
ea
l
-
wo
r
ld
d
atasets
.
Ho
wev
er
,
ar
ch
itectu
r
es
s
u
ch
as
VGG1
9
,
Den
s
eNe
t2
0
1
,
an
d
Mo
b
ileNetV2
em
er
g
e
d
later
,
p
r
o
v
id
i
n
g
r
o
b
u
s
t
f
ea
tu
r
e
ex
tr
a
ctio
n
,
h
i
g
h
class
if
icatio
n
ac
cu
r
ac
y
,
an
d
e
f
f
icien
cy
,
m
ak
in
g
it
s
u
itab
le
to
b
e
d
ep
lo
y
ed
o
n
m
o
b
ile
an
d
in
te
r
n
et
o
f
t
h
in
g
s
(
I
o
T
)
-
b
ased
s
y
s
tem
s
[
1
0
]
–
[
1
3
]
.
I
n
ter
p
r
etab
ilit
y
is
an
o
th
er
k
ey
ch
allen
g
e.
Ad
o
p
tio
n
in
ag
r
ic
u
ltu
r
e
r
e
q
u
ir
es
tr
an
s
p
ar
e
n
cy
(
e.
g
.
,
“
wh
ic
h
p
ar
ts
o
f
th
e
leaf
led
to
th
e
d
ec
is
io
n
?”
)
,
esp
ec
ially
wh
en
d
ep
l
o
y
in
g
m
o
d
els
f
o
r
c
r
itical
cr
o
p
d
is
ea
s
e
m
o
n
ito
r
in
g
.
Hig
h
class
if
icatio
n
ac
cu
r
ac
y
al
o
n
e
is
n
o
t
s
u
f
f
icien
t.
As
a
g
o
al
to
e
n
s
u
r
e
i
n
ter
p
r
etab
ilit
y
an
d
tr
u
s
t,
e
x
p
lain
a
b
le
ar
tific
ial
in
tellig
en
ce
(
XAI
)
tech
n
iq
u
es
s
u
ch
as
g
r
ad
ie
n
t
-
weig
h
ted
class
ac
tiv
atio
n
m
ap
p
in
g
(
Gr
a
d
-
C
AM
)
h
av
e
b
ee
n
ad
o
p
ted
in
c
r
ea
s
in
g
ly
to
v
is
u
alize
s
ig
n
if
ican
t
r
eg
io
n
s
in
f
lu
en
cin
g
m
o
d
el
p
r
ed
ictio
n
s
[
1
4
]
–
[
1
6
]
.
Ho
wev
er
,
I
o
T
o
f
p
r
io
r
s
tu
d
ie
s
f
o
cu
s
ed
p
r
im
ar
ily
o
n
ac
cu
r
ac
y
-
b
ased
ev
alu
atio
n
s
u
s
in
g
b
alan
ce
d
lab
o
r
at
o
r
y
d
atasets
,
wh
ich
o
f
ten
f
ail
to
g
en
er
alize
ef
f
ec
tiv
ely
u
n
d
er
r
ea
l
-
tim
e,
u
n
b
alan
ce
d
d
atasets
[
1
7
]
.
B
u
ild
in
g
o
n
th
e
g
ap
s
id
en
tifie
d
in
p
r
io
r
r
esear
c
h
,
th
e
p
r
esen
t stu
d
y
aim
s
to
p
u
r
s
u
e
f
o
u
r
p
r
im
ar
y
o
b
jectiv
es:
−
I
n
f
iv
e
v
eg
etab
le
s
p
ec
ies
(
b
it
ter
g
o
u
r
d
,
b
o
ttle
g
o
u
r
d
,
ca
u
li
f
lo
wer
,
cu
c
u
m
b
er
(
R
ash
id
)
,
a
n
d
cu
c
u
m
b
er
(
Su
ltan
a)
,
we
ev
al
u
ate
th
r
e
e
p
r
o
m
i
n
en
t
d
ee
p
-
lear
n
i
n
g
ar
ch
itectu
r
es
(
VGG1
9
,
M
o
b
ileNetV2
,
an
d
Den
s
eNe
t2
0
1
)
o
n
lab
o
r
at
o
r
y
a
n
d
r
ea
l
-
wo
r
ld
,
im
b
ala
n
ce
d
d
at
asets
co
llected
f
r
o
m
d
iv
er
s
e
s
o
u
r
ce
s
.
−
T
o
q
u
a
n
tify
m
o
d
el
g
en
er
ali
za
tio
n
u
n
d
er
im
b
alan
ce
d
c
o
n
d
itio
n
s
,
we
ap
p
ly
in
d
e
p
e
n
d
en
t
test
-
s
et
ev
alu
atio
n
an
d
r
ig
o
r
o
u
s
5
-
f
o
l
d
cr
o
s
s
-
v
alid
atio
n
.
−
Statis
t
ical
s
ig
n
if
ican
ce
test
in
g
(
p
air
ed
t
-
test
s
an
d
[
1
8
]
–
[
2
0
]
W
ilco
x
o
n
test
s
)
is
co
n
d
u
cted
to
r
eliab
ly
co
m
p
ar
e
m
o
d
el
p
er
f
o
r
m
a
n
ce
,
an
d
we
in
teg
r
ate
ex
p
lain
a
b
ilit
y
v
ia
Gr
ad
-
C
AM
to
v
is
u
alize
h
o
w
m
o
d
els
atten
d
to
d
is
ea
s
e
lesi
o
n
s
.
−
I
n
f
er
en
ce
tim
e
an
d
m
o
d
el
s
ize
ar
e
ass
ess
ed
,
wh
ich
ar
e
cr
itical
f
o
r
f
ield
-
lev
el
o
r
m
o
b
ile
u
s
e,
an
d
we
g
i
v
e
im
p
o
r
tan
ce
to
b
o
th
d
ep
lo
y
m
en
t r
ea
d
in
ess
an
d
p
r
ed
ictiv
e
p
er
f
o
r
m
an
ce
.
T
h
is
s
tu
d
y
p
r
o
v
i
d
es
a
r
eliab
le,
tr
an
s
p
ar
en
t,
an
d
r
e
p
r
o
d
u
cib
le
p
ip
elin
e
f
o
r
d
o
wn
y
m
ild
ew
d
e
tectio
n
in
p
r
ec
is
io
n
ag
r
icu
ltu
r
e
b
y
a
p
p
ly
in
g
XAI
,
r
ea
l
-
wo
r
l
d
d
ata
s
p
lits
,
tr
an
s
f
er
-
lear
n
in
g
,
an
d
s
tatis
tical
v
alid
atio
n
.
Fo
r
b
o
th
p
r
ac
titi
o
n
e
r
s
an
d
r
esear
ch
er
s
,
r
esu
lts
co
n
tr
ib
u
te
to
c
h
o
o
s
in
g
a
m
o
d
el
an
d
d
ep
lo
y
m
en
t
s
tr
ateg
y
.
T
h
is
p
ip
elin
e
s
u
p
p
o
r
ts
p
r
ac
tical
im
p
lem
en
tatio
n
an
d
in
f
o
r
m
ed
d
e
cisi
o
n
-
m
ak
in
g
i
n
p
r
ec
is
io
n
a
g
r
icu
ltu
r
e.
2.
M
E
T
H
O
D
To
en
s
u
r
e
ac
c
u
r
ac
y
a
n
d
in
ter
p
r
etab
ilit
y
,
th
e
p
r
o
p
o
s
ed
f
r
am
e
wo
r
k
f
o
r
d
o
wn
y
m
ild
ew
d
etec
tio
n
ac
r
o
s
s
m
u
ltip
le
v
eg
etab
le
s
p
ec
ies
in
t
eg
r
ates
XAI
,
cr
o
s
s
-
v
alid
atio
n
,
s
tatis
tica
l
v
alid
atio
n
,
an
d
tr
a
n
s
f
er
lear
n
in
g
.
T
h
e
co
m
p
lete
p
ip
elin
e,
f
r
o
m
d
ata
g
ath
er
in
g
to
th
e
v
is
u
al
in
ter
p
r
etatio
n
o
f
m
o
d
el
atten
tio
n
v
ia
Gr
ad
-
C
AM
,
is
p
r
o
v
id
e
d
in
Fig
u
r
e
1.
E
ac
h
c
o
m
p
o
n
en
t
of
th
is
p
ip
elin
e
wa
s
ca
r
ef
u
lly
d
esig
n
ed
to
im
p
r
o
v
e
g
en
e
r
aliza
tio
n
on
im
b
alan
ce
d
d
atasets
wh
ile
m
ain
tain
in
g
co
m
p
u
tatio
n
al
ef
f
icien
cy
f
o
r
p
r
ac
tical
f
ield
d
e
p
lo
y
m
en
t.
Fig
u
r
e
1
.
Ar
c
h
itectu
r
e
f
o
r
d
o
w
n
y
m
ild
ew
d
etec
tio
n
a
n
d
in
te
r
p
r
etatio
n
u
s
in
g
d
ee
p
lear
n
in
g
m
o
d
els
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
C
o
mp
a
r
a
tive
d
ee
p
lea
r
n
in
g
s
tu
d
y
fo
r
d
o
w
n
y
mild
ew d
etec
tio
n
in
ve
g
eta
b
les
(
S
u
p
r
ee
th
a
S
h
i
va
r
a
j
)
1721
2
.
1
.
Da
t
a
a
c
q
u
i
s
i
t
i
o
n
Fiv
e
v
eg
etab
le
d
atasets
co
m
p
r
is
ed
of
h
ea
lth
y
an
d
d
o
wn
y
m
il
d
ew
-
in
f
ec
ted
leav
es
wer
e
u
tili
ze
d
in
th
e
s
tu
d
y
.
Data
wer
e
g
ath
e
r
ed
f
r
o
m
av
ailab
le
s
o
u
r
ce
s
,
s
p
ec
if
ically
v
eg
etab
les
lik
e
b
o
ttle
g
o
u
r
d
s
,
c
u
cu
m
b
er
s
,
ca
u
liflo
wer
,
an
d
b
itter
g
o
u
r
d
s
.
T
h
e
d
if
f
e
r
en
t so
u
r
ce
s
f
r
o
m
w
h
er
e
th
e
d
ata
is
co
llected
a
r
e
li
s
ted
:
−
R
ash
id
et
a
l.
[
2
1
]
:
cu
cu
m
b
er
l
ea
f
im
ag
es
f
r
o
m
th
e
p
lan
t
leaf
f
r
esh
n
ess
an
d
d
is
ea
s
e
d
etec
tio
n
d
ataset
[
2
1
]
,
co
n
s
is
tin
g
o
f
r
ea
l
-
tim
e
d
ata.
−
Su
ltan
a
et
al
.
[
2
2
]
:
cu
c
u
m
b
er
leaf
im
ag
es
f
r
o
m
t
h
e
cu
c
u
m
b
er
d
is
ea
s
e
r
ec
o
g
n
itio
n
d
ataset
[
2
2
]
,
co
n
tain
in
g
im
ag
es c
ap
tu
r
ed
u
n
d
e
r
v
ar
i
o
u
s
en
v
ir
o
n
m
en
tal
co
n
d
itio
n
s
.
−
R
ash
id
et
a
l.
[
2
1
]
lab
o
r
ato
r
y
d
ata:
b
itter
g
o
u
r
d
,
b
o
ttle
g
o
u
r
d
,
an
d
ca
u
liflo
wer
leaf
im
a
g
e
s
co
llected
in
lab
o
r
ato
r
y
s
ettin
g
s
ag
ain
s
t b
lack
b
ac
k
g
r
o
u
n
d
s
.
T
h
e
d
ataset
is
s
u
m
m
ar
ized
in
T
ab
le
1
.
T
ab
le
1
.
An
o
v
er
v
iew
o
f
th
e
d
ataset
O
r
i
g
i
n
a
l
d
a
t
a
s
e
t
V
e
g
e
t
a
b
l
e
t
y
p
e
D
o
w
n
y
m
i
l
d
e
w
H
e
a
l
t
h
y
B
i
t
t
e
r
g
u
a
r
d
5
7
0
5
5
1
B
o
t
t
l
e
g
u
a
r
d
6
8
4
5
1
8
C
a
u
l
i
f
l
o
w
e
r
5
1
2
5
2
6
C
u
c
u
m
b
e
r
(
R
a
s
h
i
d
)
5
6
4
5
2
7
C
u
c
u
m
b
e
r
(
S
u
l
t
a
n
a
)
1
6
0
1
6
0
2
.
2
.
I
ma
g
e
prepro
ce
s
s
ing
T
o
en
s
u
r
e
co
n
s
is
ten
cy
a
n
d
r
ep
r
o
d
u
cib
ilit
y
ac
r
o
s
s
ex
p
er
i
m
en
ts
,
a
s
tan
d
ar
d
ized
p
i
p
elin
e
was
im
p
lem
en
ted
to
p
r
e
p
r
o
ce
s
s
al
l
im
ag
es.
T
o
ac
h
iev
e
th
e
in
p
u
t
r
eq
u
ir
em
en
ts
o
f
VGG1
9
,
Den
s
eNe
t2
0
1
,
an
d
Mo
b
ileNetV2
,
r
aw
R
GB
im
ag
es
co
llected
f
r
o
m
h
eter
o
g
en
eo
u
s
r
ea
l
-
f
ield
c
o
n
d
itio
n
s
wer
e
r
esized
to
2
2
4
×2
2
4
p
ix
els.
I
m
ag
eNe
t
m
e
an
an
d
s
tan
d
a
r
d
d
e
v
iatio
n
s
tati
s
tics
[
0
.
4
8
5
,
0
.
4
5
6
,
0
.
4
0
6
]
[
0
.
4
8
5
,
0
.
4
5
6
,
0
.
4
0
6
]
an
d
[
0
.
2
2
9
,
0
.
2
2
4
,
0
.
2
2
5
]
[
0
.
2
2
9
,
0
.
2
2
4
,
0
.
2
2
5
]
[
0
.
2
2
9
,
0
.
2
2
4
,
0
.
2
2
5
]
in
cr
ea
s
ed
tr
an
s
f
er
lear
n
in
g
s
tab
ilit
y
.
T
o
p
r
ev
en
t
in
f
o
r
m
atio
n
leak
a
g
e,
class
p
r
o
p
o
r
tio
n
s
ar
e
m
ain
tai
n
ed
ac
r
o
s
s
tr
ain
in
g
,
v
alid
atio
n
,
an
d
test
s
ets
b
y
s
p
litt
in
g
th
e
d
ataset
u
s
in
g
s
t
r
atif
ied
s
am
p
lin
g
p
r
io
r
to
au
g
m
en
tatio
n
.
T
h
e
v
alid
atio
n
a
n
d
test
s
ets
o
n
ly
co
m
p
r
is
ed
o
r
ig
in
al,
u
n
alter
ed
im
ag
es,
wh
ile
au
g
m
en
tatio
n
was
ap
p
lied
o
n
ly
o
n
th
e
tr
ain
i
n
g
s
et
u
s
in
g
Ker
as
g
en
er
ato
r
s
in
o
r
d
e
r
to
en
h
an
ce
r
o
b
u
s
tn
ess
an
d
ad
d
r
ess
cl
ass
im
b
alan
ce
.
T
h
e
in
teg
r
ity
o
f
p
er
f
o
r
m
an
c
e
ev
alu
atio
n
is
m
ain
tain
ed
b
y
th
is
s
tr
ict
s
ep
ar
atio
n
th
at
en
s
u
r
es
th
at
n
o
au
g
m
en
ted
v
ar
ian
ts
o
f
tr
ain
in
g
s
am
p
les
ap
p
ea
r
in
v
alid
atio
n
o
r
test
f
o
ld
s
.
E
v
er
y
v
eg
etab
le
an
d
c
r
o
s
s
-
v
alid
atio
n
f
o
l
d
wa
s
s
u
b
jecte
d
to
t
h
e
s
am
e
p
r
o
t
o
co
l.
R
an
d
o
m
r
o
tati
o
n
s
(
u
p
to
4
0
°),
h
o
r
izo
n
tal
f
lip
p
in
g
,
wid
th
an
d
h
eig
h
t
s
h
if
ts
(
±
2
0
%),
s
h
ea
r
in
g
,
an
d
zo
o
m
in
g
a
u
g
m
e
n
tatio
n
o
p
er
atio
n
s
wer
e
ca
r
r
ied
o
u
t.
Du
r
in
g
t
r
ain
in
g
,
a
u
g
m
en
te
d
s
am
p
les
wer
e
g
e
n
er
ated
d
y
n
a
m
ically
in
o
r
d
er
t
o
av
o
id
d
u
p
licate
d
ata
o
n
d
is
k
an
d
allo
w
ea
ch
im
a
g
e
to
a
p
p
ea
r
o
v
e
r
m
u
ltip
le
s
to
ch
asti
c
v
ar
iatio
n
s
ac
r
o
s
s
ep
o
ch
s
.
T
o
a
v
o
id
cr
o
s
s
-
cr
o
p
co
n
tam
in
ati
o
n
,
ea
ch
v
eg
etab
le
ca
teg
o
r
y
—
b
itter
g
o
u
r
d
,
b
o
ttle
g
o
u
r
d
,
ca
u
lif
lo
wer
,
cu
cu
m
b
e
r
(
R
ash
id
)
,
an
d
c
u
cu
m
b
e
r
(
Su
ltan
a)
—
was
p
r
o
ce
s
s
ed
in
d
i
v
id
u
ally
.
T
h
is
p
r
ep
r
o
ce
s
s
in
g
t
ec
h
n
iq
u
e
en
ab
les
g
en
er
aliza
tio
n
to
r
ea
lf
ield
v
a
r
iatio
n
s
,
in
clu
d
in
g
v
a
r
iatio
n
s
in
illu
m
in
atio
n
,
b
ac
k
g
r
o
u
n
d
co
m
p
lex
ity
,
an
d
s
y
m
p
to
m
d
iv
er
s
ity
,
wh
ile
f
ac
ilit
atin
g
o
b
jectiv
e
e
v
alu
atio
n
.
Fin
ally
,
to
ass
ess
m
o
d
el
g
en
er
aliza
tio
n
ac
r
o
s
s
d
if
f
er
en
t
s
p
lits
,
a
5
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
s
tr
ateg
y
was
em
p
lo
y
ed
.
C
o
llectiv
ely
,
th
ese
p
r
ep
r
o
ce
s
s
in
g
s
tep
s
en
s
u
r
e
th
at
th
e
tr
ain
ed
m
o
d
els
w
i
l
l
b
e
r
es
il
i
e
n
t
t
o
t
h
e
c
h
a
ll
e
n
g
e
s
t
h
a
t
e
x
i
s
t
i
n
r
e
a
l
-
w
o
r
l
d
a
g
r
i
c
u
l
t
u
r
a
l
i
m
a
g
e
r
y
,
s
u
c
h
a
s
n
o
is
e
,
li
g
h
t
i
n
g
v
a
r
i
a
ti
o
n
,
a
n
d
d
a
t
a
i
m
b
a
l
a
n
c
e
.
2
.
3
.
Da
t
a
s
pli
t
t
ing
Prio
r
to
a
n
y
d
ata
au
g
m
en
tatio
n
,
d
ataset
s
p
litt
in
g
was
p
er
f
o
r
m
ed
to
e
n
s
u
r
e
an
u
n
b
iased
ev
a
lu
atio
n
a
n
d
p
r
ev
en
t
d
ata
lea
k
ag
e.
T
h
e
d
a
taset
was
d
iv
id
ed
i
n
to
test
,
v
alid
atio
n
,
an
d
tr
ai
n
in
g
s
ets
with
a
7
0
%
tr
ain
in
g
r
atio
,
1
5
%
test
in
g
,
a
n
d
1
5
%
v
alid
atio
n
.
T
h
is
b
alan
ce
d
s
p
litt
in
g
en
s
u
r
e
d
th
at
ea
c
h
s
u
b
s
et
m
ain
tain
ed
th
e
class
d
is
tr
ib
u
tio
n
,
th
er
eb
y
en
a
b
lin
g
r
eliab
le
m
o
d
el
ev
alu
atio
n
.
T
ab
l
e
2
p
r
o
v
i
d
es
d
etails
of
t
h
e
d
atas
et
af
ter
s
p
l
it
t
i
n
g
.
2
.
4
.
M
o
del t
ra
ini
ng
a
nd
ev
a
l
ua
t
io
n
Af
ter
p
r
e
p
r
o
ce
s
s
in
g
,
th
r
ee
d
ee
p
lear
n
i
n
g
ar
ch
itectu
r
es
—
VGG1
9
,
Den
s
eNe
t2
0
1
[
2
3
]
,
[
2
4
]
a
n
d
Mo
b
ileNetV2
was
u
s
ed
to
tr
a
in
an
d
ex
am
i
n
e
th
e
p
r
e
p
ar
ed
d
atasets
f
o
r
f
iv
e
v
e
g
etab
les:
b
itter
g
o
u
r
d
,
b
o
ttle
g
o
u
r
d
,
ca
u
liflo
wer
,
cu
cu
m
b
er
(
R
ash
id
)
,
an
d
cu
c
u
m
b
e
r
(
Su
ltan
a)
.
Du
e
to
its
d
ee
p
y
et
s
t
ab
le
co
n
v
o
lu
tio
n
al
s
tr
u
ctu
r
e,
wh
ich
h
as
b
ee
n
wid
ely
v
alid
ated
in
ag
r
icu
ltu
r
al
im
ag
e
an
aly
s
is
,
VGG1
9
s
er
v
es
as
a
s
tr
o
n
g
class
ical
b
aselin
e.
Den
s
eNe
t2
0
1
was
ch
o
s
en
d
u
e
t
o
its
d
en
s
ely
co
n
n
ec
ted
d
esig
n
t
h
at
p
r
o
m
o
tes
f
ea
tu
r
e
r
eu
s
e
an
d
ef
f
icien
t
g
r
ad
ien
t
f
lo
w,
e
n
ab
li
n
g
th
e
m
o
d
el
to
d
etec
t
s
u
b
tle
an
d
f
i
n
e
-
g
r
ai
n
ed
s
y
m
p
to
m
s
o
f
d
o
wn
y
m
ild
ew
in
v
ar
io
u
s
v
e
g
etab
le
s
p
ec
ies.
I
n
o
r
d
e
r
to
ass
ess
a
co
m
p
u
tatio
n
ally
ef
f
icien
t
alter
n
ativ
e
t
h
at
ca
n
p
r
o
v
id
e
f
ast
in
f
er
en
ce
an
d
lo
w
m
em
o
r
y
u
s
ag
e,
m
ak
in
g
it
s
u
itab
le
f
o
r
r
ea
l
-
tim
e
f
ield
ap
p
licatio
n
s
,
Mo
b
i
leNe
tV2
[
2
5
]
–
[
2
7
]
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
7
1
9
-
1
7
3
2
1722
a
lig
h
tweig
h
t
ar
c
h
itectu
r
e
d
e
s
ig
n
ed
f
o
r
ed
g
e
d
e
p
lo
y
m
e
n
t,
was
in
clu
d
ed
.
T
o
le
v
er
ag
e
tr
an
s
f
er
lear
n
i
n
g
f
o
r
f
aster
co
n
v
er
g
e
n
ce
an
d
im
p
r
o
v
ed
p
er
f
o
r
m
a
n
ce
o
n
lim
ited
a
g
r
icu
ltu
r
al
d
ata,
all
m
o
d
els
wer
e
in
itialized
with
I
m
ag
eNe
t
-
p
r
etr
ai
n
ed
weig
h
ts
.
A
cu
s
to
m
ized
class
if
icatio
n
h
ea
d
co
n
s
is
tin
g
o
f
g
l
o
b
al
av
er
ag
e
p
o
o
lin
g
,
d
r
o
p
o
u
t,
an
d
a
d
en
s
e
So
f
tMa
x
o
u
tp
u
t
la
y
er
with
two
n
eu
r
o
n
s
r
ep
r
esen
tin
g
t
h
e
class
es
—
h
ea
lth
y
an
d
d
o
wn
y
m
ild
ew
—
h
as
r
ep
lace
d
ea
c
h
m
o
d
el’
s
f
in
al
f
u
lly
co
n
n
ec
ted
la
y
er
s
.
C
lass
s
ep
ar
ab
ilit
y
was
o
p
tim
ized
u
s
in
g
th
e
ca
teg
o
r
ical
cr
o
s
s
-
en
tr
o
p
y
lo
s
s
f
u
n
ctio
n
,
an
d
th
e
m
o
d
els
wer
e
em
p
lo
y
ed
u
s
in
g
th
e
Ad
a
m
o
p
tim
izer
with
a
lear
n
in
g
r
ate
o
f
1
e
-
4
.
T
o
en
s
u
r
e
r
ep
r
o
d
u
cib
ilit
y
a
n
d
tr
an
s
p
ar
en
c
y
,
all
ex
p
er
im
en
ta
l
co
n
f
ig
u
r
atio
n
s
wer
e
e
x
p
licitly
d
ef
in
e
d
.
T
o
p
r
ev
en
t o
v
er
f
itti
n
g
,
ea
ch
m
o
d
el
was tr
ain
ed
f
o
r
1
0
ep
o
c
h
s
with
a
b
atch
s
ize
o
f
8
,
u
s
in
g
e
ar
ly
s
to
p
p
in
g
wit
h
a
p
atien
ce
o
f
3
ep
o
ch
s
.
T
o
r
etain
th
e
b
est
-
p
er
f
o
r
m
in
g
weig
h
ts
b
ased
o
n
v
alid
atio
n
ac
cu
r
ac
y
,
m
o
d
e
l
ch
ec
k
p
o
i
n
tin
g
was
em
p
lo
y
e
d
.
R
ep
r
o
d
u
cib
ilit
y
was
en
s
u
r
ed
b
y
f
ix
in
g
r
an
d
o
m
s
ee
d
s
ac
r
o
s
s
Nu
m
Py
,
T
en
s
o
r
Flo
w,
an
d
Py
th
o
n
.
All
ex
p
er
im
en
ts
wer
e
ex
ec
u
ted
o
n
Go
o
g
le
C
o
lab
with
NVI
DI
A
GPU
ac
ce
ler
atio
n
.
I
n
o
r
d
er
t
o
p
r
o
v
id
e
co
n
s
is
ten
t
an
d
r
ep
ea
tab
le
e
x
p
er
im
e
n
tal
o
u
tco
m
es,
T
e
n
s
o
r
Flo
w’
s
d
ete
r
m
in
is
tic
b
eh
av
io
r
was a
d
d
itio
n
ally
en
ab
led
wh
er
e
ap
p
licab
le.
T
ab
le
2
.
An
o
v
er
v
iew
o
f
th
e
d
ataset
af
ter
d
ata
s
p
litt
in
g
D
a
t
a
s
e
t
s
p
l
i
t
s
u
mm
a
r
y
V
e
g
e
t
a
b
l
e
t
y
p
e
C
l
a
s
s
Tr
a
i
n
V
a
l
Te
st
B
i
t
t
e
r
g
o
u
r
d
H
e
a
l
t
h
y
3
8
5
82
84
D
o
w
n
y
m
i
l
d
e
w
3
9
9
85
86
B
o
t
t
l
e
g
o
u
r
d
H
e
a
l
t
h
y
3
6
2
77
79
D
o
w
n
y
m
i
l
d
e
w
4
7
8
1
0
2
1
0
4
C
a
u
l
i
f
l
o
w
e
r
H
e
a
l
t
h
y
3
6
8
78
80
D
o
w
n
y
m
i
l
d
e
w
3
5
8
76
78
C
u
c
u
m
b
e
r
(
R
a
s
h
i
d
)
H
e
a
l
t
h
y
3
6
8
79
80
D
o
w
n
y
m
i
l
d
e
w
3
9
4
84
86
C
u
c
u
m
b
e
r
(
S
u
l
t
a
n
a
)
H
e
a
l
t
h
y
1
1
2
24
24
D
o
w
n
y
m
i
l
d
e
w
1
1
2
24
24
2
.
5
.
5
-
f
o
ld
cr
o
s
s
-
v
a
lid
a
t
io
n
5
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
was
im
p
lem
en
ted
o
n
ea
c
h
v
e
g
etab
le
in
th
e
d
ataset
to
en
s
u
r
e
r
o
b
u
s
tn
ess
.
T
o
r
ed
u
ce
th
e
b
ias
ca
u
s
ed
b
y
a
s
i
n
g
le
tr
ain
-
test
s
p
lit,
ea
ch
m
o
d
el
was
tr
ain
ed
an
d
v
alid
ated
f
o
r
5
-
f
o
l
d
s
,
an
d
th
e
r
esu
lts
wer
e
ag
g
r
eg
ated
.
Acc
u
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all
,
a
n
d
F1
-
s
co
r
e
wer
e
av
er
a
g
ed
o
v
er
f
o
ld
s
in
o
r
d
er
t
o
p
r
o
v
id
e
a
th
o
r
o
u
g
h
ev
al
u
atio
n
o
f
p
r
e
d
ictiv
e
p
er
f
o
r
m
a
n
ce
.
2
.
6
.
P
er
f
o
r
m
a
nce
ev
a
lua
t
io
n
T
o
en
s
u
r
e
r
eliab
ilit
y
an
d
a
d
a
p
tab
ilit
y
ac
r
o
s
s
im
b
alan
ce
d
d
ata
s
ets,
th
e
m
o
d
el’
s
p
er
f
o
r
m
an
ce
was
ass
es
s
ed
u
s
in
g
b
o
th
in
d
e
p
en
d
en
t
test
-
s
et
ev
alu
atio
n
a
n
d
5
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
.
T
h
e
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e
wer
e
o
b
tain
ed
b
y
c
o
m
p
ar
i
n
g
th
e
n
etwo
r
k
’
s
p
r
ed
ictio
n
s
to
g
r
o
u
n
d
-
tr
u
t
h
lab
els
f
o
r
ea
ch
v
e
g
etab
le
-
m
o
d
el
p
air
.
C
o
llectiv
ely
,
th
ese
m
etr
ics
c
ap
tu
r
e
r
o
b
u
s
tn
ess
u
n
d
er
im
b
alan
ce
,
s
en
s
itiv
ity
to
d
is
ea
s
e
d
etec
tio
n
,
co
r
r
ec
tn
es
s
,
an
d
class
-
wis
e
r
eliab
ilit
y
.
C
o
m
p
u
tatio
n
al
ef
f
icien
c
y
m
etr
ics,
s
p
ec
if
ically
m
o
d
el
s
ize
(
MB),
n
u
m
b
er
o
f
tr
ain
ab
le
p
a
r
am
eter
s
,
an
d
in
f
er
en
ce
tim
e
p
er
im
a
g
e
(
m
s
)
,
wer
e
r
ep
o
r
ted
i
n
ad
d
itio
n
to
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
I
n
o
r
d
e
r
to
co
m
p
ar
e
t
h
e
d
if
f
er
e
n
t
ar
ch
itectu
r
es
f
o
r
d
ep
lo
y
m
en
t
s
ettin
g
s
in
m
o
b
ile
o
r
ed
g
e
d
ev
ices,
wh
er
e
m
em
o
r
y
an
d
laten
cy
ar
e
im
p
o
r
tan
t
f
ac
to
r
s
,
th
ese
m
etr
ics
ar
e
h
i
g
h
ly
im
p
o
r
tan
t.
Acc
u
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all
,
an
d
F1
-
s
co
r
e
ar
e
s
o
m
e
o
f
th
e
m
etr
ics
wh
ich
wer
e
av
er
ag
ed
o
v
er
th
e
5
cr
o
s
s
-
v
alid
atio
n
f
o
ld
s
,
p
r
o
v
i
d
in
g
an
in
-
d
e
p
th
u
n
d
er
s
tan
d
in
g
o
f
p
er
f
o
r
m
an
ce
s
tab
ilit
y
.
Fu
r
th
er
,
th
e
cr
o
s
s
-
f
o
l
d
v
ar
ian
ce
v
alid
ated
ea
ch
m
o
d
el’
s
co
n
s
is
ten
cy
ac
r
o
s
s
s
am
p
lin
g
d
is
tr
ib
u
tio
n
s
.
T
h
e
p
er
f
o
r
m
an
ce
in
d
icato
r
s
em
p
lo
y
ed
to
ass
ess
th
e
m
o
d
els ar
e
as f
o
llo
ws
:
−
Pre
cisi
o
n
:
a
m
o
d
el
’
s
p
r
ec
is
io
n
,
o
r
its
ca
p
ac
ity
t
o
ac
cu
r
ately
id
en
tif
y
p
o
s
itiv
e
ca
s
es
am
o
n
g
all
p
o
s
itiv
e
p
r
ed
ictio
n
s
,
is
r
ep
r
esen
ted
b
y
(
1
)
,
wh
e
r
e
T
P
is
th
e
n
u
m
b
e
r
o
f
tr
u
e
p
o
s
itiv
es,
a
n
d
FP
is
t
h
e
n
u
m
b
er
o
f
f
alse p
o
s
itiv
es
.
=
+
(
1
)
−
R
ec
all:
th
e
m
o
d
el’
s
ca
p
ac
ity
to
d
etec
t
ea
ch
an
d
ev
er
y
tr
u
e
p
o
s
itiv
e
in
s
tan
ce
is
in
d
icate
d
b
y
(
2
)
,
wh
i
c
h
also
p
r
o
v
id
es th
e
n
u
m
b
er
o
f
T
P a
n
d
FN
.
=
+
(
2
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
C
o
mp
a
r
a
tive
d
ee
p
lea
r
n
in
g
s
tu
d
y
fo
r
d
o
w
n
y
mild
ew d
etec
tio
n
in
ve
g
eta
b
les
(
S
u
p
r
ee
th
a
S
h
i
va
r
a
j
)
1723
−
F1
-
s
co
r
e
:
in
(
3
)
s
h
o
ws
th
at
th
e
F1
-
s
co
r
e
,
wh
ich
is
ca
lcu
lated
b
y
tak
in
g
th
e
h
a
r
m
o
n
ic
m
e
a
n
o
f
p
r
ec
is
io
n
an
d
r
ec
all,
is
a
b
alan
ce
d
in
d
ica
to
r
o
f
t
h
e
m
o
d
el’
s
ac
cu
r
ac
y
.
1
−
=
2
×
×
+
(
3
)
−
Acc
u
r
ac
y
:
in
(
4
)
,
wh
er
e
T
P
s
t
an
d
s
f
o
r
tr
u
e
p
o
s
itiv
es
,
T
N
f
o
r
tr
u
e
n
e
g
ativ
es
,
FP
f
o
r
f
alse
p
o
s
itiv
es,
an
d
FN f
o
r
f
alse n
eg
ativ
es,
ac
cu
r
a
cy
is
th
e
p
er
ce
n
ta
g
e
o
f
ac
cu
r
at
e
p
r
ed
ictio
n
s
o
n
th
e
test
s
et.
=
+
+
+
+
(
4
)
T
h
e
m
etr
ics
o
u
tlin
ed
in
th
e
f
o
l
lo
win
g
s
ec
tio
n
s
p
r
o
v
i
d
e
th
e
tr
ad
e
-
o
f
f
b
etwe
en
p
r
ed
ictiv
e
p
e
r
f
o
r
m
a
n
ce
an
d
d
e
p
lo
y
a
b
ilit
y
.
−
Par
am
eter
s
(
M)
:
th
e
n
etwo
r
k
’
s
to
tal
n
u
m
b
e
r
o
f
lear
n
ab
le
wei
g
h
ts
.
−
Mo
d
el
s
ize
(
MB):
th
e
am
o
u
n
t
o
f
m
em
o
r
y
r
e
q
u
ir
e
d
to
s
av
e
th
e
m
o
d
el.
−
I
n
f
er
en
ce
tim
e
(
m
s
)
:
th
e
f
o
r
wa
r
d
p
ass
laten
cy
is
th
e
av
er
a
g
e
t
im
e
r
eq
u
ir
e
d
to
p
r
o
ce
s
s
o
n
e
te
s
t im
ag
e.
2
.
7
.
St
a
t
is
t
ica
l
t
esting
Statis
t
ical
s
ig
n
if
ican
ce
test
in
g
was
in
co
r
p
o
r
ate
d
in
th
e
ev
al
u
atio
n
m
eth
o
d
to
en
s
u
r
e
th
at
d
if
f
er
en
ce
s
in
p
er
f
o
r
m
an
ce
b
etwe
en
d
ee
p
lear
n
in
g
m
o
d
els
wer
e
n
o
t
d
u
e
to
r
an
d
o
m
d
if
f
er
en
ce
s
b
u
t
m
er
ely
r
ef
lecte
d
o
r
ig
in
al
im
p
r
o
v
em
en
ts
.
Pair
ed
t
-
test
s
an
d
t
h
e
W
ilco
x
o
n
s
ig
n
ed
-
r
an
k
test
wer
e
d
ep
l
o
y
ed
to
co
m
p
ar
e
th
e
m
etr
ic
v
alu
es
f
o
r
ea
c
h
v
e
g
etab
le
–
m
o
d
el
p
air
ac
r
o
s
s
th
e
5
cr
o
s
s
-
v
alid
atio
n
f
o
l
d
s
(
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
an
d
F1
-
s
co
r
e)
.
W
h
ile
th
e
W
ilco
x
o
n
test
p
r
o
v
i
d
es
a
n
o
n
-
p
a
r
am
et
r
ic
alter
n
ativ
e
th
at
d
o
es
n
o
t
r
e
ly
o
n
d
is
tr
ib
u
tio
n
al
ass
u
m
p
tio
n
s
,
m
ak
in
g
it
s
u
itab
le
f
o
r
s
m
all
s
am
p
le
s
ize
s
lik
e
5
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
,
th
e
p
a
ir
ed
t
-
test
ass
e
s
s
e
s
wh
eth
er
th
e
m
ea
n
d
if
f
er
e
n
ce
b
etwe
en
two
m
o
d
els’
p
er
f
o
r
m
an
ce
s
co
r
es
is
s
tatis
tically
s
ig
n
if
ican
t
u
n
d
e
r
th
e
ass
u
m
p
tio
n
o
f
n
o
r
m
ality
.
T
h
e
s
e
test
s
wer
e
p
er
f
o
r
m
e
d
m
o
d
e
l
-
wis
e
(
e.
g
.
,
Den
s
eNe
t2
0
1
v
s
.
Mo
b
ileNetV2
)
o
n
ea
ch
m
etr
ic
a
n
d
v
eg
etab
le
d
ataset.
W
h
eth
er
o
n
e
ar
ch
itec
tu
r
e
co
n
s
is
ten
tly
o
u
tp
er
f
o
r
m
e
d
an
o
t
h
er
b
ey
o
n
d
ch
an
ce
(
p
<0
.
0
5
)
was
d
eter
m
i
n
ed
b
y
th
e
r
esu
ltin
g
p
-
v
alu
es.
W
h
ile
co
m
b
in
ed
with
ef
f
ec
t
-
s
ize
m
ea
s
u
r
es
s
u
ch
as
m
ea
n
p
er
f
o
r
m
a
n
ce
d
if
f
e
r
e
n
ce
s
(
(
∆M
ea
n
)
)
,
th
is
s
tatis
tic
al
an
aly
s
is
p
r
o
v
id
ed
p
er
f
o
r
m
an
ce
th
at
is
s
u
p
er
io
r
ac
r
o
s
s
s
ev
er
al
ty
p
es
o
f
o
th
er
v
eg
etab
les
,
an
d
m
o
d
el
s
u
p
er
io
r
ity
was
clea
r
ly
s
h
o
wca
s
ed
,
m
o
s
t
n
o
tab
ly
in
d
icatin
g
Den
s
eNe
t2
0
1
’
s
s
ig
n
if
ican
t a
d
v
a
n
tag
e
o
n
cu
c
u
m
b
er
d
atasets
.
2
.
8
.
E
x
pla
ina
bil
it
y
a
na
ly
s
is
Gr
ad
-
C
A
M
wa
s
em
p
lo
y
ed
to
r
ep
r
e
s
en
t
th
e
s
p
at
ia
l
r
eg
io
n
s
th
at
c
o
n
tr
ib
u
ted
m
o
s
t
to
ea
c
h
m
o
d
el’
s
p
r
ed
i
ct
io
n
s
,
r
eq
u
ir
ed
to
en
h
an
ce
th
e
d
ee
p
-
le
ar
n
in
g
m
o
d
e
l
s
’
tr
an
s
p
ar
en
cy
an
d
in
ter
p
r
e
tab
il
ity
.
B
y
co
m
p
u
t
in
g
th
e
g
r
ad
ien
t
o
f
th
e
p
r
ed
i
ct
ed
cl
a
s
s
s
co
r
e
w
i
th
r
esp
ec
t
to
th
e
ac
tiv
at
io
n
s
o
f
th
e
f
in
a
l
c
o
n
v
o
lu
tio
n
al
l
ay
er
,
Gr
ad
-
C
A
M
h
ig
h
l
ig
h
t
s
d
i
s
cr
i
m
in
a
ti
v
e
im
ag
e
r
eg
io
n
s
an
d
g
en
er
at
es
a
c
la
s
s
-
s
p
e
ci
f
i
c
h
ea
t
m
ap
.
T
h
e
b
e
s
t
p
er
f
o
r
m
in
g
ch
ec
k
p
o
in
t
o
f
e
ac
h
ar
ch
i
te
ctu
r
e
(
V
G
G1
9
,
D
en
s
eNe
t2
0
1
,
an
d
Mo
b
il
eN
et
V2
)
ac
r
o
s
s
al
l
v
eg
et
ab
l
e
te
s
t
s
et
s
wa
s
ex
a
m
in
ed
b
y
Gr
ad
-
C
A
M
f
o
r
th
i
s
s
tu
d
y
.
Fo
r
e
ac
h
v
eg
e
tab
le,
r
ep
r
e
s
en
ta
tiv
e
im
ag
e
s
f
r
o
m
b
o
th
cla
s
s
e
s
—
h
ea
lth
y
an
d
d
o
wn
y
m
ild
ew
—
wer
e
s
el
ec
t
ed
to
en
s
u
r
e
a
co
n
s
i
s
t
en
t
q
u
a
li
ta
t
iv
e
c
o
m
p
ar
is
o
n
.
T
o
s
e
e
wh
e
th
er
th
e
m
o
d
e
l
s
co
r
r
e
ct
ly
f
o
cu
s
e
d
o
n
s
y
m
p
t
o
m
a
ti
c
ar
e
as
s
u
ch
a
s
ch
lo
r
o
t
ic
p
a
tch
e
s
,
an
g
u
l
ar
le
s
io
n
s
,
o
r
d
i
s
co
lo
r
a
tio
n
ch
a
r
ac
te
r
i
s
ti
c
o
f
d
o
wn
y
m
ild
ew
,
th
e
h
ea
tm
ap
s
w
er
e
o
v
er
la
id
o
n
t
o
th
e
o
r
ig
in
a
l
R
G
B
im
ag
e
s
.
I
n
ter
p
r
e
tab
il
it
y
i
s
e
s
s
en
t
ia
l
f
o
r
d
o
m
a
in
ex
p
er
ts
an
d
f
ie
ld
-
lev
e
l
d
ec
is
io
n
-
m
ak
in
g
,
th
e
e
x
p
la
in
ab
il
i
ty
r
e
s
u
lt
s
th
u
s
r
e
in
f
o
r
c
e
th
e
r
e
li
ab
i
li
t
y
o
f
th
e
f
in
a
l
m
o
d
e
l
p
r
ed
i
ct
io
n
s
an
d
p
r
o
v
id
e
c
o
n
f
id
e
n
ce
f
o
r
r
e
al
-
wo
r
ld
ag
r
i
cu
l
tu
r
al
d
ep
lo
y
m
en
t.
2
.
9
.
Resul
t
inte
rpre
t
a
t
io
n a
n
d v
is
ua
liza
t
io
n
T
o
p
r
o
v
i
d
e
a
co
m
p
lete
ass
ess
m
en
t
o
f
p
er
f
o
r
m
an
ce
ac
r
o
s
s
v
eg
etab
les
an
d
ar
ch
itectu
r
es,
th
e
m
o
d
el
o
u
tp
u
ts
wer
e
ex
am
in
ed
u
s
in
g
a
co
m
b
in
atio
n
o
f
q
u
an
titativ
e
m
etr
ics,
s
tati
s
tical
co
m
p
ar
is
o
n
s
,
an
d
q
u
alitativ
e
v
is
u
aliza
tio
n
s
.
T
h
e
ev
alu
atio
n
m
etr
ics (
ac
cu
r
ac
y
,
p
r
ec
is
io
n
,
r
ec
all,
F1
-
s
co
r
e,
p
ar
am
eter
co
u
n
t,
m
o
d
el
s
ize,
an
d
in
f
er
en
ce
tim
e)
we
r
e
f
i
r
s
t
ag
g
r
eg
ated
in
to
co
m
p
ar
ativ
e
tab
les
to
clea
r
l
y
id
e
n
tify
th
e
tr
ad
e
-
o
f
f
s
b
etwe
e
n
co
m
p
u
tatio
n
al
ef
f
icien
c
y
a
n
d
p
r
e
d
ictiv
e
p
o
wer
.
I
n
o
r
d
er
t
o
en
h
an
ce
in
ter
p
r
eta
b
ilit
y
,
d
etailed
Gr
ad
-
C
AM
v
is
u
aliza
tio
n
s
wer
e
g
en
er
ate
d
f
o
r
th
e
h
ea
lth
y
an
d
d
o
w
n
y
m
ild
ew
class
es
o
f
all
v
eg
etab
les.
Vis
u
al
co
n
f
ir
m
atio
n
o
f
wh
eth
e
r
ea
ch
m
o
d
el
ap
p
r
o
p
r
iately
f
o
c
u
s
ed
o
n
s
y
m
p
to
m
atic
leaf
r
e
g
io
n
s
was
p
r
o
v
id
ed
b
y
h
ea
tm
ap
s
.
T
h
e
in
co
r
p
o
r
atio
n
o
f
s
tatis
tical
s
ig
n
if
ican
ce
test
i
n
g
f
u
r
th
e
r
co
n
tex
tu
alize
d
th
e
r
esu
lts
an
d
v
er
if
ied
th
at
p
er
f
o
r
m
an
ce
d
if
f
er
en
c
es
wer
e
n
o
t
in
cid
e
n
tal.
T
o
g
eth
e
r
,
th
e
s
tatis
tical
o
u
tco
m
es,
v
is
u
al
h
ea
tm
ap
s
,
an
d
s
tr
u
ctu
r
ed
ta
b
les
p
r
o
v
id
ed
an
i
n
teg
r
ated
in
ter
p
r
etiv
e
f
r
am
ew
o
r
k
th
at
p
r
o
v
i
d
ed
r
o
b
u
s
tn
ess
an
d
tr
a
n
s
p
ar
en
c
y
in
th
e
ev
alu
atio
n
o
f
th
e
p
r
o
p
o
s
ed
d
is
ea
s
e
class
if
ica
tio
n
p
ip
elin
e
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
7
1
9
-
1
7
3
2
1724
3.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
e
t
h
r
ee
d
ee
p
-
le
ar
n
i
n
g
a
r
c
h
it
ec
t
u
r
es
’
p
r
e
d
ic
ti
v
e
p
e
r
f
o
r
m
a
n
c
e,
co
m
p
u
ta
ti
o
n
al
ef
f
i
ci
en
cy
,
a
n
d
in
t
er
p
r
et
a
b
ili
ty
d
i
f
f
er
s
i
g
n
i
f
ic
a
n
tl
y
i
n
t
h
e
ex
p
er
i
m
e
n
t
al
e
v
a
lu
ati
o
n
a
c
r
o
s
s
f
iv
e
v
eg
et
ab
les
.
T
ab
l
e
3
il
lu
s
tr
ates
t
h
e
o
v
e
r
al
l
m
o
d
el
p
e
r
f
o
r
m
a
n
c
e,
w
h
ic
h
r
e
p
o
r
ts
m
et
r
ics
s
u
c
h
as
m
o
d
el
s
ize
,
in
f
er
en
ce
la
te
n
c
y
,
F1
-
s
c
o
r
e
,
ac
cu
r
a
c
y
,
p
r
ec
is
i
o
n
,
r
ec
all
,
a
n
d
p
a
r
a
m
et
e
r
c
o
u
n
t.
B
o
t
h
De
n
s
e
Net
2
0
1
a
n
d
VGG
1
9
s
h
o
w
s
tr
o
n
g
f
ea
t
u
r
e
-
le
ar
n
i
n
g
c
a
p
a
b
ili
ty
an
d
r
o
b
u
s
t
n
ess
t
o
in
tr
a
-
class
v
ar
i
a
b
ili
ty
,
d
e
m
o
n
s
tr
ati
n
g
n
e
a
r
-
p
e
r
f
ec
t
c
lass
i
f
i
ca
t
io
n
ac
r
o
s
s
al
l
v
e
g
et
ab
les
e
x
ce
p
t
cu
cu
m
b
e
r
S
u
lta
n
a
,
wit
h
a
cc
u
r
a
cy
a
n
d
F1
-
s
c
o
r
es
o
f
9
8
–
1
0
0
%
.
Des
p
i
te
h
a
v
i
n
g
o
n
l
y
2
.
5
9
M
p
a
r
a
m
et
er
s
,
Mo
b
il
eNe
tV
2
p
r
o
d
u
c
e
d
h
i
g
h
l
y
c
o
m
p
etit
iv
e
r
esu
lts
(
f
o
r
e
x
a
m
p
le
,
9
9
.
3
7
%
a
cc
u
r
ac
y
f
o
r
ca
u
l
i
f
l
o
we
r
a
n
d
9
6
.
9
9
%
f
o
r
c
u
c
u
m
b
er
R
as
h
i
d
)
,
h
i
g
h
li
g
h
ti
n
g
its
s
u
i
ta
b
il
it
y
f
o
r
d
e
p
l
o
y
m
e
n
t
o
n
e
m
b
e
d
d
e
d
o
r
m
o
b
i
le
p
la
tf
o
r
m
s
.
T
ab
le
3
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
o
f
VGG1
9
,
Den
s
eNe
t2
0
1
,
an
d
Mo
b
ileNetV2
ac
r
o
s
s
all
v
eg
etab
les
V
e
g
e
t
a
b
l
e
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
-
s
c
o
r
e
S
u
p
p
o
r
t
P
a
r
a
ms
(
M
)
S
i
z
e
(
M
B
)
I
n
f
e
r
e
n
c
e
(
m
s
)
b
i
t
t
e
r
_
g
r
o
u
n
d
D
e
n
s
e
N
e
t
2
0
1
1
0
0
1
0
0
1
0
0
1
0
0
1
7
0
1
8
.
8
1
7
7
.
9
1
6
0
5
.
2
5
b
i
t
t
e
r
_
g
r
o
u
n
d
M
o
b
i
l
e
N
e
t
V
2
9
5
.
8
8
9
6
.
1
9
9
5
.
8
8
9
5
.
8
7
1
7
0
2
.
5
9
1
2
.
9
3
5
2
1
.
6
3
b
i
t
t
e
r
_
g
r
o
u
n
d
V
G
G
1
9
1
0
0
1
0
0
1
0
0
1
0
0
1
7
0
2
0
.
1
6
7
7
.
9
9
9
7
9
.
8
3
b
o
t
t
l
e
_
g
o
u
r
d
D
e
n
s
e
N
e
t
2
0
1
1
0
0
1
0
0
1
0
0
1
0
0
1
8
3
1
8
.
8
1
7
7
.
9
1
6
0
0
.
8
b
o
t
t
l
e
_
g
o
u
r
d
M
o
b
i
l
e
N
e
t
V
2
1
0
0
1
0
0
1
0
0
1
0
0
1
8
3
2
.
5
9
1
2
.
9
3
7
7
5
.
5
6
b
o
t
t
l
e
_
g
o
u
r
d
V
G
G
1
9
1
0
0
1
0
0
1
0
0
1
0
0
1
8
3
2
0
.
1
6
7
7
.
9
9
1
0
3
0
.
2
6
c
a
u
l
i
f
l
o
w
e
r
D
e
n
s
e
N
e
t
2
0
1
9
8
.
7
3
9
8
.
7
7
9
8
.
7
3
9
8
.
7
3
1
5
8
1
8
.
8
1
7
7
.
9
1
6
7
8
.
8
7
c
a
u
l
i
f
l
o
w
e
r
M
o
b
i
l
e
N
e
t
V
2
9
9
.
3
7
9
9
.
3
8
9
9
.
3
7
9
9
.
3
7
1
5
8
2
.
5
9
1
2
.
9
3
5
7
0
.
8
4
c
a
u
l
i
f
l
o
w
e
r
V
G
G
1
9
9
8
.
7
3
9
8
.
7
7
9
8
.
7
3
9
8
.
7
3
1
5
8
2
0
.
1
6
7
7
.
9
9
1
0
1
8
.
8
4
c
u
c
u
m
b
e
r
_
r
a
s
h
i
d
D
e
n
s
e
N
e
t
2
0
1
1
0
0
1
0
0
1
0
0
1
0
0
1
6
6
1
8
.
8
1
7
7
.
9
1
8
5
4
.
9
8
c
u
c
u
m
b
e
r
_
r
a
s
h
i
d
M
o
b
i
l
e
N
e
t
V
2
9
6
.
9
9
97
9
6
.
9
9
9
6
.
9
9
1
6
6
2
.
5
9
1
2
.
9
3
5
3
9
.
7
3
c
u
c
u
m
b
e
r
_
r
a
s
h
i
d
V
G
G
1
9
9
9
.
4
9
9
.
4
9
9
.
4
9
9
.
4
1
6
6
2
0
.
1
6
7
7
.
9
9
9
8
7
.
1
c
u
c
u
m
b
e
r
_
s
u
l
t
a
n
a
D
e
n
s
e
N
e
t
2
0
1
8
5
.
4
2
8
5
.
4
8
8
5
.
4
2
8
5
.
4
1
48
1
8
.
8
1
7
7
.
9
1
8
4
0
.
6
3
c
u
c
u
m
b
e
r
_
s
u
l
t
a
n
a
M
o
b
i
l
e
N
e
t
V
2
8
9
.
5
8
8
9
.
6
5
8
9
.
5
8
8
9
.
5
8
48
2
.
5
9
1
2
.
9
3
5
7
5
.
6
8
c
u
c
u
m
b
e
r
_
s
u
l
t
a
n
a
V
G
G
1
9
9
1
.
6
7
9
1
.
9
6
9
1
.
6
7
9
1
.
6
5
48
2
0
.
1
6
7
7
.
9
9
1
7
0
7
.
0
1
Ho
wev
er
,
all
m
o
d
els
f
ailed
o
n
cu
cu
m
b
e
r
Su
ltan
a
,
p
r
o
b
ab
ly
a
s
a
r
esu
lt
o
f
th
e
test
s
et’
s
s
ig
n
if
ican
tly
s
m
aller
s
am
p
le
s
ize
(
n
=4
8
)
,
wh
ich
r
ed
u
ce
s
r
ep
r
esen
tatio
n
d
i
v
er
s
ity
an
d
r
e
n
d
er
s
o
v
er
f
itti
n
g
p
o
s
s
ib
le
.
Nev
er
th
eless
,
VGG1
9
ac
h
iev
ed
m
o
d
er
ate
r
esil
ien
ce
,
attain
in
g
a
9
1
.
6
7
%
F1
-
s
co
r
e.
VGG1
9
an
d
Den
s
eNe
t2
0
1
p
r
o
v
id
e
s
u
p
er
io
r
ac
cu
r
ac
y
i
n
h
ig
h
-
ca
p
ac
ity
e
n
v
ir
o
n
m
en
ts
,
wh
er
ea
s
Mo
b
ileNetV2
p
r
o
v
id
es
an
ac
cu
r
a
cy
–
laten
cy
tr
ad
e
-
o
f
f
th
at
is
d
esir
ab
le
f
o
r
r
ea
l
-
tim
e
o
r
e
d
g
e
-
b
a
s
ed
ag
r
icu
ltu
r
al
s
y
s
tem
s
.
Mo
b
ileNetV2
r
em
ain
ed
th
e
f
astes
t
o
v
er
all,
b
u
t
Den
s
eNe
t2
0
1
co
n
s
is
ten
tly
p
er
f
o
r
m
e
d
b
etter
th
an
VGG1
9
co
n
s
id
e
r
in
g
in
f
er
e
n
ce
tim
es
(
e.
g
.
,
6
0
5
m
s
v
s
.
9
7
9
m
s
f
o
r
b
i
tter
g
o
u
r
d
)
,
p
r
o
v
i
n
g
its
ef
f
icie
n
cy
in
r
e
u
s
in
g
f
ea
tu
r
e
m
ap
s
.
Fig
u
r
e
2
co
n
f
u
s
io
n
m
atr
ices
f
o
r
all
t
h
e
m
o
d
els
s
h
o
w
th
at
m
is
clas
s
if
icatio
n
s
wer
e
lar
g
ely
r
estricte
d
to
th
e
cu
cu
m
b
er
Su
ltan
a
s
u
b
s
et.
T
h
e
m
o
d
els
lear
n
ed
h
i
g
h
ly
d
is
cr
im
in
ativ
e
b
o
u
n
d
ar
y
f
ea
tu
r
es
f
o
r
b
o
th
h
ea
lth
y
an
d
d
o
wn
y
m
ild
ew
ca
te
g
o
r
ie
s
,
as
d
em
o
n
s
tr
ated
b
y
th
e
p
er
f
ec
tly
d
iag
o
n
al
m
atr
ices
of
al
l
o
th
er
v
eg
eta
b
les
.
T
h
is
is
f
u
r
th
er
co
n
s
is
ten
t
with
th
e
q
u
a
n
titativ
e
ev
alu
atio
n
r
ep
r
esen
ted
in
T
a
b
le
3
,
in
wh
ich
th
r
ee
v
eg
eta
b
les
r
ea
ch
ed
p
r
ec
is
io
n
an
d
r
ec
all
v
alu
es
o
f
1
0
0
%,
i
n
d
icatin
g
co
m
p
lete
s
en
s
itiv
ity
an
d
s
p
ec
if
icity
f
o
r
th
e
id
en
tific
atio
n
o
f
d
is
ea
s
e
in
t
h
o
s
e
ca
s
es.
T
h
e
n
ee
d
f
o
r
b
alan
ce
d
s
am
p
lin
g
o
r
au
g
m
e
n
tatio
n
m
eth
o
d
s
is
r
ein
f
o
r
ce
d
b
y
th
e
d
is
p
ar
ity
in
p
er
f
o
r
m
an
ce
b
etwe
en
v
eg
etab
les
with
s
p
ar
s
e
d
ata
(
e.
g
.
,
cu
cu
m
b
er
Su
ltan
a
)
a
n
d
th
o
s
e
with
ab
u
n
d
an
t d
ata
(
e.
g
.
,
b
itter
g
o
u
r
d
an
d
b
o
ttle g
o
u
r
d
)
.
Ad
d
itio
n
ally
,
th
e
5
-
f
o
ld
cr
o
s
s
-
v
alid
atio
n
r
esu
lts
s
h
o
wed
tig
h
tly
clu
s
ter
ed
p
e
r
f
o
r
m
a
n
ce
m
etr
ics
f
o
r
all
v
eg
etab
les,
co
n
f
ir
m
in
g
th
at
th
e
m
o
d
els
r
eliab
ly
g
en
er
alize
d
ac
r
o
s
s
m
u
ltip
le
s
am
p
lin
g
d
is
tr
ib
u
tio
n
s
.
Desp
ite
class
im
b
alan
ce
,
Den
s
eNe
t2
0
1
ex
h
ib
ited
th
e
least
v
ar
iab
ilit
y
,
r
ef
lectin
g
s
tab
le
le
ar
n
in
g
.
Prio
r
to
test
-
s
et
ev
alu
atio
n
,
th
ese
cr
o
s
s
-
f
o
ld
tr
e
n
d
s
v
alid
ate
th
e
tr
ain
ed
m
o
d
els’
r
o
b
u
s
tn
ess
.
Fig
u
r
e
3
s
h
o
ws
t
h
e
Gr
a
d
-
C
AM
h
e
atm
a
p
s
c
o
n
f
ir
m
t
h
at
all
th
r
e
e
ar
c
h
it
ec
t
u
r
es
c
o
r
r
e
ct
ly
m
a
n
ag
e
s
y
m
p
t
o
m
-
r
i
ch
r
e
g
i
o
n
s
ass
o
ciat
ed
wit
h
d
o
w
n
y
m
il
d
ew
in
f
e
cti
o
n
.
T
h
e
at
te
n
ti
o
n
m
a
p
s
p
r
o
d
u
c
ed
b
y
De
n
s
e
Net
2
0
1
ar
e
c
o
n
s
is
te
n
t
ly
th
e
m
o
s
t
e
x
a
c
tly
d
e
f
i
n
ed
a
n
d
b
i
o
l
o
g
i
ca
ll
y
m
ea
n
i
n
g
f
u
l
,
cl
o
s
e
l
y
ca
p
t
u
r
i
n
g
n
ec
r
o
tic
b
o
u
n
d
a
r
i
es,
v
ei
n
-
li
m
it
e
d
les
io
n
s
,
a
n
d
ch
lo
r
o
tic
p
a
tc
h
es
.
W
it
h
its
l
ig
h
t
w
eig
h
t
ar
c
h
it
ec
t
u
r
e
a
n
d
r
e
d
u
ce
d
s
p
ati
al
s
e
lec
ti
v
it
y
,
Mo
b
il
eNe
tV
2
d
em
o
n
s
t
r
at
es
a
s
li
g
h
tl
y
b
r
o
ad
er
b
u
t
s
ti
ll
ac
c
u
r
at
e
f
o
c
u
s
.
VGG
1
9
’
s
d
e
ep
er
b
u
t
u
n
r
e
g
u
la
r
i
ze
d
f
ilt
e
r
s
p
r
o
v
id
e
v
ali
d
l
esi
o
n
lo
c
aliz
ati
o
n
,
b
u
t
w
it
h
o
cc
as
io
n
al
b
a
ck
g
r
o
u
n
d
at
te
n
ti
o
n
.
S
u
b
s
t
an
tiall
y
,
al
l
v
e
g
et
ab
les
wit
h
h
e
alt
h
y
l
ea
v
es
s
h
o
w
n
e
ar
-
u
n
if
o
r
m
h
ea
t
m
ap
s
wit
h
l
o
w
a
cti
v
at
i
o
n
,
in
d
i
ca
t
in
g
l
o
w
FP
s
e
n
s
it
iv
it
y
.
Fo
r
Mo
b
il
eNe
tV
2
a
n
d
VG
G1
9
,
h
ar
d
e
r
class
es
li
k
e
c
u
c
u
m
b
e
r
S
u
lt
an
a
ex
h
i
b
i
t
d
if
f
u
s
e
d
an
d
in
co
n
s
is
t
e
n
t
a
tte
n
t
io
n
th
a
t
d
i
r
e
ctl
y
ass
o
ci
ates
wit
h
t
h
ei
r
l
o
we
r
s
tatis
tic
al
p
er
f
o
r
m
an
ce
in
T
a
b
le
4
.
Ov
er
all
,
th
e
Gr
ad
-
C
A
M
r
es
u
lts
p
r
o
v
i
d
e
s
t
r
o
n
g
in
te
r
p
r
eta
b
i
lit
y
ev
i
d
en
ce
t
h
at
De
n
s
eNe
t2
0
1
is
th
e
b
est
c
h
o
ic
e
f
o
r
h
i
g
h
-
a
cc
u
r
ac
y
d
e
p
l
o
y
m
en
t
i
n
ag
r
i
cu
lt
u
r
al
d
ec
is
i
o
n
-
s
u
p
p
o
r
t
s
y
s
te
m
s
,
as
it
n
o
t
o
n
l
y
l
ea
r
n
s
th
e
m
o
s
t
d
is
cr
im
in
ati
v
e
a
n
d
s
y
m
p
to
m
-
f
ait
h
f
u
l
f
e
at
u
r
es
b
u
t
als
o
p
e
r
f
o
r
m
s
b
e
s
t
q
u
a
n
tit
ati
v
el
y
.
Usi
n
g
p
er
-
v
eg
eta
b
l
e
p
er
f
o
r
m
a
n
ce
v
al
u
es
,
p
ai
r
e
d
t
-
tes
ts
an
d
W
ilco
x
o
n
s
i
g
n
e
d
-
r
a
n
k
tes
ts
w
er
e
c
o
n
d
u
c
te
d
ac
r
o
s
s
m
o
d
e
ls
to
ass
ess
w
h
et
h
e
r
t
h
e
o
b
s
e
r
v
e
d
d
if
f
e
r
en
ce
s
w
er
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
C
o
mp
a
r
a
tive
d
ee
p
lea
r
n
in
g
s
tu
d
y
fo
r
d
o
w
n
y
mild
ew d
etec
tio
n
in
ve
g
eta
b
les
(
S
u
p
r
ee
th
a
S
h
i
va
r
a
j
)
1725
s
tatis
t
ic
all
y
s
i
g
n
i
f
ic
an
t
o
n
la
r
g
er
d
a
tase
ts
as
p
r
o
v
id
ed
i
n
t
h
e
s
tatis
t
ic
al
c
o
m
p
a
r
is
o
n
o
f
m
o
d
els,
s
u
m
m
ar
ize
d
in
T
a
b
le
4
.
W
h
i
le
d
if
f
e
r
en
ce
s
b
et
wee
n
VGG
1
9
a
n
d
M
o
b
ileN
et
V2
we
r
e
s
i
g
n
if
ic
an
t
f
o
r
m
o
s
t
v
eg
eta
b
l
es
b
u
t
n
o
t
al
l,
b
o
t
h
tes
ts
c
o
n
f
i
r
m
e
d
s
t
atis
t
ica
l
ly
s
i
g
n
i
f
ic
a
n
t
d
i
f
f
er
e
n
ce
s
b
etw
ee
n
De
n
s
e
Net
2
0
1
a
n
d
t
h
e
o
t
h
e
r
tw
o
a
r
ch
ite
ct
u
r
es
(p
<
0
.
0
5
)
.
T
h
is
c
o
n
f
i
r
m
s
t
h
a
t
m
o
d
el
d
is
c
r
e
p
a
n
cies
we
r
e
n
o
t
d
u
e
to
s
a
m
p
li
n
g
v
a
r
ia
n
c
e
a
n
d
q
u
a
n
tit
ati
v
el
y
v
al
id
at
es D
e
n
s
eN
et
2
0
1
’
s
s
u
p
e
r
io
r
it
y
.
B
i
t
t
e
r
_
G
o
u
r
d
B
o
t
t
l
e
_
G
o
u
r
d
C
a
u
l
i
f
l
o
w
e
r
C
u
c
u
m
b
e
r
_
R
a
s
h
i
d
C
u
c
u
m
b
e
r
_
S
u
l
t
a
n
a
Fig
u
r
e
2.
C
o
n
f
u
s
io
n
m
atr
ices
of
all
v
eg
etab
le
class
es
ex
am
in
ed
d
ee
p
lear
n
in
g
m
o
d
e
l
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
7
1
9
-
1
7
3
2
1726
B
i
t
t
e
r
_
g
o
u
r
d
—
D
e
n
s
e
N
e
t
2
0
1
B
i
t
t
e
r
_
g
o
u
r
d
—
M
o
b
i
l
e
N
e
t
V
2
B
i
t
t
e
r
_
g
o
u
r
d
—
V
G
G
1
9
B
o
t
t
l
e
_
G
o
u
r
d
—
D
e
n
s
e
N
e
t
2
0
1
B
o
t
t
l
e
_
G
o
u
r
d
—
M
o
b
i
l
e
N
e
t
V
2
B
o
t
t
l
e
_
G
o
u
r
d
—
V
G
G
1
9
C
a
u
l
i
f
l
o
w
e
r
—
D
e
n
seNe
t
2
0
1
C
a
u
l
i
f
l
o
w
e
r
—
M
o
b
i
l
e
N
e
t
V
2
C
a
u
l
i
f
l
o
w
e
r
—
V
G
G
1
9
C
u
c
u
m
b
e
r
R
a
sh
i
d
—
D
e
n
s
e
N
e
t
2
0
1
C
u
c
u
m
b
e
r
R
a
sh
i
d
—
M
o
b
i
l
e
N
e
t
V
2
C
u
c
u
m
b
e
r
R
a
sh
i
d
—
V
G
G
1
9
C
u
c
u
m
b
e
r
S
u
l
t
a
n
a
—
D
e
n
seNe
t
2
0
1
C
u
c
u
m
b
e
r
S
u
l
t
a
n
a
—
M
o
b
i
l
e
N
e
t
V
2
C
u
c
u
m
b
e
r
S
u
l
t
a
n
a
—
V
G
G
1
9
Fig
u
r
e
3
.
Gr
a
d
-
C
AM
r
ep
r
esen
tatio
n
o
f
m
o
d
el
atten
tio
n
f
o
r
th
e
d
etec
tio
n
o
f
d
o
wn
y
m
ild
ew
in
v
eg
etab
les
T
a
b
l
e
4
.
S
t
at
is
t
i
ca
l
s
i
g
n
i
f
i
c
a
n
c
e
a
n
a
l
y
s
is
c
o
m
p
a
r
i
n
g
b
e
s
t
a
n
d
n
e
x
t
-
b
e
s
t
m
o
d
e
ls
a
c
r
o
s
s
v
e
g
e
ta
b
l
e
s
(
p
a
i
r
e
d
t
-
t
e
s
t
/W
il
c
o
x
o
n
)
V
e
g
e
t
a
b
l
e
M
e
t
r
i
c
B
e
st
N
e
x
t
M
e
a
n
b
e
st
(
%)
M
e
a
n
n
e
x
t
(
%)
∆
M
e
a
n
(
%)
p
-
v
a
l
u
e
S
i
g
.
I
n
t
e
r
p
r
e
t
a
t
i
o
n
b
i
t
t
e
r
g
o
u
r
d
A
c
c
u
r
a
c
y
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
9
9
.
5
8
0
.
4
2
0
.
1
8
3
5
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
b
i
t
t
e
r
g
o
u
r
d
F1
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
9
9
.
5
4
0
.
4
6
0
.
1
8
4
4
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
b
i
t
t
e
r
g
o
u
r
d
P
r
e
c
i
s
i
o
n
D
e
n
seN
e
t
2
0
1
V
G
G
1
9
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
N
o
d
i
f
f
e
r
e
n
c
e
i
n
p
r
e
c
i
si
o
n
.
b
i
t
t
e
r
g
o
u
r
d
R
e
c
a
l
l
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
N
o
d
i
f
f
e
r
e
n
c
e
i
n
r
e
c
a
l
l
.
b
o
t
t
l
e
g
o
u
r
d
A
c
c
u
r
a
c
y
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
b
o
t
t
l
e
g
o
u
r
d
F1
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
b
o
t
t
l
e
g
o
u
r
d
P
r
e
c
i
s
i
o
n
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
b
o
t
t
l
e
g
o
u
r
d
R
e
c
a
l
l
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
c
a
u
l
i
f
l
o
w
e
r
A
c
c
u
r
a
c
y
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
c
a
u
l
i
f
l
o
w
e
r
F1
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
c
a
u
l
i
f
l
o
w
e
r
P
r
e
c
i
s
i
o
n
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
c
a
u
l
i
f
l
o
w
e
r
R
e
c
a
l
l
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
c
u
c
u
m
b
e
r
r
a
sh
i
d
A
c
c
u
r
a
c
y
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
c
u
c
u
m
b
e
r
r
a
sh
i
d
F1
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
c
u
c
u
m
b
e
r
r
a
sh
i
d
P
r
e
c
i
s
i
o
n
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
c
u
c
u
m
b
e
r
r
a
sh
i
d
R
e
c
a
l
l
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
1
0
0
.
0
0
1
0
0
.
0
0
0
.
0
0
1
.
0
0
0
0
×
M
o
d
e
l
s
p
e
r
f
o
r
m
e
q
u
i
v
a
l
e
n
t
l
y
.
c
u
c
u
m
b
e
r
su
l
t
a
n
a
A
c
c
u
r
a
c
y
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
9
8
.
2
0
8
3
.
0
1
1
5
.
1
9
0
.
0
0
1
4
✓
D
e
n
seN
e
t
2
0
1
s
i
g
n
i
f
i
c
a
n
t
l
y
b
e
t
t
e
r
;
l
a
r
g
e
g
a
p
.
c
u
c
u
m
b
e
r
su
l
t
a
n
a
F1
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
9
8
.
1
2
8
5
.
3
6
1
2
.
7
6
0
.
0
0
1
7
✓
D
e
n
seN
e
t
2
0
1
si
g
n
i
f
i
c
a
n
t
l
y
b
e
t
t
e
r
o
n
F
1
.
c
u
c
u
m
b
e
r
su
l
t
a
n
a
P
r
e
c
i
s
i
o
n
D
e
n
seN
e
t
2
0
1
M
o
b
i
l
e
N
e
t
V
2
9
7
.
1
9
7
4
.
7
6
2
2
.
4
3
0
.
0
0
1
0
✓
D
e
n
seN
e
t
2
0
1
si
g
n
i
f
i
c
a
n
t
l
y
b
e
t
t
e
r
i
n
p
r
e
c
i
si
o
n
.
c
u
c
u
m
b
e
r
su
l
t
a
n
a
R
e
c
a
l
l
M
o
b
i
l
e
N
e
t
V
2
D
e
n
seN
e
t
2
0
1
1
0
0
.
0
0
9
9
.
0
9
0
.
9
1
0
.
3
7
3
9
×
N
o
si
g
n
i
f
i
c
a
n
t
d
i
f
f
e
r
e
n
c
e
i
n
r
e
c
a
l
l
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ar
tif
I
n
tell
I
SS
N:
2252
-
8
9
3
8
C
o
mp
a
r
a
tive
d
ee
p
lea
r
n
in
g
s
tu
d
y
fo
r
d
o
w
n
y
mild
ew d
etec
tio
n
in
ve
g
eta
b
les
(
S
u
p
r
ee
th
a
S
h
i
va
r
a
j
)
1727
3
.
1
.
L
ea
rning
curv
e
a
na
ly
s
is
T
h
e
tr
ain
in
g
a
n
d
v
alid
atio
n
lo
s
s
an
d
ac
cu
r
ac
y
cu
r
v
es
f
o
r
th
e
VGG1
9
,
De
n
s
eNe
t2
0
1
,
an
d
Mo
b
ileNetV2
m
o
d
els tr
ain
ed
o
n
th
e
f
iv
e
-
v
e
g
etab
le
d
ataset
ar
e
r
ep
r
esen
ted
in
Fig
u
r
e
4
.
E
f
f
ec
tiv
e
o
p
tim
izatio
n
is
o
b
s
er
v
ed
b
y
all
ar
ch
itectu
r
es’
co
n
s
is
ten
t
co
n
v
er
g
e
n
ce
,
m
o
n
o
to
n
ic
l
o
s
s
d
ec
r
ea
s
es
,
an
d
co
n
s
is
ten
t
ac
cu
r
ac
y
in
cr
ea
s
e.
W
h
ile
VGG1
9
d
em
o
n
s
tr
ates
g
r
ad
u
al
b
u
t
s
tead
y
lear
n
in
g
,
Den
s
eNe
t2
0
1
a
n
d
Mo
b
i
leNe
tV2
co
n
v
er
g
e
m
o
r
e
r
ap
id
ly
a
n
d
d
em
o
n
s
tr
ate
h
ig
h
v
alid
atio
n
ac
cu
r
ac
y
in
f
e
wer
ep
o
ch
s
.
I
m
p
o
r
tan
tly
,
all
m
o
d
els’
tr
ain
in
g
an
d
v
alid
atio
n
cu
r
v
es
alig
n
cl
o
s
ely
,
in
d
icatin
g
m
in
im
al
o
v
er
f
i
ttin
g
.
I
t
ca
n
b
e
clar
if
ie
d
b
y
th
e
ea
r
ly
s
to
p
p
in
g
s
tr
ateg
y
,
d
r
o
p
o
u
t
r
eg
u
lar
izati
o
n
,
a
n
d
ap
p
lied
d
ata
au
g
m
e
n
tatio
n
.
T
h
ese
o
u
tco
m
es
c
o
n
f
ir
m
th
e
p
r
o
p
o
s
ed
m
o
d
els’
r
o
b
u
s
tn
ess
an
d
ab
ilit
y
f
o
r
g
e
n
er
aliza
tio
n
u
n
d
e
r
r
ea
l
-
wo
r
ld
v
ar
ia
b
ilit
y
.
Fig
u
r
e
4
.
L
ea
r
n
i
n
g
cu
r
v
es
s
h
o
win
g
tr
ain
in
g
an
d
v
alid
atio
n
lo
s
s
an
d
ac
cu
r
ac
y
f
o
r
VGG1
9
,
Den
s
eNe
t2
0
1
,
an
d
Mo
b
ileNetV2
m
o
d
els
3
.
2
.
Cla
s
s
im
ba
la
nce
a
na
l
y
s
is
Mu
l
ti
p
le
im
b
ala
n
c
e
-
h
an
d
l
in
g
s
tr
a
te
g
ies
,
i
n
cl
u
d
in
g
b
aseli
n
e
t
r
ain
in
g
wit
h
o
u
t b
al
an
ci
n
g
,
cl
ass
-
we
ig
h
t
e
d
lo
s
s
,
an
d
f
o
ca
l
lo
s
s
,
we
r
e
s
y
s
tem
ati
ca
l
ly
ev
al
u
at
e
d
to
u
n
d
e
r
s
ta
n
d
t
h
e
i
m
p
ac
t
o
f
cl
ass
im
b
al
a
n
ce
o
n
m
o
d
el
p
e
r
f
o
r
m
a
n
c
e
.
T
o
e
n
s
u
r
e
a
f
ai
r
co
m
p
a
r
is
o
n
,
i
d
e
n
ti
ca
l
d
at
a
s
p
li
ts
,
n
etw
o
r
k
a
r
ch
ite
ct
u
r
es,
a
n
d
h
y
p
er
p
ar
am
ete
r
s
etti
n
g
s
w
er
e
u
s
ed
i
n
all
e
x
p
e
r
i
m
e
n
ts
.
St
r
o
n
g
b
ias
t
o
wa
r
d
m
aj
o
r
it
y
cl
ass
es
in
t
h
e
b
ase
li
n
e
m
o
d
e
l
l
e
d
t
o
p
o
o
r
m
ac
r
o
-
a
v
e
r
a
g
e
d
p
e
r
f
o
r
m
a
n
c
e
a
n
d
al
m
o
s
t
z
er
o
r
e
ca
ll
f
o
r
m
i
n
o
r
i
ty
ca
t
eg
o
r
i
es.
Si
g
n
i
f
ic
a
n
t
g
ai
n
s
w
er
e
n
o
t
g
e
n
e
r
a
te
d
b
y
ap
p
ly
in
g
c
lass
w
ei
g
h
ts
,
an
d
i
t
als
o
l
ed
to
a
s
im
i
la
r
p
r
ed
ict
io
n
co
lla
p
s
e.
C
o
n
v
e
r
s
el
y
,
b
y
em
p
h
asi
zi
n
g
h
ar
d
a
n
d
m
i
n
o
r
ity
s
am
p
les
d
u
r
in
g
tr
ai
n
i
n
g
,
f
o
ca
l
l
o
s
s
s
i
g
n
i
f
ic
an
tl
y
e
n
h
a
n
ce
d
d
is
c
r
i
m
i
n
at
io
n
ac
r
o
s
s
a
ll
cl
ass
es
.
W
it
h
a
tes
t
a
cc
u
r
ac
y
o
f
9
9
.
7
%,
a
m
a
c
r
o
-
a
v
e
r
a
g
ed
F
1
-
s
c
o
r
e
o
f
0
.
9
9
,
a
n
d
b
a
la
n
c
ed
r
e
ca
l
l
ac
r
o
s
s
a
ll
f
i
v
e
class
es,
t
h
e
D
e
n
s
eNe
t
2
0
1
m
o
d
el
t
r
ai
n
e
d
wi
th
f
o
c
al
l
o
s
s
t
h
u
s
ac
h
i
e
v
e
d
n
ea
r
-
p
er
f
e
ct
cl
ass
i
f
ic
ati
o
n
p
e
r
f
o
r
m
a
n
c
e.
Fo
ca
l
l
o
s
s
is
a
n
e
f
f
ec
t
iv
e
an
d
r
eli
ab
le
m
e
th
o
d
to
m
iti
g
a
te
cl
ass
i
m
b
al
a
n
ce
i
n
m
u
lti
-
c
lass
p
l
a
n
t
d
is
e
ase
c
lass
i
f
i
ca
t
io
n
.
T
h
e
i
m
p
ac
t
o
f
d
i
f
f
e
r
e
n
t
c
lass
i
m
b
ala
n
ce
h
a
n
d
li
n
g
s
t
r
ate
g
i
e
s
o
n
c
lass
i
f
i
ca
t
io
n
p
e
r
f
o
r
m
a
n
c
e
is
q
u
a
n
tit
ati
v
el
y
p
r
o
v
i
d
e
d
i
n
T
a
b
l
e
5
.
T
ab
le
5
.
Per
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
o
f
class
im
b
alan
ce
h
a
n
d
lin
g
s
tr
ateg
ies o
n
th
e
test
d
ataset
I
mb
a
l
a
n
c
e
st
r
a
t
e
g
y
A
c
c
u
r
a
c
y
M
a
c
r
o
p
r
e
c
i
s
i
o
n
M
a
c
r
o
r
e
c
a
l
l
M
a
c
r
o
F
1
-
sc
o
r
e
B
a
se
l
i
n
e
(
n
o
b
a
l
a
n
c
i
n
g
)
0
.
2
3
0
.
0
5
0
.
2
0
0
.
0
8
C
l
a
s
s
w
e
i
g
h
t
s
0
.
2
3
0
.
0
5
0
.
2
0
0
.
0
8
F
o
c
a
l
l
o
s
s
0
.
9
9
7
0
.
9
9
0
.
9
9
0
.
9
9
3
.
3
.
E
rr
o
r
a
na
ly
s
is
Usi
n
g
Gr
ad
-
C
AM
v
is
u
aliza
ti
o
n
s
o
n
test
s
am
p
les
th
at
ar
e
m
is
class
if
ied
,
a
q
u
alitativ
e
er
r
o
r
an
aly
s
is
was
co
n
d
u
cted
to
g
et
b
etter
in
s
ig
h
t
in
to
th
e
f
ailu
r
e
m
o
d
e
s
o
f
th
e
p
r
o
p
o
s
ed
m
o
d
els.
A
lth
o
u
g
h
th
e
o
v
er
all
class
if
icatio
n
ac
cu
r
ac
y
is
o
v
er
9
8
%,
th
er
e
r
em
ain
a
f
ew
ch
al
len
g
in
g
ca
s
es
th
at
ar
e
p
r
im
ar
il
y
ch
ar
ac
ter
ized
b
y
b
ac
k
g
r
o
u
n
d
clu
tter
,
n
o
n
-
u
n
if
o
r
m
illu
m
in
atio
n
,
an
d
s
u
b
tle
s
y
m
p
to
m
ex
p
r
ess
io
n
.
T
h
e
Gr
ad
-
C
AM
m
ap
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
2
5
2
-
8
9
3
8
I
n
t J Ar
tif
I
n
tell
,
Vo
l.
1
5
,
No
.
2
,
Ap
r
il 2
0
2
6
:
1
7
1
9
-
1
7
3
2
1728
co
r
r
esp
o
n
d
in
g
t
o
th
e
tr
u
e
an
d
p
r
ed
icted
class
es
ar
e
co
m
p
ar
e
d
in
Fig
u
r
e
5
,
wh
ich
d
ep
icts
r
e
p
r
esen
tativ
e
f
ailu
r
e
ca
s
es
f
o
r
d
if
f
er
en
t
v
eg
etab
le
c
r
o
p
s
.
T
h
e
o
r
ig
in
al
leaf
im
a
g
e
i
s
p
r
o
v
id
ed
i
n
ea
ch
e
x
am
p
le,
a
lo
n
g
with
atten
tio
n
m
ap
s
f
o
r
th
e
tr
u
e
an
d
p
r
e
d
icted
class
es
th
at
h
ig
h
lig
h
t
th
e
r
eg
io
n
s
th
at
r
esu
lted
in
in
co
r
r
e
ct
m
o
d
el
d
ec
is
io
n
s
.
T
h
is
h
ap
p
en
s
wh
en
th
e
m
o
d
e
l’
s
atten
tio
n
s
h
if
ts
f
r
o
m
lo
ca
lized
d
is
ea
s
e
s
y
m
p
to
m
s
to
n
o
n
-
d
iag
n
o
s
tic
r
eg
io
n
s
s
u
ch
as
leaf
ed
g
es,
b
ac
k
g
r
o
u
n
d
o
b
jects,
o
r
u
n
if
o
r
m
ly
tex
t
u
r
ed
ar
ea
s
.
I
n
a
d
d
itio
n
,
s
am
p
les
with
v
is
u
ally
am
b
ig
u
o
u
s
p
atter
n
s
o
r
ea
r
ly
-
s
tag
e
in
f
ec
tio
n
s
ten
d
to
p
r
o
d
u
ce
d
if
f
u
s
e
ac
tiv
atio
n
m
a
p
s
,
d
em
o
n
s
tr
atin
g
r
e
d
u
ce
d
d
is
cr
im
in
ativ
e
co
n
f
id
en
ce
.
T
h
ese
o
b
s
er
v
atio
n
s
h
ig
h
lig
h
t
r
ea
l
ch
allen
g
es
in
f
ield
-
ac
q
u
ir
ed
im
ag
e
r
y
a
n
d
s
u
g
g
est
th
at
m
is
class
if
icatio
n
s
ar
e
p
r
ed
o
m
in
an
tly
d
r
iv
e
n
b
y
s
y
m
p
to
m
h
eter
o
g
en
eity
an
d
en
v
ir
o
n
m
e
n
tal
v
ar
iab
ilit
y
r
ath
er
th
an
m
o
d
el
i
n
s
tab
ilit
y
.
B
i
t
t
e
r
g
o
u
r
d
C
a
u
l
i
f
l
o
w
e
r
C
u
c
u
m
b
e
r
_
R
a
s
h
i
d
C
u
c
u
m
b
e
r
_
S
u
l
t
a
n
a
Fig
u
r
e
5
.
Gr
a
d
-
C
AM
an
aly
s
is
f
o
r
r
e
p
r
esen
tativ
e
m
is
class
if
ic
atio
n
3
.
4
.
Abla
t
io
n
s
t
ud
y
W
e
co
n
d
u
cted
ab
latio
n
s
tu
d
ie
s
with
a
f
o
cu
s
o
n
d
ata
au
g
m
e
n
tatio
n
an
d
tr
a
n
s
f
er
lear
n
in
g
in
an
ef
f
o
r
t
to
q
u
an
tify
th
e
ef
f
ec
t
o
f
k
ey
d
esig
n
ch
o
ices.
First
,
we
ev
alu
ated
Den
s
eNe
t2
0
1
with
an
d
wit
h
o
u
t
au
g
m
en
tatio
n
u
n
d
er
th
e
id
e
n
tical
s
ettin
g
s
to
ev
alu
ate
th
e
im
p
ac
t
o
f
a
u
g
m
en
tatio
n
ac
r
o
s
s
f
iv
e
cr
o
p
d
atasets
.
T
ab
le
6
h
ig
h
lig
h
ts
th
at
au
g
m
en
tatio
n
co
n
s
is
ten
tly
s
tab
ilized
o
r
im
p
r
o
v
ed
p
e
r
f
o
r
m
an
ce
,
p
a
r
ticu
lar
ly
in
cr
o
p
s
s
u
ch
as
cu
cu
m
b
er
Su
ltan
a
with
f
ew
t
r
ain
in
g
s
am
p
les.
Seco
n
d
,
b
y
tr
ain
in
g
th
e
m
o
d
el
with
an
d
with
o
u
t
I
m
ag
eNe
t
p
r
etr
ain
in
g
,
we
ass
ess
ed
th
e
r
o
le
o
f
tr
an
s
f
er
lear
n
in
g
u
s
in
g
a
r
ep
r
esen
tativ
e
cr
o
p
(
b
itter
g
o
u
r
d
)
.
T
h
e
m
o
d
el
f
ailed
to
ac
h
iev
e
n
ea
r
-
r
an
d
o
m
p
er
f
o
r
m
an
ce
(
ac
cu
r
ac
y
:
5
0
.
6
%,
F1
-
s
co
r
e:
0
.
0
0
)
o
n
th
e
r
em
o
v
al
o
f
tr
an
s
f
er
lear
n
in
g
as
p
er
th
e
r
esu
lts
in
T
ab
le
7
,
s
in
ce
th
e
n
etwo
r
k
c
o
n
v
er
g
ed
f
o
r
p
r
e
d
ictin
g
o
n
ly
th
e
m
ajo
r
ity
class
.
T
h
e
p
r
etr
ain
ed
m
o
d
el,
in
c
o
n
tr
ast,
ac
h
iev
ed
s
tr
o
n
g
g
en
er
aliza
ti
o
n
(
Acc
u
r
ac
y
:
9
8
.
2
%,
F1
-
s
co
r
e:
9
8
.
2
%).
T
h
ese
r
esu
lts
co
n
f
ir
m
th
at
b
o
th
tr
a
n
s
f
er
lear
n
in
g
an
d
au
g
m
en
tatio
n
ar
e
cr
itical
f
o
r
r
eliab
le
p
er
f
o
r
m
an
ce
o
n
lim
ited
ag
r
icu
ltu
r
al
d
atasets
.
T
o
ass
ess
th
e
im
p
ac
t
o
f
a
r
ch
itectu
r
a
l
co
m
p
lex
ity
o
n
class
if
icatio
n
p
er
f
o
r
m
a
n
ce
,
an
ab
latio
n
s
tu
d
y
was
co
n
d
u
cted
.
T
h
e
lig
h
tweig
h
t
Mo
b
ileNetV2
m
o
d
el
o
f
f
er
s
f
aster
i
n
f
er
e
n
c
e
an
d
r
ed
u
ce
d
c
o
s
t
o
f
co
m
p
u
tatio
n
,
as
T
ab
le
3
s
u
m
m
ar
izes,
b
u
t
it
ex
h
ib
its
a
n
o
ticea
b
le
r
ed
u
ctio
n
in
p
er
f
o
r
m
an
ce
o
n
v
is
u
ally
co
m
p
lex
d
is
ea
s
e
p
atter
n
s
.
T
h
e
h
ea
v
y
weig
h
t
Den
s
eNe
t2
0
1
,
in
co
n
tr
ast
,
co
n
s
is
ten
tly
ac
h
iev
es
s
u
p
er
io
r
ac
cu
r
ac
y
an
d
F1
-
s
co
r
e
ac
r
o
s
s
all
v
eg
etab
les,
d
em
o
n
s
tr
atin
g
im
p
r
o
v
e
d
d
is
cr
im
in
ativ
e
ca
p
ac
ity
an
d
f
ea
tu
r
e
r
eu
s
e.
As
an
in
ter
m
ed
iate
b
aselin
e,
VGG1
9
co
n
f
ir
m
s
a
cl
ea
r
tr
ad
e
-
o
f
f
b
etwe
en
ac
c
u
r
a
cy
an
d
ef
f
icien
cy
.
T
h
ese
r
esu
lts
ju
s
tify
th
e
s
elec
t
io
n
o
f
Den
s
eNe
t2
0
1
f
o
r
r
eliab
le
d
iag
n
o
s
is
o
f
p
lan
t
ailm
en
ts
wh
er
e
ac
cu
r
ac
y
is
p
r
io
r
itized
o
v
er
m
in
i
m
al
m
o
d
e
l size.
Evaluation Warning : The document was created with Spire.PDF for Python.