I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
pu
t
er
E
ng
ineering
(
I
J
E
CE
)
Vo
l
.
11
,
No
.
2
,
A
p
r
il
2
0
2
1
,
p
p
.
1
5
1
0
~1
5
2
0
I
SS
N:
2
0
8
8
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/
ijece
.
v
1
1
i
2
.
p
p
1
5
1
0
-
1
5
2
0
1510
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Rea
l
-
time
huma
n
detec
tion for
elect
ricity con
serv
a
tio
n using
pruned
-
SSD and
a
rduino
Ush
a
s
uk
ha
ny
a
S
.
1
,
J
o
t
hil
a
k
s
hm
i S
.
2
1
De
p
a
rtme
n
t
o
f
Co
m
p
u
ter
S
c
ien
c
e
a
n
d
E
n
g
i
n
e
e
rin
g
,
An
n
a
m
a
lai
Un
iv
e
rsity
,
Tam
il
n
a
d
u
,
I
n
d
ia
2
De
p
a
rtme
n
t
o
f
In
fo
rm
a
ti
o
n
Tec
h
n
o
l
o
g
y
,
An
n
a
m
a
la
i
Un
iv
e
rsit
y
,
T
a
m
il
n
a
d
u
,
I
n
d
ia
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
u
l 1
,
2
0
20
R
ev
is
ed
Au
g
2
,
2
0
2
0
Acc
ep
ted
Sep
2
4
,
2
0
2
0
El
e
c
tri
c
it
y
c
o
n
se
rv
a
ti
o
n
tec
h
n
iq
u
e
s
h
a
v
e
g
a
in
e
d
m
o
re
imp
o
rta
n
c
e
in
re
c
e
n
t
y
e
a
rs.
M
a
n
y
sm
a
rt
tec
h
n
i
q
u
e
s
a
re
in
v
e
n
ted
t
o
sa
v
e
e
lec
tri
c
it
y
with
t
h
e
h
e
l
p
o
f
a
ss
isted
d
e
v
ice
s
li
k
e
se
n
so
rs.
T
h
o
u
g
h
it
sa
v
e
s
e
lec
tri
c
it
y
,
it
a
d
d
s
a
n
a
d
d
it
i
o
n
a
l
se
n
so
r
c
o
st
to
th
e
sy
ste
m
.
Th
is
wo
rk
a
ims
to
d
e
v
e
l
o
p
a
sy
ste
m
th
a
t
m
a
n
a
g
e
s
th
e
e
lec
tri
c
p
o
we
r
su
p
p
l
y
,
o
n
ly
wh
e
n
it
is
a
c
tu
a
l
ly
n
e
e
d
e
d
i.
e
.
,
t
h
e
sy
ste
m
e
n
a
b
les
th
e
p
o
we
r
su
p
p
l
y
wh
e
n
a
h
u
m
a
n
is
p
re
se
n
t
in
t
h
e
lo
c
a
ti
o
n
a
n
d
d
isa
b
les
it
o
t
h
e
rwise
.
Th
e
sy
ste
m
a
v
o
id
s
a
n
y
a
d
d
i
ti
o
n
a
l
c
o
sts
b
y
u
sin
g
t
h
e
c
lo
se
d
c
ircu
it
tele
v
isio
n
,
w
h
ich
is
in
sta
ll
e
d
i
n
m
o
st
o
f
th
e
p
lac
e
s
f
o
r
se
c
u
rit
y
re
a
so
n
s.
Hu
m
a
n
d
e
t
e
c
ti
o
n
is
d
o
n
e
b
y
a
m
o
d
ifi
e
d
-
si
n
g
le
sh
o
t
d
e
tec
ti
o
n
wit
h
a
sp
e
c
ifi
c
h
y
p
e
r
p
a
ra
m
e
ter
tu
n
in
g
m
e
th
o
d
.
F
u
rt
h
e
r
t
h
e
m
o
d
e
l
is
p
ru
n
e
d
t
o
re
d
u
c
e
th
e
c
o
m
p
u
tatio
n
a
l
c
o
st
o
f
th
e
fra
m
e
wo
rk
wh
ic
h
i
n
tu
r
n
re
d
u
c
e
s
th
e
p
ro
c
e
ss
in
g
s
p
e
e
d
o
f
th
e
n
e
tw
o
rk
d
ra
stica
ll
y
.
T
h
e
m
o
d
e
l
y
ield
s
t
h
e
o
u
t
p
u
t
to
th
e
Ard
u
i
n
o
m
icro
-
c
o
n
tro
ll
e
r
to
e
n
a
b
le
th
e
p
o
we
r
su
p
p
l
y
in
a
n
d
a
ro
u
n
d
th
e
lo
c
a
ti
o
n
o
n
ly
wh
e
n
a
h
u
m
a
n
is
d
e
tec
ted
a
n
d
d
isa
b
les
it
wh
e
n
t
h
e
h
u
m
a
n
e
x
it
s.
Th
e
m
o
d
e
l
is
e
v
a
lu
a
ted
o
n
C
HO
KEP
OIN
T
d
a
tas
e
t
a
n
d
re
a
l
-
ti
m
e
v
id
e
o
su
rv
e
il
lan
c
e
f
o
o
tag
e
.
Ex
p
e
rime
n
t
a
l
re
su
lt
s
h
a
v
e
sh
o
w
n
a
n
a
v
e
ra
g
e
a
c
c
u
ra
c
y
o
f
8
5
.
8
2
%
with
2
.
1
se
c
o
n
d
s o
f
p
r
o
c
e
ss
in
g
ti
m
e
p
e
r
fra
m
e
.
K
ey
w
o
r
d
s
:
Ar
d
u
in
o
E
lectr
ic
ity
m
an
ag
em
en
t
Hu
m
an
d
etec
tio
n
Pru
n
in
g
Sin
g
le
s
h
o
t d
etec
tio
n
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Ush
asu
k
h
an
y
a
S.
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
An
n
am
alai
Un
iv
er
s
ity
An
n
am
alai
n
ag
ar
,
T
am
iln
ad
u
,
I
n
d
ia
E
m
ail:
u
s
h
asu
k
h
an
y
a@
g
m
ail.
co
m
1.
I
NT
RO
D
U
CT
I
O
N
Un
m
an
n
ed
elec
tr
ic
p
o
wer
m
a
n
ag
em
en
t
s
y
s
tem
(
UE
PMS)
is
o
n
e
o
f
th
e
ch
allen
g
in
g
task
s
th
at
ar
e
r
eq
u
ir
ed
t
o
s
av
e
elec
tr
ical
r
eso
u
r
ce
s
in
m
o
s
t
o
f
th
e
c
o
u
n
tr
i
es.
Sev
er
al
r
esear
ch
es
ar
e
g
o
i
n
g
o
n
f
o
r
UE
PMS
with
th
e
h
elp
o
f
s
en
s
o
r
s
to
d
etec
t
th
e
p
r
esen
ce
/ab
s
en
ce
o
f
h
u
m
a
n
s
an
d
to
m
a
n
ag
e
th
e
p
o
wer
s
u
p
p
l
y
ac
co
r
d
in
g
l
y
.
Usag
e
o
f
s
en
s
o
r
s
in
cu
r
s
an
ad
d
itio
n
al
co
s
t
an
d
h
en
ce
t
h
is
wo
r
k
aim
s
to
d
e
v
elo
p
a
s
y
s
tem
th
at
m
an
ag
es
th
e
elec
tr
ical
r
eso
u
r
c
e
u
s
in
g
th
e
ex
is
tin
g
clo
s
ed
cir
cu
it
telev
is
io
n
(
C
C
T
V)
s
u
r
v
eillan
ce
ca
m
er
a.
T
h
e
s
u
r
v
eillan
ce
v
id
eo
ca
p
tu
r
ed
f
r
o
m
C
C
T
V
ca
m
er
as
is
u
s
e
d
to
d
etec
t
th
e
h
u
m
an
’
s
p
r
esen
ce
/ab
s
en
ce
to
en
ab
le/d
is
ab
le
th
e
p
o
wer
s
u
p
p
ly
th
e
r
e
b
y
a
v
o
id
in
g
ad
d
itio
n
al
co
s
t.
Hu
m
an
d
etec
tio
n
i
n
s
u
r
v
eillan
ce
ca
m
e
r
as
f
o
o
tag
e
h
as
b
ee
n
an
in
te
r
esti
n
g
[
1
]
a
n
d
ch
allen
g
in
g
[
2
]
to
p
ic
in
th
e
r
ec
en
t
y
ea
r
s
.
T
h
e
tr
ad
itio
n
al
h
an
d
-
c
r
af
te
d
m
eth
o
d
s
lik
e
lo
ca
l b
in
ar
y
p
att
er
n
(
L
B
P),
h
is
to
g
r
am
o
f
o
r
ien
t
ed
g
r
ad
ie
n
ts
(
HOG)
etc.
,
ar
e
ti
m
e
co
n
s
u
m
in
g
an
d
p
r
o
v
e
d
t
o
b
e
c
o
m
p
a
r
a
t
i
v
e
l
y
i
n
e
f
f
i
c
i
e
n
t
t
o
t
h
e
r
e
c
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
(
C
N
N
)
b
a
s
e
d
a
l
g
o
r
i
t
h
m
s
[
3
]
.
Var
io
u
s
o
b
ject
d
etec
tio
n
alg
o
r
ith
m
s
in
d
ee
p
lear
n
in
g
(
DL
)
h
av
e
s
h
o
wn
p
r
o
m
is
in
g
r
esu
lts
in
c
lass
if
y
in
g
an
d
d
etec
tin
g
th
e
l
o
ca
tio
n
o
f
th
e
o
b
jects
[
4
]
.
T
h
e
f
i
r
s
t
ca
teg
o
r
y
o
f
DNN
is
a
two
s
tag
e
ap
p
r
o
ac
h
lik
e
R
C
NN,
f
aster
-
r
eg
io
n
al
C
NN
(
f
a
s
ter
R
-
C
NN)
,
r
eg
io
n
b
ased
f
u
lly
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(R
-
FC
N)
[
5
,
6
]
etc.
,
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
R
ea
l
-
time
h
u
ma
n
d
etec
tio
n
fo
r
elec
tr
icity
co
n
s
erva
tio
n
u
s
in
g
p
r
u
n
ed
-
S
S
D
a
n
d
a
r
d
u
in
o
(
Ush
a
s
u
kh
a
n
ya
S
.
)
1511
wh
ich
p
r
o
p
o
s
es
r
eg
io
n
s
b
y
a
s
ep
a
r
ate
n
etwo
r
k
ca
lled
r
eg
io
n
p
r
o
p
o
s
al
n
etwo
r
k
an
d
t
h
e
class
if
ier
p
r
o
ce
s
s
es
th
ese
r
eg
io
n
s
f
o
r
class
if
icatio
n
[
7
]
.
T
h
e
s
ec
o
n
d
o
n
e
is
a
o
n
e
s
tag
e
a
p
p
r
o
ac
h
lik
e
y
o
u
o
n
ly
lo
o
k
o
n
ce
(
YOL
O)
an
d
s
in
g
le
s
h
o
t
m
u
ltib
o
x
d
ete
cto
r
(
SS
D)
in
wh
ich
b
o
th
t
h
e
class
p
r
o
b
ab
ilit
i
es
an
d
th
e
b
o
u
n
d
in
g
b
o
x
es
ar
e
p
r
o
d
u
ce
d
b
y
th
e
C
NN
its
elf
[
8
]
.
I
n
th
is
wo
r
k
a
s
y
s
tem
u
s
in
g
m
o
d
if
ied
-
SS
D
wh
ich
is
b
ased
o
n
v
is
u
al
g
eo
m
etr
y
g
r
o
u
p
(
VGG1
6
)
is
u
s
ed
f
o
r
h
u
m
an
d
etec
ti
o
n
i
n
s
u
r
v
eillan
ce
ca
m
er
as
[
9
]
.
T
h
e
C
NN
b
ased
n
etwo
r
k
tak
es
th
e
in
p
u
t
f
r
o
m
C
HOK
E
P
OI
NT
d
ataset,
wh
ich
h
as
f
r
a
m
es
o
f
a
s
u
r
v
eillan
ce
v
id
eo
.
T
h
e
m
o
d
el
is
in
itially
tr
ain
ed
with
th
e
s
et
o
f
h
y
p
er
-
p
ar
am
eter
s
o
b
tain
ed
f
r
o
m
o
r
th
o
g
o
n
al
ar
r
a
y
tu
n
i
n
g
m
eth
o
d
(
OAT
M)
.
T
h
e
o
p
tim
al
f
ac
to
r
s
ar
e
d
er
iv
e
d
f
r
o
m
th
e
f
ac
to
r
tab
le
o
f
th
e
O
AT
M
m
eth
o
d
.
On
ce
th
e
m
o
d
el
is
co
n
v
er
g
ed
to
th
e
m
in
im
al
lo
s
s
f
u
n
ctio
n
,
th
e
n
etwo
r
k
is
p
r
u
n
e
d
to
r
em
o
v
e
th
e
less
im
p
o
r
tan
t
p
ar
am
eter
s
o
f
th
e
n
etwo
r
k
.
Pru
n
in
g
is
a
m
eth
o
d
d
o
n
e
to
r
ed
u
ce
th
e
co
m
p
lex
ity
o
f
th
e
n
etwo
r
k
th
e
r
eb
y
m
ain
tain
in
g
th
e
ac
cu
r
ac
y
o
f
th
e
m
o
d
el.
T
h
e
id
en
tific
at
io
n
o
f
th
ese
less
im
p
o
r
tan
t
weig
h
ts
ar
e
d
o
n
e
b
y
th
e
H
-
r
a
n
k
i
n
g
al
g
o
r
ith
m
p
r
o
p
o
s
ed
b
y
L
i
n
et
a
l
.
Af
ter
p
r
u
n
i
n
g
,
th
e
n
etwo
r
k
is
r
etr
ain
ed
with
th
e
s
et
o
f
h
y
p
e
r
p
ar
am
ete
r
s
o
b
t
ain
ed
f
r
o
m
th
e
OAT
M
m
eth
o
d
.
T
h
e
p
r
o
ce
s
s
is
iter
ated
u
n
til
th
e
c
o
n
v
e
r
g
en
ce
o
r
th
e
er
r
o
r
l
o
s
s
f
u
n
ctio
n
is
s
im
ilar
to
th
e
o
n
e
o
b
tain
ed
b
y
th
e
m
o
d
el
b
e
f
o
r
e
p
r
u
n
in
g
is
d
o
n
e.
T
h
e
o
u
tp
u
t
o
f
th
e
class
if
ier
is
f
ed
to
an
Ar
d
u
in
o
m
icr
o
co
n
t
r
o
ller
to
m
an
ag
e
th
e
p
o
we
r
s
u
p
p
ly
.
Ar
d
u
i
n
o
en
ab
les
th
e
p
o
wer
s
u
p
p
ly
o
n
l
y
in
th
e
lo
ca
tio
n
(
2
.
1
m
etr
es)
in
an
d
ar
o
u
n
d
wh
er
e
a
h
u
m
an
is
d
etec
ted
an
d
d
is
ab
les
th
e
p
o
wer
wh
en
u
n
d
etec
ted
.
T
h
e
s
y
s
tem
is
also
v
alid
ated
o
n
a
r
e
al
-
tim
e
d
ataset
o
f
a
s
u
r
v
eillan
ce
v
id
e
o
in
an
in
d
o
o
r
en
v
i
r
o
n
m
e
n
t
o
f
a
liv
in
g
r
o
o
m
wh
er
e
th
e
f
o
o
ta
g
e
is
co
n
v
er
ted
to
f
r
am
es
at
th
e
r
ate
o
f
5
p
er
s
ec
o
n
d
.
T
h
e
in
te
n
s
ely
co
m
p
r
ess
ed
m
o
d
el
s
h
o
ws
p
r
o
m
is
in
g
r
esu
lts
in
p
r
ed
ictio
n
ac
cu
r
ac
y
with
r
ed
u
ce
d
t
r
ain
in
g
tim
e.
2.
RE
L
AT
E
D
WO
RK
S
Ob
ject
d
etec
tio
n
h
as
g
ain
e
d
a
lo
t
o
f
attr
ac
tio
n
b
y
th
e
r
esear
ch
er
s
in
v
ar
io
u
s
ap
p
licatio
n
s
.
Fro
m
s
m
all
cr
ac
k
d
etec
tio
n
to
h
u
m
an
d
etec
tio
n
,
lesi
o
n
d
etec
tio
n
[
1
0
]
to
d
etec
tio
n
f
r
o
m
s
atellite
im
ag
es
[
1
1
]
etc.
Ov
er
co
m
in
g
t
h
e
s
h
o
r
tf
alls
o
f
tr
ad
itio
n
al
h
an
d
-
cr
af
ted
m
et
h
o
d
s
,
DL
h
as
ac
h
iev
ed
en
o
r
m
o
u
s
g
r
o
wth
in
th
ese
d
etec
tio
n
s
.
Ob
ject
d
etec
tio
n
in
s
u
r
v
eillan
ce
v
id
eo
s
is
o
n
e
o
f
th
e
m
o
s
t
ch
a
llen
g
in
g
task
s
d
u
e
to
th
e
lack
o
f
clar
ity
in
f
r
am
es.
I
t
all
s
tar
ted
with
th
e
r
eg
io
n
al
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
(R
-
C
NN)
w
h
ich
u
s
es
s
elec
tiv
e
s
ea
r
ch
to
d
etec
t
th
e
lo
ca
tio
n
o
f
th
e
o
b
ject
with
a
b
o
u
n
d
in
g
-
b
o
x
[
1
2
]
.
Ap
p
r
o
x
im
ately
2
0
0
0
ca
n
d
id
ate
r
eg
i
o
n
s
ar
e
e
x
tr
ac
ted
in
th
is
p
r
o
ce
s
s
wh
ich
is
ex
ten
s
iv
ely
tim
e
c
o
n
s
u
m
in
g
.
T
h
is
is
o
v
er
c
o
m
e
b
y
s
p
atial
p
y
r
am
id
p
o
o
lin
g
n
et
(
SP
P
-
Net)
o
n
th
e
f
ea
tu
r
e
m
ap
s
[
1
3
]
an
d
f
u
r
t
h
er
b
y
f
aster
-
R
C
NN
[
1
4
]
wh
ich
u
s
es
a
s
ep
ar
ate
n
etwo
r
k
ca
lled
r
eg
i
o
n
p
r
o
p
o
s
al
n
etwo
r
k
(
R
PN)
to
g
en
er
ate
c
an
d
id
ate
r
eg
io
n
s
.
A
s
y
s
tem
f
o
r
f
ac
ial
ex
p
r
ess
io
n
r
ec
o
g
n
itio
n
was d
ev
elo
p
e
d
f
o
r
an
em
o
tio
n
al
au
d
io
an
d
v
id
eo
d
ata
[
1
5
]
u
s
in
g
f
aster
R
-
C
NN
.
Sev
er
al
m
o
d
if
ied
v
er
s
io
n
s
o
f
f
aster
R
-
C
NN
wer
e
d
ev
elo
p
e
d
to
im
p
r
o
v
e
th
e
p
r
ed
ictio
n
a
cc
u
r
ac
y
at
m
in
im
al
co
s
t
[
1
6
]
.
T
h
o
u
g
h
all
th
ese
tech
n
iq
u
es
h
av
e
im
p
r
o
v
e
d
th
e
ac
cu
r
ac
y
,
it
s
ti
ll
s
er
v
es
as
a
tim
e
co
n
s
u
m
in
g
task
as
it
r
eq
u
ir
es
two
s
tag
es
f
o
r
p
r
ed
ictio
n
.
T
h
i
s
was
o
v
er
co
m
e
b
y
th
e
YOL
O
m
o
d
el
wh
er
e
t
h
e
en
tire
p
r
o
ce
s
s
is
ca
r
r
ied
o
u
t
b
y
a
s
in
g
le
n
e
u
r
al
n
etwo
r
k
th
at
m
a
k
es
o
p
tim
izatio
n
q
u
ite
ea
s
ier
[
1
7
]
.
A
n
ad
v
an
ce
d
v
er
s
io
n
o
f
o
n
e
s
tag
e
ap
p
r
o
ac
h
is
a
s
in
g
le
SS
D
s
y
s
tem
,
wh
ich
is
ac
h
iev
in
g
p
r
o
m
i
s
in
g
r
esu
lts
in
r
ea
l
-
tim
e
s
u
r
v
e
illan
ce
v
id
eo
s
[
1
8
]
.
On
e
s
u
ch
ap
p
licatio
n
was
d
e
v
elo
p
ed
to
d
etec
t
s
m
all
o
b
jects
u
s
in
g
co
n
tex
tu
al
in
f
o
r
m
atio
n
in
SS
D
at
in
cr
ea
s
ed
s
p
ee
d
[
1
9
]
.
W
h
en
c
o
m
p
a
r
e
d
to
all
th
e
ab
o
v
e
lis
ted
d
etec
ti
o
n
alg
o
r
ith
m
s
,
SS
D
h
as
ac
h
iev
ed
r
elativ
ely
p
r
o
m
i
s
in
g
r
esu
lts
at
in
cr
ea
s
ed
s
p
ee
d
b
y
ap
p
ly
in
g
p
r
e
d
ictio
n
f
ilter
s
o
n
e
v
er
y
f
ea
tu
r
e
m
ap
p
r
o
d
u
ce
d
.
T
h
o
u
g
h
SS
D
ac
h
iev
es
g
o
o
d
r
esu
lts
,
o
n
e
o
f
th
e
m
ajo
r
ch
allen
g
i
n
g
task
s
o
f
DL
is
th
e
h
ig
h
tr
ain
in
g
tim
e
r
eq
u
ir
ed
b
y
th
e
m
o
d
el
to
lear
n
[
2
0
,
2
1
]
.
Hen
c
e
th
e
p
r
o
p
o
s
ed
wo
r
k
u
s
es
SS
D
ar
ch
itect
u
r
e
with
a
s
p
ec
if
ic
h
y
p
er
-
p
ar
am
eter
tu
n
i
n
g
m
eth
o
d
to
r
ed
u
ce
th
e
tr
ai
n
i
n
g
tim
e
o
f
th
e
n
etwo
r
k
[
2
2
]
.
As
th
e
d
e
ep
n
etwo
r
k
d
esig
n
ed
f
o
r
an
y
r
ea
l
-
tim
e
ap
p
licatio
n
in
v
o
lv
es
h
ig
h
co
m
p
u
tatio
n
al
c
o
s
t,
p
r
u
n
in
g
is
d
o
n
e
in
m
an
y
ap
p
licatio
n
s
to
k
ee
p
th
e
m
o
d
e
l
s
im
p
le.
D
ata
-
d
ep
en
d
en
t
a
n
d
d
ata
-
in
d
ep
e
n
d
en
t
m
eth
o
d
s
a
r
e
th
e
two
tech
n
iq
u
es
ad
o
p
ted
t
o
ev
alu
ate
th
e
im
p
o
r
tan
ce
o
f
t
h
e
weig
h
ts
am
o
n
g
wh
ich
o
p
tim
al
b
r
ain
d
am
a
g
e
[
2
3
]
,
o
p
tim
al
b
r
ain
s
u
r
g
eo
n
[
2
4
]
,
ab
s
o
lu
te
v
al
u
e
m
eth
o
d
[
2
5
]
,
L
ASSO
r
eg
r
ess
io
n
m
eth
o
d
[
2
6
]
ar
e
ce
r
tain
r
e
n
o
wn
ed
tech
n
i
q
u
es.
An
in
ter
esti
n
g
m
eth
o
d
b
y
Min
g
b
ao
L
in
et
a
l
.
h
as
b
ee
n
d
e
v
el
o
p
ed
wh
ich
u
s
es
a
H
-
R
an
k
f
ilter
p
r
u
n
in
g
m
eth
o
d
to
p
r
u
n
e
th
e
m
o
d
el
b
y
ca
lcu
la
tin
g
th
e
r
an
k
o
f
ea
ch
an
d
ev
er
y
p
ar
am
eter
.
T
h
e
tech
n
iq
u
e
th
en
r
e
-
ar
r
a
n
g
es
th
e
r
an
k
ed
p
ar
am
eter
s
in
d
ec
r
ea
s
in
g
o
r
d
er
an
d
f
in
ally
elim
in
ates th
e
least im
p
o
r
tan
t p
a
r
am
eter
s
.
On
e
o
f
th
e
im
p
o
r
tan
t
to
o
ls
in
liter
atu
r
e
to
co
n
tr
o
l
th
e
p
o
wer
s
u
p
p
ly
is
th
e
A
r
d
u
in
o
m
icr
o
-
co
n
tr
o
ller
wh
ich
h
as
its
u
s
ag
e
in
a
wid
e
r
an
g
e
o
f
a
p
p
licatio
n
s
.
A
s
y
s
t
em
to
q
u
an
tify
th
e
en
e
r
g
y
o
f
th
e
g
iv
en
lo
ad
a
n
d
p
lan
ap
p
r
o
p
r
iate
en
er
g
y
c
o
n
s
e
r
v
atio
n
p
o
licies
was
also
d
esig
n
ed
[
2
7
]
.
An
o
t
h
er
a
p
p
licatio
n
was
d
ev
elo
p
ed
f
o
r
s
m
ar
t
h
o
m
e
en
er
g
y
m
a
n
ag
em
en
t
s
y
s
tem
s
(
SHEM
S)
[
2
8
]
to
en
ab
le/d
is
ab
le
th
e
p
o
wer
s
u
p
p
ly
wh
en
a
h
u
m
an
is
d
etec
ted
/u
n
d
etec
ted
r
esp
ec
ti
v
ely
.
All
th
ese
s
y
s
tem
s
u
s
e
s
en
s
o
r
s
wh
ich
in
cu
r
s
ad
d
itio
n
al
co
s
t.
Hen
ce
o
u
r
s
y
s
tem
u
s
es a
m
o
d
if
ied
-
SS
D
t
o
d
e
tect
h
u
m
a
n
s
in
ex
is
tin
g
C
C
T
V
s
u
r
v
eillan
ce
f
o
o
tag
e
with
an
Ar
d
u
i
n
o
m
icr
o
-
co
n
tr
o
ller
to
m
an
ag
e
th
e
p
o
wer
s
u
p
p
ly
ac
c
o
r
d
in
g
ly
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
11
,
No
.
2
,
Ap
r
il 2
0
2
1
:
1
5
1
0
-
1520
1512
3.
P
RO
P
O
SE
D
F
RA
M
E
WO
R
K
A
mo
d
if
ied
SS
D
is
d
ev
elo
p
e
d
in
th
is
s
eg
m
en
t
with
an
o
r
th
o
g
o
n
al
ar
r
ay
b
ased
t
u
n
in
g
m
eth
o
d
to
r
ed
u
ce
th
e
tr
ain
in
g
tim
e
with
th
e
s
et
o
f
o
b
tain
ed
f
ac
to
r
v
alu
es.
T
h
e
m
o
d
el
is
f
u
r
th
er
p
r
u
n
ed
to
r
ed
u
ce
th
e
co
m
p
lex
ity
o
f
t
h
e
DNN
m
o
d
e
l
wh
ich
in
t
u
r
n
r
ed
u
ce
s
t
h
e
co
m
p
u
ta
tio
n
c
o
s
t
in
ten
s
ely
.
T
h
e
wo
r
k
in
g
m
o
d
el
o
f
th
e
p
r
o
p
o
s
ed
ar
c
h
itectu
r
e
is
g
i
v
en
in
Fig
u
r
e
1
.
T
h
e
o
u
tp
u
t
o
f
th
e
m
o
d
el
is
em
b
ed
d
e
d
with
an
Ar
d
u
in
o
m
icr
o
-
co
n
tr
o
ller
to
m
an
ag
e
th
e
p
o
wer
s
u
p
p
ly
o
f
th
e
e
n
v
ir
o
n
m
en
t.
Fig
u
r
e
1
.
W
o
r
k
i
n
g
m
o
d
el
o
f
p
r
u
n
ed
OAT
M
b
ased
SSD
3
.
1
.
Da
t
a
s
et
s
a
nd
da
t
a
a
ug
m
ent
a
t
io
n
T
h
e
p
r
o
p
o
s
ed
s
y
s
tem
is
ev
alu
ated
u
s
in
g
s
tan
d
ar
d
C
H
OKE
POI
NT
d
ataset,
co
n
s
is
tin
g
6
4
.
2
0
4
f
r
a
m
es
o
f
a
s
u
r
v
eillan
ce
v
id
e
o
in
an
in
d
o
o
r
en
v
ir
o
n
m
en
t.
T
h
e
s
y
s
tem
is
also
ev
al
u
ated
o
n
a
r
ea
l
-
tim
e
s
u
r
v
eillan
ce
v
id
eo
o
f
a
li
v
in
g
r
o
o
m
w
h
ich
is
co
n
v
er
ted
at
th
e
r
ate
o
f
5
f
r
am
es
p
er
s
ec
o
n
d
(
f
p
s
)
.
I
t
c
o
n
s
is
ts
o
f
9
0
0
0
f
r
am
es
t
o
tally
o
u
t
o
f
wh
ich
7
4
2
0
f
r
a
m
es
(
in
clu
d
in
g
au
g
m
e
n
ted
f
r
a
m
es)
ar
e
co
n
s
id
er
ed
f
o
r
tr
ain
i
n
g
an
d
th
e
r
em
ain
in
g
3
1
8
0
ar
e
c
o
n
s
id
er
ed
f
o
r
test
in
g
.
Valid
atio
n
s
et
co
n
s
is
ts
o
f
3
7
1
im
ag
es
to
tu
n
e
th
e
h
y
p
e
r
-
p
ar
a
m
eter
s
.
Au
g
m
en
tatio
n
is
also
d
o
n
e
to
m
ak
e
th
e
n
etwo
r
k
m
o
r
e
r
o
b
u
s
t.
Sh
r
in
k
in
g
an
d
cr
o
p
p
i
n
g
o
f
t
h
e
o
r
ig
i
n
al
tr
ain
in
g
im
ag
es
ar
e
d
o
n
e
.
Gam
m
a
co
r
r
ec
tio
n
is
ap
p
lied
to
cr
ea
te
v
ar
i
atio
n
s
in
in
ten
s
ities
an
d
b
r
ig
h
t
n
ess
o
f
th
e
f
r
am
es
u
s
in
g
two
s
ets
o
f
v
alu
es
(
0
.
5
an
d
3
.
0
)
a
n
d
(
0
.
5
a
n
d
2
.
0
)
f
o
r
all
th
e
th
r
ee
ch
a
n
n
els
in
d
iv
i
d
u
ally
[
2
9
]
.
Hid
e
an
d
s
ee
k
[
3
0
]
is
f
in
ally
u
s
ed
t
o
d
iv
id
e
th
e
en
tire
im
ag
e
in
t
o
p
atch
es
with
a
d
iv
is
io
n
n
u
m
b
e
r
o
f
4
to
m
ak
e
th
e
n
etwo
r
k
lear
n
th
e
f
in
e
en
t
ities
o
f
th
e
p
atch
es.
T
h
e
s
u
b
p
atch
es
ar
e
h
id
d
en
with
a
p
r
o
b
ab
ili
ty
o
f
0
.
2
5
.
I
n
b
o
th
th
e
ca
s
es
th
e
p
o
s
itiv
e
b
o
u
n
d
i
n
g
b
o
x
es
ar
e
th
e
o
n
es
with
t
h
e
m
atch
s
co
r
e
g
r
ea
ter
t
h
an
0
.
5
with
th
at
o
f
its
g
r
o
u
n
d
tr
u
th
an
d
th
e
r
em
ain
i
n
g
ar
e
co
n
s
id
er
ed
n
eg
ativ
e.
All
th
e
e
x
p
er
im
e
n
ts
ar
e
co
n
d
u
cte
d
in
GPU
-
NVI
DI
A
T
I
T
AN
X
1
2
GB
,
R
AM
-
3
2
G
B
DDR4
.
3
.
2
.
M
o
del c
o
ns
t
ruct
io
n
T
h
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
u
s
e
s
a
m
o
d
if
ied
SS
D
tech
n
iq
u
e
wh
ich
is
a
f
ee
d
-
f
o
r
war
d
c
o
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
b
ased
o
n
V
GG1
6
.
T
h
e
s
tr
u
ctu
r
e
h
as
a
tr
u
n
ca
ted
p
a
r
t
o
f
VGG
1
6
with
an
ad
d
itio
n
al
co
n
v
o
l
u
tio
n
al
s
tr
u
ctu
r
e
attac
h
ed
to
its
en
d
.
I
t e
lim
in
ates th
e
n
ee
d
f
o
r
R
PN in
f
aster
R
-
C
N
N
th
er
eb
y
in
cr
ea
s
i
n
g
th
e
d
etec
tio
n
s
p
ee
d
d
r
asti
ca
lly
[
3
1
]
.
I
n
s
tead
o
f
R
PN,
it
u
s
es
m
u
lti
-
s
ca
le
f
ea
tu
r
es
an
d
d
ef
au
lt
b
o
x
es
to
m
ak
e
its
p
r
ed
ic
tio
n
ac
cu
r
ac
y
b
e
in
p
ar
with
th
e
av
er
ag
e
ac
cu
r
ac
y
o
b
tain
ed
b
y
f
aster
R
-
C
NN
[
3
2
]
.
T
h
e
en
tire
ar
ch
itectu
r
e
o
f
th
e
p
r
o
p
o
s
ed
s
y
s
tem
is
g
iv
en
in
Fig
u
r
e
1
.
T
h
e
in
p
u
t
is
a
co
lo
u
r
im
ag
e
I
wh
ich
is
f
ed
in
t
o
th
e
in
itial
b
ase
n
etwo
r
k
co
n
s
is
tin
g
o
f
f
iv
e
co
n
v
o
lu
tio
n
al
lay
er
s
(
=
5
)
as
in
VGG
-
1
6
wh
ich
d
etec
ts
th
e
ed
g
es,
b
lo
b
s
,
tex
tu
r
e
an
d
o
b
ject
p
ar
ts
.
T
h
e
s
ix
th
a
n
d
s
ev
e
n
th
d
en
s
e
lay
e
r
s
o
f
VGG
-
1
6
ar
e
r
ep
lace
d
with
co
n
v
o
l
u
tio
n
al
la
y
er
s
wh
ile
r
e
m
o
v
in
g
th
e
eig
h
t
h
a
n
d
th
e
o
th
er
d
r
o
p
o
u
t
la
y
er
s
.
R
eL
U
is
th
e
ac
tiv
atio
n
f
u
n
ctio
n
u
s
ed
at
ev
er
y
co
n
v
o
lu
tio
n
al
la
y
er
ex
ce
p
t t
h
e
o
u
t
p
u
t la
y
er
.
=
1
1
+
2
2
+
⋯
+
(
1
)
wh
er
e
Y
is
th
e
f
ea
tu
r
e
m
ap
,
W
is
th
e
wei
g
h
t
ass
o
ciate
d
w
i
th
in
p
u
t
I
.
E
ac
h
o
f
th
ese
co
n
v
o
lu
tio
n
al
lay
er
s
wi
th
s
ize
×
an
d
ch
an
n
els
p
r
o
d
u
ce
s
a
f
ea
tu
r
e
m
a
p
o
n
wh
ich
t
h
e
p
r
ed
ictio
n
lay
er
(
class
i
f
icatio
n
lay
er
)
o
f
s
ize
3
×
3
×
is
ap
p
lied
.
T
h
e
v
alu
es
p
r
o
d
u
ce
d
ar
e
with
r
esp
ec
t
to
ea
ch
lo
ca
tio
n
o
f
th
e
f
ea
tu
r
e
m
ap
th
er
eb
y
y
ield
in
g
(
+
4
)
o
u
tp
u
ts
wh
er
e
c
r
ep
r
esen
ts
th
e
clas
s
s
co
r
es
with
f
o
u
r
o
f
f
s
ets
f
o
r
z
f
ilter
s
.
T
h
e
f
ea
tu
r
e
m
ap
is
th
en
s
u
b
s
eq
u
en
tly
p
r
o
ce
s
s
ed
b
y
o
th
er
co
n
v
o
lu
tio
n
a
l
lay
er
s
f
r
o
m
till
th
e
f
in
al
lay
er
w
h
er
e
f
o
r
ea
c
h
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
R
ea
l
-
time
h
u
ma
n
d
etec
tio
n
fo
r
elec
tr
icity
co
n
s
erva
tio
n
u
s
in
g
p
r
u
n
ed
-
S
S
D
a
n
d
a
r
d
u
in
o
(
Ush
a
s
u
kh
a
n
ya
S
.
)
1513
f
ea
tu
r
e
m
ap
(
+
4
)
v
al
u
es
ar
e
p
r
o
d
u
ce
d
[
3
3
]
.
Ap
p
r
o
p
r
iate
d
ef
au
lt
b
o
x
es
ar
e
t
o
b
e
s
elec
ted
f
o
r
b
etter
p
r
ed
ictio
n
an
d
h
en
c
e
th
e
g
r
o
u
n
d
tr
u
th
b
o
x
is
m
atch
ed
with
t
h
e
d
ef
a
u
lt
b
o
x
u
s
in
g
J
ac
ca
r
d
o
v
er
lap
m
et
h
o
d
with
a
th
r
esh
o
ld
v
alu
e
s
et
to
0
.
5
.
T
h
is
y
ield
s
a
lar
g
e
n
u
m
b
e
r
o
f
n
eg
ativ
e
s
am
p
les o
u
t o
f
wh
ich
th
e
s
am
p
les with
th
e
h
ig
h
est
co
n
f
id
en
ce
l
o
s
s
is
alo
n
e
s
elec
ted
,
th
e
r
eb
y
m
ak
in
g
a
r
atio
o
f
1
:3
f
o
r
p
o
s
itiv
e
t
o
n
eg
ativ
e
s
am
p
les.
Fin
ally
,
s
ig
m
o
id
f
u
n
ctio
n
is
u
s
ed
at
th
e
o
u
tp
u
t la
y
er
to
m
ak
e
a
b
in
ar
y
class
if
icatio
n
o
f
th
e
f
r
am
es a
n
d
th
e
n
o
n
-
m
ax
im
u
m
s
u
p
p
r
ess
io
n
s
(
NM
S)
m
ak
es
o
b
ject
d
etec
tio
n
s
.
T
h
e
weig
h
ts
o
f
th
e
n
etwo
r
k
a
r
e
in
itialize
d
u
s
in
g
th
e
“HE
”
in
itializatio
n
tech
n
iq
u
e
b
ased
o
n
th
e
f
o
r
m
u
la
(
)
=
2
_
(
2
)
wh
er
e
f
an
_
in
r
e
p
r
ese
n
ts
th
e
n
u
m
b
er
o
f
in
p
u
t
u
n
its
.
I
t
p
r
o
v
i
d
es
a
co
n
tr
o
lled
in
itializatio
n
,
th
er
eb
y
in
cr
ea
s
in
g
th
e
co
n
v
er
g
e
n
ce
r
ate.
L
ea
r
n
i
n
g
r
ate,
m
o
m
en
t
u
m
,
d
r
o
p
o
u
t
an
d
weig
h
t
d
ec
ay
ar
e
th
e
f
o
u
r
h
y
p
er
-
p
ar
am
eter
s
co
n
s
id
er
ed
f
o
r
tu
n
i
n
g
an
d
ar
e
d
o
n
e
b
y
o
r
th
o
g
o
n
al
ar
r
a
y
tu
n
in
g
m
eth
o
d
(
OAT
M)
b
ased
o
n
T
a
g
u
ch
i’
s
f
ac
to
r
tab
le.
T
h
is
m
eth
o
d
y
ield
s
a
ce
r
tain
s
et
o
f
f
ac
to
r
v
alu
es
wh
ich
m
ak
es
th
e
m
o
d
el’
s
co
n
v
er
g
en
ce
ea
s
ier
an
d
f
aster
.
T
h
e
h
y
p
er
-
p
ar
am
eter
s
ar
e
r
ep
r
esen
ted
as
f
ac
t
o
r
s
,
wh
ile
th
e
co
r
r
esp
o
n
d
in
g
v
alu
es
a
r
e
d
ef
in
ed
as
lev
els
in
th
e
tab
le.
T
h
e
o
r
th
o
g
o
n
al
ar
r
ay
tab
le
is
g
e
n
er
ated
u
s
in
g
W
eib
u
ll
++
s
o
f
twar
e
with
4
f
ac
t
o
r
s
th
at
h
as
9
r
o
ws.
T
h
e
m
o
d
el
is
tr
ain
ed
with
th
e
f
r
am
es
o
f
th
e
tr
ain
in
g
d
atasets
wh
ich
is
d
iv
id
ed
in
to
1
0
b
atch
es.
T
h
e
m
o
d
el
is
iter
ated
f
o
r
9
0
ep
o
ch
s
with
th
e
h
y
p
e
r
p
ar
am
ete
r
s
o
b
tain
e
d
b
y
th
e
OAT
M
m
eth
o
d
an
d
th
e
ac
cu
r
ac
y
f
o
r
ea
ch
s
et
o
f
h
y
p
e
r
p
ar
am
eter
s
is
esti
m
a
ted
.
T
h
e
r
o
w
with
th
e
h
ig
h
est
ac
cu
r
ac
y
is
co
n
s
id
er
ed
an
d
th
e
co
r
r
esp
o
n
d
i
n
g
v
alu
es
ar
e
th
e
o
p
tim
al
v
alu
es
with
wh
ich
th
e
h
y
p
er
p
ar
am
ete
r
s
ar
e
tu
n
ed
.
E
ac
h
an
d
ev
er
y
b
atch
is
iter
ated
f
o
r
4
ep
o
ch
s
an
d
t
h
e
a
v
er
ag
e
o
f
t
h
ese
f
o
u
r
e
x
p
er
im
e
n
ts
is
co
n
s
id
er
ed
as
th
e
ac
c
u
r
ac
y
o
f
a
s
elec
ted
s
in
g
le
s
et
o
f
hy
p
er
p
ar
am
eter
s
.
T
h
e
ac
cu
r
ac
ies
an
d
th
e
co
r
r
e
s
p
o
n
d
in
g
v
al
u
es
f
o
r
th
ese
s
elec
ted
h
y
p
er
p
a
r
am
eter
s
ar
e
p
r
esen
ted
in
T
ab
le
1
.
L
ev
el
5
in
T
ab
le
1
o
b
tain
s
th
e
h
ig
h
est
ac
cu
r
ac
y
r
ate
an
d
h
en
ce
th
e
c
o
r
r
esp
o
n
d
in
g
v
alu
es
ar
e
t
h
e
o
p
tim
al
v
alu
es
f
o
r
t
h
is
p
h
ase
o
f
SS
D.
T
h
e
s
p
ec
if
icatio
n
s
o
f
th
e
co
n
s
tr
u
cted
m
o
d
el
lik
e
th
e
k
er
n
el
s
ize,
in
p
u
t
an
d
o
u
tp
u
t
d
im
en
s
io
n
s
o
f
th
e
f
ea
tu
r
e
m
ap
s
o
f
ev
er
y
lay
er
,
th
e
to
tal
p
ar
am
eter
s
u
s
ed
in
ev
er
y
lay
er
etc
ar
e
g
iv
en
in
T
a
b
le
2
.
T
ab
le
1
.
Or
th
o
g
o
n
al
ar
r
a
y
tab
l
e
Ex
p
.
N
o
F
a
c
t
o
r
1
F
a
c
t
o
r
2
F
a
c
t
o
r
3
F
a
c
t
o
r
4
A
c
c
u
r
a
c
y
(
%)
1
0
.
0
1
0
.
9
9
9
0
.
3
0
.
0
0
1
8
5
.
2
1
2
0
.
0
0
1
0
.
9
9
0
.
2
5
0
.
0
0
3
8
4
.
3
2
3
0
.
0
1
0
.
9
0
.
2
0
0
.
0
0
5
8
6
.
1
2
4
0
.
0
0
1
0
.
9
9
9
0
.
2
5
0
.
0
0
5
8
5
.
7
2
5
0
.
0
0
1
0
.
9
9
0
.
2
0
0
.
0
0
1
8
7
.
2
4
6
0
.
0
0
1
0
.
9
0
.
3
0
.
0
0
3
8
6
.
8
8
7
0
.
0
0
0
1
0
.
9
9
9
0
.
2
0
0
.
0
0
3
8
5
.
3
2
8
0
.
0
0
0
1
0
.
9
9
0
.
3
0
.
0
0
5
8
6
.
9
1
9
0
.
0
0
0
1
0
.
9
0
.
2
5
0
.
0
0
1
8
5
.
4
7
T
ab
le
2
.
Sp
ec
if
icatio
n
s
o
f
th
e
co
n
s
tr
u
cted
m
o
d
el
S
S
D
LA
Y
ER
X
Y
K
e
r
n
e
l
si
z
e
C
in
C
out
P
a
r
a
me
t
e
r
s
C
o
n
v
1
_
1
2
4
0
0
4
0
0
0
3
3
64
1
.
7
2
8
C
o
n
v
1
_
2
2
4
0
0
4
0
0
0
3
64
64
36
.
864
C
o
n
v
2
_
1
1
2
0
0
2
0
0
0
3
64
1
2
8
73
.
728
C
o
n
v
2
_2
1
2
0
0
2
0
0
0
3
1
2
8
1
2
8
1
4
7
.
4
5
6
C
o
n
v
3
_
1
6
0
0
1
0
0
0
3
1
2
8
2
5
6
2
9
4
.
9
1
2
C
o
n
v
3
_
2
6
0
0
1
0
0
0
3
2
5
6
2
5
6
5
8
9
.
8
2
4
C
o
n
v
3
_
3
6
0
0
1
0
0
0
3
2
5
6
2
5
6
5
8
9
.
8
2
4
C
o
n
v
4
_
1
3
0
0
5
0
0
3
2
5
6
5
1
2
1
.
1
7
9
.
6
4
8
C
o
n
v
4
_
2
3
0
0
5
0
0
3
5
1
2
5
1
2
2
.
3
5
9
.
2
9
6
C
o
n
v
4
_
3
3
0
0
5
0
0
3
5
1
2
5
1
2
2
.
3
5
9
.
2
9
6
C
o
n
v
5
_3
1
5
0
2
5
0
3
5
1
2
1
.
0
2
4
4
.
7
1
8
.
5
9
2
C
o
n
v
5
_
3
1
5
0
2
5
0
3
1
.
0
2
4
1
.
0
2
4
9
.
4
3
7
.
1
8
4
C
o
n
v
5
_
3
1
5
0
2
5
0
3
1
.
0
2
4
1
.
0
2
4
9
.
4
3
7
.
1
8
4
S
S
D
_
C
o
n
v
6
75
1
2
5
1
1
.
0
2
4
2
.
0
4
8
2
.
0
9
7
.
1
5
2
S
S
D
_
C
o
n
v
7
75
1
2
5
1
2
.
0
4
8
4
.
0
9
6
8
.
3
8
8
.
6
0
8
S
S
D
_
C
o
n
v
8
38
63
2
4
.
0
9
6
8
.
1
9
2
1
3
4
.
2
1
7
.
728
S
S
D
_
C
o
n
v
9
38
63
2
8
.
1
9
2
16
.
384
5
3
6
.
8
7
0
.
912
S
S
D
_
C
o
n
v
1
0
19
32
2
16
.
384
32
.
768
2
.
1
4
7
.
4
8
3
.
648
S
S
D
_
C
o
n
v
1
1
19
32
2
32
.
768
32
.
768
4
.
2
9
4
.
9
6
7
.
296
D
e
t
e
c
t
i
o
n
s
1
1
1
32
.
768
21
6
8
8
.
1
2
8
N
M
S
1
1
1
32
.
768
84
2
.
7
5
2
.
5
1
2
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
11
,
No
.
2
,
Ap
r
il 2
0
2
1
:
1
5
1
0
-
1520
1514
T
h
e
p
ar
am
eter
s
ar
e
ca
lcu
lated
u
s
i
n
g
th
e
f
o
r
m
u
la;
=
_
2
×
ℎ
×
ℎ
(
3
)
wh
er
e
th
e
k
er
n
el_
s
ize,
c
h
an
n
e
l
in
,
an
d
ch
a
n
n
el
out
ar
e
d
ef
in
e
d
as
th
e
k
er
n
el
s
ize
o
f
th
e
weig
h
t
f
ilter
,
n
u
m
b
er
o
f
in
p
u
t
ch
an
n
els,
an
d
n
u
m
b
er
o
f
o
u
tp
u
t
c
h
an
n
els
in
ea
ch
c
o
n
v
o
lu
tio
n
lay
e
r
.
X
an
d
Y
r
ep
r
e
s
en
t
th
e
h
o
r
izo
n
tal
an
d
v
er
tical
d
im
en
s
io
n
s
o
f
th
e
f
ea
tu
r
e
m
ap
s
in
ea
ch
co
n
v
o
lu
tio
n
lay
er
.
T
h
e
am
o
u
n
t
o
f
c
o
m
p
u
tatio
n
is
d
ir
ec
tly
p
r
o
p
o
r
tio
n
al
to
th
e
n
u
m
b
e
r
o
f
p
ar
am
eter
s
o
f
th
e
n
etwo
r
k
in
e
ac
h
an
d
ev
e
r
y
lay
er
[
3
4
]
.
T
h
er
ef
o
r
e,
r
e
d
u
cin
g
t
h
e
n
u
m
b
er
o
f
p
ar
am
eter
s
in
b
o
t
h
co
n
v
o
lu
tio
n
an
d
FC
lay
er
s
h
elp
s
ac
h
iev
e
th
e
g
o
al.
T
h
er
e
ar
e
v
ar
io
u
s
s
tr
ateg
ies
to
d
ec
r
ea
s
e
th
e
m
o
d
el
s
ize,
s
u
ch
as
p
r
u
n
i
n
g
[
3
5
]
a
n
d
q
u
a
n
tizatio
n
[
3
6
]
etc.
,
o
u
t
o
f
wh
ich
p
r
u
n
in
g
is
s
elec
ted
f
o
r
o
u
r
wo
r
k
as it h
as y
ie
ld
ed
p
r
o
m
is
in
g
r
esu
lts
in
v
a
r
io
u
s
a
p
p
licatio
n
s
[
3
7
]
.
3
.
3
.
P
runin
g
T
h
e
SS
D
n
etwo
r
k
ex
is
ts
with
a
s
et
o
f
N
c
o
n
v
o
lu
tio
n
al
lay
er
s
wh
er
e
i
th
co
n
v
o
lu
tio
n
a
l
lay
er
is
r
ep
r
esen
ted
b
y
N
i
.
Pru
n
in
g
is
d
ef
in
ed
as
th
e
r
em
o
v
al
o
f
f
ilt
er
s
o
r
p
ar
am
eter
s
wh
ich
ar
e
co
n
s
id
e
r
ed
to
b
e
less
im
p
o
r
tan
t.
Hen
ce
,
th
e
en
tire
s
et
o
f
f
ilter
s
is
d
iv
i
d
ed
i
n
to
an
d
wh
ich
r
ep
r
esen
ts
th
e
im
p
o
r
t
an
t
f
ilter
an
d
l
ess
im
p
o
r
tan
t
f
ilter
s
et
r
esp
ec
tiv
ely
.
T
h
e
t
o
tal
n
u
m
b
er
o
f
im
p
o
r
tan
t
a
n
d
less
im
p
o
r
ta
n
t
f
ilt
er
s
is
d
ep
icted
b
y
K
an
d
L
r
esp
ec
tiv
ely
w
h
er
e
K+
L
=
Q
r
ep
r
esen
tin
g
th
e
to
tal
f
ilter
s
.
I
n
s
p
ir
ed
b
y
[
3
8
]
,
we
p
er
f
o
r
m
p
r
u
n
in
g
in
th
e
s
am
e
way
wh
er
e
f
ilter
p
r
u
n
in
g
in
g
en
er
al
is
f
o
r
m
u
lated
as
;
∑
∑
(
)
=
1
=
1
(
4
)
s
u
ch
th
at
∑
=
=
1
(
5
)
I
n
(
4
)
wh
e
r
e
ij
r
e
p
r
esen
ts
1
i
f
th
e
weig
h
ts
ar
e
lab
elled
as
K
an
d
0
o
th
er
wis
e.
T
h
e
im
p
o
r
tan
ce
o
f
t
h
e
f
ilter
is
m
ea
s
u
r
ed
b
y
(
)
.
Hen
ce
th
e
o
b
je
ctiv
e
is
to
m
in
im
ize
th
e
eq
u
ati
o
n
to
r
em
o
v
e
L
.
As
ea
ch
an
d
ev
er
y
f
ea
tu
r
e
m
ap
o
f
t
h
e
N
i
th
lay
er
p
lay
s
d
if
f
er
en
t r
o
les,
e
q
u
a
tio
n
4
h
as b
e
en
r
ef
o
r
m
u
lated
as;
∑
∑
−
(
)
=
1
=
1
[
Ẋ
(
(
,
:
,
:
)
)
]
(
6
)
wh
er
e
Y
r
ep
r
esen
ts
th
e
f
ea
t
u
r
e
m
ap
s
,
I
is
th
e
in
p
u
t im
ag
e
s
a
m
p
led
f
r
o
m
d
is
tr
ib
u
tio
n
D(
I
)
s
u
ch
th
at
;
∑
=
=
1
(
7
)
a
ls
o
th
e
ev
alu
at
io
n
o
f
th
e
f
ilter
is
d
ef
in
ed
as;
[
Ẋ
(
(
,
:
,
:
)
)
]
=
(
(
,
:
,
:
)
)
(
8
)
s
in
g
le
v
alu
e
d
ec
o
m
p
o
s
itio
n
is
ap
p
lied
wh
er
e
;
(
,
:
,
:
)
=
∑
=
1
(
9
)
s
u
ch
th
at;
∑
+
∑
=
ѓ
+
1
ѓ
=
1
(
1
0
)
wh
en
r
’
<
r,
i
m
i
an
d
n
i
ar
e
t
o
p
,
lef
t
an
d
r
ig
h
t
s
in
g
u
lar
v
alu
es
r
esp
ec
tiv
ely
.
T
h
u
s
,
th
e
r
an
k
o
f
ea
c
h
an
d
e
v
er
y
p
ar
am
eter
is
ca
lcu
lated
an
d
th
e
tech
n
iq
u
e
r
e
-
a
r
r
an
g
es
th
e
r
a
n
k
ed
p
ar
am
eter
s
in
d
ec
r
ea
s
in
g
o
r
d
e
r
wh
ich
f
in
ally
elim
in
ates
th
e
least
im
p
o
r
tan
t
p
ar
am
eter
s
(
b
o
tto
m
m
o
s
t)
.
T
h
e
m
o
d
el
is
tr
ain
e
d
ag
ain
with
th
e
h
y
p
er
p
ar
am
eter
s
o
b
tain
ed
b
y
OAT
M
m
eth
o
d
f
o
r
all
th
e
f
a
cto
r
v
alu
es.
T
h
e
o
p
tim
al
s
et
o
f
h
y
p
er
p
ar
am
eter
s
o
b
tain
ed
in
th
is
p
h
ase
is
d
i
f
f
er
en
t
as
th
e
n
etwo
r
k
is
ac
tu
ally
p
r
u
n
e
d
.
T
h
e
H
-
R
an
k
alg
o
r
ith
m
is
ag
ain
im
p
lem
en
te
d
an
d
th
e
p
r
o
ce
s
s
is
iter
ated
u
n
til
th
e
er
r
o
r
f
u
n
c
tio
n
is
s
im
ilar
to
th
e
o
n
e
o
b
ta
in
ed
b
y
th
e
m
o
d
el
with
o
u
t p
r
u
n
in
g
.
T
h
e
e
n
tire
wo
r
k
in
g
m
o
d
el
o
f
th
e
p
r
u
n
in
g
m
eth
o
d
o
l
o
g
y
is
g
iv
e
n
in
Fig
u
r
e
2
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
R
ea
l
-
time
h
u
ma
n
d
etec
tio
n
fo
r
elec
tr
icity
co
n
s
erva
tio
n
u
s
in
g
p
r
u
n
ed
-
S
S
D
a
n
d
a
r
d
u
in
o
(
Ush
a
s
u
kh
a
n
ya
S
.
)
1515
Fig
u
r
e
2
.
Pru
n
in
g
SS
D
m
o
d
el
3
.
4
.
T
ra
ini
ng
a
nd
t
esting
pro
ce
ss
T
h
e
f
r
am
es
a
r
e
r
esized
t
o
3
0
0
×3
0
0
t
o
f
ee
d
it
in
to
th
e
m
o
d
if
i
ed
-
SS
D
n
etwo
r
k
.
T
h
e
d
ataset
is
d
iv
id
ed
in
to
a
tr
ain
in
g
s
et
an
d
test
in
g
s
et
in
th
e
r
atio
o
f
7
0
:3
0
.
T
h
e
tr
ain
in
g
s
et
co
n
s
is
ts
o
f
4
7
.
2
9
9
f
r
am
es
in
wh
ich
th
e
v
alid
atio
n
s
et
is
a
s
u
b
class
co
n
tain
in
g
5
%
o
f
th
e
tr
ain
in
g
f
r
am
es
(
2
3
6
4
f
r
a
m
es)
to
tu
n
e
th
e
h
y
p
e
r
-
p
ar
a
m
eter
s
o
f
th
e
n
etwo
r
k
.
T
h
is
is
f
o
llo
wed
b
y
p
r
o
ce
s
s
in
g
o
f
th
e
test
in
g
s
et
(
2
0
.
2
7
1
f
r
am
es).
I
n
itializin
g
th
e
weig
h
ts
o
f
th
e
n
etwo
r
k
is
d
o
n
e
u
s
in
g
“
HE
”
in
itializatio
n
tech
n
i
q
u
e.
T
h
e
f
r
am
es
o
f
th
e
tr
ain
in
g
s
et
ar
e
f
ed
in
to
t
h
e
n
etwo
r
k
an
d
th
e
ex
p
e
r
im
en
ts
ar
e
r
u
n
f
o
r
all
th
e
lev
els
o
f
T
ab
le
1
(
OAT
M)
.
T
h
e
m
o
d
el
is
th
en
p
r
u
n
ed
b
y
r
em
o
v
in
g
th
e
less
im
p
o
r
tan
t
p
ar
am
eter
s
o
f
th
e
n
etwo
r
k
u
s
in
g
th
e
H
-
r
an
k
alg
o
r
ith
m
.
T
h
e
n
etwo
r
k
is
r
etr
ain
ed
with
th
e
s
et
o
f
hyp
er
p
ar
am
ete
r
s
o
b
tain
ed
b
y
th
e
OAT
M
m
e
th
o
d
.
T
h
e
p
r
o
ce
s
s
is
iter
ated
u
n
til
th
e
er
r
o
r
lo
s
s
f
u
n
ctio
n
is
s
im
ilar
to
th
e
o
n
e
o
b
tain
ed
b
y
th
e
m
o
d
el
b
ef
o
r
e
p
r
u
n
in
g
.
A
v
alid
atio
n
s
et
a
n
d
test
s
et
is
p
ass
ed
af
ter
tr
ain
in
g
th
e
n
etwo
r
k
an
d
th
e
ex
p
er
im
e
n
t
is
iter
ated
90
e
p
o
ch
s
,
an
d
th
e
m
ea
n
av
e
r
ag
e
p
r
ec
is
io
n
(
m
AP)
is
tak
en
f
o
r
ea
ch
lev
el.
T
h
e
h
ig
h
est
m
AP
o
b
tain
ed
f
o
r
t
h
e
test
s
et
is
8
7
.
2
1
%
b
e
f
o
r
e
p
r
u
n
in
g
an
d
8
5
.
8
2
%
af
ter
p
r
u
n
in
g
.
C
o
m
p
r
ess
io
n
r
ate
o
f
4
2
%
is
ac
h
iev
ed
b
y
p
r
u
n
in
g
wh
ich
r
ed
u
ce
s
th
e
test
in
g
tim
e
to
2
.
1
s
ec
o
n
d
s
f
r
o
m
4
.
5
s
ec
o
n
d
s
o
f
an
u
n
-
p
r
u
n
ed
m
o
d
el.
3
.
5
.
E
v
a
lua
t
io
n
T
h
e
p
r
e
d
icted
v
alu
es
an
d
th
e
g
r
o
u
n
d
tr
u
th
v
alu
es
a
r
e
r
e
p
r
esen
ted
as
=
{
p
cx
,
p
cy
,
p
w
,
p
h
}
an
d
=
{
g
cx
,
g
cy
,
g
w
,
g
h
}
r
esp
ec
tiv
ely
.
T
h
e
lo
s
s
f
u
n
ctio
n
is
th
e
to
tal
lo
s
s
ca
lcu
lated
b
y
s
u
m
m
in
g
th
e
c
lass
if
icatio
n
an
d
th
e
r
e
g
r
ess
io
n
lo
s
s
wh
ich
is
d
en
o
ted
b
y
:
=
(
,
,
,
)
=
1
(
(
,
)
+
(
,
)
)
(
1
1
)
wh
er
e
c
r
e
p
r
esen
ts
th
e
ce
n
tr
e
o
f
t
h
e
b
o
u
n
d
in
g
b
o
x
,
p
is
th
e
p
r
e
d
icted
v
al
u
e,
g
is
th
e
g
r
o
u
n
d
tr
u
th
o
f
th
e
b
o
u
n
d
in
g
b
o
x
es,
N
is
th
e
n
u
m
b
e
r
o
f
m
atch
ed
d
ef
a
u
lt
b
o
x
es
an
d
t
h
e
r
eg
r
ess
io
n
l
o
s
s
is
ca
lcu
lated
u
s
in
g
th
e
f
o
r
m
u
la
g
iv
e
n
b
y
[
3
9
,
40]
:
(
,
)
=
∑
ℎ
1
(
−
)
{
,
,
,
ℎ
)
(
1
2
)
wh
er
e
ℎ
1
(
)
=
{
0
.
5
2
,
|
|
<1
an
d
z
-
0
.
5
o
th
e
r
wis
e.
T
h
e
class
if
icatio
n
lo
s
s
b
ased
o
n
cr
o
s
s
en
tr
o
p
y
is
g
iv
en
b
y
(
,
)
=
−
(
)
−
(
1
−
)
(
1
−
)
(
1
3
)
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
11
,
No
.
2
,
Ap
r
il 2
0
2
1
:
1
5
1
0
-
1520
1516
m
AP
an
d
m
ea
n
F1
in
d
ex
(
m
F1
)
ar
e
th
e
m
etr
ics
u
s
ed
to
ev
alu
ate
th
e
d
etec
tio
n
ac
cu
r
ac
y
o
f
th
e
en
tire
m
o
d
el
an
d
s
u
b
lay
e
r
s
r
esp
ec
tiv
ely
.
T
h
e
av
er
ag
e
p
r
ec
i
s
io
n
(
AP)
is
g
iv
en
b
y
th
e
eq
u
atio
n
=
∫
(
)
1
0
(
1
4
)
wh
er
e
p
an
d
r
d
en
o
tes p
r
e
cisi
o
n
an
d
r
ec
all
r
esp
ec
tiv
ely
.
m
F1
is
g
iv
en
b
y
th
e
eq
u
atio
n
;
1
=
1
∑
1
=
1
(
1
5
)
wh
er
e
F1
is
g
iv
en
b
y
th
e
e
q
u
at
io
n
;
1
=
2
×
×
+
(
1
6
)
an
d
p
r
ec
is
io
n
an
d
r
ec
all
ar
e
g
i
v
en
b
y
th
e
eq
u
atio
n
s
;
=
+
=
+
(
1
7
)
T
P,
FP
an
d
FN d
en
o
tes tr
u
e
p
o
s
itiv
e,
f
alse
p
o
s
itiv
e
an
d
f
als
e
n
eg
ativ
e
r
esp
ec
tiv
ely
.
3.
6
.
P
o
wer
s
up
ply
m
a
na
g
em
ent
T
h
e
o
u
tp
u
t
o
f
th
e
p
r
u
n
ed
SS
D
is
f
ed
to
th
e
Ar
d
u
in
o
m
ic
r
o
co
n
tr
o
ller
wh
ich
is
co
n
n
ec
t
ed
to
th
e
elec
tr
ical
s
u
p
p
ly
o
f
a
r
o
o
m
o
r
a
n
y
in
d
o
o
r
en
v
ir
o
n
m
e
n
t.
T
h
e
co
n
tr
o
ller
en
ab
les
th
e
p
o
wer
s
u
p
p
l
y
i
n
a
n
d
ar
o
u
n
d
th
e
lo
ca
tio
n
(
2
m
ts
)
wh
er
e
a
h
u
m
an
is
d
etec
ted
an
d
d
is
ab
les
it,
if
th
e
h
u
m
an
is
u
n
d
etec
ted
.
T
h
er
ef
o
r
e,
th
e
elec
tr
ic
r
eso
u
r
ce
is
u
tili
ze
d
o
n
ly
wh
e
n
an
d
w
h
er
e
it
is
ac
tu
ally
n
ee
d
ed
a
n
d
s
av
es
th
e
r
eso
u
r
ce
ef
f
icien
tly
.
Usi
n
g
th
e
p
r
o
p
o
s
ed
f
r
am
ewo
r
k
,
th
e
a
v
er
ag
e
m
o
n
t
h
ly
co
n
s
u
m
p
tio
n
o
f
elec
tr
icity
f
o
r
a
r
esid
en
tial
en
v
ir
o
n
m
en
t
is
r
ed
u
ce
d
to
7
2
f
r
o
m
9
0
u
n
its
(
k
W
h
)
,
wh
ich
is
n
ea
r
l
y
o
n
e
q
u
ar
ter
o
f
th
e
to
tal
elec
tr
icity
co
n
s
u
m
p
tio
n
.
4.
RE
SU
L
T
S AN
D
D
I
SCU
SS
I
O
NS
T
h
e
p
r
e
d
ictio
n
ac
c
u
r
ac
ies
o
f
t
h
e
m
o
d
if
i
ed
-
SS
D
m
o
d
el
wh
ic
h
is
b
ased
o
n
OAT
M
tec
h
n
iq
u
e
f
o
r
b
o
th
th
e
d
atasets
ar
e
g
iv
en
in
T
ab
le
3
.
T
h
e
v
alid
atio
n
an
d
test
d
ata
ac
cu
r
ac
ies
ar
e
m
ea
s
u
r
ed
an
d
it
is
f
o
u
n
d
th
at
th
e
p
r
ed
ictio
n
ac
c
u
r
ac
y
ac
h
iev
ed
b
y
m
o
d
if
ied
-
SS
D
is
v
er
y
clo
s
e
to
th
at
o
f
th
e
o
r
ig
in
al
SS
D
b
u
t
th
e
tr
ain
in
g
tim
e
is
ex
ten
s
iv
ely
r
ed
u
ce
d
in
m
o
d
if
ied
-
SS
D,
th
er
eb
y
in
cr
ea
s
in
g
th
e
p
r
o
ce
s
s
in
g
s
p
ee
d
d
r
asti
ca
lly
.
T
h
e
av
er
ag
e
lo
s
s
o
f
v
alid
atio
n
an
d
test
d
ata
s
ets
b
y
m
o
d
i
f
ied
-
SS
D
(
with
o
u
t
p
r
u
n
in
g
)
is
1
3
.
4
7
5
%
f
o
r
b
o
t
h
th
e
d
atasets
.
T
h
e
g
r
ap
h
ical
r
e
p
r
esen
ta
tio
n
o
f
it
is
g
iv
en
in
Fig
u
r
e
3
.
SS
D
with
o
th
er
p
r
u
n
in
g
m
eth
o
d
s
lik
e
Sp
ar
s
e
s
tr
u
ctu
r
e
s
elec
tio
n
an
d
g
en
er
ativ
e
ad
v
e
r
s
ar
ial
lear
n
in
g
(
GAL
)
h
av
e
a
ls
o
b
ee
n
im
p
lem
en
ted
f
o
r
o
u
r
p
r
ed
ictio
n
b
u
t
th
e
r
esu
lts
o
f
SS
D
with
H
-
r
an
k
p
r
u
n
in
g
t
o
p
s
o
th
e
r
tech
n
i
q
u
es.
T
h
e
r
e
s
u
lts
o
f
v
a
r
io
u
s
p
r
u
n
in
g
t
ec
h
n
iq
u
es
in
te
r
m
s
o
f
ac
cu
r
ac
y
,
co
m
p
r
ess
io
n
r
ate
an
d
f
lo
atin
g
-
p
o
i
n
t
o
p
e
r
atio
n
s
(
FLOPs)
ar
e
r
e
p
r
esen
ted
i
n
T
ab
le
4
.
Ou
r
m
o
d
el
out
-
p
er
f
o
r
m
s
th
e
tr
ain
in
g
s
p
e
ed
o
f
th
e
m
o
d
el
cr
ea
ted
b
y
m
u
lti
-
lay
e
r
p
r
u
n
in
g
f
r
am
ewo
r
k
[
4
1
]
,
as
th
e
latter
im
p
lem
en
ts
th
r
ee
s
tag
es
o
f
p
r
u
n
in
g
,
o
n
l
y
af
ter
co
m
p
letely
t
r
ain
in
g
t
h
e
n
etwo
r
k
f
r
o
m
th
e
s
cr
atch
.
Hen
ce
th
e
tr
ain
in
g
tim
e
i
n
cr
ea
s
es
m
an
y
f
o
ld
s
in
th
is
m
eth
o
d
,
wh
er
ea
s
t
h
e
f
o
r
m
er
(
o
u
r
m
o
d
el)
tr
ai
n
s
th
e
n
etwo
r
k
in
itiall
y
b
ased
o
n
OAT
M
m
eth
o
d
wh
ich
r
ed
u
ce
s
th
e
tr
ain
in
g
tim
e
d
r
asti
ca
lly
.
Seco
n
d
ly
,
o
u
r
m
o
d
el
u
s
es
H
-
r
an
k
alg
o
r
ith
m
t
o
f
o
cu
s
o
n
l
o
w
-
r
a
n
k
f
ea
t
u
r
e
m
ap
s
r
ath
e
r
th
a
n
elim
in
atin
g
th
e
ze
r
o
weig
h
ts
b
ased
o
n
s
p
ar
s
ity
s
tatis
t
ics with
a
n
eg
lig
ib
le
lo
s
s
o
f
0
.
8
3
% a
t 4
2
.
0
1
%
co
m
p
r
ess
io
n
.
T
ab
le
3
.
R
esu
lts
o
f
b
ase
lin
e
SS
D
an
d
m
o
d
if
ie
d
-
SSD
M
o
d
e
l
s
D
a
t
a
s
e
t
s
C
l
a
s
ses
P
r
e
c
i
s
i
o
n
(
%)
R
e
c
a
l
l
(
%)
F
1
(
%)
Tr
a
i
n
i
n
g
t
i
me
(
sec
)
S
S
D
(
b
a
se
l
i
n
e
)
C
H
O
K
EPO
I
N
T
H
u
ma
n
8
9
.
1
8
6
.
7
8
7
.
8
5
8
3
4
6
N
o
n
-
H
u
ma
n
8
8
.
9
8
5
.
3
8
7
.
0
R
e
a
l
-
t
i
me
H
u
ma
n
8
7
.
1
8
5
.
6
8
6
.
3
2
1
2
3
6
N
o
n
-
H
u
ma
n
8
6
.
9
8
4
.
3
8
5
.
5
Mo
d
i
f
i
e
d
-
SSD
C
H
O
K
EPO
I
N
T
H
u
ma
n
8
8
.
6
8
6
.
4
8
7
.
4
5
6
2
1
N
o
n
-
H
u
ma
n
8
8
.
8
8
5
.
1
8
7
.
3
R
e
a
l
-
t
i
me
H
u
ma
n
8
7
.
1
8
5
.
1
8
6
.
1
1
7
2
1
N
o
n
-
H
u
ma
n
8
6
.
5
8
4
.
3
8
5
.
3
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
R
ea
l
-
time
h
u
ma
n
d
etec
tio
n
fo
r
elec
tr
icity
co
n
s
erva
tio
n
u
s
in
g
p
r
u
n
ed
-
S
S
D
a
n
d
a
r
d
u
in
o
(
Ush
a
s
u
kh
a
n
ya
S
.
)
1517
(
a)
(
b
)
Fig
u
r
e
3
.
(
a
)
L
o
s
s
o
f
C
HOKE
POI
NT
d
ataset
,
b
)
lo
s
s
o
f
r
ea
l
-
tim
e
d
at
as
et
As
o
u
r
f
r
am
ewo
r
k
u
s
es
,
lo
w
r
eso
lu
tio
n
in
d
o
o
r
C
C
T
V
im
ag
es,
d
u
e
to
d
u
lln
ess
,
th
e
ac
cu
r
ac
y
is
af
f
ec
ted
o
n
f
u
r
t
h
er
p
r
u
n
in
g
an
d
h
e
n
ce
we
s
to
p
p
r
u
n
in
g
at
th
is
lev
el.
T
h
o
u
g
h
th
e
r
esu
lts
o
f
Pru
n
ed
-
SS
D
s
ee
m
s
lig
h
tly
less
er
th
an
th
e
tr
ad
i
ti
o
n
al
SS
D,
th
is
d
if
f
er
en
ce
ca
n
b
e
ig
n
o
r
e
d
wh
en
c
o
m
p
a
r
ed
to
th
e
p
r
o
ce
s
s
in
g
s
p
ee
d
o
f
th
e
p
r
u
n
e
d
-
SSD
wh
ich
is
d
r
asti
ca
lly
im
p
r
o
v
ed
.
Pru
n
in
g
r
e
d
u
ce
s
th
e
m
o
d
el’
s
co
m
p
lex
ity
,
w
h
ich
ev
en
tu
ally
in
cr
ea
s
es
th
e
p
r
e
d
i
ctio
n
s
p
ee
d
o
f
th
e
test
d
ata
f
r
o
m
4
.
4
s
ec
o
n
d
s
o
n
a
n
av
er
a
g
e
b
y
u
n
-
p
r
u
n
ed
SS
D
to
2
.
1
s
ec
o
n
d
s
b
y
p
r
u
n
ed
-
SS
D.
T
h
e
p
r
ed
ictio
n
s
p
ee
d
an
d
th
e
lo
s
s
f
u
n
ctio
n
o
f
b
o
th
SS
D
with
an
d
with
o
u
t
p
r
u
n
in
g
ar
e
g
iv
en
i
n
Fig
u
r
e
4
.
T
ab
le
4
.
C
o
m
p
a
r
is
o
n
o
f
H
-
r
an
k
p
r
u
n
in
g
with
o
th
er
tec
h
n
iq
u
es
Te
c
h
n
i
q
u
e
s
M
a
p
C
o
m
p
r
e
ss
i
o
n
C
o
m
p
u
t
a
t
i
o
n
F
LO
P
s
S
S
D
(
B
a
se
l
i
n
e
)
8
6
.
6
5
N
i
l
1
0
0
%
3
4
3
.
6
0
M
S
S
D
w
i
t
h
S
S
S
8
3
.
2
1
%
3
8
.
7
1
%
2
2
%
2
1
0
.
1
6
M
S
S
D
w
i
t
h
G
A
L
8
4
.
9
1
%
4
0
.
2
0
%
2
1
.
2
%
2
0
8
.
0
2
M
S
S
D
w
i
t
h
H
-
R
a
n
k
8
5
.
8
2
%
4
2
.
0
1
%
1
8
.
9
%
1
4
4
.
3
1
M
(
a)
(
b
)
Fig
u
r
e
4
.
(
a
)
E
r
r
o
r
r
ate
o
f
SS
D
with
o
u
t a
n
d
with
p
r
u
n
in
g
,
(
b
)
d
etec
tio
n
s
p
ee
d
o
f
test
d
ata
T
h
e
p
r
o
ce
s
s
in
g
s
p
ee
d
o
f
th
e
test
d
ata
s
et
is
co
m
p
ar
ed
with
v
ar
io
u
s
p
r
elim
i
n
ar
y
m
o
d
e
ls
an
d
th
e
r
esu
lts
ar
e
p
r
esen
te
d
in
Fig
u
r
e
5
.
Am
o
n
g
all
t
h
e
o
th
er
tech
n
i
q
u
es,
th
e
p
r
o
p
o
s
e
d
m
o
d
if
ied
-
p
r
u
n
e
d
SS
D
ex
ce
ls
b
y
y
ield
in
g
t
h
e
lo
west
p
r
o
ce
s
s
in
g
s
p
ee
d
o
f
2
.
2
s
ec
o
n
d
s
o
n
an
av
er
ag
e
.
T
h
e
m
o
d
el
d
etec
ts
th
e
h
u
m
an
alo
n
g
with
th
e
lo
ca
tio
n
r
ep
r
esen
ted
b
y
a
b
o
u
n
d
in
g
b
o
x
wh
ic
h
is
g
iv
en
in
Fig
u
r
e
6
.
T
h
is
is
th
en
p
r
o
ce
s
s
ed
b
y
th
e
co
n
tr
o
ller
a
n
d
th
e
r
e
s
u
l
t
s
o
f
t
h
e
c
o
n
t
r
o
l
l
e
r
a
r
e
g
i
v
e
n
i
n
F
i
g
u
r
e
7.
A
s
t
h
e
r
o
t
a
t
i
o
n
o
f
t
h
e
f
a
n
(
w
h
e
n
p
o
w
e
r
e
n
a
b
l
e
d
)
w
i
l
l
n
o
t
b
e
clea
r
ly
v
is
ib
le
in
an
im
ag
e,
we
h
a
v
e
tak
en
two
b
u
lb
s
,
o
n
e
r
ep
r
esen
tin
g
th
e
l
ig
h
t
an
d
th
e
o
th
er
r
ep
r
esen
tin
g
th
e
f
an
i
n
F
ig
u
r
e
7
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8
7
0
8
I
n
t J E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
11
,
No
.
2
,
Ap
r
il 2
0
2
1
:
1
5
1
0
-
1520
1518
Fig
u
r
e
5
.
Pro
ce
s
s
in
g
s
p
ee
d
s
o
f
v
ar
io
u
s
tech
n
iq
u
es
(
a)
(
b
)
Fig
u
r
e
6
.
(
a
)
L
o
ca
tio
n
o
f
h
u
m
an
d
etec
ted
b
y
p
r
u
n
ed
SS
D
,
(
b
)
h
u
m
a
n
u
n
d
etec
ted
b
y
p
r
u
n
ed
SSD
(
a)
(
b
)
Fig
u
r
e
7
.
(
a
)
E
n
a
b
lin
g
th
e
r
eso
u
r
ce
o
n
h
u
m
a
n
’
s
p
r
esen
ce
,
(
b
)
d
is
ab
lin
g
th
e
r
eso
u
r
ce
wh
e
n
h
u
m
an
ex
its
5.
CO
NCLU
SI
O
N
UE
PMS
is
o
n
e
o
f
th
e
m
o
s
t
ess
en
tial
s
er
v
ices
in
d
ay
to
d
a
y
life
d
u
e
to
th
e
d
ep
letio
n
o
f
r
eso
u
r
ce
s
.
Am
o
n
g
v
a
r
io
u
s
ex
is
tin
g
m
eth
o
d
s
,
th
is
s
y
s
tem
u
s
es
th
e
e
x
is
tin
g
C
C
T
V
f
o
o
tag
e
to
d
etec
t
h
u
m
an
s
an
d
e
n
ab
le
th
e
p
o
wer
s
u
p
p
l
y
o
n
ly
i
n
th
e
l
o
ca
tio
n
(
2
m
ts
)
wh
er
e
h
u
m
a
n
s
ar
e
d
etec
ted
.
T
h
e
s
y
s
tem
u
s
es
m
o
d
if
ie
d
-
SS
D
f
o
r
h
u
m
an
d
etec
tio
n
with
a
s
p
e
cif
ic
h
y
p
er
p
ar
am
eter
tu
n
in
g
to
d
ec
r
ea
s
e
th
e
tr
ain
in
g
tim
e
o
f
th
e
m
o
d
el.
T
h
e
m
o
d
el
is
f
u
r
th
er
p
r
u
n
ed
b
y
th
e
H
-
r
an
k
alg
o
r
ith
m
to
d
ec
r
ea
s
e
th
e
co
m
p
u
tatio
n
al
co
s
t
th
er
e
b
y
in
cr
ea
s
in
g
th
e
p
r
o
ce
s
s
in
g
s
p
ee
d
o
f
th
e
n
etwo
r
k
.
An
Ar
d
u
in
o
m
icr
o
-
co
n
tr
o
ller
is
u
s
ed
to
m
a
n
a
g
e
th
e
p
o
wer
s
u
p
p
l
y
o
f
th
e
s
y
s
tem
.
T
h
e
p
r
o
p
o
s
ed
ar
c
h
itectu
r
e
ac
h
iev
es
an
a
v
er
ag
e
p
r
e
d
ictio
n
ac
cu
r
ac
y
o
f
8
5
.
8
2
%
with
a
m
u
ch
r
e
d
u
ce
d
co
m
p
r
ess
io
n
r
ate
o
f
4
2
%
o
f
th
e
o
r
ig
in
al
n
etwo
r
k
.
I
t
is
ev
id
e
n
t
th
at
th
e
p
r
o
p
o
s
ed
s
y
s
tem
s
av
es
n
ea
r
ly
o
n
e
th
ir
d
o
f
th
e
to
tal
elec
tr
icity
co
n
s
u
m
p
tio
n
.
As
th
e
s
y
s
tem
is
d
ev
elo
p
ed
o
n
l
y
f
o
r
in
d
o
o
r
en
v
ir
o
n
m
en
ts
,
co
n
s
id
er
in
g
lar
g
er
p
lace
s
lik
e
m
alls
o
r
o
u
t
d
o
o
r
e
n
v
ir
o
n
m
en
ts
,
o
cc
lu
s
io
n
h
an
d
lin
g
in
t
h
ese
p
lace
s
co
u
ld
b
e
tak
en
as
o
n
e
o
f
th
e
f
u
tu
r
e
d
ir
ec
tio
n
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
n
t J E
lec
&
C
o
m
p
E
n
g
I
SS
N:
2088
-
8
7
0
8
R
ea
l
-
time
h
u
ma
n
d
etec
tio
n
fo
r
elec
tr
icity
co
n
s
erva
tio
n
u
s
in
g
p
r
u
n
ed
-
S
S
D
a
n
d
a
r
d
u
in
o
(
Ush
a
s
u
kh
a
n
ya
S
.
)
1519
RE
F
E
R
E
NC
E
S
[1
]
A.
Ra
g
h
u
n
a
n
d
a
n
,
M
o
h
a
n
a
,
P
.
Ra
g
h
a
v
a
n
d
H.
V.
R.
Ara
d
h
y
a
,
“
Ob
j
e
c
t
De
tec
ti
o
n
Alg
o
rit
h
m
s
fo
r
V
i
d
e
o
S
u
r
v
e
il
lan
c
e
Ap
p
li
c
a
ti
o
n
s,”
2
0
1
8
In
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Co
mm
u
n
ica
ti
o
n
a
n
d
S
i
g
n
a
l
Pr
o
c
e
ss
in
g
(ICCS
P)
,
Ch
e
n
n
a
i
,
2
0
1
8
,
p
p
.
0
5
6
3
-
0
5
6
8
.
[2
]
Xio
n
g
we
i,
Wu
,
S
a
h
o
o
,
Do
y
e
n
a
n
d
Ho
i,
S
tev
e
n,
“
Re
c
e
n
t
Ad
v
a
n
c
e
s
in
De
e
p
Lea
rn
in
g
f
o
r
Ob
j
e
c
t
De
tec
ti
o
n
,
”
Ne
u
ro
c
o
mp
u
ti
n
g
,
v
o
l
.
3
9
6
,
p
p
.
3
9
-
64,
2
0
1
9
.
[3
]
Din
g
,
S
.
,
e
t
a
l.
,
“
S
u
r
v
S
u
rf:
h
u
m
a
n
re
tri
e
v
a
l
o
n
la
rg
e
s
u
rv
e
il
la
n
c
e
v
id
e
o
d
a
ta,”
M
u
lt
ime
d
i
a
T
o
o
ls
a
n
d
A
p
p
l
ica
ti
o
n
s
,
v
o
l.
7
6
,
p
p
.
6
5
2
1
-
6
5
4
9
,
2
0
1
7
.
[4
]
J
.
Da
i,
Y.
Li
,
K.
He
,
a
n
d
J.
S
u
n
,
“
R
-
F
CN:
o
b
jec
t
d
e
tec
ti
o
n
v
ia reg
i
o
n
-
b
a
s
e
d
fu
ll
y
c
o
n
v
o
lu
ti
o
n
a
l
n
e
two
rk
s,”
N
I
P
S
'
1
6
:
P
r
o
c
e
e
d
i
n
g
s
o
f
t
h
e
3
0
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
N
e
u
r
a
l
I
n
f
o
r
m
a
t
i
o
n
P
r
o
c
e
s
s
i
n
g
S
y
s
t
e
m
s
,
2
0
1
6
,
p
p
.
3
7
9
-
387.
[5
]
M
a
h
m
o
u
d
,
Ha
n
a
n
a
n
d
M
e
n
g
a
sh
,
Ha
n
a
n
,
“
A
n
o
v
e
l
tec
h
n
i
q
u
e
f
o
r
a
u
to
m
a
ted
c
o
n
c
e
a
led
fa
c
e
d
e
tec
ti
o
n
in
su
r
v
e
il
lan
c
e
v
id
e
o
s,”
Per
so
n
a
l
a
n
d
Ub
i
q
u
it
o
u
s
Co
mp
u
t
in
g
,
p
p
.
1
-
1
2
,
2
0
2
0
.
[6
]
Z.
Ca
i,
Q.
F
a
n
,
R.
S
.
F
e
ris,
a
n
d
N.
Va
sc
o
n
c
e
lo
s,
“
A
u
n
ifi
e
d
m
u
lt
i
-
sc
a
le
d
e
e
p
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
f
o
r
fa
st
o
b
jec
t
d
e
t
e
c
ti
o
n
,
”
C
o
n
fer
e
n
c
e
:
E
u
ro
p
e
a
n
Co
n
fer
e
n
c
e
o
n
Co
mp
u
ter
Vi
sio
n
,
2
0
1
6
,
p
p
.
3
5
4
-
3
7
0
.
[7
]
Alb
e
rto
Ca
stil
l
o
,
S
i
h
a
m
Tab
ik
,
F
ra
n
c
isc
o
P
é
re
z
,
Ro
b
e
rt
o
O
lmo
s,
F
ra
n
c
isc
o
He
rre
ra
,
“
Brig
h
t
n
e
ss
g
u
i
d
e
d
p
re
p
ro
c
e
ss
in
g
f
o
r
a
u
t
o
m
a
ti
c
c
o
ld
ste
e
l
we
a
p
o
n
d
e
tec
ti
o
n
i
n
su
rv
e
il
la
n
c
e
v
id
e
o
s
wit
h
d
e
e
p
lea
rn
in
g
,
”
Ne
u
ro
c
o
mp
u
ti
n
g
,
v
o
l
.
3
3
0
,
p
p
.
1
5
1
-
1
6
1
,
2
0
1
9
.
[8
]
J.
Re
d
m
o
n
a
n
d
A.
F
a
rh
a
d
i,
“
YOL
O9
0
0
0
:
b
e
tt
e
r,
fa
ste
r,
stro
n
g
e
r,
”
C
VP
R
,
2
0
1
6
.
[9
]
Yu
n
d
o
n
g
L
i
,
e
t
a
l.
,
“
M
u
lt
i
-
b
lo
c
k
S
S
D
b
a
se
d
o
n
sm
a
ll
o
b
jec
t
d
e
t
e
c
ti
o
n
fo
r
UA
V
ra
il
wa
y
sc
e
n
e
su
rv
e
il
lan
c
e
,
”
Ch
in
e
se
J
o
u
r
n
a
l
o
f
Aer
o
n
a
u
ti
c
s
,
v
o
l.
3
3
,
n
o
.
6
,
p
p
.
1
7
4
7
-
1
7
5
5
,
2
0
2
0
.
[1
0
]
Cru
z
-
Ro
a
A.
A.,
Are
v
a
lo
Ov
a
ll
e
J.
E.
,
M
a
d
a
b
h
u
s
h
i
A.,
G
o
n
z
á
lez
Os
o
rio
F
.
A.
,
“
A
De
e
p
Lea
rn
in
g
Arc
h
it
e
c
tu
re
fo
r
Im
a
g
e
Re
p
re
se
n
tatio
n
,
”
M
e
d
ica
l
I
ma
g
e
c
o
mp
u
ti
n
g
a
n
d
c
o
m
p
u
ter
-
a
s
siste
d
in
ter
v
e
n
ti
o
n
,
v
o
l.
1
6
,
p
p
.
4
0
3
-
4
1
0
,
2
0
1
3
.
[1
1
]
F
re
it
a
s S
.
,
Alm
e
id
a
C
.
,
S
il
v
a
H
.
,
e
t
a
l.
,
“
S
u
p
e
rv
ise
d
c
las
sifica
ti
o
n
fo
r
h
y
p
e
rsp
e
c
tral
ima
g
in
g
i
n
UA
V
m
a
rit
ime
targ
e
t
d
e
tec
ti
o
n
,
”
2
0
1
8
IEE
E
i
n
ter
n
a
ti
o
n
a
l
c
o
n
fer
e
n
c
e
o
n
a
u
to
n
o
m
o
u
s
r
o
b
o
t
sy
ste
ms
a
n
d
c
o
mp
e
ti
t
io
n
s
,
To
rre
s
Ve
d
ra
s
,
2
0
1
8
,
p
p
.
8
4
-
9
0
.
[1
2
]
Ra
fiq
u
e
,
M
.
A.,
P
e
d
ry
c
z
,
W.
a
n
d
Je
o
n
,
M
.
,
“
Ve
h
icle
li
c
e
n
se
p
late
d
e
tec
ti
o
n
u
sin
g
re
g
i
o
n
-
b
a
se
d
c
o
n
v
o
lu
ti
o
n
a
l
n
e
u
ra
l
n
e
two
rk
s,”
S
o
f
t
Co
m
p
u
ti
n
g
,
v
o
l.
2
2
,
p
p
.
6
4
2
9
-
6
4
4
0
,
20
18
.
[1
3
]
He
K
.
M
.
,
Zh
a
n
g
X
.
Y
.
,
Re
n
S
.
Q
.
,
e
t
a
l.
,
“
S
p
a
ti
a
l
p
y
ra
m
id
p
o
o
li
n
g
i
n
d
e
e
p
c
o
n
v
o
l
u
ti
o
n
a
l
n
e
tw
o
rk
s
fo
r
v
isu
a
l
r
e
c
o
g
n
i
t
i
o
n
,
”
i
n
I
E
E
E
T
r
a
n
s
a
c
t
i
o
n
s
o
n
P
a
t
t
e
r
n
A
n
a
l
y
s
i
s
a
n
d
M
a
c
h
i
n
e
I
n
t
e
l
l
i
g
e
n
c
e
,
v
o
l
.
3
7
,
n
o
.
9
,
p
p
.
1
9
0
4
-
19
16
,
2
0
1
5
.
[1
4
]
Re
n
,
Yu
n
,
Z
h
u
,
Ch
a
n
g
re
n
a
n
d
X
iao
,
S
h
u
n
p
i
n
g
,
“
Ob
jec
t
De
tec
ti
o
n
Ba
se
d
o
n
F
a
st/F
a
ste
r
RCNN
E
m
p
lo
y
i
n
g
F
u
ll
y
Co
n
v
o
l
u
ti
o
n
a
l
Arc
h
it
e
c
t
u
re
s,”
M
a
th
e
ma
ti
c
a
l
Pro
b
lem
s in
E
n
g
in
e
e
ri
n
g
,
v
o
l.
2
0
1
8
,
p
p
.
1
-
7
,
2
0
1
8
.
[1
5
]
Li
,
Jia
x
i
n
g
,
e
t
a
l.
,
“
F
a
c
ial
E
x
p
re
ss
io
n
Re
c
o
g
n
it
i
o
n
with
F
a
ste
r
R
-
CNN
,
”
Pro
c
e
d
ia
C
o
mp
u
ter
S
c
i
e
n
c
e
,
v
o
l
.
1
0
7
,
p
p
.
1
3
5
-
1
4
0
,
2
0
1
7
.
[1
6
]
B.
Li
u
,
W.
Z
h
a
o
a
n
d
Q.
S
u
n
,
"
S
tu
d
y
o
f
o
b
jec
t
d
e
tec
ti
o
n
b
a
se
d
o
n
F
a
ste
r
R
-
CNN
,
"
2
0
1
7
C
h
i
n
e
se
Au
to
ma
t
io
n
Co
n
g
re
ss
(CAC)
,
Jin
a
n
,
2
0
1
7
,
p
p
.
6
2
3
3
-
6
2
3
6
.
[1
7
]
J.
Re
d
m
o
n
,
S
.
Di
v
v
a
la,
R.
G
irsh
i
c
k
a
n
d
A.
F
a
rh
a
d
i,
"
Yo
u
On
l
y
L
o
o
k
On
c
e
:
Un
ifi
e
d
,
Re
a
l
-
Ti
m
e
Ob
j
e
c
t
De
tec
ti
o
n
,
"
2
0
1
6
IE
EE
C
o
n
fer
e
n
c
e
o
n
Co
m
p
u
ter
Vi
sio
n
a
n
d
Pa
tt
e
rn
Rec
o
g
n
it
i
o
n
(CV
PR
),
Las
Ve
g
a
s,
NV
,
2
0
1
6
,
p
p
.
7
7
9
-
7
8
8
.
[1
8
]
C.
F
u
,
W.
Li
u
,
A.
Ra
n
g
a
,
A.
Ty
a
g
i,
a
n
d
A.
C.
Be
r
g
,
“
DSS
D
:
De
c
o
n
v
o
l
u
ti
o
n
a
l
si
n
g
le sh
o
t
d
e
tec
to
r,
”
Co
RR
,
2
0
1
7
.
[1
9
]
G
u
ime
i
Ca
o
,
Xu
e
m
e
i
Xie
,
Wen
z
h
e
Ya
n
g
,
Qu
a
n
L
iao
,
G
u
a
n
g
m
in
g
S
h
i
,
a
n
d
Jin
ji
a
n
W
u
,
"
F
e
a
tu
re
-
fu
se
d
S
S
D:
fa
st
d
e
tec
ti
o
n
f
o
r
sm
a
ll
o
b
jec
ts,"
N
in
th
In
ter
n
a
ti
o
n
a
l
C
o
n
fer
e
n
c
e
o
n
Gr
a
p
h
ic
a
n
d
Ima
g
e
Pro
c
e
ss
in
g
(ICGIP
2
0
1
7
)
,
2
0
1
8
,
p
p
.
1
-
8
.
[
2
0
]
A
.
G
.
H
o
w
a
r
d
,
“
S
o
m
e
i
m
p
r
o
v
e
m
e
n
t
s
o
n
d
e
e
p
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
b
a
s
e
d
i
m
a
g
e
c
l
a
s
s
i
f
i
c
a
t
i
o
n
,
”
C
o
R
R
,
2
0
1
3
.
[2
1
]
D.
P
.
Ki
n
g
m
a
a
n
d
J.
Ba
,
“
Ad
a
m
:
A m
e
th
o
d
f
o
r
st
o
c
h
a
stic o
p
ti
m
iza
ti
o
n
,
”
IC
L
R
,
2
0
1
5
.
[2
2
]
Zh
a
n
g
,
Xia
n
g
,
Ch
e
n
,
Xia
o
c
o
n
g
,
Ya
o
,
Li
n
a
,
G
e
,
Ch
a
n
g
,
Do
n
g
,
M
a
n
q
in
g
,
“
De
e
p
Ne
u
ra
l
Ne
two
r
k
Hy
p
e
rp
a
ra
m
e
ter
Op
ti
m
iza
ti
o
n
wi
th
Ort
h
o
g
o
n
a
l
Ar
ra
y
Tu
n
in
g
,
”
Ne
u
r
a
l
In
fo
rm
a
t
io
n
Pro
c
e
ss
in
g
,
p
p
.
2
8
7
-
2
9
5
,
2
0
1
9
.
[2
3
]
Y.
LeCu
n
,
J
.
S
.
De
n
k
e
r,
a
n
d
S
.
A.
S
o
ll
a
,
“
Op
ti
m
a
l
b
ra
in
d
a
m
a
g
e
,
”
Ad
v
a
n
c
e
s
in
Ne
u
ra
l
I
n
fo
rm
a
t
io
n
Pr
o
c
e
ss
in
g
S
y
ste
ms
2
(NIPS
1
9
8
9
)
,
p
p
.
5
9
8
-
6
0
5
,
1
9
9
0
.
[2
4
]
B.
Ha
ss
ib
i
a
n
d
D.
G
.
S
t
o
rk
,
“
S
e
c
o
n
d
o
rd
e
r
d
e
riv
a
ti
v
e
s
fo
r
n
e
tw
o
r
k
p
ru
n
in
g
:
O
p
ti
m
a
l
b
ra
in
su
rg
e
o
n
,
”
A
d
v
a
n
c
e
s
i
n
Ne
u
ra
l
In
fo
rm
a
t
io
n
Pro
c
e
ss
in
g
S
y
ste
ms
5
(NIPS
1
9
9
2
)
,
p
p
.
1
6
4
-
1
7
1
,
1
9
9
3
.
[2
5
]
H.
Li
,
A.
Ka
d
a
v
,
I.
Du
r
d
a
n
o
v
ic,
H.
S
a
m
e
t,
a
n
d
H.
P
.
G
ra
f,
“
P
ru
n
in
g
fil
ter
s
fo
r
e
fficie
n
t
c
o
n
v
Ne
t
s,”
ICL
R
2
0
1
7
c
o
n
fer
e
n
c
e
su
b
miss
io
n
,
2
0
1
6
.
[2
6
]
R.
Ti
b
s
h
iran
i,
“
Re
g
re
ss
io
n
s
h
rin
k
a
g
e
a
n
d
se
lec
ti
o
n
v
ia
th
e
las
so
,
”
J
o
u
rn
a
l
o
f
t
h
e
Ro
y
a
l
S
ta
ti
stic
a
l
S
o
c
iety
.
S
e
rie
s
B
(M
e
th
o
d
o
lo
g
ica
l)
,
v
o
l.
5
8
,
n
o
.
1
,
p
p
.
2
6
7
-
2
8
8
,
1
9
9
6
.
[2
7
]
X.
Re
n
,
H
.
G
u
o
,
G
.
He
,
X.
Xu
,
C.
Di
a
n
d
S
.
L
i,
"
Co
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
Ne
two
r
k
Ba
se
d
o
n
P
ri
n
c
i
p
a
l
Co
m
p
o
n
e
n
t
An
a
ly
sis
In
it
ializa
ti
o
n
fo
r
Im
a
g
e
Clas
sifica
ti
o
n
,
"
2
0
1
6
IEE
E
I
n
ter
n
a
ti
o
n
a
l
Co
n
fer
e
n
c
e
o
n
Da
ta
S
c
ien
c
e
in
Cy
b
e
rs
p
a
c
e
(DS
C)
,
Ch
a
n
g
sh
a
,
2
0
1
6
,
p
p
.
3
2
9
-
3
3
4
.
[2
8
]
Hu
Q
.
,
Li
F
.
,
“
Ha
rd
wa
re
d
e
sig
n
o
f
sm
a
rt
h
o
m
e
e
n
e
rg
y
m
a
n
a
g
e
m
e
n
t
sy
s
tem
wi
t
h
d
y
n
a
m
ic
p
rice
r
e
sp
o
n
se
,
”
IE
E
E
T
ra
n
sa
c
ti
o
n
s
o
n
S
m
a
rt Grid
,
v
o
l.
4
,
n
o
.
4
,
p
p
.
1
8
7
8
-
18
87
,
2
0
1
3
.
[2
9
]
Ja
n
g
,
Y
o
u
n
g
k
y
o
o
n
,
G
u
n
e
s,
Ha
ti
c
e
a
n
d
P
a
tras
,
I
o
a
n
n
is,
“
Re
g
istrati
o
n
-
fre
e
F
a
c
e
-
S
S
D:
S
i
n
g
le
sh
o
t
a
n
a
ly
sis
o
f
sm
il
e
s,
fa
c
ial
a
tt
rib
u
tes
,
a
n
d
a
ffe
c
t
in
th
e
wild
,
”
Co
m
p
u
ter
Vi
si
o
n
a
n
d
Im
a
g
e
Un
d
e
rs
ta
n
d
in
g
,
p
p
.
1
-
1
4
,
2
0
1
9
.
[3
0
]
S
in
g
h
,
K.
K.
a
n
d
Lee
,
Y.
J.,
“
Hid
e
-
a
n
d
-
se
e
k
:
F
o
rc
in
g
a
n
e
two
rk
t
o
b
e
m
e
ti
c
u
l
o
u
s
fo
r
we
a
k
ly
s
u
p
e
rv
ise
d
o
b
jec
t
a
n
d
a
c
t
i
o
n
l
o
c
a
l
i
z
a
t
i
o
n
,
”
2
0
1
7
I
E
E
E
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
V
i
s
i
o
n
(
I
C
C
V
)
,
V
e
n
i
ce
,
2
0
17
,
p
p
.
3
5
4
4
-
3553
.
[3
1
]
Zh
o
u
Z
.
,
S
a
n
d
e
rs
J
.
W
.
,
J
o
h
n
so
n
J
.
M
.
,
e
t
a
l.
,
“
C
o
m
p
u
ter
-
a
id
e
d
De
tec
ti
o
n
o
f
Bra
in
M
e
tas
tas
e
s
in
T
1
-
we
ig
h
ted
M
RI
f
o
r
S
t
e
r
e
o
t
a
c
t
i
c
R
a
d
i
o
s
u
r
g
e
r
y
U
s
i
n
g
D
e
e
p
L
e
a
r
n
i
n
g
S
i
n
g
l
e
-
S
h
o
t
D
e
t
e
c
t
o
r
s
,
”
R
a
d
i
o
l
o
gy
,
v
o
l
.
2
9
5
,
n
o
.
2
,
p
p
.
4
0
7
-
4
1
5
,
2020
.
[3
2
]
W.
Li
u
,
D.
An
g
u
e
lo
v
,
D.
Erh
a
n
,
C.
S
z
e
g
e
d
y
,
S
.
E.
Re
e
d
,
C.
F
u
,
a
n
d
A.
C.
Be
rg
,
“
S
S
D:
sin
g
l
e
sh
o
t
m
u
l
ti
b
o
x
d
e
tec
to
r,
”
Eu
r
o
p
e
a
n
C
o
n
fer
e
n
c
e
o
n
C
o
mp
u
ter
Vi
sio
n
-
ECCV
2
0
1
6
,
v
o
l
.
9
9
0
5
,
2
0
1
6
,
p
p
.
2
1
-
3
7
.
Evaluation Warning : The document was created with Spire.PDF for Python.