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.
15
,
No
.
2
,
A
p
r
il
2
0
2
6
,
p
p
.
115
5
~
1
1
6
5
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
1
5
5
-
116
5
1155
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
i
.
ia
esco
r
e.
co
m
G
ra
dien
t
-
ba
sed s
t
o
cha
stic dept
h wi
th
co
nv
o
lutiona
l
neura
l
network
for coco
nut
t
ree
l
ea
f
disea
se cla
ss
ificatio
n
K
a
v
it
ha
M
a
g
a
di G
o
pa
la
k
ris
hn
a
1
,
Ra
v
ipra
k
a
s
h M
a
denu
r
L
ing
a
ra
j
u
2
,
Ana
nd
a
B
a
bu
J
a
y
a
cha
nd
ra
3
1
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
,
K
a
l
p
a
t
a
r
u
I
n
st
i
t
u
t
e
o
f
Te
c
h
n
o
l
o
g
y
,
V
i
sv
e
sv
a
r
a
y
a
Te
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
si
t
y
,
B
e
l
a
g
a
v
i
,
I
n
d
i
a
2
D
e
p
a
r
t
me
n
t
o
f
A
r
t
i
f
i
c
i
a
l
i
n
t
e
l
l
i
g
e
n
c
e
a
n
d
M
a
c
h
i
n
e
L
e
a
r
n
i
n
g
,
K
a
l
p
a
t
a
r
u
I
n
st
i
t
u
t
e
o
f
Te
c
h
n
o
l
o
g
y
,
V
i
s
v
e
s
v
a
r
a
y
a
T
e
c
h
n
o
l
o
g
i
c
a
l
U
n
i
v
e
r
s
i
t
y
,
B
e
l
a
g
a
v
i
,
I
n
d
i
a
3
D
e
p
a
r
t
me
n
t
o
f
I
n
f
o
r
mat
i
o
n
S
c
i
e
n
c
e
a
n
d
E
n
g
i
n
e
e
r
i
n
g
,
M
a
l
n
a
d
C
o
l
l
e
g
e
o
f
E
n
g
i
n
e
e
r
i
n
g
,
H
a
ssa
n
,
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
Mar
15
,
2
0
2
5
R
ev
is
ed
J
an
20
,
2
0
2
6
Acc
ep
ted
Feb
6
,
2
0
2
6
Th
e
c
o
c
o
n
u
t
p
a
lm
(
Co
c
o
s
n
u
c
if
e
ra
)
is
v
it
a
l
p
lan
tatio
n
c
ro
p
,
v
a
l
u
e
d
fo
r
t
h
e
ir
d
iffere
n
t
u
se
s,
ra
n
g
i
n
g
fro
m
t
h
e
ir
fru
i
t
to
it
s
tru
n
k
.
I
n
re
c
e
n
t
ti
m
e
s,
it
h
a
s
b
e
e
n
o
b
se
rv
e
d
t
h
a
t
m
a
n
y
c
o
c
o
n
u
t
tree
s
a
re
a
ffe
c
ted
b
y
d
ise
a
se
s
t
h
a
t
re
d
u
c
e
p
ro
d
u
c
ti
o
n
a
n
d
we
a
k
e
n
t
h
e
stre
n
g
t
h
o
f
t
h
e
c
o
c
o
n
u
t.
T
h
e
c
las
sifica
ti
o
n
o
f
c
o
c
o
n
u
t
lea
f
d
ise
a
se
s
is
c
h
a
ll
e
n
g
in
g
b
e
c
a
u
se
o
f
i
n
tra
-
c
las
s
a
n
d
in
ter
-
c
las
s
v
a
riab
il
it
y
.
Th
is
re
se
a
rc
h
in
tr
o
d
u
c
e
s
th
e
g
ra
d
ien
t
-
b
a
se
d
sto
c
h
a
stic
d
e
p
t
h
(G
S
D)
with
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
r
k
(CNN
)
tec
h
n
i
q
u
e
to
c
o
c
o
n
u
t
lea
f
d
ise
a
se
c
las
sific
a
ti
o
n
to
o
v
e
rc
o
m
e
th
e
se
c
h
a
ll
e
n
g
e
s.
Th
e
G
S
D
tec
h
n
i
q
u
e
is
in
c
o
rp
o
ra
ted
i
n
to
e
v
e
r
y
lay
e
r
o
f
t
h
e
CNN
,
wh
e
re
it
c
a
lcu
late
s
th
e
p
ro
b
a
b
il
it
y
u
sin
g
g
ra
d
ien
t
m
a
g
n
it
u
d
e
s
a
n
d
s
k
ip
s
lay
e
rs
th
a
t
c
o
n
tri
b
u
te
m
in
ima
ll
y
t
o
t
h
e
c
las
sifica
ti
o
n
.
Th
e
ima
g
e
s
a
re
se
g
m
e
n
ted
u
si
n
g
t
h
e
G
ra
b
Cu
t
se
g
m
e
n
tatio
n
a
lg
o
rit
h
m
,
w
h
ich
iso
late
s
th
e
lea
f
fro
m
th
e
b
a
c
k
g
ro
u
n
d
u
sin
g
g
r
a
p
h
-
b
a
se
d
se
g
m
e
n
tatio
n
,
h
e
lp
i
n
g
t
o
d
iffere
n
ti
a
te
b
e
twe
e
n
v
a
rio
u
s
d
ise
a
se
c
la
ss
e
s.
Th
e
G
S
D
with
CNN
a
lg
o
rit
h
m
o
b
ta
in
s
a
n
a
c
c
u
ra
c
y
o
f
9
6
.
4
2
%
,
p
re
c
isio
n
o
f
9
6
.
1
5
%
,
re
c
a
ll
o
f
9
5
.
8
7
%
,
a
n
d
F
1
-
sc
o
re
o
f
9
5
.
9
3
%
,
wh
i
le
c
o
m
p
a
rin
g
wit
h
e
x
isti
n
g
a
lg
o
rit
h
m
s.
K
ey
w
o
r
d
s
:
C
o
co
n
u
t tr
ee
d
is
ea
s
e
C
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
Dee
p
lear
n
in
g
Gr
ab
C
u
t seg
m
en
tatio
n
Gr
ad
ien
t m
ag
n
itu
d
es
Gr
ad
ien
t
-
b
ased
s
to
ch
asti
c
d
ep
t
h
I
n
ter
-
an
d
in
tr
a
-
class
v
ar
iab
ilit
y
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
:
Kav
ith
a
Ma
g
ad
i G
o
p
ala
k
r
is
h
n
a
Dep
ar
tm
en
t o
f
C
o
m
p
u
ter
Scie
n
ce
an
d
E
n
g
in
ee
r
in
g
,
Kalp
atar
u
I
n
s
titu
te
o
f
T
ec
h
n
o
l
o
g
y
Vis
v
esv
ar
ay
a
T
ec
h
n
o
l
o
g
ical
Un
iv
er
s
ity
B
elag
av
i,
I
n
d
ia
E
m
ail:
k
av
ith
a.
m
g
@
k
itti
p
tu
r
.
a
c.
in
1.
I
NT
RO
D
UCT
I
O
N
I
n
r
ec
e
n
t
y
ea
r
s
,
a
s
ig
n
if
ican
t d
ec
lin
e
in
ag
r
ic
u
ltu
r
al
p
r
o
d
u
ctiv
ity
h
as
p
o
s
ed
a
th
r
ea
t
to
th
e
g
lo
b
al
f
o
o
d
s
u
p
p
ly
[
1
]
.
T
h
e
co
c
o
n
u
t
tr
ee
is
a
v
ital
p
lan
tatio
n
cr
o
p
d
u
e
t
o
its
d
iv
er
s
e
an
d
u
s
ef
u
l
p
r
o
d
u
cts
[
2
]
.
Glo
b
ally
,
in
m
an
y
r
eg
i
o
n
s
,
th
e
ex
ce
p
tio
n
al
n
u
tr
itio
n
al
v
alu
e
a
n
d
v
ar
iety
o
f
co
co
n
u
t
-
b
ased
p
r
o
d
u
cts
h
av
e
m
ad
e
it
ess
en
tial
f
o
r
d
aily
life
a
n
d
n
u
m
er
o
u
s
in
d
u
s
tr
ies
[
3
]
.
Desp
ite
th
e
r
is
in
g
d
em
an
d
f
r
o
m
co
n
s
u
m
e
r
s
,
m
a
n
y
co
co
n
u
t
f
a
r
m
er
s
s
tr
u
g
g
le
to
m
a
k
e
en
d
s
m
ee
t
d
u
e
to
h
ig
h
lab
o
r
co
s
ts
an
d
th
e
tim
e
-
co
n
s
u
m
in
g
p
r
o
ce
s
s
o
f
p
r
o
d
u
cin
g
c
o
co
n
u
t
p
r
o
d
u
cts
[
4
]
.
C
o
co
n
u
ts
ar
e
a
r
ich
s
o
u
r
ce
o
f
v
itam
in
s
,
m
in
er
als,
an
d
f
ib
er
,
m
ak
i
n
g
t
h
em
a
h
ea
lth
y
s
n
ac
k
[
5
]
.
Ad
d
itio
n
ally
,
th
e
y
s
er
v
e
as
a
n
im
p
o
r
tan
t
s
o
u
r
ce
o
f
n
u
tr
itio
n
an
d
ar
e
wid
ely
u
s
ed
in
th
e
p
r
o
d
u
ctio
n
o
f
f
u
el,
f
o
o
d
,
a
n
d
m
e
d
icin
es
[
6
]
.
T
h
e
co
co
n
u
t
f
r
u
it,
wh
ich
d
ev
elo
p
s
o
n
co
co
n
u
t
p
alm
s
,
h
as
b
ee
n
u
s
ed
f
o
r
ce
n
t
u
r
ies,
an
d
h
u
m
a
n
s
h
av
e
r
elied
o
n
it
f
o
r
th
eir
b
asic
n
ee
d
s
[
7
]
.
I
n
r
e
ce
n
t
tim
es,
p
eo
p
le
o
f
v
ar
i
o
u
s
f
aith
s
an
d
cu
ltu
r
es
h
av
e
r
ec
o
g
n
ized
co
co
n
u
t
as
a
v
alu
ab
le
m
ed
icin
al
an
d
n
u
tr
itio
n
al
r
eso
u
r
ce
[
8
]
.
Ho
wev
er
,
c
er
tain
d
is
ea
s
es
th
at
attac
k
th
e
co
c
o
n
u
t
lea
v
es
wea
k
en
th
em
o
v
e
r
tim
e,
lea
d
in
g
t
o
cr
o
p
lo
s
s
es.
Far
m
er
s
u
s
in
g
co
n
v
en
tio
n
al
co
co
n
u
t
f
ar
m
in
g
tech
n
iq
u
es
ca
n
o
n
ly
o
b
s
er
v
e
th
e
l
o
west
leav
es
o
f
t
h
e
p
lan
t
cr
o
wn
s
f
r
o
m
th
e
g
r
o
u
n
d
[
9
]
.
C
o
n
v
en
tio
n
al
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.
15
,
No
.
2
,
Ap
r
il
20
26
:
115
5
-
1
1
6
5
1156
im
ag
in
g
tech
n
iq
u
es,
wh
ich
ca
p
tu
r
e
p
h
o
to
s
f
r
o
m
h
ig
h
altitu
d
es
o
v
er
ex
ten
s
iv
e
ar
ea
s
,
a
r
e
b
o
th
ch
allen
g
i
n
g
an
d
tim
e
-
co
n
s
u
m
in
g
[
1
0
]
.
T
h
is
p
r
o
ce
s
s
r
eq
u
ir
es
m
an
u
al
in
v
o
lv
e
m
en
t,
wh
ich
is
co
s
tly
,
tim
e
-
co
n
s
u
m
in
g
,
a
n
d
p
r
o
n
e
to
er
r
o
r
s
.
Mo
r
eo
v
e
r
,
th
e
tr
ee
s
ar
e
s
u
s
ce
p
tib
le
to
v
ar
io
u
s
d
is
ea
s
es
th
at
r
ed
u
ce
th
eir
p
r
o
d
u
ctiv
ity
an
d
ca
n
u
ltima
tely
lead
to
th
eir
d
ea
t
h
[
1
1
]
–
[
1
3
]
.
E
ar
ly
d
etec
tio
n
an
d
d
iag
n
o
s
is
o
f
th
ese
d
is
e
ases
ar
e
cr
u
cial
f
o
r
ef
f
ec
tiv
e
d
is
ea
s
e
m
an
ag
e
m
en
t
an
d
p
r
ev
e
n
tin
g
f
u
r
th
e
r
s
p
r
ea
d
[
1
4
]
.
Ma
c
h
in
e
l
ea
r
n
in
g
(
M
L
)
-
b
ased
alg
o
r
ith
m
s
ar
e
m
ajo
r
ly
u
tili
ze
d
in
ag
r
icu
ltu
r
e
to
class
if
y
p
lan
ts
,
tr
ee
s
,
an
d
cr
o
p
s
,
as
well
as
to
id
e
n
tify
d
is
ea
s
es
[
1
5
]
.
Dee
p
l
ea
r
n
in
g
(
DL
)
-
b
ased
al
g
o
r
ith
m
s
ar
e
ef
f
ec
tiv
e
n
eu
r
al
n
etwo
r
k
s
f
o
r
h
an
d
lin
g
lar
g
e
d
atasets
,
h
elp
in
g
to
ac
h
iev
e
b
etter
o
u
tco
m
es.
Kav
ith
am
an
i
an
d
Um
aM
ah
es
war
i
[
1
6
]
em
p
l
o
y
ed
d
r
o
n
es
to
ca
p
tu
r
e
im
ag
es
o
f
d
am
a
g
ed
a
n
d
h
ea
lth
y
co
co
n
u
t
p
alm
s
.
T
h
e
im
a
g
es
wer
e
an
aly
ze
d
f
o
r
ab
n
o
r
m
al
r
eg
io
n
s
,
s
u
ch
as
s
y
m
p
to
m
s
o
f
illn
ess
o
r
wh
itef
ly
in
f
estatio
n
,
u
s
in
g
s
eg
m
en
tatio
n
tech
n
i
q
u
es.
T
h
e
p
r
ed
ictio
n
a
n
d
class
if
icatio
n
o
f
d
is
ea
s
es
f
o
cu
s
ed
o
n
tr
ain
i
n
g
a
d
ee
p
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
et
wo
r
k
(
DC
NN)
,
wh
ich
d
etec
ted
is
s
u
es
lik
e
r
o
o
t
b
leed
in
g
,
b
l
ad
e
p
o
llu
tio
n
,
an
d
in
s
ec
t
in
f
estatio
n
b
y
u
tili
zin
g
th
e
s
eg
m
en
ted
r
e
g
io
n
s
.
T
h
e
m
eth
o
d
q
u
ic
k
ly
lo
ca
ted
a
b
n
o
r
m
al
b
o
u
n
d
ar
ies
u
s
in
g
s
eg
m
en
tatio
n
tech
n
iq
u
es.
Ho
wev
er
,
th
e
m
eth
o
d
s
tr
u
g
g
le
d
to
d
if
f
er
en
tiate
b
etwe
en
th
e
d
if
f
er
en
t
d
is
ea
s
e
class
es
d
u
e
to
lim
ited
n
u
m
b
er
o
f
lay
er
s
in
n
etwo
r
k
.
Me
g
alin
g
am
et
a
l.
[
1
7
]
s
u
g
g
ested
th
e
m
o
d
if
ied
I
n
ce
p
tio
n
Net
-
b
ased
h
y
p
er
tu
n
in
g
s
u
p
p
o
r
t v
ec
to
r
m
ac
h
in
e
(
MI
N
-
SVM)
to
class
if
icatio
n
o
f
co
co
n
u
t tr
ee
,
wh
ich
r
elied
o
n
m
o
r
p
h
o
lo
g
ical
elem
en
ts
s
u
c
h
as
in
clin
atio
n
,
h
eig
h
t,
a
n
d
o
r
ien
tatio
n
.
Featu
r
es
wer
e
ex
tr
ac
ted
f
r
o
m
p
r
e
-
p
r
o
ce
s
s
ed
im
ag
es
b
y
f
o
u
r
d
if
f
er
en
t
co
n
v
o
l
u
tio
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN)
alg
o
r
ith
m
s
,
in
clu
d
ed
v
is
u
al
g
eo
m
etr
y
g
r
o
u
p
(
VGG)
,
R
esNet,
I
n
ce
p
tio
n
Net,
a
n
d
MI
N
-
SVM.
T
h
ese
ca
p
tu
r
ed
attr
ib
u
tes
wer
e
class
if
ied
b
y
SVM.
T
h
e
p
r
o
p
o
s
ed
m
eth
o
d
,
wh
ich
in
clu
d
ed
th
e
I
n
ce
p
tio
n
Net
f
ea
tu
r
e
ex
tr
ac
t
o
r
,
y
ield
ed
g
o
o
d
r
esu
lts
in
class
if
icatio
n
.
Ho
wev
er
,
th
e
s
u
g
g
ested
m
eth
o
d
d
o
es
n
o
t
en
h
an
ce
th
e
co
n
t
r
ast
o
f
th
e
im
a
g
e,
wh
ich
lim
its
th
e
m
o
d
el'
s
p
er
f
o
r
m
an
ce
.
T
h
ite
et
a
l.
[
1
8
]
i
n
tr
o
d
u
ce
d
th
e
Mo
b
ileNetV2
,
R
esNet5
0
,
an
d
VGG1
6
ar
ch
itectu
r
es,
t
h
at
wer
e
r
en
o
wn
ed
to
its
ab
ilit
ies
in
im
ag
e
r
ec
o
g
n
itio
n
task
s
.
I
n
th
e
p
r
o
p
o
s
ed
m
eth
o
d
,
th
e
d
ataset
was
p
r
e
-
p
r
o
ce
s
s
ed
,
an
d
d
ata
au
g
m
e
n
tatio
n
tech
n
i
q
u
es
wer
e
ap
p
lied
to
im
p
r
o
v
e
g
en
er
aliza
tio
n
an
d
f
i
n
e
-
tu
n
e
m
o
d
el
to
a
s
p
ec
if
ic
class
if
icatio
n
task
.
Me
th
o
d
ac
h
iev
ed
h
ig
h
class
if
icatio
n
p
er
f
o
r
m
an
ce
with
ex
ce
llen
t
ac
c
u
r
ac
y
.
Ho
wev
er
,
th
e
m
eth
o
d
f
aile
d
to
elim
in
ate
n
o
is
e
in
th
e
im
ag
es,
wh
ic
h
r
ed
u
ce
d
t
h
e
d
is
ea
s
e
class
if
ic
atio
n
p
er
f
o
r
m
an
c
e.
Ma
n
o
h
ar
an
et
a
l
.
[
1
9
]
d
ev
elo
p
ed
th
e
YOL
Ov
9
m
et
h
o
d
f
o
r
id
en
tify
in
g
m
ac
r
o
an
d
m
icr
o
n
u
tr
ien
t
d
ef
icien
cies
o
n
co
c
o
n
u
t
tr
ee
s
an
d
in
tr
o
d
u
c
ed
th
e
im
a
g
e
a
n
aly
s
is
-
b
ased
s
ev
er
ity
d
etec
tio
n
(
I
ASD)
f
o
r
a
s
s
es
s
in
g
s
ev
er
ity
o
f
th
ese
d
ef
icien
cies.
T
h
e
s
ev
er
ity
in
d
e
x
ca
lcu
latio
n
m
o
d
el
(
SI
C
M)
was
u
s
ed
to
ca
lcu
late
s
e
v
er
ity
in
d
ex
(
SI)
o
f
th
e
d
ef
icien
cies.
T
o
ev
er
y
i
d
en
tifie
d
d
ef
icien
c
y
,
r
elev
an
t
f
er
tili
ze
r
an
d
th
eir
ap
p
licatio
n
q
u
an
tity
wer
e
an
aly
ze
d
.
T
h
e
d
ev
elo
p
e
d
m
et
h
o
d
u
tili
ze
d
v
ar
i
o
u
s
au
g
m
e
n
tatio
n
an
d
p
r
e
-
p
r
o
ce
s
s
in
g
alg
o
r
ith
m
s
to
im
p
r
o
v
e
im
ag
e
q
u
ality
,
m
a
x
im
ize
co
u
n
t o
f
im
ag
es,
an
d
m
in
im
ize
o
v
e
r
f
itti
n
g
.
Ho
wev
er
,
th
e
m
eth
o
d
h
ad
lim
ited
f
ea
tu
r
e
r
ep
r
esen
tatio
n
,
wh
ic
h
h
in
d
er
e
d
m
o
d
el'
s
ab
ilit
y
to
d
if
f
er
e
n
tiate
am
o
n
g
t
h
e
v
ar
io
u
s
d
is
ea
s
e
class
es.
Div
y
an
th
et
a
l.
[
2
0
]
p
r
esen
t
ed
Fas
ter
r
eg
io
n
-
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
(
Fa
s
ter
R
-
C
NN)
to
d
etec
tin
g
co
co
n
u
t
clu
s
ter
s
,
b
o
th
n
o
n
-
o
cc
lu
d
e
d
an
d
th
o
s
e
with
leaf
o
cc
lu
s
io
n
.
T
o
en
h
an
ce
ac
cu
r
ac
y
,
th
e
atten
tio
n
m
ec
h
a
n
is
m
was
im
p
lem
en
ted
in
Fas
ter
R
-
C
NN
m
eth
o
d
.
Pre
s
en
ted
m
eth
o
d
p
r
o
v
id
ed
ess
en
tial
d
ata
f
o
r
h
ar
v
esti
n
g
co
c
o
n
u
ts
e
f
f
ec
tiv
ely
an
d
s
af
ely
.
Ho
wev
e
r
,
t
h
e
m
eth
o
d
h
ad
lo
wer
class
if
icatio
n
p
e
r
f
o
r
m
an
ce
d
u
e
to
lim
ited
f
ea
tu
r
e
r
ep
r
ese
n
tatio
n
.
Fro
m
th
e
a
b
o
v
e
a
n
aly
s
is
,
th
e
ex
is
tin
g
alg
o
r
ith
m
s
h
av
e
lim
itatio
n
s
,
s
u
ch
as
s
tr
u
g
g
lin
g
to
d
if
f
er
e
n
tiate
b
etwe
en
d
if
f
e
r
en
t
d
is
ea
s
e
cl
ass
es,
n
o
t
en
h
an
cin
g
th
e
co
n
tr
ast
o
f
th
e
im
ag
e
,
f
ailin
g
to
elim
in
ate
n
o
is
e,
an
d
h
av
in
g
lim
ited
f
ea
tu
r
e
r
ep
r
esen
tatio
n
an
d
class
if
icatio
n
p
er
f
o
r
m
an
ce
d
u
e
to
is
s
u
es lik
e
in
tr
a
-
class
v
ar
iab
ili
ty
an
d
in
ter
-
class
s
im
ilar
ity
.
T
o
m
itig
ate
th
ese
lim
itatio
n
s
,
th
is
ar
ticle
d
ev
elo
p
s
g
r
ad
ien
t
-
b
ased
s
to
ch
asti
c
d
ep
th
(
GSD)
with
C
NN
ap
p
r
o
ac
h
,
wh
ich
ef
f
ec
tiv
ely
class
if
ies
th
e
d
if
f
er
en
t
d
is
ea
s
e
class
e
s
.
I
n
co
r
p
o
r
atin
g
th
e
GSD
in
th
e
C
NN
m
eth
o
d
s
k
ip
s
th
e
less
co
n
tr
ib
u
tin
g
lay
er
s
,
m
i
n
im
izin
g
t
r
ain
in
g
tim
e
wh
ile
ac
h
iev
i
n
g
h
ig
h
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
I
n
th
e
p
r
e
-
p
r
o
ce
s
s
in
g
p
h
ase,
th
e
co
n
tr
ast
ad
ap
tiv
e
h
is
to
g
r
am
e
q
u
aliza
t
io
n
(
C
L
AHE
)
an
d
Gau
s
s
ian
b
lu
r
tech
n
iq
u
es a
r
e
u
s
ed
to
en
h
an
ce
im
ag
e
co
n
tr
ast an
d
elim
in
ate
n
o
is
e,
r
esp
ec
tiv
ely
.
T
h
r
o
u
g
h
th
ese
p
r
o
ce
s
s
es,
th
e
is
s
u
es
o
f
in
ter
-
class
s
im
ilar
ity
an
d
in
tr
a
-
class
v
ar
iab
ilit
y
ar
e
m
in
im
ize
d
,
an
d
th
e
r
ep
r
esen
tatio
n
o
f
f
ea
tu
r
es
f
o
r
d
is
ea
s
e
class
if
i
ca
tio
n
is
en
h
an
ce
d
.
T
h
e
co
n
tr
ib
u
tio
n
s
o
f
th
is
r
esear
ch
ar
e
o
u
tlin
ed
as
f
o
llo
ws:
C
L
AHE
an
d
Gau
s
s
ian
b
lu
r
tech
n
iq
u
es a
r
e
em
p
l
o
y
ed
in
t
h
e
p
r
e
-
p
r
o
ce
s
s
in
g
s
tag
e
to
en
h
an
ce
co
lo
r
co
n
tr
ast an
d
r
em
o
v
e
n
o
is
e
f
r
o
m
th
e
im
ag
es.
T
h
e
Gr
ab
C
u
t
s
eg
m
en
tatio
n
alg
o
r
ith
m
is
u
s
ed
to
s
eg
m
en
t
th
e
im
ag
es,
is
o
latin
g
th
e
leaf
p
o
r
tio
n
s
f
r
o
m
th
e
b
ac
k
g
r
o
u
n
d
a
n
d
im
p
r
o
v
in
g
th
e
v
is
ib
ilit
y
o
f
th
e
d
is
ea
s
e,
wh
ich
h
elp
s
d
if
f
er
en
tiate
b
etwe
en
th
e
v
ar
i
o
u
s
d
is
ea
s
e
class
es.
G
SD w
ith
C
N
N
ap
p
r
o
a
ch
is
d
ev
elo
p
e
d
to
class
if
y
v
ar
io
u
s
d
is
ea
s
e
class
e
s
with
h
ig
h
ac
cu
r
ac
y
a
n
d
r
e
d
u
ce
d
tr
ain
in
g
tim
e.
T
h
e
GSD
tech
n
iq
u
e
is
in
co
r
p
o
r
ated
in
t
o
ea
ch
lay
e
r
o
f
th
e
n
etwo
r
k
,
wh
ich
s
k
ip
s
less
e
f
f
icien
t
lay
er
s
b
y
ca
lcu
latin
g
p
r
o
b
ab
ilit
ies
u
s
in
g
g
r
ad
ien
t
m
ag
n
itu
d
es.
T
h
is
p
r
o
ce
s
s
m
in
im
izes tr
ain
in
g
ti
m
e
an
d
e
n
h
an
ce
s
class
if
icatio
n
p
er
f
o
r
m
an
ce
with
h
ig
h
e
r
ac
c
u
r
ac
y
.
T
h
is
m
an
u
s
cr
ip
t
is
o
r
g
an
ized
as
f
o
llo
ws:
s
ec
tio
n
2
s
h
o
ws
p
r
o
ce
s
s
o
f
p
r
o
p
o
s
ed
tech
n
iq
u
e.
Sectio
n
3
d
is
cu
s
s
es r
esu
lt
s
o
f
th
e
p
r
o
p
o
s
ed
tech
n
iq
u
e.
Sectio
n
4
co
n
clu
d
es a
m
an
u
s
cr
ip
t.
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
Gra
d
ien
t
-
b
a
s
ed
s
to
ch
a
s
tic
d
e
p
th
w
ith
co
n
vo
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k
fo
r
…
(
K
a
vith
a
Ma
g
a
d
i G
o
p
a
la
kris
h
n
a
)
1157
2.
M
E
T
H
O
D
An
ef
f
icien
t
DL
m
o
d
el
is
d
e
v
elo
p
ed
t
o
class
if
y
v
ar
io
u
s
d
is
ea
s
e
clas
s
es
wi
th
h
ig
h
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
Data
s
et
u
tili
ze
d
o
n
t
h
is
wo
r
k
is
th
e
c
o
c
o
n
u
t
tr
ee
d
is
ea
s
e
d
ataset
,
an
d
th
e
im
a
g
es
ar
e
p
r
e
-
p
r
o
ce
s
s
ed
u
s
in
g
th
e
C
L
A
HE
an
d
Gau
s
s
ian
b
lu
r
tech
n
i
q
u
es,
wh
ich
en
h
an
ce
im
a
g
e
c
o
n
tr
ast
an
d
r
em
o
v
e
n
o
is
e.
Pre
-
p
r
o
ce
s
s
ed
im
ag
es
ar
e
th
en
s
eg
m
en
ted
b
y
Gr
ab
C
u
t
s
eg
m
en
tatio
n
tech
n
iq
u
e,
wh
ich
s
eg
m
en
ts
th
e
im
ag
es
ef
f
ec
tiv
ely
th
r
o
u
g
h
a
n
iter
ativ
e
p
r
o
c
ess
.
Fin
ally
,
th
e
s
eg
m
en
ted
im
ag
es
ar
e
class
if
ied
u
s
in
g
th
e
d
ev
elo
p
e
d
GSD
-
C
NN
alg
o
r
ith
m
,
ac
h
ie
v
in
g
h
ig
h
class
if
icat
io
n
p
er
f
o
r
m
a
n
ce
.
Fig
u
r
e
1
illu
s
tr
ates
th
e
p
r
o
ce
s
s
o
f
co
co
n
u
t tr
ee
d
is
ea
s
e
class
if
i
ca
tio
n
.
Fig
u
r
e
1
.
Pro
ce
s
s
o
f
co
co
n
u
t tr
ee
d
is
ea
s
e
class
if
ica
tio
n
2
.
1
.
Da
t
a
c
o
llect
io
n
Data
p
r
esen
t
in
co
co
n
u
t
tr
ee
d
is
ea
s
e
d
ataset
i
s
g
ath
er
ed
f
r
o
m
f
ar
m
s
in
th
e
d
is
tr
icts
o
f
Pu
n
e,
I
n
d
ia,
T
alu
k
a
-
Sh
ir
u
r
,
an
d
Ken
d
u
r
.
T
h
e
im
ag
es
ar
e
ca
p
tu
r
e
d
o
n
d
if
f
er
en
t
s
ce
n
a
r
io
s
,
in
clu
d
ed
leav
es
in
its
n
atu
r
al
en
v
ir
o
n
m
en
t
an
d
d
etac
h
ed
f
r
o
m
th
e
p
lan
t.
T
h
e
d
ataset
in
clu
d
es
5
7
9
8
im
a
g
es,
with
ev
er
y
class
co
n
tain
in
g
v
ar
y
in
g
n
u
m
b
er
o
f
im
ag
es
[
2
1
]
.
T
ab
le
1
p
r
esen
ts
th
e
d
ataset
d
escr
ip
tio
n
.
T
ab
le
1
.
Data
s
et
d
escr
ip
tio
n
N
a
me
o
f
d
i
s
e
a
se
N
u
mb
e
r
o
f
sam
p
l
e
s
B
u
d
r
o
t
4
7
0
B
u
d
r
o
o
t
d
r
o
p
p
i
n
g
5
1
4
S
t
e
m
b
l
e
e
d
i
n
g
1
0
0
6
Le
a
f
r
o
t
1
6
7
3
G
r
a
y
l
e
a
f
sp
o
t
2
1
3
5
To
t
a
l
5
7
9
8
2
.
2
.
P
re
-
pro
ce
s
s
ing
2
.
2
.
1
.
I
m
a
g
e
deno
is
ing
T
h
is
is
a
s
ig
n
if
ican
t
p
r
e
-
p
r
o
ce
s
s
in
g
p
h
ase
f
o
r
r
e
d
u
cin
g
d
is
to
r
tio
n
s
o
r
n
o
is
e
in
th
e
i
m
ag
e
[
2
2
]
.
C
o
n
v
en
tio
n
al
d
e
n
o
is
in
g
tech
n
iq
u
es
f
o
cu
s
o
n
s
m
o
o
th
i
n
g
im
ag
es
th
r
o
u
g
h
ap
p
ly
in
g
u
n
e
q
u
a
l
weig
h
ts
to
p
ix
els,
wh
ich
ar
e
in
v
er
s
ely
p
r
o
p
o
r
tio
n
al
to
t
h
eir
d
is
tan
ce
f
r
o
m
ce
n
tr
al
p
ix
el
o
n
im
ag
e.
Par
ticu
l
ar
ly
,
th
e
Gau
s
s
ian
f
ilter
is
lin
ea
r
s
m
o
o
th
i
n
g
f
ilter
wh
ich
m
in
im
izes
th
e
weig
h
ts
ap
p
lied
t
o
p
ix
els
as
th
e
d
is
tan
ce
f
r
o
m
th
e
ce
n
tr
al
p
ix
el
in
cr
ea
s
es,
d
ep
en
d
ed
o
n
Gau
s
s
ian
f
u
n
ctio
n
.
Ma
th
em
a
tical
f
o
r
m
u
la
f
o
r
a
n
in
p
u
t
p
ix
el
in
th
e
Gau
s
s
ian
f
ilter
is
g
iv
en
in
(
1
)
.
I
n
th
e
(
1
)
,
th
e
=
√
(
−
)
2
+
(
−
)
2
r
esp
ec
tiv
e
p
ix
el
d
is
tan
ce
f
r
o
m
ce
n
tr
al
p
ix
el.
(
,
)
=
1
√
2
−
2
2
2
(
1
)
2
.
2
.
2
.
Co
ntr
a
s
t
a
da
ptiv
e
his
t
o
g
ra
m
equa
liza
t
io
n
C
L
AHE
is
in
tr
o
d
u
ce
d
to
p
r
o
v
i
d
e
a
m
o
r
e
n
atu
r
al
ap
p
ea
r
an
ce
f
o
r
en
h
an
cin
g
im
ag
es.
Ad
d
itio
n
ally
,
th
e
th
r
esh
o
ld
elem
en
t
is
u
tili
ze
d
to
lim
it
co
n
tr
ast
en
h
an
ce
m
en
t
to
ch
o
s
en
ar
ea
s
.
T
o
o
b
tain
th
is
,
th
e
R
G
B
co
lo
r
s
p
ac
e
is
co
n
v
er
ted
to
h
u
e,
s
at
u
r
atio
n
,
a
n
d
v
a
lu
e
(
HSV)
.
T
h
e
v
alu
e
co
m
p
o
n
en
t
is
p
r
o
ce
s
s
ed
th
r
o
u
g
h
C
L
AHE
,
wh
ile
th
e
h
u
e
an
d
s
atu
r
atio
n
co
m
p
o
n
en
ts
r
em
ain
u
n
ch
an
g
ed
.
Fin
ally
,
C
L
AHE
is
em
p
lo
y
ed
t
h
r
o
u
g
h
r
ed
is
tr
ib
u
tin
g
th
e
g
r
ey
lev
els b
ac
k
to
th
e
R
GB
co
lo
r
s
p
ac
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.
15
,
No
.
2
,
Ap
r
il
20
26
:
115
5
-
1
1
6
5
1158
2
.
3
.
Seg
m
ent
a
t
i
o
n
Gr
ab
C
u
t
s
eg
m
en
tatio
n
ap
p
r
o
ac
h
is
en
h
an
ce
m
en
t
o
f
Gr
ap
h
C
u
t
tech
n
iq
u
e.
Gr
ab
C
u
t
tech
n
iq
u
e
u
s
es
tex
tu
r
e
(
co
l
o
r
)
an
d
b
o
u
n
d
ar
y
(
co
n
tr
ast)
d
ata
in
im
a
g
es,
allo
win
g
in
ter
ac
tio
n
with
u
s
er
s
b
y
r
ec
tan
g
u
lar
b
o
x
es.
I
t
ev
alu
ates
p
ix
els
with
in
b
o
x
an
d
class
if
ies
p
ix
els
o
u
ts
i
d
e
b
o
x
as
b
ac
k
g
r
o
u
n
d
.
Af
te
r
s
ev
er
al
iter
atio
n
s
,
alg
o
r
ith
m
ac
q
u
ir
es
th
e
d
esire
d
s
eg
m
en
tatio
n
[
2
3
]
.
T
h
e
co
n
v
en
tio
n
al
Gr
ab
C
u
t
tech
n
iq
u
e
p
er
f
o
r
m
s
iter
ativ
e
o
p
tim
izatio
n
in
s
eg
m
en
tatio
n
,
s
im
p
lify
in
g
th
e
u
s
er
in
te
r
ac
tio
n
r
eq
u
ir
ed
to
s
eg
m
en
t
o
b
jects.
I
t
u
tili
ze
s
r
ec
tan
g
u
lar
b
o
u
n
d
in
g
b
o
x
to
ev
alu
ate
p
ix
els
in
s
id
e
b
o
x
an
d
b
ac
k
g
r
o
u
n
d
p
ix
els
o
u
ts
id
e.
E
n
er
g
y
f
u
n
ctio
n
o
f
Gr
ab
C
u
t
is
g
iv
en
in
(
2
)
.
T
h
e
f
u
n
ctio
n
d
ef
in
es
r
e
g
io
n
al
in
f
o
r
m
atio
n
te
r
m
o
f
en
er
g
y
f
u
n
ctio
n
,
d
en
o
tes
h
is
to
g
r
am
m
o
d
el,
is
th
e
tr
an
s
p
ar
en
cy
c
o
ef
f
icien
t,
a
n
d
d
en
o
tes
a
s
in
g
le
p
ix
el.
T
h
e
m
ath
em
atica
l
f
o
r
m
u
la
f
o
r
is
g
iv
en
in
(
3
)
.
T
h
e
m
ath
e
m
atica
l f
o
r
m
u
la
f
o
r
s
m
o
o
th
in
g
item
(
,
)
is
g
iv
en
as (
4
)
.
(
,
,
)
=
(
,
,
)
+
(
,
)
(
2
)
(
,
,
)
=
∑
−
ℎ
(
;
)
(
3
)
(
,
)
=
∑
(
,
)
−
1
(
,
)
∈
[
≠
]
−
(
−
)
2
(
4)
I
n
(
4
)
,
r
ep
r
esen
ts
a
g
r
o
u
p
o
f
ad
jace
n
t
p
ix
el
p
air
s
,
th
e
(
)
r
ep
r
esen
ts
E
u
clid
ea
n
d
is
tan
ce
a
m
o
n
g
ad
jace
n
t
p
ix
els
an
d
c
h
o
o
s
in
g
en
s
u
r
e
d
co
r
r
ec
t
s
witch
in
g
am
o
n
g
h
ig
h
an
d
lo
w
c
o
n
tr
as
t
o
n
e
x
p
o
n
en
tial
d
ir
ec
tio
n
th
at
p
r
o
m
o
ted
th
e
c
o
n
s
is
ten
cy
o
n
t
h
e
s
am
e
g
r
ay
s
ca
le
ar
ea
s
.
W
h
ile
co
n
s
tan
t
=
0
,
s
m
o
o
th
n
ess
ter
m
en
s
u
r
es
s
m
o
o
th
n
ess
ev
er
y
wh
e
r
e
d
ef
in
ed
th
r
o
u
g
h
co
n
s
tan
t.
T
h
is
en
er
g
y
en
s
u
r
es
co
h
er
en
c
e
in
ar
ea
s
with
th
e
s
am
e
g
r
ey
-
lev
el.
Pra
ctica
lly
,
b
etter
o
u
tco
m
es
ar
e
attain
ed
th
r
o
u
g
h
d
eter
m
in
in
g
th
e
p
i
x
e
ls
o
f
n
eig
h
b
o
r
in
g
,
wh
eth
er
ad
jace
n
t
h
o
r
izo
n
tally
o
r
v
er
tically
.
T
h
e
s
eg
m
en
tatio
n
p
r
o
ce
s
s
o
f
G
r
ab
C
u
t im
ag
e
is
d
escr
ib
ed
as:
i)
I
n
itializatio
n
–
u
s
er
attain
s
th
e
in
itial
tr
im
ap
im
ag
e
th
r
o
u
g
h
d
ir
ec
tly
ch
o
o
s
in
g
tar
g
et,
wh
e
r
e
en
tire
p
ix
els
o
u
ts
id
e
b
o
x
ar
e
co
n
s
id
er
ed
b
a
ck
g
r
o
u
n
d
p
ix
els,
a
n
d
e
n
tire
p
ix
els
in
s
id
e
b
o
x
ar
e
tak
en
as
ta
r
g
et
p
ix
els
.
W
h
en
∈
,
p
ix
el
lab
el
=
0
,
an
d
wh
e
n
∈
,
th
e
p
ix
el
lab
el
=
1
.
Gau
s
s
ian
m
ix
tu
r
e
m
o
d
el
(
GM
M)
to
f
o
r
eg
r
o
u
n
d
a
n
d
b
a
ck
g
r
o
u
n
d
is
in
itiated
with
=
0
an
d
=
1
.
Pix
els
ar
e
n
ex
t
clu
s
ter
ed
in
to
class
es,
co
r
r
esp
o
n
d
in
g
t
o
Gau
s
s
ian
m
o
d
els
in
GM
M,
u
tili
zin
g
K
-
m
ea
n
s
tech
n
iq
u
e.
Pro
ce
s
s
o
f
iter
ativ
e
r
ed
u
ctio
n
is
d
escr
ib
e
d
.
ii)
E
m
p
lo
y
c
o
m
p
o
n
en
ts
o
f
Ga
u
s
s
ian
in
GM
M
f
o
r
e
v
er
y
p
ix
el.
F
o
r
ev
er
y
in
,
=
(
,
,
,
)
.
iii)
Fo
r
g
iv
en
im
a
g
e
d
ata
,
iter
ativ
ely
o
p
tim
ize
GM
M
p
ar
am
ete
r
s
=
(
,
,
,
)
.
iv
)
E
v
alu
ate
s
eg
m
en
tatio
n
-
r
elied
o
n
Gib
b
s
en
er
g
y
ter
m
b
y
ev
alu
atin
g
th
e
g
r
a
p
h
an
d
m
ea
s
u
r
in
g
weig
h
ts
t
-
lin
k
s
an
d
n
-
lin
k
s
.
Nex
t,
m
a
x
-
f
lo
w/m
in
-
cu
t te
ch
n
iq
u
e
is
u
tili
ze
d
to
s
eg
m
en
tatio
n
.
v)
R
ep
ea
t p
r
o
ce
s
s
u
n
til co
n
v
e
r
g
e
n
ce
o
cc
u
r
s
.
vi
)
Utilize
b
o
r
d
er
cu
t
o
u
t a
n
d
m
atc
h
in
g
f
o
r
s
m
o
o
t
h
in
g
a
n
d
p
o
s
t
-
p
r
o
ce
s
s
in
g
in
s
eg
m
en
ted
b
o
u
n
d
ar
ies.
2
.
4
.
Cla
s
s
if
ica
t
io
n
Seg
m
en
ted
im
ag
es
a
r
e
f
e
d
to
th
e
class
if
icatio
n
s
tag
e
to
clas
s
if
y
th
e
v
a
r
io
u
s
class
es
o
f
co
co
n
u
t
leaf
d
is
ea
s
es.
I
n
th
is
p
h
ase,
th
e
G
SD
tech
n
iq
u
e
is
in
clu
d
ed
in
C
NN,
s
k
ip
p
in
g
th
e
lay
er
th
at
less
co
n
tr
ib
u
tes
to
co
co
n
u
t
leaf
d
is
ea
s
e
class
if
ica
tio
n
an
d
im
p
r
o
v
es
class
if
icati
o
n
p
er
f
o
r
m
an
ce
.
T
h
is
s
ec
tio
n
ex
p
lain
s
th
e
p
r
o
ce
s
s
o
f
C
NN
an
d
GSD
tech
n
iq
u
es.
2
.
4
.
1
.
Co
nv
o
lutio
na
l la
y
er
s
C
o
n
v
o
lu
tio
n
al
lay
er
is
a
p
r
im
ar
y
c
o
m
p
o
n
en
t
o
f
C
NN
wh
ich
in
clu
d
es
a
g
r
o
u
p
o
f
in
d
i
v
id
u
al
f
ilter
s
.
E
v
er
y
f
ilter
is
in
d
iv
id
u
ally
co
n
v
o
lv
e
d
to
im
ag
e
an
d
f
ea
tu
r
e
m
ap
ac
q
u
ir
e
d
[
2
4
]
,
[
2
5
]
.
Ge
n
e
r
ally
,
th
e
s
ize
o
f
th
e
im
ag
e
×
is
co
n
v
o
lv
ed
with
a
f
ilter
s
ize
o
f
×
ℎ
,
wh
er
e
th
e
m
ath
e
m
atica
l
f
o
r
m
u
lae
f
o
r
o
u
tp
u
t
f
ea
tu
r
e
m
ap
s
ize
is
g
iv
en
in
(
5
)
a
n
d
(
6
)
.
=
−
+
2
+
1
(
5
)
ℎ
=
−
ℎ
+
2
ℎ
ℎ
+
1
(
6
)
W
h
er
e
an
d
ℎ
d
ef
in
es
p
a
d
d
in
g
o
f
ze
r
o
s
o
n
wid
th
an
d
h
eig
h
t,
r
esp
ec
tiv
ely
.
T
h
e
an
d
ℎ
ar
e
s
tr
id
es
i
n
h
o
r
izo
n
tal
a
n
d
v
er
tical
d
ir
ec
ti
o
n
s
.
E
ac
h
r
esu
ltin
g
f
ea
t
u
r
e
m
ap
is
ac
h
iev
ed
th
r
o
u
g
h
co
n
v
o
l
u
tio
n
o
f
in
p
u
t
m
ap
s
to
lin
ea
r
f
ilter
,
in
clu
d
in
g
b
ias
ter
m
,
an
d
n
ex
t
n
o
n
-
lin
ea
r
f
u
n
ctio
n
is
ap
p
lied
.
Ma
th
em
atica
l
f
o
r
m
u
la
f
o
r
o
u
tp
u
t
is
g
en
er
ally
g
iv
e
n
in
(
7
)
.
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
Gra
d
ien
t
-
b
a
s
ed
s
to
ch
a
s
tic
d
e
p
th
w
ith
co
n
vo
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k
fo
r
…
(
K
a
vith
a
Ma
g
a
d
i G
o
p
a
la
kris
h
n
a
)
1159
=
(
∑
−
1
×
+
∈
)
(
7
)
W
h
er
e
is
th
e
n
u
m
b
er
o
f
lay
er
s
,
d
en
o
tes
a
co
n
v
o
lu
tio
n
al
k
er
n
el,
d
en
o
tes
b
ias,
d
en
o
t
es
g
r
o
u
p
o
f
in
p
u
t m
ap
s
,
a
n
d
(
.
)
is
ac
tiv
atio
n
f
u
n
ctio
n
.
Activ
atio
n
f
u
n
ctio
n
is
cr
u
cial
o
n
C
NNs,
en
ab
lin
g
th
em
to
lear
n
an
d
p
r
o
ce
s
s
co
m
p
lex
task
s
.
T
h
e
ac
tiv
atio
n
f
u
n
ctio
n
is
a
n
o
n
-
l
in
ea
r
tr
an
s
f
o
r
m
atio
n
em
p
lo
y
e
d
to
i
n
p
u
t.
I
t
g
e
n
er
ally
d
eter
m
in
es
wh
eth
er
d
ata
r
ec
eiv
ed
b
y
n
e
u
r
o
n
is
r
elev
an
t
en
o
u
g
h
to
p
r
o
v
i
d
e
o
u
tp
u
t
o
r
if
it
s
h
o
u
ld
b
e
ig
n
o
r
e
d
.
I
n
t
h
is
wo
r
k
,
r
ec
tifie
d
lin
ea
r
u
n
it
(
R
eL
U)
is
u
s
ed
,
as
it
allo
w
s
f
o
r
ea
s
y
b
ac
k
p
r
o
p
ag
atio
n
o
f
er
r
o
r
s
an
d
en
ab
les
m
an
y
lay
er
s
o
f
n
eu
r
o
n
s
to
b
e
ac
tiv
ated
t
h
r
o
u
g
h
R
eL
U
f
u
n
ctio
n
.
Prim
ar
y
b
en
ef
its
o
f
R
eL
U
p
r
o
ce
s
s
ac
r
o
s
s
o
th
er
ac
tiv
atio
n
f
u
n
ctio
n
s
is
th
at
it
d
o
esn
’
t
ac
t
iv
ate
en
tire
n
eu
r
o
n
s
s
im
u
ltan
e
o
u
s
ly
.
T
h
is
m
ea
n
s
th
at,
in
an
y
g
iv
en
tim
e,
o
n
ly
s
u
b
s
et
o
f
n
eu
r
o
n
s
is
ac
tiv
ated
,
m
ak
es
n
etwo
r
k
s
p
ar
s
e,
ef
f
ic
ien
t,
an
d
ea
s
ier
to
co
m
p
u
te.
F
o
r
n
eg
ativ
e
i
n
p
u
ts
,
g
r
ad
ien
ts
ar
e
ze
r
o
,
a
n
d
weig
h
t
s
ar
e
n
o
t
u
p
d
ated
i
n
b
ac
k
p
r
o
p
ag
atio
n
.
T
h
at
lead
to
d
ea
d
n
e
u
r
o
n
s
th
at
n
ev
er
g
et
ac
tiv
ated
,
an
d
its
m
ath
em
atica
l f
o
r
m
u
la
is
g
iv
e
n
in
(
8
)
a
n
d
(
9
)
.
(
)
=
(
0
,
)
(
8
)
(
)
=
{
,
≥
0
0
,
<
0
(
9
)
I
n
itial
co
n
v
o
lu
tio
n
al
lay
e
r
ca
p
tu
r
es
v
ar
io
u
s
lo
w
-
le
v
el
f
ea
tu
r
es,
lik
e
ed
g
es
an
d
tex
tu
r
es.
B
y
s
tack
in
g
m
u
ltip
le
co
n
v
o
lu
tio
n
al
lay
e
r
s
,
n
etwo
r
k
lear
n
s
m
o
r
e
g
l
o
b
al
f
ea
tu
r
es.
Her
e,
th
e
GSD
lay
er
is
ap
p
lied
to
ea
ch
co
n
v
o
l
u
tio
n
al
lay
er
t
o
im
p
r
o
v
e
f
ea
tu
r
e
r
e
p
r
esen
tatio
n
.
2
.
4
.
2
.
P
o
o
lin
g
la
y
er
s
T
h
e
p
o
o
lin
g
lay
er
is
in
clu
d
ed
am
o
n
g
co
n
s
ec
u
tiv
e
co
n
v
o
lu
tio
n
al
lay
er
s
i
n
a
C
NN.
Pu
r
p
o
s
e
o
f
p
o
o
lin
g
lay
er
is
to
m
in
im
ize
s
p
atial
s
ize
o
f
r
ep
r
esen
tatio
n
.
T
h
is
p
r
o
ce
s
s
m
in
im
izes
n
u
m
b
e
r
o
f
p
a
r
am
eter
s
an
d
co
m
p
u
tatio
n
al
co
s
t
o
n
m
o
d
el,
wh
ich
h
elp
s
i
n
h
an
d
lin
g
o
v
er
f
itti
n
g
.
Po
o
lin
g
la
y
er
p
r
o
ce
s
s
o
n
ea
c
h
i
n
p
u
t
lay
e
r
in
d
iv
id
u
ally
,
r
esizin
g
it
s
p
atially
th
r
o
u
g
h
p
o
o
lin
g
p
r
o
ce
s
s
.
T
h
e
m
o
s
t
co
m
m
o
n
f
o
r
m
o
f
th
e
p
o
o
lin
g
lay
er
,
to
f
ilter
s
ize
o
f
2
×2
an
d
s
tr
id
e
o
f
2
,
d
o
wn
-
s
am
p
les
ea
ch
d
e
p
th
s
lice
o
f
in
p
u
t
ac
r
o
s
s
th
e
wid
th
an
d
h
eig
h
t
o
f
ac
tiv
atio
n
s
.
I
n
th
is
wo
r
k
,
th
e
m
ax
p
o
o
lin
g
lay
er
is
u
s
ed
,
as
i
t
p
r
o
v
id
es
g
o
o
d
r
esu
lts
an
d
r
e
d
u
ce
s
co
m
p
u
tatio
n
a
l
co
s
t
f
o
r
th
e
u
p
p
er
la
y
er
s
th
r
o
u
g
h
r
em
o
v
in
g
n
o
n
-
m
ax
im
al
v
alu
es,
o
f
f
er
in
g
f
o
r
m
o
f
tr
a
n
s
latio
n
in
v
ar
ian
ce
.
Ad
d
itio
n
ally
,
it
o
f
f
er
s
ex
tr
a
r
o
b
u
s
tn
ess
f
o
r
p
o
s
itio
n
al
v
ar
iati
o
n
s
an
d
e
f
f
ec
tiv
ely
r
ed
u
ce
s
th
e
d
im
en
s
io
n
ality
o
f
in
ter
m
ed
iate
r
ep
r
esen
tatio
n
s
.
2
.
4
.
3
.
F
ull
y
c
o
nn
ec
t
ed
la
y
er
s
T
h
is
m
ea
n
s
th
at
n
eu
r
o
n
in
p
r
io
r
lay
er
is
in
teg
r
ated
to
ea
c
h
n
e
u
r
o
n
o
n
p
r
esen
t la
y
er
.
I
ts
ac
tiv
atio
n
s
ar
e
ca
lcu
lated
th
r
o
u
g
h
m
at
r
ix
m
u
l
tip
licatio
n
,
f
o
llo
wed
th
r
o
u
g
h
b
ias
o
f
f
s
et.
Neu
r
o
n
s
in
f
in
al
f
u
lly
co
n
n
ec
ted
lay
e
r
co
r
r
esp
o
n
d
to
am
o
u
n
t
o
f
class
es
f
o
r
p
r
ed
ictio
n
.
Ma
n
y
o
f
attr
ib
u
tes
f
r
o
m
co
n
v
o
l
u
tio
n
a
l
an
d
s
u
b
s
am
p
lin
g
lay
er
s
ar
e
u
s
ef
u
l
to
class
if
icat
io
n
,
an
d
in
teg
r
atio
n
o
f
th
ese
f
ea
tu
r
es
p
lay
s
a
k
ey
r
o
le.
Fu
l
ly
co
n
n
ec
ted
lay
e
r
in
teg
r
ates
en
tire
attr
ib
u
tes
ex
tr
ac
ted
f
r
o
m
p
r
io
r
co
n
v
o
lu
tio
n
al
an
d
s
u
b
s
am
p
lin
g
lay
er
s
.
L
ast
f
u
lly
co
n
n
ec
ted
lay
er
u
tili
ze
s
So
f
tMa
x
ac
tiv
atio
n
f
u
n
ctio
n
,
w
h
ich
is
p
r
im
ar
il
y
u
s
ed
f
o
r
m
u
lti
-
class
class
if
ic
atio
n
.
2
.
4
.
4
.
Sto
cha
s
t
ic
depth
t
ec
hn
iqu
e
Sh
allo
w
m
o
d
els
en
ab
le
q
u
ick
f
o
r
war
d
an
d
b
ac
k
war
d
d
ata
p
r
o
p
ag
atio
n
,
allo
win
g
th
em
to
b
e
tr
ain
ed
in
s
h
o
r
ter
tim
e.
T
h
o
u
g
h
,
b
ec
au
s
e
o
f
co
m
p
le
x
s
p
atial
d
is
tr
ib
u
tio
n
s
o
f
p
ix
els
an
d
m
u
lti
-
s
ca
le
s
em
an
tic
d
ata,
s
h
allo
w
n
etwo
r
k
s
ar
e
in
s
u
f
f
icien
t
to
lear
n
in
g
an
d
ca
p
tu
r
in
g
p
r
im
ar
y
f
ea
tu
r
es.
Du
r
in
g
th
e
tr
ain
in
g
p
r
o
ce
s
s
,
is
s
u
es
s
u
ch
as lo
n
g
tr
ain
in
g
ti
m
es,
d
ata
atten
u
atio
n
d
u
r
in
g
f
o
r
war
d
p
r
o
p
a
g
atio
n
,
an
d
v
an
is
h
in
g
g
r
a
d
ien
ts
ar
is
e.
T
o
o
v
e
r
co
m
e
th
ese
ch
allen
g
es,
a
s
to
ch
asti
c
d
ep
th
la
y
er
is
in
co
r
p
o
r
ate
d
in
to
th
e
C
N
N.
Usi
n
g
a
s
h
allo
w
n
etwo
r
k
d
u
r
in
g
tr
ain
in
g
im
p
r
o
v
es
d
ata
tr
an
s
f
er
with
in
th
e
n
etwo
r
k
,
r
esu
ltin
g
in
a
d
e
ep
er
n
etwo
r
k
with
en
h
an
ce
d
g
en
er
aliza
tio
n
ca
p
ab
ilit
ies.
T
h
e
n
etwo
r
k
with
r
an
d
o
m
d
e
p
th
is
ac
h
iev
ed
th
r
o
u
g
h
in
cl
u
d
in
g
s
to
ch
asti
c
d
ep
th
d
r
o
p
b
l
o
ck
,
th
at
is
s
am
e
to
d
r
o
p
o
u
t
p
r
o
c
ess
.
Ma
th
em
atica
l
f
o
r
m
u
la
f
o
r
in
v
er
ted
r
esid
u
al
ar
ch
itectu
r
e
is
g
iv
en
in
(
1
0
)
,
wh
er
e
a
r
an
d
o
m
v
a
r
iab
le
f
o
llo
ws
B
er
n
o
u
lli
d
is
tr
ib
u
tio
n
f
o
r
in
v
er
ted
r
esid
u
al
p
r
o
ce
s
s
.
I
n
th
e
(
1
0
)
,
=
1
,
wh
e
r
e
t
h
e
f
u
lly
in
v
er
ted
r
esid
u
al
ar
ch
itectu
r
e
is
r
etain
ed
,
wh
ile
=
0
is
wh
en
th
e
r
esid
u
al
ar
c
h
itectu
r
e
is
n
o
t
ac
tiv
ated
,
wh
ile
th
e
wh
o
le
ar
ch
itectu
r
e
is
d
eg
en
e
r
ated
f
o
r
id
en
tify
i
n
g
th
e
p
r
o
ce
s
s
,
an
d
its
m
ath
em
atica
l
f
o
r
m
u
la
is
g
iv
en
as
(
1
1
)
.
W
h
er
e
=
0
with
p
r
o
b
ab
ilit
y
o
f
1
−
an
d
=
1
with
p
r
o
b
ab
ilit
y
o
f
,
wh
ich
r
e
p
r
esen
ts
s
u
r
v
iv
in
g
p
r
o
b
a
b
ilit
y
o
f
a
r
esid
u
al
m
o
d
u
le.
Ma
th
e
m
atica
l
f
o
r
m
u
la
f
o
r
th
e
lin
ea
r
d
ec
ay
f
u
n
ctio
n
,
as g
iv
en
in
(
1
2
)
.
+
1
=
(
,
)
+
(
)
(
1
0
)
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.
15
,
No
.
2
,
Ap
r
il
20
26
:
115
5
-
1
1
6
5
1160
+
1
=
(
)
(
1
1
)
=
1
−
(
1
−
)
(
1
2
)
T
h
is
s
ettin
g
allo
ws
s
h
allo
w
l
ay
er
s
to
c
ap
tu
r
e
lo
w
-
lev
el
f
e
atu
r
es
to
u
tili
ze
in
later
lay
er
s
,
with
th
e
s
h
allo
w
lay
er
s
n
o
t
b
ein
g
d
is
ca
r
d
ed
f
r
e
q
u
en
tly
.
B
ec
au
s
e
o
f
r
an
d
o
m
d
r
o
p
p
i
n
g
s
ch
em
e,
ce
r
tain
lay
er
s
o
n
n
etwo
r
k
ar
e
n
o
t
ac
tiv
ated
in
ev
er
y
tr
ai
n
in
g
iter
atio
n
,
wh
ich
ef
f
icien
tly
d
e
v
elo
p
s
en
s
e
m
b
le
o
f
in
e
x
p
licit
ap
p
r
o
ac
h
es.
T
r
ain
in
g
p
r
o
ce
s
s
to
r
a
n
d
o
m
d
ep
t
h
s
co
m
b
i
n
es
r
esid
u
al
m
eth
o
d
s
at
v
a
r
io
u
s
d
ep
th
s
,
r
an
d
o
m
ly
ex
tr
ac
tin
g
d
ee
p
r
esid
u
al
f
ea
tu
r
es
f
r
o
m
s
h
allo
w
r
esid
u
al
f
ea
tu
r
es.
T
h
is
en
s
u
r
es
f
ea
tu
r
e
d
iv
er
s
en
ess
,
r
ed
u
ce
s
o
v
er
f
itti
n
g
,
a
n
d
e
n
h
an
ce
s
g
e
n
er
aliza
tio
n
ca
p
ab
ilit
y
o
f
m
e
th
o
d
.
Netwo
r
k
’
s
tr
ain
in
g
tim
e,
wh
en
em
p
l
o
y
in
g
s
to
ch
asti
c
d
ep
th
,
n
o
lo
n
g
e
r
m
ax
im
izes
to
n
etwo
r
k
d
ep
th
b
u
t
is
b
ased
o
n
ex
p
ec
ted
n
etwo
r
k
d
ep
th
.
I
n
tr
ain
in
g
,
ℎ
in
v
er
ted
r
esid
u
al
b
lo
c
k
h
a
v
e
p
r
o
b
a
b
ilit
y
o
f
in
ac
tiv
e,
m
a
k
in
g
ef
f
icien
t
i
n
v
er
ted
r
esid
u
a
l
b
lo
ck
r
a
n
d
o
m
v
ar
iab
le.
Ma
th
em
atica
l
f
o
r
m
u
la
f
o
r
th
is
is
g
iv
en
in
(
1
3
)
.
I
n
li
n
ea
r
d
ec
ay
r
u
le,
=
0
.
5
an
d
f
o
r
s
u
f
f
i
cien
t
lar
g
e
,
co
u
n
t
o
f
ac
tiv
ated
r
esid
u
al
b
l
o
ck
s
in
tr
ain
in
g
in
cr
ea
s
es.
(
)
=
∑
=
1
(
1
3
)
2
.
4
.
5
.
G
ra
dient
-
ba
s
ed
s
t
o
cha
s
t
ic
depth
t
ec
hn
i
qu
e
T
h
e
GSD
s
k
ip
s
lay
er
s
in
th
e
tr
ain
in
g
p
r
o
ce
s
s
b
ased
o
n
t
h
eir
g
r
ad
ie
n
t
m
ag
n
itu
d
e,
m
ak
i
n
g
a
m
o
r
e
in
f
o
r
m
e
d
an
d
ad
a
p
tiv
e
tech
n
i
q
u
e
wh
en
co
m
p
ar
e
d
to
th
e
tr
a
d
itio
n
al
s
to
ch
asti
c
d
ep
th
lay
er
.
T
h
e
d
etailed
s
tep
s
o
f
th
is
p
r
o
ce
s
s
ar
e
s
u
m
m
ar
ize
d
in
Alg
o
r
ith
m
1
.
E
ac
h
c
o
n
v
o
lu
tio
n
al
lay
er
is
ev
alu
ated
b
ased
o
n
its
g
r
ad
ien
t
m
ag
n
itu
d
e,
wh
ich
ca
lc
u
lates
its
co
n
tr
ib
u
tio
n
to
t
h
e
lo
s
s
f
u
n
ctio
n
.
Du
r
i
n
g
th
e
b
ac
k
war
d
p
ass
in
g
,
=
‖
∇
‖
2
,
wh
er
e
∇
r
ep
r
esen
ts
th
e
lo
s
s
g
r
ad
ien
ts
co
r
r
esp
o
n
d
in
g
to
th
e
l
ay
er
p
ar
am
eter
s
.
T
h
e
d
ec
is
io
n
o
f
s
k
ip
p
in
g
is
p
er
f
o
r
m
ed
d
y
n
am
ically
;
if
is
less
th
an
th
e
th
r
esh
o
ld
,
t
h
e
la
y
er
is
s
k
ip
p
ed
with
a
p
r
o
b
ab
il
ity
(
)
,
wh
ich
d
ec
r
ea
s
es a
s
in
cr
ea
s
es.
T
h
e
C
NN
s
u
f
f
er
s
f
r
o
m
v
a
n
is
h
in
g
g
r
ad
ien
ts
an
d
lo
n
g
tr
ain
in
g
tim
es.
T
o
a
d
d
r
ess
th
ese
p
r
o
b
lem
s
,
s
to
ch
asti
c
d
ep
th
is
in
tr
o
d
u
ce
d
as
a
r
eg
u
lar
izatio
n
m
eth
o
d
th
a
t
r
an
d
o
m
ly
s
k
ip
s
s
o
m
e
lay
er
s
d
u
r
in
g
th
e
tr
ai
n
in
g
p
r
o
ce
s
s
.
C
o
n
v
en
tio
n
al
s
to
ch
a
s
tic
d
ep
th
in
v
o
lv
es
r
an
d
o
m
s
k
ip
p
in
g
with
a
f
ix
ed
p
r
o
b
a
b
i
lity
.
T
h
is
r
an
d
o
m
s
k
ip
p
in
g
d
is
ca
r
d
s
lay
e
r
s
with
o
u
t
co
n
s
id
er
in
g
th
eir
s
ig
n
if
ic
an
ce
,
wh
ich
m
a
y
elim
in
ate
i
m
p
o
r
tan
t
lay
e
r
s
an
d
r
ed
u
ce
p
er
f
o
r
m
a
n
ce
.
Ad
d
itio
n
ally
,
th
e
f
ix
ed
p
r
o
b
a
b
ilit
y
r
em
ain
s
s
tatic
an
d
d
o
es
n
o
t
ad
ap
t
to
th
e
lear
n
in
g
p
r
o
ce
s
s
.
T
o
m
itig
ate
th
ese
ch
allen
g
es,
th
e
GSD
is
p
r
o
p
o
s
ed
,
wh
ich
d
y
n
am
ically
d
ec
id
es
wh
eth
er
to
s
k
i
p
a
lay
er
b
ased
o
n
its
g
r
ad
ie
n
t m
a
g
n
itu
d
e.
T
h
e
GSD
is
ap
p
lied
to
co
n
v
o
lu
tio
n
al
lay
er
s
,
s
elec
tiv
ely
s
k
ip
p
in
g
la
y
er
s
d
u
r
in
g
th
e
tr
ain
in
g
p
r
o
ce
s
s
.
Alg
o
r
ith
m
1
.
C
NN
with
GSD
I
n
p
u
t: I
n
p
u
t im
a
g
es
Ou
tp
u
t: C
lass
if
ied
d
is
ea
s
e
ty
p
e
I
m
ag
es a
r
e
p
r
e
-
p
r
o
ce
s
s
ed
u
s
in
g
C
L
AHE
an
d
Gau
s
s
ian
b
lu
r
t
ec
h
n
iq
u
e.
Gr
ab
C
u
t seg
m
en
tatio
n
is
u
s
ed
to
is
o
late
co
co
n
u
t le
a
v
es
Featu
r
es a
r
e
ex
tr
ac
ted
u
s
in
g
C
NN
with
GSD
‒
Fo
r
ev
er
y
la
y
er
,
‖
‖
is
m
ea
s
u
r
ed
‒
Dy
n
am
ic
th
r
esh
o
l
d
is
ex
ec
u
ted
‒
T
h
e
s
k
ip
lay
er
with
p
r
o
b
ab
ilit
y
(
)
is
u
s
ed
f
o
r
f
u
r
t
h
er
p
r
o
ce
s
s
in
g
T
h
e
o
u
tp
u
t
is
f
ed
to
th
e
f
u
lly
co
n
n
ec
ted
lay
e
r
s
f
o
r
class
if
icatio
n
.
L
o
s
s
an
d
b
ac
k
p
r
o
p
a
g
atio
n
ar
e
ex
ec
u
te
d
.
T
h
e
weig
h
ts
ar
e
u
p
d
ated
a
n
d
t
h
e
s
k
ip
p
in
g
p
r
o
b
ab
ilit
ies ar
e
a
d
ju
s
ted
d
y
n
a
m
ically
.
3.
E
XP
E
R
I
M
E
N
T
A
L
ANA
L
Y
SI
S
T
h
e
d
e
v
elo
p
ed
alg
o
r
ith
m
is
s
im
u
lated
in
a
p
y
th
o
n
e
n
v
ir
o
n
m
en
t
an
d
th
e
r
e
q
u
ir
ed
co
n
f
ig
u
r
atio
n
s
b
ein
g
i5
p
r
o
ce
s
s
o
r
,
win
d
o
ws
1
0
(
6
4
b
it)
an
d
8
GB
R
AM
.
Per
f
o
r
m
a
n
ce
o
f
d
e
v
elo
p
ed
alg
o
r
ith
m
is
v
alid
ated
i
n
ter
m
s
o
f
v
ar
io
u
s
m
etr
ics
in
c
lu
d
in
g
,
F
1
-
s
co
r
e,
p
r
ec
is
io
n
,
r
ec
all
,
an
d
ac
cu
r
ac
y
.
Ma
th
em
atica
l
f
o
r
m
u
la
f
o
r
m
etr
ics is
g
iv
en
f
r
o
m
(
1
4
)
to
(
1
7
)
.
=
+
+
+
+
×
100
(
1
4
)
=
+
×
100
(
1
5
)
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
Gra
d
ien
t
-
b
a
s
ed
s
to
ch
a
s
tic
d
e
p
th
w
ith
co
n
vo
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k
fo
r
…
(
K
a
vith
a
Ma
g
a
d
i G
o
p
a
la
kris
h
n
a
)
1161
=
+
×
100
(
1
6
)
1
−
=
2
×
×
+
×
100
(
1
7
)
I
n
T
a
b
le
2
,
p
er
f
o
r
m
an
ce
o
f
d
ev
elo
p
ed
alg
o
r
ith
m
is
v
alid
at
ed
u
s
in
g
v
a
r
io
u
s
class
es
o
f
d
a
tasets
with
d
if
f
er
en
t
m
etr
ics.
T
h
e
class
es
n
am
ely
,
b
u
d
r
o
t,
b
u
t
r
o
o
t
d
r
o
p
p
in
g
,
s
tem
b
leed
in
g
,
leaf
r
o
t,
an
d
g
r
ay
leaf
s
p
o
t
ar
e
ev
alu
ated
in
T
a
b
le
2
.
T
h
e
d
ev
el
o
p
ed
m
eth
o
d
o
b
tain
s
an
av
er
a
g
e
ac
cu
r
ac
y
o
f
9
6
.
4
2
%,
9
6
.
1
5
%
o
f
p
r
ec
is
io
n
,
9
5
.
8
7
%
o
f
r
ec
all
,
a
n
d
9
5
.
9
3
%
o
f
F
1
-
s
co
r
e
o
n
cla
s
s
-
wis
e
r
esu
lt
s
.
T
ab
le
3
p
r
esen
ts
th
e
p
er
f
o
r
m
an
ce
o
f
Gr
ab
C
ut
-
b
ased
s
eg
m
en
tati
o
n
tech
n
iq
u
e,
as
ev
alu
ated
o
n
th
e
co
co
n
u
t
tr
ee
d
is
ea
s
e
d
ataset
.
T
h
e
v
ar
io
u
s
ex
is
tin
g
alg
o
r
ith
m
s
u
s
ed
to
e
v
alu
ate
th
e
Gr
a
b
C
u
t
tech
n
iq
u
e
ar
e:
K
-
m
ea
n
s
clu
s
ter
in
g
,
ac
tiv
e
co
n
to
u
r
f
ield
s
,
co
n
d
itio
n
al
r
an
d
o
m
f
ield
s
,
an
d
g
r
ap
h
-
ba
s
ed
s
eg
m
en
tatio
n
a
lg
o
r
ith
m
s
.
T
h
e
Gr
ab
C
u
t
s
eg
m
en
tatio
n
alg
o
r
ith
m
o
b
tain
s
9
6
.
4
2
% a
cc
u
r
ac
y
,
9
6
.
1
5
% p
r
ec
is
io
n
,
9
5
.
8
7
% r
ec
all
,
a
n
d
9
5
.
9
3
%
F
1
-
s
co
r
e
.
T
ab
le
2
.
Per
f
o
r
m
an
ce
o
f
class
-
wis
e
r
esu
lts
C
l
a
s
ses
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
(
%)
B
u
d
r
o
t
9
6
.
2
0
9
6
.
0
0
9
5
.
7
0
9
5
.
8
0
B
u
d
r
o
o
t
d
r
o
p
p
i
n
g
9
6
.
3
5
9
6
.
1
0
9
5
.
8
0
9
5
.
8
5
S
t
e
m
b
l
e
e
d
i
n
g
9
6
.
5
0
9
6
.
2
0
9
5
.
9
0
9
5
.
9
0
Le
a
f
r
o
t
9
6
.
4
5
9
6
.
2
5
9
6
.
0
0
9
6
.
0
0
G
r
a
y
l
e
a
f
sp
o
t
9
6
.
6
0
9
6
.
2
0
9
5
.
9
5
9
6
.
1
0
A
v
e
r
a
g
e
9
6
.
4
2
9
6
.
1
5
9
5
.
8
7
9
5
.
9
3
T
ab
le
3
.
Per
f
o
r
m
an
ce
o
f
s
eg
m
en
tatio
n
alg
o
r
ith
m
s
M
e
t
h
o
d
s
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
(
%)
K
-
me
a
n
s
c
l
u
s
t
e
r
i
n
g
9
3
.
2
9
9
3
.
0
2
9
2
.
9
3
9
2
.
9
9
A
c
t
i
v
e
c
o
n
t
o
u
r
f
i
e
l
d
s
9
3
.
8
7
9
3
.
5
4
9
3
.
1
8
9
3
.
3
2
C
o
n
d
i
t
i
o
n
a
l
r
a
n
d
o
m fi
e
l
d
s
9
4
.
0
8
9
3
.
8
5
9
3
.
4
5
9
3
.
6
5
G
r
a
p
h
-
b
a
s
e
d
se
g
me
n
t
a
t
i
o
n
9
4
.
5
7
9
4
.
2
8
9
3
.
7
8
9
3
.
9
6
G
r
a
b
C
u
t
9
6
.
4
2
9
6
.
1
5
9
5
.
8
7
9
5
.
9
3
T
ab
le
4
p
r
esen
ts
C
NN
p
er
f
o
r
m
an
ce
with
o
u
t
th
e
u
s
e
o
f
GS
D
tech
n
iq
u
e
o
n
th
e
c
o
co
n
u
t
tr
ee
d
is
ea
s
e
d
ataset
.
T
h
e
d
if
f
er
e
n
t
ex
is
tin
g
alg
o
r
ith
m
s
u
s
ed
to
e
v
alu
at
e
th
e
C
NN
with
o
u
t
GSD
tec
h
n
iq
u
e
ar
e:
e
x
tr
em
e
lear
n
in
g
m
a
ch
in
e
s
(
E
L
M
)
,
C
ap
s
Net,
f
u
lly
co
n
n
ec
ted
n
eu
r
al
n
etw
o
r
k
(
FC
NN)
,
an
d
m
u
lti
-
lay
er
p
er
ce
p
tr
o
n
(
ML
P).
C
o
n
v
en
tio
n
al
C
NN
att
ain
s
9
5
.
2
7
%
ac
cu
r
ac
y
,
9
5
.
0
1
%
p
r
ec
is
io
n
,
9
4
.
7
8
%
r
ec
all
,
a
n
d
9
4
.
9
2
%
F
1
-
s
co
r
e
,
o
u
tp
er
f
o
r
m
in
g
th
e
ex
is
tin
g
ap
p
r
o
ac
h
es.
I
n
T
ab
le
5
,
p
er
f
o
r
m
an
ce
o
f
th
e
C
NN
with
GSD
te
ch
n
iq
u
e
is
v
alid
ate
d
u
s
in
g
d
if
f
er
en
t m
etr
ics o
n
th
e
co
co
n
u
t tr
ee
d
is
ea
s
e
d
ataset
.
T
ab
le
4
.
Per
f
o
r
m
an
ce
C
NN
wi
th
o
u
t G
SD m
o
d
el
M
e
t
h
o
d
s
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
(
%)
ELM
9
3
.
5
3
9
3
.
3
8
9
3
.
1
9
9
3
.
2
2
C
a
p
sN
e
t
9
3
.
8
9
9
3
.
6
6
9
3
.
5
7
9
3
.
6
1
F
C
N
N
9
4
.
2
1
9
4
.
0
5
9
3
.
8
8
9
3
.
9
1
M
LP
9
4
.
7
5
9
4
.
5
3
9
4
.
3
2
9
4
.
4
2
C
N
N
9
5
.
2
7
9
5
.
0
1
9
4
.
7
8
9
4
.
9
2
T
ab
le
5
.
Per
f
o
r
m
an
ce
o
f
C
NN
with
GSD
tech
n
iq
u
e
M
e
t
h
o
d
s
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
(
%)
C
a
p
sN
e
t
9
5
.
0
2
9
4
.
8
6
9
4
.
5
9
9
4
.
6
5
F
C
N
N
9
5
.
3
1
9
5
.
1
8
9
4
.
8
3
9
4
.
9
2
M
LP
9
5
.
7
5
9
5
.
5
2
9
5
.
3
7
9
5
.
4
2
C
N
N
9
6
.
0
6
9
5
.
8
4
9
5
.
6
4
9
5
.
7
7
G
S
D
–
C
N
N
9
6
.
4
2
9
6
.
1
5
9
5
.
8
7
9
5
.
9
3
T
h
e
p
er
f
o
r
m
an
ce
o
f
th
e
C
NN
with
GSD
tech
n
iq
u
e
is
co
m
p
ar
ed
with
s
ev
er
al
ex
is
tin
g
alg
o
r
ith
m
s
,
in
clu
d
in
g
C
ap
s
Net
,
FC
NN,
ML
P,
an
d
tr
ad
itio
n
al
C
NN.
T
h
e
GSD
with
C
NN
alg
o
r
ith
m
ac
h
iev
e
d
9
6
.
4
2
%
ac
cu
r
ac
y
,
9
6
.
1
5
%
p
r
ec
is
io
n
,
9
5
.
8
7
%
r
ec
all,
an
d
9
5
.
9
3
%
F1
-
s
co
r
e
wh
en
c
o
m
p
ar
e
d
to
th
ese
ex
is
tin
g
alg
o
r
ith
m
s
.
Fig
u
r
e
2
s
h
o
ws
a
cc
u
r
ac
y
v
s
.
e
p
o
ch
s
g
r
ap
h
f
o
r
th
e
d
e
v
elo
p
ed
alg
o
r
ith
m
an
d
Fig
u
r
e
3
s
h
o
ws
lo
s
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.
15
,
No
.
2
,
Ap
r
il
20
26
:
115
5
-
1
1
6
5
1162
v
s
.
ep
o
ch
s
g
r
ap
h
f
o
r
th
e
d
ev
el
o
p
ed
al
g
o
r
ith
m
.
T
ab
le
6
p
r
esen
ts
th
e
k
-
f
o
ld
v
alid
atio
n
r
esu
lts
f
o
r
th
e
p
r
o
p
o
s
e
d
alg
o
r
ith
m
an
d
T
ab
le
7
p
r
esen
t
s
th
e
p
er
f
o
r
m
an
ce
o
f
th
e
p
r
e
-
p
r
o
ce
s
s
in
g
tech
n
iq
u
es.
Fig
u
r
e
2
.
Acc
u
r
ac
y
v
s
e
p
o
ch
s
g
r
ap
h
f
o
r
t
h
e
d
ev
el
o
p
ed
al
g
o
r
i
th
m
Fig
u
r
e
3
.
L
o
s
s
v
s
e
p
o
ch
s
g
r
a
p
h
f
o
r
t
h
e
d
e
v
elo
p
ed
alg
o
r
ith
m
T
ab
le
6
.
Per
f
o
r
m
an
ce
o
f
d
if
f
er
en
t K
-
f
o
ld
v
alu
es
K
-
v
a
l
u
e
s
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
(
%)
K
=2
9
3
.
0
2
9
2
.
7
8
9
2
.
4
7
9
2
.
1
5
K
=3
9
3
.
6
7
9
3
.
3
7
9
3
.
0
2
9
2
.
7
8
K
=4
9
4
.
3
3
9
4
.
0
5
9
3
.
7
9
9
3
.
2
3
K
=5
9
6
.
4
2
9
6
.
1
5
9
5
.
8
7
9
5
.
9
3
T
ab
le
7
.
Per
f
o
r
m
an
ce
o
f
p
r
e
-
p
r
o
ce
s
s
in
g
tech
n
iq
u
es
M
e
t
h
o
d
s
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
(
%)
M
e
d
i
a
n
f
i
l
t
e
r
9
4
.
8
9
9
4
.
5
6
9
4
.
2
7
9
4
.
4
0
B
i
l
a
t
e
r
a
l
f
i
l
t
e
r
9
5
.
2
8
9
5
.
0
2
9
4
.
6
8
9
4
.
9
3
W
a
v
e
l
e
t
-
b
a
se
d
c
o
n
t
r
a
st
e
n
h
a
n
c
e
m
e
n
t
9
5
.
6
7
9
5
.
2
1
9
5
.
0
5
9
5
.
1
7
A
n
i
s
o
t
r
o
p
i
c
d
i
f
f
u
si
o
n
9
6
.
0
3
9
5
.
8
7
9
5
.
4
3
9
5
.
6
5
C
LA
H
E
a
n
d
G
a
u
ss
i
a
n
b
l
u
r
9
6
.
4
2
9
6
.
1
5
9
5
.
8
7
9
5
.
9
3
3
.
1
.
Co
m
pa
riso
n
a
na
ly
s
is
I
n
th
is
s
ec
tio
n
,
p
er
f
o
r
m
an
ce
o
f
d
ev
elo
p
ed
GSD
-
SNN
alg
o
r
ith
m
is
co
m
p
ar
ed
to
ex
is
tin
g
alg
o
r
ith
m
s
lik
e
DC
NN
[
1
6
]
,
M
I
N
-
SVM
[
1
7
]
,
an
d
R
esNet
-
50
[
1
8
]
on
co
co
n
u
t
tr
ee
leaf
im
ag
es
.
T
h
e
GSD
with
C
NN
alg
o
r
ith
m
ac
h
iev
e
d
9
6
.
4
2
%
ac
cu
r
ac
y
,
9
6
.
1
5
%
p
r
ec
is
io
n
,
9
5
.
8
7
%
r
ec
all,
an
d
9
5
.
9
3
%
F1
-
s
co
r
e
wh
en
co
m
p
ar
ed
to
th
ese
ex
is
tin
g
a
lg
o
r
ith
m
s
.
T
h
e
GSD
tech
n
iq
u
e
is
in
co
r
p
o
r
ate
d
in
to
ea
ch
lay
er
o
f
t
h
e
C
NN
n
etwo
r
k
,
wh
ich
s
k
ip
s
th
e
less
ef
f
icien
t
lay
er
s
b
y
ca
lcu
latin
g
p
r
o
b
ab
ilit
ies
u
s
in
g
g
r
ad
ie
n
t
m
ag
n
itu
d
es.
T
h
is
p
r
o
ce
s
s
m
in
im
izes
tr
ain
in
g
t
im
e
an
d
im
p
r
o
v
es
class
if
icatio
n
p
e
r
f
o
r
m
an
ce
with
h
ig
h
ac
cu
r
ac
y
.
T
ab
le
8
p
r
esen
ts
co
m
p
ar
ativ
e
a
n
aly
s
is
o
f
d
ev
el
o
p
ed
m
o
d
el.
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
Gra
d
ien
t
-
b
a
s
ed
s
to
ch
a
s
tic
d
e
p
th
w
ith
co
n
vo
lu
tio
n
a
l n
eu
r
a
l
n
etw
o
r
k
fo
r
…
(
K
a
vith
a
Ma
g
a
d
i G
o
p
a
la
kris
h
n
a
)
1163
T
ab
le
8
.
C
o
m
p
a
r
ativ
e
an
aly
s
is
o
f
d
e
v
elo
p
ed
m
o
d
el
M
e
t
h
o
d
s
D
a
t
a
s
e
t
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
(
%)
D
C
N
N
[
1
6
]
C
o
c
o
n
u
t
t
r
e
e
l
e
a
f
i
ma
g
e
s
9
5
.
7
1
NA
NA
NA
M
I
N
-
S
V
M
[
1
7
]
9
5
.
3
5
95
95
95
R
e
sN
e
t
-
5
0
[
1
8
]
94
9
2
.
2
5
9
3
.
4
2
9
3
.
8
8
P
r
o
p
o
se
d
G
S
D
-
C
N
N
9
6
.
4
2
9
6
.
1
5
9
5
.
8
7
9
5
.
9
3
3
.
2
.
Dis
cus
s
io
n
T
h
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
d
ev
elo
p
ed
a
lg
o
r
i
th
m
i
s
ev
a
lu
ate
d
o
n
co
co
n
u
t
tr
ee
d
i
s
ea
s
e
d
a
ta
s
e
t
u
s
in
g
d
if
f
e
r
en
t
m
e
tr
i
c
s
.
Per
f
o
r
m
an
ce
o
f
d
ev
elo
p
ed
a
lg
o
r
ith
m
i
s
a
s
s
e
s
s
ed
b
a
s
ed
o
n
c
la
s
s
-
wi
s
e
r
e
s
u
l
t
s
,
u
s
in
g
s
eg
m
en
ta
tio
n
-
b
a
s
ed
m
eth
o
d
s
an
d
c
la
s
s
i
f
i
ca
tio
n
a
lg
o
r
ith
m
s
.
Ad
d
i
tio
n
al
ly
,
t
h
e
p
er
f
o
r
m
an
ce
o
f
d
ev
elo
p
ed
alg
o
r
it
h
m
i
s
co
m
p
a
r
ed
w
ith
ex
i
s
t
in
g
a
lg
o
r
ith
m
s
:
D
C
NN
[
1
6
]
,
MI
N
-
S
VM
[
1
7
]
,
an
d
R
e
s
N
et
-
50
[
1
8
]
.
T
h
e
ex
i
s
t
in
g
alg
o
r
i
th
m
s
h
av
e
l
im
i
ta
tio
n
s
as
n
o
t
ed
f
u
r
th
er
:
s
t
r
u
g
g
le
s
in
d
i
f
f
er
en
t
ia
ti
n
g
d
if
f
er
en
t
d
i
s
ea
s
e
c
la
s
s
e
s
,
n
o
en
h
an
c
em
en
t
o
f
im
ag
e
co
n
tr
a
s
t,
f
a
ilu
r
e
to
el
im
in
a
te
i
m
ag
e
n
o
i
s
e,
lo
wer
r
ep
r
e
s
en
ta
tio
n
o
f
f
ea
tu
r
e
an
d
cla
s
s
if
i
ca
tio
n
p
er
f
o
r
m
an
ce
b
e
ca
u
s
e
o
f
i
s
s
u
e
s
li
k
e
in
tr
a
cl
as
s
v
ar
iab
i
li
ty
an
d
i
n
t
er
cl
a
s
s
s
im
il
ar
i
ty
.
T
o
m
it
ig
at
e
th
e
s
e
l
im
i
ta
ti
o
n
s
,
t
h
i
s
r
es
ea
r
c
h
p
r
o
p
o
s
e
s
G
S
D
ap
p
r
o
a
ch
w
it
h
C
NN
m
e
th
o
d
,
wh
ich
i
s
c
la
s
s
i
f
i
ed
in
to
d
if
f
er
en
t
d
i
s
ea
s
e
cl
a
s
s
e
s
ef
f
ec
tiv
ely
.
B
y
in
co
r
p
o
r
a
tin
g
th
e
G
S
D
in
t
h
e
C
NN
m
et
h
o
d
,
th
e
l
es
s
c
o
n
tr
ib
u
t
in
g
l
ay
er
is
s
k
ip
p
ed
,
wh
i
le
th
e
t
r
a
in
in
g
ti
m
e
i
s
m
in
im
iz
ed
w
ith
a
h
ig
h
cla
s
s
if
i
ca
tio
n
p
er
f
o
r
m
an
c
e.
I
n
th
e
p
r
e
-
p
r
o
ce
s
s
in
g
p
h
as
e,
t
h
e
C
L
AH
E
an
d
Gau
s
s
i
an
b
lu
r
te
ch
n
iq
u
e
s
ar
e
u
s
ed
to
en
h
an
ce
im
ag
e
co
n
tr
a
s
t
a
n
d
el
im
in
at
e
n
o
i
s
e,
r
es
p
ec
ti
v
e
ly
.
T
h
r
o
u
g
h
th
e
s
e
p
r
o
ce
s
s
e
s
,
th
e
i
s
s
u
e
o
f
in
t
er
-
cl
as
s
s
im
il
ar
i
ty
an
d
in
tr
a
-
c
l
as
s
v
ar
iab
il
ity
ar
e
el
im
in
at
ed
an
d
th
e
r
ep
r
e
s
en
ta
tio
n
o
f
f
ea
tu
r
e
s
f
o
r
d
i
s
e
a
s
e
cl
as
s
if
i
ca
tio
n
ar
e
en
h
an
ce
d
.
T
h
e
G
S
D
wi
th
C
NN
alg
o
r
it
h
m
o
b
ta
in
s
ac
c
u
r
a
cy
o
f
9
6
.
4
2
%,
p
r
e
ci
s
io
n
o
f
9
6
.
1
5
%,
r
ec
al
l
o
f
9
5
.
8
7
%
,
a
n
d
F
1
-
s
c
o
r
e
o
f
9
5
.
9
3
%
wh
i
le
co
m
p
ar
in
g
wi
th
ex
i
s
tin
g
al
g
o
r
i
th
m
s
.
H
ig
h
p
er
f
o
r
m
a
n
ce
o
f
p
r
o
p
o
s
ed
GS
D
-
C
NN
i
s
at
tr
ib
u
t
ed
to
th
e
i
r
ad
ap
tiv
e
lay
er
-
s
k
ip
p
in
g
m
e
c
h
an
i
s
m
d
ep
en
d
ed
o
n
g
r
ad
ie
n
t
m
ag
n
itu
d
e
s
,
th
a
t
p
r
e
s
er
v
e
s
h
ig
h
in
f
o
r
m
a
ti
v
e
f
ea
tu
r
e
r
ep
r
e
s
en
ta
tio
n
s
wh
en
m
in
im
iz
in
g
r
ed
u
n
d
an
t
co
m
p
u
ta
tio
n
s
.
U
n
l
ik
e
tr
ad
i
tio
n
a
l
s
to
ch
a
s
t
ic
d
ep
t
h
,
g
r
ad
i
en
t
-
a
war
e
s
k
ip
p
in
g
s
tr
ateg
y
p
r
ev
en
t
s
el
im
in
a
tio
n
o
f
e
s
s
en
ti
al
lay
er
s
,
b
y
en
h
an
c
in
g
f
ea
tu
r
e
d
i
s
cr
im
i
n
a
tio
n
ac
r
o
s
s
in
te
r
-
c
l
as
s
v
a
r
i
ab
i
li
ty
an
d
in
t
r
a
-
cl
as
s
v
ar
ia
b
i
li
ty
.
Mo
r
eo
v
er
,
c
o
m
b
in
a
tio
n
o
f
C
L
A
HE
an
d
Gau
s
s
ian
b
lu
r
im
p
r
o
v
e
s
d
i
s
ea
s
e
v
i
s
ib
i
li
ty
an
d
n
o
is
e
s
u
p
p
r
e
s
s
io
n
,
wh
en
Gr
ap
C
u
t
s
e
g
m
en
t
at
io
n
en
s
u
r
e
s
lea
r
n
in
g
o
n
d
is
ea
s
e
-
af
f
ec
ted
r
eg
io
n
s
,
th
e
s
e
i
m
p
r
o
v
s
c
l
as
s
if
i
ca
tio
n
ac
c
u
r
a
cy
.
R
at
h
e
r
th
an
i
t
s
s
tr
o
n
g
p
er
f
o
r
m
an
c
e,
p
r
o
p
o
s
ed
m
o
d
e
l
s
h
o
w
s
m
i
s
c
la
s
s
if
i
ca
ti
o
n
,
e
s
p
ec
i
al
ly
wh
e
r
e
d
i
s
ea
s
e
s
y
m
p
to
m
s
o
v
er
l
ap
p
in
g
v
i
s
u
a
l
p
at
ter
n
s
wi
th
s
u
b
t
le
t
ex
tu
r
e
v
ar
ia
tio
n
s
.
Sam
e
co
l
o
r
an
d
s
t
r
u
c
tu
r
a
l
ch
a
r
ac
te
r
i
s
ti
cs
a
m
o
n
g
ce
r
t
ain
d
i
s
ea
s
e
c
la
s
s
e
s
l
ik
e
le
af
r
o
t
an
d
g
r
ay
le
af
s
p
o
t
lead
s
to
cl
a
s
s
i
f
ic
at
io
n
am
b
i
g
u
i
ty
.
4.
CO
NCLU
SI
O
N
T
h
is
r
esear
ch
d
ev
el
o
p
s
a
D
L
-
b
ased
alg
o
r
ith
m
f
o
r
co
c
o
n
u
t
tr
ee
leaf
d
is
ea
s
e
class
if
icatio
n
.
T
h
e
co
co
n
u
t tr
ee
d
is
ea
s
e
d
ataset
u
s
ed
in
clu
d
es d
is
ea
s
e
-
lab
eled
im
ag
es.
T
h
e
im
ag
es in
d
ataset
ar
e
p
r
e
-
p
r
o
ce
s
s
ed
b
y
C
L
AHE
an
d
Gau
s
s
ian
b
lu
r
t
ec
h
n
iq
u
es,
wh
ich
en
h
an
ce
i
m
ag
e
co
n
tr
ast
an
d
elim
in
ate
n
o
is
e,
wh
ich
h
elp
s
d
if
f
er
en
tiate
d
if
f
e
r
en
t
d
is
ea
s
e
class
e
s
.
T
h
en
,
th
e
im
ag
es
ar
e
s
eg
m
en
ted
u
s
in
g
Gr
ab
C
u
t
alg
o
r
ith
m
,
wh
ich
is
o
lates
th
e
leaf
im
a
g
es
f
r
o
m
th
e
b
ac
k
g
r
o
u
n
d
th
r
o
u
g
h
a
n
ite
r
ativ
e
p
r
o
ce
s
s
to
e
n
h
an
ce
f
ea
t
u
r
e
r
e
p
r
esen
tatio
n
.
Fin
ally
,
class
if
icat
io
n
is
p
er
f
o
r
m
ed
u
s
in
g
th
e
d
ev
el
o
p
ed
GSD
-
C
NN
tech
n
iq
u
e,
wh
ich
ex
tr
ac
ts
h
ier
ar
ch
ical
f
ea
tu
r
es
th
r
o
u
g
h
th
e
co
n
v
o
lu
ti
o
n
al
lay
er
s
an
d
class
if
ies
th
e
d
if
f
er
en
t
d
is
ea
s
e
class
es
with
h
ig
h
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
T
h
e
GSD
tech
n
i
q
u
e
is
in
co
r
p
o
r
ate
d
in
to
ev
er
y
lay
er
o
f
th
e
C
NN,
s
k
ip
p
in
g
la
y
er
s
th
at
co
n
tr
ib
u
te
less
to
cla
s
s
if
icatio
n
an
d
m
in
im
izin
g
th
e
m
o
d
el'
s
tr
ain
in
g
tim
e
wh
ile
m
ain
tain
in
g
h
ig
h
cla
s
s
if
icatio
n
ac
cu
r
ac
y
.
T
h
e
GSD
with
C
NN
alg
o
r
ith
m
ac
h
iev
ed
9
6
.
4
2
%
ac
c
u
r
ac
y
,
9
6
.
1
5
%
p
r
ec
is
io
n
,
9
5
.
8
7
%
r
ec
all,
an
d
9
5
.
9
3
%
F1
-
s
co
r
e
wh
en
co
m
p
ar
e
d
to
e
x
is
tin
g
alg
o
r
ith
m
s
.
I
n
f
u
tu
r
e
wo
r
k
,
o
th
er
DL
-
b
ased
m
o
d
els
ca
n
b
e
ex
p
lo
r
ed
to
f
u
r
th
er
e
n
h
an
ce
class
if
icatio
n
p
er
f
o
r
m
an
ce
.
ACK
NO
WL
E
DG
M
E
N
T
S
W
e
wo
u
ld
lik
e
to
g
r
atef
u
l
to
o
u
r
C
o
lleag
u
e'
s
f
o
r
th
eir
ass
is
tan
ce
with
d
ata
co
llectio
n
an
d
f
o
r
th
ei
r
in
s
ig
h
tf
u
l d
is
cu
s
s
io
n
s
.
F
UNDING
I
NF
O
R
M
A
T
I
O
N
Au
th
o
r
s
s
tate
n
o
f
u
n
d
in
g
in
v
o
lv
ed
.
AUTHO
R
CO
NT
RI
B
UT
I
O
NS ST
A
T
E
M
E
N
T
T
h
is
jo
u
r
n
al
u
s
es
th
e
C
o
n
tr
ib
u
to
r
R
o
les
T
ax
o
n
o
m
y
(
C
R
ed
iT)
to
r
ec
o
g
n
ize
in
d
iv
id
u
al
au
th
o
r
co
n
tr
ib
u
tio
n
s
,
r
ed
u
ce
au
th
o
r
s
h
ip
d
is
p
u
tes,
an
d
f
ac
ilit
ate
co
llab
o
r
atio
n
.
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.
15
,
No
.
2
,
Ap
r
il
20
26
:
115
5
-
1
1
6
5
1164
Na
m
e
o
f
Aut
ho
r
C
M
So
Va
Fo
I
R
D
O
E
Vi
Su
P
Fu
Kav
ith
a
Ma
g
ad
i
Go
p
alak
r
is
h
n
a
✓
✓
✓
✓
✓
✓
✓
✓
✓
✓
R
av
ip
r
ak
ash
Ma
d
en
u
r
L
in
g
ar
aju
✓
✓
✓
✓
✓
✓
✓
An
an
d
a
B
ab
u
J
ay
ac
h
an
d
r
a
✓
✓
✓
✓
✓
✓
✓
✓
C
:
C
o
n
c
e
p
t
u
a
l
i
z
a
t
i
o
n
M
:
M
e
t
h
o
d
o
l
o
g
y
So
:
So
f
t
w
a
r
e
Va
:
Va
l
i
d
a
t
i
o
n
Fo
:
Fo
r
mal
a
n
a
l
y
s
i
s
I
:
I
n
v
e
s
t
i
g
a
t
i
o
n
R
:
R
e
so
u
r
c
e
s
D
:
D
a
t
a
C
u
r
a
t
i
o
n
O
:
W
r
i
t
i
n
g
-
O
r
i
g
i
n
a
l
D
r
a
f
t
E
:
W
r
i
t
i
n
g
-
R
e
v
i
e
w
&
E
d
i
t
i
n
g
Vi
:
Vi
su
a
l
i
z
a
t
i
o
n
Su
:
Su
p
e
r
v
i
s
i
o
n
P
:
P
r
o
j
e
c
t
a
d
mi
n
i
st
r
a
t
i
o
n
Fu
:
Fu
n
d
i
n
g
a
c
q
u
i
si
t
i
o
n
CO
NF
L
I
C
T
O
F
I
N
T
E
R
E
S
T
ST
A
T
E
M
E
NT
T
h
e
au
th
o
r
s
d
ec
lar
e
n
o
co
n
f
lict o
f
in
ter
est.
DATA AV
AI
L
AB
I
L
I
T
Y
T
h
e
d
ata
th
at
s
u
p
p
o
r
t
th
e
f
in
d
in
g
s
o
f
t
h
is
s
tu
d
y
ar
e
o
p
en
ly
av
ailab
le
i
n
Me
n
d
ele
y
Data
at
h
ttp
s
://d
o
i.o
r
g
/1
0
.
1
7
6
3
2
/
g
h
5
6
wb
s
n
j5
.
1
,
r
e
f
er
en
ce
n
u
m
b
er
[
2
1
]
.
RE
F
E
R
E
NC
E
S
[
1
]
R
.
H
a
y
a
t
i
,
A
.
A
.
M
u
n
a
w
a
r
,
E
.
Lu
k
i
t
a
n
i
n
g
si
h
,
N
.
E
a
r
l
i
a
,
T.
K
a
r
m
a
,
a
n
d
R
.
I
d
r
o
e
s,
“
C
o
mb
i
n
a
t
i
o
n
o
f
P
C
A
w
i
t
h
LD
A
a
n
d
S
V
M
c
l
a
ss
i
f
i
e
r
s:
a
m
o
d
e
l
f
o
r
d
e
t
e
r
m
i
n
i
n
g
t
h
e
g
e
o
g
r
a
p
h
i
c
a
l
o
r
i
g
i
n
o
f
c
o
c
o
n
u
t
i
n
t
h
e
c
o
a
st
a
l
p
l
a
n
t
a
t
i
o
n
,
A
c
e
h
P
r
o
v
i
n
c
e
,
I
n
d
o
n
e
si
a
,
”
C
a
s
e
S
t
u
d
i
e
s
i
n
C
h
e
m
i
c
a
l
a
n
d
En
v
i
r
o
n
m
e
n
t
a
l
En
g
i
n
e
e
r
i
n
g
,
v
o
l
.
9
,
2
0
2
4
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
c
sc
e
e
.
2
0
2
3
.
1
0
0
5
5
2
.
[
2
]
R
.
G
a
j
j
a
r
,
N
.
G
a
j
j
a
r
,
V
.
J
.
Th
a
k
o
r
,
N
.
P
.
P
a
t
e
l
,
a
n
d
S
.
R
u
p
a
r
e
l
i
a
,
“
R
e
a
l
-
t
i
m
e
d
e
t
e
c
t
i
o
n
a
n
d
i
d
e
n
t
i
f
i
c
a
t
i
o
n
o
f
p
l
a
n
t
l
e
a
f
d
i
s
e
a
s
e
s
u
si
n
g
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
o
n
a
n
e
mb
e
d
d
e
d
p
l
a
t
f
o
r
m,”
T
h
e
V
i
su
a
l
C
o
m
p
u
t
e
r
,
v
o
l
.
3
8
,
n
o
.
8
,
p
p
.
2
9
2
3
–
2
9
3
8
,
2
0
2
2
,
d
o
i
:
1
0
.
1
0
0
7
/
s
0
0
3
7
1
-
021
-
0
2
1
6
4
-
9.
[
3
]
L.
L
i
u
,
W
.
Y
a
n
,
a
n
d
B
.
L
i
u
,
“
Tr
a
n
s
c
r
i
p
t
o
me
se
q
u
e
n
c
i
n
g
o
f
C
o
c
o
s
n
u
c
i
f
e
ra
l
e
a
v
e
s
i
n
r
e
s
p
o
n
s
e
t
o
R
h
y
n
c
h
o
p
h
o
r
u
s
f
e
r
ru
g
i
n
e
u
s
i
n
f
e
s
t
a
t
i
o
n
,
”
Fr
o
n
t
i
e
rs
i
n
G
e
n
e
t
i
c
s
,
v
o
l
.
1
4
,
2
0
2
3
,
d
o
i
:
1
0
.
3
3
8
9
/
f
g
e
n
e
.
2
0
2
3
.
1
1
1
5
3
9
2
.
[
4
]
L.
Lu
e
t
a
l
.
,
“
I
n
t
e
g
r
a
t
e
d
t
r
a
n
scri
p
t
o
mi
c
a
n
d
m
e
t
a
b
o
l
o
m
i
c
a
n
a
l
y
ses
r
e
v
e
a
l
k
e
y
m
e
t
a
b
o
l
i
c
p
a
t
h
w
a
y
s
i
n
r
e
s
p
o
n
se
t
o
p
o
t
a
ss
i
u
m
d
e
f
i
c
i
e
n
c
y
i
n
c
o
c
o
n
u
t
(
C
o
c
o
s
n
u
c
i
f
e
r
a
L.
)
s
e
e
d
l
i
n
g
s,
”
Fr
o
n
t
i
e
rs
i
n
P
l
a
n
t
S
c
i
e
n
c
e
,
v
o
l
.
1
4
,
2
0
2
3
,
d
o
i
:
1
0
.
3
3
8
9
/
f
p
l
s.
2
0
2
3
.
1
1
1
2
2
6
4
.
[
5
]
A
.
S
u
n
p
a
p
a
o
e
t
a
l
.
,
“
M
o
r
p
h
o
l
o
g
i
c
a
l
a
n
d
mo
l
e
c
u
l
a
r
i
d
e
n
t
i
f
i
c
a
t
i
o
n
o
f
p
l
a
n
t
p
a
t
h
o
g
e
n
i
c
f
u
n
g
i
a
ss
o
c
i
a
t
e
d
w
i
t
h
d
i
r
t
y
p
a
n
i
c
l
e
d
i
se
a
se
i
n
c
o
c
o
n
u
t
s (
C
o
c
o
s
n
u
c
i
f
e
r
a
)
i
n
T
h
a
i
l
a
n
d
,
”
J
o
u
rn
a
l
o
f
F
u
n
g
i
,
v
o
l
.
8
,
n
o
.
4
,
2
0
2
2
,
d
o
i
:
1
0
.
3
3
9
0
/
j
o
f
8
0
4
0
3
3
5
.
[
6
]
V
.
B
a
l
a
f
a
s,
E
.
K
a
r
a
n
t
o
u
m
a
n
i
s,
M
.
L
o
u
t
a
,
a
n
d
N
.
P
l
o
sk
a
s,
“
M
a
c
h
i
n
e
l
e
a
r
n
i
n
g
a
n
d
d
e
e
p
l
e
a
r
n
i
n
g
f
o
r
p
l
a
n
t
d
i
se
a
se
c
l
a
ss
i
f
i
c
a
t
i
o
n
a
n
d
d
e
t
e
c
t
i
o
n
,
”
I
EEE
Ac
c
e
ss
,
v
o
l
.
1
1
,
p
p
.
1
1
4
3
5
2
–
1
1
4
3
7
7
,
2
0
2
3
,
d
o
i
:
1
0
.
1
1
0
9
/
A
C
C
ESS
.
2
0
2
3
.
3
3
2
4
7
2
2
.
[
7
]
S
.
K
.
B
r
a
r
,
R
.
S
h
a
r
m
a
,
S
.
V
a
t
s,
a
n
d
V
.
K
u
k
r
e
j
a
,
“
A
sm
a
r
t
a
p
p
r
o
a
c
h
t
o
c
o
c
o
n
u
t
l
e
a
f
sp
o
t
d
i
se
a
se
c
l
a
ss
i
f
i
c
a
t
i
o
n
u
s
i
n
g
c
o
m
p
u
t
e
r
v
i
s
i
o
n
a
n
d
d
e
e
p
l
e
a
r
n
i
n
g
t
e
c
h
n
i
q
u
e
,
”
i
n
2
0
2
3
Wo
r
l
d
C
o
n
f
e
re
n
c
e
o
n
C
o
m
m
u
n
i
c
a
t
i
o
n
&
C
o
m
p
u
t
i
n
g
,
2
0
2
3
,
p
p
.
1
–
6
,
d
o
i
:
1
0
.
1
1
0
9
/
W
C
O
N
F
5
8
2
7
0
.
2
0
2
3
.
1
0
2
3
5
2
5
1
.
[
8
]
A
.
B
h
u
j
e
l
,
N
.
-
E
.
K
i
m
,
E
.
A
r
u
l
mo
z
h
i
,
J.
K
.
B
a
s
a
k
,
a
n
d
H
.
-
T.
K
i
m,
“
A
l
i
g
h
t
w
e
i
g
h
t
a
t
t
e
n
t
i
o
n
-
b
a
se
d
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
f
o
r
t
o
ma
t
o
l
e
a
f
d
i
s
e
a
s
e
c
l
a
ss
i
f
i
c
a
t
i
o
n
,
”
Ag
r
i
c
u
l
t
u
re
,
v
o
l
.
1
2
,
n
o
.
2
,
2
0
2
2
,
d
o
i
:
1
0
.
3
3
9
0
/
a
g
r
i
c
u
l
t
u
r
e
1
2
0
2
0
2
2
8
.
[
9
]
K
.
Jas
p
i
n
,
S
.
S
e
l
v
a
n
,
P
.
S
.
P
a
c
k
i
a
n
a
t
h
a
n
,
a
n
d
P
.
K
u
mar,
“
C
o
n
v
o
l
u
t
i
o
n
n
e
u
r
a
l
n
e
t
w
o
r
k
b
a
s
e
d
c
l
a
ss
i
f
i
c
a
t
i
o
n
o
f
p
l
a
n
t
l
e
a
f
d
i
s
e
a
s
e
i
ma
g
e
s,
”
i
n
Ap
p
l
i
c
a
t
i
o
n
s
o
f
Ar
t
i
f
i
c
i
a
l
I
n
t
e
l
l
i
g
e
n
c
e
a
n
d
M
a
c
h
i
n
e
L
e
a
rn
i
n
g
,
S
i
n
g
a
p
o
r
e
:
S
p
r
i
n
g
e
r
,
2
0
2
2
,
p
p
.
5
1
1
–
5
2
7
,
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
9
8
1
-
19
-
4
8
3
1
-
2
_
4
2
.
[
1
0
]
P
.
R
.
D
.
S
i
l
v
a
,
C
.
N
.
P
e
r
e
r
a
,
B
.
W
.
B
a
h
d
e
r
,
a
n
d
R
.
N
.
A
t
t
a
n
a
y
a
k
e
,
“
N
e
st
e
d
P
C
R
-
b
a
se
d
r
a
p
i
d
d
e
t
e
c
t
i
o
n
o
f
p
h
y
t
o
p
l
a
sma
l
e
a
f
w
i
l
t
d
i
s
e
a
se
o
f
c
o
c
o
n
u
t
i
n
S
r
i
La
n
k
a
a
n
d
sy
s
t
e
m
i
c
m
o
v
e
me
n
t
o
f
t
h
e
p
a
t
h
o
g
e
n
,
”
P
a
t
h
o
g
e
n
s
,
v
o
l
.
1
2
,
n
o
.
2
,
2
0
2
3
,
d
o
i
:
1
0
.
3
3
9
0
/
p
a
t
h
o
g
e
n
s
1
2
0
2
0
2
9
4
.
[
1
1
]
H
.
D
.
M
a
f
u
k
i
d
z
e
,
G
.
O
w
o
m
u
g
i
s
h
a
,
D
.
O
t
i
m,
A
.
N
e
c
h
i
b
v
u
t
e
,
C
.
N
y
a
mh
e
r
e
,
a
n
d
F
.
M
a
z
u
n
g
a
,
“
A
d
a
p
t
i
v
e
t
h
r
e
s
h
o
l
d
i
n
g
o
f
C
N
N
f
e
a
t
u
r
e
s
f
o
r
ma
i
z
e
l
e
a
f
d
i
sea
s
e
c
l
a
ss
i
f
i
c
a
t
i
o
n
a
n
d
se
v
e
r
i
t
y
e
s
t
i
m
a
t
i
o
n
,
”
A
p
p
l
i
e
d
S
c
i
e
n
c
e
s
,
v
o
l
.
1
2
,
n
o
.
1
7
,
2
0
2
2
,
d
o
i
:
1
0
.
3
3
9
0
/
a
p
p
1
2
1
7
8
4
1
2
.
[
1
2
]
W
.
L
i
,
L
.
Zh
u
,
a
n
d
J.
Li
u
,
“
P
L
-
D
I
N
O
:
a
n
i
m
p
r
o
v
e
d
t
r
a
n
sf
o
r
mer
-
b
a
se
d
me
t
h
o
d
f
o
r
p
l
a
n
t
l
e
a
f
d
i
se
a
se
d
e
t
e
c
t
i
o
n
,
”
A
g
ri
c
u
l
t
u
re
,
v
o
l
.
1
4
,
n
o
.
5
,
2
0
2
4
,
d
o
i
:
1
0
.
3
3
9
0
/
a
g
r
i
c
u
l
t
u
r
e
1
4
0
5
0
6
9
1
.
[
1
3
]
J
.
L
i
e
t
a
l
.
,
“
G
e
n
o
m
e
-
w
i
d
e
i
d
e
n
t
i
f
i
c
a
t
i
o
n
,
c
l
a
s
s
i
f
i
c
a
t
i
o
n
a
n
d
e
x
p
r
e
s
s
i
o
n
a
n
a
l
y
s
i
s
o
f
M
Y
B
g
e
n
e
f
a
m
i
l
y
i
n
c
o
c
o
n
u
t
(
C
o
c
o
s
n
u
c
i
f
e
r
a
L
.
)
,
”
Fro
n
t
i
e
rs
i
n
P
l
a
n
t
S
c
i
e
n
c
e
,
v
o
l
.
1
4
,
2
0
2
4
,
d
o
i
:
1
0
.
3
3
8
9
/
f
p
l
s
.
2
0
2
3
.
1
2
6
3
5
9
5
.
[
1
4
]
A
.
P
a
l
a
n
a
n
d
a
a
n
d
W
.
K
i
m
p
a
n
,
“
C
l
a
ssi
f
i
c
a
t
i
o
n
o
f
a
d
u
l
t
e
r
a
t
e
d
p
a
r
t
i
c
l
e
i
m
a
g
e
s
i
n
c
o
c
o
n
u
t
o
i
l
u
si
n
g
d
e
e
p
l
e
a
r
n
i
n
g
a
p
p
r
o
a
c
h
e
s,
”
Ap
p
l
i
e
d
S
c
i
e
n
c
e
s
,
v
o
l
.
1
2
,
n
o
.
2
,
2
0
2
2
,
d
o
i
:
1
0
.
3
3
9
0
/
a
p
p
1
2
0
2
0
6
5
6
.
[
1
5
]
Y
.
C
h
e
n
e
t
a
l
.
,
“
D
F
C
A
N
e
t
:
a
n
o
v
e
l
l
i
g
h
t
w
e
i
g
h
t
c
o
n
v
o
l
u
t
i
o
n
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
m
o
d
e
l
f
o
r
c
o
r
n
d
i
sea
s
e
i
d
e
n
t
i
f
i
c
a
t
i
o
n
,
”
A
g
ri
c
u
l
t
u
re
,
v
o
l
.
1
2
,
n
o
.
1
2
,
2
0
2
2
,
d
o
i
:
1
0
.
3
3
9
0
/
a
g
r
i
c
u
l
t
u
r
e
1
2
1
2
2
0
4
7
.
[
1
6
]
V
.
K
a
v
i
t
h
a
ma
n
i
a
n
d
S
.
U
m
a
M
a
h
e
s
w
a
r
i
,
“
I
n
v
e
s
t
i
g
a
t
i
o
n
o
f
d
e
e
p
l
e
a
r
n
i
n
g
f
o
r
w
h
i
t
e
f
l
y
i
d
e
n
t
i
f
i
c
a
t
i
o
n
i
n
c
o
c
o
n
u
t
t
r
e
e
l
e
a
v
e
s,”
I
n
t
e
l
l
i
g
e
n
t
S
y
s
t
e
m
s
w
i
t
h
A
p
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
2
0
,
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
sw
a
.
2
0
2
3
.
2
0
0
2
9
0
.
[
1
7
]
R
.
K
.
M
e
g
a
l
i
n
g
a
m
,
S
.
K
.
M
a
n
o
h
a
r
a
n
,
D
.
H
.
T.
A
.
B
a
b
u
,
G
.
S
r
i
r
a
m,
K
.
Lo
k
e
sh
,
a
n
d
S
.
K
.
S
u
d
h
e
e
s
h
,
“
C
o
c
o
n
u
t
t
r
e
e
s
c
l
a
ss
i
f
i
c
a
t
i
o
n
b
a
s
e
d
o
n
h
e
i
g
h
t
,
i
n
c
l
i
n
a
t
i
o
n
,
a
n
d
o
r
i
e
n
t
a
t
i
o
n
u
si
n
g
M
I
N
-
S
V
M
a
l
g
o
r
i
t
h
m,
”
N
e
u
ra
l
C
o
m
p
u
t
i
n
g
a
n
d
Ap
p
l
i
c
a
t
i
o
n
s
,
v
o
l
.
3
5
,
n
o
.
1
6
,
p
p
.
1
2
0
5
5
–
1
2
0
7
1
,
2
0
2
3
,
d
o
i
:
1
0
.
1
0
0
7
/
s0
0
5
2
1
-
0
2
3
-
0
8
3
3
9
-
w.
[
1
8
]
S
.
T
h
i
t
e
,
Y
.
S
u
r
y
a
w
a
n
s
h
i
,
K
.
P
a
t
i
l
,
a
n
d
P
.
C
h
u
m
c
h
u
,
“
C
o
c
o
n
u
t
(
C
o
c
o
s
n
u
c
i
f
e
r
a
)
t
r
e
e
d
i
se
a
se
d
a
t
a
se
t
:
a
d
a
t
a
set
f
o
r
d
i
se
a
se
d
e
t
e
c
t
i
o
n
a
n
d
c
l
a
ss
i
f
i
c
a
t
i
o
n
f
o
r
m
a
c
h
i
n
e
l
e
a
r
n
i
n
g
a
p
p
l
i
c
a
t
i
o
n
s,
”
D
a
t
a
i
n
Bri
e
f
,
v
o
l
.
5
1
,
2
0
2
3
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
d
i
b
.
2
0
2
3
.
1
0
9
6
9
0
.
Evaluation Warning : The document was created with Spire.PDF for Python.