I
nte
rna
t
io
na
l J
o
urna
l o
f
Adv
a
nces in Applie
d Science
s
(
I
J
AAS)
Vo
l.
1
4
,
No
.
1
,
Ma
r
ch
2
0
2
5
,
p
p
.
28
~
38
I
SS
N:
2252
-
8
8
1
4
,
DOI
:
1
0
.
1
1
5
9
1
/ijaas
.
v
1
4
.
i
1
.
pp
28
-
38
28
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
a
a
s
.
ia
esco
r
e.
co
m
Betta fish
species
cla
ss
ificatio
n usin
g
light weight
dee
p learning
a
lg
o
rithm
Da
nis
ha
h H
a
na
M
u
ha
m
m
a
d
M
uh
a
im
in L
im
,
No
riza
n M
a
t
Dia
h,
Z
a
ida
h Ib
ra
him
,
Z
o
lid
a
h K
a
s
ira
n
S
c
h
o
o
l
o
f
C
o
m
p
u
t
i
n
g
S
c
i
e
n
c
e
s,
C
o
l
l
e
g
e
o
f
C
o
mp
u
t
i
n
g
,
I
n
f
o
r
ma
t
i
c
s
a
n
d
M
a
t
h
e
ma
t
i
c
s,
U
n
i
v
e
r
s
i
t
i
T
e
k
n
o
l
o
g
i
M
A
R
A
,
S
e
l
a
n
g
o
r
,
M
a
l
a
y
si
a
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Ma
r
1
2
,
2
0
2
4
R
ev
is
ed
Sep
2
7
,
2
0
2
4
Acc
ep
ted
Oct
8
,
2
0
2
4
Be
tt
a
fish
se
ll
e
rs
a
n
d
b
re
e
d
e
rs
o
ft
e
n
fa
c
e
c
h
a
ll
e
n
g
e
s
in
a
c
c
u
ra
tely
id
e
n
ti
f
y
in
g
Be
tt
a
fish
sp
e
c
ies
d
u
e
t
o
v
a
riatio
n
s
in
c
o
l
o
rs,
p
a
tt
e
rn
s,
a
n
d
s
h
a
p
e
s,
l
e
a
d
in
g
to
p
o
ten
ti
a
l
fi
n
a
n
c
ial
lo
ss
e
s
a
n
d
d
e
c
e
p
ti
v
e
tra
n
sa
c
ti
o
n
s.
T
o
a
d
d
re
ss
th
is
issu
e
,
we
d
e
v
e
lo
p
e
d
a
m
o
b
i
le
a
p
p
li
c
a
ti
o
n
t
h
a
t
e
m
p
l
o
y
s
M
o
b
il
e
Ne
t,
a
d
e
e
p
lea
rn
i
n
g
(DL)
tec
h
n
i
q
u
e
,
to
c
las
sify
Be
tt
a
fish
s
p
e
c
ies
.
Th
e
d
a
tas
e
t,
a
c
q
u
ired
fr
o
m
o
n
li
n
e
sto
re
s,
c
o
m
p
r
ise
s
4
0
0
ima
g
e
s,
with
1
0
0
ima
g
e
s
re
p
re
se
n
ti
n
g
e
a
c
h
o
f
th
e
fo
u
r
st
u
d
ied
Be
tt
a
fish
sp
e
c
ie
s:
c
o
m
b
tail,
d
e
lt
a
t
a
il
,
sp
a
d
e
tail
,
a
n
d
v
e
il
tail
.
P
r
io
r
to
m
o
d
e
l
imp
lem
e
n
tatio
n
,
t
h
e
d
a
tas
e
t
u
n
d
e
rg
o
e
s
p
re
-
p
ro
c
e
ss
in
g
with
d
a
ta
a
u
g
m
e
n
tati
o
n
tec
h
n
i
q
u
e
s,
in
c
l
u
d
i
n
g
r
o
tati
o
n
,
s
h
e
a
r,
z
o
o
m
-
in
,
h
o
rizo
n
tal
fl
ip
,
a
n
d
b
r
ig
h
tn
e
ss
a
d
ju
stm
e
n
ts,
e
n
h
a
n
c
i
n
g
th
e
m
o
d
e
l
p
e
rfo
rm
a
n
c
e
.
Train
in
g
u
ti
li
z
e
s
8
0
%
o
f
t
h
e
d
a
ta,
with
t
h
e
re
m
a
i
n
in
g
2
0
%
a
ll
o
c
a
ted
fo
r
tes
ti
n
g
.
T
h
re
e
d
isti
n
c
t
M
o
b
il
e
Ne
t
m
o
d
e
ls
a
re
d
e
v
e
lo
p
e
d
f
o
r
m
a
le
s
,
fe
m
a
le
s
,
a
n
d
b
o
t
h
g
e
n
d
e
rs
c
o
m
b
in
e
d
,
a
c
h
iev
i
n
g
a
c
c
u
ra
c
ies
o
f
7
0
,
8
3
.
7
5
,
a
n
d
6
5
%
,
re
sp
e
c
ti
v
e
ly
.
Th
e
se
train
e
d
m
o
d
e
ls
a
re
th
e
f
o
u
n
d
a
ti
o
n
f
o
r
a
m
o
b
il
e
a
p
p
li
c
a
ti
o
n
d
e
v
e
l
o
p
e
d
fo
r
th
e
An
d
r
o
i
d
p
latf
o
rm
th
a
t
e
n
a
b
les
u
se
rs,
p
a
rti
c
u
larly
Be
tt
a
fish
se
ll
e
rs
,
a
n
d
b
re
e
d
e
rs,
to
e
fficie
n
tl
y
c
las
sify
Be
tt
a
fish
sp
e
c
ies
,
e
m
p
o
we
rin
g
th
e
m
to
s
e
t
a
c
c
u
ra
te
p
rice
s
b
a
s
e
d
o
n
th
e
id
e
n
ti
fied
sp
e
c
ies
.
K
ey
w
o
r
d
s
:
B
etta
f
is
h
C
las
s
if
icatio
n
Dee
p
lear
n
in
g
L
ig
h
t
weig
h
t
Mo
b
ile
N
et
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
:
No
r
izan
Ma
t D
iah
Sch
o
o
l o
f
C
o
m
p
u
tin
g
Scien
ce
s
,
C
o
lleg
e
o
f
C
o
m
p
u
tin
g
,
I
n
f
o
r
m
atics a
n
d
Ma
th
em
atics
Un
iv
er
s
iti T
ek
n
o
lo
g
i M
AR
A
4
0
4
5
0
Sh
ah
Alam
,
Selan
g
o
r
,
Ma
lay
s
ia
E
m
ail:
n
o
r
izan
2
8
9
@
u
itm
.
ed
u
.
m
y
1.
I
NT
RO
D
UCT
I
O
N
I
n
r
ec
en
t
y
ea
r
s
,
th
e
h
o
b
b
y
o
f
co
llectin
g
an
d
b
r
ee
d
i
n
g
B
etta
f
is
h
h
as
ev
o
lv
ed
in
to
a
lu
cr
ativ
e
v
en
tu
r
e
f
o
r
m
an
y
en
t
h
u
s
iast
s
[
1
]
.
T
h
e
allu
r
e
o
f
B
et
ta
f
is
h
lies
in
th
eir
ca
p
tiv
atin
g
b
ea
u
ty
,
esp
ec
i
ally
ev
id
en
t
in
th
eir
u
n
iq
u
e
s
h
ap
es
an
d
d
is
tin
ctiv
e
tail
p
atter
n
s
.
Ho
wev
e
r
,
b
r
e
ed
er
s
o
f
ten
en
co
u
n
ter
ch
alle
n
g
es
in
ac
c
u
r
ately
class
if
y
in
g
d
if
f
er
en
t
B
etta
f
is
h
s
p
ec
ies,
s
u
ch
as
C
r
o
wn
tail
B
etta,
Vei
ltail
B
etta,
Half
Mo
o
n
B
etta,
an
d
Do
u
b
letail
B
etta
[
2
]
.
I
n
ad
d
iti
o
n
,
b
ec
au
s
e
th
e
m
ale
B
etta
f
i
s
h
h
as
m
o
r
e
attr
ac
tiv
e
a
n
d
c
o
lo
r
f
u
l
f
ea
tu
r
es
th
a
n
th
e
f
em
ales,
th
ey
ar
e
wo
r
th
m
o
r
e
ec
o
n
o
m
ically
.
T
h
is
class
if
icatio
n
is
s
u
e
h
as
p
r
o
m
p
ted
th
e
n
ee
d
to
im
p
lem
en
t
B
etta
f
is
h
s
p
ec
ies r
ec
o
g
n
itio
n
u
s
in
g
d
ee
p
lea
r
n
in
g
(
DL
)
tailo
r
ed
ex
p
licitly
f
o
r
m
o
b
ile
ap
p
licatio
n
s
.
T
h
e
p
er
v
asiv
e
p
r
esen
ce
o
f
m
o
b
ile
d
ev
ices
in
o
u
r
d
aily
liv
es
is
u
n
d
en
iab
le,
with
a
p
r
o
jecte
d
o
wn
er
s
h
ip
r
ate
o
f
o
v
e
r
9
0
%
am
o
n
g
ad
u
lts
in
d
ev
elo
p
ed
co
u
n
tr
ies
b
y
th
e
en
d
o
f
2
0
2
3
[
3
]
.
T
h
e
s
u
cc
ess
o
f
DL
in
v
ar
io
u
s
m
ac
h
in
e
lear
n
in
g
(
ML
)
task
s
h
as
f
u
elled
th
e
in
teg
r
atio
n
o
f
th
is
tech
n
o
lo
g
y
in
t
o
m
o
b
ile
ap
p
licatio
n
s
.
R
ec
o
g
n
i
z
in
g
th
is
tr
en
d
,
im
p
lem
en
tin
g
DL
f
o
r
B
etta
f
is
h
s
p
ec
ie
s
cla
s
s
i
f
icatio
n
in
m
o
b
ile
ap
p
licatio
n
s
b
ec
o
m
es e
s
s
en
tial.
B
etta
f
is
h
,
also
k
n
o
wn
as
Fig
h
tin
g
f
is
h
,
h
as
g
ain
ed
p
o
p
u
lar
i
ty
am
o
n
g
f
is
h
en
th
u
s
iast
s
d
u
e
to
its
ea
s
e
o
f
ca
r
e
an
d
v
ib
r
an
t
ae
s
th
etics.
T
h
e
h
o
b
b
y
o
f
co
llectin
g
th
es
e
f
is
h
h
as
tr
an
s
f
o
r
m
ed
in
to
a
p
r
o
f
itab
le
s
o
u
r
ce
o
f
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
B
etta
fis
h
s
p
ec
ies cla
s
s
if
ica
tio
n
u
s
in
g
lig
h
t weig
h
t
d
ee
p
…
(
Da
n
is
h
a
h
Ha
n
a
Mu
h
a
mma
d
Mu
h
a
imin
Lim
)
29
in
co
m
e,
with
in
d
iv
i
d
u
al
f
is
h
f
etch
in
g
h
ig
h
p
r
ices,
r
ea
c
h
in
g
u
p
to
R
M7
0
0
[
4
]
.
T
h
e
u
n
iq
u
e
co
lo
r
p
atter
n
s
a
n
d
s
h
ap
es
o
f
B
etta
f
is
h
co
n
tr
ib
u
te
to
t
h
eir
m
a
r
k
et
v
alu
e,
r
esu
ltin
g
in
an
in
cr
ea
s
in
g
d
em
a
n
d
a
n
d
s
u
b
s
eq
u
en
t
r
is
e
in
s
ellin
g
p
r
ices.
Ho
wev
e
r
,
t
h
e
v
ar
iab
ilit
y
in
tail
s
h
ap
es,
co
lo
r
s
,
an
d
p
atter
n
s
p
o
s
es
a
ch
all
en
g
e
f
o
r
s
eller
s
an
d
b
u
y
er
s
alik
e.
No
t
all
en
th
u
s
iast
s
ca
n
ac
cu
r
ately
r
ec
o
g
n
i
z
e
th
e
s
p
ec
ies
o
f
B
etta
f
is
h
,
le
ad
in
g
to
p
o
ten
tial
f
in
an
cial
lo
s
s
es
f
o
r
s
eller
s
an
d
th
e
r
is
k
o
f
b
u
y
e
r
s
p
ay
in
g
in
f
la
ted
p
r
ices f
o
r
m
is
id
en
tifie
d
f
is
h
[
5
]
.
T
h
is
r
esear
ch
f
o
c
u
s
es
o
n
d
e
v
e
lo
p
in
g
a
m
o
b
ile
a
p
p
licatio
n
f
o
r
B
etta
f
is
h
s
p
ec
ies
class
if
ic
atio
n
u
s
in
g
lig
h
tweig
h
t D
L
m
o
d
els to
ad
d
r
ess
th
ese
ch
allen
g
es.
T
h
e
o
b
j
ec
tiv
e
is
to
p
r
o
v
id
e
a
u
s
er
-
f
r
ien
d
ly
to
o
l to
id
en
tif
y
B
etta
f
is
h
s
p
ec
ie
s
ac
cu
r
ately
,
em
p
o
wer
in
g
s
eller
s
an
d
b
u
y
e
r
s
.
C
o
n
s
id
er
in
g
th
e
co
m
p
u
tatio
n
al
co
n
s
tr
ain
ts
o
f
m
o
b
ile
d
e
v
ices,
in
co
r
p
o
r
atin
g
lig
h
tweig
h
t
m
o
d
els
is
cr
u
cial,
an
d
it
aim
s
t
o
o
p
tim
i
z
e
th
e
class
if
icatio
n
p
r
o
ce
s
s
f
o
r
r
ea
l
-
tim
e
a
p
p
licatio
n
s
in
th
e
d
y
n
am
ic
B
etta
f
is
h
m
ar
k
et.
No
t
m
u
ch
r
esear
ch
h
as
b
ee
n
c
o
n
d
u
cte
d
o
n
B
etta
f
is
h
class
if
i
ca
tio
n
,
an
d
n
o
p
u
b
licly
av
ailab
le
d
ataset
ca
n
b
e
u
s
ed
f
o
r
co
m
p
ar
ativ
e
a
n
aly
s
is
.
T
h
e
ML
ap
p
r
o
ac
h
h
as
b
ee
n
ap
p
lied
t
o
id
en
tify
f
iv
e
s
p
ec
ies o
f
B
etta
f
is
h
b
y
ex
tr
ac
tin
g
g
r
ey
-
lev
el
co
-
o
cc
u
r
r
en
ce
m
atr
ix
(
GL
C
M)
f
e
atu
r
es
as
in
p
u
t
to
th
e
K
-
n
ea
r
est
n
eig
h
b
o
r
(K
-
NN)
class
if
ier
[
6
]
.
W
it
h
6
0
p
er
s
o
n
al
co
llectio
n
im
ag
es
p
er
s
p
ec
ies
f
o
r
th
e
d
ata
s
et,
wh
ich
to
tals
u
p
to
3
0
0
im
ag
es,
ex
ce
llen
t
class
if
icatio
n
h
as
b
ee
n
ac
h
iev
ed
;
h
o
we
v
er
,
th
e
f
is
h
h
as
to
b
e
in
a
s
p
ec
if
ic
an
g
u
lar
d
ir
ec
tio
n
.
Mo
o
k
d
ar
s
an
it
an
d
Mo
o
k
d
ar
s
an
it
[
7
]
co
n
d
u
cted
r
esear
ch
to
cr
ea
te
a
r
e
g
io
n
-
b
ased
c
o
n
v
o
lu
ti
o
n
al
n
eu
r
al
n
etwo
r
k
(R
-
C
NN)
m
o
d
el
n
a
m
ed
“Siam
Fis
h
Net”
th
at
class
if
ies
th
e
b
r
ee
d
o
f
an
u
n
k
n
o
wn
B
etta
f
is
h
im
ag
e
b
ased
s
o
lely
o
n
t
h
e
im
a
g
e
its
elf
.
T
h
e
r
esear
ch
er
s
f
o
r
m
u
late
d
th
is
m
o
d
el
u
s
in
g
a
d
ataset
o
f
8
7
,
5
6
0
B
etta
f
is
h
im
ag
es
r
ep
r
esen
tin
g
1
2
d
if
f
er
en
t
b
r
ee
d
s
o
f
B
ettas.
T
h
e
f
in
d
in
g
s
r
ev
ea
led
th
at
th
e
m
o
d
el
a
ch
iev
ed
an
a
v
er
ag
e
pr
ec
is
io
n
o
f
8
4
%,
in
d
icatin
g
i
ts
ef
f
ec
tiv
en
ess
in
ac
cu
r
ately
id
en
tify
in
g
t
h
e
b
r
ee
d
o
f
B
etta
f
is
h
.
An
o
th
er
ML
ap
p
r
o
ac
h
u
s
ed
th
e
G
ab
o
r
f
e
atu
r
e
a
n
d
ar
tific
ial
n
e
u
r
al
n
etwo
r
k
class
if
ier
,
b
u
t
th
e
r
esu
lts
w
er
e
n
o
t
en
c
o
u
r
a
g
in
g
[
8
]
.
DL
m
eth
o
d
s
s
u
ch
as
R
esNet
-
5
0
h
a
v
e
b
ee
n
u
tili
z
ed
f
o
r
B
etta
f
is
h
class
if
icatio
n
b
ased
o
n
p
er
s
o
n
al
d
ata
co
llectio
n
an
d
ac
h
iev
ed
8
0
%
ac
cu
r
ac
y
[
9
]
.
R
esNet
-
5
0
is
a
h
ea
v
y
weig
h
t
DL
th
at
co
n
s
is
ts
o
f
4
8
co
n
v
o
lu
tio
n
al
lay
er
s
,
o
n
e
Ma
x
Po
o
l
lay
er
,
a
n
d
o
n
e
av
er
a
g
e
p
o
o
l
lay
er
.
Ho
wev
er
,
a
h
ea
v
y
weig
h
t
DL
u
s
u
ally
n
ee
d
s
h
ig
h
s
to
r
ag
e
an
d
h
ig
h
-
p
o
wer
d
ev
ic
es
[
1
0
]
,
wh
ic
h
m
a
y
b
e
a
b
a
r
r
ie
r
f
o
r
u
s
er
s
,
esp
ec
ially
s
m
all
p
e
t
s
h
o
p
s
.
T
h
er
ef
o
r
e,
th
is
r
esear
ch
p
r
o
p
o
s
es
to
u
s
e
a
lig
h
tweig
h
t
DL
m
o
d
el
th
at
ca
n
ac
h
iev
e
h
ig
h
ac
c
u
r
ac
y
a
t
m
in
im
al
co
s
t
an
d
m
em
o
r
y
r
eq
u
ir
e
m
en
ts
wh
ile
s
till
b
ein
g
co
m
p
etitiv
e
with
h
ea
v
y
weig
h
t m
o
d
els.
T
h
is
p
ap
er
is
o
r
g
an
i
z
e
d
as
f
o
llo
ws.
T
h
e
n
ex
t
s
ec
tio
n
d
is
cu
s
s
es
th
e
wo
r
k
s
r
elate
d
to
DL
.
Sectio
n
3
ex
p
lain
s
th
e
class
if
icatio
n
m
eth
o
d
u
tili
z
ed
in
th
is
r
esear
ch
,
th
e
d
ataset
u
s
ed
,
an
d
th
e
v
a
r
io
u
s
ex
p
er
im
en
tal
r
esu
lts
b
ased
o
n
f
in
e
-
tu
n
in
g
v
ar
io
u
s
h
y
p
er
p
a
r
am
eter
s
.
Secti
o
n
4
d
is
cu
s
s
es
th
e
r
esu
lt
a
n
al
y
s
is
,
f
o
llo
wed
b
y
a
co
n
clu
s
io
n
in
t
h
e
last
s
ec
tio
n
.
2.
B
ACK
G
RO
UND
S
T
UD
Y
Op
er
atin
g
DL
m
o
d
els
o
n
ed
g
e
d
ev
ices
p
o
s
es
s
ig
n
if
ican
t
ch
allen
g
es
d
u
e
to
lim
ited
r
es
o
u
r
ce
s
an
d
co
m
p
u
tatio
n
al
ca
p
ab
ilit
ies.
Usi
n
g
lig
h
tweig
h
t
DL
m
o
d
els
h
as
em
er
g
ed
as
a
c
r
u
cial
s
tr
ateg
y
to
a
d
d
r
ess
th
is
.
L
ig
h
tweig
h
t
alg
o
r
ith
m
s
ar
e
d
e
s
ig
n
ed
to
b
e
c
o
m
p
u
tatio
n
ally
ef
f
icien
t
with
a
s
m
all
m
em
o
r
y
f
o
o
t
p
r
in
t,
m
ak
in
g
th
em
s
u
itab
le
f
o
r
d
e
p
lo
y
m
e
n
t
o
n
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
d
ev
i
ce
s
lik
e
m
o
b
ile
p
h
o
n
es,
in
ter
n
et
o
f
th
in
g
(
I
o
T
)
d
ev
ices,
an
d
e
d
g
e
co
m
p
u
tin
g
p
latf
o
r
m
s
[
1
1
]
,
[
1
2
]
.
T
h
ese
lig
h
tweig
h
t
m
o
d
els
aim
to
r
ed
u
ce
co
m
p
u
tatio
n
al
d
e
m
an
d
s
b
y
o
p
tim
i
z
in
g
n
etwo
r
k
s
tr
u
ctu
r
es
an
d
em
p
lo
y
in
g
ef
f
icien
t
b
u
ild
in
g
m
eth
o
d
s
.
T
h
e
co
n
ce
p
t
o
f
lig
h
tweig
h
t
alg
o
r
ith
m
s
m
in
im
i
z
es
th
e
n
u
m
b
er
o
f
p
ar
am
eter
s
an
d
co
m
p
u
tatio
n
s
,
m
ak
in
g
t
h
em
id
ea
l
f
o
r
r
ea
l
-
tim
e
p
r
o
ce
s
s
in
g
o
n
d
e
v
ices
with
lim
ited
co
m
p
u
tatio
n
al
r
eso
u
r
ce
s
[
1
3
]
.
Ad
d
itio
n
ally
,
th
ese
m
o
d
els
ty
p
ically
h
av
e
s
m
aller
m
e
m
o
r
y
f
o
o
tp
r
i
n
ts
,
an
ad
v
an
tag
e
o
u
s
f
ea
tu
r
e
f
o
r
d
ev
i
ce
s
with
lim
ited
r
an
d
o
m
-
a
cc
ess
m
em
o
r
y
(
R
AM
)
.
Sev
er
al
n
o
tab
le
lig
h
tweig
h
t
c
o
n
v
o
lu
ti
o
n
al
n
eu
r
al
n
etwo
r
k
(
C
NN)
m
o
d
els
h
a
v
e
b
ee
n
p
r
o
p
o
s
ed
t
o
ad
d
r
ess
th
ese
ch
allen
g
es.
Sq
u
ee
ze
Net,
in
tr
o
d
u
ce
d
b
y
th
e
B
er
k
eley
a
n
d
Stan
f
o
r
d
r
esear
ch
team
s
in
2
0
1
6
,
an
d
Mo
b
ileNet,
p
r
esen
ted
b
y
th
e
Go
o
g
le
team
in
2
0
1
7
,
ar
e
n
o
te
wo
r
th
y
ex
a
m
p
les.
Sh
u
f
f
leNe
t,
p
r
o
p
o
s
ed
b
y
I
g
n
o
r
e
T
ec
h
n
o
lo
g
y
in
2
0
1
7
,
an
d
E
f
f
icien
tNet,
in
tr
o
d
u
ce
d
b
y
Go
o
g
le
in
2
0
1
8
,
f
u
r
th
er
co
n
tr
ib
u
t
e
to
th
e
ar
s
en
al
o
f
lig
h
tweig
h
t
m
o
d
els.
T
h
ese
m
o
d
els
o
p
tim
i
z
e
n
etwo
r
k
s
tr
u
c
tu
r
es,
d
ec
r
ea
s
e
th
e
n
u
m
b
er
o
f
p
ar
am
eter
s
,
an
d
en
h
an
ce
ac
c
u
r
ac
y
,
ev
en
ac
h
i
ev
in
g
f
u
ll
co
n
v
o
lu
tio
n
ac
cu
r
ac
y
[
1
4
]
–
[
1
6
]
.
Sh
u
f
f
leNe
t
u
t
ili
z
es
b
o
th
g
r
o
u
p
co
n
v
o
l
u
tio
n
a
n
d
c
h
an
n
el
s
h
u
f
f
le
o
p
er
atio
n
s
to
s
im
p
lify
p
o
i
n
twis
e
co
n
v
o
lu
tio
n
s
.
Gr
o
u
p
c
o
n
v
o
lu
ti
o
n
d
iv
id
es
th
e
in
p
u
t
c
h
an
n
els
in
t
o
g
r
o
u
p
s
,
r
ed
u
cin
g
co
m
p
u
tatio
n
al
co
m
p
lex
ity
.
C
h
an
n
el
s
h
u
f
f
le
is
th
en
ap
p
lied
to
ex
ch
an
g
e
in
f
o
r
m
atio
n
b
etwe
e
n
g
r
o
u
p
s
,
p
r
o
m
o
tin
g
i
n
f
o
r
m
atio
n
f
lo
w
a
n
d
m
ain
tain
in
g
m
o
d
el
ef
f
icien
cy
[
1
7
]
,
[
1
8
]
u
s
in
g
g
r
o
u
p
co
n
v
o
lu
tio
n
an
d
ch
an
n
el
s
h
u
f
f
le
to
s
im
p
lify
p
o
in
twis
e
co
n
v
o
lu
tio
n
.
Mic
r
o
Net’
s
m
icr
o
-
f
ac
to
r
i
z
ed
co
n
v
o
l
u
tio
n
an
d
ad
ju
s
tin
g
n
o
d
e
c
o
n
n
ec
tiv
ity
an
d
n
etwo
r
k
wid
th
aim
to
b
ala
n
c
e
m
o
d
el
ef
f
icien
cy
an
d
ex
p
r
ess
iv
e
p
o
wer
[
1
9
]
.
Mo
b
ileNet
s
tan
d
s
o
u
t
as
a
p
a
r
ticu
lar
ly
lig
h
tweig
h
t
d
ee
p
C
NN.
I
t
is
s
m
aller
an
d
f
aster
t
h
an
m
an
y
well
-
k
n
o
wn
class
if
icatio
n
m
o
d
els,
m
ak
in
g
it
s
u
itab
le
f
o
r
im
ag
e
d
etec
tio
n
,
f
ac
e
attr
ib
u
tes,
an
d
im
a
g
e
an
al
y
s
is
[
2
0
]
.
M
o
b
ileNet
u
tili
z
es
a
s
i
m
p
lifie
d
ar
ch
itectu
r
e
with
d
ep
th
-
wis
e
s
ep
ar
ab
le
co
n
v
o
lu
t
io
n
s
,
p
r
o
v
id
in
g
a
n
ef
f
icien
t
s
o
lu
tio
n
f
o
r
b
o
th
m
o
b
ile
an
d
em
b
ed
d
ed
d
e
v
ices
[
2
1
]
,
[
2
2
]
.
T
h
e
a
d
v
an
tag
es
o
f
Mo
b
ileNet
lie
in
its
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
4
,
No
.
1
,
Ma
r
c
h
2
0
2
5
:
28
-
38
30
r
ea
l
-
tim
e
p
er
f
o
r
m
a
n
ce
o
n
t
ask
s
lik
e
im
ag
e
class
if
icatio
n
,
o
b
ject
d
etec
tio
n
,
an
d
s
eg
m
en
tatio
n
.
T
h
e
ar
ch
itectu
r
e’
s
f
lex
ib
ilit
y
allo
ws
u
s
er
s
to
co
n
tr
o
l
th
e
tr
a
d
e
-
o
f
f
b
etwe
en
m
o
d
el
ac
cu
r
ac
y
an
d
co
m
p
u
tatio
n
al
ef
f
icien
cy
th
r
o
u
g
h
h
y
p
er
p
ar
a
m
eter
s
lik
e
th
e
wid
th
an
d
r
eso
lu
tio
n
m
u
ltip
lier
s
.
T
h
is
a
d
ap
tab
ilit
y
p
r
o
v
es
in
v
alu
ab
le
w
h
en
o
p
tim
i
z
in
g
m
o
d
els
f
o
r
s
p
ec
if
ic
d
ep
lo
y
m
en
t
s
ce
n
ar
io
s
,
o
f
f
er
in
g
lo
w
-
laten
cy
r
esp
o
n
s
es
f
o
r
v
ar
io
u
s
ap
p
licati
o
n
s
[
2
0
]
.
I
n
th
e
co
n
tex
t
o
f
t
h
is
r
esea
r
ch
,
th
e
f
o
cu
s
is
o
n
c
o
n
s
tr
u
ctin
g
a
n
etwo
r
k
b
ased
o
n
Mo
b
ileNet
f
o
r
B
etta
f
is
h
s
p
ec
ies
class
if
icatio
n
.
Mo
b
ileNet’
s
m
ain
co
n
tr
ib
u
tio
n
is
th
e
p
r
o
p
o
s
al
o
f
d
ee
p
s
ep
a
r
ab
le
co
n
v
o
lu
tio
n
,
a
d
ec
o
m
p
o
s
itio
n
f
o
r
m
s
ig
n
if
ica
n
tly
r
ed
u
ci
n
g
co
m
p
u
tatio
n
al
c
o
m
p
lex
ity
a
n
d
m
o
d
e
l
s
ize.
T
h
is
r
esear
ch
aim
s
to
lev
er
ag
e
Mo
b
ileNet’
s
lig
h
tweig
h
t
d
esig
n
to
cr
ea
te
an
ef
f
icien
t
B
etta
f
i
s
h
s
p
ec
ies
cla
s
s
if
icatio
n
s
y
s
tem
,
ca
ter
in
g
to
th
e
u
n
iq
u
e
d
em
an
d
s
o
f
ed
g
e
d
ev
ices
an
d
c
o
n
tr
ib
u
tin
g
to
r
ea
l
-
tim
e
ap
p
licati
o
n
s
in
th
e
f
ield
o
f
aq
u
atic
s
p
ec
ies
id
en
tific
atio
n
.
3.
M
E
T
H
O
D
AND
M
AT
E
R
I
A
L
T
h
is
co
n
ce
p
tu
al
f
r
am
ewo
r
k
o
u
tlin
es
th
e
s
y
s
tem
atic
p
r
o
ce
d
u
r
es
em
p
lo
y
ed
in
th
e
r
esear
c
h
,
d
elin
ea
tin
g
k
ey
p
h
ases
en
co
m
p
ass
in
g
d
ataset
co
llectio
n
,
p
r
e
-
p
r
o
ce
s
s
in
g
,
m
o
d
el
a
r
ch
itectu
r
e
d
esig
n
,
tr
ain
in
g
,
ev
alu
atio
n
,
an
d
e
x
p
er
im
e
n
tatio
n
.
I
t
s
er
v
es
as
a
s
tr
u
ctu
r
e
d
g
u
id
e,
o
f
f
er
in
g
i
n
s
ig
h
t
in
to
th
e
an
ticip
ated
m
eth
o
d
s
an
d
m
ater
ials
d
ep
lo
y
ed
i
n
th
e
s
tu
d
y
.
3
.
1
.
Da
t
a
c
o
llect
io
n
T
h
is
s
tu
d
y
co
llected
3
0
0
im
ag
es
o
f
s
ev
e
n
ty
p
es
o
f
B
etta
f
is
h
f
r
o
m
B
etta
f
is
h
s
eller
s
th
at
d
o
E
-
co
m
m
er
ce
in
L
az
ad
a,
I
n
s
tag
r
am
,
an
d
Face
b
o
o
k
.
T
h
ese
d
at
a
ar
e
all
in
.
jp
g
f
o
r
m
at,
s
m
alle
r
th
an
th
e
.
p
n
g
f
ile.
T
ab
le
1
lis
ts
th
e
co
llected
d
ata
an
d
to
tal
im
ag
es
f
o
r
e
v
er
y
B
e
tta
f
is
h
s
p
ec
ies.
T
h
e
n
u
m
b
e
r
o
f
im
ag
es
f
o
r
co
m
b
tail,
d
elta
tail,
an
d
d
o
u
b
le
tail
is
3
4
,
r
esp
ec
tiv
ely
.
Fo
r
ty
-
s
ix
im
ag
es
wer
e
co
llected
ea
ch
f
o
r
s
p
ad
e
tail
an
d
v
eil
tail.
C
r
o
wn
tail
an
d
h
alf
m
o
o
n
tail
h
av
e
6
4
an
d
5
8
im
ag
es,
r
esp
ec
tiv
ely
.
Ov
er
all,
3
1
6
im
ag
es
wer
e
o
b
tain
ed
f
r
o
m
o
n
lin
e
s
to
r
es.
T
h
e
d
ata
w
as th
en
au
g
m
e
n
ted
,
an
d
th
e
im
a
g
es we
r
e
r
esized
as p
ar
t o
f
th
e
p
r
e
-
p
r
o
ce
s
s
in
g
.
T
ab
le
1
.
T
h
e
to
tal
n
u
m
b
er
o
f
i
m
ag
es a
cq
u
ir
ed
f
o
r
ea
ch
s
p
ec
i
es
B
e
t
t
a
F
i
s
h
S
p
e
c
i
e
s
N
u
mb
e
r
o
f
I
mag
e
s
C
o
m
b
Ta
i
l
34
C
r
o
w
n
T
a
i
l
64
D
e
l
t
a
T
a
i
l
34
D
o
u
b
l
e
T
a
i
l
34
H
a
l
f
mo
o
n
T
a
i
l
58
S
p
a
d
e
Ta
i
l
46
V
e
i
l
Ta
i
l
46
To
t
a
l
3
1
6
3
.
2
.
Da
t
a
pre
-
pro
ce
s
s
ing
Data
p
r
e
-
p
r
o
ce
s
s
in
g
is
th
e
m
o
s
t
s
ig
n
if
ican
t
an
d
in
f
lu
en
tial
f
ac
to
r
in
th
e
g
en
er
ali
z
atio
n
p
er
f
o
r
m
an
ce
o
f
a
s
u
p
er
v
is
ed
ML
alg
o
r
it
h
m
[
2
3
]
.
Af
ter
th
e
d
ataset
was
co
llected
,
all
th
e
im
ag
es
wer
e
r
esized
to
224
×
2
4
4
p
ix
els
f
o
r
f
itti
n
g
in
to
Mo
b
ileNet.
T
h
en
,
as
s
h
o
wn
i
n
Fig
u
r
e
1
,
th
e
d
ataset
was
au
g
m
en
ted
d
u
e
to
th
e
s
m
all
am
o
u
n
t
o
f
d
ata
b
y
u
s
in
g
Fig
u
r
e
1
(
a)
a
r
o
tatio
n
r
an
g
e
o
f
0
.
2
,
Fig
u
r
e
1
(
b
)
a
s
h
ea
r
r
an
g
e
o
f
0
.
2
,
Fig
u
r
e
1
(
c)
a
zo
o
m
-
i
n
r
an
g
e
o
f
0
.
2
,
Fig
u
r
e
1
(
d
)
a
h
o
r
izo
n
tal
f
lip
is
eq
u
al
to
tr
u
e,
an
d
Fig
u
r
e
1
(
e)
a
b
r
ig
h
tn
ess
r
an
g
e
o
f
0
.
5
to
1
.
5
.
(a
)
(b
)
(c
)
(
d)
(e
)
Fig
u
r
e
1
.
Au
g
m
en
tatio
n
co
d
es
an
d
s
am
p
le
r
esu
lts
o
f
(
a)
r
o
tatio
n
r
an
g
e
,
(
b
)
s
h
ea
r
r
an
g
e
,
(
c)
zo
o
m
-
in
r
an
g
e
,
(
d
)
h
o
r
izo
n
tal
f
lip
is
eq
u
al
,
an
d
(
e)
b
r
ig
h
tn
ess
r
an
g
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
B
etta
fis
h
s
p
ec
ies cla
s
s
if
ica
tio
n
u
s
in
g
lig
h
t weig
h
t
d
ee
p
…
(
Da
n
is
h
a
h
Ha
n
a
Mu
h
a
mma
d
Mu
h
a
imin
Lim
)
31
I
t
is
cr
u
cial
to
au
g
m
en
t
im
ag
e
d
ata
p
r
o
p
er
l
y
to
in
cr
ea
s
e
ac
cu
r
ac
y
an
d
p
r
ev
e
n
t
o
v
er
f
itti
n
g
[
2
4
]
.
T
h
e
p
ar
am
eter
s
o
f
au
g
m
e
n
tatio
n
w
er
e
ad
ju
s
ted
f
o
u
r
ti
m
es
to
g
ain
m
o
r
e
tr
ain
in
g
d
atasets
.
Af
te
r
au
g
m
en
tatio
n
,
th
e
to
tal
n
u
m
b
er
o
f
B
etta
f
is
h
im
ag
es
f
o
r
ev
er
y
s
p
ec
ies
h
as
in
cr
ea
s
ed
s
ig
n
if
ican
tly
.
B
ef
o
r
e
th
e
d
ata
au
g
m
en
tatio
n
p
r
o
ce
s
s
,
th
e
to
tal
n
u
m
b
er
o
f
i
m
ag
es
f
o
r
ev
e
r
y
s
p
ec
ies
was
m
o
s
tly
lo
wer
t
h
an
1
0
0
.
B
ased
o
n
T
a
b
le
2
,
a
to
tal
o
f
7
0
0
im
a
g
es
wer
e
cr
ea
ted
f
r
o
m
th
e
o
r
i
g
in
al
3
1
6
im
a
g
es,
wh
er
e
ea
ch
o
f
th
e
s
ev
e
n
class
es
co
n
tain
s
1
0
0
im
ag
es.
Fro
m
ea
ch
o
f
t
h
e
s
ev
en
class
es,
8
0
im
ag
es
(
8
0
%)
wer
e
u
s
ed
f
o
r
tr
ain
in
g
,
an
d
2
0
im
a
g
es
(
2
0
%)
wer
e
u
s
ed
f
o
r
test
in
g
.
Fig
u
r
e
2
s
h
o
ws a
s
n
ip
p
et
o
f
th
e
B
etta
f
is
h
d
ataset
th
at
was d
iv
id
ed
ac
co
r
d
in
g
ly
.
T
ab
le
2
.
T
h
e
to
tal
n
u
m
b
er
o
f
B
etta
f
is
h
im
ag
es b
ef
o
r
e
an
d
a
f
ter
au
g
m
e
n
tatio
n
B
e
t
t
a
F
i
s
h
S
p
e
c
i
e
s
To
t
a
l
N
u
m
b
e
r
o
f
I
mag
e
s
B
e
f
o
r
e
D
a
t
a
A
u
g
me
n
t
a
t
i
o
n
A
f
t
e
r
D
a
t
a
A
u
g
m
e
n
t
a
t
i
o
n
C
o
m
b
Ta
i
l
34
2
4
0
C
r
o
w
n
T
a
i
l
64
2
4
0
D
e
l
t
a
T
a
i
l
34
2
4
0
D
o
u
b
l
e
T
a
i
l
34
2
4
0
H
a
l
f
mo
o
n
T
a
i
l
58
2
4
0
S
p
a
d
e
Ta
i
l
46
2
4
0
V
e
i
l
Ta
i
l
46
2
4
0
To
t
a
l
3
1
6
1
6
8
0
Fig
u
r
e
2
.
T
h
e
s
n
ip
p
et
o
f
th
e
B
etta
f
is
h
d
ataset
3
.
3
.
M
o
del
a
rc
hite
ct
ure
T
h
is
r
esear
ch
u
s
ed
Mo
b
ileNet
to
class
if
y
B
etta
f
i
s
h
s
p
ec
ies
b
y
s
p
ec
if
ic
ar
ch
itectu
r
al
co
n
f
i
g
u
r
atio
n
s
.
T
h
e
p
r
im
a
r
y
p
u
r
p
o
s
e
o
f
c
h
o
o
s
in
g
Mo
b
ileNet
is
to
ad
d
r
e
s
s
th
e
ch
allen
g
es
ass
o
ciate
d
with
d
ep
lo
y
in
g
DL
m
o
d
els
o
n
r
eso
u
r
ce
-
c
o
n
s
tr
ain
ed
d
ev
ices,
s
u
ch
as
m
o
b
ile
p
h
o
n
es
wh
ile
ac
h
iev
in
g
ac
cu
r
ate
an
d
ef
f
icien
t
B
etta
f
is
h
class
if
icat
io
n
.
Mo
b
ileNet’
s
ar
ch
itectu
r
e
is
ch
ar
ac
ter
i
z
ed
b
y
its
lig
h
tweig
h
t
d
esig
n
,
m
ak
in
g
it
well
-
s
u
ited
f
o
r
r
ea
l
-
tim
e
im
a
g
e
p
r
o
ce
s
s
in
g
o
n
d
e
v
ices w
ith
lim
ited
co
m
p
u
tatio
n
al
ca
p
ab
ilit
ies.
T
h
e
cr
it
ical
in
n
o
v
atio
n
i
n
Mo
b
ileNet
is
th
e
u
s
e
o
f
d
e
p
th
wis
e
s
ep
ar
ab
le
co
n
v
o
lu
ti
o
n
s
,
wh
ich
s
ig
n
if
ican
tly
r
ed
u
ce
s
th
e
n
u
m
b
er
o
f
p
a
r
am
eter
s
an
d
co
m
p
u
tatio
n
s
co
m
p
ar
ed
to
tr
a
d
itio
n
al
co
n
v
o
lu
tio
n
al
lay
er
s
[
2
5
]
–
[
2
7
]
.
T
h
is
d
esig
n
en
ab
les
Mo
b
ileNet
to
m
ain
tain
s
a
tis
f
ac
to
r
y
ac
cu
r
ac
y
wh
ile
s
ig
n
if
ican
tly
lo
wer
in
g
th
e
m
o
d
el’
s
s
ize,
m
ak
in
g
it
p
r
ac
tical
f
o
r
d
ep
lo
y
m
en
t
o
n
m
o
b
ile
p
latf
o
r
m
s
.
An
o
v
e
r
v
iew
o
f
th
e
Mo
b
ileNet
ar
ch
itectu
r
e
is
illu
s
tr
ated
in
Fig
u
r
e
3
.
I
t
co
n
s
is
ts
o
f
2
8
lay
er
s
,
i
n
clu
d
in
g
a
d
ee
p
co
n
v
o
lu
tio
n
lay
e
r
,
1
×1
p
o
i
n
t c
o
n
v
o
lu
tio
n
lay
er
,
b
atch
n
o
r
m
,
R
eL
U,
av
er
ag
e
co
l
lectin
g
lay
er
,
an
d
So
f
tMa
x
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
4
,
No
.
1
,
Ma
r
c
h
2
0
2
5
:
28
-
38
32
Fig
u
r
e
3
Mo
b
ileNet
a
r
ch
itectu
r
e
[
2
8
]
,
[
2
9
]
3
.
4
.
E
x
peri
m
ent
a
l
a
pp
ro
a
ch
a
nd
o
ptim
i
z
a
t
io
n
f
o
r
M
o
bil
eNe
t
t
ra
in
ing
Sev
er
al
ex
p
er
im
en
ts
h
a
v
e
b
ee
n
m
eticu
lo
u
s
ly
co
n
d
u
cted
to
cr
ea
te
a
r
eliab
le
an
d
ac
c
u
r
ate
m
o
d
el
f
o
r
B
etta
f
is
h
s
p
ec
ies
clas
s
if
icati
o
n
u
s
in
g
Mo
b
ileNet.
T
h
ese
ex
p
er
im
en
ts
aim
ed
to
en
h
an
c
e
th
e
class
if
icatio
n
ac
cu
r
ac
y
th
r
o
u
g
h
two
p
r
im
ar
y
ap
p
r
o
ac
h
es:
in
cr
ea
s
in
g
th
e
n
u
m
b
er
o
f
tr
ain
in
g
im
a
g
es
an
d
f
in
e
-
tu
n
in
g
f
o
u
r
k
e
y
h
y
p
er
-
p
ar
am
eter
s
,
n
am
el
y
ep
o
ch
,
d
r
o
p
o
u
t
lay
er
,
p
o
o
lin
g
lay
er
,
an
d
b
atch
s
ize.
T
h
e
cu
lm
in
atio
n
o
f
th
ese
ef
f
o
r
ts
in
v
o
lv
ed
eig
h
t
d
is
tin
ct
ex
p
er
im
e
n
ts
,
ea
ch
co
n
tr
i
b
u
tin
g
v
alu
ab
le
in
s
ig
h
ts
to
th
e
o
v
e
r
all
m
o
d
el
p
er
f
o
r
m
an
ce
.
3
.
4
.
1
.
E
x
perim
ent
1
:
c
o
m
pa
r
is
o
n bet
wee
n t
wo
s
et
s
o
f
da
t
a
s
et
s
I
n
th
is
e
x
p
er
im
e
n
t,
th
e
r
e
ar
e
t
wo
s
ets
o
f
d
atasets
.
I
n
t
h
e
f
i
r
s
t
s
et,
th
e
test
in
g
im
ag
es
co
n
s
is
t
o
f
o
n
ly
th
e
o
r
ig
in
al
im
a
g
es,
wh
ile
in
th
e
s
ec
o
n
d
s
et,
th
e
test
in
g
im
ag
es
co
n
s
is
t
o
f
a
co
m
b
in
atio
n
o
f
th
e
o
r
ig
in
al
a
n
d
au
g
m
en
ted
im
ag
es.
T
h
e
ex
p
er
im
en
t
in
v
o
lv
ed
s
ev
en
class
es
r
ep
r
esen
tin
g
d
if
f
e
r
en
t
B
ett
a
f
is
h
s
p
ec
ies
an
d
7
0
0
im
ag
es.
T
h
e
d
iv
is
io
n
b
et
wee
n
tr
ain
in
g
an
d
test
in
g
d
ata
s
e
ts
wa
s
ex
ec
u
ted
with
8
0
%
f
o
r
tr
ain
in
g
an
d
2
0
%
f
o
r
test
in
g
,
ad
h
er
i
n
g
to
DL
m
o
d
el
d
ev
elo
p
m
en
t
s
tan
d
ar
d
s
[
3
0
]
.
A
h
y
p
er
p
ar
a
m
eter
ca
lled
“e
p
o
ch
s
”
d
eter
m
in
es
h
o
w
m
an
y
tim
es
th
e
lear
n
in
g
alg
o
r
ith
m
will
r
u
n
o
v
er
t
h
e
tr
a
in
in
g
d
ataset
[
3
1
]
,
[
3
2
]
.
T
h
is
ex
p
er
im
en
t
was
r
u
n
f
o
r
8
0
ep
o
ch
s
,
an
d
it
to
o
k
tw
o
h
o
u
r
s
f
o
r
ea
ch
m
o
d
el
t
o
lea
r
n
.
T
ab
le
3
illu
s
tr
ates
t
h
e
r
esu
lts
wh
ich
s
h
o
w
th
at
th
e
m
o
d
el
with
co
m
b
in
ed
im
ag
es
(
v
alid
atio
n
ac
cu
r
ac
y
o
f
0
.
3
3
5
7
)
p
er
f
o
r
m
s
b
etter
th
an
th
e
o
n
e
with
th
e
o
r
ig
in
al
im
a
g
es
(
v
alid
atio
n
ac
cu
r
ac
y
o
f
0
.
2
7
8
6
)
.
Ho
wev
er
,
an
o
v
er
f
itti
n
g
p
r
o
b
lem
o
cc
u
r
s
wh
er
e
th
e
v
alu
e
o
f
th
e
v
alid
atio
n
ac
cu
r
a
cy
is
v
er
y
m
u
c
h
l
o
wer
th
a
n
th
e
tr
ain
in
g
ac
cu
r
ac
y
.
T
h
er
e
f
o
r
e,
in
th
e
n
ex
t
ex
p
er
im
en
t,
a
d
r
o
p
o
u
t la
y
er
is
ad
d
ed
.
T
ab
le
3
.
E
x
p
er
im
en
t
1
tr
ain
in
g
r
esu
lts
o
n
th
e
o
r
ig
in
al
im
a
g
e
d
ataset
an
d
co
m
b
in
ed
d
ataset
D
a
t
a
s
e
t
Tr
a
i
n
i
n
g
A
c
c
u
r
a
c
y
Tr
a
i
n
i
n
g
L
o
ss
V
a
l
i
d
a
t
i
o
n
A
c
c
u
r
a
c
y
V
a
l
i
d
a
t
i
o
n
L
o
ss
O
r
i
g
i
n
a
l
I
mag
e
s
0
.
9
0
3
6
0
.
3
2
9
8
0
.
2
7
8
6
4
.
5
0
2
8
C
o
m
b
i
n
e
d
I
mag
e
s (O
r
i
g
i
n
a
l
a
n
d
A
u
g
men
t
e
d
)
0
.
9
1
2
5
0
.
3
0
0
7
0
.
3
3
5
7
6
.
4
9
4
4
3
.
4
.
2
.
E
x
perim
ent
2
:
c
o
m
pa
r
is
o
n bet
wee
n u
s
ing
a
dro
po
u
t
la
y
er
a
nd
wit
ho
ut
a
dro
po
u
t
la
y
er
Dr
o
p
o
u
t
is
an
ef
f
icien
t
way
t
o
r
ed
u
ce
o
v
er
f
itti
n
g
[
3
3
]
.
I
t
r
an
d
o
m
ly
s
ets
in
p
u
t
u
n
its
to
0
with
a
p
r
e
-
d
eter
m
in
ed
p
e
r
ce
n
tag
e
at
ea
ch
s
tep
d
u
r
in
g
tr
ain
i
n
g
tim
e
[
3
4
]
.
T
h
e
d
ataset
f
o
r
th
is
ex
p
er
im
e
n
t is th
e
s
am
e
as in
E
x
p
er
im
en
t
1
.
T
h
is
ex
p
er
im
e
n
t
was
r
u
n
f
o
r
4
0
ep
o
ch
s
,
an
d
it
to
o
k
o
n
e
h
o
u
r
f
o
r
ea
ch
m
o
d
el
to
lear
n
.
T
ab
le
4
illu
s
tr
ates
th
e
o
u
tco
m
e
o
f
th
e
ex
p
er
im
e
n
t
an
d
it
in
d
icate
s
th
at
th
e
m
o
d
el
with
co
m
b
in
ed
im
ag
es
(
v
alid
atio
n
ac
cu
r
ac
y
o
f
0
.
3
6
4
3
)
s
h
o
wed
a
b
etter
r
esu
lt
th
an
t
h
e
m
o
d
el
with
o
r
i
g
in
al
im
ag
es
(
v
al
id
atio
n
ac
cu
r
ac
y
o
f
0
.
2
3
5
7
)
wh
e
n
a
d
r
o
p
o
u
t
lay
e
r
was
ad
d
ed
.
Fu
r
th
er
m
o
r
e,
th
e
p
er
f
o
r
m
an
ce
h
as
s
lig
h
tly
im
p
r
o
v
ed
c
o
m
p
ar
e
d
to
E
x
p
er
im
en
t
1
.
Ho
wev
er
,
o
v
er
f
itti
n
g
s
till
o
cc
u
r
s
.
Hen
ce
,
th
e
to
tal
n
u
m
b
er
o
f
im
ag
es
is
ad
d
ed
with
d
if
f
er
e
n
t
p
o
o
lin
g
lay
er
s
in
E
x
p
er
im
en
t
3
.
T
ab
le
4
.
E
x
p
er
im
en
t
2
tr
ain
in
g
r
esu
lts
o
n
th
e
o
r
ig
in
al
im
a
g
e
d
ataset
an
d
co
m
b
in
ed
d
ataset
with
a
d
r
o
p
o
u
t la
y
er
D
a
t
a
s
e
t
Tr
a
i
n
i
n
g
A
c
c
u
r
a
c
y
Tr
a
i
n
i
n
g
L
o
ss
V
a
l
i
d
a
t
i
o
n
A
c
c
u
r
a
c
y
V
a
l
i
d
a
t
i
o
n
L
o
ss
O
r
i
g
i
n
a
l
I
mag
e
s
0
.
6
4
6
4
1
.
0
4
0
4
0
.
2
3
5
7
4
.
5
4
6
6
C
o
m
b
i
n
e
d
I
mag
e
s
(
O
r
i
g
i
n
a
l
a
n
d
A
u
g
m
e
n
t
e
d
)
0
.
5
8
7
5
1
.
2
1
6
9
0
.
3
6
4
3
3
.
8
3
9
3
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
B
etta
fis
h
s
p
ec
ies cla
s
s
if
ica
tio
n
u
s
in
g
lig
h
t weig
h
t
d
ee
p
…
(
Da
n
is
h
a
h
Ha
n
a
Mu
h
a
mma
d
Mu
h
a
imin
Lim
)
33
3
.
4
.
3
.
E
x
perim
ent
3
:
c
o
m
pa
r
is
o
n bet
wee
n m
a
x
po
o
lin
g
a
n
d a
v
er
a
g
e
po
o
lin
g
Ma
x
po
o
lin
g
an
d
av
er
a
g
e
p
o
o
lin
g
ar
e
th
e
two
ty
p
es
o
f
p
o
o
li
n
g
lay
e
r
s
.
T
h
e
m
ax
im
u
m
v
alu
e
f
r
o
m
th
e
p
o
r
tio
n
o
f
th
e
im
a
g
e
th
at
th
e
k
er
n
el
(
f
ilter
)
h
as
co
v
er
ed
is
r
etu
r
n
ed
b
y
m
a
x
p
o
o
lin
g
[
3
5
]
,
[
3
6
]
.
On
t
h
e
co
n
tr
ar
y
,
th
e
a
v
er
ag
e
o
f
all
t
h
e
v
alu
es
f
r
o
m
th
e
p
o
r
tio
n
o
f
th
e
im
ag
e
co
v
e
r
ed
b
y
th
e
k
e
r
n
el
is
r
etu
r
n
ed
b
y
av
er
ag
e
p
o
o
lin
g
[
3
7
]
.
T
wo
th
i
n
g
s
wer
e
s
elec
ted
,
wh
ich
ar
e
t
h
e
p
o
o
l
s
ize
an
d
a
s
tr
id
e,
to
p
er
f
o
r
m
m
ax
p
o
o
lin
g
an
d
av
e
r
ag
e
p
o
o
lin
g
.
T
h
e
s
tr
id
e
co
n
tr
o
ls
h
o
w
m
an
y
p
ix
els
th
e
win
d
o
w
will
m
o
v
e
ac
r
o
s
s
th
e
im
ag
e
p
o
o
lin
g
[
3
8
]
,
[
3
9
]
.
T
h
is
ex
p
er
im
en
t
co
m
p
ar
ed
m
ax
p
o
o
lin
g
an
d
av
e
r
ag
e
p
o
o
lin
g
k
er
n
el
s
ize
3
×3
w
ith
a
p
o
o
l
s
ize
o
f
7
an
d
s
tr
id
e
1
.
Si
n
ce
th
e
two
p
r
ev
io
u
s
ex
p
er
im
en
ts
h
a
v
e
p
r
o
v
en
th
at
a
co
m
b
in
atio
n
d
ataset
o
f
o
r
ig
in
al
im
a
g
es
an
d
au
g
m
e
n
ted
im
ag
es,
with
a
d
r
o
p
o
u
t
lay
er
is
a
b
etter
o
p
tio
n
,
th
is
ca
s
e
is
al
s
o
ap
p
lied
in
E
x
p
er
im
en
t
3
.
Mo
r
eo
v
er
,
1
6
8
0
im
ag
es
wer
e
ad
d
ed
,
2
0
0
f
o
r
tr
ain
in
g
(
8
0
%)
an
d
4
0
f
o
r
test
in
g
(
2
0
%)
f
o
r
ea
ch
o
f
th
e
s
ev
e
n
class
es.
At
th
e
en
d
o
f
th
is
ex
p
er
im
en
t,
two
m
o
d
els
wer
e
tr
ai
n
ed
.
T
h
is
ex
p
er
im
en
t
was
r
u
n
f
o
r
te
n
e
p
o
ch
s
,
an
d
it
to
o
k
f
iv
e
h
o
u
r
s
f
o
r
ea
c
h
m
o
d
el
to
lear
n
.
T
ab
le
5
illu
s
tr
ates
th
e
r
esu
lts
o
f
th
e
ex
p
er
im
en
ts
.
B
y
r
ef
er
r
in
g
to
T
ab
le
5
,
we
c
an
s
ee
th
at
t
h
e
m
o
d
el
with
m
ax
p
o
o
lin
g
is
b
etter
th
an
th
e
m
o
d
el
with
a
v
e
r
ag
e
p
o
o
lin
g
s
in
ce
th
er
e
is
n
o
o
v
er
f
itti
n
g
f
o
r
m
a
x
p
o
o
lin
g.
Ov
er
f
itti
n
g
o
cc
u
r
s
with
av
er
ag
e
p
o
o
lin
g
wh
er
e
t
h
e
tr
ain
in
g
ac
cu
r
ac
y
is
h
ig
h
er
th
an
its
v
alid
atio
n
a
cc
u
r
ac
y
.
Ho
wev
e
r
,
th
e
ac
cu
r
a
cy
ac
h
iev
ed
b
y
th
e
m
o
d
el
with
m
ax
p
o
o
lin
g
was
n
o
t h
ig
h
.
T
h
u
s
,
d
if
f
er
e
n
t p
o
o
lin
g
s
izes w
er
e
ex
p
er
im
e
n
ted
w
ith
an
d
co
m
p
ar
ed
in
th
e
n
ex
t e
x
p
er
im
en
t.
T
ab
le
5
.
E
x
p
er
im
en
t
3
tr
ain
in
g
r
esu
lts
o
n
m
ax
p
o
o
lin
g
an
d
a
v
er
ag
e
p
o
o
lin
g
P
o
o
l
i
n
g
L
a
y
e
r
Tr
a
i
n
i
n
g
A
c
c
u
r
a
c
y
Tr
a
i
n
i
n
g
L
o
ss
V
a
l
i
d
a
t
i
o
n
A
c
c
u
r
a
c
y
V
a
l
i
d
a
t
i
o
n
L
o
ss
M
a
x
P
o
o
l
i
n
g
0
.
2
0
0
0
2
.
2
2
2
7
0
.
2
0
0
0
2
.
2
8
4
2
A
v
e
r
a
g
e
P
o
o
l
i
n
g
0
.
2
7
3
6
1
.
8
5
6
7
0
.
2
4
6
4
2
.
2
0
2
7
3
.
4
.
4
.
E
x
perim
ent
4
:
c
o
m
pa
r
is
o
n bet
wee
n po
o
l size
I
n
th
e
p
r
e
v
io
u
s
ex
p
er
im
en
t,
th
e
p
o
o
l
s
ize
u
s
ed
was
7
.
I
n
t
h
is
ex
p
er
im
en
t,
a
c
o
m
p
a
r
is
o
n
b
et
wee
n
m
ax
p
o
o
lin
g
an
d
av
er
ag
e
p
o
o
lin
g
with
a
p
o
o
l
s
ize
o
f
6
is
p
er
f
o
r
m
ed
.
T
h
is
ex
p
er
im
en
t
was
r
u
n
f
o
r
ten
ep
o
c
h
s
an
d
to
o
k
f
i
v
e
h
o
u
r
s
f
o
r
ea
c
h
m
o
d
el
to
lear
n
.
T
ab
le
6
l
is
ts
th
e
r
esu
lts
an
d
th
e
m
o
d
el
with
m
a
x
p
o
o
lin
g
p
er
f
o
r
m
s
s
lig
h
tly
b
etter
th
an
th
e
m
o
d
el
with
av
er
ag
e
p
o
o
lin
g
s
in
ce
th
e
o
v
er
f
itti
n
g
t
h
at
o
cc
u
r
s
b
y
m
ax
p
o
o
lin
g
is
less
th
an
av
er
a
g
e
p
o
o
lin
g
.
Ho
we
v
er
,
th
e
ac
cu
r
ac
y
was
s
till
n
o
t
h
ig
h
.
T
h
u
s
,
in
t
h
e
n
e
x
t
e
x
p
er
im
en
t,
d
if
f
e
r
en
t
d
r
o
p
o
u
t h
y
p
er
p
a
r
am
eter
s
wer
e
co
m
p
ar
e
d
.
T
ab
le
6
.
E
x
p
er
im
en
t
4
tr
ain
in
g
r
esu
lts
o
n
m
ax
p
o
o
lin
g
an
d
a
v
er
ag
e
p
o
o
lin
g
with
a
p
o
o
l size
o
f
6
P
o
o
l
i
n
g
L
a
y
e
r
Tr
a
i
n
i
n
g
A
c
c
u
r
a
c
y
Tr
a
i
n
i
n
g
L
o
ss
V
a
l
i
d
a
t
i
o
n
A
c
c
u
r
a
c
y
V
a
l
i
d
a
t
i
o
n
L
o
ss
M
a
x
P
o
o
l
i
n
g
0
.
4
1
9
3
1
.
5
6
4
4
0
.
2
4
2
9
3
.
0
6
1
7
A
v
e
r
a
g
e
P
o
o
l
i
n
g
0
.
5
1
5
7
1
.
3
2
2
3
0
.
2
5
0
0
5
.
5
5
2
1
3
.
4
.
5
.
E
x
perim
ent
5
:
c
o
m
pa
r
is
o
n bet
wee
n dro
po
ut
hy
per
pa
ra
m
et
er
E
x
p
er
im
en
t
5
c
o
m
p
a
r
ed
th
e
d
r
o
p
o
u
t
h
y
p
e
r
p
ar
am
eter
s
o
f
0
.
2
an
d
0
.
5
.
T
h
is
ex
p
e
r
im
en
t
w
as
r
u
n
f
o
r
ten
ep
o
c
h
s
an
d
to
o
k
f
iv
e
h
o
u
r
s
f
o
r
ea
c
h
m
o
d
el
t
o
lear
n
.
T
ab
le
7
illu
s
tr
ates
th
at
th
e
m
o
d
el
with
a
d
r
o
p
o
u
t
h
y
p
er
p
ar
am
eter
o
f
0
.
2
(
v
alid
a
tio
n
ac
cu
r
ac
y
o
f
0
.
2
4
2
9
)
s
h
o
ws
b
etter
r
esu
lts
th
an
th
e
m
o
d
el
with
a
d
r
o
p
o
u
t
h
y
p
er
p
ar
am
eter
o
f
0
.
5
(
v
alid
atio
n
ac
cu
r
ac
y
o
f
0
.
2
)
.
T
h
e
f
o
llo
win
g
ex
p
e
r
im
en
t
co
m
p
ar
ed
two
d
if
f
er
e
n
t
h
y
p
er
p
ar
am
eter
v
alu
es o
f
th
e
b
atch
s
ize
to
d
eter
m
in
e
wh
ich
wo
u
ld
ar
r
iv
e
at
a
b
etter
v
alid
at
io
n
ac
cu
r
ac
y
.
T
ab
le
7
.
E
x
p
er
im
en
t
5
tr
ain
in
g
r
es
u
lts
o
n
d
r
o
p
o
u
t h
y
p
er
p
a
r
a
m
eter
s
D
r
o
p
o
u
t
H
y
p
e
r
p
a
r
a
me
t
e
r
s
Tr
a
i
n
i
n
g
A
c
c
u
r
a
c
y
Tr
a
i
n
i
n
g
L
o
ss
V
a
l
i
d
a
t
i
o
n
A
c
c
u
r
a
c
y
V
a
l
i
d
a
t
i
o
n
L
o
ss
0
.
2
0
.
4
7
9
3
1
.
4
1
9
3
0
.
2
4
2
9
2
.
4
6
6
2
0
.
5
0
.
2
0
0
0
2
.
2
2
2
7
0
.
2
0
0
0
2
.
2
8
4
2
3
.
4
.
6
.
E
x
perim
ent
6
:
c
o
m
pa
r
is
o
n bet
wee
n ba
t
ch
s
ize
hy
pe
rpa
ra
m
et
er
T
h
e
b
atch
s
ize,
a
g
r
ad
ie
n
t
d
es
ce
n
t
h
y
p
e
r
p
ar
am
ete
r
,
d
eter
m
in
es
h
o
w
m
an
y
tr
ain
in
g
s
am
p
l
es
m
u
s
t
b
e
ex
am
in
ed
b
e
f
o
r
e
th
e
m
o
d
el'
s
in
ter
n
al
p
ar
am
eter
s
ar
e
u
p
d
at
ed
[
4
0
]
,
[
4
1
]
.
E
x
p
e
r
im
en
t
6
c
o
m
p
ar
ed
t
h
e
b
atch
s
ize
h
y
p
er
p
a
r
am
eter
s
o
f
2
an
d
5
.
T
h
is
ex
p
e
r
im
en
t
was
r
u
n
f
o
r
ten
ep
o
c
h
s
an
d
to
o
k
f
iv
e
h
o
u
r
s
f
o
r
ea
c
h
m
o
d
el
to
lear
n
.
T
ab
le
8
s
h
o
ws
th
at
th
e
m
o
d
el
with
a
b
atch
s
ize
h
y
p
er
p
ar
am
eter
o
f
5
p
r
o
d
u
ce
s
b
etter
r
esu
lts
(
v
alid
atio
n
ac
cu
r
ac
y
o
f
0
.
2
4
6
4
)
th
an
th
e
m
o
d
el
with
a
b
atc
h
s
ize
h
y
p
er
p
ar
am
et
er
o
f
2
(
v
alid
atio
n
ac
cu
r
ac
y
o
f
0
.
2
)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
4
,
No
.
1
,
Ma
r
c
h
2
0
2
5
:
28
-
38
34
T
ab
le
8
.
E
x
p
er
im
en
t
6
tr
ain
in
g
r
esu
lts
o
n
b
atch
s
ize
h
y
p
er
p
ar
am
eter
s
B
a
t
c
h
S
i
z
e
H
y
p
e
r
p
a
r
a
m
e
t
e
r
s
Tr
a
i
n
i
n
g
A
c
c
u
r
a
c
y
Tr
a
i
n
i
n
g
L
o
ss
V
a
l
i
d
a
t
i
o
n
A
c
c
u
r
a
c
y
V
a
l
i
d
a
t
i
o
n
L
o
ss
2
0
.
2
0
0
0
2
.
2
2
2
7
0
.
2
0
0
0
2
.
2
8
4
2
5
0
.
2
5
0
7
3
.
2
2
7
9
0
.
2
4
6
4
4
.
4
2
9
8
3
.
4
.
7
.
E
x
perim
ent
7
:
re
t
ra
in
ex
perim
ent
3
wit
h 2
0
0
epo
ch
s
Sin
ce
th
e
b
est
m
o
d
el
s
o
f
ar
was
ac
h
iev
ed
in
ex
p
er
i
m
en
t
3
wh
er
e
th
e
m
o
d
el
was
tr
ain
ed
with
m
ax
p
o
o
lin
g
with
a
p
o
o
l
s
ize
o
f
7
,
d
r
o
p
o
u
t
h
y
p
er
p
ar
am
eter
o
f
0
.
5
,
an
d
b
atch
s
ize
h
y
p
e
r
p
ar
am
eter
o
f
2
,
t
h
is
ex
p
er
im
en
t
u
s
ed
th
e
s
am
e
h
y
p
er
p
ar
am
eter
s
b
u
t
with
2
0
0
e
p
o
ch
s
.
I
t
to
o
k
t
h
r
ee
d
ay
s
f
o
r
th
is
m
o
d
el
to
lear
n
,
b
u
t
th
e
v
alid
atio
n
ac
cu
r
ac
y
p
r
o
d
u
ce
d
was
lo
wer
t
h
an
in
e
x
p
er
im
en
t
3
.
Sin
ce
th
is
m
o
d
el
d
i
d
n
o
t
p
r
o
d
u
ce
h
ig
h
v
alid
atio
n
ac
cu
r
a
cy
as
s
h
o
w
n
in
T
ab
le
9
,
t
h
e
co
llected
d
ata
was
an
aly
z
ed
a
g
ain
.
I
t
was
f
o
u
n
d
th
at
th
e
b
ac
k
g
r
o
u
n
d
im
ag
es
an
d
th
e
ap
p
ea
r
an
ce
s
o
f
th
e
m
ale
B
etta
f
is
h
an
d
f
em
ale
B
etta
f
is
h
d
if
f
er
s
ig
n
if
ican
tly
.
T
h
er
ef
o
r
e,
in
th
e
n
ex
t
ex
p
e
r
im
en
t,
th
e
d
ataset
was
s
ep
ar
at
ed
ac
co
r
d
in
g
to
th
e
B
etta
f
is
h
g
e
n
d
er
,
m
ale
a
n
d
f
em
ale,
with
f
o
u
r
class
es o
n
ly
d
u
e
to
th
e
lack
o
f
f
em
ale
B
etta
f
is
h
im
ag
es,
tim
e,
an
d
h
ar
d
war
e
co
n
s
tr
ain
ts
.
T
ab
le
9
.
E
x
p
er
im
en
t
7
tr
ain
in
g
r
esu
lts
o
n
r
etr
ain
ex
p
e
r
im
en
t
3
with
2
0
0
ep
o
c
h
s
Ep
o
c
h
Tr
a
i
n
i
n
g
A
c
c
u
r
a
c
y
Tr
a
i
n
i
n
g
L
o
ss
V
a
l
i
d
a
t
i
o
n
A
c
c
u
r
a
c
y
V
a
l
i
d
a
t
i
o
n
L
o
ss
2
0
0
0
.
9
9
9
3
0
.
0
0
4
3
0
.
4
4
6
4
5
.
1
3
0
1
3
.
4
.
8
.
E
x
perim
ent
8
:
t
hree
mo
dels
T
h
r
ee
m
o
d
els
wer
e
tr
ain
ed
i
n
th
is
ex
p
e
r
im
en
t:
th
e
m
ale
B
etta
f
is
h
d
ataset,
th
e
f
em
ale
B
etta
f
is
h
d
ataset,
an
d
th
e
co
m
b
in
atio
n
o
f
th
e
m
ale
an
d
f
em
ale
B
etta
f
is
h
d
ataset.
T
h
e
d
ataset
was
r
ed
u
ce
d
to
4
0
0
im
ag
es
in
ea
ch
m
o
d
el.
B
ased
o
n
ex
p
er
im
e
n
t
3
,
it
tu
r
n
e
d
o
u
t
th
at
m
ax
p
o
o
lin
g
,
p
o
o
l
s
ize
o
f
7
,
d
r
o
p
o
u
t
o
f
0
.
5
,
an
d
b
atch
s
ize
o
f
2
s
h
o
wed
th
e
b
est
r
esu
lt.
T
h
e
r
ef
o
r
e,
th
ese
h
y
p
e
r
p
ar
am
ete
r
s
wer
e
u
s
ed
in
th
is
ex
p
er
im
en
t.
B
esid
es,
th
ese
m
o
d
els
u
s
ed
s
to
ch
asti
c
g
r
ad
ien
t
d
escen
t
(
SGD)
as
an
o
p
tim
i
z
er
an
d
So
f
tMa
x
as
a
n
ac
tiv
atio
n
f
u
n
ctio
n
.
T
h
is
ex
p
e
r
im
en
t
was
r
u
n
f
o
r
8
0
ep
o
ch
s
an
d
to
o
k
th
r
ee
h
o
u
r
s
f
o
r
ea
ch
m
o
d
el
to
lea
r
n
.
As
s
h
o
wn
in
T
ab
le
1
0
,
th
e
v
alid
at
io
n
ac
cu
r
ac
y
was m
u
ch
im
p
r
o
v
ed
co
m
p
ar
ed
t
o
th
e
p
r
ev
io
u
s
ex
p
er
im
en
ts
.
T
ab
le
1
0
.
E
x
p
er
im
e
n
t
8
tr
ai
n
in
g
r
esu
lts
o
n
th
r
ee
m
o
d
els
M
o
d
e
l
Tr
a
i
n
i
n
g
A
c
c
u
r
a
c
y
Tr
a
i
n
i
n
g
L
o
ss
V
a
l
i
d
a
t
i
o
n
A
c
c
u
r
a
c
y
V
a
l
i
d
a
t
i
o
n
L
o
ss
M
a
l
e
0
.
9
8
7
5
0
.
0
2
3
7
0
.
7
0
0
0
7
.
4
7
6
5
F
e
mal
e
0
.
9
5
6
3
0
.
1
7
2
1
0
.
8
3
7
5
0
.
9
6
2
2
M
a
l
e
a
n
d
F
e
ma
l
e
0
.
9
6
5
6
0
.
1
0
1
6
0
.
6
5
0
0
2
.
0
2
3
4
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
N
T
h
is
s
tu
d
y
co
n
d
u
cted
a
co
m
p
r
eh
en
s
iv
e
s
er
ies
o
f
eig
h
t
ex
p
er
im
en
ts
to
o
p
tim
i
z
e
B
etta
f
i
s
h
s
p
ec
ies
class
if
icatio
n
th
r
o
u
g
h
DL
,
y
i
eld
in
g
1
4
d
is
tin
ct
m
o
d
els.
T
h
e
p
r
im
ar
y
o
b
jectiv
e
was
to
d
ev
elo
p
a
r
o
b
u
s
t
class
if
icatio
n
s
y
s
tem
to
clas
s
if
y
B
etta
f
i
s
h
s
p
ec
ies.
T
h
e
s
e
ex
p
er
im
en
ts
wer
e
ex
ec
u
ted
with
m
eticu
lo
u
s
atten
tio
n
to
f
ac
to
r
s
s
u
ch
as
d
ata
au
g
m
en
tatio
n
,
h
y
p
er
p
ar
a
m
eter
tu
n
in
g
,
an
d
th
e
u
tili
z
atio
n
o
f
th
e
Mo
b
ileNet
m
o
d
el
f
o
r
its
ef
f
icien
c
y
in
la
r
g
e
-
s
ca
le
im
ag
e
class
if
icatio
n
p
r
o
ce
s
s
in
g
.
T
h
e
c
u
lm
in
atio
n
o
f
th
ese
e
f
f
o
r
t
s
is
en
ca
p
s
u
lated
in
t
h
e
co
m
p
ar
is
o
n
o
f
th
e
1
4
m
o
d
els,
d
etailed
i
n
T
ab
le
1
1
.
W
ith
in
T
a
b
le
1
1
,
th
e
p
er
f
o
r
m
an
ce
o
f
ea
ch
m
o
d
el
is
s
cr
u
tin
i
z
ed
,
an
d
th
e
f
in
d
in
g
s
r
ev
ea
l
th
at
th
e
f
in
al
th
r
ee
m
o
d
els
f
r
o
m
ex
p
er
im
en
t
8
ex
h
ib
it
th
e
m
o
s
t
p
r
o
m
is
in
g
r
esu
lts
.
T
h
e
e
n
s
u
i
n
g
s
ec
tio
n
d
elv
es
i
n
to
a
d
etailed
an
aly
s
is
o
f
th
ese
o
u
tco
m
es,
s
h
ed
d
in
g
lig
h
t
o
n
th
e
k
ey
f
ac
to
r
s
in
f
lu
en
ci
n
g
th
e
s
u
p
er
io
r
p
er
f
o
r
m
an
ce
o
f
th
e
s
elec
ted
m
o
d
els.
T
h
e
B
etta
f
is
h
clas
s
if
icatio
n
m
o
d
el,
em
p
lo
y
in
g
Mo
b
ileNet
ar
ch
itectu
r
e,
d
em
o
n
s
tr
ates
ex
ce
p
tio
n
al
ac
cu
r
ac
y
b
y
u
tili
z
in
g
m
ax
p
o
o
lin
g
,
a
p
o
o
l
s
ize
o
f
7
,
a
d
r
o
p
o
u
t
r
ate
o
f
0
.
5
,
an
d
a
b
atch
s
ize
o
f
2
,
co
u
p
led
wit
h
SGD
o
p
tim
i
z
er
an
d
So
f
tMa
x
as
th
e
ac
tiv
atio
n
f
u
n
ctio
n
ac
r
o
s
s
8
0
ep
o
ch
s
.
T
h
e
v
alid
atio
n
ac
cu
r
ac
y
f
o
r
m
ale
B
etta
f
is
h
r
ea
ch
es
0
.
7
,
wh
ile
f
em
ale
B
etta
f
i
s
h
ac
h
iev
es
a
n
im
p
r
ess
iv
e
0
.
8
3
7
5
.
Ho
wev
er
,
co
m
b
in
in
g
b
o
th
m
ale
an
d
f
em
ale
im
ag
es
r
esu
lt
s
in
a
s
lig
h
tly
lo
wer
ac
c
u
r
ac
y
o
f
0
.
6
5
.
N
o
tab
ly
,
t
h
e
d
ec
is
io
n
to
em
p
lo
y
s
ep
ar
ate
m
o
d
els
f
o
r
m
ale
an
d
f
em
ale
class
if
icatio
n
p
r
o
v
es
ad
v
an
ta
g
eo
u
s
,
h
ig
h
lig
h
tin
g
th
e
s
u
b
s
t
an
tial
d
if
f
er
e
n
ce
s
in
s
h
ap
e
an
d
co
lo
r
b
etwe
en
m
ale
an
d
f
em
ale
B
etta
f
is
h
th
at
i
m
p
ac
t
ac
cu
r
ate
class
if
icatio
n
.
I
t
u
n
d
er
s
co
r
es
th
e
im
p
o
r
tan
ce
o
f
tailo
r
e
d
m
o
d
els f
o
r
d
is
tin
ct
g
en
d
er
s
to
o
p
tim
i
z
e
class
if
icatio
n
p
er
f
o
r
m
a
n
ce
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
B
etta
fis
h
s
p
ec
ies cla
s
s
if
ica
tio
n
u
s
in
g
lig
h
t weig
h
t
d
ee
p
…
(
Da
n
is
h
a
h
Ha
n
a
Mu
h
a
mma
d
Mu
h
a
imin
Lim
)
35
T
ab
le
1
1
.
Su
m
m
ar
y
o
f
e
x
p
er
i
m
en
tal
r
esu
lts
with
v
ar
io
u
s
tr
a
in
in
g
m
o
d
els
Ex
p
e
r
i
m
e
n
t
D
a
t
a
s
e
t
A
c
c
u
r
a
c
y
Lo
ss
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
Tr
a
i
n
i
n
g
V
a
l
i
d
a
t
i
o
n
1
O
r
i
g
i
n
a
l
0
.
9
0
3
6
0
.
2
7
8
6
0
.
3
2
9
8
4
.
5
0
2
8
C
o
m
b
i
n
e
d
0
.
9
1
2
5
0
.
3
3
5
7
0
.
3
0
0
7
6
.
4
9
4
4
2
O
r
i
g
i
n
a
l
0
.
6
4
6
4
0
.
2
3
5
7
1
.
0
4
0
4
4
.
5
4
6
6
C
o
m
b
i
n
e
d
0
.
5
8
7
5
0
.
3
6
4
3
1
.
2
1
6
9
3
.
8
3
9
3
3
C
o
m
b
i
n
e
d
(
M
a
x
P
o
o
l
i
n
g
)
0
.
2
0
0
0
0
.
2
0
0
0
2
.
2
2
2
7
2
.
2
8
4
2
C
o
m
b
i
n
e
d
(
A
v
e
r
a
g
e
P
o
o
l
i
n
g
)
0
.
2
7
3
6
0
.
2
4
6
4
1
.
8
5
6
7
2
.
2
0
2
7
4
C
o
m
b
i
n
e
d
(
M
a
x
P
o
o
l
i
n
g
,
P
o
o
l
S
i
z
e
=
6
)
0
.
4
1
9
3
0
.
2
4
2
9
1
.
5
6
4
4
3
.
0
6
1
7
C
o
m
b
i
n
e
d
(
A
v
e
r
a
g
e
P
o
o
l
i
n
g
,
P
o
o
l
S
i
z
e
=
6
)
0
.
5
1
5
7
0
.
2
5
0
0
1
.
3
2
2
3
5
.
5
5
2
1
5
C
o
m
b
i
n
e
d
(
D
r
o
p
o
u
t
(
0
.
2
)
)
0
.
4
7
9
3
0
.
2
4
2
9
1
.
4
1
9
3
2
.
4
6
6
2
C
o
m
b
i
n
e
d
(
D
r
o
p
o
u
t
(
0
.
5
)
)
0
.
2
0
0
0
0
.
2
0
0
0
2
.
2
2
2
7
2
.
2
8
4
2
6
C
o
m
b
i
n
e
d
(
B
a
t
c
h
S
i
z
e
=
2
)
0
.
2
0
0
0
0
.
2
0
0
0
2
.
2
2
2
7
2
.
2
8
4
2
C
o
m
b
i
n
e
d
(
B
a
t
c
h
S
i
z
e
=
5
)
0
.
2
5
0
7
0
.
2
4
6
4
3
.
2
2
7
9
4
.
4
2
9
8
7
C
o
m
b
i
n
e
d
(
Ep
o
c
h
=
2
0
0
)
0
.
9
9
9
3
0
.
4
4
6
4
0
.
0
0
4
3
5
.
1
3
0
1
8
M
a
l
e
0
.
9
8
7
5
0
.
7
0
0
0
0
.
0
2
3
7
7
.
4
7
6
5
F
e
mal
e
0
.
9
5
6
3
0
.
8
3
7
5
0
.
1
7
2
1
0
.
9
6
2
2
M
a
l
e
a
n
d
F
e
ma
l
e
0
.
9
6
5
6
0
.
6
5
0
0
0
.
1
0
1
6
2
.
0
2
3
4
4
.
1
.
M
o
bil
e
a
pp
lica
t
io
n
dev
elo
pm
ent
T
h
e
m
o
b
ile
ap
p
licatio
n
d
ev
el
o
p
m
en
t
in
v
o
lv
es
h
ar
n
ess
in
g
th
e
ca
p
ab
ilit
ies
o
f
th
e
tr
ain
ed
m
o
d
els
to
p
r
o
d
u
ce
a
u
s
er
-
f
r
ien
d
l
y
ap
p
licatio
n
tailo
r
ed
f
o
r
B
etta
f
is
h
s
p
ec
ies
class
if
icat
io
n
o
n
th
e
An
d
r
o
id
p
latf
o
r
m
.
I
n
teg
r
atin
g
a
n
in
tu
itiv
e
in
ter
f
a
ce
with
in
th
e
ap
p
licatio
n
en
ab
les
u
s
er
s
to
s
ea
m
les
s
ly
ca
p
tu
r
e
o
r
u
p
lo
ad
im
a
g
es
f
o
r
i
n
s
tan
tan
eo
u
s
s
p
ec
ies
id
en
tific
atio
n
.
T
h
is
f
ea
tu
r
e
-
r
ic
h
ap
p
licatio
n
aim
s
t
o
p
r
o
v
id
e
u
s
er
s
with
a
s
tr
aig
h
tf
o
r
war
d
an
d
en
g
a
g
in
g
ex
p
er
ien
ce
,
allo
win
g
th
e
m
to
ac
tiv
ely
p
ar
ticip
ate
in
t
h
e
r
ea
l
-
tim
e
id
en
tific
atio
n
o
f
B
etta
f
is
h
s
p
ec
ies ea
s
ily
an
d
ac
cu
r
ately
.
Fig
u
r
e
4
illu
s
tr
ates
a
s
am
p
le
in
ter
f
ac
e
o
f
th
e
B
etta
f
is
h
s
p
ec
ies
class
if
icatio
n
m
o
b
ile
ap
p
licatio
n
.
Fig
u
r
e
4
(
a
)
illu
s
tr
ates
th
e
m
ai
n
p
ag
e,
wh
ich
c
o
n
s
is
ts
o
f
two
b
u
tto
n
s
,
n
am
ely
‘
B
u
k
a
K
a
mera
’
to
ca
p
tu
r
e
th
e
im
ag
e
o
f
th
e
B
etta
f
is
h
u
s
in
g
th
e
m
o
b
ile’
s
ca
m
er
a
an
d
‘
B
u
ka
Ga
leri
’
to
s
elec
t
a
s
av
ed
i
m
ag
e
s
to
r
ed
in
th
e
g
aller
y
to
class
if
y
th
e
s
p
ec
ies.
Fig
u
r
e
4
(
b
)
will
b
e
d
is
p
lay
ed
if
a
u
s
er
clic
k
s
th
e
‘
B
u
ka
K
a
mera
’
b
u
tto
n
,
wh
ile
Fig
u
r
e
4
(
c
)
wi
ll
b
e
s
h
o
wn
i
f
a
u
s
er
clic
k
s
th
e
‘
B
u
ka
Ga
l
eri
’
b
u
tto
n
.
User
s
ca
n
also
p
r
ess
th
e
b
ac
k
ar
r
o
w
b
u
tto
n
to
n
av
ig
ate
t
o
th
e
p
r
e
v
io
u
s
p
a
g
e.
Fig
u
r
e
5
s
h
o
ws
t
h
e
r
esu
lt
p
ag
e,
wh
er
e
t
h
e
n
a
m
e
o
f
th
e
B
etta
f
is
h
s
p
ec
ies
an
d
its
r
an
g
e
o
f
p
r
ices
in
R
in
g
g
it
Ma
lay
s
ia
(
R
M)
ar
e
s
h
o
wn
.
T
h
e
u
s
er
ca
n
th
en
r
ep
ea
t
th
is
p
r
o
ce
s
s
f
o
r
th
e
n
ex
t
B
etta
f
is
h
im
ag
e
class
if
icatio
n
.
(
a)
(
b
)
(
c)
Fig
u
r
e
4
.
Sam
p
le
i
n
ter
f
ac
e
o
f
t
h
e
m
o
b
ile
a
p
p
licatio
n
of
(
a)
m
ain
p
ag
e,
(
b
)
p
ag
e
to
ca
p
tu
r
e
t
h
e
im
ag
e
af
ter
click
in
g
th
e
‘
B
u
ka
K
a
mera
’
b
u
tto
n
,
an
d
(
c)
p
ag
e
t
o
s
elec
t a
n
im
ag
e
af
ter
click
i
n
g
th
e
‘
B
u
k
a
Ga
leri
’
b
u
tto
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2252
-
8
8
1
4
I
n
t J Ad
v
Ap
p
l Sci
,
Vo
l.
1
4
,
No
.
1
,
Ma
r
c
h
2
0
2
5
:
28
-
38
36
Fig
u
r
e
5
.
Sam
p
le
i
n
ter
f
ac
e
f
o
r
th
e
r
esu
lt p
ag
e
5.
CO
NCLU
SI
O
N
AND
F
U
T
U
RE
WO
RK
I
n
co
n
clu
s
io
n
,
th
is
r
esear
c
h
s
u
cc
ess
f
u
lly
ex
p
lo
r
es
an
d
im
p
lem
en
ts
th
e
Mo
b
ileNet
ar
c
h
itectu
r
e
to
class
if
y
B
etta
f
is
h
s
p
ec
ies,
ad
d
r
ess
in
g
ch
allen
g
es
f
ac
ed
b
y
b
r
ee
d
er
s
a
n
d
s
eller
s
in
ac
cu
r
at
ely
id
en
tif
y
in
g
th
ese
aq
u
atic
s
p
ec
ies.
Mo
b
ileNet,
wh
ich
u
s
es
d
ep
th
wis
e
s
ep
ar
ab
le
co
n
v
o
lu
tio
n
s
,
is
an
e
f
f
i
cien
t
s
o
lu
tio
n
f
o
r
d
ep
lo
y
m
e
n
t
o
n
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
d
ev
ices
.
T
h
e
c
r
u
cial
f
in
d
in
g
is
th
at
h
av
in
g
s
ep
ar
ate
m
o
d
els
f
o
r
m
ale
an
d
f
em
ale
class
if
icatio
n
s
im
p
r
o
v
es
ac
cu
r
ac
y
s
ig
n
if
ican
tly
,
h
ig
h
lig
h
tin
g
th
e
n
u
a
n
ce
d
d
if
f
er
e
n
ce
s
in
s
h
ap
e
an
d
co
lo
r
b
etwe
en
g
en
d
er
s
.
T
h
is
r
esear
ch
n
o
t
o
n
ly
co
n
tr
i
b
u
tes
to
th
e
ad
v
an
ce
m
en
t
o
f
th
e
B
ett
a
f
is
h
clas
s
if
ica
tio
n
tech
n
iq
u
e
b
u
t
also
p
r
o
v
id
es
a
p
r
ac
tical
to
o
l
f
o
r
in
d
u
s
tr
y
s
tak
eh
o
ld
er
s
to
en
h
an
ce
th
e
ir
d
ec
is
io
n
-
m
ak
in
g
p
r
o
ce
s
s
es a
n
d
r
ed
u
ce
m
o
n
etar
y
lo
s
s
es a
s
s
o
ciate
d
with
m
is
cl
ass
if
icatio
n
.
C
o
n
ce
r
n
in
g
f
u
tu
r
e
wo
r
k
,
s
ev
er
al
av
en
u
es
ex
is
t
f
o
r
en
h
a
n
c
in
g
th
e
B
etta
f
is
h
s
p
ec
ies
c
la
s
s
if
icatio
n
m
o
d
el
b
ased
o
n
Mo
b
ileNet
ar
ch
itectu
r
e.
E
x
p
a
n
d
in
g
th
e
d
at
aset
to
in
clu
d
e
a
b
r
o
ad
er
s
p
e
ctr
u
m
o
f
B
etta
f
is
h
im
ag
es
with
d
iv
er
s
e
c
o
lo
r
s
,
p
atter
n
s
,
an
d
tail
s
h
ap
es
ca
n
co
n
tr
i
b
u
te
to
th
e
im
p
r
o
v
em
en
t
o
f
m
o
d
el
g
en
er
ali
z
atio
n
an
d
r
e
d
u
ce
th
e
o
v
er
f
itti
n
g
p
r
o
b
lem
.
C
o
n
s
id
er
in
g
th
e
p
o
ten
tial
b
en
ef
its
,
u
tili
zin
g
p
r
e
-
tr
ain
ed
m
o
d
els,
an
d
ex
p
e
r
im
en
tin
g
with
ad
v
an
ce
d
d
ata
a
u
g
m
e
n
tatio
n
tech
n
i
q
u
es
co
u
l
d
f
u
r
th
er
e
n
h
an
ce
th
e
m
o
d
el'
s
r
o
b
u
s
tn
ess
.
I
n
v
esti
g
atin
g
en
s
em
b
le
lear
n
in
g
ap
p
r
o
ac
h
es
a
n
d
in
teg
r
atin
g
u
s
e
r
f
ee
d
b
ac
k
m
ec
h
an
is
m
s
in
th
e
m
o
b
ile
ap
p
licatio
n
ca
n
c
o
n
tr
ib
u
te
to
co
n
tin
u
o
u
s
m
o
d
el
im
p
r
o
v
em
en
t.
O
p
tim
i
z
atio
n
f
o
r
r
ea
l
-
tim
e
d
ep
lo
y
m
en
t,
co
m
p
atib
ilit
y
with
v
ar
io
u
s
m
o
b
ile
d
e
v
ices,
an
d
co
llab
o
r
at
io
n
with
a
q
u
atic
e
x
p
er
ts
to
i
n
co
r
p
o
r
ate
d
o
m
ain
k
n
o
wled
g
e
ar
e
cr
u
cial
co
n
s
id
er
atio
n
s
.
T
h
ese
f
u
tu
r
e
d
ir
ec
tio
n
s
aim
to
r
e
f
in
e
th
e
m
o
d
el
’
s
ac
cu
r
ac
y
,
u
s
ab
ilit
y
,
an
d
ad
a
p
tab
ilit
y
to
b
e
n
ef
it
B
etta
f
is
h
b
r
ee
d
er
s
an
d
s
eller
s
.
ACK
NO
WL
E
DG
M
E
N
T
S
T
h
an
k
s
ar
e
d
u
e
to
t
h
e
C
o
lleg
e
o
f
C
o
m
p
u
tin
g
,
I
n
f
o
r
m
atics
an
d
Ma
th
em
atics
at
U
n
iv
er
s
iti
T
ek
n
o
lo
g
i
MA
R
A,
Sh
ah
Alam
,
Selan
g
o
r
,
Ma
lay
s
ia,
f
o
r
th
eir
s
u
p
p
o
r
t in
f
ac
ilit
atin
g
th
is
wo
r
k
.
RE
F
E
R
E
NC
E
S
[
1
]
A
.
S
i
n
h
a
a
n
d
P
.
K
.
P
a
n
d
e
y
,
Bre
e
d
i
n
g
a
n
d
c
u
l
t
u
re
o
f
f
r
e
s
h
w
a
t
e
r
o
rn
a
m
e
n
t
a
l
f
i
sh
.
L
o
n
d
o
n
:
C
R
C
P
r
e
ss,
2
0
2
3
.
d
o
i
:
1
0
.
1
2
0
1
/
9
7
8
1
0
0
3
4
5
6
8
5
8
.
[
2
]
K
.
T
a
n
t
a
y
a
k
u
l
a
n
d
W
.
P
a
n
i
c
h
p
a
t
t
a
n
a
k
u
l
,
“
A
c
o
m
p
a
r
a
t
i
v
e
s
t
u
d
y
o
f
ma
c
h
i
n
e
l
e
a
r
n
i
n
g
f
o
r
I
O
S
b
a
s
e
d
o
n
S
i
a
m
B
e
t
t
a
m
o
b
i
l
e
a
p
p
l
i
c
a
t
i
o
n
,
”
i
n
2
0
2
0
-
5
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
I
n
f
o
rm
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
(
I
n
C
I
T)
,
I
EEE,
O
c
t
.
2
0
2
0
,
p
p
.
1
0
4
–
1
0
9
.
d
o
i
:
1
0
.
1
1
0
9
/
I
n
C
I
T5
0
5
8
8
.
2
0
2
0
.
9
3
1
0
9
3
2
.
[
3
]
J.
W
a
n
g
,
B
.
C
a
o
,
P
.
Y
u
,
L.
S
u
n
,
W
.
B
a
o
,
a
n
d
X
.
Zh
u
,
“
D
e
e
p
l
e
a
r
n
i
n
g
t
o
w
a
r
d
s
m
o
b
i
l
e
a
p
p
l
i
c
a
t
i
o
n
s,”
i
n
2
0
1
8
I
EEE
3
8
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
D
i
s
t
ri
b
u
t
e
d
C
o
m
p
u
t
i
n
g
S
y
st
e
m
s
(
I
C
D
C
S
)
,
I
EEE,
J
u
l
.
2
0
1
8
,
p
p
.
1
3
8
5
–
1
3
9
3
.
d
o
i
:
1
0
.
1
1
0
9
/
I
C
D
C
S
.
2
0
1
8
.
0
0
1
3
9
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J Ad
v
Ap
p
l Sci
I
SS
N:
2252
-
8
8
1
4
B
etta
fis
h
s
p
ec
ies cla
s
s
if
ica
tio
n
u
s
in
g
lig
h
t weig
h
t
d
ee
p
…
(
Da
n
is
h
a
h
Ha
n
a
Mu
h
a
mma
d
Mu
h
a
imin
Lim
)
37
[
4
]
B
e
r
n
a
ma
,
“
F
i
g
h
t
i
n
g
f
i
sh
s
t
a
l
l
a
t
t
r
a
c
t
s
l
a
r
g
e
c
r
o
w
d
,
”
N
e
w
S
t
r
a
i
t
s
Ti
mes.
A
c
c
e
sse
d
:
O
c
t
.
1
8
,
2
0
2
2
.
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
s
:
/
/
w
w
w
.
n
s
t
.
c
o
m
.
m
y
/
n
e
w
s/
n
a
t
i
o
n
/
2
0
2
2
/
0
3
/
7
8
3
5
7
0
/
f
i
g
h
t
i
n
g
-
f
i
s
h
-
st
a
l
l
-
a
t
t
r
a
c
t
s
-
l
a
r
g
e
-
c
r
o
w
d
[
5
]
Y
.
H
e
n
i
n
g
t
y
a
s
,
F
.
R
a
h
m
i
,
a
n
d
K
.
M
u
l
u
d
i
,
“
I
mp
l
e
me
n
t
a
t
i
o
n
o
f
d
e
n
s
i
t
y
-
b
a
s
e
d
c
l
u
st
e
r
i
n
g
i
n
b
e
t
t
a
f
i
sh
i
ma
g
e
seg
m
e
n
t
a
t
i
o
n
(
i
n
I
n
d
o
n
e
si
a
n
:
I
m
p
l
e
m
e
n
t
a
s
i
d
e
n
si
t
y
-
b
a
s
e
d
c
l
u
st
e
ri
n
g
p
a
d
a
s
e
g
m
e
n
t
a
s
i
c
i
t
r
a
b
e
t
t
a
f
i
s
h
)
,
”
J
.
T
e
k
n
o
i
n
f
o
,
v
o
l
.
1
6
,
n
o
.
1
,
p
.
8
,
J
a
n
.
2
0
2
2
,
do
i
:
1
0
.
3
3
3
6
5
/
j
t
i
.
v
1
6
i
1
.
1
2
7
3
.
[
6
]
Z.
A
b
i
d
i
n
,
R
u
sl
i
y
a
w
a
t
i
,
P
e
r
ma
t
a
,
F
.
A
r
i
a
n
y
,
I
.
S
o
l
e
h
u
d
i
n
,
a
n
d
A
.
Ju
n
a
i
d
i
,
“
B
e
t
t
a
f
i
s
h
i
ma
g
e
i
d
e
n
t
i
f
i
c
a
t
i
o
n
u
s
i
n
g
f
e
a
t
u
r
e
e
x
t
r
a
c
t
i
o
n
G
LC
M
a
n
d
k
-
n
e
a
r
e
st
n
e
i
g
h
b
o
u
r
c
l
a
ssi
f
i
c
a
t
i
o
n
,
”
i
n
2
0
2
2
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
I
n
f
o
rm
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
R
e
se
a
rc
h
a
n
d
I
n
n
o
v
a
t
i
o
n
(
I
C
I
T
RI
)
,
I
EEE,
N
o
v
.
2
0
2
2
,
p
p
.
1
5
6
–
1
6
1
.
d
o
i
:
1
0
.
1
1
0
9
/
I
C
I
TR
I
5
6
4
2
3
.
2
0
2
2
.
9
9
7
0
2
0
9
.
[
7
]
L.
M
o
o
k
d
a
r
s
a
n
i
t
a
n
d
P
.
M
o
o
k
d
a
r
sa
n
i
t
,
“
S
i
a
mF
i
s
h
N
e
t
:
t
h
e
d
e
e
p
i
n
v
e
st
i
g
a
t
i
o
n
o
f
s
i
a
m
e
se
f
i
g
h
t
i
n
g
f
i
s
h
e
s
,
”
I
n
t
.
J
.
A
p
p
l
.
C
o
m
p
u
t
.
T
e
c
h
n
o
l
.
I
n
f
.
S
y
st
.
,
v
o
l
.
8
,
n
o
.
2
,
2
0
1
9
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
2
0
3
.
1
5
8
.
9
8
.
1
2
/
a
c
t
i
sj
o
u
r
n
a
l
/
i
n
d
e
x
.
p
h
p
/
I
JA
C
TI
S
/
a
r
t
i
c
l
e
/
v
i
e
w
/
2
5
6
[
8
]
S
.
H
i
d
a
y
a
t
,
A
.
Y
.
R
a
h
ma
n
,
a
n
d
I
st
i
a
d
i
,
“
B
e
t
t
a
f
i
s
h
i
ma
g
e
c
l
a
ssi
f
i
c
a
t
i
o
n
u
si
n
g
a
r
t
i
f
i
c
i
a
l
n
e
u
r
a
l
n
e
t
w
o
r
k
s
w
i
t
h
G
a
b
o
r
e
x
t
r
a
c
t
i
o
n
f
e
a
t
u
r
e
s,
”
i
n
2
0
2
2
I
EE
E
I
n
t
e
rn
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
C
y
b
e
r
n
e
t
i
c
s
a
n
d
C
o
m
p
u
t
a
t
i
o
n
a
l
I
n
t
e
l
l
i
g
e
n
c
e
(
C
y
b
e
rn
e
t
i
c
s
C
o
m
)
,
I
EEE,
J
u
n
.
2
0
2
2
,
p
p
.
2
7
0
–
2
7
3
.
d
o
i
:
1
0
.
1
1
0
9
/
C
y
b
e
r
n
e
t
i
c
sC
o
m
5
5
2
8
7
.
2
0
2
2
.
9
8
6
5
5
0
9
.
[
9
]
D
.
M
.
H
i
b
b
a
n
a
n
d
W
.
F
.
A
l
M
a
k
i
,
“
C
l
a
ssi
f
i
c
a
t
i
o
n
o
f
o
r
n
a
me
n
t
a
l
b
e
t
t
a
f
i
s
h
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
me
t
h
o
d
a
n
d
g
r
a
b
c
u
t
seg
m
e
n
t
a
t
i
o
n
,
”
i
n
2
0
2
1
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
D
a
t
a
S
c
i
e
n
c
e
a
n
d
I
t
s A
p
p
l
i
c
a
t
i
o
n
s
(
I
C
o
D
S
A)
,
I
EEE,
O
c
t
.
2
0
2
1
,
p
p
.
1
0
2
–
1
0
9
.
d
o
i
:
1
0
.
1
1
0
9
/
I
C
o
D
S
A
5
3
5
8
8
.
2
0
2
1
.
9
6
1
7
2
1
3
.
[
1
0
]
O
.
U
k
w
a
n
d
u
,
H
.
H
i
n
d
y
,
a
n
d
E.
U
k
w
a
n
d
u
,
“
A
n
e
v
a
l
u
a
t
i
o
n
o
f
l
i
g
h
t
w
e
i
g
h
t
d
e
e
p
l
e
a
r
n
i
n
g
t
e
c
h
n
i
q
u
e
s
i
n
me
d
i
c
a
l
i
ma
g
i
n
g
f
o
r
h
i
g
h
p
r
e
c
i
si
o
n
C
O
V
I
D
-
1
9
d
i
a
g
n
o
st
i
c
s
,
”
H
e
a
l
t
h
c
.
An
a
l
.
,
v
o
l
.
2
,
p
.
1
0
0
0
9
6
,
N
o
v
.
2
0
2
2
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
h
e
a
l
t
h
.
2
0
2
2
.
1
0
0
0
9
6
.
[
1
1
]
S
.
S
i
n
g
h
,
P
.
K
.
S
h
a
r
ma,
S
.
Y
.
M
o
o
n
,
a
n
d
J.
H
.
P
a
r
k
,
“
A
d
v
a
n
c
e
d
l
i
g
h
t
w
e
i
g
h
t
e
n
c
r
y
p
t
i
o
n
a
l
g
o
r
i
t
h
ms
f
o
r
I
o
T
d
e
v
i
c
e
s:
s
u
r
v
e
y
,
c
h
a
l
l
e
n
g
e
s
a
n
d
s
o
l
u
t
i
o
n
s,
”
J
.
Am
b
i
e
n
t
I
n
t
e
l
l
.
H
u
m
a
n
i
z
.
C
o
m
p
u
t
.
,
v
o
l
.
1
5
,
n
o
.
2
,
p
p
.
1
6
2
5
–
1
6
4
2
,
F
e
b
.
2
0
2
4
,
d
o
i
:
1
0
.
1
0
0
7
/
s
1
2
6
5
2
-
0
1
7
-
0
4
9
4
-
4.
[
1
2
]
M
.
G
.
S
.
M
u
r
s
h
e
d
,
C
.
M
u
r
p
h
y
,
D
.
H
o
u
,
N
.
K
h
a
n
,
G
.
A
n
a
n
t
h
a
n
a
r
a
y
a
n
a
n
,
a
n
d
F
.
H
u
ssa
i
n
,
“
M
a
c
h
i
n
e
l
e
a
r
n
i
n
g
a
t
t
h
e
n
e
t
w
o
r
k
e
d
g
e
:
a
su
r
v
e
y
,
”
AC
M
C
o
m
p
u
t
.
S
u
rv
.
,
v
o
l
.
5
4
,
n
o
.
8
,
p
p
.
1
–
3
7
,
N
o
v
.
2
0
2
2
,
d
o
i
:
1
0
.
1
1
4
5
/
3
4
6
9
0
2
9
.
[
1
3
]
K
.
K
i
m,
S
.
-
J.
Ja
n
g
,
J
.
P
a
r
k
,
E.
Le
e
,
a
n
d
S
.
-
S
.
L
e
e
,
“
L
i
g
h
t
w
e
i
g
h
t
a
n
d
e
n
e
r
g
y
-
e
f
f
i
c
i
e
n
t
d
e
e
p
l
e
a
r
n
i
n
g
a
c
c
e
l
e
r
a
t
o
r
f
o
r
r
e
a
l
-
t
i
m
e
o
b
j
e
c
t
d
e
t
e
c
t
i
o
n
o
n
e
d
g
e
d
e
v
i
c
e
s
,
”
S
e
n
s
o
rs
,
v
o
l
.
2
3
,
n
o
.
3
,
p
.
1
1
8
5
,
J
a
n
.
2
0
2
3
,
d
o
i
:
1
0
.
3
3
9
0
/
s
2
3
0
3
1
1
8
5
.
[
1
4
]
A
.
H
o
w
a
r
d
e
t
a
l
.
,
“
S
e
a
r
c
h
i
n
g
f
o
r
M
o
b
i
l
e
N
e
t
v
3
,
”
i
n
2
0
1
9
I
EEE
/
C
V
F
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
C
o
m
p
u
t
e
r
Vi
s
i
o
n
(
I
C
C
V)
,
I
EEE
,
O
c
t
.
2
0
1
9
,
p
p
.
1
3
1
4
–
1
3
2
4
.
d
o
i
:
1
0
.
1
1
0
9
/
I
C
C
V
.
2
0
1
9
.
0
0
1
4
0
.
[
1
5
]
J.
Y
a
n
g
,
L.
Z
h
a
n
g
,
X
.
Ta
n
g
,
a
n
d
M
.
H
a
n
,
“
C
o
d
n
N
e
t
:
a
l
i
g
h
t
w
e
i
g
h
t
C
N
N
a
r
c
h
i
t
e
c
t
u
r
e
f
o
r
d
e
t
e
c
t
i
o
n
o
f
C
O
V
I
D
-
1
9
i
n
f
e
c
t
i
o
n
,
”
A
p
p
l
.
S
o
f
t
C
o
m
p
u
t
.
,
v
o
l
.
1
3
0
,
p
.
1
0
9
6
5
6
,
N
o
v
.
2
0
2
2
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
a
s
o
c
.
2
0
2
2
.
1
0
9
6
5
6
.
[
1
6
]
M
.
Ta
n
a
n
d
Q
.
V
.
L
e
,
“
Ef
f
i
c
i
e
n
t
N
e
t
:
r
e
t
h
i
n
k
i
n
g
m
o
d
e
l
sc
a
l
i
n
g
f
o
r
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,
”
i
n
Pr
o
c
e
e
d
i
n
g
s
o
f
t
h
e
3
6
t
h
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
r
e
n
c
e
o
n
Ma
c
h
i
n
e
L
e
a
r
n
i
n
g
,
L
o
n
g
B
e
a
c
h
,
C
a
l
i
f
o
r
n
i
a
:
P
M
L
R
9
7
,
M
a
y
2
0
1
9
.
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s/
1
9
0
5
.
1
1
9
4
6
[
1
7
]
N
.
M
a
,
X
.
Z
h
a
n
g
,
H
.
-
T.
Z
h
e
n
g
,
a
n
d
J.
S
u
n
,
“
S
h
u
f
f
l
e
N
e
t
v
2
:
p
r
a
c
t
i
c
a
l
g
u
i
d
e
l
i
n
e
s
f
o
r
e
f
f
i
c
i
e
n
t
C
N
N
a
r
c
h
i
t
e
c
t
u
r
e
d
e
si
g
n
,
”
i
n
L
e
c
t
u
re
N
o
t
e
s
i
n
C
o
m
p
u
t
e
r
S
c
i
e
n
c
e
,
C
o
m
p
u
t
e
r
V
.
,
S
p
r
i
n
g
e
r
,
C
h
a
m,
2
0
1
8
,
p
p
.
1
2
2
–
1
3
8
.
d
o
i
:
1
0
.
1
0
0
7
/
9
7
8
-
3
-
0
3
0
-
0
1
2
6
4
-
9
_
8
.
[
1
8
]
X
.
Zh
a
n
g
,
X
.
Zh
o
u
,
M
.
L
i
n
,
a
n
d
J.
S
u
n
,
“
S
h
u
f
f
l
e
N
e
t
:
a
n
e
x
t
r
e
m
e
l
y
e
f
f
i
c
i
e
n
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
f
o
r
mo
b
i
l
e
d
e
v
i
c
e
s,
”
i
n
2
0
1
8
I
EEE
/
C
V
F
C
o
n
f
e
r
e
n
c
e
o
n
C
o
m
p
u
t
e
r
V
i
si
o
n
a
n
d
Pa
t
t
e
rn
Re
c
o
g
n
i
t
i
o
n
,
I
EEE,
J
u
n
.
2
0
1
8
,
p
p
.
6
8
4
8
–
6
8
5
6
.
d
o
i
:
1
0
.
1
1
0
9
/
C
V
P
R
.
2
0
1
8
.
0
0
7
1
6
.
[
1
9
]
Y
.
L
i
e
t
a
l
.
,
“
M
i
c
r
o
N
e
t
:
i
m
p
r
o
v
i
n
g
i
m
a
g
e
r
e
c
o
g
n
i
t
i
o
n
w
i
t
h
e
x
t
r
e
m
e
l
y
l
o
w
f
l
o
p
s
,
”
i
n
2
0
2
1
I
EE
E/
C
VF
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
C
o
m
p
u
t
e
r
Vi
s
i
o
n
(
I
C
C
V)
,
I
EEE,
O
c
t
.
2
0
2
1
,
p
p
.
4
5
8
–
4
6
7
.
d
o
i
:
1
0
.
1
1
0
9
/
I
C
C
V
4
8
9
2
2
.
2
0
2
1
.
0
0
0
5
2
.
[
2
0
]
K
.
D
e
y
,
M
.
M
.
H
a
ssa
n
,
M
.
M
.
R
a
n
a
,
a
n
d
M
.
H
.
H
e
n
a
,
“
B
a
n
g
l
a
d
e
sh
i
i
n
d
i
g
e
n
o
u
s
f
i
s
h
c
l
a
ss
i
f
i
c
a
t
i
o
n
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,”
i
n
2
0
2
1
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
n
f
e
re
n
c
e
o
n
I
n
f
o
rm
a
t
i
o
n
T
e
c
h
n
o
l
o
g
y
(
I
C
I
T
)
,
I
EEE,
Ju
l
.
2
0
2
1
,
p
p
.
8
9
9
–
9
0
4
.
d
o
i
:
1
0
.
1
1
0
9
/
I
C
I
T5
2
6
8
2
.
2
0
2
1
.
9
4
9
1
6
8
1
.
[
2
1
]
U
.
K
u
l
k
a
r
n
i
,
M
.
S
.
M
.
,
S
.
V
.
G
u
r
l
a
h
o
s
u
r
,
a
n
d
G
.
B
h
o
g
a
r
,
“
Q
u
a
n
t
i
z
a
t
i
o
n
f
r
i
e
n
d
l
y
M
o
b
i
l
e
N
e
t
(
q
f
-
M
o
b
i
l
e
N
e
t
)
a
r
c
h
i
t
e
c
t
u
r
e
f
o
r
v
i
s
i
o
n
b
a
s
e
d
a
p
p
l
i
c
a
t
i
o
n
s
o
n
e
m
b
e
d
d
e
d
p
l
a
t
f
o
r
ms,”
N
e
u
r
a
l
N
e
t
w
o
rks
,
v
o
l
.
1
3
6
,
p
p
.
2
8
–
3
9
,
A
p
r
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
n
e
u
n
e
t
.
2
0
2
0
.
1
2
.
0
2
2
.
[
2
2
]
G
.
L
u
,
W
.
Z
h
a
n
g
,
a
n
d
Z.
W
a
n
g
,
“
O
p
t
i
mi
z
i
n
g
d
e
p
t
h
w
i
se
se
p
a
r
a
b
l
e
c
o
n
v
o
l
u
t
i
o
n
o
p
e
r
a
t
i
o
n
s
o
n
G
P
U
s,”
I
E
E
E
T
ra
n
s.
P
a
ra
l
l
e
l
D
i
st
ri
b
.
S
y
st
.
,
v
o
l
.
3
3
,
n
o
.
1
,
p
p
.
7
0
–
8
7
,
J
a
n
.
2
0
2
2
,
d
o
i
:
1
0
.
1
1
0
9
/
TPD
S
.
2
0
2
1
.
3
0
8
4
8
1
3
.
[
2
3
]
K
.
M
a
h
a
r
a
n
a
,
S
.
M
o
n
d
a
l
,
a
n
d
B
.
N
e
mad
e
,
“
A
r
e
v
i
e
w
:
d
a
t
a
p
r
e
-
p
r
o
c
e
s
si
n
g
a
n
d
d
a
t
a
a
u
g
m
e
n
t
a
t
i
o
n
t
e
c
h
n
i
q
u
e
s,
”
G
l
o
b
.
T
r
a
n
si
t
i
o
n
s
Pro
c
.
,
v
o
l
.
3
,
n
o
.
1
,
p
p
.
9
1
–
9
9
,
Ju
n
.
2
0
2
2
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
g
l
t
p
.
2
0
2
2
.
0
4
.
0
2
0
.
[
2
4
]
N
.
E.
K
h
a
l
i
f
a
,
M
.
L
o
e
y
,
a
n
d
S
.
M
i
r
j
a
l
i
l
i
,
“
A
c
o
m
p
r
e
h
e
n
si
v
e
s
u
r
v
e
y
o
f
r
e
c
e
n
t
t
r
e
n
d
s
i
n
d
e
e
p
l
e
a
r
n
i
n
g
f
o
r
d
i
g
i
t
a
l
i
m
a
g
e
s
a
u
g
me
n
t
a
t
i
o
n
,
”
A
rt
i
f
.
I
n
t
e
l
l
.
Re
v
.
,
v
o
l
.
5
5
,
n
o
.
3
,
p
p
.
2
3
5
1
–
2
3
7
7
,
M
a
r
.
2
0
2
2
,
d
o
i
:
1
0
.
1
0
0
7
/
s1
0
4
6
2
-
0
2
1
-
1
0
0
6
6
-
4.
[
2
5
]
S
.
B
o
u
g
u
e
z
z
i
,
H
.
B
e
n
F
r
e
d
j
,
T.
B
e
l
a
b
e
d
,
C
.
V
a
l
d
e
r
r
a
m
a
,
H
.
F
a
i
e
d
h
,
a
n
d
C
.
S
o
u
a
n
i
,
“
A
n
e
f
f
i
c
i
e
n
t
f
p
g
a
-
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
f
o
r
c
l
a
ssi
f
i
c
a
t
i
o
n
:
a
d
-
M
o
b
i
l
e
N
e
t
,
”
E
l
e
c
t
r
o
n
i
c
s
,
v
o
l
.
1
0
,
n
o
.
1
8
,
p
.
2
2
7
2
,
S
e
p
.
2
0
2
1
,
d
o
i
:
1
0
.
3
3
9
0
/
e
l
e
c
t
r
o
n
i
c
s1
0
1
8
2
2
7
2
.
[
2
6
]
K
.
K
C
,
Z.
Y
i
n
,
M
.
W
u
,
a
n
d
Z.
W
u
,
“
D
e
p
t
h
w
i
se
s
e
p
a
r
a
b
l
e
c
o
n
v
o
l
u
t
i
o
n
a
r
c
h
i
t
e
c
t
u
r
e
s
f
o
r
p
l
a
n
t
d
i
se
a
se
c
l
a
ssi
f
i
c
a
t
i
o
n
,
”
C
o
m
p
u
t
.
El
e
c
t
r
o
n
.
A
g
r
i
c
.
,
v
o
l
.
1
6
5
,
p
.
1
0
4
9
4
8
,
O
c
t
.
2
0
1
9
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
c
o
m
p
a
g
.
2
0
1
9
.
1
0
4
9
4
8
.
[
2
7
]
W
.
W
a
n
g
,
Y
.
H
u
,
T.
Z
o
u
,
H
.
Li
u
,
J.
W
a
n
g
,
a
n
d
X
.
W
a
n
g
,
“
A
n
e
w
i
ma
g
e
c
l
a
ss
i
f
i
c
a
t
i
o
n
a
p
p
r
o
a
c
h
v
i
a
i
mp
r
o
v
e
d
M
o
b
i
l
e
N
e
t
mo
d
e
l
s
w
i
t
h
l
o
c
a
l
r
e
c
e
p
t
i
v
e
f
i
e
l
d
e
x
p
a
n
si
o
n
i
n
sh
a
l
l
o
w
l
a
y
e
r
s,”
C
o
m
p
u
t
.
I
n
t
e
l
l
.
N
e
u
r
o
sc
i
.
,
v
o
l
.
2
0
2
0
,
p
p
.
1
–
1
0
,
A
u
g
.
2
0
2
0
,
d
o
i
:
1
0
.
1
1
5
5
/
2
0
2
0
/
8
8
1
7
8
4
9
.
[
2
8
]
A
.
G
.
H
o
w
a
r
d
e
t
a
l
.
,
“
M
o
b
i
l
e
N
e
t
s
:
e
f
f
i
c
i
e
n
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
s
f
o
r
m
o
b
i
l
e
v
i
s
i
o
n
a
p
p
l
i
c
a
t
i
o
n
s,
”
a
rXi
v
,
A
p
r
.
2
0
1
7
,
[
O
n
l
i
n
e
]
.
A
v
a
i
l
a
b
l
e
:
h
t
t
p
:
/
/
a
r
x
i
v
.
o
r
g
/
a
b
s/
1
7
0
4
.
0
4
8
6
1
[
2
9
]
W
.
W
a
n
g
,
Y
.
L
i
,
T
.
Z
o
u
,
X
.
W
a
n
g
,
J
.
Y
o
u
,
a
n
d
Y
.
L
u
o
,
“
A
n
o
v
e
l
i
ma
g
e
c
l
a
ss
i
f
i
c
a
t
i
o
n
a
p
p
r
o
a
c
h
v
i
a
d
e
n
s
e
-
M
o
b
i
l
e
N
e
t
m
o
d
e
l
s
,
”
Mo
b
.
I
n
f
.
S
y
s
t
.
,
v
o
l
.
2
0
2
0
,
p
p
.
1
–
8
,
J
a
n
.
2
0
2
0
,
d
o
i
:
1
0
.
1
1
5
5
/
2
0
2
0
/
7
6
0
2
3
8
4
.
[
3
0
]
S
.
W
.
P
.
Li
st
i
o
,
“
P
e
r
f
o
r
m
a
n
c
e
o
f
d
e
e
p
l
e
a
r
n
i
n
g
i
n
c
e
p
t
i
o
n
m
o
d
e
l
a
n
d
M
o
b
i
l
e
N
e
t
mo
d
e
l
o
n
g
e
n
d
e
r
p
r
e
d
i
c
t
i
o
n
t
h
r
o
u
g
h
e
y
e
i
ma
g
e
,
”
S
i
n
k
r
o
n
,
v
o
l
.
7
,
n
o
.
4
,
p
p
.
2
5
9
3
–
2
6
0
1
,
N
o
v
.
2
0
2
2
,
d
o
i
:
1
0
.
3
3
3
9
5
/
si
n
k
r
o
n
.
v
7
i
4
.
1
1
8
8
7
.
[
3
1
]
S
.
K
a
u
r
,
H
.
A
g
g
a
r
w
a
l
,
a
n
d
R
.
R
a
n
i
,
“
H
y
p
e
r
-
p
a
r
a
met
e
r
o
p
t
i
mi
z
a
t
i
o
n
o
f
d
e
e
p
l
e
a
r
n
i
n
g
m
o
d
e
l
f
o
r
p
r
e
d
i
c
t
i
o
n
o
f
P
a
r
k
i
n
s
o
n
’
s
d
i
s
e
a
se
,
”
Ma
c
h
.
V
i
s.
Ap
p
l
.
,
v
o
l
.
3
1
,
n
o
.
5
,
p
.
3
2
,
Ju
l
.
2
0
2
0
,
d
o
i
:
1
0
.
1
0
0
7
/
s
0
0
1
3
8
-
0
2
0
-
0
1
0
7
8
-
1.
[
3
2
]
Y
.
N
.
K
u
n
a
n
g
,
S
.
N
u
r
ma
i
n
i
,
D
.
S
t
i
a
w
a
n
,
a
n
d
B
.
Y
.
S
u
p
r
a
p
t
o
,
“
A
t
t
a
c
k
c
l
a
ssi
f
i
c
a
t
i
o
n
o
f
a
n
i
n
t
r
u
si
o
n
d
e
t
e
c
t
i
o
n
s
y
st
e
m
u
si
n
g
d
e
e
p
l
e
a
r
n
i
n
g
a
n
d
h
y
p
e
r
p
a
r
a
met
e
r
o
p
t
i
m
i
z
a
t
i
o
n
,
”
J
.
I
n
f
.
S
e
c
u
r
.
A
p
p
l
.
,
v
o
l
.
5
8
,
p
.
1
0
2
8
0
4
,
M
a
y
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
j
i
sa.
2
0
2
1
.
1
0
2
8
0
4
.
[3
3]
C
.
H
a
,
V
.
-
D
.
Tr
a
n
,
L.
N
g
o
V
a
n
,
a
n
d
K
.
Th
a
n
,
“
E
l
i
mi
n
a
t
i
n
g
o
v
e
r
f
i
t
t
i
n
g
o
f
p
r
o
b
a
b
i
l
i
s
t
i
c
t
o
p
i
c
mo
d
e
l
s
o
n
sh
o
r
t
a
n
d
n
o
i
s
y
t
e
x
t
:
t
h
e
r
o
l
e
o
f
d
r
o
p
o
u
t
,
”
I
n
t
.
J
.
A
p
p
r
o
x
.
R
e
a
s
o
n
.
,
v
o
l
.
1
1
2
,
p
p
.
8
5
–
1
0
4
,
S
e
p
.
2
0
1
9
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
i
j
a
r
.
2
0
1
9
.
0
5
.
0
1
0
.
[
3
4
]
J.
Z
h
a
n
g
,
Y
.
S
u
n
,
L.
G
u
o
,
H
.
G
a
o
,
X
.
H
o
n
g
,
a
n
d
H
.
S
o
n
g
,
“
A
n
e
w
b
e
a
r
i
n
g
f
a
u
l
t
d
i
a
g
n
o
s
i
s
me
t
h
o
d
b
a
sed
o
n
m
o
d
i
f
i
e
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,
”
C
h
i
n
e
se
J
.
Ae
r
o
n
a
u
t
.
,
v
o
l
.
3
3
,
n
o
.
2
,
p
p
.
4
3
9
–
4
4
7
,
F
e
b
.
2
0
2
0
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
c
j
a
.
2
0
1
9
.
0
7
.
0
1
1
.
[
3
5
]
J.
H
y
u
n
,
H
.
S
e
o
n
g
,
a
n
d
E
.
K
i
m,
“
U
n
i
v
e
r
sa
l
p
o
o
l
i
n
g
–
a
n
e
w
p
o
o
l
i
n
g
m
e
t
h
o
d
f
o
r
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,”
E
x
p
e
rt
S
y
st
.
Ap
p
l
.
,
v
o
l
.
1
8
0
,
p
.
1
1
5
0
8
4
,
O
c
t
.
2
0
2
1
,
d
o
i
:
1
0
.
1
0
1
6
/
j
.
e
sw
a
.
2
0
2
1
.
1
1
5
0
8
4
.
[
3
6
]
S
.
S
.
H
u
s
a
i
n
a
n
d
M
.
B
o
b
e
r
,
“
R
E
M
A
P
:
mu
l
t
i
-
l
a
y
e
r
e
n
t
r
o
p
y
-
g
u
i
d
e
d
p
o
o
l
i
n
g
o
f
d
e
n
se
C
N
N
f
e
a
t
u
r
e
s
f
o
r
i
m
a
g
e
r
e
t
r
i
e
v
a
l
,
”
I
EE
E
T
ra
n
s
.
I
m
a
g
e
Pr
o
c
e
ss.
,
v
o
l
.
2
8
,
n
o
.
1
0
,
p
p
.
5
2
0
1
–
5
2
1
3
,
O
c
t
.
2
0
1
9
,
d
o
i
:
1
0
.
1
1
0
9
/
TI
P
.
2
0
1
9
.
2
9
1
7
2
3
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.