I
nte
rna
t
io
na
l J
o
urna
l o
f
E
lect
rica
l a
nd
Co
m
p
ute
r
E
ng
in
ee
ring
(
I
J
E
CE
)
Vo
l.
8
,
No
.
4
,
A
u
g
u
s
t
201
8
,
p
p
.
2
1
2
6
~
2
1
3
8
I
SS
N:
2
0
8
8
-
8708
,
DOI
: 1
0
.
1
1
5
9
1
/
i
j
ec
e
.
v8
i
4
.
p
p
2
1
2
6
-
2138
2126
J
o
ur
na
l ho
m
ep
a
g
e
:
h
ttp
:
//ia
e
s
co
r
e
.
co
m/
jo
u
r
n
a
ls
/in
d
ex
.
p
h
p
/
I
JE
C
E
Ag
e Inv
a
ria
nt
Fa
ce Re
co
g
nition
u
s
ing
C
o
nv
o
lutiona
l
N
eura
l
N
e
tw
o
rk
M
rudu
la
Ni
m
ba
rt
e
,
K
is
ho
r
B
ho
y
a
r
IT
De
p
a
rt
m
e
n
t,
YCCE Na
g
p
u
r,
M
H,
In
d
ia
Art
icle
I
nfo
AB
ST
RAC
T
A
r
ticle
his
to
r
y:
R
ec
eiv
ed
Oct
2
9
,
2
0
1
7
R
ev
i
s
ed
Dec
23
,
2
0
1
7
A
cc
ep
ted
A
p
r
7
,
2
0
1
8
In
th
e
re
c
e
n
t
y
e
a
rs,
f
a
c
e
re
c
o
g
n
it
io
n
a
c
ro
ss
a
g
in
g
h
a
s
b
e
c
o
m
e
v
e
r
y
p
o
p
u
lar
a
n
d
c
h
a
ll
e
n
g
in
g
tas
k
in
th
e
a
re
a
o
f
fa
c
e
r
e
c
o
g
n
it
io
n
.
M
a
n
y
re
s
e
a
r
c
h
e
rs
h
a
v
e
c
o
n
tri
b
u
ted
i
n
th
is
a
re
a
,
b
u
t
stil
l
t
h
e
re
is
a
si
g
n
if
ica
n
t
g
a
p
to
f
il
l
in
.
S
e
lec
ti
o
n
o
f
f
e
a
tu
re
e
x
trac
ti
o
n
a
n
d
c
l
a
ss
if
ic
a
ti
o
n
a
lg
o
rit
h
m
s
p
la
y
s
a
n
im
p
o
rtan
t
ro
le
in
th
is
a
re
a
.
De
e
p
Lea
rn
in
g
w
it
h
Co
n
v
o
lu
ti
o
n
a
l
Ne
u
ra
l
Ne
tw
o
rk
s
p
ro
v
id
e
s
u
s
a
c
o
m
b
in
a
ti
o
n
o
f
f
e
a
tu
re
e
x
trac
ti
o
n
a
n
d
c
las
sif
ica
ti
o
n
in
a
sin
g
le
st
ru
c
tu
re
.
I
n
th
is
p
a
p
e
r,
w
e
h
a
v
e
p
re
se
n
ted
a
n
o
v
e
l
id
e
a
o
f
7
-
L
a
y
e
r
CNN
a
rc
h
it
e
c
tu
re
f
o
r
so
lv
in
g
th
e
p
ro
b
lem
o
f
a
g
in
g
f
o
r
re
c
o
g
n
izin
g
f
a
c
ial
i
m
a
g
e
s
a
c
ro
ss
a
g
in
g
.
We
h
a
v
e
d
o
n
e
e
x
ten
siv
e
e
x
p
e
ri
m
e
n
tatio
n
s
to
tes
t
th
e
p
e
rf
o
r
m
a
n
c
e
o
f
th
e
p
ro
p
o
se
d
sy
ste
m
u
sin
g
tw
o
sta
n
d
a
rd
d
a
tas
e
ts
F
G
NET
a
n
d
M
ORPH
(A
lb
u
m
II).
Ra
n
k
-
1
re
c
o
g
n
it
io
n
a
c
c
u
ra
c
y
o
f
o
u
r
p
ro
p
o
se
d
s
y
ste
m
is
7
6
.
6
%
o
n
F
G
NE
T
a
n
d
9
2
.
5
%
o
n
M
OR
P
H
(A
lb
u
m
II).
Ex
p
e
ri
m
e
n
tal
re
su
lt
s
sh
o
w
th
e
sig
n
if
ica
n
t
im
p
ro
v
e
m
e
n
t
o
v
e
r
a
v
a
il
a
b
le
sta
te
-
of
-
th
e
-
a
rts
w
it
h
th
e
p
ro
p
o
se
d
CNN
a
rc
h
it
e
c
tu
re
a
n
d
t
h
e
c
las
sif
i
e
r.
K
ey
w
o
r
d
:
Ag
i
n
g
m
o
d
el
A
I
F
R
C
NN
Dee
p
l
ea
r
n
in
g
Face
r
ec
o
g
n
itio
n
Co
p
y
rig
h
t
©
2
0
1
8
In
stit
u
te o
f
A
d
v
a
n
c
e
d
E
n
g
i
n
e
e
rin
g
a
n
d
S
c
ien
c
e
.
Al
l
rig
h
ts
re
se
rv
e
d
.
C
o
r
r
e
s
p
o
nd
ing
A
uth
o
r
:
Mr
u
d
u
la
Ni
m
b
ar
te
,
Dep
ar
t
m
en
t o
f
I
n
f
o
r
m
atio
n
T
e
ch
n
o
lo
g
y
,
Yesh
w
a
n
tr
ao
C
h
av
a
n
C
o
lleg
e
o
f
E
n
g
in
ee
r
i
n
g
,
Nag
p
u
r
,
M
H,
I
n
d
ia
.
E
m
ail:
m
r
u
d
u
la_
n
i
m
b
ar
te@
r
e
d
if
f
m
ail.
co
m
1.
I
NT
RO
D
UCT
I
O
N
R
ec
o
g
n
izi
n
g
a
n
id
en
tit
y
o
f
a
p
er
s
o
n
f
r
o
m
f
ac
ia
l
i
m
a
g
es
is
a
ch
allen
g
i
n
g
a
n
d
in
ter
esti
n
g
p
r
o
b
lem
i
n
m
an
y
r
ea
l
-
w
o
r
ld
ap
p
licatio
n
s
.
Var
iatio
n
s
in
p
ar
a
m
eter
s
li
k
e
h
ea
d
p
o
s
e,
ex
p
r
ess
io
n
s
,
b
r
ig
h
t
n
e
s
s
an
d
ag
i
n
g
m
ak
e
s
it
d
if
f
icu
lt
to
r
ec
o
g
n
iz
e
a
p
er
s
o
n
f
r
o
m
th
e
d
atab
ases
[1
]
-
[
3
]
.
Ma
n
y
r
esear
ch
er
s
ar
e
w
o
r
k
i
n
g
o
n
t
h
ese
p
ar
am
eter
s
to
im
p
r
o
v
e
th
e
p
er
f
o
r
m
a
n
ce
o
f
th
e
s
y
s
te
m
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
.
Ag
in
g
in
its
elf
,
is
v
er
y
co
m
p
lica
ted
p
r
o
ce
s
s
.
I
t
v
ar
ie
s
f
r
o
m
p
er
s
o
n
to
p
er
s
o
n
[
4
]
.
Ag
ai
n
,
it
is
h
i
g
h
l
y
d
ep
en
d
en
t
o
n
f
ac
to
r
s
li
k
e
g
eo
g
r
ap
h
ical
lo
ca
tio
n
,
li
v
i
n
g
s
t
y
le,
ea
tin
g
h
ab
it
s
,
u
s
e
o
f
co
s
m
etic
s
,
p
h
y
s
ical
an
d
m
e
n
tal
h
ea
lth
etc.
o
f
a
p
er
s
o
n
[
5
]
-
[
7
]
.
So
,
c
o
n
s
id
er
in
g
all
t
h
ese
f
ac
to
r
s
,
r
ec
o
g
n
iz
in
g
an
id
e
n
tit
y
o
f
a
p
er
s
o
n
o
v
er
a
s
p
an
o
f
y
ea
r
s
i
s
r
ea
ll
y
d
if
f
ic
u
lt.
I
t
h
as
m
an
y
r
ea
l
-
w
o
r
ld
ap
p
licatio
n
s
lik
e
f
i
n
d
in
g
m
is
s
in
g
ch
i
ld
r
en
,
p
ass
p
o
r
t
r
en
e
w
al
s
y
s
te
m
,
d
r
iv
in
g
lice
n
s
e
r
en
e
w
a
l s
y
s
te
m
,
f
i
n
d
i
n
g
cr
i
m
i
n
als,
p
r
o
v
id
in
g
s
ec
u
r
ities
to
VI
P
'
s
etc.
[
8
]
-
[
1
0
].
On
e
m
o
r
e
i
m
p
o
r
tan
t
i
s
s
u
e
in
t
h
is
ap
p
licatio
n
is
ab
o
u
t
t
h
e
ch
an
g
e
s
i
n
t
h
e
i
m
ag
e
s
o
f
th
e
s
a
m
e
p
er
s
o
n
o
v
er
ag
in
g
i.e
.
in
tr
a
-
s
u
b
j
ec
t
v
ar
iatio
n
s
a
n
d
s
i
m
ilar
it
y
i
n
th
e
i
m
a
g
es
o
f
o
t
h
er
p
er
s
o
n
s
i.e
.
in
ter
-
s
u
b
j
ec
t
s
i
m
ilar
ities
.
T
h
is
ap
p
licatio
n
b
asicall
y
co
n
s
is
t
s
o
f
t
w
o
ca
te
g
o
r
ies:
Fi
r
s
t
is
Ag
e
I
n
v
ar
ian
t
Face
R
ec
o
g
n
it
io
n
(
A
I
F
R
)
an
d
s
ec
o
n
d
i
s
Ag
e
I
n
v
ar
ian
t
Face
Ver
i
f
icatio
n
(
A
I
FV)
.
Face
r
ec
o
g
n
it
io
n
i
s
g
e
n
e
r
all
y
co
n
s
id
er
ed
as
m
u
lti
-
clas
s
p
r
o
b
lem
a
n
d
f
ac
e
v
er
if
icat
io
n
is
b
i
n
ar
y
-
clas
s
p
r
o
b
lem
o
f
cla
s
s
i
f
icat
io
n
.
A
I
F
R
f
o
c
u
s
e
s
o
n
r
ec
o
g
n
izi
n
g
a
s
i
n
g
le
i
n
p
u
t
i
m
ag
e
o
f
a
p
er
s
o
n
w
i
th
av
a
ilab
l
e
i
m
a
g
e
s
o
f
t
h
e
s
a
m
e
p
er
s
o
n
i
n
g
aller
y
.
W
h
er
ea
s
,
A
I
FV tr
y
to
ch
ec
k
th
e
id
e
n
tit
y
o
f
t
w
o
g
iv
e
n
i
m
a
g
es,
w
h
eth
er
th
e
y
ar
e
s
a
m
e
o
r
d
if
f
er
e
n
t.
T
h
o
u
g
h
v
ar
io
u
s
m
eth
o
d
o
lo
g
i
es
h
a
v
e
b
ee
n
p
r
o
p
o
s
ed
b
y
r
esear
ch
er
s
to
s
o
l
v
e
t
h
is
p
r
o
b
le
m
an
d
i
m
p
r
o
v
e
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
,
th
er
e
is
s
til
l
a
s
co
p
e
to
i
m
p
r
o
v
e
ac
cu
r
ac
y
.
I
n
t
h
i
s
p
ap
er
,
w
e
h
a
v
e
f
o
cu
s
ed
o
n
Ag
e
I
n
v
ar
ia
n
t
Face
R
ec
o
g
n
i
tio
n
p
r
o
b
lem
b
y
u
s
i
n
g
C
o
n
v
o
lu
tio
n
al
Ne
u
r
al
Net
w
o
r
k
(
C
NN)
,
n
a
m
ed
as
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:
2
0
8
8
-
8708
A
g
e
I
n
v
a
r
ia
n
t F
a
ce
R
ec
o
g
n
iti
o
n
u
s
in
g
C
o
n
v
o
lu
tio
n
a
l Neu
r
a
l Netw
o
r
k
(
M
r
u
d
u
la
N
imb
a
r
te)
2127
A
I
F
R
-
C
N
N.
Dee
p
lear
n
in
g
u
s
in
g
C
N
N
h
as
b
e
co
m
e
v
er
y
p
o
p
u
lar
n
o
w
ad
a
y
s
.
I
t
h
as
t
h
e
ad
v
an
ta
g
e
th
at,
it
p
r
o
v
id
es
f
ea
tu
r
e
e
x
tr
ac
tio
n
a
n
d
class
if
icatio
n
i
n
a
s
i
n
g
le
s
tr
u
ctu
r
e.
B
u
t,
f
o
r
d
ec
id
in
g
t
h
e
d
etails
o
f
th
e
C
NN
ar
ch
itect
u
r
e,
th
er
e
is
n
o
s
ta
n
d
ar
d
r
u
le
o
r
an
y
lo
g
ic.
R
esear
ch
er
s
p
r
o
p
o
s
ed
th
eir
m
et
h
o
d
o
lo
g
ies
u
s
i
n
g
C
NN
also
;
it
is
f
o
c
u
s
ed
o
n
l
y
o
n
t
h
ei
r
o
w
n
ar
c
h
itect
u
r
e
n
o
t
o
n
an
y
s
p
ec
if
ic
r
ea
s
o
n
b
eh
i
n
d
u
s
i
n
g
t
h
o
s
e
d
e
tails
.
T
h
es
e
d
etails
i
n
clu
d
e:
(
a
)
n
u
m
b
er
o
f
la
y
er
s
i
n
t
h
e
n
et
w
o
r
k
,
(
b
)
s
eq
u
en
ce
o
f
th
e
s
e
la
y
er
s
,
(
c
)
d
i
m
e
n
s
io
n
s
o
f
t
h
e
f
ilter
s
ap
p
lied
,
an
d
(
d
)
n
u
m
b
er
o
f
n
eu
r
o
n
s
u
s
ed
etc.
Hen
ce
,
w
e
al
s
o
p
r
o
p
o
s
ed
o
u
r
o
w
n
m
et
h
o
d
o
lo
g
y
to
d
esig
n
th
e
C
NN
ar
ch
i
tectu
r
e
f
o
r
A
I
F
R
.
T
h
e
m
a
in
co
n
tr
ib
u
tio
n
s
o
f
th
i
s
p
ap
er
ar
e:
(
a
)
n
o
v
el
7
-
la
y
er
C
NN
ar
ch
itectu
r
e
f
o
r
A
I
FR
,
a
n
d
(
b
)
th
e
u
s
e
o
f
s
m
aller
i
m
a
g
e
s
ize
o
f
3
2
х
3
2
p
ix
e
ls
to
r
ed
u
ce
ti
m
e
an
d
s
p
ac
e
co
m
p
le
x
it
y
.
T
h
e
r
est
o
f
t
h
e
p
ap
er
is
o
r
g
an
ized
as
f
o
llo
w
s
.
Seco
n
d
s
ec
tio
n
i
n
cl
u
d
es
t
h
e
r
elate
d
wo
r
k
d
o
n
e
in
t
h
is
ar
ea
.
T
h
e
n
e
x
t
s
ec
t
io
n
i.e
.
t
h
ir
d
g
iv
e
s
co
m
p
le
te
d
etails
o
f
t
h
e
p
r
o
p
o
s
ed
m
eth
o
d
o
lo
g
y
f
o
r
ag
e
i
n
v
ar
ia
n
t
f
ac
e
r
ec
o
g
n
i
tio
n
.
I
t
is
f
o
llo
w
ed
b
y
s
ec
tio
n
f
o
u
r
,
f
o
r
ex
p
er
i
m
e
n
tal
d
etails
u
s
i
n
g
b
o
th
s
ta
n
d
ar
d
d
atasets
F
GNE
T
an
d
MO
R
P
H(
A
lb
u
m
I
I
)
.
Fin
a
ll
y
,
f
i
f
th
s
ec
tio
n
p
r
ese
n
ts
co
n
c
lu
s
i
o
n
.
2.
RE
L
AT
E
D
WO
RK
T
h
is
s
ec
tio
n
p
r
esen
t
s
th
e
r
elat
ed
w
o
r
k
in
th
is
ar
ea
.
So
m
e
o
f
th
e
r
esear
ch
er
s
f
o
cu
s
ed
th
ei
r
w
o
r
k
o
n
f
ac
e
id
en
ti
f
icatio
n
o
r
r
ec
o
g
n
iti
o
n
p
r
o
b
lem
an
d
o
th
er
s
o
n
f
ac
e
v
er
if
icatio
n
p
r
o
b
lem
.
T
h
is
p
r
o
b
lem
is
b
asical
l
y
ca
teg
o
r
ized
in
t
w
o
t
y
p
es:
G
en
er
ativ
e
an
d
Dis
cr
i
m
i
n
ati
v
e
Me
th
o
d
s
.
Gen
er
ati
v
e
m
et
h
o
d
s
n
ee
d
to
d
ev
elo
p
s
y
n
t
h
etic
i
m
a
g
es
o
f
t
h
e
p
er
s
o
n
at
th
e
r
eq
u
ir
ed
ag
e
an
d
th
en
p
er
f
o
r
m
m
atc
h
i
n
g
o
f
t
h
o
s
e
i
m
a
g
es
w
it
h
g
i
v
e
n
i
m
a
g
e.
Di
s
cr
i
m
i
n
ati
v
e
m
et
h
o
d
s
n
ee
d
t
h
eir
o
w
n
w
a
y
f
o
r
f
e
atu
r
e
e
x
tr
ac
tio
n
a
n
d
cla
s
s
i
f
ica
tio
n
p
u
r
p
o
s
e
s
o
th
a
t
t
w
o
i
m
ag
e
s
o
f
s
a
m
e
p
er
s
o
n
ar
e
m
atch
ed
.
2
.
1
.
G
ener
a
t
iv
e
m
et
ho
ds
R
ec
en
t
l
y
,
th
e
m
e
th
o
d
i
n
[
1
1
]
p
r
esen
ted
h
ier
ar
ch
ical
m
o
d
el
b
ased
o
n
t
w
o
-
le
v
el
lear
n
i
n
g
w
it
h
n
e
w
f
ea
t
u
r
e
d
escr
ip
to
r
ca
lled
as
L
o
ca
l
P
atter
n
Selectio
n
(
L
P
S)
f
o
r
s
o
lv
in
g
t
h
e
p
r
o
b
le
m
o
f
a
g
i
n
g
f
ac
e
r
ec
o
g
n
i
tio
n
.
T
h
e
m
et
h
o
d
in
[
1
2
]
,
f
o
cu
s
ed
o
n
th
e
r
o
le
o
f
f
ac
ial
as
y
m
m
etr
y
i
n
r
ec
o
g
n
izi
n
g
ag
e
-
s
ep
ar
ated
f
ac
e
i
m
a
g
es
b
ased
o
n
m
atc
h
i
n
g
-
s
co
r
e
s
p
ac
e
(
MSS)
.
I
n
[
1
3
]
,
au
th
o
r
s
u
s
ed
m
in
i
m
al
s
e
t
o
f
g
eo
m
etr
ic
f
ea
t
u
r
es
f
o
r
ag
e
i
n
v
ar
ia
n
t
f
ac
e
r
ec
o
g
n
itio
n
.
I
t
w
as
b
ased
o
n
s
elec
ted
f
ea
t
u
r
e
p
o
in
t
s
a
n
d
p
er
f
o
r
m
an
ce
e
v
al
u
ated
o
n
FG
NE
T
d
ataset.
P
ar
k
et
a
l
.
[
14
]
p
r
o
p
o
s
ed
a
g
en
er
i
c
m
et
h
o
d
th
at
co
n
s
is
ts
o
f
a
3
-
D
a
g
in
g
m
o
d
el
to
i
m
p
r
o
v
e
t
h
e
f
ac
e
r
ec
o
g
n
it
io
n
p
er
f
o
r
m
a
n
ce
.
T
h
e
y
u
s
ed
p
o
s
e
co
r
r
ec
tio
n
s
tep
an
d
s
ep
ar
ate
m
o
d
elin
g
f
o
r
s
h
ap
e
an
d
tex
tu
r
e.
2
.
2
.
Dis
cr
i
m
i
na
t
iv
e
m
Go
n
g
et
a
l
.
[1
5
]
p
r
esen
ted
a
n
o
v
el
f
ea
t
u
r
e
d
escr
ip
to
r
n
a
m
ed
as
m
a
x
i
m
u
m
en
tr
o
p
y
f
ea
t
u
r
e
d
escr
ip
to
r
(
ME
FD)
to
r
ec
o
g
n
ize
a
g
e
i
n
v
ar
ia
n
t
f
ac
e
i
m
a
g
es.
I
t
is
a
d
is
cr
i
m
i
n
ati
v
e
f
ea
tu
r
e
d
esc
r
ip
to
r
.
T
o
im
p
r
o
v
e
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
a
n
e
w
f
e
atu
r
e
-
m
atc
h
i
n
g
f
r
a
m
e
w
o
r
k
is
also
p
r
esen
ted
as
I
d
en
tit
y
Fac
to
r
A
n
al
y
s
is
(
I
F
A
)
.
A
li
et
a
l
.
[1
6
]
f
o
cu
s
ed
o
n
a
co
m
b
i
n
atio
n
o
f
s
h
ap
e
an
d
tex
tu
r
e
f
ea
t
u
r
es
f
o
r
ag
e
-
i
n
v
ar
ia
n
t
f
ac
e
r
ec
o
g
n
itio
n
.
T
h
ey
ad
o
p
ted
p
h
ase
co
n
g
r
u
e
n
c
y
f
ea
tu
r
e
f
o
r
s
h
ap
e
an
d
L
B
P
v
ar
ian
ce
f
o
r
tex
t
u
r
e
f
ea
t
u
r
e.
B
o
u
ch
af
f
r
a
[
1
7
]
in
tr
o
d
u
ce
d
a
n
o
v
el
f
r
a
m
e
w
o
r
k
to
r
ed
u
ce
d
im
e
n
s
io
n
alit
y
an
d
ex
tr
ac
ti
n
g
to
p
o
lo
g
ical
f
ea
tu
r
e
s
s
u
c
h
as
s
h
ap
e
f
o
r
ag
e
in
v
ar
ia
n
t
f
ac
e
r
ec
o
g
n
iti
o
n
.
I
t
is
a
co
m
b
i
n
atio
n
o
f
Ker
n
elize
d
R
ad
ial
b
asis
f
u
n
ctio
n
(
K
R
B
F)
f
o
r
d
i
m
en
s
io
n
al
it
y
r
ed
u
c
tio
n
,
c
o
n
s
tr
u
ct
io
n
o
f
α
-
s
h
ap
e
f
o
r
f
ea
t
u
r
e
ex
tr
ac
tio
n
a
n
d
m
ix
t
u
r
e
m
u
l
ti
n
o
m
ial
d
is
tr
ib
u
tio
n
s
f
o
r
o
b
j
ec
t c
lass
if
i
ca
tio
n
.
T
an
d
o
n
et
a
l
.
[1
8
]
attem
p
ted
a
n
o
v
el
ap
p
r
o
ac
h
u
s
i
n
g
L
B
P
o
f
p
ar
ticu
lar
r
eg
io
n
as
R
O
I
f
o
r
ag
e
in
v
ar
ia
n
t
f
ac
e
r
ec
o
g
n
i
tio
n
.
C
h
i
-
s
q
u
ar
e
m
ea
s
u
r
e
is
u
s
ed
a
s
a
d
is
s
i
m
ilar
it
y
m
ea
s
u
r
e
to
ca
l
cu
late
t
h
e
d
i
s
tan
ce
b
et
w
e
e
n
t
w
o
h
i
s
to
g
r
a
m
s
.
Ya
d
av
et
a
l
.
[1
9
]
p
r
esen
ted
a
s
y
s
te
m
to
i
m
p
r
o
v
e
th
e
r
es
u
lts
o
f
f
ac
e
r
ec
o
g
n
itio
n
ac
r
o
s
s
ag
e
p
r
o
g
r
ess
io
n
b
y
u
s
in
g
b
ac
ter
ia
f
o
r
ag
i
n
g
f
u
s
io
n
al
g
o
r
ith
m
.
I
t
r
ed
u
ce
s
th
e
ag
in
g
ef
f
ec
t
s
b
y
a
co
m
b
i
n
atio
n
o
f
L
B
P
f
ea
t
u
r
es
o
f
g
lo
b
al
an
d
lo
ca
l
f
ac
ial
r
e
g
i
o
n
s
b
y
u
s
i
n
g
b
ac
ter
ia
f
o
r
ag
in
g
f
u
s
io
n
al
g
o
r
ith
m
.
Xiao
et
a
l
.
[2
0
]
p
r
esen
ted
a
n
o
v
el
m
et
h
o
d
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
u
s
i
n
g
a
co
m
b
in
a
tio
n
o
f
te
x
t
u
r
e
an
d
s
h
ap
e
d
escr
ip
to
r
s
,
ca
lled
as
B
iv
ie
w
f
ac
e
r
ec
o
g
n
itio
n
alg
o
r
it
h
m
.
F
o
r
tex
tu
r
e
f
ea
t
u
r
e
s
u
b
s
p
ac
e
le
ar
n
in
g
m
et
h
o
d
s
ar
e
u
s
ed
an
d
g
r
ap
h
i
s
co
n
s
tr
u
cted
f
o
r
s
h
ap
e
to
p
o
lo
g
y
f
o
r
f
ac
e
i
m
ag
e
s
.
L
i
et
a
l
.
[2
1
]
p
r
o
p
o
s
e
d
a
d
is
cr
im
i
n
ati
v
e
ap
p
r
o
ac
h
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
o
v
er
ag
i
n
g
.
I
n
t
h
i
s
m
o
d
el,
t
h
e
y
u
s
ed
Scale
-
I
n
v
ar
ia
n
t
f
ea
t
u
r
e
tr
an
s
f
o
r
m
(
SI
FT
)
an
d
Mu
lti
-
s
ca
le
L
o
ca
l
B
in
ar
y
P
at
ter
n
s
(
ML
B
P
)
as
f
ea
t
u
r
e
d
escr
ip
to
r
s
an
d
m
u
l
tip
le
L
D
A
-
b
ased
class
if
ier
to
g
en
er
ate
a
d
ec
is
io
n
v
ia
f
u
s
io
n
r
u
le.
L
i
n
g
et
a
l
.
[
22
]
p
r
o
p
o
s
ed
a
d
is
cr
im
i
n
ati
v
e
m
e
th
o
d
f
o
r
f
ac
e
v
er
if
icatio
n
ac
r
o
s
s
ag
e
p
r
o
g
r
ess
io
n
.
I
n
t
h
eir
s
t
u
d
y
,
t
h
e
y
u
s
ed
Gr
ad
ien
t
Or
ien
tatio
n
(
GO
)
an
d
G
r
ad
ien
t
Or
ien
tatio
n
P
y
r
a
m
id
(
GOP
)
as f
ea
tu
r
e
d
es
cr
ip
to
r
an
d
Su
p
p
o
r
t V
ec
to
r
Ma
ch
in
e
(
S
VM
)
as a
class
if
ier
.
2
.
3
.
Usi
ng
co
nv
o
lutio
na
l neura
l net
w
o
rk
s
(
CNN)
R
ec
en
t
l
y
C
NN
h
a
v
e
b
ec
o
m
e
a
v
er
y
p
o
p
u
lar
tech
n
iq
u
e
f
o
r
C
o
m
p
u
ter
Vis
io
n
ap
p
licatio
n
s
.
Ma
n
y
r
esear
ch
er
s
u
s
ed
C
N
N
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
ap
p
licatio
n
s
.
I
n
[
23
]
,
a
m
eth
o
d
is
p
r
o
p
o
s
ed
u
s
in
g
a
f
u
s
io
n
o
f
2
-
D
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
8
,
No
.
4
,
A
u
g
u
s
t 2
0
1
8
:
2
1
2
6
–
2
1
3
8
2128
f
ac
e
i
m
a
g
es
a
n
d
m
o
tio
n
h
i
s
t
o
r
y
i
m
a
g
e
(
MH
I
)
f
o
r
f
ac
e
r
e
co
g
n
itio
n
b
ased
o
n
7
-
la
y
er
d
ee
p
lear
n
in
g
n
e
u
r
al
n
et
w
o
r
k
.
I
n
[
24
]
,
au
th
o
r
s
p
r
esen
ted
t
h
e
n
o
v
el
u
s
e
o
f
d
e
e
p
lear
n
in
g
u
s
i
n
g
C
NN
f
o
r
au
to
m
a
tic
f
ea
t
u
r
e
ex
tr
ac
tio
n
f
o
r
r
o
u
s
t
f
ac
e
r
ec
o
g
n
itio
n
ac
r
o
s
s
ti
m
e
lap
s
e.
T
h
e
y
u
s
ed
VGG
-
v
er
y
-
d
ee
p
1
6
la
y
e
r
C
NN
ar
ch
itect
u
r
e
in
t
h
eir
ex
p
er
i
m
e
n
t
s
.
L
i
et
a
l
.
[
25
]
,
p
r
o
p
o
s
ed
a
n
e
w
d
ee
p
C
NN
m
o
d
el
f
o
r
ag
e
-
i
n
v
ar
ia
n
t
f
ac
e
v
er
if
icat
io
n
w
i
t
h
7
-
la
y
er
C
NN
ar
ch
itect
u
r
e.
P
ar
k
h
i
et
a
l
.
[
26
]
p
r
esen
ted
a
m
o
d
el
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
ei
th
er
f
r
o
m
a
s
i
n
g
le
i
m
a
g
e
o
r
f
r
o
m
a
s
er
ies
o
f
f
ac
es
tr
ac
ed
f
r
o
m
v
id
eo
.
I
t
w
a
s
1
1
-
la
y
er
ar
ch
i
tectu
r
e
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
.
Su
n
et
a
l
.
[
27
]
p
r
o
p
o
s
ed
t
w
o
v
er
y
d
ee
p
n
eu
r
a
l
n
et
w
o
r
k
ar
c
h
itect
u
r
es
f
o
r
f
ac
e
r
ec
o
g
n
it
io
n
n
a
m
ed
as
Dee
p
I
D3
n
et1
an
d
Dee
p
I
D3
n
et2
.
Hal
f
f
ea
t
u
r
es
f
r
o
m
Dee
p
I
d
3
n
et1
an
d
o
th
er
h
a
lf
f
r
o
m
n
et2
ar
e
co
n
ca
ten
ated
i
n
to
a
lo
n
g
f
ea
t
u
r
e
v
ec
to
r
in
th
is
m
et
h
o
d
.
Hu
et
a
l
.
[
28
]
p
r
o
p
o
s
ed
t
h
r
ee
C
NN
ar
ch
i
tectu
r
e
s
an
d
co
n
d
u
cted
ex
te
n
s
i
v
e
ev
alu
a
tio
n
o
f
C
NN
-
b
ased
f
ac
e
r
ec
o
g
n
itio
n
s
y
s
te
m
.
T
h
ese
ar
ch
itectu
r
es
ar
e:
s
m
al
l
(
C
NN
-
S),
m
ed
iu
m
(
C
NN
-
M)
an
d
lar
g
e
(
C
NN
-
L
)
.
T
h
e
y
u
s
ed
L
FW
d
ataset
f
o
r
ex
p
er
i
m
en
tat
io
n
.
Xin
h
u
a
et
a
l
.
[
29
]
f
o
cu
s
ed
o
n
f
ac
e
r
ec
o
g
n
itio
n
p
r
o
b
le
m
u
s
in
g
C
NN
o
n
L
FW
d
ataset.
T
h
ey
u
s
ed
So
b
el
o
p
e
r
ato
r
to
im
p
r
o
v
e
r
esu
lt
ac
cu
r
a
c
y
.
T
aig
m
an
et
a
l
.
[
30
]
p
r
o
p
o
s
ed
a
9
-
lay
er
d
ee
p
n
eu
r
al
n
e
t
w
o
r
k
f
o
r
f
ac
e
v
er
i
f
i
ca
tio
n
p
r
o
b
lem
w
h
er
e
t
h
e
y
u
s
ed
alig
n
m
en
t
s
tep
an
d
r
ep
r
ese
n
tatio
n
s
tep
to
ap
p
ly
a
p
iece
w
is
e
a
f
f
i
n
e
tr
a
n
s
f
o
r
m
atio
n
.
Yi
et
al
.
[
31
]
d
ev
elo
p
ed
an
e
f
f
ec
t
iv
e
r
ep
r
esen
tati
o
n
s
f
o
r
b
o
th
f
ac
e
id
en
ti
f
icatio
n
an
d
v
er
i
f
icatio
n
w
it
h
d
ee
p
lear
n
i
n
g
n
a
m
ed
a
s
Dee
p
I
D2
.
Ma
n
y
r
esear
ch
er
s
p
r
esen
ted
t
h
eir
w
o
r
k
o
n
A
I
F
R
u
s
in
g
v
ar
io
u
s
m
et
h
o
d
s
as
d
is
cu
s
s
ed
ab
o
v
e,
b
u
t
o
n
l
y
f
e
w
s
tu
d
ie
s
r
ep
o
r
ted
f
o
r
A
g
e
I
n
v
ar
ia
n
t
Face
R
ec
o
g
n
itio
n
s
p
ec
iall
y
u
s
in
g
C
o
n
v
o
l
u
tio
n
al
Ne
u
r
al
Net
w
o
r
k
.
3.
P
RO
P
O
SE
D
M
E
T
H
O
DO
L
O
G
Y
F
O
R
AG
E
I
N
VARI
A
NT
F
ACE R
E
CO
G
N
I
T
I
O
N
T
h
is
s
ec
tio
n
d
escr
ib
es
th
e
p
r
o
p
o
s
ed
m
et
h
o
d
o
lo
g
y
f
o
r
ag
e
in
v
ar
ian
t
f
ac
e
r
ec
o
g
n
iti
o
n
u
s
i
n
g
C
o
n
v
o
lu
tio
n
al
Ne
u
r
al
Net
w
o
r
k
s
(
A
I
F
R
-
C
NN)
.
T
h
is
n
et
w
o
r
k
i
s
d
esi
g
n
ed
f
o
r
t
h
e
r
ec
o
g
n
i
tio
n
o
f
th
e
p
er
s
o
n
h
av
i
n
g
s
o
m
e
a
g
i
n
g
v
ar
iatio
n
s
.
T
h
e
o
v
er
all
p
r
o
ce
s
s
co
n
tain
s
th
e
s
a
m
e
tr
ad
itio
n
al
s
tep
s
:
I
m
ag
e
P
r
ep
r
o
ce
s
s
in
g
,
Featu
r
e
E
x
tr
ac
tio
n
a
n
d
C
lass
if
icatio
n
.
I
m
a
g
e
p
r
ep
r
o
ce
s
s
in
g
s
tep
s
i
m
p
r
o
v
e
t
h
e
p
er
f
o
r
m
an
ce
o
f
th
e
s
y
s
te
m
.
W
e
u
s
ed
th
r
ee
b
asic
p
r
ep
r
o
ce
s
s
i
n
g
s
tep
s
.
Feat
u
r
e
ex
tr
ac
tio
n
is
th
e
p
r
o
ce
s
s
o
f
ca
p
t
u
r
in
g
t
h
e
d
esire
d
f
ea
t
u
r
e
d
escr
ip
to
r
s
u
s
i
n
g
C
NN
r
ath
er
th
an
e
x
tr
ac
tin
g
it
m
a
n
u
a
ll
y
.
I
n
t
h
i
s
m
o
d
el,
w
e
u
s
ed
7
-
la
y
er
C
NN
ar
ch
i
tectu
r
e
.
C
las
s
i
f
icatio
n
i
s
r
eq
u
ir
ed
to
r
ec
o
g
n
ize
t
h
e
id
e
n
tit
y
o
f
t
h
e
p
er
s
o
n
.
T
h
is
w
o
r
k
i
n
cl
u
d
es
m
u
lti
-
clas
s
clas
s
i
f
icatio
n
p
r
o
b
lem
.
T
h
e
o
v
er
all
p
r
o
ce
s
s
f
o
r
A
I
FR
-
C
NN
is
s
h
o
w
n
in
Fig
u
r
e
1.
Fig
u
r
e
1
.
B
asic B
lo
ck
Diag
r
am
f
o
r
P
r
o
p
o
s
ed
A
I
FR
-
C
NN
3
.
1
.
I
m
a
g
e
p
re
pro
ce
s
s
i
ng
Stan
d
ar
d
d
ataset
s
m
a
y
h
a
v
e
i
m
ag
e
s
o
f
d
i
f
f
er
e
n
t
s
izes
an
d
i
llu
m
i
n
atio
n
.
Hen
ce
,
it
m
a
y
le
ad
to
s
o
m
e
r
ec
o
g
n
itio
n
p
r
o
b
lem
s
.
I
m
a
g
e
p
r
ep
r
o
ce
s
s
in
g
h
elp
s
to
k
ee
p
o
u
r
d
ataset
in
n
o
r
m
a
lized
f
o
r
m
at.
I
t
in
clu
d
e
s
d
etec
tio
n
an
d
cr
o
p
p
in
g
o
f
f
ac
i
al
p
o
r
tio
n
f
r
o
m
t
h
e
g
i
v
en
i
m
a
g
e.
Fo
r
th
i
s
p
u
r
p
o
s
e,
w
e
u
s
ed
p
o
p
u
lar
Vio
la
J
o
n
es
alg
o
r
ith
m
f
o
r
f
ac
e
d
etec
tio
n
.
T
h
e
n
ex
t
s
tep
i
s
to
co
n
v
er
t
t
h
e
R
GB
i
m
a
g
e
t
o
g
r
a
y
s
ca
le
i
m
ag
e.
L
ater
,
i
m
ag
e
s
ar
e
r
esized
to
3
2
х
3
2
p
ix
els
an
d
6
4
х
6
4
p
ix
els
f
o
r
s
o
m
e
ex
p
er
i
m
e
n
ts
.
I
n
t
h
i
s
w
o
r
k
,
w
e
h
a
v
e
n
o
t
p
er
f
o
r
m
ed
an
y
Fe
a
t
u
r
e
E
x
t
r
a
c
t
i
o
n
7
-
L
a
y
e
r
C
N
N
A
r
c
h
i
t
e
c
t
u
r
e
(
3
C
+2
S
+
2
F)
C
l
a
ss
i
f
i
c
a
t
i
o
n
S
V
M
C
l
a
ss
i
f
i
e
r
Im
a
g
e
Pr
e
p
r
o
c
e
ssi
n
g
Fa
c
e
D
e
t
e
c
t
i
o
n
&
C
r
o
p
p
i
n
g
R
GB
Gr
a
y
Fa
c
e
R
e
s
i
z
e
32
х
3
2
I/
p
i
m
a
g
e
o
f
a
p
e
r
so
n
a
t
a
g
e
2
8
R
e
c
o
g
n
i
z
e
d
Im
a
g
e
o
f
t
h
e
sa
m
e
p
e
r
so
n
a
t
a
g
e
6
1
T
r
a
i
n
i
n
g
D
a
t
a
set
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:
2
0
8
8
-
8708
A
g
e
I
n
v
a
r
ia
n
t F
a
ce
R
ec
o
g
n
iti
o
n
u
s
in
g
C
o
n
v
o
lu
tio
n
a
l Neu
r
a
l Netw
o
r
k
(
M
r
u
d
u
la
N
imb
a
r
te)
2129
co
m
p
lica
ted
p
r
ep
r
o
ce
s
s
in
g
s
t
ep
s
as
h
is
to
g
r
a
m
n
o
r
m
aliza
ti
o
n
an
d
h
ea
d
-
p
o
s
e
co
r
r
ec
tio
n
.
Fig
u
r
e
2
illu
s
tr
ate
s
p
r
ep
r
o
ce
s
s
in
g
s
tep
s
o
n
a
s
a
m
p
le
i
m
a
g
e
f
r
o
m
FGNE
T
d
ataset.
I
n
p
u
t
I
mag
e
4
0
4
х
5
0
4
C
r
o
p
p
e
d
F
a
c
e
3
0
7
х
307
G
r
a
y
C
o
l
o
r
F
a
c
e
3
0
7
х
307
R
e
si
z
e
d
F
a
c
e
32
х
32
Fig
u
r
e
2
.
Step
w
i
s
e
P
r
ep
r
o
ce
s
s
in
g
f
o
r
a
Sa
m
p
le
I
m
ag
e
f
r
o
m
FGNE
T
Data
s
et
3
.
2
.
Net
w
o
rk
a
rc
hite
ct
ure
f
o
r
f
e
a
t
u
re
ex
t
ra
ct
io
n
Af
ter
b
asic
p
r
ep
r
o
ce
s
s
i
n
g
s
te
p
s
,
th
e
n
ex
t
s
tep
i
s
to
e
x
tr
ac
t
f
ea
t
u
r
es
as
p
er
o
u
r
r
eq
u
ir
e
m
e
n
ts
.
I
n
o
u
r
p
r
o
p
o
s
ed
w
o
r
k
(
A
I
FR
-
C
NN)
,
f
o
r
th
i
s
p
u
r
p
o
s
e
w
e
u
s
ed
d
ee
p
lear
n
in
g
ap
p
r
o
ac
h
u
s
i
n
g
C
o
n
v
o
l
u
tio
n
a
l
Neu
r
al
Net
w
o
r
k
(
C
N
N)
.
I
t
h
a
s
m
a
n
y
ad
v
an
ta
g
e
s
.
Firs
t,
f
ea
tu
r
e
e
x
t
r
ac
tio
n
an
d
cla
s
s
i
f
ica
tio
n
ar
e
co
n
ce
r
n
s
o
f
C
N
N
its
el
f
w
it
h
s
i
n
g
le
s
tr
u
ct
u
r
e.
Se
co
n
d
,
th
is
n
et
w
o
r
k
ex
tr
ac
ts
d
ee
p
er
2
-
D
f
ea
tu
r
es.
T
h
ir
d
,
it
is
f
u
ll
y
ad
ap
tiv
e
a
n
d
in
v
ar
ia
n
t to
lo
ca
l a
n
d
g
eo
m
etr
i
c
ch
an
g
es i
n
t
h
e
i
m
a
g
e.
T
h
r
ee
t
y
p
es
o
f
m
ai
n
la
y
er
s
ar
e
th
er
e
i
n
a
C
NN:
(
a
)
C
o
n
v
o
lu
tio
n
la
y
er
,
(
b)
P
o
o
lin
g
l
a
y
er
(
S
u
b
-
s
a
m
p
li
n
g
)
,
an
d
(
c
)
Ou
tp
u
t
la
y
er
.
Feed
-
f
o
r
w
ar
d
s
tr
u
c
tu
r
e
is
u
s
ed
to
ar
r
an
g
e
th
e
s
e
la
y
er
s
i
n
th
e
n
e
t
w
o
r
k
.
E
ac
h
co
n
v
o
lu
tio
n
la
y
er
is
f
o
llo
w
ed
b
y
a
p
o
o
lin
g
la
y
er
,
w
h
er
ea
s
la
s
t
co
n
v
o
l
u
tio
n
l
a
y
er
is
f
o
llo
wed
b
y
o
u
tp
u
t
la
y
er
.
C
o
n
v
o
lu
tio
n
a
n
d
p
o
o
lin
g
la
y
e
r
s
ar
e
2
-
D
la
y
er
s
w
h
er
ea
s
o
u
t
p
u
t
la
y
er
is
1
-
D
la
y
er
.
E
v
er
y
2
-
D
la
y
er
o
f
a
C
NN
co
n
tain
s
s
ev
er
al
p
lan
e
s
.
A
p
lan
e
o
f
a
2
-
D
la
y
er
co
n
s
is
t
s
o
f
2
-
D
ar
r
a
y
o
f
n
eu
r
o
n
s
.
Featu
r
e
m
ap
i
s
t
h
e
o
u
tp
u
t
o
f
a
p
la
n
e.
I
n
A
I
F
R
-
C
NN,
w
e
p
r
o
p
o
s
ed
a
7
-
la
y
er
ar
ch
itec
tu
r
e
th
at
i
n
cl
u
d
es
3
co
n
v
o
l
u
tio
n
l
a
y
er
s
(
C
1
,
C
3
,
C
5
)
,
2
p
o
o
lin
g
la
y
er
s
(
P
2
,
P4
)
an
d
2
f
u
ll
y
co
n
n
ec
ted
o
u
tp
u
t
la
y
er
s
(
F6
,
F7
)
.
T
h
e
ar
ch
itec
tu
r
e
f
o
r
p
r
o
p
o
s
ed
m
et
h
o
d
o
lo
g
y
is
s
h
o
w
n
in
Fig
u
r
e
3.
Fig
u
r
e
3
.
7
-
L
a
y
e
r
C
NN
A
r
c
h
it
ec
tu
r
e
f
o
r
A
I
FR
-
C
NN
In
p
u
t
I
m
a
g
e
L
a
y
e
r
1
-
C
1
:
C
o
n
v
o
l
u
t
i
o
n
L
a
y
e
r
2
-
P2
:
Po
o
l
i
n
g
L
a
y
e
r
3
-
C
3
:
C
o
n
v
o
l
u
t
i
o
n
L
a
y
e
r
4
-
P4
:
Po
o
l
i
n
g
L
a
y
e
r
5
-
C
5
:
C
o
n
v
o
l
u
t
i
o
n
L
a
y
e
r
6
-
F6
:
Fu
l
l
y
C
o
n
n
e
c
t
e
d
L
a
y
e
r
7
-
F7
:
Ful
l
y
C
o
n
n
e
c
t
e
d
Ou
t
put
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
8
,
No
.
4
,
A
u
g
u
s
t 2
0
1
8
:
2
1
2
6
–
2
1
3
8
2130
3
.
2
.
1
.
Co
nv
o
lutio
n
l
a
y
er
E
ac
h
p
lan
e
o
f
co
n
v
o
lu
t
io
n
l
a
y
er
is
ass
o
ciate
d
w
it
h
o
n
e
o
r
m
o
r
e
f
ea
t
u
r
e
m
ap
s
o
f
ea
r
lier
la
y
er
.
C
o
n
v
o
lu
tio
n
m
a
s
k
is
u
s
ed
as a
n
ass
o
ciate
d
co
n
n
ec
tio
n
w
h
ic
h
is
a
2
-
D
w
ei
g
h
t
m
atr
i
x
o
f
ad
j
u
s
tab
le
e
n
tr
ie
s
.
T
h
e
co
n
v
o
lu
tio
n
is
co
m
p
u
ted
in
e
ac
h
p
lan
e
b
et
w
ee
n
its
2
-
D
in
p
u
ts
an
d
its
co
n
v
o
lu
t
io
n
m
a
s
k
.
T
h
e
o
u
tp
u
ts
o
f
co
n
v
o
lu
tio
n
la
y
er
s
ar
e
ad
d
ed
to
g
eth
er
w
it
h
a
n
ad
j
u
s
tab
le
s
ca
lar
,
ca
lled
as
b
ias.
L
astl
y
,
a
n
ac
tiv
atio
n
f
u
n
ctio
n
is
ap
p
lied
to
o
b
tain
t
h
e
p
lan
e
s
o
u
t
p
u
t.
T
h
is
o
u
tp
u
t
o
f
ea
c
h
p
l
an
e
is
k
n
o
w
n
as
f
ea
t
u
r
e
m
ap
.
A
co
n
v
o
l
u
tio
n
la
y
er
m
a
y
h
av
e
o
n
e
o
r
m
o
r
e
m
ap
s
.
E
ac
h
o
f
t
h
ese
f
ea
t
u
r
e
m
ap
s
is
co
n
n
ec
ted
to
e
x
ac
tl
y
o
n
e
p
la
n
e
o
f
n
e
x
t
la
y
er
i.e
.
s
u
b
-
s
a
m
p
lin
g
la
y
er
.
E
ac
h
p
lan
e
in
la
s
t
co
n
v
o
lu
tio
n
la
y
er
is
ass
o
ciate
d
w
i
t
h
f
ea
t
u
r
e
m
ap
o
f
ex
ac
tl
y
o
n
e
p
r
ec
ed
in
g
la
y
er
.
E
ac
h
p
la
n
e
i
n
th
e
co
n
v
o
l
u
tio
n
la
y
er
p
r
o
d
u
ce
s
o
n
e
s
ca
lar
o
u
tp
u
t;
th
e
s
e
o
u
tp
u
ts
f
r
o
m
all
p
lan
e
s
ar
e
g
iv
e
n
to
o
u
tp
u
t
la
y
er
.
T
h
e
p
u
r
p
o
s
e
o
f
th
is
la
y
er
is
to
ex
t
r
ac
t
lo
w
-
le
v
el
f
ea
t
u
r
es
s
u
c
h
as
ed
g
es
an
d
tex
t
u
r
e.
Fe
atu
r
e
m
ap
o
f
co
n
v
o
l
u
tio
n
l
a
y
er
is
ca
lcu
la
ted
as
:
(
∑
)
(
1
)
W
h
er
e
is
th
e
co
n
v
o
lu
t
io
n
m
as
k
,
is
th
e
b
ias ter
m
,
an
d
is
th
e
lis
t o
f
p
lan
e
s
[
32
].
3
.
2
.
2
.
P
o
o
lin
g
(
s
ub
-
s
a
m
pli
ng
)
la
y
er
T
h
e
d
im
e
n
s
io
n
a
lit
y
o
f
ea
ch
f
e
atu
r
e
m
ap
is
r
ed
u
ce
d
b
y
s
p
atia
l
p
o
o
lin
g
b
y
r
etai
n
i
n
g
t
h
e
m
o
s
t
v
alu
ab
l
e
in
f
o
r
m
atio
n
.
I
t
ca
n
b
e
o
f
t
h
r
ee
d
if
f
er
en
t
t
y
p
e
s
:
Ma
x
p
o
o
lin
g
-
ta
k
es
t
h
e
l
ar
g
est
ele
m
e
n
t,
Av
er
ag
e
p
o
o
lin
g
-
ta
k
es
t
h
e
av
er
ag
e
o
f
th
e
ele
m
e
n
t
s
,
an
d
Su
m
p
o
o
lin
g
-
tak
e
s
th
e
s
u
m
o
f
all
th
e
el
e
m
en
ts
.
T
h
e
m
a
in
f
u
n
ctio
n
o
f
p
o
o
lin
g
i
s
to
r
ed
u
ce
th
e
s
p
atial
s
ize
o
f
th
e
i
n
p
u
t
r
ep
r
esen
tatio
n
p
r
o
g
r
ess
iv
el
y
.
I
t
h
elp
s
to
m
a
k
e
th
e
in
p
u
t
r
ep
r
ese
n
tatio
n
s
s
m
aller
an
d
m
o
r
e
co
n
v
en
ien
t.
A
p
o
o
l
in
g
an
d
p
r
ec
ed
in
g
co
n
v
o
lu
tio
n
la
y
er
s
h
a
v
e
t
h
e
s
a
m
e
n
u
m
b
er
o
f
p
la
n
es.
T
h
is
r
esu
lt
is
th
e
n
p
as
s
ed
t
h
r
o
u
g
h
t
h
e
ac
tiv
at
io
n
f
u
n
ct
io
n
to
p
r
o
d
u
ce
th
e
o
u
tp
u
t
s
.
T
h
is
f
ea
t
u
r
e
m
ap
i
s
co
n
n
ec
ted
to
o
n
e
o
r
m
o
r
e
p
la
n
es
o
f
t
h
e
n
ex
t
co
n
v
o
lu
t
io
n
la
y
er
.
I
t
m
ak
es
th
e
o
u
tp
u
t
o
f
co
n
v
o
lu
tio
n
la
y
er
m
o
r
e
r
o
b
u
s
t
to
lo
ca
l d
is
to
r
tio
n
s
.
Featu
r
e
m
ap
o
f
s
u
b
-
s
a
m
p
l
in
g
la
y
er
is
ca
l
cu
lated
as
(
)
(
2
)
w
h
er
e
is
m
atr
ix
o
b
tain
ed
b
y
s
u
m
m
i
n
g
all
f
o
u
r
p
ix
els
o
f
a
b
lo
ck
,
is
th
e
w
eig
h
t
an
d
is
th
e
b
ias
ter
m
[
32
]
.
3
.
2
.
3
.
O
utput
la
y
er
(
f
ull
y
co
nn
ec
t
e
d la
y
er
)
I
n
A
I
F
R
-
C
NN,
th
e
o
u
tp
u
t
la
y
er
is
co
n
s
tr
u
c
ted
f
r
o
m
s
ig
m
o
i
d
al
n
e
u
r
o
n
.
Ge
n
er
all
y
,
th
e
o
u
t
p
u
ts
o
f
t
h
i
s
la
y
er
ar
e
t
h
e
o
u
tp
u
t
s
o
f
t
h
e
n
et
w
o
r
k
.
I
n
th
e
o
u
tp
u
t
la
y
er
,
s
o
f
t
m
a
x
ac
tiv
a
tio
n
f
u
n
ctio
n
is
u
s
ed
b
y
tr
ad
itio
n
a
l
m
u
lti
la
y
er
p
er
ce
p
tio
n
.
Ot
h
er
class
i
f
ier
s
li
k
e
SVM
ca
n
a
ls
o
b
e
u
s
ed
.
T
h
ese
f
u
ll
y
co
n
n
ec
te
d
la
y
er
s
ca
p
tu
r
e
th
e
co
r
r
elatio
n
s
b
et
w
ee
n
f
ea
t
u
r
es
o
f
v
ar
io
u
s
p
ar
ts
o
f
th
e
f
ac
e
lik
e
s
h
ap
e
an
d
lo
ca
tio
n
o
f
ey
e
s
an
d
m
o
u
t
h
.
T
h
e
C
o
n
v
o
lu
tio
n
a
n
d
P
o
o
lin
g
la
y
er
s
i
n
co
m
b
i
n
atio
n
ar
e
u
s
ed
f
o
r
f
ea
t
u
r
e
ex
tr
ac
tio
n
wh
ile
f
u
ll
y
co
n
n
ec
ted
la
y
e
r
s
ar
e
u
s
ed
f
o
r
class
i
f
icati
o
n
.
T
h
e
o
u
tp
u
t o
f
s
i
g
m
o
id
al
n
eu
r
o
n
is
ca
lcu
lated
as
(
∑
)
(
3
)
w
h
er
e
is
th
e
n
u
m
b
er
o
f
o
u
t
p
u
t
s
i
g
m
o
id
al
n
e
u
r
o
n
s
,
is
w
ei
g
h
t
f
r
o
m
f
ea
t
u
r
e
m
ap
m
o
f
th
e
last
co
n
v
o
lu
tio
n
la
y
er
to
n
eu
r
o
n
n
o
f
th
e
o
u
tp
u
t la
y
er
,
an
d
is
th
e
b
ias o
f
n
e
u
r
o
n
n
a
s
s
o
ciate
d
w
i
th
la
y
er
L
[
32
].
3
.
2
.
4
.
7
-
L
a
y
er
Arc
hite
ct
ure
f
o
r
AI
F
R
-
CNN
I
n
o
u
r
i
m
p
le
m
e
n
tatio
n
,
w
e
u
s
ed
7
-
la
y
er
C
NN
ar
c
h
itect
u
r
e
f
o
r
a
g
e
i
n
v
ar
ia
n
t
f
ac
e
r
ec
o
g
n
i
tio
n
a
s
s
h
o
w
n
in
F
ig
u
r
e
4
.
T
h
is
n
etw
o
r
k
ar
ch
itect
u
r
e
co
n
s
is
ts
o
f
s
eq
u
en
ce
o
f
co
n
v
o
l
u
tio
n
,
s
u
b
-
s
a
m
p
li
n
g
a
n
d
f
u
ll
y
co
n
n
ec
ted
o
u
tp
u
t
la
y
er
s
.
T
h
e
co
n
v
o
lu
tio
n
la
y
er
s
u
s
e
f
il
ter
s
o
f
5
х
5
w
h
er
ea
s
s
u
b
-
s
a
m
p
li
n
g
w
it
h
2
х
2
.
T
h
e
in
p
u
t
to
th
is
n
et
w
o
r
k
is
3
2
х
3
2
p
ix
e
l
s
g
r
a
y
s
ca
le
i
m
a
g
e.
T
h
is
i
m
a
g
e
u
s
ed
f
o
r
p
er
f
o
r
m
i
n
g
co
n
v
o
l
u
t
io
n
w
it
h
a
f
ilter
o
f
5
х
5
p
ix
els
i
n
s
ize.
C
o
n
v
o
l
u
ti
o
n
is
a
li
n
ea
r
o
p
er
atio
n
an
d
p
er
f
o
r
m
s
ele
m
e
n
t
w
is
e
m
atr
i
x
m
u
l
tip
licatio
n
a
n
d
ad
d
itio
n
.
T
h
e
f
ilter
ed
i
m
a
g
e
o
f
2
8
х
2
8
p
ix
els
is
o
b
tain
ed
b
y
co
n
v
o
lv
in
g
a
5
х
5
f
ilter
w
i
th
3
2
х
3
2
p
i
x
e
ls
i
m
a
g
e.
I
t
h
as
6
d
i
s
ti
n
ct
p
lan
e
s
.
T
h
ese
6
p
lan
es
w
i
ll
g
en
er
ate
6
s
ep
ar
ate
f
ea
t
u
r
e
m
ap
s
as
o
u
tp
u
t
o
f
f
ir
s
t
co
n
v
o
l
u
tio
n
la
y
er
C
1
as
2
8
х
2
8
х
6
m
atr
i
x
.
L
a
y
er
2
is
s
u
b
-
s
a
m
p
li
n
g
la
y
er
S2
.
T
h
e
p
o
o
lin
g
o
p
er
atio
n
w
e
u
s
ed
is
s
u
m
m
in
g
an
d
r
eg
io
n
s
ar
e
2
х
2
p
ix
el
s
.
I
t
r
esu
lt
s
i
n
r
ed
u
ce
d
f
ea
tu
r
e
m
ap
b
y
a
f
ac
to
r
o
f
2
i
n
b
o
th
d
i
m
en
s
io
n
s
a
n
d
w
e
o
b
tain
1
4
х
1
4
х
6
m
atr
i
x
.
Nex
t
la
y
er
is
a
n
o
th
er
co
n
v
o
l
u
tio
n
l
a
y
er
C
3
,
an
d
w
e
ap
p
lied
a
f
il
ter
o
f
s
a
m
e
s
ize
o
f
5
х
5
.
I
n
th
i
s
la
y
er
,
w
e
h
a
v
e
1
6
d
is
tin
ct
p
lan
e
s
o
f
1
0
х
1
0
.
L
a
y
er
4
is
ag
ai
n
s
u
b
-
s
a
m
p
l
in
g
la
y
er
S4
w
i
th
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:
2
0
8
8
-
8708
A
g
e
I
n
v
a
r
ia
n
t F
a
ce
R
ec
o
g
n
iti
o
n
u
s
in
g
C
o
n
v
o
lu
tio
n
a
l Neu
r
a
l Netw
o
r
k
(
M
r
u
d
u
la
N
imb
a
r
te)
2131
2
х
2
b
lo
ck
r
ed
u
ce
s
f
ea
t
u
r
e
m
a
p
b
y
f
ac
to
r
o
f
2
an
d
g
i
v
es
m
atr
ix
o
f
s
ize
5
х
5
х
1
6
.
L
ast
co
n
v
o
l
u
tio
n
la
y
er
C
5
p
r
o
d
u
ce
s
f
ea
tu
r
e
m
ap
o
f
1
2
0
s
ca
lar
v
alu
es.
Fi
n
all
y
,
th
e
n
ex
t
la
y
er
s
ar
e
t
w
o
f
u
ll
y
co
n
n
ec
ted
lay
er
s
(
F6
an
d
F7
)
w
h
er
e
ea
c
h
o
u
t
p
u
t
u
n
it
i
s
co
n
n
ec
ted
to
all
in
p
u
ts
.
F6
co
n
ta
i
n
s
8
4
a
n
d
F7
co
n
tai
n
s
1
0
n
eu
r
o
n
s
.
T
h
e
o
u
tp
u
t
o
f
last
f
u
ll
y
co
n
n
ec
ted
la
y
er
F7
i
s
p
r
o
v
id
ed
to
th
e
class
i
f
ier
.
T
ab
le
1
s
h
o
w
s
th
e
d
etail
s
o
f
t
h
e
C
NN
ar
ch
itect
u
r
e
u
s
ed
ea
r
lier
f
o
r
f
ac
e
r
ec
o
g
n
i
tio
n
p
r
o
b
le
m
.
So
m
e
o
f
th
e
m
ar
e
o
n
l
y
f
o
r
f
ac
e
r
ec
o
g
n
itio
n
an
d
v
er
i
f
icatio
n
w
h
ile
s
o
m
e
o
f
t
h
e
m
ar
e
f
o
r
a
g
e
in
v
ar
ian
t
f
ac
e
r
ec
o
g
n
itio
n
.
I
t
s
h
o
w
s
all
t
h
e
d
etails
as
s
ize
o
f
i
n
p
u
t
i
m
ag
e
tak
en
,
n
u
m
b
er
o
f
la
y
er
s
u
s
ed
in
th
e
ar
ch
itec
tu
r
e,
d
atasets
u
s
ed
a
n
d
le
n
g
t
h
o
f
f
e
atu
r
e
v
ec
to
r
.
Fr
o
m
th
is
tab
le
i
t
is
o
b
s
er
v
ed
t
h
at
s
o
m
e
w
o
r
k
h
as
b
ee
n
d
o
n
e
f
o
r
f
ac
e
r
ec
o
g
n
i
tio
n
,
b
u
t
f
o
r
ag
e
in
v
ar
ia
n
t
f
ac
e
r
ec
o
g
n
i
tio
n
th
er
e
is
s
ti
ll
a
lo
t
o
f
s
co
p
e
to
im
p
r
o
v
e
t
h
e
p
er
f
o
r
m
a
n
ce
.
Ag
ai
n
,
v
ar
iatio
n
s
ar
e
th
er
e
i
n
ea
ch
ar
c
h
itect
u
r
e
in
s
ize
o
f
i
m
ag
e
a
n
d
n
u
m
b
er
o
f
la
y
er
s
i
n
t
h
e
ar
ch
itect
u
r
e.
So
,
th
e
s
ize
o
f
f
e
atu
r
e
v
ec
to
r
v
ar
ies.
An
d
,
th
er
e
is
n
o
s
p
ec
i
f
ic
r
ea
s
o
n
m
e
n
tio
n
ed
b
eh
in
d
s
elec
ti
n
g
th
ese
p
ar
a
m
eter
s
.
Fig
u
r
e
4
.
7
-
L
a
y
er
A
r
ch
itect
u
r
e
f
o
r
A
I
F
R
u
s
in
g
C
NN
T
ab
le
1
.
C
o
m
p
ar
ativ
e
De
tails
o
f
State
-
of
-
th
e
-
A
r
t
s
f
o
r
Face
R
ec
o
g
n
itio
n
u
s
i
n
g
o
n
l
y
C
N
N
I
n
p
u
t
I
mag
e
1
A
r
c
h
i
t
e
c
t
u
r
e
2
D
a
t
a
se
t
3
F
e
a
t
u
r
e
V
e
c
t
o
r
F
a
c
e
R
e
c
o
g
n
i
t
i
o
n
/
V
e
r
i
f
i
c
a
t
i
o
n
u
si
n
g
C
N
N
D
e
e
p
F
a
c
e
[
30
]
1
5
2
x
1
5
2
x
3
7
L
a
y
e
r
(
2
C
+
1
M
+
2
L
+
1
F
)
S
F
C
,
L
F
W
,
Y
TF
4
0
9
6
D
e
e
p
F
a
c
e
R
e
c
o
g
n
i
t
i
o
n
[
26
]
2
2
4
x
2
2
4
x
1
1
4
L
a
y
e
r
(
8
C
+
3
M
+
3
F)
L
F
W
,
Y
TF
4
0
9
6
D
e
e
p
I
D
3
9
x
3
1
x
1
8
L
a
y
e
r
s
(
4
C
+
3
M
+
1
F
)
L
F
W
1
6
0
D
e
e
p
I
D
2
[
31
]
5
5
x
4
7
x
3
8
L
a
y
e
r
s
(
4
C
+
3
M
+
1
F
)
L
F
W
1
6
0
D
e
e
p
I
D
3
[
27
]
-
(
1
0
C
+
4
P
)
L
F
W
3
0
0
W
h
e
n
F
R
me
e
t
s w
i
t
h
D
e
e
p
L
e
a
r
n
i
n
g
[
28
]
5
8
x
5
8
x
1
C
N
N
-
S
(
3
C
+
1
F
)
C
N
N
-
M
(
3
C
+
1
F
)
C
N
N
-
L
(
4
C
+
1
F
)
L
F
W
1
6
0
F
R
b
a
se
d
o
n
D
e
e
p
N
e
u
r
a
l
N
e
t
w
o
r
k
[
29
]
6
0
x
4
8
x
1
6
L
a
y
e
r
(
2
C
+
2
P
+
2
F
)
L
F
W
-
F
u
si
o
n
F
R
b
a
se
d
d
e
e
p
l
e
a
r
n
i
n
g
[
23
]
1
0
0
x
1
0
0
x
1
7
L
a
y
e
r
P
r
i
v
a
t
e
2
0
0
0
0
A
g
e
-
I
n
v
a
r
i
a
n
t
F
a
c
e
R
e
c
o
g
n
i
t
i
o
n
/
V
e
r
i
f
i
c
a
t
i
o
n
u
si
n
g
C
N
N
F
R
a
c
r
o
ss t
i
me
l
a
p
se
u
s
i
n
g
C
N
N
[
24
]
2
2
4
х
2
2
4
х
1
1
6
L
a
y
e
r
(
1
6
C
+
5
P
+
3
F
)
F
G
N
E
T
,
M
O
R
P
H
4
0
9
6
D
e
e
p
Jo
i
n
t
L
e
a
r
n
i
n
g
a
p
p
r
o
a
c
h
f
o
r
A
I
F
V
[
25
]
1
8
0
х
1
3
0
х
3
7
L
a
y
e
r
(
2
C
+
2
M
+
3
F
)
M
O
R
P
H
4
0
0
P
r
o
p
o
se
d
A
I
F
R
-
C
N
N
3
2
х
3
2
х
1
7
L
a
y
e
r
(
3
C
+
2
M
+
2
F
)
F
G
N
E
T
,
M
O
R
P
H
9
6
1
1
I
n
p
u
t
i
m
a
g
e
i
s re
p
r
e
se
n
t
e
d
a
s W
i
d
t
h
х
H
e
i
g
h
t
х
C
h
a
n
n
e
l
s,
1
a
n
d
3
me
a
n
G
r
a
y
a
n
d
R
G
B
i
m
a
g
e
s re
sp
e
c
t
i
v
e
l
y
.
2
U
p
p
e
r
c
a
se
l
e
t
t
e
r
s
C
,
P
,
M
,
L
a
n
d
F
r
e
p
r
e
se
n
t
C
o
n
v
o
l
u
t
i
o
n
a
l
,
P
o
o
l
i
n
g
,
M
a
x
-
P
o
o
l
i
n
g
,
L
o
c
a
l
l
y
c
o
n
n
e
c
t
e
d
a
n
d
f
u
l
l
y
c
o
n
n
e
c
t
e
d
l
a
y
e
r
s
r
e
sp
e
c
t
i
v
e
l
y
.
3
L
F
W
-
L
a
b
e
l
e
d
F
a
c
e
s
i
n
t
h
e
W
i
l
d
d
a
t
a
se
t
,
Y
TF
-
Y
o
u
T
u
b
e
F
a
c
e
s,
S
F
C
-
S
o
c
i
a
l
F
a
c
e
C
l
a
ssi
f
i
c
a
t
i
o
n
,
F
G
N
E
T
-
F
a
c
e
a
n
d
G
e
st
u
r
e
f
e
a
t
u
r
e
r
e
c
o
g
n
i
t
i
o
n
N
ET
w
o
r
k
,
M
O
R
P
H
-
M
O
R
P
H
(
a
l
b
u
m I
I
)
.
3
.
3
.
Cla
s
s
if
ica
t
io
n
t
ec
hn
iq
ues
I
n
t
h
is
w
o
r
k
,
w
e
u
s
e
m
u
lt
i
-
cla
s
s
Su
p
p
o
r
t
Vec
to
r
Ma
c
h
i
n
e
(
S
VM
)
as
a
clas
s
i
f
ier
f
o
r
t
h
e
id
e
n
ti
f
icatio
n
o
f
a
p
er
s
o
n
o
v
er
lo
n
g
p
er
io
d
.
I
t
is
a
s
u
p
er
v
is
ed
lear
n
i
n
g
alg
o
r
it
h
m
as
d
ata
lab
els
ar
e
av
ailab
le.
T
h
e
y
ar
e
ef
f
ec
tiv
e
i
n
h
i
g
h
d
i
m
e
n
s
io
n
al
s
p
ac
es.
I
t
is
also
m
e
m
o
r
y
e
f
f
icie
n
t
an
d
v
er
s
a
tile
in
n
atu
r
e.
I
n
an
o
t
h
er
6
28
28
C
1
:
6
@
2
8
х
28
16
10
10
C
3
:
1
6
@
1
0
х
10
1
2
0
1
1
C
5
:
1
2
0
@
1
х1
6
14
14
S
2
:
6
@
1
4
х
14
F
6
:
8
4
16
5
5
S
4
:
1
6
@
5
х5
32
32
I
n
p
u
t
I
mag
e
32
х
32
5
5
K
e
r
n
e
l
F
7
:
1
0
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
8
,
No
.
4
,
A
u
g
u
s
t 2
0
1
8
:
2
1
2
6
–
2
1
3
8
2132
ex
p
er
i
m
e
n
t,
w
e
u
s
e
E
u
clid
ian
Dis
ta
n
ce
w
it
h
Nea
r
est
Neig
h
b
o
r
(
NN)
as
a
cla
s
s
i
f
icat
io
n
r
u
l
e
f
o
r
r
ec
o
g
n
izi
n
g
a
p
er
s
o
n
ac
r
o
s
s
ag
i
n
g
.
4.
E
XP
E
R
I
M
E
NT
A
L
DE
T
AI
L
S
I
n
th
is
s
ec
tio
n
,
w
e
d
escr
ib
e
th
e
i
m
p
le
m
e
n
tatio
n
a
n
d
ex
p
er
i
m
e
n
tal
d
etails
f
o
r
A
I
FR
-
C
N
N.
I
n
o
u
r
ex
p
er
i
m
e
n
ts
,
w
e
u
s
ed
L
OP
O(
L
ea
v
e
On
e
P
er
s
o
n
Ou
t)
s
ch
e
m
e
f
o
r
test
i
n
g
,
w
h
er
e
o
n
e
p
er
s
o
n
f
r
o
m
d
ata
s
et
is
k
ep
t o
u
t
f
o
r
test
i
n
g
.
I
n
o
u
r
ea
r
lier
ex
p
er
i
m
en
t
s
,
w
e
u
s
ed
3
-
f
o
ld
cr
o
s
s
v
alid
atio
n
t
h
at
n
ee
d
s
t
o
k
ee
p
3
i
m
a
g
es
o
f
th
e
s
a
m
e
p
er
s
o
n
at
d
i
f
f
er
en
t
a
g
es
i
n
th
r
ee
s
ep
ar
ate
f
o
ld
er
s
.
I
n
t
h
is
ap
p
r
o
ac
h
,
in
s
tead
o
f
k
e
ep
in
g
o
n
e
i
m
a
g
e
o
f
th
e
p
er
s
o
n
i
n
te
s
tin
g
f
o
ld
er
,
w
e
k
ep
t
t
w
o
i
m
ag
e
s
o
f
t
h
e
s
a
m
e
p
er
s
o
n
h
av
in
g
a
g
ap
o
f
at
least
1
0
y
ea
r
s
.
T
h
e
r
ea
s
o
n
b
eh
i
n
d
d
o
in
g
s
o
is
to
av
o
id
r
ep
etitiv
e
te
s
tin
g
u
s
i
n
g
d
if
f
er
e
n
t
f
o
ld
er
s
.
As
all
i
m
a
g
es
ar
e
i
n
a
s
i
n
g
le
test
i
n
g
f
o
ld
er
,
all
p
er
s
o
n
s
ar
e
co
n
s
id
er
ed
as
d
if
f
er
en
t
in
d
iv
id
u
al
s
.
R
e
m
ai
n
in
g
i
m
a
g
es
o
f
th
e
p
er
s
o
n
ar
e
co
n
s
id
er
ed
f
o
r
tr
ain
i
n
g
,
i
n
o
r
d
er
to
av
o
id
th
e
s
a
m
e
p
er
s
o
n
in
b
o
th
th
e
f
o
ld
er
s
.
W
e
u
s
e
R
a
n
k
-
1
r
ec
o
g
n
itio
n
r
ate
as
a
p
er
f
o
r
m
an
ce
e
v
alu
a
tio
n
p
ar
am
eter
.
T
h
e
ex
p
er
i
m
e
n
t
s
ar
e
p
er
f
o
r
m
ed
o
n
M
A
T
L
A
B
2
0
1
5
a
(
6
4
-
b
it)
w
i
t
h
2
.
6
0
GHz
I
n
tel(
R
)
C
OR
E
(
T
M)
i
-
5
C
P
U
a
n
d
8
GB
o
f
R
A
M.
4
.
1
.
Da
t
a
s
et
s
W
e
u
s
e
t
w
o
p
u
b
licl
y
a
v
ailab
l
e
d
atasets
FGNE
T
[
33
]
an
d
MO
R
P
H
(
A
lb
u
m
I
I
)
[3
4
]
f
o
r
A
I
F
R
-
C
N
N.
B
o
th
th
e
d
ataset
s
co
n
tain
m
a
n
y
i
m
ag
e
s
o
f
th
e
s
a
m
e
p
er
s
o
n
h
av
i
n
g
v
ar
iatio
n
i
n
a
g
e,
ex
p
r
e
s
s
io
n
s
,
ill
u
m
i
n
atio
n
an
d
h
ea
d
p
o
s
itio
n
.
FGNE
T
c
o
n
s
is
t
s
o
f
1
0
0
2
im
a
g
es
o
f
8
2
s
u
b
j
ec
ts
.
I
t
in
clu
d
es
1
2
i
m
a
g
es
o
f
a
p
er
s
o
n
i
n
av
er
ag
e.
I
t
h
a
s
a
g
e
r
a
n
g
e
b
et
w
ee
n
0
to
6
9
y
ea
r
s
.
MO
R
P
H
A
lb
u
m
I
I
co
n
tai
n
s
m
o
r
e
t
h
a
n
5
5
0
0
0
i
m
ag
e
s
o
f
1
3
0
0
0
s
u
b
j
ec
ts
.
I
t
in
cl
u
d
es
a
g
e
r
a
n
g
e
b
et
w
ee
n
1
6
to
7
7
y
ea
r
s
.
Fig
u
r
e
5
a
n
d
Fi
g
u
r
e
6
,
s
h
o
w
s
o
m
e
s
a
m
p
le
i
m
a
g
es
w
h
er
e
w
e
ca
n
o
b
s
er
v
e
th
e
v
ar
iatio
n
s
in
ill
u
m
in
at
io
n
,
ex
p
r
ess
io
n
,
h
ea
d
p
o
s
itio
n
an
d
ag
e
in
b
o
th
th
e
d
atasets
.
Fig
u
r
e
5
.
Sa
m
p
le
i
m
ag
e
s
f
r
o
m
FGNE
T
d
ataset
w
it
h
ag
e
v
alu
es
[
33
]
Fig
u
r
e
6
.
Sa
m
p
le
i
m
ag
e
s
f
r
o
m
MO
R
P
H
(
A
lb
u
m
I
I
)
d
ata
s
et
w
it
h
a
g
e
v
al
u
es
[
34
]
4
.
2
.
E
x
peri
m
e
nts o
n F
G
N
E
T
d
a
t
a
s
et
I
n
o
u
r
ex
p
er
i
m
en
t
s
,
w
e
u
s
ed
t
o
tal
9
8
0
im
a
g
es
o
f
8
2
s
u
b
j
ec
ts
f
r
o
m
FGNE
T
d
ataset
f
o
r
A
I
FR
-
C
N
N
a
m
o
n
g
w
h
ic
h
8
5
2
im
a
g
es
ar
e
u
s
ed
f
o
r
tr
ain
i
n
g
a
n
d
1
2
8
f
o
r
test
in
g
.
W
e
p
er
f
o
r
m
ed
v
ar
io
u
s
ex
p
er
i
m
e
n
ts
u
s
i
n
g
th
i
s
d
atase
t.
Firstl
y
,
w
e
u
s
ed
i
m
a
g
es
o
f
s
ize
3
2
х
3
2
af
ter
p
er
f
o
r
m
i
n
g
all
p
r
ep
r
o
ce
s
s
in
g
s
tep
s
.
T
h
ese
i
m
ag
e
s
in
cl
u
d
e
h
ea
d
p
o
s
e
v
ar
iatio
n
s
a
ls
o
.
W
e
u
s
ed
R
an
k
-
1
r
ec
o
g
n
it
i
o
n
as
a
p
er
f
o
r
m
an
ce
m
ea
s
u
r
i
n
g
p
ar
a
m
eter
.
Fro
m
th
e
r
esu
l
ts
o
b
tain
ed
,
it
is
f
o
u
n
d
th
at
9
8
im
a
g
e
s
f
r
o
m
1
2
8
ar
e
tr
u
l
y
r
ec
o
g
n
ized
.
I
t
in
d
icate
s
7
6
.
6
%
R
an
k
-
1
r
ec
o
g
n
itio
n
.
Seco
n
d
l
y
,
w
e
u
s
e
d
th
e
s
a
m
e
p
r
o
ce
d
u
r
e
an
d
s
a
m
e
n
u
m
b
er
o
f
i
m
a
g
es
b
u
t
w
ith
d
if
f
er
e
n
t
i
m
ag
e
s
ize
6
4
х
6
4
.
I
n
th
is
ex
p
er
i
m
en
t
8
7
i
m
a
g
es
f
r
o
m
1
2
8
g
o
t
co
r
r
ec
tly
r
ec
o
g
n
ized
.
I
t
s
h
o
w
s
6
8
.
8
%
R
an
k
-
1
r
ec
o
g
n
itio
n
ac
cu
r
ac
y
.
I
t
m
a
y
b
ec
au
s
e
o
f
t
h
e
n
et
w
o
r
k
ar
ch
itect
u
r
e
th
a
t
i
s
n
o
t
en
o
u
g
h
ca
p
ab
le
f
o
r
th
i
s
i
m
a
g
e
s
ize.
I
n
t
h
e
n
ex
t
e
x
p
er
i
m
e
n
t,
w
e
u
s
ed
o
n
l
y
s
tr
ai
g
h
t
p
o
s
e(
f
r
o
n
ta
l)
i
m
ag
es.
W
e
elim
i
n
ated
n
o
n
-
f
r
o
n
ta
l
i
m
ag
e
s
f
r
o
m
o
u
r
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:
2
0
8
8
-
8708
A
g
e
I
n
v
a
r
ia
n
t F
a
ce
R
ec
o
g
n
iti
o
n
u
s
in
g
C
o
n
v
o
lu
tio
n
a
l Neu
r
a
l Netw
o
r
k
(
M
r
u
d
u
la
N
imb
a
r
te)
2133
d
ataset.
Usi
n
g
F
GNE
T
d
ataset
w
e
e
x
tr
ac
ted
6
5
4
f
r
o
n
tal
i
m
a
g
es
f
o
r
tr
ain
i
n
g
a
n
d
9
6
f
r
o
n
tal
i
m
a
g
es
f
o
r
test
in
g
.
I
n
th
i
s
ex
p
er
i
m
en
t,
w
e
u
s
ed
r
esized
i
m
a
g
es
o
f
3
2
х
3
2
p
ix
els
as
it
h
as
co
m
p
ar
ati
v
e
g
o
o
d
p
er
f
o
r
m
a
n
ce
.
I
n
t
h
i
s
ca
s
e,
6
1
.
2
%
R
an
k
-
1
r
ec
o
g
n
i
ti
o
n
is
o
b
tai
n
ed
f
o
r
a
g
e
i
n
v
ar
ia
n
t
f
ac
e
r
ec
o
g
n
itio
n
u
s
i
n
g
p
r
o
p
o
s
ed
m
et
h
o
d
o
lo
g
y
.
I
n
th
e
la
s
t
e
x
p
er
i
m
e
n
t,
w
e
test
ed
o
u
r
p
r
o
p
o
s
ed
s
y
s
te
m
w
i
th
o
u
t
u
s
i
n
g
SVM
as
a
class
i
f
ier
.
Her
e
f
o
r
class
i
f
icatio
n
,
w
e
u
s
ed
E
u
clid
ian
Dis
tan
ce
w
i
th
Nea
r
es
t
Ne
ig
h
b
o
r
(
NN)
class
i
f
icatio
n
r
u
l
e.
A
s
t
h
e
i
m
ag
e
o
f
3
2
x
3
2
p
ix
els
g
i
v
es
b
etter
r
e
s
u
lt
s
,
w
e
u
s
ed
th
e
s
a
m
e
s
ize
in
t
h
is
e
x
p
er
i
m
e
n
tatio
n
.
He
r
e,
w
e
g
et
R
a
n
k
-
1
r
ec
o
g
n
itio
n
as
7
5
%.
T
ab
le
2
,
T
ab
le
3
an
d
T
ab
le
4
s
h
o
w
th
is
co
m
p
ar
ati
v
e
an
al
y
s
i
s
f
o
r
R
an
k
-
1
r
ec
o
g
n
itio
n
u
s
i
n
g
FGNE
T
d
ataset.
T
ab
le
2
.
C
o
m
p
ar
ativ
e
R
an
k
-
1
R
ec
o
g
n
itio
n
o
n
d
if
f
er
en
t i
m
a
g
e
s
izes f
r
o
m
FGNE
T
Data
s
et
(
A
ll i
m
a
g
es
w
it
h
v
ar
iatio
n
i
n
h
ea
d
p
o
s
e)
I
mag
e
S
i
z
e
T
r
a
i
n
i
n
g
I
mag
e
s
T
e
st
i
n
g
I
mag
e
s
R
a
n
k
-
1
R
e
c
o
g
n
i
t
i
o
n
3
2
х
3
2
8
5
2
1
2
8
7
6
.
6
%
6
4
х
6
4
8
5
2
1
2
8
6
8
.
8
%
T
ab
le
3
.
R
an
k
-
1
R
ec
o
g
n
itio
n
o
n
On
l
y
Fro
n
tal
I
m
a
g
es
f
r
o
m
FGNE
T
Data
s
et
I
mag
e
S
i
z
e
T
r
a
i
n
i
n
g
I
mag
e
s
T
e
st
i
n
g
I
mag
e
s
R
a
n
k
-
1
R
e
c
o
g
n
i
t
i
o
n
3
2
х
3
2
6
5
4
96
6
1
.
2
%
T
ab
le
4
.
C
o
m
p
ar
ativ
e
R
an
k
-
1
R
ec
o
g
n
itio
n
w
i
th
SVM/NN
f
r
o
m
FGNE
T
Data
s
et
(
A
ll i
m
a
g
es
w
it
h
v
ar
iatio
n
i
n
h
ea
d
p
o
s
e)
Fig
u
r
e
7
s
h
o
w
s
co
r
r
ec
tl
y
r
ec
o
g
n
ized
r
esu
lts
f
o
r
s
o
m
e
s
a
m
p
l
e
i
m
a
g
es
u
s
i
n
g
A
I
F
R
-
C
N
N
o
n
F
GNE
T
d
ataset.
First
co
l
u
m
n
s
h
o
w
s
th
e
i
m
a
g
es
u
s
ed
f
o
r
test
i
n
g
a
n
d
r
e
m
ain
in
g
co
lu
m
n
s
s
h
o
w
t
h
e
i
m
ag
e
s
o
f
th
e
s
a
m
e
p
er
s
o
n
at
d
if
f
er
e
n
t
ag
e
s
av
ai
l
ab
le
in
tr
ain
i
n
g
f
o
ld
er
.
W
e
u
s
ed
SVM,
th
e
s
u
p
er
v
i
s
ed
lear
n
in
g
al
g
o
r
ith
m
f
o
r
class
i
f
icatio
n
as
lab
els
ar
e
a
v
a
ilab
le,
it
s
h
o
w
s
t
h
e
cla
s
s
lab
el
to
w
h
ic
h
te
s
t
i
m
a
g
e
b
elo
n
g
s
.
I
t
is
o
b
s
er
v
ed
f
r
o
m
th
e
r
es
u
lt
s
t
h
at
A
I
F
R
-
C
NN
i
s
c
ap
ab
le
to
r
ec
o
g
n
ize
i
m
a
g
es
o
f
t
h
e
s
a
m
e
p
er
s
o
n
at
d
i
f
f
er
en
t
ag
e
s
.
Hen
ce
,
it
i
s
o
n
e
o
f
t
h
e
g
o
o
d
ap
p
r
o
ac
h
es.
Fro
m
Fi
g
u
r
e
7
,
w
e
ca
n
s
ee
F
GNE
T
d
ataset
co
n
tain
s
m
a
n
y
i
m
a
g
es
o
f
th
e
s
a
m
e
s
u
b
j
ec
t
w
it
h
lar
g
e
a
g
e
g
ap
a
n
d
f
o
r
m
o
r
e
a
g
e
g
ap
m
o
r
e
v
ar
i
atio
n
s
ar
e
t
h
er
e
i
n
t
h
e
i
m
a
g
es.
Fi
g
u
r
e
8
s
h
o
w
s
th
e
R
an
k
-
1
r
esu
lts
o
f
s
o
m
e
i
m
a
g
es
o
th
er
th
an
t
h
e
i
m
a
g
es
f
r
o
m
s
ta
n
d
ar
d
d
atas
et
al
so
.
Fo
r
th
is
,
w
e
ad
d
ed
o
u
r
s
a
m
p
le
i
m
a
g
es a
n
d
te
s
ted
o
n
A
I
F
R
-
C
N
N.
Fig
u
r
e
7
.
So
m
e
C
o
r
r
ec
tl
y
R
ec
o
g
n
ized
r
esu
lts
u
s
i
n
g
p
r
o
p
o
s
e
d
m
et
h
o
d
o
n
FGNE
T
Dat
aset.
T
h
e
n
u
m
b
er
s
b
elo
w
th
e
s
e
i
m
ag
e
s
s
h
o
w
t
h
e
ag
e
o
f
t
h
e
p
er
s
o
n
I
mag
e
T
r
a
i
n
i
n
g
I
mag
e
s
T
e
st
i
n
g
I
mag
e
s
R
a
n
k
-
1
R
e
c
o
g
n
i
t
i
o
n
3
2
х
3
2
w
i
t
h
S
V
M
8
5
2
1
2
8
7
6
.
6
%
3
2
х
3
2
w
i
t
h
N
N
8
5
2
1
2
8
7
5
%
Evaluation Warning : The document was created with Spire.PDF for Python.
I
SS
N
:
2
0
8
8
-
8708
I
n
t J
E
lec
&
C
o
m
p
E
n
g
,
Vo
l.
8
,
No
.
4
,
A
u
g
u
s
t 2
0
1
8
:
2
1
2
6
–
2
1
3
8
2134
Fig
u
r
e
8.
So
m
e
C
o
r
r
ec
tl
y
R
ec
o
g
n
ized
r
esu
lts
u
s
i
n
g
p
r
o
p
o
s
e
d
m
et
h
o
d
o
n
o
u
r
o
w
n
i
m
ag
e
s
4
.
3
.
E
x
peri
m
e
nts o
n M
O
RP
H
(
Albu
m
I
I
)
da
t
a
s
et
Fo
r
p
r
o
p
o
s
ed
ag
e
in
v
ar
ian
t
f
a
ce
r
ec
o
g
n
itio
n
u
s
i
n
g
C
NN,
w
e
u
s
ed
an
o
th
er
p
u
b
licl
y
a
v
ailab
le
d
ataset,
MO
R
P
H(
Alb
u
m
I
I
)
.
Usi
n
g
th
i
s
d
ataset,
w
e
p
er
f
o
r
m
ed
th
r
ee
ex
p
er
i
m
e
n
ts
.
First
l
y
,
w
e
h
a
v
e
u
s
ed
1
0
0
5
i
m
a
g
e
s
o
f
2
5
5
s
u
b
j
ec
ts
w
i
th
all
h
ea
d
p
o
s
es.
Am
o
n
g
t
h
ese
i
m
ag
e
s
7
5
0
i
m
ag
e
s
ar
e
u
s
ed
f
o
r
tr
ai
n
i
n
g
an
d
2
5
5
i
m
a
g
es
u
s
ed
f
o
r
tes
ti
n
g
.
R
e
s
u
lts
s
h
o
w
9
2
.
5
%
R
an
k
-
1
r
ec
o
g
n
itio
n
u
s
i
n
g
C
NN.
Seco
n
d
l
y
,
w
e
u
s
ed
2
0
8
4
f
r
o
n
tal
im
a
g
e
s
o
f
5
7
5
s
u
b
j
ec
ts
.
1
5
0
9
im
ag
e
s
f
o
r
tr
ain
in
g
an
d
5
7
5
im
ag
e
s
f
o
r
test
in
g
.
I
n
th
i
s
ca
s
e,
w
e
o
b
tain
9
2
.
8
%
R
an
k
-
1
r
ec
o
g
n
itio
n
.
I
n
th
e
la
s
t
ex
p
er
i
m
en
t,
as
in
FGNE
T
d
ataset
,
w
e
te
s
ted
p
er
f
o
r
m
an
ce
u
s
i
n
g
C
NN
w
it
h
E
u
cl
id
ian
Dis
ta
n
ce
an
d
Nea
r
est
Neig
h
b
o
r
(
NN)
as
class
if
ier
s
.
Her
e
w
e
o
b
tain
9
1
.
3
%
R
an
k
-
1
r
e
co
g
n
itio
n
.
T
ab
le
5
,
T
ab
le
6
,
T
a
b
le
7
an
d
T
ab
le
8
d
e
m
o
n
s
tr
ate
th
i
s
co
m
p
ar
is
o
n
.
T
ab
le
5
.
R
an
k
-
1
R
ec
o
g
n
itio
n
u
s
i
n
g
MO
R
P
H
(
A
lb
u
m
I
I
)
Data
s
et
w
it
h
SVM
(
A
ll I
m
ag
e
s
)
I
mag
e
s
T
r
a
i
n
i
n
g
I
mag
e
s
T
e
st
i
n
g
I
mag
e
s
R
a
n
k
-
1
R
e
c
o
g
n
i
t
i
o
n
A
l
l
(
F
r
o
n
t
a
l
+
N
o
n
-
f
r
o
n
t
a
l
)
7
5
0
2
5
5
9
2
.
5
%
T
ab
le
6
.
R
an
k
-
1
R
ec
o
g
n
itio
n
u
s
i
n
g
MO
R
P
H
(
A
lb
u
m
I
I
)
Data
s
et
w
it
h
SVM
(
On
l
y
Fro
n
tal
I
m
ag
e
s
)
I
mag
e
s
T
r
a
i
n
i
n
g
I
mag
e
s
T
e
st
i
n
g
I
mag
e
s
R
a
n
k
-
1
R
e
c
o
g
n
i
t
i
o
n
O
n
l
y
F
r
o
n
t
a
l
1
5
0
9
5
7
5
9
2
.
8
%
T
ab
le
7
.
C
o
m
p
ar
ativ
e
R
an
k
-
1
R
ec
o
g
n
itio
n
u
s
i
n
g
MO
R
P
H
(
A
lb
u
m
I
I
)
Data
s
et
w
i
th
N
N
I
mag
e
S
i
z
e
T
r
a
i
n
i
n
g
I
mag
e
s
T
e
st
i
n
g
I
mag
e
s
R
a
n
k
-
1
R
e
c
o
g
n
i
t
i
o
n
3
2
x
3
2
7
5
0
2
5
5
9
1
.
3
%
6
4
x
6
4
7
5
0
2
5
5
9
0
.
2
%
T
ab
le
8
.
C
o
m
p
ar
ativ
e
R
an
k
-
1
R
ec
o
g
n
itio
n
u
s
i
n
g
MO
R
P
H
(
A
lb
u
m
I
I
)
Data
s
et
w
i
th
SVM
an
d
NN
I
mag
e
S
i
z
e
T
r
a
i
n
i
n
g
I
mag
e
s
T
e
st
i
n
g
I
mag
e
s
R
a
n
k
-
1
R
e
c
o
g
n
i
t
i
o
n
3
2
x
3
2
w
i
t
h
S
V
M
7
5
0
2
5
5
9
2
.
5
%
3
2
x
3
2
w
i
t
h
N
N
7
5
0
2
5
5
9
1
.
3
%
Fig
u
r
e
9
s
h
o
w
s
s
o
m
e
co
r
r
ec
tl
y
r
ec
o
g
n
ized
r
esu
lts
f
o
r
A
I
F
R
-
C
NN
o
n
MO
R
P
H(
A
lb
u
m
I
I
)
d
ataset.
Fro
m
th
is
f
i
g
u
r
e,
w
e
o
b
s
er
v
e
t
h
at
t
h
er
e
is
n
o
m
o
r
e
v
ar
iatio
n
in
t
h
e
ag
e
o
f
th
e
p
er
s
o
n
as
co
m
p
ar
ed
to
FGNE
T
d
ataset.
Seco
n
d
l
y
,
MO
R
P
H
d
ataset
co
n
tai
n
s
le
s
s
n
u
m
b
er
o
f
i
m
ag
e
s
p
er
p
er
s
o
n
.
Fig
u
r
e
9.
Sa
m
p
le
C
o
r
r
ec
tl
y
R
e
c
o
g
n
ized
r
esu
lts
u
s
i
n
g
p
r
o
p
o
s
ed
m
et
h
o
d
o
n
MO
R
P
H
Data
s
e
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:
2
0
8
8
-
8708
A
g
e
I
n
v
a
r
ia
n
t F
a
ce
R
ec
o
g
n
iti
o
n
u
s
in
g
C
o
n
v
o
lu
tio
n
a
l Neu
r
a
l Netw
o
r
k
(
M
r
u
d
u
la
N
imb
a
r
te)
2135
Fig
u
r
e
10
s
h
o
w
s
t
h
e
C
u
m
u
la
tiv
e
Ma
tch
C
h
ar
ac
ter
is
t
ic(
C
MC)
c
u
r
v
e
o
f
t
h
e
p
r
o
p
o
s
ed
A
I
F
R
-
C
N
N
w
it
h
d
i
f
f
er
en
t
m
et
h
o
d
s
as
m
e
n
tio
n
ed
u
s
i
n
g
FG
NE
T
an
d
MO
R
P
H
(
Alb
u
m
I
I
)
d
ataset.
Fi
g
u
r
e
1
1
s
h
o
w
s
t
h
e
co
m
p
ar
ati
v
e
p
er
f
o
r
m
a
n
ce
an
al
y
s
i
s
o
f
p
r
o
p
o
s
ed
A
I
FR
-
C
NN
o
v
er
FGNE
T
an
d
MO
R
P
H
d
ataset.
Fig
u
r
e
1
0
.
C
u
m
u
lat
iv
e
Ma
tc
h
C
h
ar
ac
ter
is
tic
(
C
M
C
)
C
u
r
v
e
s
w
it
h
d
i
f
f
er
e
n
t
m
eth
o
d
s
o
n
B
o
t
h
Data
s
et
s
Fig
u
r
e
1
1.
C
o
m
p
ar
ativ
e
P
er
f
o
r
m
an
ce
An
al
y
s
is
(
R
a
n
k
-
1
R
ec
o
g
n
i
tio
n
)
o
f
F
GNE
T
Vs M
OR
P
H
Data
s
et
4
.
4
.
O
v
er
a
ll
co
m
pa
ra
t
iv
e
dis
cus
s
io
ns
Her
e
in
th
i
s
s
ec
tio
n
,
w
e
co
m
p
ar
e
o
u
r
p
r
o
p
o
s
ed
m
et
h
o
d
o
lo
g
y
w
it
h
e
x
is
tin
g
s
tate
-
of
-
t
h
e
-
ar
ts
.
Face
r
ec
o
g
n
itio
n
is
v
er
y
v
ast
ar
ea
in
t
h
e
f
ield
o
f
i
m
a
g
e
p
r
o
ce
s
s
in
g
a
n
d
p
atter
n
r
ec
o
g
n
itio
n
.
T
h
er
e
ar
e
v
ar
io
u
s
p
ar
am
eter
s
th
at
m
ak
e
it
r
ea
ll
y
d
if
f
ic
u
lt
li
k
e
v
ar
iatio
n
s
in
h
e
ad
p
o
s
itio
n
,
f
ac
ial
ex
p
r
ess
io
n
s
an
d
ag
in
g
ef
f
ec
ts
.
Ma
n
y
r
esear
ch
er
s
p
r
o
p
o
s
ed
th
eir
m
et
h
o
d
o
lo
g
ies
f
o
r
s
o
l
v
i
n
g
th
e
p
r
o
b
le
m
o
f
r
ec
o
g
n
izi
n
g
f
a
cial
i
m
ag
e
s
ac
r
o
s
s
ag
in
g
.
I
t
g
en
er
all
y
i
n
clu
d
e
s
t
h
e
s
tep
s
:
f
ac
e
d
etec
tio
n
,
p
r
ep
r
o
ce
s
s
in
g
,
f
ea
t
u
r
e
ex
tr
ac
tio
n
an
d
class
i
f
icat
io
n
.
P
er
f
o
r
m
a
n
ce
o
f
th
e
s
y
s
te
m
i
s
d
ir
ec
tl
y
d
ep
en
d
en
t
o
n
alg
o
r
it
h
m
s
u
s
ed
f
o
r
f
ea
t
u
r
e
ex
tr
ac
tio
n
an
d
class
i
f
icatio
n
.
B
u
t,
th
e
b
ea
u
t
y
o
f
C
o
n
v
o
lu
tio
n
Neu
r
al
Ne
t
w
o
r
k
s
i
s
t
h
at,
it
p
r
o
v
id
es
f
ea
tu
r
e
e
x
tr
ac
tio
n
a
n
d
class
if
icatio
n
i
n
a
s
in
g
le
s
tr
u
ct
u
r
e.
A
lt
h
o
u
g
h
C
N
N
is
a
v
er
y
p
o
w
er
f
u
l
to
o
l,
it
m
ak
es d
i
f
f
icu
lt to
d
ec
id
e
n
u
m
b
e
r
o
f
la
y
er
s
,
n
u
m
b
er
o
f
n
e
u
r
o
n
s
,
an
d
t
h
e
s
ize
o
f
i
n
p
u
t
i
m
a
g
e
p
r
o
v
id
ed
to
C
NN
ar
ch
itectu
r
e.
U
n
f
o
r
t
u
n
ate
l
y
,
th
er
e
is
n
o
w
a
y
o
r
f
o
r
m
u
la
av
a
ilab
le.
No
b
o
d
y
f
o
cu
s
ed
o
n
t
h
ese
i
s
s
u
es
r
at
h
er
th
e
y
p
r
o
p
o
s
ed
ar
ch
itectu
r
e
b
y
th
eir
o
w
n
w
a
y
.
W
e
also
f
o
llo
w
t
h
e
s
a
m
e
p
r
o
ce
s
s
t
o
d
ec
id
e
n
u
m
b
er
o
f
la
y
er
s
,
t
h
e
ir
d
i
m
en
s
io
n
s
,
an
d
s
ize
o
f
th
e
i
m
a
g
e
p
r
o
v
id
ed
as
in
p
u
t to
C
NN.
Fig
u
r
e
1
2
an
d
Fi
g
u
r
e
1
3
s
h
o
w
s
o
m
e
o
f
t
h
e
f
ailed
R
a
n
k
-
1
r
etr
iev
al
r
es
u
lts
f
r
o
m
FG
NE
T
an
d
MO
R
P
H
(
A
lb
u
m
I
I
)
r
e
s
p
ec
tiv
el
y
.
First
r
o
w
s
h
o
w
s
t
h
e
in
p
u
t
i
m
a
g
es
u
s
ed
f
o
r
tes
tin
g
,
s
ec
o
n
d
r
o
w
s
h
o
w
s
th
e
o
u
tp
u
t
o
f
o
u
r
m
et
h
o
d
i.e
.
f
ailed
to
r
ec
o
g
n
ize
co
r
r
ec
tl
y
an
d
t
h
ir
d
r
o
w
s
h
o
w
s
th
e
g
r
o
u
n
d
tr
u
t
h
i
m
ag
es
av
ailab
le
i
n
t
h
e
g
aller
y
.
I
t
is
s
ee
n
f
r
o
m
t
h
e
r
e
s
u
lt
s
t
h
at,
th
er
e
ar
e
m
o
r
e
in
tr
a
-
clas
s
d
if
f
er
en
ce
s
an
d
i
n
ter
-
class
s
i
m
i
lar
ities
i
n
b
o
th
th
e
d
atasets
.
Ma
n
u
all
y
a
l
s
o
,
it is
d
if
f
icu
lt to
id
en
ti
f
y
t
h
e
p
er
s
o
n
s
,
as so
m
e
o
f
th
e
m
lo
o
k
s
i
m
ilar
to
o
th
er
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.