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.
16
,
No
.
4
,
A
u
g
u
s
t
20
26
,
p
p
.
2
0
4
2
~
2
0
6
0
I
SS
N:
2088
-
8
7
0
8
,
DOI
: 1
0
.
1
1
5
9
1
/ijece.
v
16
i
4
.
pp
2
0
4
2
-
206
0
2042
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ij
ec
e.
ia
esco
r
e.
co
m
Co
mpa
ra
tive perf
o
rma
nce ana
ly
sis
of lightweig
h
t
fac
e
identif
i
ca
tion a
lg
o
rithm
Wuy
un
Wa
ng
1
,
Su
c
ha
da
Sitj
o
ng
s
a
t
a
po
rn
2
1
Th
e
E
l
e
c
t
r
i
c
a
l
E
n
g
i
n
e
e
r
i
n
g
G
r
a
d
u
a
t
e
P
r
o
g
r
a
m
,
F
a
c
u
l
t
y
o
f
E
n
g
i
n
e
e
r
i
n
g
a
n
d
Te
c
h
n
o
l
o
g
y
,
M
a
h
a
n
a
k
o
r
n
U
n
i
v
e
r
si
t
y
o
f
Te
c
h
n
o
l
o
g
y
I
n
t
e
r
n
a
t
i
o
n
a
l
C
o
l
l
e
g
e
(
M
U
TI
C
)
,
M
a
h
a
n
a
k
o
r
n
U
n
i
v
e
r
si
t
y
o
f
Te
c
h
n
o
l
o
g
y
,
B
a
n
g
k
o
k
,
Th
a
i
l
a
n
d
2
D
e
p
a
r
t
me
n
t
o
f
El
e
c
t
r
o
n
i
c
E
n
g
i
n
e
e
r
i
n
g
,
S
c
h
o
o
l
o
f
El
e
c
t
r
i
c
a
l
a
n
d
El
e
c
t
r
o
n
i
c
En
g
i
n
e
e
r
i
n
g
(
S
E
E)
,
F
a
c
u
l
t
y
o
f
En
g
i
n
e
e
r
i
n
g
a
n
d
Te
c
h
n
o
l
o
g
y
,
M
a
h
a
n
a
k
o
r
n
U
n
i
v
e
r
s
i
t
y
o
f
T
e
c
h
n
o
l
o
g
y
,
B
a
n
g
k
o
k
,
Th
a
i
l
a
n
d
Art
icle
I
nfo
AB
S
T
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
J
an
1
7
,
2
0
2
6
R
ev
is
ed
Ma
y
7
,
2
0
2
6
Acc
ep
ted
J
u
l 2
2
,
2
0
2
6
Wi
th
t
h
e
wid
e
a
p
p
li
c
a
ti
o
n
o
f
fa
c
e
re
c
o
g
n
it
io
n
in
re
so
u
rc
e
-
c
o
n
stra
in
e
d
sc
e
n
a
rio
s
li
k
e
m
o
b
il
e
a
n
d
e
m
b
e
d
d
e
d
d
e
v
ice
s,
li
g
h
twe
ig
h
t
a
l
g
o
rit
h
m
s
h
a
v
e
b
e
c
o
m
e
a
re
se
a
r
c
h
fo
c
u
s,
b
u
t
e
x
isti
n
g
stu
d
ies
lac
k
m
u
lt
i
-
d
i
m
e
n
sio
n
a
l,
sc
e
n
a
rio
-
b
a
se
d
p
e
rfo
rm
a
n
c
e
c
o
m
p
a
riso
n
s.
T
h
is
p
a
p
e
r
st
u
d
ies
th
e
p
e
rfo
rm
a
n
c
e
e
v
a
lu
a
ti
o
n
a
n
d
a
p
p
li
c
a
ti
o
n
a
d
a
p
tati
o
n
o
f
li
g
h
twe
ig
h
t
fa
c
e
re
c
o
g
n
it
i
o
n
a
lg
o
rit
h
m
s,
in
n
o
v
a
ti
v
e
ly
b
u
i
ld
s
a
sc
e
n
a
ri
o
-
b
a
se
d
e
v
a
lu
a
ti
o
n
sy
ste
m
,
v
e
rifi
e
s
th
e
p
e
rfo
rm
a
n
c
e
imp
ro
v
e
m
e
n
t
o
f
c
o
m
b
i
n
in
g
t
ra
d
it
io
n
a
l
a
lg
o
rit
h
m
s
with
M
o
b
il
e
Ne
t,
a
n
d
c
o
n
stru
c
ts
a
n
e
fficie
n
t,
sta
b
le
a
n
d
lo
w
-
c
o
st
sy
ste
m
.
It
e
lab
o
ra
tes
o
n
fa
c
e
re
c
o
g
n
it
io
n
p
ri
n
c
ip
les
,
in
c
l
u
d
i
n
g
k
e
y
li
n
k
s
o
f
fa
c
e
d
e
tec
ti
o
n
,
fe
a
tu
re
e
x
trac
ti
o
n
a
n
d
m
a
tch
in
g
,
i
n
tr
o
d
u
c
e
s
trad
it
io
n
a
l
a
lg
o
rit
h
m
s
su
c
h
a
s
Ei
g
e
n
fa
c
e
s,
F
ish
e
rfa
c
e
s
a
n
d
LBP
H,
a
n
d
f
o
c
u
se
s
o
n
M
o
b
i
leN
e
t’s
c
h
a
ra
c
teristics
:
re
d
u
c
in
g
c
o
m
p
u
tati
o
n
a
n
d
p
a
ra
m
e
ters
v
ia
d
e
p
th
wise
se
p
a
ra
b
le
c
o
n
v
o
lu
ti
o
n
,
a
n
d
a
d
j
u
sta
b
le
wid
t
h
a
n
d
re
so
lu
t
io
n
.
F
o
u
r
c
o
m
p
a
ra
ti
v
e
e
x
p
e
rime
n
ts
v
e
rify
th
e
"
trad
it
io
n
a
l
a
lg
o
rit
h
m
s
+
M
o
b
i
leN
e
t"
h
y
b
rid
stra
teg
y
.
Re
su
l
ts
sh
o
w
t
h
e
c
o
m
b
in
a
ti
o
n
a
c
h
ie
v
e
s
9
8
.
1
%
a
c
c
u
ra
c
y
,
4
.
3
p
e
rc
e
n
tag
e
p
o
in
ts
h
ig
h
e
r
t
h
a
n
sin
g
le
M
o
b
i
leN
e
t;
LBP
H
+
M
o
b
i
leN
e
t
b
a
lan
c
e
s
p
e
rfo
rm
a
n
c
e
a
n
d
re
so
u
r
c
e
c
o
n
su
m
p
ti
o
n
b
e
st,
wit
h
1
1
0
M
B
m
e
m
o
ry
,
4
0
%
CP
U
u
sa
g
e
a
n
d
3
1
5
m
s
p
ro
c
e
ss
in
g
ti
m
e
.
Th
e
h
y
b
rid
stra
teg
y
imp
ro
v
e
s
a
c
c
u
ra
c
y
a
n
d
e
fficie
n
c
y
in
d
iffere
n
t
sc
e
n
a
rio
s,
a
imin
g
to
p
r
o
v
i
d
e
a
sc
ien
ti
fic
b
a
sis
fo
r
th
e
e
n
g
i
n
e
e
rin
g
a
p
p
l
ica
ti
o
n
a
n
d
s
u
b
se
q
u
e
n
t
o
p
ti
m
i
z
a
ti
o
n
o
f
li
g
h
tw
e
i
g
h
t
fa
c
e
re
c
o
g
n
it
i
o
n
a
l
g
o
rit
h
m
s,
a
n
d
s
u
p
p
o
rt
in
g
a
lg
o
rit
h
m
se
lec
ti
o
n
a
n
d
p
e
rf
o
rm
a
n
c
e
imp
ro
v
e
m
e
n
t
in
re
so
u
rc
e
-
c
o
n
stra
in
e
d
sc
e
n
a
rio
s.
K
ey
w
o
r
d
s
:
Face
r
ec
o
g
n
itio
n
al
g
o
r
ith
m
Mo
b
ileNet
Mo
d
el
o
p
tim
izatio
n
Mu
lti
-
d
im
en
s
io
n
al
ev
alu
atio
n
s
y
s
tem
Per
f
o
r
m
an
ce
c
o
m
p
a
r
is
o
n
R
eso
u
r
ce
-
co
n
s
tr
ain
ed
s
ce
n
ar
i
o
s
T
h
is i
s
a
n
o
p
e
n
a
c
c
e
ss
a
rticle
u
n
d
e
r th
e
CC B
Y
-
SA
li
c
e
n
se
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Su
ch
ad
a
Sit
jo
n
g
s
atap
o
r
n
Dep
ar
tm
en
t o
f
E
lectr
o
n
ic
E
n
g
in
ee
r
in
g
,
Ma
h
an
ak
o
r
n
I
n
s
titu
te
o
f
I
n
n
o
v
ati
o
n
,
Facu
lty
o
f
E
n
g
in
ee
r
in
g
a
n
d
T
ec
h
n
o
lo
g
y
,
Ma
h
an
ak
o
r
n
Un
i
v
er
s
ity
o
f
T
ec
h
n
o
lo
g
y
B
an
g
k
o
k
,
T
h
ailan
d
E
m
ail:
s
s
u
ch
ad
a@
m
u
t.a
c.
th
1.
I
NT
RO
D
UCT
I
O
N
As
an
im
p
o
r
tan
t
d
ir
ec
tio
n
in
co
m
p
u
ter
v
is
io
n
an
d
b
io
m
etr
ic
id
en
tific
atio
n
,
f
ac
e
r
ec
o
g
n
itio
n
tech
n
o
lo
g
y
h
as
a
d
v
an
ce
d
r
em
ar
k
ab
ly
i
n
r
ec
e
n
t
y
ea
r
s
.
Or
i
g
in
atin
g
in
th
e
1
9
6
0
s
,
it
h
as
ev
o
lv
ed
f
r
o
m
s
im
p
l
e
tem
p
late
m
atch
in
g
an
d
m
an
u
al
f
ea
tu
r
e
e
x
tr
ac
tio
n
t
o
au
t
o
m
atic
r
ec
o
g
n
itio
n
u
s
in
g
co
m
p
u
t
er
v
is
io
n
a
n
d
d
ee
p
lear
n
in
g
.
T
h
e
in
tr
o
d
u
ctio
n
o
f
d
ee
p
lear
n
in
g
in
th
e
ea
r
ly
2
1
s
t c
en
tu
r
y
g
r
ea
tly
im
p
r
o
v
ed
its
ac
cu
r
ac
y
,
p
r
o
m
o
ti
n
g
wid
e
ap
p
licatio
n
in
s
ec
u
r
ity
m
o
n
ito
r
in
g
,
id
en
tity
a
u
th
en
tic
atio
n
,
s
m
ar
t
p
ay
m
e
n
t,
an
d
ac
c
ess
co
n
tr
o
l.Wi
th
th
e
r
ap
id
d
ev
elo
p
m
en
t
o
f
I
o
T
,
m
o
b
ile
d
ev
ices,
an
d
ed
g
e
co
m
p
u
tin
g
,
d
em
a
n
d
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
in
r
eso
u
r
ce
-
co
n
s
tr
ain
ed
s
ce
n
a
r
io
s
is
g
r
o
win
g
.
Dev
ices
lik
e
s
m
ar
t
ca
m
e
r
as
an
d
m
o
b
ile
te
r
m
in
als
h
a
v
e
lim
ited
h
ar
d
war
e
r
eso
u
r
ce
s
(
co
m
p
u
tin
g
p
o
wer
,
m
em
o
r
y
,
s
to
r
ag
e)
,
m
ak
in
g
tr
ad
itio
n
al
co
m
p
le
x
alg
o
r
ith
m
s
h
ar
d
to
d
ep
lo
y
.
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
C
o
mp
a
r
a
tive
p
erfo
r
ma
n
ce
a
n
a
lysi
s
o
f lig
h
tw
eig
h
t fa
ce
id
en
t
ifica
tio
n
a
lg
o
r
ith
m
(
Wu
yu
n
Wa
n
g
)
2043
L
i
g
h
twei
g
h
t
al
g
o
r
i
th
m
s
t
h
u
s
e
m
e
r
g
e
d
:
b
y
o
p
ti
m
i
zi
n
g
n
etw
o
r
k
s
t
r
u
ct
u
r
es
a
n
d
r
e
d
u
ci
n
g
p
a
r
a
m
et
er
s
/c
o
m
p
u
t
ati
o
n
,
th
ey
r
u
n
e
f
f
icien
tly
o
n
co
n
s
tr
ain
ed
d
ev
ices
wh
ile
en
s
u
r
in
g
ac
ce
p
tab
le
ac
cu
r
ac
y
,
m
ee
tin
g
r
ea
l
-
tim
e
an
d
lo
w
-
p
o
wer
r
eq
u
ir
em
en
ts
.
C
o
m
p
ar
in
g
lig
h
tweig
h
t
alg
o
r
ith
m
p
er
f
o
r
m
an
ce
is
o
f
g
r
ea
t
th
eo
r
etica
l
an
d
p
r
ac
tical
s
ig
n
if
ican
ce
.
T
h
eo
r
etica
lly
,
s
tu
d
y
in
g
in
d
ica
to
r
s
lik
e
ac
cu
r
ac
y
,
r
ec
all,
an
d
m
em
o
r
y
o
cc
u
p
a
n
cy
clar
if
ies
alg
o
r
ith
m
s
tr
en
g
th
s
,
wea
k
n
ess
es,
an
d
b
o
ttlen
ec
k
s
,
s
u
p
p
o
r
tin
g
o
p
tim
izatio
n
.
Pra
ctica
lly
,
it
h
elp
s
d
ev
elo
p
er
s
s
elec
t
o
p
tim
al
alg
o
r
ith
m
s
b
ased
o
n
n
ee
d
s
,
im
p
r
o
v
in
g
s
y
s
tem
s
tab
ilit
y
,
r
ed
u
cin
g
co
s
ts
,
an
d
e
x
p
an
d
in
g
a
p
p
licatio
n
s
co
p
e.
Fo
r
eig
n
r
esear
ch
o
n
lig
h
twei
g
h
t
f
ac
e
r
ec
o
g
n
itio
n
s
tar
ted
ea
r
ly
with
f
r
u
itfu
l
r
esu
lts
.
E
a
r
ly
s
tu
d
ies
f
o
cu
s
ed
o
n
o
p
tim
izin
g
tr
ad
it
io
n
al
alg
o
r
ith
m
s
(
p
r
i
n
cip
al
co
m
p
o
n
en
t
an
aly
s
is
(
PC
A)
,
li
n
ea
r
d
is
cr
im
in
an
t
an
aly
s
is
(
L
DA)
)
to
r
ed
u
ce
co
m
p
u
tatio
n
al
co
m
p
lex
ity
.
W
ith
d
ee
p
lear
n
in
g
,
lig
h
twe
ig
h
t
n
etwo
r
k
s
lik
e
Mo
b
ileNet
(
d
ep
t
h
wis
e
s
ep
ar
a
b
le
co
n
v
o
lu
tio
n
f
o
r
r
ed
u
ce
d
c
o
m
p
u
tatio
n
[
1
]
)
an
d
Sh
u
f
f
leNe
t
(
ch
an
n
el
s
h
u
f
f
lin
g
f
o
r
h
ig
h
er
ef
f
icien
c
y
)
wer
e
ap
p
lied
.
Sch
o
lar
s
also
ex
p
lo
r
ed
r
o
b
u
s
tn
ess
,
g
en
er
aliza
tio
n
,
an
d
m
u
lti
-
m
o
d
al
f
u
s
io
n
,
s
u
ch
as c
o
m
b
in
in
g
f
ac
e
r
ec
o
g
n
itio
n
with
f
in
g
er
p
r
i
n
t/iris r
ec
o
g
n
itio
n
[
2
]
.
Do
m
esti
c
r
esear
ch
h
as
also
m
ad
e
g
r
ea
t
p
r
o
g
r
ess
.
R
esear
c
h
in
s
titu
tio
n
s
an
d
u
n
iv
er
s
ities
p
r
o
p
o
s
ed
in
n
o
v
ativ
e
alg
o
r
ith
m
s
,
o
p
ti
m
izin
g
f
o
r
co
m
p
le
x
d
o
m
est
ic
s
ce
n
ar
io
s
(
v
ar
iab
le
lig
h
ti
n
g
,
d
iv
er
s
e
p
o
s
es,
o
cc
lu
s
io
n
s
)
to
im
p
r
o
v
e
a
d
ap
t
ab
ilit
y
[
3
]
.
D
o
m
esti
c
en
ter
p
r
i
s
es
p
r
o
m
o
ted
in
d
u
s
tr
ial
ap
p
li
ca
tio
n
in
s
ec
u
r
ity
,
f
in
an
ce
,
an
d
tr
an
s
p
o
r
tatio
n
,
ac
h
iev
in
g
g
o
o
d
s
o
cial
an
d
ec
o
n
o
m
ic
b
en
ef
its
.
Ho
wev
er
,
ex
is
tin
g
r
esear
ch
h
a
s
s
h
o
r
tco
m
in
g
s
:
in
c
o
n
s
is
ten
t
ev
alu
atio
n
in
d
icato
r
s
an
d
d
atas
ets
lead
to
lack
o
f
u
n
if
ie
d
p
e
r
f
o
r
m
an
ce
c
o
m
p
ar
is
o
n
s
tan
d
ar
d
s
;
b
ala
n
cin
g
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
an
d
r
es
o
u
r
ce
c
o
n
s
u
m
p
tio
n
r
em
ain
s
an
u
r
g
en
t iss
u
e.
T
h
is
s
tu
d
y
aim
s
to
d
esig
n
an
d
im
p
lem
en
t
a
d
ee
p
lear
n
i
n
g
-
b
ased
f
ac
e
r
ec
o
g
n
itio
n
atten
d
a
n
ce
s
y
s
tem
to
im
p
r
o
v
e
au
to
m
atio
n
,
ac
cu
r
ac
y
,
an
d
s
ec
u
r
ity
,
ad
d
r
ess
in
g
th
e
lim
itatio
n
s
o
f
tr
ad
itio
n
al
atten
d
an
ce
m
eth
o
d
s
.
I
t
v
er
if
ies
th
e
p
er
f
o
r
m
an
ce
im
p
r
o
v
em
e
n
t
o
f
co
m
b
in
in
g
tr
ad
i
tio
n
al
alg
o
r
ith
m
s
with
Mo
b
ile
Net
v
er
s
u
s
s
in
g
le
-
alg
o
r
ith
m
tr
ain
i
n
g
,
to
b
u
ild
a
n
ef
f
icien
t,
s
tab
le,
lo
w
-
co
s
t sy
s
tem
.
Aim
in
g
at
ex
is
tin
g
s
y
s
tem
l
im
itatio
n
s
,
th
is
s
tu
d
y
p
r
o
p
o
s
es
alg
o
r
ith
m
o
p
tim
izatio
n
s
c
h
em
es
to
en
h
an
ce
illu
m
in
atio
n
ad
a
p
tab
i
lity
,
d
ata
au
g
m
en
tatio
n
,
m
o
d
el
lig
h
tweig
h
tin
g
,
an
d
p
r
i
v
ac
y
p
r
o
tectio
n
,
en
s
u
r
in
g
ef
f
icien
t,
ac
cu
r
ate
r
ec
o
g
n
itio
n
in
co
m
p
lex
en
v
ir
o
n
m
en
ts
.
T
o
b
o
o
s
t
g
e
n
er
aliza
tio
n
,
lig
h
t
weig
h
t
m
o
d
els
(
e.
g
.
,
M
o
b
ileNet)
ar
e
in
tr
o
d
u
ce
d
with
q
u
an
ti
za
tio
n
an
d
p
r
u
n
in
g
[
4
]
,
r
ed
u
cin
g
r
eso
u
r
ce
o
cc
u
p
atio
n
an
d
im
p
r
o
v
in
g
r
ea
l
-
tim
e
p
er
f
o
r
m
an
ce
.
T
h
ese
o
p
tim
izatio
n
s
in
cr
ea
s
e
lo
w
-
co
m
p
u
tin
g
-
p
o
we
r
d
ev
ice
s
p
ee
d
b
y
3
0
% a
n
d
r
ed
u
ce
co
m
p
u
tatio
n
b
y
5
0
%.
2.
F
UNDA
M
E
N
T
AL
S O
F
F
ACE RE
CO
G
NIT
I
O
N
AL
G
O
RIT
H
M
2
.
1
.
O
v
er
v
iew
o
f
t
he
princip
le
o
f
f
a
ce
re
co
g
nitio
n t
ec
hn
o
l
o
g
y
As
an
im
p
o
r
ta
n
t
p
ar
t
o
f
b
i
o
m
etr
ic
id
en
tific
atio
n
,
f
ac
e
r
ec
o
g
n
itio
n
tech
n
o
lo
g
y
’
s
b
as
ic
p
r
o
ce
s
s
in
clu
d
es k
ey
lin
k
s
: f
ac
e
d
etec
t
io
n
,
f
ea
tu
r
e
ex
tr
ac
tio
n
,
an
d
f
ea
tu
r
e
co
m
p
a
r
is
o
n
.
Face
d
etec
tio
n
,
th
e
f
ir
s
t
s
tep
o
f
f
ac
e
r
ec
o
g
n
itio
n
,
aim
s
to
ac
cu
r
ately
lo
ca
te
f
ac
es
in
co
m
p
l
ex
im
ag
es
o
r
v
id
eo
s
.
E
a
r
ly
m
eth
o
d
s
r
elie
d
o
n
tr
a
d
itio
n
al
m
a
ch
in
e
lear
n
in
g
(
e
.
g
.
,
Haa
r
ca
s
ca
d
e
class
if
ier
with
Ad
a
b
o
o
s
t)
,
wh
ich
wer
e
s
en
s
itiv
e
to
illu
m
i
n
atio
n
an
d
p
o
s
e
ch
a
n
g
es,
with
lim
ited
ac
cu
r
ac
y
a
n
d
r
o
b
u
s
tn
ess
[
5
]
.
Ma
in
s
tr
ea
m
m
eth
o
d
s
n
o
w
u
s
e
C
NN
-
b
ased
alg
o
r
ith
m
s
lik
e
SS
D
an
d
YO
L
O,
wh
ich
au
t
o
m
atica
lly
lear
n
f
ac
ial
f
ea
tu
r
es
to
ac
h
iev
e
ef
f
icien
t
an
d
ac
cu
r
ate
d
etec
tio
n
in
co
m
p
lex
b
ac
k
g
r
o
u
n
d
s
[
6
]
;
f
o
r
ex
am
p
le,
SS
D
en
ab
les
m
u
lti
-
s
ca
le
f
ac
e
d
etec
tio
n
v
ia
m
u
lti
-
s
ca
le
f
ea
tu
r
e
m
ap
p
r
ed
ictio
n
,
im
p
r
o
v
in
g
r
ec
all
an
d
ac
cu
r
ac
y
.
Featu
r
e
ex
tr
ac
tio
n
,
a
k
ey
s
tep
,
ex
tr
ac
ts
r
ep
r
esen
tativ
e
f
ea
tu
r
es
f
r
o
m
d
etec
ted
f
ac
es.
T
r
ad
itio
n
al
m
eth
o
d
s
(
p
r
in
cip
al
co
m
p
o
n
en
t
an
aly
s
is
(
PC
A)
,
lin
ea
r
d
is
cr
im
in
an
t
an
aly
s
is
(
L
DA)
,
a
n
d
l
o
ca
l
b
in
ar
y
p
atter
n
s
(
L
B
P)
)
h
av
e
lim
ited
f
ea
tu
r
e
e
x
p
r
ess
io
n
f
o
r
co
m
p
lex
f
ac
e
v
ar
iatio
n
s
.
Dee
p
lear
n
in
g
m
eth
o
d
s
(
e.
g
.
,
Face
Net,
Dee
p
I
D)
h
a
v
e
r
e
v
o
lu
tio
n
ized
th
is
:
Face
Net
u
s
es
tr
ip
let
lo
s
s
to
m
ap
f
ac
es
in
to
h
ig
h
-
d
im
en
s
io
n
al
s
p
ac
e,
en
s
u
r
in
g
s
im
ilar
f
ea
tu
r
es
f
o
r
th
e
s
am
e
id
e
n
tity
an
d
d
is
tin
ct
f
ea
tu
r
es
f
o
r
d
if
f
e
r
en
t
i
d
en
titi
es
[
7
]
;
Dee
p
I
D
lear
n
s
ab
s
tr
ac
t d
is
cr
im
in
ativ
e
f
ea
tu
r
e
s
v
ia
m
u
lti
-
lay
er
C
NNs,
s
ig
n
if
ican
tly
im
p
r
o
v
in
g
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
[
8
]
.
Featu
r
e
co
m
p
a
r
is
o
n
,
th
e
f
in
al
s
tep
,
m
atch
es
ex
t
r
ac
ted
f
ea
t
u
r
es
with
d
atab
ase
f
ea
tu
r
es
t
o
ca
lcu
late
s
im
ilar
ity
an
d
d
eter
m
in
e
id
en
t
ity
[
9
]
.
C
o
m
m
o
n
m
et
h
o
d
s
in
cl
u
d
e
E
u
clid
ea
n
d
is
tan
ce
(
m
ea
s
u
r
in
g
s
p
atial
lin
ea
r
d
is
tan
ce
,
with
s
m
aller
v
alu
es
i
n
d
icatin
g
h
i
g
h
er
s
im
ilar
ity
)
a
n
d
co
s
in
e
s
im
ilar
ity
(
m
ea
s
u
r
in
g
v
ec
to
r
a
n
g
le,
with
v
alu
es
clo
s
er
t
o
1
in
d
icatin
g
h
ig
h
er
s
im
ilar
ity
)
[
1
0
]
,
[
1
1
]
.
A
s
im
ilar
ity
th
r
esh
o
ld
is
u
s
u
ally
s
et;
id
en
titi
es
ar
e
m
atch
ed
if
th
e
r
esu
lt e
x
ce
ed
s
t
h
e
th
r
esh
o
ld
.
2.
2
.
I
ntr
o
du
ct
io
n o
f
t
ra
diti
o
na
l f
a
ce
re
co
g
nitio
n a
lg
o
rit
hm
s
2
.
2
.
1
.
E
ig
enf
a
ce
s
Featu
r
e
f
ac
e
[
1
2
]
,
also
r
ef
er
r
ed
to
as
a
f
ea
tu
r
e
v
ec
to
r
s
et,
is
a
co
llect
io
n
o
f
v
ec
to
r
s
d
e
s
ig
n
ed
to
id
en
tify
h
u
m
a
n
f
ac
ial
f
ea
tu
r
es,
wh
ich
is
d
er
iv
e
d
f
r
o
m
m
ath
e
m
atica
l
an
d
s
tatis
tical
p
r
o
ce
s
s
i
n
g
o
f
d
iv
er
s
e
f
ac
ial
im
ag
es.
As
a
d
etec
tio
n
an
d
r
ec
o
g
n
itio
n
m
eth
o
d
f
o
r
d
eter
m
i
n
in
g
th
e
v
ar
ian
ce
o
f
im
ag
e
d
a
tasets
,
f
ea
tu
r
e
f
ac
e
em
p
lo
y
s
th
ese
v
ar
ian
ce
s
to
en
co
d
e
an
d
d
ec
o
d
e
f
ac
ial
f
ea
tu
r
es.
A
s
e
t
o
f
f
ea
tu
r
e
f
ac
es
co
n
s
titu
tes
a
co
llectio
n
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.
16
,
No
.
4
,
Au
g
u
s
t
20
26
:
2
0
4
2
-
2060
2044
o
f
u
n
iq
u
e
f
ea
tu
r
es,
with
th
e
n
u
m
b
er
o
f
f
ac
ial
im
ag
es
d
eter
m
in
ed
b
y
s
tatis
tical
an
aly
s
is
o
f
ex
ten
s
iv
e
d
atasets
.
Sin
ce
th
is
m
eth
o
d
u
tili
ze
s
s
tat
is
tical
d
atab
ases
in
s
tead
o
f
d
ig
ital
im
ag
es
d
ir
ec
tly
,
f
ac
ial
f
e
atu
r
es
ar
e
ass
ig
n
ed
m
ath
em
atica
l
v
alu
es,
an
d
a
n
y
h
u
m
a
n
f
ac
e
ca
n
b
e
r
eg
a
r
d
ed
as
a
co
m
b
in
atio
n
o
f
th
ese
v
alu
es
in
d
if
f
er
e
n
t
p
r
o
p
o
r
tio
n
s
.
2
.
2
.
2
.
F
is
herf
a
ce
s
Fis
h
er
f
ac
es
[
1
3
]
is
o
n
e
o
f
th
e
m
o
s
t
p
r
ev
alen
t
f
ac
ial
r
ec
o
g
n
iti
o
n
alg
o
r
ith
m
s
an
d
is
d
ee
m
ed
s
u
p
er
io
r
to
n
u
m
er
o
u
s
o
th
e
r
alter
n
ativ
es.
As
an
im
p
r
o
v
em
en
t
o
f
th
e
E
i
g
en
f
ac
es
alg
o
r
ith
m
,
it
is
g
en
er
a
lly
co
n
s
id
er
ed
m
o
r
e
ef
f
ec
tiv
e
th
an
th
e
eig
en
f
ac
e
m
eth
o
d
in
tr
ain
in
g
f
o
r
class
d
if
f
er
en
ce
s
.
I
ts
k
ey
ad
v
an
tag
es
lie
in
its
in
ter
p
o
latio
n
an
d
ex
tr
a
p
o
latio
n
ca
p
ab
ilit
ies,
as
well
as
its
r
o
b
u
s
tn
ess
to
v
ar
iatio
n
s
in
illu
m
in
atio
n
a
n
d
f
a
cial
ex
p
r
ess
io
n
s
.
I
t
h
as
b
ee
n
r
ep
o
r
te
d
th
at
th
e
ac
cu
r
ac
y
o
f
th
e
Fis
h
er
f
ac
es
alg
o
r
ith
m
r
ea
ch
es
u
p
to
9
3
%,
with
PC
A
em
p
lo
y
ed
f
o
r
p
r
ep
r
o
ce
s
s
in
g
.
2
.
2
.
3
.
L
o
ca
l bina
ry
pa
t
t
er
ns
his
t
o
g
ra
m
s
L
o
ca
l
B
in
ar
y
Patter
n
His
to
g
r
am
(
L
B
PH)
[
1
4
]
is
a
s
im
p
le
y
et
ef
f
ec
tiv
e
tex
tu
r
e
o
p
er
ato
r
in
co
m
p
u
ter
v
is
io
n
.
B
y
d
ef
in
in
g
th
e
n
ei
g
h
b
o
r
h
o
o
d
o
f
ea
ch
p
ix
el,
th
e
p
ix
el
th
r
esh
o
ld
s
in
th
e
im
ag
e
a
r
e
m
ar
k
ed
,
a
n
d
th
e
r
esu
ltin
g
v
alu
es
ar
e
tr
ea
ted
as
b
in
ar
y
n
u
m
b
er
s
.
I
n
th
e
l
ea
r
n
in
g
p
h
ase,
th
e
L
B
PH
al
g
o
r
ith
m
g
e
n
er
ates
h
is
to
g
r
am
lab
els
a
n
d
class
if
ies
im
ag
es
at
ea
ch
s
tag
e.
E
ac
h
h
is
to
g
r
am
c
o
r
r
esp
o
n
d
s
to
an
im
ag
e
in
th
e
tr
ai
n
in
g
s
et;
th
u
s
,
th
e
ac
tu
al
r
ec
o
g
n
itio
n
p
r
o
ce
s
s
in
v
o
lv
es c
o
m
p
a
r
in
g
an
y
two
h
is
to
g
r
a
m
-
b
ased
im
a
g
e
r
ep
r
esen
tatio
n
s
.
3.
I
NT
RO
D
UCT
I
O
N
AND
O
P
T
I
M
I
Z
AT
I
O
N
O
F
L
I
G
H
T
WE
I
G
H
T
NE
T
WO
RK
3
.
1
.
Cha
ra
c
t
er
is
t
ics o
f
M
o
bil
eNe
t
m
o
del
Mo
b
ileNet
is
a
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)
p
r
o
p
o
s
ed
b
y
Go
o
g
le,
s
p
ec
if
ically
d
esig
n
ed
f
o
r
m
o
b
ile
d
e
v
ices
an
d
e
m
b
ed
d
ed
s
y
s
tem
s
.
C
o
m
p
ar
ed
with
tr
ad
itio
n
al
co
n
v
o
lu
tio
n
al
n
eu
r
al
n
etwo
r
k
s
,
Mo
b
ileNet
s
ig
n
if
ican
tly
r
ed
u
ce
s
th
e
co
m
p
u
tatio
n
al
co
m
p
lex
ity
o
f
th
e
m
o
d
el
wh
ile
m
ain
tain
in
g
h
ig
h
p
er
f
o
r
m
a
n
ce
.
I
t
g
r
ea
tly
d
ec
r
ea
s
es
th
e
n
etwo
r
k
’
s
p
ar
a
m
eter
co
u
n
t
a
n
d
co
m
p
u
tatio
n
al
lo
ad
th
r
o
u
g
h
th
e
tech
n
iq
u
e
o
f
Dep
th
wis
e
Sep
ar
ab
le
C
o
n
v
o
lu
tio
n
s
[
1
5
]
,
[
1
6
]
.
3.
1
.
1
.
Dept
hwis
e
s
epa
ra
ble c
o
nv
o
lutio
n
T
h
e
tr
ad
itio
n
al
co
n
v
o
lu
tio
n
o
p
er
atio
n
p
er
f
o
r
m
s
s
tan
d
a
r
d
co
n
v
o
lu
tio
n
ca
lcu
latio
n
s
o
n
th
e
in
p
u
t
im
ag
e,
wh
ic
h
r
e
q
u
ir
es
a
lar
g
e
am
o
u
n
t
o
f
co
m
p
u
tatio
n
.
Par
tic
u
lar
ly
wh
e
n
t
h
e
n
u
m
b
e
r
o
f
ch
an
n
els
is
lar
g
e,
th
e
p
ar
am
eter
s
o
f
th
e
c
o
n
v
o
lu
tio
n
k
er
n
el
also
in
cr
ea
s
e
ex
p
o
n
e
n
tially
.
Ho
wev
er
,
d
ep
th
wis
e
s
ep
ar
ab
le
co
n
v
o
lu
tio
n
d
ec
o
m
p
o
s
es
th
e
s
tan
d
ar
d
co
n
v
o
lu
tio
n
i
n
to
two
s
tep
s
[
1
7
]
:
i)
Dep
th
wis
e
C
o
n
v
o
lu
tio
n
:
ea
ch
in
p
u
t
c
h
an
n
el
is
co
n
v
o
l
v
ed
in
d
e
p
en
d
e
n
tly
,
an
d
ea
ch
co
n
v
o
l
u
tio
n
k
er
n
el
o
n
ly
p
r
o
ce
s
s
es
d
ata
f
r
o
m
o
n
e
in
p
u
t
ch
an
n
el,
th
er
e
b
y
s
ig
n
if
ican
tly
r
ed
u
cin
g
th
e
co
m
p
u
tatio
n
al
lo
ad
;
an
d
ii)
Po
in
twis
e
C
o
n
v
o
lu
tio
n
:
n
am
ely
,
1
×1
co
n
v
o
lu
tio
n
,
wh
ich
is
r
esp
o
n
s
ib
le
f
o
r
f
u
s
in
g
th
e
o
u
tp
u
ts
o
f
d
ep
th
wis
e
co
n
v
o
l
u
tio
n
ac
r
o
s
s
ch
an
n
els.
T
h
e
co
m
p
u
tatio
n
al
v
o
lu
m
e
o
f
th
is
s
tep
is
s
m
all,
en
ab
lin
g
f
u
r
th
e
r
r
ed
u
ctio
n
in
co
m
p
u
tatio
n
wh
ile
p
r
eser
v
in
g
th
e
m
o
d
el
’
s
p
er
f
o
r
m
an
ce
.
T
h
e
ca
lcu
latio
n
am
o
u
n
t
o
f
s
ta
n
d
ar
d
c
o
n
v
o
lu
tio
n
ca
n
b
e
e
x
p
r
ess
ed
as f
o
llo
ws
:
=
×
×
×
×
×
(
1
)
D
f
is
th
e
s
ize
o
f
co
n
v
o
lu
tio
n
k
er
n
el
(
s
u
c
h
as
3
×
3
3
\
tim
es
3
3
×3
)
,
C
in
is
th
e
n
u
m
b
er
o
f
in
p
u
t
ch
a
n
n
els,
C
out
is
th
e
n
u
m
b
er
o
f
o
u
t
p
u
t c
h
a
n
n
els,
an
d
D
o
is
th
e
s
p
atial
s
ize
o
f
o
u
tp
u
t f
ea
tu
r
e
m
ap
.
Dep
th
s
ep
ar
ab
le
co
n
v
o
lu
tio
n
s
o
lv
es
th
e
s
tan
d
ar
d
v
o
lu
m
e
in
teg
r
al
in
to
Dep
th
wis
e
C
o
n
v
o
lu
tio
n
an
d
Po
in
twis
e
C
o
n
v
o
lu
tio
n
,
an
d
it
s
ca
lcu
latio
n
am
o
u
n
t
ca
n
b
e
e
x
p
r
ess
ed
as
:
FLO
P
S
de
pthwis
e
=
D
f
×
D
f
×
C
in
×
D
o
×
D
o
(
2
)
FLO
P
S
point
wis
e
=
C
in
×
C
out
×
D
o
×
D
o
(
3
)
T
h
er
ef
o
r
e,
th
e
to
tal
am
o
u
n
t
o
f
ca
lcu
latio
n
is
:
FLO
P
S
s
e
pa
r
a
ble
=
(
D
f
×
D
f
×
C
in
+
C
in
×
C
out
)
×
D
o
×
D
o
(
4
)
C
o
m
p
ar
is
o
n
o
f
ca
lc
u
latio
n
am
o
u
n
t
:
C
o
m
p
a
r
ed
with
s
tan
d
ar
d
c
o
n
v
o
lu
tio
n
,
th
e
co
m
p
u
tatio
n
al
co
m
p
lex
ity
o
f
d
ee
p
s
ep
ar
ab
le
co
n
v
o
l
u
tio
n
is
r
ed
u
ce
d
b
y
ab
o
u
t
:
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
C
o
mp
a
r
a
tive
p
erfo
r
ma
n
ce
a
n
a
lysi
s
o
f lig
h
tw
eig
h
t fa
ce
id
en
t
ifica
tio
n
a
lg
o
r
ith
m
(
Wu
yu
n
Wa
n
g
)
2045
=
1
+
1
2
(
5
)
Fo
r
th
e
3
×3
co
n
v
o
lu
tio
n
k
e
r
n
el,
ass
u
m
in
g
th
at
C
out
is
m
u
ch
lar
g
e
r
th
a
n
1
,
th
e
ca
lcu
lat
io
n
am
o
u
n
t
ca
n
b
e
r
ed
u
ce
d
b
y
a
b
o
u
t
8
-
9
tim
es,
w
h
ich
is
th
e
k
ey
f
ac
to
r
o
f
lig
h
tw
eig
h
t M
o
b
ileNet
[
1
8
]
.
3
.
1
.
2
.
Adj
us
t
a
ble w
idth a
nd
re
s
o
lutio
n
Mo
b
ileNet
also
p
o
s
s
ess
es
two
k
ey
ch
ar
ac
ter
is
tics
:
an
ad
ju
s
t
ab
le
W
id
th
Mu
ltip
lier
a
n
d
a
n
ad
ju
s
tab
le
R
eso
lu
tio
n
Mu
ltip
lier
[
1
9
]
.
Us
er
s
ca
n
ad
ju
s
t
th
ese
two
p
ar
a
m
eter
s
ac
co
r
d
in
g
to
s
p
ec
if
ic
a
p
p
licatio
n
s
ce
n
ar
io
s
:
i
)
Ad
ju
s
tab
le
W
id
th
:
B
y
ad
ju
s
tin
g
th
e
wid
th
co
ef
f
icien
t
(
α
)
,
th
e
wid
th
o
f
th
e
n
etwo
r
k
ca
n
b
e
co
n
tr
o
lled
,
th
er
eb
y
b
alan
cin
g
m
o
d
el
ac
c
u
r
ac
y
an
d
co
m
p
u
tatio
n
al
r
es
o
u
r
ce
s
.
A
s
m
aller
wid
th
co
e
f
f
icien
t
r
ed
u
ce
s
th
e
m
o
d
el’
s
co
m
p
u
tatio
n
al
an
d
s
to
r
ag
e
r
eq
u
i
r
em
en
ts
.
ii
)
Ad
ju
s
tab
le
R
eso
lu
tio
n
:
B
y
ad
ju
s
tin
g
th
e
r
eso
lu
tio
n
co
ef
f
icien
t
(
ρ
)
,
th
e
r
eso
lu
tio
n
o
f
th
e
in
p
u
t
im
ag
e
ca
n
b
e
c
o
n
t
r
o
lled
,
f
u
r
th
e
r
o
p
tim
izin
g
t
h
e
co
m
p
u
tatio
n
al
lo
ad
.
Alth
o
u
g
h
lo
wer
r
eso
lu
tio
n
m
a
y
s
lig
h
tly
r
ed
u
ce
ac
cu
r
ac
y
,
it
ca
n
s
ig
n
if
ican
tly
en
h
an
ce
p
r
o
c
ess
in
g
s
p
ee
d
u
n
d
er
th
e
co
n
d
itio
n
o
f
lim
ited
h
ar
d
w
ar
e
r
eso
u
r
ce
s
.
T
h
e
ad
ju
s
tm
en
t
o
f
th
ese
two
p
ar
am
eter
s
en
d
o
ws
Mo
b
ileNe
t
with
g
o
o
d
ad
ap
tab
ilit
y
ac
r
o
s
s
v
ar
io
u
s
d
ev
ices
an
d
a
p
p
licatio
n
s
,
allo
win
g
it
to
b
e
o
p
tim
ize
d
in
ac
c
o
r
d
an
ce
with
ac
tu
al
h
a
r
d
war
e
ca
p
ab
ilit
ies.
T
ab
le
1
p
r
esen
ts
a
co
m
p
ar
ativ
e
an
aly
s
is
o
f
d
if
f
er
e
n
t m
o
d
els
[
2
0
]
–
[
2
5
]
.
T
ab
le
1
.
Mo
d
el
co
m
p
ar
ativ
e
a
n
aly
s
is
tab
le
M
o
d
e
l
P
a
r
a
me
t
e
r
q
u
a
n
t
i
t
y
(
mi
l
l
i
o
n
s
)
C
o
m
p
u
t
a
t
i
o
n
a
l
c
o
m
p
l
e
x
i
t
y
(
F
LO
P
s,
b
i
l
l
i
o
n
s
)
To
p
-
1
a
c
c
u
r
a
c
y
(
%)
M
o
d
e
l
s
i
z
e
(
MB
)
VGG
-
16
1
3
8
15
.
5
71
.
5
5
2
8
R
e
sN
e
t
-
50
25
.
6
3
.
8
76
.
0
98
R
e
sN
e
t
-
101
44
.
5
7
.
6
77
.
4
1
7
1
M
o
b
i
l
e
N
e
t
V
1
4
.
2
0
.
57
70
.
6
16
M
o
b
i
l
e
N
e
t
V
2
3
.
4
0
.
32
72
.
0
14
M
o
b
i
l
e
N
e
t
V
3
2
.
9
0
.
22
73
.
0
t
w
e
l
v
e
3
.
2
.
Alg
o
rit
hm
ic
des
cr
iptio
n o
f
t
he
s
m
a
r
t
s
wit
ching
m
ec
ha
nis
m
T
h
e
f
o
llo
win
g
is
th
e
co
d
e
im
p
lem
en
tin
g
th
e
r
o
llb
ac
k
s
tr
ateg
y
.
Alg
o
r
itm
1
.
T
r
ad
itio
n
al
alg
o
r
it
h
m
p
r
io
r
ity
-
b
ased
b
ac
k
tr
ac
k
i
n
g
s
tr
ateg
y
Input: face_roi (face region), face_bbox (bounding box), frame_id
Output: (algorithm_name, predicted_id, confidence) or NULL
processed ← Resize(face_roi, 100×100)
processed ← CLAHE(processed, clipLimit=2.0, gridSize=8×8)
track_id ← GridQuantize(face_bbox, gridSize=20px)
// Stage 1: Try LBPH (fastest, ~1.8ms)
(id, conf) ← LBPH_Model.predict(processed)
IF conf ≤ T_lbph (default 50):
RETURN (‘lbph’, MapID(id), conf)
lbph_fail_count[track_id] += 1
// Stage 2: Environment assessment
brightness ← Median(processed)
aspect_ratio ← bbox.w / bbox.h
env_stress ← (brightness < 80) OR (aspect_ratio
∉
[0.80, 1.35])
need_fallback ← (lbph_fail_count ≥ 1) OR
(env_stress AND stability_count ≥ 2)
IF need_fallback:
(id, conf) ← Eigen_Model.predict(processed)
IF conf ≤ T_eigen (default 8000): RETURN (‘eigen’, id, conf)
(id, conf) ← Fisher_Model.predict(processed)
IF conf ≤ T_fisher (default 1000): RETURN (‘fisher’, id, conf)
RETURN NULL // All algorithms failed
4.
E
XP
E
R
I
M
E
N
T
I
M
P
L
E
M
E
NT
A
T
I
O
N
AND
R
E
S
UL
T
ANALY
SI
S
4
.
1
.
E
x
perim
ent
a
l
s
et
up
4
.
1
.
1
.
E
x
perim
ent
a
l desig
n
T
h
is
ex
p
er
im
en
t
em
p
lo
y
s
a
h
y
b
r
id
s
tr
ateg
y
co
m
b
in
in
g
"E
ig
en
f
ac
e,
Fis
h
er
f
ac
e,
L
B
PH"
an
d
"M
o
b
ileNet"
.
T
wo
p
u
b
licly
a
v
ailab
le
d
atasets
wer
e
u
tili
ze
d
t
o
estab
lis
h
an
Op
en
-
Set
I
d
en
ti
f
icatio
n
ev
alu
atio
n
f
r
am
ewo
r
k
.
a.
Kn
o
wn
I
d
e
n
titi
es
——
E
x
ten
d
ed
Yale
B
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.
16
,
No
.
4
,
Au
g
u
s
t
20
26
:
2
0
4
2
-
2060
2046
T
h
is
d
ataset
co
m
p
r
is
es
3
8
ca
teg
o
r
ies
o
f
f
ac
ial
id
en
titi
es,
wi
th
ea
ch
ca
teg
o
r
y
co
n
tain
in
g
ap
p
r
o
x
im
ately
6
4
f
ac
ial
im
ag
e
s
ca
p
tu
r
ed
u
n
d
er
d
if
f
e
r
en
t
lig
h
tin
g
an
g
les.
T
h
e
ex
p
er
im
en
t
u
s
ed
a
f
i
x
ed
r
an
d
o
m
s
ee
d
(
2
0
2
6
0
3
0
2
)
to
en
s
u
r
e
r
ep
r
o
d
u
cib
ilit
y
,
an
d
8
id
en
titi
es
wer
e
r
an
d
o
m
ly
s
elec
ted
as
s
tu
d
y
s
u
b
jects
f
o
r
h
ier
ar
ch
ical
s
am
p
lin
g
b
ased
o
n
lig
h
tin
g
co
n
d
itio
n
s
.
T
h
e
tr
ai
n
in
g
s
et
in
clu
d
es
2
5
im
ag
es
p
er
id
en
tity
—
9
u
n
d
er
s
im
p
le
lig
h
tin
g
,
8
u
n
d
er
m
o
d
er
ate
lig
h
tin
g
,
an
d
8
u
n
d
er
ch
allen
g
in
g
lig
h
tin
g
co
n
d
itio
n
s
—
wh
ile
th
e
test
s
et
co
n
tain
s
3
0
im
ag
es
p
er
i
d
e
n
tity
,
with
1
0
im
a
g
es
ea
ch
f
o
r
s
im
p
le,
m
o
d
er
ate,
a
n
d
c
h
allen
g
in
g
lig
h
tin
g
s
ce
n
ar
io
s
.
All sam
p
les f
r
o
m
th
e
tr
ain
in
g
an
d
test
s
ets ar
e
co
m
p
letely
is
o
lated
with
n
o
d
ata
o
v
er
lap
.
b.
Un
k
n
o
wn
I
d
en
titi
es
——
L
FW
T
h
is
ex
p
er
im
e
n
t
u
tili
ze
d
th
e
L
ab
eled
Face
s
in
th
e
W
ild
(
L
F
W
-
d
ee
p
f
u
n
n
eled
)
d
ataset
as
a
n
au
x
iliar
y
d
ata
s
o
u
r
ce
,
r
an
d
o
m
ly
s
elec
tin
g
1
2
0
in
d
i
v
id
u
als
with
u
n
k
n
o
wn
id
en
titi
es
an
d
e
x
tr
ac
tin
g
a
s
in
g
le
f
ac
ial
im
ag
e
f
r
o
m
ea
ch
f
o
r
q
u
an
titativ
e
ev
a
lu
atio
n
o
f
th
e
f
alse
ac
ce
p
tan
ce
r
ate
(
FAR
)
m
etr
ic
in
o
p
en
-
s
et
r
ec
o
g
n
itio
n
task
s
.
T
h
e
ex
p
e
r
im
en
tal
d
esig
n
wa
s
b
ased
o
n
t
h
e
f
o
llo
win
g
as
s
u
m
p
tio
n
s
:
tr
ad
itio
n
al
alg
o
r
it
h
m
s
ex
h
ib
it
f
aster
r
esp
o
n
s
e
tim
es
u
n
d
er
s
p
ec
if
i
c
co
n
d
itio
n
s
(
e.
g
.
,
n
o
r
m
al
lig
h
tin
g
an
d
f
r
o
n
tal
v
iewin
g
a
n
g
les),
wh
er
ea
s
th
e
Mo
b
ileNet
alg
o
r
ith
m
d
em
o
n
s
tr
ates
g
r
ea
ter
r
o
b
u
s
tn
ess
in
co
m
p
lex
s
ce
n
ar
i
o
s
(
e.
g
.
,
lo
w
-
l
ig
h
t
co
n
d
itio
n
s
an
d
v
ar
y
in
g
v
iewin
g
an
g
les).
T
h
r
o
u
g
h
ap
p
r
o
p
r
iate
alg
o
r
ith
m
c
o
m
b
in
atio
n
s
a
n
d
s
witch
in
g
s
tr
ateg
ies,
th
eo
r
etica
l
im
p
r
o
v
em
e
n
ts
in
ac
cu
r
ac
y
a
n
d
p
r
o
ce
s
s
in
g
ef
f
icien
cy
ca
n
b
e
ac
h
iev
ed
,
p
r
o
v
i
d
in
g
a
s
cien
tific
b
asis
f
o
r
alg
o
r
ith
m
s
elec
tio
n
in
atten
d
an
ce
s
y
s
tem
s
.
T
ab
les
2
an
d
3
p
r
esen
t
th
e
h
y
p
er
p
ar
a
m
eter
co
n
f
ig
u
r
atio
n
s
an
d
s
p
ec
if
ic
h
ar
d
war
e/so
f
twar
e
en
v
ir
o
n
m
e
n
t p
ar
am
eter
s
f
o
r
th
is
ex
p
er
im
en
t,
r
esp
ec
tiv
ely
.
T
ab
le
2
.
Hy
p
er
p
ar
a
m
eter
co
n
f
ig
u
r
atio
n
P
a
r
a
me
t
e
r
V
a
l
u
e
M
a
i
n
b
a
c
k
b
o
n
e
n
e
t
w
o
r
k
M
o
b
i
l
e
N
e
t
V
3
-
La
r
g
e
Emb
e
d
d
e
d
v
e
c
t
o
r
d
i
m
e
n
s
i
o
n
1
9
2
D
i
st
a
n
c
e
m
e
t
r
i
c
C
o
s
i
n
e
D
i
st
a
n
c
e
R
e
c
o
g
n
i
t
i
o
n
t
h
r
e
sh
o
l
d
0
.
3
2
7
9
（
C
a
l
i
b
r
a
t
e
d
a
n
d
o
p
t
i
m
i
z
e
d
u
s
i
n
g
F
1
）
I
n
p
u
t
si
z
e
1
1
2
×
1
1
2
A
l
i
g
n
m
e
n
t
m
e
t
h
o
d
5
p
t
_
a
f
f
i
n
e
N
o
r
mal
i
z
a
t
i
o
n
[
-
1
,
1
]
Li
n
e
a
r
sc
a
l
i
n
g
+
L
2
r
e
g
u
l
a
r
i
z
a
t
i
o
n
C
r
o
p
p
i
n
g
marg
i
n
0
.
2
(
2
0
%)
T
ab
le
3
.
Har
d
war
e
an
d
s
o
f
twar
e
en
v
ir
o
n
m
en
t
p
ar
am
eter
s
P
r
o
j
e
c
t
C
o
n
f
i
g
u
r
e
P
y
t
h
o
n
3
.
8
+
Te
n
s
o
r
F
l
o
w
2
.
1
3
.
0
(
C
P
U
)
K
e
r
a
s
2
.
1
3
.
1
O
p
e
n
C
V
4
.
9
.
0
.
8
0
(
o
p
e
n
c
v
-
c
o
n
t
r
i
b
-
p
y
t
h
o
n
)
N
u
mP
y
1
.
2
4
.
3
sci
k
i
t
-
l
e
a
r
n
1
.
3
.
2
d
l
i
b
1
9
.
2
4
.
6
(
b
i
n
a
r
y
)
f
a
c
e
-
r
e
c
o
g
n
i
t
i
o
n
1
.
3
.
0
I
n
f
e
r
e
n
c
e
D
e
v
i
c
e
C
P
U
o
n
l
y
4
.
1
.
2
.
H
y
perpa
ra
m
et
er
Co
nf
ig
ura
t
io
n
Sin
ce
t
h
e
h
y
b
r
i
d
f
r
a
m
e
wo
r
k
em
p
l
o
y
s
t
r
a
d
it
io
n
al
f
a
ce
r
e
co
g
n
iti
o
n
al
g
o
r
it
h
m
s
as
th
e
f
i
r
s
t
-
s
ta
g
e
class
i
f
i
er
,
t
h
e
d
ec
is
i
o
n
th
r
esh
o
l
d
s
f
o
r
ea
c
h
al
g
o
r
i
th
m
m
u
s
t
b
e
p
r
ed
ef
in
e
d
b
e
f
o
r
e
c
o
n
d
u
ct
in
g
c
o
m
p
a
r
a
ti
v
e
ex
p
e
r
i
m
e
n
ts
.
T
a
b
l
e
4
s
u
m
m
a
r
i
ze
s
th
e
th
r
esh
o
l
d
v
al
u
es
u
s
e
d
f
o
r
E
i
g
e
n
f
ac
e,
Fis
h
e
r
f
ac
e,
a
n
d
L
B
PH i
n
t
h
is
s
t
u
d
y
.
T
ab
le
4
.
T
r
a
d
itio
n
al
alg
o
r
ith
m
th
r
esh
o
ld
A
l
g
o
r
i
t
h
m
Th
r
e
s
h
o
l
d
M
e
a
n
i
n
g
LB
P
H
50
C
o
n
f
i
d
e
n
c
e
l
e
v
e
l
≤
Th
i
s
v
a
l
u
e
i
n
d
i
c
a
t
e
s a
ma
t
c
h
Ei
g
e
n
8
0
0
0
C
o
n
f
i
d
e
n
c
e
l
e
v
e
l
≤
Th
i
s
v
a
l
u
e
i
n
d
i
c
a
t
e
s a
ma
t
c
h
F
i
sh
e
r
1
0
0
0
C
o
n
f
i
d
e
n
c
e
l
e
v
e
l
≤
Th
i
s
v
a
l
u
e
i
n
d
i
c
a
t
e
s a
ma
t
c
h
Fo
u
r
g
r
o
u
p
s
o
f
co
m
p
ar
is
o
n
s
c
h
em
es
wer
e
d
esig
n
ed
in
th
e
e
x
p
er
im
en
t,
with
ea
ch
g
r
o
u
p
a
d
o
p
tin
g
th
e
co
m
p
ar
is
o
n
m
o
d
e
o
f
"m
ix
e
d
s
tr
ateg
y
v
s
s
in
g
le
Mo
b
ileNet"
:
a.
C
ase
1
:
E
ig
en
f
ac
e
an
d
M
o
b
ile
n
et
v
s
Mo
b
ilen
et
I
n
th
is
h
y
b
r
id
s
ch
em
e,
th
e
E
ig
en
f
ac
e
alg
o
r
ith
m
is
g
iv
e
n
p
r
io
r
ity
f
o
r
r
ap
id
r
ec
o
g
n
itio
n
.
W
h
en
E
ig
en
f
ac
e
r
ec
o
g
n
itio
n
f
ails
o
r
th
e
co
n
f
id
en
ce
is
in
s
u
f
f
icie
n
t,
it
a
u
to
m
atica
lly
s
witch
es
to
t
h
e
M
o
b
ileNet
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
C
o
mp
a
r
a
tive
p
erfo
r
ma
n
ce
a
n
a
lysi
s
o
f lig
h
tw
eig
h
t fa
ce
id
en
t
ifica
tio
n
a
lg
o
r
ith
m
(
Wu
yu
n
Wa
n
g
)
2047
alg
o
r
ith
m
f
o
r
in
-
d
ep
th
r
ec
o
g
n
itio
n
;
C
o
n
tr
o
l
Gr
o
u
p
B
is
s
et
u
p
,
wh
ich
o
n
ly
u
s
es
th
e
Mo
b
ileNet
alg
o
r
ith
m
f
o
r
r
ec
o
g
n
itio
n
.
T
h
e
E
ig
e
n
f
ac
e
al
g
o
r
ith
m
ac
h
ie
v
es
f
ast
p
r
o
ce
s
s
in
g
s
p
ee
d
u
n
d
er
n
o
r
m
al
lig
h
ti
n
g
co
n
d
itio
n
s
,
an
d
co
m
b
in
ed
with
t
h
e
d
ee
p
f
ea
tu
r
e
ex
tr
ac
tio
n
ca
p
a
b
ilit
y
o
f
Mo
b
ileNet,
th
e
ex
p
ec
ted
r
ec
o
g
n
itio
n
ac
c
u
r
ac
y
ca
n
b
e
im
p
r
o
v
e
d
.
b.
C
ase
2
:
Fis
h
er
f
ac
e
an
d
Mo
b
il
en
et
v
s
Mo
b
ilen
et
I
n
th
is
h
y
b
r
id
s
ch
em
e,
Fis
h
er
f
ac
e
alg
o
r
ith
m
is
u
s
ed
f
ir
s
t,
an
d
its
ad
v
an
tag
es
o
f
lin
ea
r
d
i
s
cr
im
in
an
t
an
aly
s
is
ar
e
u
s
ed
to
s
ep
ar
ate
c
ateg
o
r
ies
.
W
h
en
it
f
ails
,
it
i
s
s
witch
ed
to
Mo
b
ilen
et,
an
d
th
e
B
co
n
tr
o
l
g
r
o
u
p
is
s
et,
an
d
o
n
l
y
Mo
b
ilen
et
al
g
o
r
ith
m
is
u
s
ed
f
o
r
id
en
tific
atio
n
.
Fis
h
er
f
ac
e
alg
o
r
ith
m
p
e
r
f
o
r
m
s
well
in
m
u
lti
-
p
er
s
o
n
s
ce
n
es,
an
d
th
e
ex
p
ec
te
d
ac
cu
r
ac
y
ca
n
b
e
im
p
r
o
v
ed
a
f
ter
co
m
b
in
in
g
with
Mo
b
ileNet
.
c.
C
ase
3
:
L
B
P
H
an
d
Mo
b
ilen
et
v
s
Mo
b
ilen
et
I
n
th
is
h
y
b
r
id
s
ch
e
m
e,
L
B
PH
alg
o
r
ith
m
is
u
s
ed
f
ir
s
t,
a
n
d
b
y
u
s
in
g
its
r
o
b
u
s
tn
ess
to
illu
m
in
atio
n
ch
an
g
es,
it
is
s
w
itch
ed
to
Mo
b
ileNet
u
n
d
er
co
m
p
le
x
illu
m
in
atio
n
co
n
d
itio
n
s
,
an
d
th
e
B
co
n
tr
o
l
g
r
o
u
p
is
s
et,
an
d
o
n
ly
Mo
b
ileNet
alg
o
r
ith
m
is
u
s
ed
f
o
r
r
ec
o
g
n
itio
n
.
T
h
e
L
B
PH
alg
o
r
ith
m
is
s
tab
le
in
th
e
lig
h
t
c
h
an
g
in
g
s
ce
n
e,
an
d
th
e
e
x
p
ec
ted
a
cc
u
r
ac
y
ca
n
b
e
im
p
r
o
v
ed
a
f
ter
co
m
b
in
in
g
with
M
o
b
ileNet
.
d.
C
ase
4
:
T
r
ad
itio
n
al
alg
o
r
ith
m
co
m
b
in
atio
n
an
d
M
o
b
ileNet
v
s
Mo
b
ileNet
I
n
th
is
h
y
b
r
id
s
ch
em
e,
a
c
o
m
p
lete
co
m
b
in
atio
n
o
f
tr
ad
it
io
n
al
alg
o
r
ith
m
s
,
n
am
ely
E
i
g
en
f
ac
e
+
Fis
h
er
f
ac
e
+
L
B
PH,
i
s
u
s
ed
as
th
e
f
ir
s
t
lay
er
o
f
r
ec
o
g
n
itio
n
.
W
h
en
th
is
co
m
b
in
ed
r
ec
o
g
n
itio
n
f
ails
,
Mo
b
ileNet
is
ac
tiv
ated
as
th
e
s
ec
o
n
d
lay
er
o
f
r
ec
o
g
n
itio
n
;
C
o
n
tr
o
l
Gr
o
u
p
B
is
s
et
u
p
,
wh
ich
o
n
ly
u
s
es
th
e
Mo
b
ileNet
alg
o
r
ith
m
f
o
r
r
e
co
g
n
itio
n
.
T
h
e
co
m
b
in
atio
n
o
f
tr
ad
itio
n
al
al
g
o
r
ith
m
s
a
ch
iev
es
v
er
y
h
ig
h
r
ec
o
g
n
itio
n
ac
c
u
r
ac
y
,
an
d
th
e
ex
p
ec
ted
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
ca
n
b
e
s
ig
n
if
ican
tly
im
p
r
o
v
ed
af
ter
in
teg
r
atio
n
with
Mo
b
ileNet.
4
.
1
.
3
.
F
o
ur
ex
perim
ent
a
l scena
rio
s
illu
s
t
ra
t
e
t
he
po
int
B
ased
o
n
th
e
E
x
ten
d
e
d
Yale
B
d
ataset
'
s
il
lu
m
in
atio
n
a
n
g
le
an
n
o
tatio
n
s
,
th
e
test
im
ag
es
ar
e
ca
teg
o
r
ized
in
to
th
r
ee
d
if
f
icu
lt
y
s
ce
n
ar
io
s
p
lu
s
o
n
e
b
atch
p
r
o
ce
s
s
in
g
s
ce
n
ar
io
in
T
ab
le
5
.
T
ab
le
5
.
Scen
e
d
escr
ip
tio
n
tab
le
S
c
e
n
e
C
l
a
s
si
f
y
i
n
g
R
u
l
e
s
P
h
y
s
i
c
a
l
M
e
a
n
i
n
g
s
Ea
sy
(
n
o
r
ma
l
l
i
g
h
t
i
n
g
)
max
(
|
a
z
i
m
u
t
h
|
,
|
e
l
e
v
a
t
i
o
n
|
)
≤
2
0
°
Th
e
l
i
g
h
t
so
u
r
c
e
i
s
p
o
s
i
t
i
o
n
e
d
o
n
t
h
e
f
r
o
n
t
o
r
n
e
a
r
t
h
e
f
r
o
n
t
,
p
r
o
v
i
d
i
n
g
u
n
i
f
o
r
m
i
l
l
u
mi
n
a
t
i
o
n
a
n
d
l
o
w
r
e
c
o
g
n
i
t
i
o
n
d
i
f
f
i
c
u
l
t
y
.
M
e
d
i
u
m (m
o
d
e
r
a
t
e
l
i
g
h
t
)
2
0
°
<
ma
x
≤
5
0
°
Th
e
l
i
g
h
t
so
u
r
c
e
i
s
o
f
f
set
,
c
r
e
a
t
i
n
g
n
o
t
i
c
e
a
b
l
e
sh
a
d
o
w
s
a
n
d
p
a
r
t
i
a
l
l
y
o
b
s
c
u
r
i
n
g
f
a
c
i
a
l
f
e
a
t
u
r
e
s.
H
a
r
d
(
l
o
w
l
i
g
h
t
)
max
>
5
0
°
Ex
t
r
e
me
f
r
o
n
t
l
i
g
h
t
i
n
g
/
b
a
c
k
l
i
g
h
t
i
n
g
,
l
a
r
g
e
-
a
r
e
a
s
h
a
d
o
w
s
o
r
o
v
e
r
e
x
p
o
s
u
r
e
mak
e
i
d
e
n
t
i
f
i
c
a
t
i
o
n
e
x
t
r
e
me
l
y
d
i
f
f
i
c
u
l
t
.
b
a
t
c
h
p
r
o
c
e
ss
M
i
x
e
d
f
o
r
a
l
l
sce
n
a
r
i
o
s
Te
st
t
h
e
t
h
r
o
u
g
h
p
u
t
a
n
d
s
t
a
b
i
l
i
t
y
o
f
m
u
l
t
i
-
u
ser
c
o
n
t
i
n
u
o
u
s rec
o
g
n
i
t
i
o
n
4
.
1
.
4
.
Rea
s
o
ns
f
o
r
s
elec
t
ing
t
he
ev
a
lua
t
io
n ind
ica
t
o
rs
T
h
is
s
tu
d
y
s
elec
ts
Acc
u
r
ac
y
,
Ma
cr
o
-
F1
,
Fals
e
Acc
ep
tan
ce
R
ate
(
FAR
)
,
Fals
e
R
ejec
tio
n
R
ate
(
FR
R
)
,
AUC,
an
d
E
q
u
al
E
r
r
o
r
R
ate
(
E
E
R
)
as e
v
alu
atio
n
m
etr
ics:
−
Acc
u
r
ac
y
is
th
e
o
v
er
all
r
ec
o
g
n
itio
n
co
r
r
ec
tn
ess
r
ate,
in
tu
itiv
ely
r
ef
lectin
g
th
e
r
ec
o
g
n
itio
n
p
er
f
o
r
m
a
n
ce
o
f
k
n
o
wn
id
e
n
titi
es in
clo
s
ed
-
s
et
s
ce
n
ar
io
s
;
−
Ma
cr
o
-
F1
co
m
p
r
eh
en
s
iv
ely
b
alan
ce
s
p
r
ec
is
io
n
an
d
r
ec
a
ll,
av
o
id
s
ev
alu
atio
n
b
ias
ca
u
s
ed
b
y
class
im
b
alan
ce
,
an
d
eq
u
ally
co
n
s
id
er
s
th
e
r
ec
o
g
n
itio
n
e
f
f
ec
t
o
f
ea
ch
id
en
tity
,
s
u
itab
le
f
o
r
f
air
n
e
s
s
ev
alu
atio
n
in
m
u
lti
-
id
en
tity
s
ce
n
ar
io
s
;
−
FAR
r
ep
r
esen
ts
th
e
p
r
o
p
o
r
tio
n
o
f
u
n
k
n
o
w
n
id
en
titi
es
in
co
r
r
ec
tly
ac
ce
p
ted
,
co
r
r
esp
o
n
d
in
g
to
th
e
s
ec
u
r
ity
r
is
k
o
f
p
r
o
x
y
clo
c
k
-
in
i
n
atte
n
d
an
ce
s
y
s
tem
s
,
an
d
is
u
s
ed
t
o
m
ea
s
u
r
e
t
h
e
ab
ilit
y
o
f
o
p
en
-
s
et
r
ec
o
g
n
itio
n
task
s
to
r
ejec
t u
n
f
am
iliar
id
e
n
titi
es;
−
FR
R
in
d
icate
s
th
e
p
r
o
b
a
b
ilit
y
th
at
k
n
o
w
n
leg
itima
te
id
en
tit
ies
ar
e
in
co
r
r
ec
tly
r
ejec
ted
,
d
i
r
ec
tly
af
f
ec
tin
g
u
s
er
ex
p
er
ien
ce
,
an
d
is
u
s
ed
t
o
ev
alu
ate
th
e
m
o
d
el’
s
ab
ilit
y
to
ac
ce
p
t le
g
itima
te
u
s
er
s
;
−
AUC
is
th
e
ar
ea
u
n
d
er
th
e
R
OC
cu
r
v
e,
n
o
t
af
f
ec
ted
b
y
th
e
class
if
icatio
n
th
r
esh
o
ld
,
an
d
ca
n
ev
alu
ate
th
e
o
v
er
all
d
is
cr
im
in
atio
n
p
er
f
o
r
m
an
ce
o
f
th
e
m
o
d
el
f
r
o
m
a
th
r
es
h
o
ld
-
f
r
ee
p
er
s
p
ec
tiv
e;
−
E
E
R
is
th
e
eq
u
al
er
r
o
r
r
ate
wh
en
FAR
eq
u
als
F
R
R
,
a
g
en
e
r
al
s
tan
d
ar
d
m
etr
ic
in
th
e
f
ield
o
f
b
io
m
etr
ics,
f
ac
ilit
atin
g
h
o
r
izo
n
tal
p
er
f
o
r
m
an
ce
co
m
p
a
r
is
o
n
with
o
th
e
r
al
g
o
r
ith
m
m
o
d
els.
4.
2
.
E
x
perim
ent
a
l im
plem
e
nta
t
io
n
4.
2
.
1
.
T
ra
ini
ng
s
t
a
g
e
I
n
th
is
e
x
p
er
im
en
t,
th
e
tr
ain
i
n
g
p
h
ase
was
d
esig
n
e
d
u
n
d
e
r
th
e
f
r
am
ewo
r
k
o
f
t
h
e
m
ix
e
d
s
tr
ateg
y
,
aim
in
g
to
v
e
r
if
y
th
e
p
er
f
o
r
m
an
ce
im
p
r
o
v
em
en
t
e
f
f
ec
t
o
f
t
h
e
co
m
b
i
n
atio
n
o
f
tr
ad
itio
n
al
alg
o
r
ith
m
s
an
d
th
e
Mo
b
ileNet
alg
o
r
ith
m
co
m
p
ar
e
d
with
s
in
g
le
-
alg
o
r
ith
m
tr
ain
in
g
.
I
n
th
e
tr
ain
in
g
p
h
ase,
th
e
ex
p
er
im
en
t f
o
llo
we
d
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.
16
,
No
.
4
,
Au
g
u
s
t
20
26
:
2
0
4
2
-
2060
2048
th
e
p
r
in
ci
p
le
o
f
"
g
r
ad
u
al
tr
ain
in
g
".
On
th
e
b
asis
o
f
m
ain
tai
n
in
g
t
h
e
s
tab
ilit
y
o
f
th
e
ex
is
tin
g
tr
ain
i
n
g
p
r
o
ce
s
s
,
th
e
m
ix
ed
s
tr
ateg
y
tr
ain
i
n
g
m
ec
h
an
is
m
o
f
"h
ier
a
r
ch
ical
tr
ai
n
in
g
+
co
llab
o
r
ativ
e
o
p
tim
izatio
n
"
was
g
r
ad
u
ally
in
tr
o
d
u
ce
d
to
en
s
u
r
e
th
e
r
el
iab
ilit
y
an
d
r
ep
r
o
d
u
cib
ilit
y
o
f
th
e
tr
ain
in
g
r
esu
lts
.
T
h
r
o
u
g
h
th
e
in
tellig
en
t
alg
o
r
ith
m
s
elec
tio
n
m
ec
h
an
is
m
,
co
llab
o
r
ativ
e
tr
ain
in
g
o
f
tr
ad
itio
n
al
alg
o
r
ith
m
s
an
d
d
ee
p
lear
n
in
g
alg
o
r
ith
m
s
was
r
ea
lized
.
T
h
e
tr
ain
i
n
g
s
tr
ateg
y
was
b
ased
o
n
th
e
f
o
llo
win
g
co
r
e
id
ea
s
:
tr
a
d
itio
n
al
a
lg
o
r
ith
m
s
h
a
v
e
t
h
e
ad
v
an
tag
e
o
f
f
ast
c
o
n
v
e
r
g
en
ce
in
s
p
ec
if
ic
s
ce
n
es,
wh
ile
th
e
M
o
b
ileNet
alg
o
r
ith
m
h
as
s
tr
o
n
g
er
f
ea
tu
r
e
lear
n
in
g
a
b
ilit
y
in
co
m
p
lex
s
ce
n
es.
I
n
th
eo
r
y
,
b
o
th
t
r
ain
in
g
ef
f
icien
c
y
an
d
m
o
d
el
p
e
r
f
o
r
m
a
n
ce
ca
n
b
e
im
p
r
o
v
e
d
th
r
o
u
g
h
r
ea
s
o
n
a
b
le
tr
ain
in
g
co
m
b
in
atio
n
an
d
p
ar
a
m
eter
o
p
tim
izatio
n
.
Fig
u
r
e
1
p
r
esen
ts
th
e
co
m
p
a
r
is
o
n
o
f
tr
ain
in
g
tim
e
b
etwe
e
n
tr
ad
itio
n
al
alg
o
r
ith
m
s
an
d
Mo
b
ileNet
u
n
d
er
d
if
f
e
r
en
t
d
ata
s
ca
les,
wh
ile
Fig
u
r
e
2
s
h
o
ws
th
e
co
m
p
ar
is
o
n
o
f
tr
ain
in
g
tim
e
b
etwe
en
th
e
h
y
b
r
id
s
tr
ateg
y
an
d
Mo
b
ileNet,
with
th
e
d
ata
s
ca
le
r
an
g
in
g
f
r
o
m
5
0
to
2
,
0
0
0
im
ag
es,
illu
s
tr
atin
g
th
e
g
r
o
wth
tr
en
d
o
f
tr
ain
in
g
tim
e.
Alth
o
u
g
h
th
e
tr
ain
i
n
g
ti
m
e
o
f
th
e
h
y
b
r
id
s
tr
ateg
y
is
s
lig
h
tly
lo
n
g
er
t
h
an
th
at
o
f
th
e
s
in
g
le
Mo
b
ileNet,
th
r
o
u
g
h
h
ier
a
r
ch
ical
tr
ain
in
g
an
d
co
llab
o
r
ativ
e
o
p
tim
izatio
n
,
th
e
c
o
m
b
in
ati
o
n
o
f
L
B
PH
an
d
Mo
b
ileNet
ac
h
iev
es
a
1
0
%
-
1
5
%
o
p
tim
i
za
tio
n
in
tr
ain
in
g
ef
f
icien
c
y
,
wh
ich
v
er
if
ies
th
e
ef
f
ec
tiv
e
n
ess
o
f
th
e
h
y
b
r
id
s
tr
ateg
y
tr
ain
in
g
m
ec
h
an
is
m
.
T
ab
le
6
p
r
esen
ts
th
e
co
m
p
ar
is
o
n
o
f
tr
ain
in
g
co
n
v
er
g
en
ce
s
p
ee
d
am
o
n
g
d
if
f
er
en
t
h
y
b
r
id
s
ch
em
es.
Fig
u
r
e
1
.
C
o
m
p
a
r
is
o
n
d
iag
r
am
o
f
tr
ain
in
g
tim
e
o
f
s
in
g
le
alg
o
r
ith
m
Fig
u
r
e
2
.
C
o
m
p
a
r
is
o
n
ch
a
r
t o
f
tr
ain
in
g
tim
e
o
f
h
y
b
r
id
al
g
o
r
it
h
m
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
C
o
mp
a
r
a
tive
p
erfo
r
ma
n
ce
a
n
a
lysi
s
o
f lig
h
tw
eig
h
t fa
ce
id
en
t
ifica
tio
n
a
lg
o
r
ith
m
(
Wu
yu
n
Wa
n
g
)
2049
T
ab
le
6
.
C
o
m
p
a
r
is
o
n
o
f
tr
ain
i
n
g
co
n
v
er
g
e
n
ce
s
p
ee
d
C
a
se
Tr
a
d
i
t
i
o
n
a
l
a
l
g
o
r
i
t
h
m
t
i
m
e
(
s
)
M
o
b
i
l
e
N
e
t
t
i
me
(
s
)
To
t
a
l
t
i
me
(
s
)
1
2
.
3
6
.
2
8
.
5
2
4
.
1
5
.
4
9
.
5
3
1
.
8
6
.
0
7
.
8
4
8
.
2
2
.
0
10
.
2
Fro
m
Fig
u
r
es
1
an
d
2
an
d
T
a
b
le
6
,
it
ca
n
b
e
c
o
n
clu
d
ed
th
a
t
th
e
L
B
PH
alg
o
r
ith
m
ex
h
ib
it
s
th
e
m
o
s
t
s
tab
le
co
n
v
er
g
e
n
ce
am
o
n
g
tr
ad
itio
n
al
alg
o
r
ith
m
s
,
r
eq
u
ir
i
n
g
alm
o
s
t
n
o
iter
ativ
e
p
r
o
ce
s
s
.
T
h
e
co
n
v
er
g
e
n
ce
p
r
o
ce
s
s
es
o
f
th
e
E
i
g
en
f
ac
e
a
n
d
Fis
h
er
f
ac
e
alg
o
r
ith
m
s
ar
e
s
t
ab
le
an
d
p
r
ed
ictab
le.
Du
e
to
th
e
ad
o
p
tio
n
o
f
a
p
r
e
-
tr
ain
ed
m
o
d
el,
th
e
c
o
n
v
e
r
g
en
ce
p
r
o
ce
s
s
o
f
th
e
Mo
b
ileNet
alg
o
r
ith
m
is
r
elativ
ely
co
m
p
lex
,
b
u
t
its
o
v
er
all
s
tab
ilit
y
is
g
o
o
d
.
T
h
r
o
u
g
h
th
e
in
tellig
en
t
alg
o
r
ith
m
s
elec
tio
n
m
ec
h
a
n
is
m
,
th
e
co
n
v
er
g
en
ce
o
f
t
h
e
h
y
b
r
id
s
tr
ateg
y
is
m
o
r
e
s
tab
le,
with
a
tr
ain
in
g
s
u
cc
ess
r
ate
o
f
o
v
er
9
8
%.
Fig
u
r
es
3
a
n
d
4
a
r
e
g
r
ap
h
s
o
f
tr
ain
i
n
g
l
o
s
s
an
d
c
o
m
p
ar
i
s
o
n
o
f
co
n
v
er
g
e
n
ce
s
p
ee
d
,
r
esp
ec
tiv
ely
,
illu
s
tr
atin
g
th
e
tr
ain
in
g
p
r
o
ce
s
s
o
f
th
e
Mo
b
ileNet
m
o
d
el
o
v
er
3
0
ep
o
ch
s
:
th
e
tr
ain
in
g
l
o
s
s
d
ec
r
ea
s
ed
f
r
o
m
2
.
8
4
7
to
0
.
0
5
3
(
a
9
8
.
1
%
d
ec
li
n
e)
,
an
d
th
e
v
alid
atio
n
lo
s
s
d
ec
r
ea
s
ed
f
r
o
m
2
.
8
9
1
to
0
.
0
5
2
,
d
em
o
n
s
tr
atin
g
g
o
o
d
co
n
v
er
g
en
ce
.
C
o
m
p
ar
e
th
e
c
o
n
v
er
g
e
n
ce
o
f
th
r
ee
lear
n
in
g
r
at
es a
r
e
0
.
0
0
1
,
0
.
0
0
0
5
an
d
0
.
0
0
0
1
.
Fig
u
r
e
3
.
Mo
b
ilit
y
tr
ain
in
g
lo
s
s
cu
r
v
e
Fig
u
r
e
4
.
C
o
m
p
a
r
is
o
n
g
r
ap
h
o
f
co
n
v
e
r
g
en
ce
s
p
ee
d
Fig
u
r
e
5
is
an
o
v
e
r
-
f
itti
n
g
d
et
ec
tio
n
an
aly
s
is
d
iag
r
am
,
f
r
o
m
wh
ich
it
ca
n
b
e
o
b
s
er
v
ed
th
at
th
e
m
o
d
el
h
as
g
o
o
d
g
e
n
er
aliza
tio
n
ab
ilit
y
.
Fig
u
r
e
6
is
a
co
m
p
ar
is
o
n
d
iag
r
am
o
f
tr
ain
i
n
g
an
d
v
alid
at
io
n
ac
cu
r
ac
y
.
T
h
e
cu
r
v
es
o
f
tr
ai
n
in
g
lo
s
s
an
d
v
al
id
atio
n
lo
s
s
in
d
icate
s
lig
h
t
o
v
er
-
f
itti
n
g
,
with
th
e
g
a
p
b
etwe
en
tr
ain
in
g
ac
c
u
r
ac
y
an
d
v
alid
atio
n
ac
cu
r
ac
y
b
ein
g
ap
p
r
o
x
im
ately
5
%,
wh
ich
is
r
e
aso
n
ab
le
an
d
c
o
n
f
ir
m
s
th
e
m
o
d
el’
s
s
tab
ilit
y
.
T
h
e
f
in
al
lo
s
s
g
ap
is
o
n
ly
0
.
0
0
1
,
in
d
icatin
g
a
lo
w
r
is
k
o
f
o
v
er
-
f
itt
in
g
.
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.
16
,
No
.
4
,
Au
g
u
s
t
20
26
:
2
0
4
2
-
2060
2050
Fig
u
r
e
5
.
Ov
e
r
-
f
itti
n
g
d
etec
tio
n
an
aly
s
is
d
iag
r
am
Fig
u
r
e
6
.
C
o
m
p
a
r
is
o
n
ch
a
r
t o
f
tr
ain
in
g
an
d
v
er
if
icatio
n
ac
cu
r
ac
y
4
.
2
.
2
.
Rea
s
o
nin
g
s
t
a
g
e
I
n
th
is
ex
p
er
im
e
n
t,
th
e
r
ea
s
o
n
in
g
p
h
ase
was
d
esig
n
ed
u
n
d
er
th
e
f
r
am
ew
o
r
k
o
f
th
e
m
i
x
e
d
s
tr
ateg
y
,
aim
in
g
to
v
er
if
y
th
e
p
er
f
o
r
m
a
n
ce
im
p
r
o
v
em
en
t
ef
f
ec
t
o
f
c
o
m
b
in
ed
r
ea
s
o
n
in
g
with
tr
ad
iti
o
n
al
alg
o
r
ith
m
s
an
d
th
e
Mo
b
ileNet
alg
o
r
ith
m
co
m
p
ar
ed
with
s
in
g
le
-
alg
o
r
ith
m
r
ea
s
o
n
in
g
.
T
h
e
r
ea
s
o
n
in
g
p
h
ase
e
x
p
er
im
e
n
t
f
o
llo
wed
th
e
p
r
in
ci
p
le
o
f
"in
t
ellig
en
t
s
witch
in
g
".
On
t
h
e
b
asis
o
f
m
ain
tain
in
g
th
e
s
tab
il
ity
o
f
t
h
e
ex
is
tin
g
r
ea
s
o
n
in
g
p
r
o
ce
s
s
,
th
e
m
ix
ed
s
tr
ateg
y
r
ea
s
o
n
in
g
m
ec
h
an
is
m
was
g
r
ad
u
ally
in
tr
o
d
u
ce
d
to
e
n
s
u
r
e
th
e
r
eliab
ilit
y
an
d
r
ep
r
o
d
u
cib
ilit
y
o
f
th
e
r
e
aso
n
in
g
r
esu
lts
.
T
ab
le
7
p
r
e
s
en
ts
a
co
m
p
ar
is
o
n
o
f
s
in
g
le
-
f
r
am
e
p
r
o
ce
s
s
in
g
p
er
f
o
r
m
an
ce
am
o
n
g
d
if
f
e
r
en
t
s
ch
em
es:
Sch
em
e
1
in
cr
ea
s
ed
b
y
5
0
%,
Sch
em
e
2
b
y
6
0
%,
Sch
em
e
3
b
y
2
6
%,
an
d
Sch
em
e
4
b
y
8
0
%.
I
n
ter
m
s
o
f
s
in
g
le
-
f
r
am
e
p
r
o
ce
s
s
in
g
p
er
f
o
r
m
an
ce
,
th
e
p
r
o
ce
s
s
in
g
d
u
r
atio
n
o
f
th
e
h
y
b
r
id
s
tr
ateg
y
is
s
ig
n
if
ican
tl
y
lo
n
g
er
,
b
u
t
th
e
ef
f
icien
c
y
o
f
th
e
h
y
b
r
id
s
tr
ateg
y
is
o
b
v
io
u
s
ly
im
p
r
o
v
e
d
in
b
atch
p
r
o
ce
s
s
in
g
.
T
h
e
b
atch
p
r
o
ce
s
s
in
g
ef
f
icien
cy
a
n
aly
s
is
o
f
th
e
h
y
b
r
i
d
s
tr
ateg
y
is
s
h
o
w
n
in
T
ab
le
8
,
an
d
th
e
r
ea
l
-
tim
e
p
er
f
o
r
m
an
ce
ev
alu
at
io
n
is
p
r
esen
ted
in
T
ab
le
9
.
T
ab
le
7
.
Sin
g
le
f
r
am
e
p
r
o
ce
s
s
in
g
p
er
f
o
r
m
a
n
ce
tim
etab
le
C
a
se
Tr
a
d
i
t
i
o
n
a
l
a
l
g
o
r
i
t
h
m
t
i
m
e
(
ms
)
M
o
b
i
l
e
N
e
t
t
i
me
(
s
)
To
t
a
l
t
i
me
(
s
)
1.
1
2
5
2
5
0
3
7
5
2.
1
5
0
2
5
0
4
0
0
3.
65
2
5
0
3
1
5
4.
1
1
3
2
5
0
4
5
0
T
ab
le
8
.
C
o
m
p
a
r
is
o
n
tab
le
o
f
b
atch
p
r
o
ce
s
s
in
g
tim
e
C
a
se
S
i
n
g
l
e
f
r
a
me
p
r
o
c
e
ssi
n
g
t
i
m
e
(
ms
)
B
a
t
c
h
p
r
o
c
e
ssi
n
g
t
i
m
e
(
ms
)
I
n
c
r
e
a
se
(
%)
1
3
7
5
2
8
0
25
%
2
4
0
0
3
0
0
25
%
3
3
1
5
2
5
0
21
%
4
4
5
0
3
2
0
29
%
T
ab
le
9
.
FP
S
p
er
f
o
r
m
an
ce
e
v
a
lu
atio
n
tab
le
C
a
se
S
i
n
g
l
e
f
r
a
me
m
o
d
e
(
FPS
)
B
a
t
c
h
m
o
d
e
(
FPS
)
1
2
.
7
8
.
9
2
2
.
5
8
.
3
3
3
.
2
10
.
0
4
2
.
2
7
.
8
Fig
u
r
es
7
to
1
0
p
r
esen
t
th
e
c
o
m
p
ar
is
o
n
o
f
th
e
r
ea
l
-
tim
e
p
e
r
f
o
r
m
a
n
ce
o
f
t
r
ad
itio
n
al
alg
o
r
ith
m
s
,
th
e
r
ea
l
-
tim
e
p
er
f
o
r
m
an
ce
o
f
th
e
h
y
b
r
id
s
tr
ateg
y
,
th
e
b
atch
p
er
f
o
r
m
an
ce
o
f
t
h
e
h
y
b
r
id
s
tr
ateg
y
,
an
d
th
e
atten
u
atio
n
r
ate
o
f
3
0
-
m
i
n
u
te
r
u
n
n
in
g
p
e
r
f
o
r
m
an
ce
,
r
esp
ec
tiv
ely
,
p
r
o
v
id
i
n
g
c
o
m
p
r
e
h
en
s
iv
e
d
ata
s
u
p
p
o
r
t
f
o
r
th
e
p
er
f
o
r
m
an
ce
ev
alu
atio
n
o
f
th
e
h
y
b
r
id
s
tr
ateg
y
.
Acc
o
r
d
i
n
g
to
th
e
d
ata,
b
ased
o
n
th
e
3
0
-
m
in
u
te
co
n
tin
u
o
u
s
r
u
n
n
in
g
test
d
ata,
th
e
h
y
b
r
id
s
tr
ateg
y
ex
h
i
b
its
ex
ce
llen
t
s
ta
b
ilit
y
in
r
ea
l
-
tim
e
p
e
r
f
o
r
m
an
c
e.
T
h
e
co
m
b
i
n
atio
n
o
f
L
B
PH
an
d
Mo
b
ileNet
ac
h
i
ev
es
th
e
b
est
r
ea
l
-
tim
e
p
e
r
f
o
r
m
an
ce
,
with
t
h
e
s
m
allest
s
tan
d
ar
d
d
ev
iatio
n
(
0
.
1
7
)
an
d
a
p
er
f
o
r
m
an
ce
atten
u
atio
n
r
ate
o
f
1
8
.
8
%,
wh
ich
v
er
if
ies
th
e
s
tab
ilit
y
an
d
r
eliab
ilit
y
o
f
t
h
e
h
y
b
r
id
s
tr
ateg
y
in
lo
n
g
-
te
r
m
o
p
er
atio
n
.
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
C
o
mp
a
r
a
tive
p
erfo
r
ma
n
ce
a
n
a
lysi
s
o
f lig
h
tw
eig
h
t fa
ce
id
en
t
ifica
tio
n
a
lg
o
r
ith
m
(
Wu
yu
n
Wa
n
g
)
2051
Fig
u
r
e
7
.
T
r
ad
itio
n
al
alg
o
r
ith
m
r
ea
l
-
tim
e
p
er
f
o
r
m
a
n
ce
Fig
u
r
e
8
.
Mix
e
d
s
tr
ateg
y
r
ea
l
-
tim
e
p
er
f
o
r
m
an
ce
Fig
u
r
e
9
.
L
i
n
e
ch
ar
t
o
f
b
atc
h
p
er
f
o
r
m
a
n
ce
o
f
m
ix
ed
s
tr
ateg
y
Evaluation Warning : The document was created with Spire.PDF for Python.