T
E
L
KO
M
NIK
A
, V
ol
.
17
,
No.
6,
Dec
em
be
r
20
1
9,
p
p.3
1
1
0~
3
1
1
9
IS
S
N: 1
69
3
-
6
93
0
,
accr
ed
ited
F
irst
Gr
ad
e b
y K
em
en
r
istekdikti,
Decr
ee
No: 2
1/E/
K
P
T
/20
18
DOI:
10.12928/TE
LK
OM
N
IK
A
.v
1
7
i
6
.
12857
◼
31
10
Rec
ei
v
ed
A
pril
6
, 2
01
9
; Re
v
i
s
ed
J
ul
y
2
, 2
0
1
9
;
A
c
c
ep
t
e
d
J
ul
y
1
8
, 2
01
9
Rege
nerati
ng
fa
ce ima
ges f
r
om
mu
lt
i
-
s
pect
ral
palm
i
mag
es us
ing
mu
lt
iple
fu
si
on
met
ho
d
s
Raid Raf
i O
m
ar
A
l
-
N
ima,
M
o
a
t
as
e
m Y
as
e
en
A
l
-
Rid
h
a
*
, Fa
r
q
ad H
amid
A
b
d
u
lr
aheem
T
e
c
h
n
i
c
a
l
En
g
i
n
e
e
ri
n
g
Co
l
l
e
g
e
,
M
o
s
u
l
,
Nort
h
e
r
n
T
e
c
h
n
i
c
a
l
U
n
i
v
e
rs
i
ty
,
M
o
s
u
l
,
Ira
q
*C
o
rre
s
p
o
n
d
i
n
g
a
u
th
o
r,
e
-
m
a
i
l
:
m
o
a
l
ri
d
h
a
@gm
a
i
l
.c
o
m
Ab
strac
t
Th
i
s
p
a
p
e
r
e
s
ta
b
l
i
s
h
e
d
a
r
e
l
a
ti
o
n
s
h
i
p
b
e
twe
e
n
m
u
l
t
i
-
s
p
e
c
tr
a
l
p
a
l
m
i
m
a
g
e
s
a
n
d
a
fa
c
e
i
m
a
g
e
b
a
s
e
d
o
n
m
u
l
ti
p
l
e
fu
s
i
o
n
m
e
th
o
d
s
.
T
h
e
fi
rs
t
fu
s
i
o
n
m
e
th
o
d
to
b
e
c
o
n
s
i
d
e
re
d
i
s
a
fe
a
t
u
re
e
x
tra
c
ti
o
n
b
e
twe
e
n
d
i
ff
e
r
e
n
t
m
u
l
ti
-
s
p
e
c
tr
a
l
p
a
l
m
i
m
a
g
e
s
,
w
h
e
re
m
u
l
ti
-
s
p
e
c
tra
l
CASIA
d
a
t
a
b
a
s
e
was
u
s
e
d
.
Th
e
s
e
c
o
n
d
fu
s
i
o
n
m
e
th
o
d
t
o
b
e
c
o
n
s
i
d
e
re
d
i
s
a
s
c
o
r
e
f
u
s
i
o
n
b
e
twe
e
n
t
wo
p
a
rt
s
o
f
a
n
o
u
t
p
u
t
fa
c
e
i
m
a
g
e
.
O
u
r m
e
th
o
d
s
u
g
g
e
s
ts
t
h
a
t
b
o
t
h
ri
g
h
t
a
n
d
l
e
ft
h
a
n
d
s
a
re
u
s
e
d
,
a
n
d
th
a
t
e
a
c
h
h
a
n
d
a
i
m
s
to
p
r
o
d
u
c
e
a
s
i
g
n
i
fi
c
a
n
t
p
a
r
t
o
f
a
fa
c
e
i
m
a
g
e
b
y
u
s
i
n
g
a
M
u
l
ti
-
L
a
y
e
r
Pe
rc
e
p
tr
o
n
(M
L
P)
n
e
tw
o
rk
.
T
h
i
s
wil
l
l
e
a
d
to
th
e
s
e
c
o
n
d
fu
s
i
o
n
p
a
rt
t
o
re
c
o
n
s
tru
c
t
th
e
fu
l
l
-
fa
c
e
i
m
a
g
e
,
i
n
o
rd
e
r
to
e
x
a
m
i
n
e
i
t
s
a
p
p
e
a
r
a
n
c
e
.
Th
i
s
to
p
o
l
o
g
y
p
ro
v
i
d
e
d
i
n
te
re
s
t
i
n
g
re
s
u
l
t
s
o
f
Eq
u
a
l
Erro
r R
a
te
(EER) e
q
u
a
l
t
o
1
.9
9
%
.
Key
w
ords
:
fa
c
e
,
fu
s
i
o
n
,
M
L
P
n
e
u
ra
l
n
e
tw
o
rk
,
p
a
l
m
p
ri
n
t
Copy
righ
t
©
2
0
1
9
Uni
v
e
rsi
t
a
s
Ahm
a
d
D
a
hl
a
n.
All
rig
ht
s
r
e
s
e
rve
d
.
1.
Int
r
o
d
u
ctio
n
B
i
om
etri
c
ap
p
l
i
c
at
i
o
ns
are
c
urr
en
tl
y
ex
ten
s
i
v
e
l
y
us
ed
i
n
t
he
c
as
e
of
r
ec
og
ni
t
i
on
,
i
de
nti
f
i
c
at
i
on
or
au
t
he
n
ti
c
a
ti
on
s
y
s
tem
s
.
T
ha
t
i
s
be
c
au
s
e
b
i
ol
og
i
c
al
bi
om
etri
c
s
ha
v
e
un
i
q
ue
c
ha
r
ac
teri
s
ti
c
s
s
uc
h
as
a
f
a
c
e
an
d
pa
l
m
prin
t
[1]
.
F
ac
es
prov
i
de
c
ha
l
l
en
g
es
,
s
tarti
n
g
w
i
th
ex
tr
ac
ti
ng
the
m
ai
n
an
d
i
m
po
r
tan
t
pa
r
ts
,
the
n
us
i
ng
t
he
s
e
p
arts
to
r
ec
on
s
tr
uc
t
i
m
ag
es
of
the
f
ac
e
au
tom
ati
c
al
l
y
.
I
t
i
s
ap
p
arent
tha
t
the
i
m
ag
e
of
the
f
ac
e
ha
s
m
an
y
de
tai
l
s
,
f
or
i
ns
ta
nc
e
e
y
e
bro
w
s
,
e
y
es
,
no
s
e,
m
ou
th,
ea
r
s
a
nd
th
e
b
ou
n
da
r
y
of
the
f
a
c
e
.
T
hu
s
,
au
to
-
es
ta
bl
i
s
hi
ng
the
i
m
ag
e
of
the
f
ac
e t
o
be
c
l
ea
r
e
no
u
gh
to
r
ec
og
ni
z
e
i
s
r
ea
l
l
y
c
om
pl
i
c
at
ed
[
2].
In
the
l
i
te
r
atu
r
e,
s
ev
eral
s
tu
di
es
c
on
s
i
de
r
e
ge
ne
r
ati
ng
f
ac
e
f
ea
tu
r
e
s
fr
o
m
an
o
the
r
bi
o
m
et
r
i
c
c
ha
r
ac
teri
s
ti
c
.
T
hi
s
pe
r
ha
p
s
s
t
art
s
fr
o
m
the
w
or
k
o
f
S
ağ
i
r
oğ
l
u
an
d
Ö
z
k
a
y
a,
w
he
r
e
an
i
nte
l
l
i
ge
n
t
s
y
s
te
m
o
f
ge
ne
r
ati
ng
ey
e
s
f
ea
tu
r
e
s
fr
o
m
O
nl
y
f
i
ng
erpr
i
n
ts
w
as
p
r
e
s
en
ted
[3
]
.
Ö
z
k
ay
a
an
d
S
ağ
i
r
oğ
l
u
al
s
o
p
r
e
s
en
ted
a
s
tud
y
of
pr
od
uc
i
ng
f
a
c
e
bo
r
de
r
s
fr
o
m
fi
ng
e
r
p
r
i
nt
s
b
y
e
m
pl
oy
i
ng
the
A
r
ti
f
i
c
i
al
Neu
r
al
Netw
or
k
(
A
N
N)
[4]
.
T
he
n,
S
ağ
i
r
oğ
l
u
an
d Öz
k
ay
a p
ub
l
i
s
he
d e
x
te
nd
ed
s
tud
y
f
or
produ
c
i
ng
m
ai
n
f
a
c
e
fea
tu
r
e
s
fr
o
m
f
i
ng
e
r
p
r
i
nt
s
[5]
.
A
f
te
r
tha
t,
Chi
tr
av
an
s
hi
et
al
.
pr
o
po
s
ed
s
u
c
h
an
i
nte
r
e
s
ti
ng
w
or
k
o
f
ge
ne
r
a
t
i
ng
m
ai
n
fac
e
c
ha
r
a
c
te
r
i
s
ti
c
s
f
r
o
m
pa
l
m
f
ea
tu
r
e
s
[6
].
A
l
-
Ni
m
a
et
al
.
ap
proa
c
he
d
a
no
v
el
w
or
k
f
o
r
p
r
ed
i
c
ti
ng
f
ul
l
fa
c
e
fea
tu
r
e
s
fr
o
m
s
i
gn
atu
r
e
s
,
w
he
r
e
t
hi
s
w
as
the
f
i
r
s
t
s
tud
y
of
pr
od
uc
i
ng
ph
y
s
i
ol
og
i
c
al
bi
o
m
etri
c
s
fr
o
m
be
ha
v
i
ou
r
al
bi
o
m
e
tr
i
c
s
[7]
.
Y
an
g
et
al
.
r
ec
on
s
tr
u
c
te
d
f
a
c
e
i
m
ag
e
s
b
y
uti
l
i
z
i
ng
a
Ma
ni
f
ol
d
Con
s
t
r
ai
ne
d
C
on
v
ol
uti
on
al
S
pa
r
s
e
Codi
ng
[8
]
.
Lu
et
al
.
c
on
s
t
r
u
c
ted
hi
g
h
r
e
s
ol
uti
on
fac
e
i
m
ag
e
s
fr
o
m
l
ow
r
e
s
ol
u
ti
on
f
a
c
e
i
ma
ge
s
by
us
i
ng
c
on
di
ti
on
al
c
y
c
l
e
G
en
erati
v
e
A
dv
ers
a
r
i
al
Netw
or
k
(
G
A
N)
[9]
.
J
i
an
g
et
al
.
c
on
s
i
de
r
ed
a
s
tud
y
of
ge
ne
r
ati
ng
3D
f
a
c
e
s
fr
o
m
t
he
ge
o
m
et
r
y
de
tai
l
s
o
f
a
s
i
ng
l
e
f
ac
e
i
m
ag
e
[10
].
Li
et
al
.
prod
uc
e
d
ph
oto
r
ea
l
i
s
ti
c
fa
c
es
f
o
r
r
e
c
og
ni
z
i
ng
f
ac
i
al
s
k
e
tc
he
s
[11
].
Ma
i
e
t
al
.
i
l
l
u
s
t
r
ate
s
a
Nei
gh
bo
r
l
y
de
-
c
on
v
ol
uti
on
al
ne
u
r
al
N
etw
or
k
(
NbNe
t)
m
eth
od
t
o
r
ep
r
od
u
c
e
f
a
c
e
i
m
ag
e
s
fr
o
m
de
ep
te
m
pl
a
tes
[12
].
A
l
-
Ni
m
a
et
al
.
e
x
pl
oi
ted
ha
nd
-
d
or
s
al
i
ma
ge
s
to
g
en
era
te
ful
l
f
a
c
e
d
eta
i
l
s
b
y
us
i
n
g
B
ac
k
P
r
op
a
ga
ti
on
N
eu
r
al
(
B
P
N)
an
d
Ca
s
c
ad
e
-
F
o
r
w
ard
Neural
(
C
F
N
)
Ne
tw
or
k
s
[13
]
.
Up
to
no
w
, i
t
ha
s
be
en
no
ted
tha
t th
e
r
e i
s
no
r
e
f
e
r
en
c
e i
n rel
ati
on
t
o r
ep
r
od
u
c
e
f
a
c
e i
m
ag
e
s
f
r
o
m
m
u
l
ti
-
s
p
ec
tr
a
l
ha
nd
i
m
ag
e
s
.
B
e
c
au
s
e
v
ei
n
pa
tte
r
n
s
r
eq
ui
r
e
s
ub
s
tan
ti
a
l
eff
or
ts
to
ac
qu
i
r
e
,
a
s
y
s
tem
s
e
c
urit
y
c
an
be
i
n
c
r
ea
s
ed
f
u
r
the
r
.
T
h
er
ef
o
r
e,
thi
s
i
s
s
ue
ha
s
be
en
ad
d
r
es
s
ed
i
n
ou
r
w
or
k
.
T
he
ai
m
s
an
d
c
on
tr
i
bu
ti
on
s
of
th
i
s
pa
p
er ar
e a
s
f
ol
l
o
w
s
:
-
P
r
op
os
i
ng
r
ob
us
t f
ac
e reg
e
ne
r
ati
on
s
y
s
t
em
th
at
c
an
r
e
c
on
s
tr
uc
t th
e f
ul
l
d
eta
i
l
s
of
f
ac
e i
m
ag
es
.
-
E
x
pl
o
i
t
i
ng
m
uti
-
s
pe
c
tr
al
i
m
a
ge
s
of
r
i
g
ht
a
nd
l
ef
t
h
an
ds
as
r
eq
u
i
r
ed
i
n
pu
ts
.
T
hi
s
wo
ul
d
i
nc
r
e
as
e
the
s
ec
urit
y
an
d
an
t
i
-
s
po
of
i
ng
of
th
e
s
ug
g
es
ted
s
y
s
t
em
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
Rege
ne
r
at
i
ng
fa
c
e
i
m
ag
es
f
r
om
mu
l
t
i
-
s
pe
c
tr
al
pa
l
m
i
ma
ge
s
... (
Ra
i
d R
afi
O
ma
r
A
l
-
N
i
ma
)
3111
T
he
e
m
pl
o
y
e
d
ob
j
ec
ti
v
es
c
an
be
s
tat
ed
as
f
ol
l
o
w
s
:
f
i
r
s
tl
y
,
i
m
ag
e
pre
proc
es
s
i
ng
s
tep
s
wer
e
h
an
d
l
e
d
to
ex
tr
ac
t
a
n
d
en
ha
nc
e
th
e
h
an
d
i
m
ag
e
s
.
T
he
s
e
prepr
oc
es
s
i
n
gs
wer
e
e
l
i
m
i
na
te
d
or
c
r
op
pe
d,
m
orpho
l
og
i
c
al
op
erat
i
o
ns
,
ad
d
i
n
g
to
p
-
h
at
c
ha
r
ac
teri
s
t
i
c
s
an
d
an
un
s
ha
r
p
f
i
l
ter.
S
ec
on
dl
y
,
a
f
ea
ture
f
us
i
on
be
t
w
e
en
m
ul
ti
-
s
pe
c
tr
a
l
i
m
ag
es
to
c
om
bi
ne
th
e
f
ea
tur
e
s
,
where
Haar
wav
el
e
t
f
us
i
on
ba
s
ed
on
t
h
e
m
ea
n
r
ul
e
was
us
ed
.
T
hi
r
dl
y
,
a
wav
el
et
tr
an
s
f
orm
w
as
ap
p
l
i
ed
f
or
the
en
h
an
c
ed
a
nd
f
us
ed
i
m
ag
e.
F
ou
r
t
hl
y
,
the
M
LP
ne
u
r
al
n
et
wor
k
s
w
ere
tr
a
i
n
ed
b
y
c
o
ns
i
de
r
i
ng
a
r
i
g
ht
h
an
d
to
pred
i
c
t
the
i
nn
er
f
ac
e
i
m
ag
e
an
d
th
e
l
ef
t
ha
nd
to
pred
i
c
t
the
ou
t
er
f
ac
e
i
m
ag
e.
F
i
f
thl
y
,
a
s
c
ore
f
us
i
on
w
as
uti
l
i
z
ed
t
o
c
ol
l
ec
t
th
e
f
ac
e
i
m
ag
e
ac
c
ordi
ng
to
th
e
m
ax
i
m
u
m
or
ad
di
ng
r
ul
e.
S
i
x
th
l
y
,
the
s
am
e
proc
es
s
i
ng
s
s
ho
ul
d
be
f
ol
l
o
wed
to
tes
t
t
he
M
LP
s
us
i
ng
di
f
f
erent
da
ta.
F
i
na
l
l
y
,
t
he
l
as
t
d
ec
i
s
i
on
wa
s
tak
en
an
d
he
nc
e,
F
i
g
ures
1
an
d
2
de
m
on
s
tr
ate
t
he
b
l
oc
k
di
ag
r
a
m
s
f
or
ou
r
propos
e
d
m
eth
od
.
T
hi
s
n
e
w
s
ug
ge
s
te
d
to
po
l
og
y
wi
l
l
i
nc
r
ea
s
e
th
e
s
ec
urit
y
an
d
ef
f
i
c
ti
v
e
ne
s
s
of
th
e b
i
om
etri
c
s
y
s
tem
.
T
hi
s
pa
pe
r
i
s
org
an
i
z
e
d
as
f
ol
l
o
w
s
:
t
he
f
i
r
s
t
s
ec
ti
o
n
i
s
the
i
ntro
du
c
ti
on
,
pri
or
wor
k
an
d
the
pro
po
s
ed
m
eth
od
.
T
he
s
ec
on
d
s
ec
t
i
on
i
s
th
e
ha
n
d
i
m
ag
es
ex
tr
ac
ti
o
n
f
ol
l
o
wed
b
y
en
ha
nc
em
en
t.
T
he
thi
r
d
s
ec
ti
on
i
s
i
n
r
el
at
i
o
n
to
the
t
wo
t
y
p
es
of
f
us
i
on
s
,
f
ea
tures
an
d
the
s
c
ore.
T
he
f
ou
r
th
s
ec
ti
on
i
s
f
or
the
ML
P
s
arti
f
i
c
i
a
l
ne
ural
ne
t
wor
k
s
.
T
he
f
i
f
th
s
ec
ti
o
n
i
s
f
or
the
r
es
ul
ts
an
d
di
s
c
us
s
i
on
s
,
wi
th
the
f
i
n
al
s
ec
ti
on
be
i
ng
th
e
c
on
c
l
us
i
on
.
F
i
gu
r
e
1.
T
he
b
l
oc
k
di
ag
r
a
m
o
f
predi
c
ti
n
g f
ac
e f
r
o
m
ha
nd
i
m
ag
es
ba
s
ed
on
ML
P
ne
ural
ne
t
wor
k
s
an
d m
ul
ti
pl
e f
us
i
o
n
s
F
i
gu
r
e
2
.
T
he
b
l
oc
k
di
ag
r
a
m
o
f
th
e p
r
ep
r
oc
es
s
i
n
g s
te
ps
2.
Imag
e
E
xtrac
t
ion
a
n
d
E
n
h
ancemen
t
Data
ex
tr
ac
ti
on
an
d
an
al
y
s
i
s
are
c
on
s
i
de
r
ed
as
th
e
m
o
s
t
c
r
i
ti
c
al
an
d
es
s
en
ti
al
pro
bl
em
s
i
n
Im
ag
e
P
r
oc
es
s
i
ng
(
IM)
an
d
A
r
t
i
f
i
c
i
a
l
Int
e
l
l
i
g
en
c
e
(
A
I)
.
In
th
i
s
pa
p
er,
prepr
oc
e
s
s
i
ng
s
tep
s
are
ad
op
ted
t
o
ex
tr
ac
t,
en
h
an
c
e
an
d
no
r
m
al
i
z
e
th
e
da
t
a
i
n
order
to
b
e
prep
ared
f
or
the
(
ML
P
s
)
ne
t
w
ork
s
.
CA
S
I
A
m
ul
ti
-
s
pe
c
tr
al
pa
l
m
i
m
ag
es
da
tab
as
e
are
em
pl
o
y
e
d
to
bu
i
l
d
a
r
el
at
i
o
ns
hi
p
w
i
th
the
O
RL
f
ac
e
i
m
ag
es
da
tab
as
e.
T
he
m
ul
ti
-
s
pe
c
tr
al
i
m
ag
es
of
the
pa
l
m
c
on
s
i
s
t
of
7
20
0
j
pg
i
m
ag
es
f
or
10
0
d
i
f
f
erent
pe
o
pl
e.
A
l
l
t
he
s
e
i
m
ag
es
are
8
-
b
i
ts
gra
y
-
s
c
al
e
w
i
t
h
s
i
x
el
ec
tr
om
ag
ne
ti
c
s
pe
c
tr
um
s
s
tarti
ng
wi
th
4
60
nm
,
63
0
nm
,
70
0
n
m
,
850
nm
,
94
0
nm
an
d
w
hi
te
r
es
pe
c
t
i
v
el
y
.
T
hi
s
v
al
ua
bl
e
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 16
93
-
6
93
0
T
E
L
KO
M
NIK
A
V
ol
.
17
,
No
.
6,
D
ec
em
be
r
20
19
:
31
1
0
-
31
1
9
3112
da
ta
ba
s
e
i
nc
l
ud
e
d
i
m
ag
es
f
or
bo
t
h
r
i
g
ht
a
nd
l
ef
t
ha
nd
s
.
A
CCD
c
am
era
i
s
po
s
i
t
i
on
e
d
at
t
he
bo
tt
o
m
of
a
ha
nd
wi
t
h
s
om
e
s
pe
c
tr
um
l
i
gh
ts
.
T
he
r
e
wer
e
n
o
p
eg
s
or
r
es
tr
i
c
t
po
s
i
ti
on
s
f
or
the
p
al
m
i
n
the
de
v
i
c
e,
a
l
th
ou
g
h
i
t
ha
d
a
un
i
f
or
m
ba
c
k
ground
c
ol
ou
r
.
T
wo
s
es
s
i
on
s
w
ere
or
ga
n
i
z
ed
t
o
c
ap
ture
t
he
i
m
ag
es
.
E
ac
h
s
es
s
i
on
to
o
k
s
na
p
s
ho
ts
of
three
m
u
l
ti
-
s
pe
c
tr
um
i
m
ag
es
.
T
he
i
nte
r
v
al
p
erio
d
be
t
w
e
en
ea
c
h s
es
s
i
o
n
w
as
m
ore tha
n o
ne
m
on
th
[14
]
.
A
nu
m
be
r
of
m
orphol
og
i
c
al
op
erati
on
s
w
ere
ad
o
pte
d
af
ter
the
i
m
ag
e
e
l
i
m
i
na
ti
on
to
r
ed
uc
e
an
y
no
i
s
e
an
d
m
ai
nta
i
n
th
e
ha
nd
i
m
ag
e.
T
o
be
gi
n
wi
th,
t
he
c
r
op
pi
ng
i
m
ag
e
w
as
an
8
-
b
i
t
gra
y
s
c
al
e
de
no
te
d
as
(
,
)
:
2
→
[
0
,
255
]
whi
c
h
ne
e
de
d
t
o
b
e
c
on
v
erted
to
a
bi
n
ar
y
i
m
ag
e
de
f
i
n
ed
as
(
,
)
:
2
→
{
0
,
1
}
[1
5].
Con
s
eq
ue
n
tl
y
,
to
c
on
v
ert
the
i
m
ag
e
f
r
o
m
8
-
bi
t
gra
y
s
c
a
l
e
to
a
bi
n
ar
y
i
m
ag
e a
th
r
es
ho
l
d
w
as
ex
ec
ut
ed
i
n (1)
:
(
,
)
=
{
1
(
,
)
>
0
(
,
)
≤
(
1
)
f
or
the
b
i
na
r
y
i
m
ag
e
(
,
)
an
d
a
s
tr
uc
turi
ng
e
l
em
en
t
ℎ
(
,
)
,
the
e
r
os
i
on
⊝
an
d
d
i
l
ati
on
⨁
are
de
no
ted
as
[1
6
]:
(
⊝
ℎ
)
(
,
)
=
{
(
+
,
+
)
−
ℎ
(
,
)
}
(
2
)
(
⨁
ℎ
)
(
,
)
=
{
(
−
,
−
)
+
ℎ
(
,
)
}
(
3
)
s
m
al
l
w
h
i
te
ob
j
ec
ts
s
ho
u
l
d
be
r
em
ov
ed
f
r
om
the
bi
na
r
y
i
m
ag
e,
w
h
i
l
e
ho
l
d
i
ng
th
e
v
er
y
l
arge
area.
A
n o
pe
n
m
orphol
og
i
c
al
op
erati
o
n m
a
y
s
ol
v
e t
hi
s
i
s
s
ue
. S
ee
(
4):
(
ℎ
)
=
(
⊝
ℎ
)
⨁
ℎ
≤
(
4
)
where:
i
s
a
s
pe
c
i
f
i
c
area.
Nev
erthe
l
es
s
,
the
r
e
wi
l
l
s
ti
l
l
be
s
o
m
e
s
m
al
l
ob
j
ec
ts
c
o
nn
ec
te
d
to
the
l
arges
t
ha
n
d
area,
whi
c
h
wi
l
l
no
t
be
d
el
e
ted
b
y
(
4).
T
o
r
e
m
ov
e
th
es
e
ob
j
ec
ts
,
a
c
om
pl
e
m
en
t
i
m
ag
e
̂
de
f
i
ne
d
as
̂
(
,
)
:
2
→
{
1
,
0
}
w
as
prod
uc
ed
f
r
om
the
l
as
t
op
erati
on
.
A
m
aj
or
m
orp
ho
l
og
i
c
a
l
op
era
ti
on
was
pe
r
f
or
m
ed
c
on
s
ec
uti
v
el
y
,
to
c
l
ea
r
t
he
u
ne
x
p
ec
ted
no
i
s
e
b
y
s
ett
i
n
g
th
em
to
1's
i
f
the
n
ei
gh
b
ou
r
ho
od
m
aj
orit
i
es
are
o
ne
s
[1
7,
1
8].
H
en
c
e,
the
pa
l
m
i
m
ag
e
w
i
t
h
f
i
ng
ers
i
s
e
as
i
l
y
c
r
ea
ted
b
y
c
om
bi
ni
ng
t
he
o
r
i
gi
n
al
i
m
ag
e
w
i
t
h t
he
c
om
pl
em
en
t i
m
a
ge
as
de
s
c
r
i
b
ed
i
n
(
5):
(
,
)
=
{
(
,
)
̂
(
,
)
=
0
+
̂
(
,
)
̂
(
,
)
=
1
(
5
)
where:
(
,
)
i
s
the
ne
w
c
r
ea
ted
i
m
ag
e,
s
i
s
a
s
m
al
l
s
c
al
ar
v
al
ue
a
nd
̂
(
,
)
i
s
the
r
es
ul
t
i
ng
bi
n
ar
y
i
m
ag
e
af
ter
(
4).
A
n
ex
am
pl
e
of
a
h
an
d
i
m
ag
e
be
f
ore
an
d
af
ter
th
e
m
orphol
og
i
c
al
proc
es
s
i
ng
s
i
s
gi
v
en
i
n
F
i
g
ure
3.
(
a)
(
b)
(
c
)
(
d)
(
e)
(
f
)
F
i
gu
r
e
3
.
A
ha
nd
i
m
ag
e b
ef
ore an
d a
f
ter th
e m
orphol
o
gi
c
al
op
erati
on
s
: (a)
i
np
ut
i
m
ag
e,
(
b) el
i
m
i
na
t
i
ng
an
d c
on
v
ert
i
ng
to
th
e
bi
na
r
y
i
m
ag
e,
(
c
)
r
em
ov
i
ng
t
he
s
m
al
l
are
as
,
(
d) c
on
v
ert
i
n
g t
o t
he
c
om
pl
em
en
t i
m
ag
e,
(
e) r
em
ov
i
ng
the
u
ne
x
pe
c
t
ed
n
oi
s
e
an
d
(
d) c
o
m
bi
ni
ng
t
he
or
i
g
i
na
l
i
m
ag
e w
i
t
h a
c
l
ea
r
ba
c
k
ground
A
s
e
ac
h
m
ul
ti
-
s
pe
c
tr
a
l
i
m
ag
e
ha
s
s
pe
c
i
f
i
c
f
ea
tur
es
,
en
ha
nc
em
en
ts
an
d
f
ea
ture
f
us
i
on
s
are
i
m
po
r
tan
t
to
c
o
l
l
ec
t
as
m
an
y
ha
v
e
c
h
arac
ter
i
s
ti
c
s
.
O
bta
i
n
i
ng
top
-
h
at
v
a
l
ue
s
f
r
o
m
ea
c
h
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
Rege
ne
r
at
i
ng
fa
c
e
i
m
ag
es
f
r
om
mu
l
t
i
-
s
pe
c
tr
al
pa
l
m
i
ma
ge
s
... (
Ra
i
d R
afi
O
ma
r
A
l
-
N
i
ma
)
3113
i
m
ag
e,
f
ol
l
o
w
e
d
b
y
a
dd
i
n
g
th
em
to
t
he
r
es
ul
t
i
n
g
orig
i
na
l
i
m
ag
e
(
,
)
ap
pe
are
d
to
be
a
hi
g
h
-
qu
al
i
t
y
en
h
an
c
em
en
t.
Ini
t
i
a
l
l
y
,
a
s
tr
uc
tur
i
ng
el
e
m
en
t
i
s
c
r
e
ate
d
as
a
di
s
k
s
ha
pe
of
on
e
’
s
v
a
l
ue
s
[
18
]
:
=
2
×
(
6
)
where:
R
i
s
r
a
di
us
of
p
i
x
el
s
.
Cons
e
qu
e
ntl
y
,
th
e
t
op
-
h
at
d
eta
i
l
s
are
i
s
o
l
ate
d
f
r
om
the
i
m
ag
e
as
s
ho
w
n
i
n t
hi
s
e
qu
at
i
o
n
[1
9]
:
=
–
(
)
(
7
)
where:
i
s
the
i
m
ag
e
af
ter
the
to
p
-
ha
t
f
i
l
ter
an
d
i
s
the
s
tr
uc
turi
ng
e
l
em
en
t.
He
nc
e,
the
s
e
top
-
ha
t
f
ea
tures
are
ad
de
d
to
th
e h
a
nd
i
m
ag
e a
c
c
ordi
n
g t
o (8)
[
18
]
:
=
+
(
8
)
then
,
an
un
s
ha
r
p
f
i
l
ter
i
s
ap
p
l
i
e
d
to
en
ha
nc
e
the
de
tai
l
s
of
the
e
dg
es
.
S
o,
an
i
s
produc
e
d a
s
f
ol
l
o
w
s
[2
0]
:
(
,
)
=
(
,
)
+
(
,
)
(
9
)
where:
i
s
a
p
os
i
ti
v
e
s
c
al
e
f
ac
tor
a
nd
(
,
)
i
s
th
e
c
orr
ec
ti
on
s
i
gn
a
l
,
c
a
l
c
u
l
ate
d
as
a
n
o
ut
pu
t
of
a h
i
gh
p
as
s
f
i
l
ter
[2
0]
.
A
n
ex
am
pl
e o
f
th
e i
m
ag
e e
nh
a
nc
em
en
t p
r
oc
es
s
i
ng
i
s
gi
v
e
n i
n
F
i
g
ure
4.
(
a)
(
b)
(
c
)
(
d)
F
i
gu
r
e
4
.
A
n e
x
am
pl
e
of
i
m
ag
e
en
h
an
c
em
en
t p
r
oc
es
s
i
ng
s
: (a)
a
ha
n
d i
m
ag
e
n
ee
d
s
to
be
en
ha
nc
ed
, (
b) top
-
h
at
i
m
ag
e d
et
ai
l
s
, (c
)
ad
d
i
ng
the
top
-
ha
t d
e
tai
l
s
to
the
orig
i
n
al
h
an
d
i
m
ag
e
an
d (d)
s
h
arpen
ed
i
m
ag
e e
nh
an
c
em
en
t
3
.
F
u
sion
F
us
i
on
b
et
w
e
en
d
i
f
f
erent
bi
om
etri
c
f
ea
ture
ac
qu
i
r
em
en
ts
c
ou
l
d
be
c
on
s
i
de
r
ed
as
a
s
i
gn
i
f
i
c
an
t
m
eth
od
to
i
nc
r
ea
s
e
the
ab
i
l
i
t
y
an
d
s
ec
ur
i
t
y
of
the
s
y
s
tem
s
[21]
.
T
w
o
f
us
i
on
m
eth
od
s
are
us
ed
:
f
ea
ture
f
us
i
on
b
a
s
ed
on
the
wav
el
e
t
w
i
t
h
m
e
an
r
ul
es
to
ex
tr
ac
t
th
e
h
an
d
f
ea
tures
fr
o
m
the
m
ul
ti
-
s
pe
c
tr
a
l
i
m
ag
es
a
nd
s
c
ore f
us
i
on
t
o c
om
bi
ne
or c
ol
l
ec
t th
e
f
i
na
l
f
ac
e i
m
ag
e.
3.1
.
F
e
atu
r
e
F
u
sion
A
f
us
i
on
be
t
ween
i
m
ag
es
i
s
c
on
s
i
de
r
ed
as
a
n
i
nt
eres
ti
n
g
tec
hn
i
qu
e
,
s
o
as
to
i
nc
r
ea
s
e
the
l
e
v
el
of
ef
f
i
c
i
en
c
y
of
an
y
bi
om
etri
c
s
y
s
t
em
.
Col
l
ec
ti
n
g
m
ul
ti
pl
e
i
nf
orm
ati
on
f
r
om
di
ff
erent
i
m
ag
es
i
s
a
prob
l
em
w
h
i
c
h
c
a
n
b
e
s
ol
v
ed
b
y
th
e
f
us
i
o
n
tec
h
ni
qu
e.
M
ergi
ng
t
wo
i
m
ag
es
i
nt
o
on
e
i
n
di
v
i
du
a
l
i
m
ag
e i
s
a
t
y
p
e o
f
f
us
i
on
a
b
i
l
i
t
y
tha
t p
r
o
v
i
de
s
m
ore da
ta
i
n
a s
i
n
gl
e
i
m
ag
e
[2
2]
.
Mu
l
t
i
-
s
pe
c
tr
a
l
h
an
d
i
m
ag
e
s
ha
v
e
m
an
y
c
h
arac
teri
s
t
i
c
s
f
r
o
m
pa
l
m
,
f
i
ng
ers
an
d
ha
n
d
ge
om
etr
y
to
v
e
i
n,
l
i
ne
s
an
d
s
m
al
l
pa
tte
r
ns
.
F
us
i
on
be
t
ween
ea
c
h
t
wo
m
ul
ti
-
s
pe
c
tr
al
i
m
ag
e
t
y
p
es
wi
l
l
m
ai
nta
i
n
a
nd
c
ov
er
th
e
c
om
bi
na
ti
on
i
nf
orm
ati
on
.
I
n
th
i
s
pa
p
er,
i
m
ag
e
s
pe
c
tr
um
s
of
46
0
n
m
wer
e
c
om
bi
ne
d
wi
th
th
e
i
m
a
ge
s
p
ec
tr
um
s
of
63
0
nm
.
S
i
m
i
l
arl
y
,
70
0
nm
i
m
ag
es
w
er
e
m
erged
wi
th
850
nm
,
w
h
i
l
s
t
94
0
nm
wer
e
f
us
ed
wi
th
the
w
h
i
te
l
i
gh
t
i
m
ag
e.
T
hi
s
f
ea
ture
f
u
s
i
on
m
eth
od
i
s
i
m
pl
em
en
ted
b
y
w
a
v
e
l
et
f
us
i
on
wi
th
t
he
Ha
ar
s
i
gn
al
,
4
-
l
ev
el
s
an
d
m
ea
n
r
ul
e
f
or
the
bo
th
ap
prox
i
m
ati
on
s
an
d
de
t
ai
l
s
pa
r
ts
. S
ee
F
i
gu
r
e
5
(
a).
A
c
c
ordi
n
g
to
th
i
s
f
us
i
on
,
f
ou
r
c
oe
ff
i
c
i
en
ts
wi
l
l
b
e
g
en
erat
ed
f
or
ea
c
h
i
m
ag
e
af
ter
the
wa
v
el
et
tr
an
s
f
or
m
(
on
e
a
pp
r
ox
i
m
ati
on
a
nd
thre
e
d
eta
i
l
s
)
.
Im
O
T
U1
an
d
Im
O
T
U2
are
t
w
o
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 16
93
-
6
93
0
T
E
L
KO
M
NIK
A
V
ol
.
17
,
No
.
6,
D
ec
em
be
r
20
19
:
31
1
0
-
31
1
9
3114
m
ul
ti
-
s
pe
c
tr
al
pa
l
m
i
m
ag
es
,
(
LL
1
)
i
s
the
a
pp
r
ox
i
m
ati
o
n
an
d
(
LH
1
,
HL
1
,
HH
1
)
are
the
de
t
ai
l
s
of
the
f
i
r
s
t s
pe
c
tr
u
m
i
m
ag
e.
Li
k
ew
i
s
e,
(
LL
2
)
i
s
th
e a
pp
r
ox
i
m
ati
on
an
d (LH
2
,
HL
2
, HH
2
)
are th
e d
et
ai
l
s
of
th
e s
ec
on
d
s
pe
c
tr
um
i
m
ag
e.
Her
ea
f
ter, (
1
0) i
s
us
ed
to
c
om
bi
n
e t
h
e i
nf
orm
ati
on
:
3
=
(
(
1
,
2
)
;
(
1
,
1
,
1
,
2
,
2
,
2
)
)
(
1
0)
where:
I
DW
T
i
s
th
e i
nv
ers
e o
f
th
e 2
D
w
a
v
e
l
et
tr
a
ns
f
orm
an
d
AV
i
s
th
e
av
erag
e.
(
a)
(
b)
F
i
gu
r
e
5.
F
us
i
on
m
eth
od
s
: (
a) f
ea
ture f
us
i
on
m
eth
od
be
twee
n
t
w
o s
p
ec
tr
al
i
m
ag
es
(
94
0n
m
an
d
w
h
i
te)
ba
s
ed
o
n t
he
w
a
v
e
l
et
m
eth
od
an
d
(
b) s
c
ore f
us
i
on
m
eth
od
be
t
w
e
en
t
wo f
ac
e i
m
ag
e p
a
r
ts
af
ter the
ML
P
ne
ura
l
n
et
wor
k
s
3.2
.
S
cor
e
F
u
sion
S
c
ore
m
atc
hi
ng
l
e
v
e
l
f
us
i
on
i
s
a
m
eth
od
tha
t
i
s
u
s
ed
ex
ten
s
i
v
e
l
y
wi
th
th
e
m
ul
ti
pl
e
bi
om
etri
c
m
od
el
s
[2
3]
.
It
i
s
pe
r
f
or
m
ed
af
ter
the
au
th
en
ti
c
ati
on
proc
es
s
i
ng
,
where
ea
c
h
ou
tpu
t
i
s
c
al
c
ul
a
ted
i
n
di
v
i
du
al
l
y
an
d
s
ub
s
eq
ue
ntl
y
,
a
c
om
bi
na
t
i
o
n
s
c
orin
g
l
e
v
e
l
i
s
ap
pl
i
ed
[
24
]
.
T
he
O
RL
da
ta
ba
s
e
of
f
ac
e i
m
ag
es
ar
e u
s
e
d
i
n
th
i
s
p
ap
er.
T
hi
s
d
ata
b
as
e
i
s
pro
du
c
ed
i
n
A
T
&
T
l
ab
orator
i
es
at
Cam
brid
ge
Un
i
v
ers
i
t
y
thr
ou
gh
the
c
ol
l
ab
orat
i
on
of
three
gro
up
s
(
S
pe
ec
h
,
V
i
s
i
o
n
an
d
R
ob
o
ti
c
s
)
.
It
c
on
s
i
s
ts
of
4
00
i
m
ag
es
f
r
o
m
40
pe
op
l
e
an
d
e
ac
h
pe
r
s
on
ha
s
a
d
i
f
f
erent
ex
pres
s
i
on
[
25
]
.
T
he
c
r
i
ti
c
al
prob
l
em
i
n o
ur wor
k
i
s
ge
ne
r
a
ti
n
g
f
ac
e i
m
a
ge
s
us
i
ng
s
c
ore
l
e
v
e
l
f
us
i
o
n
.
A
f
ter
w
ards
, a
de
c
i
s
i
o
n
i
s
t
ak
en
ac
c
ordi
ng
to
th
e
m
os
t
c
l
ea
r
an
d
di
s
ti
nc
t
i
m
ag
e.
T
o
ex
pl
ai
n
i
n
m
ore
d
eta
i
l
s
,
t
w
o
es
s
en
ti
a
l
pa
r
ts
of
a
f
ac
e
i
m
ag
e
are
i
nt
en
d
ed
to
b
e
pre
di
c
te
d
b
y
us
i
ng
ML
P
ne
ur
al
ne
t
w
ork
s
. T
hi
s
i
s
the
m
i
dd
l
e
p
art
of
a
f
ac
e
i
m
ag
e,
whi
c
h
m
ai
nl
y
c
o
n
s
i
s
ts
of
the
e
y
es
,
n
os
e
an
d
m
ou
th.
T
hu
s
,
the
r
i
g
ht
-
ha
nd
i
m
ag
es
are
us
ed
to
pred
i
c
t
thi
s
m
i
dd
l
e
pa
r
t.
T
he
ou
t
er
pa
r
t
of
a
f
ac
e,
w
h
i
c
h
c
o
m
m
on
l
y
ha
s
th
e
ea
r
s
,
ha
i
r
an
d
l
o
wer
j
aw
em
pl
o
y
s
the
l
ef
t
-
hand
i
m
ag
es
to
pre
di
c
t
thi
s
bo
un
d
ar
y
pa
r
t.
S
ub
s
e
qu
e
ntl
y
,
a
s
c
ore
f
us
i
on
i
s
pe
r
f
or
m
ed
b
y
us
i
n
g
the
m
ax
i
m
u
m
or
ad
di
n
g
r
u
l
e
t
o
c
on
s
tr
uc
t
the
f
ac
e i
m
ag
e.
F
i
gu
r
e
5 (b
)
s
ho
w
s
t
he
i
d
ea
of
th
e s
c
ore f
us
i
on
l
e
v
el
.
A
s
s
um
e
FA
1
i
s
the
i
nn
er
(
m
i
dd
l
e)
f
ac
e
i
m
ag
e
pa
r
t
an
d
FA
2
i
s
th
e
ou
ter
f
ac
e
i
m
ag
e
pa
r
t.
Nex
t,
FA
i
s
th
e f
us
ed
i
m
ag
e f
r
o
m
th
e m
i
dd
l
e a
nd
t
he
b
ou
nd
ar
y
pa
r
ts
. S
ee
(
1
1):
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
Rege
ne
r
at
i
ng
fa
c
e
i
m
ag
es
f
r
om
mu
l
t
i
-
s
pe
c
tr
al
pa
l
m
i
ma
ge
s
... (
Ra
i
d R
afi
O
ma
r
A
l
-
N
i
ma
)
3115
=
(
1
;
2
)
(
11
)
A
c
c
ordi
n
g
to
th
i
s
tec
hn
i
qu
e
the
s
ec
urit
y
of
the
s
y
s
tem
wi
l
l
i
nc
r
ea
s
e
be
c
au
s
e
t
wo
ha
nd
i
m
ag
es
are
r
eq
ui
r
e
d t
o
pro
v
e t
he
f
ac
e i
m
ag
e.
4.
Neu
r
al
N
etw
o
r
k
A
r
ti
f
i
c
i
a
l
Ne
ural
Net
wor
k
(
A
NN)
i
s
on
e
of
the
m
os
t
po
p
ul
ar
t
y
p
es
of
the
A
r
t
i
f
i
c
i
al
Int
e
l
l
i
ge
nc
e
(
A
I)
.
In
r
ec
en
t
y
ea
r
s
,
i
t
h
as
be
c
om
e
wi
de
s
pr
ea
d
i
n
di
f
f
erent
f
i
el
ds
.
B
i
om
etri
c
v
erif
i
c
at
i
o
n
,
i
d
en
t
i
f
i
c
ati
o
n
an
d
c
l
as
s
i
f
i
c
ati
o
n
are
ex
am
pl
es
of
the
s
e
f
i
el
ds
.
T
he
r
e
are
t
wo
m
ai
n
t
y
p
es
of
A
NN,
s
up
erv
i
s
e
d
an
d
un
s
u
pe
r
v
i
s
ed
.
E
ac
h
on
e
of
t
he
s
e
t
y
p
es
att
em
pts
to
s
i
m
ul
ate
a
s
i
gn
i
f
i
c
an
t
tas
k
i
n
a
hu
m
an
brai
n.
Mo
r
e
ov
er,
the
r
e
are
t
w
o
es
s
en
ti
al
s
tag
e
s
i
n
an
y
A
N
N:
the
l
ea
r
n
i
ng
s
ta
ge
a
nd
t
he
tes
ti
n
g
s
tag
e.
In
th
e
f
i
r
s
t
s
tag
e,
th
e
n
et
w
ork
l
ea
r
ns
t
he
i
np
uts
a
nd
ge
ne
r
ate
s
s
p
ec
i
f
i
c
w
e
i
gh
ts
t
o
m
an
ag
e
th
e
pr
ob
l
em
.
In
t
he
s
ec
o
nd
s
tag
e
,
t
he
A
N
N
de
a
l
s
w
i
th
oth
er
i
np
uts
whi
c
h
ha
v
e
no
t
be
e
n
s
ee
n
be
f
ore
[26]
.
In
ou
r
wo
r
k
,
s
up
erv
i
s
e
d
ML
P
ne
ural
ne
t
wor
k
s
ar
e
i
n
v
es
ti
ga
te
d a
nd
a
da
pte
d
t
o a
c
hi
ev
e t
h
ei
r
t
as
k
s
.
F
i
r
s
t
of
al
l
,
the
i
np
u
t
d
ata
of
ea
c
h
i
m
ag
e n
ee
d
t
o
b
e p
r
e
pa
r
ed
f
or
the
M
LP
ne
t
wor
k
.
T
hu
s
,
ea
c
h
i
np
ut
i
m
ag
e
i
s
s
eg
m
en
ted
i
n
to
s
p
ec
i
f
i
c
m
atri
c
es
wi
th
di
f
f
erent
s
i
z
es
of
5
×
5,
7×
7
,
…,
13
×
1
3
pi
x
el
s
.
T
hi
s
w
i
l
l
en
s
ure
pr
ov
i
di
ng
d
i
f
f
erent
ov
er
l
ap
s
be
t
w
e
en
t
he
m
atr
i
c
es
.
A
Coef
f
i
c
i
en
t
of
v
ari
an
c
e
i
s
c
al
c
u
l
ate
d t
o
ea
c
h s
eg
m
en
t a
s
i
l
l
um
i
na
te
d i
n
(
12
)
-
(
14
)
[2
7]
:
=
1
∑
=
1
(
12
)
=
√
1
−
1
∑
(
−
)
2
=
1
(
13
)
=
(
14
)
where
:
n
i
s
the
n
um
be
r
of
pi
x
el
s
i
n
ea
c
h
s
e
gm
en
t,
s
eg
i
s
the
m
atri
x
of
5×
5,
7
×
7,
9×
9,
11
×
11
or
13
×
13
p
i
x
el
s
,
M
ea
n
i
s
the
av
era
ge
,
S
D
i
s
the
s
ta
nd
ard
de
v
i
ati
on
an
d
CV
i
s
th
e
c
oe
f
f
i
c
i
en
t
of
v
ari
an
c
e.
T
he
ad
v
a
nta
g
es
of
us
i
ng
the
C
V
are:
no
d
i
m
en
s
i
o
n
un
i
ts
c
an
be
c
o
ns
i
d
ered,
a
l
l
v
a
l
ue
s
are p
os
i
t
i
v
e,
t
he
d
i
f
f
erenc
es
w
i
l
l
b
e
gi
v
en
as
s
m
al
l
r
ati
o
v
a
l
ue
s
(
t
hi
s
w
i
l
l
a
v
o
i
d
th
e
A
NNs
o
v
erl
oa
d
i
n
th
e
ne
x
t
s
tag
e),
t
he
v
ari
an
c
es
be
t
ween
t
he
s
am
e
v
ec
tor
t
y
p
e
c
an
be
c
a
l
c
ul
a
te
d
(
thi
s
w
i
l
l
b
e
v
a
l
ua
bl
e
f
or
the
s
am
e
target
i
n
the
tr
a
i
n
i
ng
s
ta
ge
)
,
an
d
the
v
aria
nc
es
be
t
w
e
en
th
e
di
f
f
erent
v
ec
tor
t
y
p
es
c
an
be
d
ete
r
m
i
ne
d
(
thi
s
wi
l
l
be
us
ef
ul
f
or
the
s
ta
ge
of
f
us
i
on
be
t
ween
t
wo
di
ff
erent
t
y
pe
s
)
.
T
he
s
ec
on
d
s
tep
,
i
s
arr
an
gi
n
g
th
e
C
V
v
al
ue
s
i
nto
a
one
-
d
i
m
en
s
i
on
a
l
v
ec
tor
f
o
r
ea
c
h
i
m
ag
e.
T
he
f
i
na
l
i
np
u
t p
r
ep
arat
i
on
i
s
m
ap
pi
n
g t
h
e i
np
ut
da
t
a i
n
[0
,1]
r
an
g
e a
s
s
ho
wn
i
n (1
5)
[28]
:
=
(
ma
x
(
)
−
min
(
)
)
×
(
−
min
(
)
)
(
ma
x
(
)
−
min
(
)
)
+
min
(
)
(
15
)
5.
Re
sult
s a
n
d
D
isc
s
u
ss
i
o
n
s
T
he
pe
r
f
or
m
an
c
e
of
the
propos
e
d
m
eth
od
i
s
ex
a
m
i
ne
d
an
d
c
om
pa
r
ed
w
i
t
h
ot
he
r
wor
k
.
T
he
da
t
ab
as
es
are
c
ol
l
ec
ted
as
wel
l
as
org
an
i
z
e
d
i
nto
gro
up
s
.
A
n
i
np
ut
grou
p
of
40
20
m
ul
ti
-
s
pe
c
tr
al
i
m
ag
e
i
s
us
e
d
i
n
the
A
NNs
tr
a
i
ni
ng
s
tag
e
a
nd
an
o
the
r
i
np
ut
grou
p
of
80
4
m
ul
ti
-
s
pe
c
tr
al
i
m
ag
es
u
ti
l
i
z
ed
i
n
A
NNs
t
es
ti
ng
s
t
ag
e.
I
n
ad
d
i
ti
on
,
e
ac
h o
ne
of
t
he
t
wo
grou
ps
ha
v
e
be
e
n
s
e
pa
r
at
ed
i
n
to
t
wo
o
the
r
grou
ps
;
t
he
l
ef
t
h
an
d
an
d
r
i
gh
t
-
h
an
d
gr
ou
ps
.
In
the
tr
a
i
n
i
ng
s
tag
e,
ea
c
h h
an
d
gro
up
c
o
nta
i
ne
d
20
10
i
m
ag
es
.
In
t
h
e
tes
ti
ng
s
t
ag
e
,
ea
c
h
ha
nd
group
c
o
ns
i
s
ted
of
40
2
i
m
ag
es
.
B
ot
h
tr
a
i
n
i
n
g
an
d
tes
t
i
n
g
s
tag
es
of
the
ML
P
ne
t
w
ork
att
em
pte
d
to
predi
c
t
a
c
l
ea
r
an
d
ea
s
i
l
y
r
ec
og
ni
z
e
d
pa
r
t
of
a
f
ac
i
al
i
m
ag
e,
w
h
ere
ea
c
h
i
n
di
v
i
du
al
pa
r
t
ha
s
i
ts
o
w
n
ML
P
.
A
l
l
tr
a
i
n
i
n
gs
ha
v
e
b
ee
n
ac
c
om
pl
i
s
he
d
b
y
t
h
e
a
l
g
orit
hm
of
the
S
c
a
l
ed
Co
nj
ug
a
te
G
r
ad
i
en
t
(
S
CG
)
,
w
hi
c
h
i
s
de
s
c
r
i
be
d
i
n
[29]
.
E
x
am
pl
es
of
tr
ai
ni
ng
c
urv
es
are
gi
v
e
n
i
n
F
i
g
ure
6.
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 16
93
-
6
93
0
T
E
L
KO
M
NIK
A
V
ol
.
17
,
No
.
6,
D
ec
em
be
r
20
19
:
31
1
0
-
31
1
9
3116
T
he
m
i
ni
m
u
m
Me
an
S
qu
are
E
r
r
or (
M
S
E
)
i
n
a
l
l
tra
i
n
i
ng
i
s
eq
ua
l
t
o
0.0
0
00
1
. Fur
t
he
r
m
ore, trai
ni
n
g
f
or
bo
th
t
he
r
i
gh
t
a
nd
l
ef
t
h
an
ds
s
uc
c
ee
d
ed
i
n
ac
h
i
e
v
i
ng
t
he
m
i
ni
m
u
m
err
or
wi
th
the
ap
pro
pria
t
e
nu
m
be
r
of
ep
oc
hs
.
(
a)
(
b)
F
i
gu
r
e
6.
A
n e
x
am
pl
e
of
a
t
r
ai
ni
ng
c
ur
v
e f
or: (
a) th
e ri
g
ht
ha
nd
m
ul
ti
-
s
pe
c
tr
a
l
i
m
ag
es
to
pre
di
c
t
the
i
nn
er part
of
a
f
ac
e i
m
a
ge
a
nd
(
b)
t
he
l
ef
t h
an
d m
ul
ti
-
s
pe
c
tr
al
i
m
ag
es
to
pre
di
c
t
th
e
ou
ter
p
art of
a
f
ac
e i
m
ag
e
F
or
the
r
e
gres
s
i
on
tes
t,
b
oth
tr
ai
ni
n
gs
a
tta
i
ne
d
45
de
gre
es
or
Regres
s
i
on
(
R
)
eq
u
al
to
1
be
t
ween
th
e
ML
P
ou
t
pu
ts
an
d
targ
ets
.
S
e
e
F
i
g
u
r
e
7.
T
o
a
na
l
y
s
e
F
i
g
ure
7,
th
e
no
n
-
l
i
n
ea
r
r
el
at
i
on
s
h
i
ps
wer
e
es
t
ab
l
i
s
he
d
be
t
ween
t
w
o
b
i
om
etri
c
s
,
whi
c
h
are
t
he
r
i
gh
t
an
d
l
ef
t
ha
n
d
wi
th
the
i
nn
er
an
d
ou
t
er
pa
r
t
of
a
f
ac
e
,
r
es
pe
c
ti
v
e
l
y
.
T
hi
s
r
el
at
i
o
ns
hi
p
i
s
th
e
ba
s
e
f
or
predi
c
t
i
ng
a
f
ul
l
f
ac
e
i
m
ag
e
f
r
o
m
i
np
uts
,
whi
c
h
ha
v
e
n
ev
er
be
en
s
e
en
b
ef
ore.
F
r
om
thi
s
po
i
nt
,
pred
i
c
ti
n
g
pa
r
ts
of
s
o
m
e
f
ac
e
i
m
ag
es
are
s
ho
wn
i
n
F
i
gu
r
e
s
8
(
a
an
d
b).
W
hi
l
s
t,
the
c
o
m
bi
na
ti
on
b
et
ween
e
ac
h
t
w
o
pa
r
ts
are d
i
s
pl
a
y
e
d
i
n
F
i
gu
r
e
8 (c
)
.
(
a)
(
b)
F
i
gu
r
e
7.
R
eg
r
es
s
i
o
n t
es
t f
or: (
a) a
r
i
gh
t
-
ha
nd
tr
ai
n
i
n
g
an
d (b)
a
l
ef
t
-
hand
tr
a
i
n
i
n
g
A
s
c
ou
l
d
be
ob
s
er
v
ed
,
v
er
y
c
l
ea
r
f
ac
e
pa
r
ts
c
a
n
b
e
c
o
m
bi
ne
d
i
f
n
e
w
k
no
w
n
v
ec
to
r
s
are
i
ntrod
uc
ed
t
o
the
tr
ai
n
ed
ML
P
s
.
C
on
v
ers
el
y
,
t
he
M
L
P
s
c
an
no
t
r
ec
o
gn
i
z
e
a
f
ac
e
i
m
ag
e
c
l
ea
r
l
y
when
n
e
w
un
k
no
w
n
v
ec
tor
s
are
tes
ted
.
F
i
gu
r
e
s
9
(
a,
b
an
d
c
)
ha
v
e
ex
am
pl
es
of
un
c
l
ea
r
f
ac
e
i
m
ag
es
.
A
s
m
en
ti
on
ed
be
f
ore,
us
i
ng
th
e
t
w
o
m
ai
n
f
ac
e
pa
r
ts
w
i
l
l
i
nc
r
ea
s
e
t
he
s
ec
uri
t
y
of
0
50
100
150
0
50
100
150
T
a
r
g
e
t
O
u
t
p
u
t
~
=
1
*
T
a
r
g
e
t
+
-
1
.
5
e
-
0
5
T
r
a
i
n
i
n
g
:
R=
1
D
at
a
F
it
Y
=
T
0
50
100
150
200
0
50
100
150
200
T
a
r
g
e
t
O
u
t
p
u
t
~
=
1
*
T
a
r
g
e
t
+
-
6
.
5
e
-
0
6
T
r
a
i
n
i
n
g
:
R=
1
D
at
a
F
it
Y
=
T
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
Rege
ne
r
at
i
ng
fa
c
e
i
m
ag
es
f
r
om
mu
l
t
i
-
s
pe
c
tr
al
pa
l
m
i
ma
ge
s
... (
Ra
i
d R
afi
O
ma
r
A
l
-
N
i
ma
)
3117
the
s
y
s
tem
.
Mo
r
eo
v
er,
pro
du
c
i
n
g
a
c
l
ea
r
f
ac
e
i
m
ag
e
wi
l
l
as
s
i
s
t
a
n
y
i
n
ex
pe
r
i
en
c
ed
i
nd
i
v
i
du
al
t
o
ea
s
i
l
y
au
th
ori
z
e
or i
d
en
t
i
f
y
t
he
c
erta
i
n p
ers
on
.
In
t
he
f
i
n
al
de
c
i
s
i
o
n
proc
es
s
,
the
i
m
ag
e
whi
c
h
i
s
the
m
os
t
c
l
ea
r
an
d
n
ea
r
es
t
to
the
s
pe
c
i
f
i
c
v
ec
tor
wi
l
l
be
c
on
s
i
d
ered
as
a
'tr
ue
'
a
nd
t
he
i
m
ag
e
whi
c
h
i
s
the
m
os
t
di
s
torte
d
an
d
f
urthes
t
f
r
o
m
the
s
pe
c
i
f
i
c
i
nf
orm
ati
on
wi
l
l
be
c
on
s
i
de
r
ed
as
a
'f
al
s
e'.
T
hu
s
,
t
w
o
t
y
p
es
of
c
l
as
s
i
f
i
c
ati
on
w
ere
de
t
erm
i
ne
d.
Her
ea
f
ter,
t
he
t
hr
ee
m
ai
n
pa
r
am
ete
r
s
are
c
al
c
ul
at
ed
:
F
al
s
e
A
c
c
ep
tan
c
e
R
ate
(
F
A
R)
,
F
al
s
e
Rej
ec
ti
on
Ra
te
(
F
RR)
an
d
E
E
R.
T
he
f
i
r
s
t
tw
o
p
aram
ete
r
s
are
c
al
c
ul
a
ted
ac
c
ordi
n
g t
o
th
e
f
ol
l
o
wi
ng
eq
u
at
i
on
s
:
=
ℎ
(
16
)
=
ℎ
(
17
)
the
E
E
R
pa
r
am
ete
r
i
s
c
al
c
ul
ate
d
f
r
om
the
eq
ua
l
i
t
y
be
t
w
e
en
F
A
R
an
d
F
RR
t
h
r
ou
gh
di
f
f
erent
thres
ho
l
d
v
a
l
ue
s
, s
e
e Fi
gu
r
e
10.
(
a)
(
b)
(
c
)
(
a)
(
b)
(
c
)
F
i
gu
r
e
8
. C
l
ea
r
f
ac
es
pa
r
ts
i
m
ag
es
:
(
a) i
nn
er f
ac
es
pa
r
ts
, (b)
ou
ter f
ac
es
pa
r
ts
an
d
(
c
)
a
c
o
m
bi
na
t
i
on
be
t
ween
the
t
wo p
arts
F
i
gu
r
e
9.
D
i
s
torte
d f
ac
es
pa
r
ts
i
m
ag
es
:
(
a) i
nn
er f
ac
es
pa
r
ts
, (b)
ou
ter f
ac
es
pa
r
ts
an
d
(
c
)
a
c
o
m
bi
na
t
i
on
be
t
ween
the
t
wo p
arts
F
i
gu
r
e
10
. F
A
R
v
ers
us
FR
R to
ac
q
ui
r
e
E
E
R
T
o
c
on
f
i
r
m
the
eff
i
c
i
en
c
y
of
ou
r
m
eth
od
,
c
om
pa
r
i
s
on
s
wi
th
oth
er
wor
k
i
nc
l
ud
i
ng
the
s
tat
e
-
of
-
the
-
art
ha
v
e
b
ee
n
es
ta
bl
i
s
he
d
i
n
T
ab
l
e
1
.
T
he
r
ea
s
on
of
s
el
ec
ti
n
g
t
he
t
w
o
s
tu
di
es
Evaluation Warning : The document was created with Spire.PDF for Python.
◼
IS
S
N: 16
93
-
6
93
0
T
E
L
KO
M
NIK
A
V
ol
.
17
,
No
.
6,
D
ec
em
be
r
20
19
:
31
1
0
-
31
1
9
3118
i
n
[7,
13
]
i
s
tha
t
bo
th
of
the
s
e
wor
k
s
c
on
c
en
tr
ate
d
on
r
eg
en
er
ati
ng
f
ul
l
de
t
ai
l
s
of
f
ac
e
i
m
ag
es
.
In
ad
di
t
i
o
n,
a
l
l
of
the
r
ep
orted
w
ork
s
i
n
T
ab
l
e
1
ha
v
e
u
s
e
d
the
O
R
L
f
ac
e
i
m
ag
es
da
tab
as
e.
F
r
om
thi
s
tab
l
e
i
t
c
an
b
e
s
ee
n
t
h
at
s
i
m
pl
e
s
tat
i
s
ti
c
s
w
ere
us
ed
w
i
t
h
A
N
N
tec
hn
i
q
ue
s
i
n
[7,
13
],
where
the
s
y
s
tem
s
tr
en
gth
l
e
v
e
l
of
the
prop
os
ed
s
y
s
t
em
s
is
hi
gh
.
In
t
hi
s
s
tud
y
,
t
wo
f
us
i
on
m
eth
od
s
an
d
two
m
ul
ti
-
s
pe
c
tr
a
l
h
an
d
i
m
ag
es
(
r
i
gh
t
a
nd
l
ef
t)
ha
v
e
be
e
n
em
pl
o
y
e
d
t
o
r
e
ge
n
erate
f
ul
l
f
ac
e
de
t
ai
l
s
.
T
he
r
ef
ore,
the
s
tr
en
gth
l
e
v
e
l
i
s
v
er
y
-
h
i
gh
as
s
po
of
i
ng
th
e
s
ug
g
es
ted
s
y
s
t
em
i
s
s
o
di
f
f
i
c
ul
t.
T
he
E
E
R
v
a
l
u
es
ha
v
e
b
ee
n
r
ec
or
de
d
to
10
%
f
o
r
[7]
a
nd
6
.43
%
an
d
2.
86
%
f
or
[13
].
In
th
i
s
prop
os
ed
ap
pr
oa
c
h,
the
E
E
R
v
al
ue
ha
s
b
ee
n
r
ec
orded
e
qu
al
t
o
(
1.
99
%)
.
S
o,
ou
r
wor
k
ap
pe
ars
to
h
av
e m
ore ac
c
urate res
ut
l
s
th
a
n o
t
he
r
s
tud
i
es
.
T
ab
l
e
1
. C
om
pa
r
i
s
on
s
of
V
ario
us
Fac
e
Reg
en
erati
on
Me
th
od
s
M
o
d
e
l
Fac
e
R
e
g
e
n
e
r
a
t
ing
M
e
t
h
o
d
s
S
y
s
t
e
m
S
t
r
e
n
g
t
h
L
e
v
e
l
EER
Al
-
N
i
m
a
e
t
a
l
.
[
7
]
S
i
m
p
le
s
t
a
t
i
s
t
ic
s
w
it
h
t
h
e
M
L
P
H
igh
10%
Al
-
N
i
m
a
e
t
a
l
.
[
1
3
]
S
i
m
p
le
s
t
a
t
i
s
t
ic
s
w
it
h
t
h
e
B
P
N
H
igh
6
.
4
3
%
S
i
m
p
le
s
t
a
t
i
s
t
ic
s
w
it
h
t
h
e
D
FN
H
igh
2
.
8
6
%
P
r
o
p
o
s
e
d
a
p
p
r
o
a
c
h
Tw
o
f
u
s
ion
m
e
t
h
o
d
s
w
it
h
t
h
e
M
L
P
V
e
r
y
-
H
igh
1
.
9
9
%
6.
Co
n
clus
ion
A
s
tr
ate
g
y
to
pre
di
c
t
f
ac
e
i
m
ag
e
wi
t
h
h
i
g
h
-
l
e
v
e
l
s
ec
ur
i
t
y
was
prod
uc
ed
i
n
th
i
s
p
ap
er,
where
two
no
n
-
l
i
n
ea
r
r
e
l
at
i
o
ns
hi
ps
ha
v
e
be
en
es
tab
l
i
s
he
d.
F
i
r
s
tl
y
,
a
r
e
l
at
i
on
s
hi
p
be
t
ween
the
r
i
g
ht
-
ha
nd
f
ea
tures
a
nd
the
m
i
dd
l
e
pa
r
t
of
the
f
ac
i
al
i
m
ag
e,
whi
c
h
i
n
g
en
er
a
l
c
on
ta
i
n
the
e
y
es
,
no
s
e
an
d
m
ou
th.
S
ec
o
nd
l
y
,
a
r
el
a
ti
on
s
h
i
p
be
t
w
ee
n
the
l
ef
t
-
ha
n
d
f
ea
tures
an
d
the
b
ou
nd
ar
y
p
art of
th
e
f
ac
e i
m
ag
e.
In
th
i
s
pa
pe
r
, t
wo
l
ev
e
l
s
of
f
us
i
on
are ex
am
i
ne
d;
th
e f
ea
ture
l
e
v
el
a
nd
the
s
c
ore
l
e
v
e
l
.
T
he
i
de
a
be
h
i
nd
us
i
ng
th
e
f
ea
tu
r
e
l
e
v
e
l
i
s
t
o
c
om
bi
n
e
a
nd
en
h
an
c
e
the
m
ul
ti
-
s
pe
c
tr
a
l
h
an
d
c
ha
r
ac
teri
s
ti
c
s
,
w
h
i
l
s
t,
th
e
s
c
ore
l
e
v
el
i
s
us
ed
o
n
t
he
f
ac
e
i
m
ag
es
t
o
c
ol
l
ec
t
an
d rec
o
ns
tr
uc
t c
l
ea
r
de
t
ai
l
s
of
a
f
ac
e.
T
he
s
ug
ge
s
ted
a
pp
r
o
ac
h
s
tr
uc
ture
c
on
f
i
r
m
ed
i
ts
eff
i
c
i
en
c
y
a
nd
r
ob
us
tne
s
s
.
T
he
pe
r
f
or
m
an
c
e
of
ov
eral
l
tec
hn
i
qu
e
was
be
nc
hm
ar
k
e
d
to
E
E
R
=
1.9
9%
du
r
i
ng
the
tes
ti
ng
s
t
ag
e,
where
f
ul
l
f
ac
e
de
t
ai
l
s
wer
e
r
ec
on
s
tr
uc
ted
.
In
a
dd
i
t
i
o
n,
the
propos
e
d
s
y
s
tem
i
nc
r
ea
s
es
the
a
nti
-
s
p
oo
f
i
ng
,
s
tr
en
gt
h
an
d
s
ec
ur
i
t
y
l
ev
el
s
.
T
hi
s
i
s
be
c
au
s
e
t
wo
m
ul
ti
-
s
pe
c
tr
a
l
i
m
ag
es
of
the
two h
an
ds
(
l
ef
t a
nd
r
i
gh
t)
are r
eq
u
i
r
ed
to
r
e
ge
n
erate
al
l
f
ac
e d
eta
i
l
s
.
A
c
kno
w
ledg
men
t
“
P
orti
o
ns
of
the
r
es
ea
r
c
h
i
n
thi
s
p
ap
er
us
e
t
he
C
A
S
I
A
-
MS
-
P
a
l
m
prin
t
V
1
c
ol
l
ec
t
ed
b
y
the
Ch
i
n
es
e
A
c
ad
em
y
of
S
c
i
en
c
es
'
Ins
ti
t
ute
o
f
A
uto
m
ati
on
(
CA
S
IA
)
”
.
In
ad
d
i
t
i
on
,
an
ac
k
no
wl
e
dg
m
en
t i
s
g
i
v
e
n t
o
A
T
&
T
l
ab
orator
i
es
of
C
am
brid
ge
Un
i
v
ers
i
t
y
.
Ref
er
en
ce
s
[1
]
L
i
u
Y
F,
L
i
n
CY
,
G
u
o
J
M
.
Im
p
a
c
t
o
f
t
h
e
L
i
p
s
f
o
r
B
i
o
m
e
tri
c
s
.
I
EEE
Tra
n
s
a
c
ti
o
n
s
o
n
Im
a
g
e
P
ro
c
e
s
s
i
n
g
.
2012
;
21
:
3092
-
3
1
0
1
.
[2
]
Ö
z
k
a
y
a
N
,
Sa
ğ
i
ro
ğ
l
u
Ş
.
G
e
n
e
ra
ti
n
g
O
n
e
Bi
o
m
e
tri
c
Fe
a
t
u
re
fro
m
An
o
t
h
e
r:
Fa
c
e
s
fro
m
Fi
n
g
e
rp
ri
n
ts
.
Se
n
s
o
r
s
.
2
0
1
0
;
1
0
(
5
)
:
4
2
0
6
-
4
2
3
7
.
[3
]
Sa
ğ
i
ro
ğ
l
u
Ş
,
Ö
z
k
a
y
a
N
.
An
In
t
e
l
l
i
g
e
n
t
a
n
d
Au
to
m
a
t
i
c
Ey
e
G
e
n
e
ra
ti
o
n
S
y
s
t
e
m
fro
m
O
n
l
y
Fi
n
g
e
rp
r
i
n
ts
.
Pro
c
e
e
d
i
n
g
s
o
f
In
fo
r
m
a
t
i
o
n
Se
c
u
ri
ty
a
n
d
Cry
p
to
l
o
g
y
Con
fe
re
n
c
e
w
i
th
In
te
r
n
a
ti
o
n
a
l
Pa
rti
c
i
p
a
ti
o
n
.
An
k
a
ra
.
2
0
0
8
:
2
3
1
-
2
3
6
.
[4
]
Ö
z
k
a
y
a
N,
Sa
ğ
i
r
o
ğ
l
u
Ş.
In
te
l
l
i
g
e
n
t
fa
c
e
b
o
rd
e
r
g
e
n
e
r
a
ti
o
n
s
y
s
t
e
m
fro
m
fi
n
g
e
rp
r
i
n
t
s
.
2
0
0
8
IEE
E
In
te
rn
a
ti
o
n
a
l
Con
fe
r
e
n
c
e
o
n
Fu
z
z
y
Sy
s
te
m
s
(IEEE
W
o
rl
d
Con
g
re
s
s
o
n
Co
m
p
u
ta
ti
o
n
a
l
I
n
te
l
l
i
g
e
n
c
e
.
2008
:
2
1
6
9
-
2176
.
[5
]
Sa
ğ
i
ro
ğ
l
u
Ş
,
Ö
z
k
a
y
a
N
.
A
n
i
n
t
e
l
l
i
g
e
n
t
fa
c
e
fe
a
tu
re
s
g
e
n
e
ra
t
i
o
n
s
y
s
t
e
m
fr
o
m
fi
n
g
e
rp
ri
n
ts
.
T
u
rk
J
El
e
c
En
g
&
Com
p
Sc
i
.
2
0
0
9
;
17
(
2
)
:
1
8
3
-
2
0
3
.
[6
]
Chi
tra
v
a
n
s
h
i
A,
S
i
n
g
h
A,
So
l
a
n
k
i
S,
K
u
m
a
r
A,
Sh
a
rm
a
R
.
G
e
n
e
ra
t
i
n
g
Fa
c
e
Fe
a
tu
re
s
b
y
P
a
l
m
Sc
a
n
:
Rec
o
g
n
i
z
i
n
g
a
Pe
rs
o
n
U
n
i
q
u
e
l
y
.
In
te
r
n
a
t
i
o
n
a
l
J
o
u
r
n
a
l
o
f
Ap
p
l
i
e
d
In
fo
rm
a
ti
o
n
S
y
s
t
e
m
s
(
IJ
AIS)
.
2
0
1
2
;
2
(
5
)
:
3
6
-
41
.
[7
]
Al
-
Nim
a
RR
,
Dl
a
y
SS,
W
o
o
W
L
.
A
New
Ap
p
ro
a
c
h
to
Pre
d
i
c
ti
n
g
Ph
y
s
i
c
a
l
B
i
o
m
e
tri
c
s
fro
m
B
e
h
a
v
i
o
u
r
a
l
Bi
o
m
e
tri
c
s
.
W
o
r
l
d
Ac
a
d
e
m
y
o
f
Sc
i
e
n
c
e
,
En
g
i
n
e
e
ri
n
g
a
n
d
Te
c
h
n
o
l
o
g
y
,
In
te
r
n
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
Com
p
u
te
r,
El
e
c
tr
i
c
a
l
,
Au
t
o
m
a
ti
o
n
,
Con
tr
o
l
a
n
d
I
n
fo
rm
a
t
i
o
n
En
g
i
n
e
e
r
i
n
g
.
2
0
1
4
;
8
(
1
1
)
:
1
9
9
6
-
2
0
0
1
.
Evaluation Warning : The document was created with Spire.PDF for Python.
T
E
L
KO
M
NIK
A
IS
S
N: 1
69
3
-
6
93
0
◼
Rege
ne
r
at
i
ng
fa
c
e
i
m
ag
es
f
r
om
mu
l
t
i
-
s
pe
c
tr
al
pa
l
m
i
ma
ge
s
... (
Ra
i
d R
afi
O
ma
r
A
l
-
N
i
ma
)
3119
[8
]
Y
a
n
g
L
,
L
i
C,
Han
J
,
Ch
e
n
C
,
Y
e
Q
,
Zh
a
n
g
B
,
Cao
X
,
L
i
u
W
.
Im
a
g
e
Rec
o
n
s
tr
u
c
t
i
o
n
v
i
a
M
a
n
i
fo
l
d
Con
s
tra
i
n
e
d
Con
v
o
l
u
ti
o
n
a
l
Sp
a
rs
e
C
o
d
i
n
g
fo
r
I
m
a
g
e
Se
t
s
.
I
EEE
J
o
u
r
n
a
l
o
f
S
e
l
e
c
te
d
To
p
i
c
s
i
n
S
i
g
n
a
l
Pro
c
e
s
s
i
n
g
.
2017
;
11
(
7
)
:
1
0
7
2
-
81
.
[9
]
L
u
Y
,
T
a
i
Y
W
,
T
a
n
g
CK
.
Con
d
i
t
i
o
n
a
l
c
y
c
l
e
g
a
n
fo
r
a
t
tri
b
u
te
g
u
i
d
e
d
fa
c
e
i
m
a
g
e
g
e
n
e
ra
t
i
o
n
.
a
r
Xi
v
p
re
p
ri
n
t
a
rXi
v
:
1
7
0
5
.0
9
9
6
6
.
2
0
1
7
.
[1
0
]
J
i
a
n
g
L
,
Zh
a
n
g
J
,
Den
g
B,
L
i
H,
L
i
u
L
.
L
i
u
.
3
D
Fa
c
e
Re
c
o
n
s
tru
c
ti
o
n
W
i
t
h
G
e
o
m
e
try
Det
a
i
l
s
Fr
o
m
a
Si
n
g
l
e
I
m
a
g
e
.
IEEE
Tra
n
s
a
c
ti
o
n
s
o
n
I
m
a
g
e
Pr
o
c
e
s
s
i
n
g
.
2
0
1
8
;
27
(
10
)
:
4
7
5
6
-
70
.
[1
1
]
L
i
X
,
Y
a
n
g
X
,
Su
H,
Z
h
o
u
Q
,
Zh
e
n
g
S
.
Rec
o
g
n
i
z
i
n
g
F
a
c
i
a
l
Sk
e
tc
h
e
s
b
y
G
e
n
e
ra
t
i
n
g
P
h
o
to
r
e
a
l
i
s
ti
c
Fa
c
e
s
G
u
i
d
e
d
b
y
De
s
c
r
i
p
ti
v
e
A
tt
ri
b
u
t
e
s
.
IEEE
A
c
c
e
s
s
.
2
0
1
8
;
6
:
7
7
5
6
8
-
80
.
[1
2
]
M
a
i
G, Ca
o
K,
Y
u
e
n
PC,
J
a
i
n
AK
.
O
n
t
h
e
Re
c
o
n
s
tru
c
ti
o
n
o
f
Fa
c
e
Im
a
g
e
s
f
ro
m
De
e
p
F
a
c
e
T
e
m
p
l
a
te
s
.
IEEE
Tra
n
s
a
c
t
i
o
n
s
o
n
Pa
tt
e
rn
An
a
l
y
s
i
s
a
n
d
M
a
c
h
i
n
e
I
n
te
l
l
i
g
e
n
c
e
.
2
0
1
9
;
41
(
5
)
:
1
1
8
8
-
2
0
2
.
[1
3
]
Al
-
Nim
a
RRO
,
Ab
d
u
l
r
a
h
e
e
m
FH
,
Al
-
Rid
h
a
MY
.
Us
i
n
g
Ha
n
d
-
D
o
rs
a
l
Im
a
g
e
s
to
Rep
ro
d
u
c
e
Fa
c
e
Im
a
g
e
s
b
y
Ap
p
l
y
i
n
g
Ba
c
k
p
ro
p
a
g
a
ti
o
n
a
n
d
C
a
s
c
a
d
e
-
Fo
rward
Neu
ra
l
Net
work
s
.
2
n
d
In
te
r
n
a
ti
o
n
a
l
C
o
n
fe
re
n
c
e
o
n
El
e
c
tri
c
a
l
,
Co
m
m
u
n
i
c
a
ti
o
n
,
Com
p
u
t
e
r,
Po
w
e
r
a
n
d
Con
t
ro
l
En
g
i
n
e
e
ri
n
g
ICECC
PCE19
.
M
o
s
u
l
.
2
0
1
9
.
[1
4
]
Pa
u
l
C,
Cou
rtn
e
y
R,
Sa
n
s
o
n
-
F
i
s
h
e
r
R,
Car
e
y
M
,
Hil
l
D,
Si
m
m
o
n
s
J
,
Ro
s
e
S
.
A
ra
n
d
o
m
i
z
e
d
c
o
n
tr
o
l
l
e
d
tri
a
l
o
f
t
h
e
e
ff
e
c
t
i
v
e
n
e
s
s
o
f
a
p
re
-
re
c
r
u
i
t
m
e
n
t
p
r
i
m
e
r
l
e
tt
e
r
to
i
n
c
re
a
s
e
p
a
rti
c
i
p
a
ti
o
n
i
n
a
s
tu
d
y
o
f
c
o
l
o
r
e
c
t
a
l
s
c
re
e
n
i
n
g
a
n
d
s
u
rv
e
i
l
l
a
n
c
e
.
BM
C
M
e
d
i
c
a
l
Re
s
e
a
r
c
h
M
e
th
o
d
o
l
o
g
y
.
2
0
1
4
;
1
4
(
4
4
)
.
[1
5
]
He
Z
.
Hi
e
ra
r
c
h
i
c
a
l
Col
o
r
a
n
t
-
Ba
s
e
d
Dir
e
c
t
B
i
n
a
ry
Se
a
rc
h
Hal
ft
o
n
i
n
g
.
IEEE
Tr
a
n
s
a
c
ti
o
n
s
o
n
Im
a
g
e
Pro
c
e
s
s
i
n
g
.
2010
;
19
(
7
)
:
1
8
2
4
-
1836
.
[1
6
]
Ba
i
M
R,
Kri
s
h
n
a
VV,
Sr
e
e
De
v
i
J
.
A
n
e
w
M
o
rp
h
o
l
o
g
i
c
a
l
A
p
p
ro
a
c
h
fo
r
No
i
s
e
Rem
o
v
a
l
c
u
m
Ed
g
e
Det
e
c
t
i
o
n
.
IJ
CSI
I
n
te
rn
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
Com
p
u
te
r
Sc
i
e
n
c
e
I
s
s
u
e
s
.
2
0
1
0
;
7
(6
)
:1
8
7
.
[1
7
]
J
a
m
i
l
N,
Se
m
b
o
k
T
M
,
Ba
k
a
r
ZA
.
Noi
s
e
re
m
o
v
a
l
a
n
d
e
n
h
a
n
c
e
m
e
n
t
o
f
b
i
n
a
r
y
i
m
a
g
e
s
u
s
i
n
g
m
o
rp
h
o
l
o
g
i
c
a
l
o
p
e
r
a
ti
o
n
s
.
In
t
e
rn
a
ti
o
n
a
l
Sy
m
p
o
s
i
u
m
o
n
In
fo
r
m
a
ti
o
n
T
e
c
h
n
o
l
o
g
y
IT
Si
m
2
0
0
8
.
2
0
0
8
;
4
:
1
-
6.
[1
8
]
T
h
e
M
a
th
W
o
rk
s
I
n
c
.
Im
a
g
e
Pr
o
c
e
s
s
i
n
g
T
o
o
l
b
o
x
Fo
r Us
e
wi
t
h
M
ATL
AB
,
Ve
r.3
e
d
.,
1
9
9
3
-
2
0
0
1
.
[1
9
]
Ke
k
re
HB
,
M
u
k
h
e
rj
e
e
P
,
W
a
d
h
w
a
S
.
I
m
a
g
e
R
e
tri
e
v
a
l
w
i
th
Sh
a
p
e
Fe
a
tu
re
s
Ex
tra
c
te
d
u
s
i
n
g
M
o
rp
h
o
l
o
g
i
c
a
l
O
p
e
ra
t
o
rs
w
i
th
BT
C
.
In
te
r
n
a
ti
o
n
a
l
J
o
u
r
n
a
l
o
f
Com
p
u
t
e
r
A
p
p
l
i
c
a
ti
o
n
s
.
2
0
1
0
;
12
(3
)
:
0975
–
8
8
8
7
.
[2
0
]
Po
l
e
s
e
l
A,
Ra
m
p
o
n
i
G
,
M
a
th
e
w
s
VJ
.
Im
a
g
e
e
n
h
a
n
c
e
m
e
n
t
v
i
a
a
d
a
p
ti
v
e
u
n
s
h
a
rp
m
a
s
k
i
n
g
.
IE
E
E
tra
n
s
a
c
t
i
o
n
s
o
n
i
m
a
g
e
p
ro
c
e
s
s
i
n
g
.
2000
;
9
(3
)
:
505
-
5
1
0
.
[2
1
]
Al
-
Nim
a
RR
O
.
Si
g
n
a
l
p
r
o
c
e
s
s
i
n
g
a
n
d
m
a
c
h
i
n
e
l
e
a
rn
i
n
g
te
c
h
n
i
q
u
e
s
fo
r
h
u
m
a
n
v
e
ri
f
i
c
a
ti
o
n
b
a
s
e
d
o
n
fi
n
g
e
r
t
e
x
tu
re
s
.
Ph
D t
h
e
s
i
s
.
Sc
h
o
o
l
o
f
En
g
i
n
e
e
ri
n
g
,
New
c
a
s
t
l
e
Un
i
v
e
rs
i
ty
.
2
0
1
7
.
[2
2
]
Sa
l
e
D,
So
n
a
re
P,
J
o
s
h
i
M
A
.
PCA
Ba
s
e
d
Im
a
g
e
Fu
s
i
o
n
fo
r
M
u
l
ti
s
p
e
c
tr
a
l
Pa
l
m
En
h
a
n
c
e
m
e
n
t
.
In
te
rn
a
ti
o
n
a
l
J
o
u
r
n
a
l
o
f
Ad
v
a
n
c
e
d
Re
s
e
a
r
c
h
i
n
El
e
c
tr
i
c
a
l
,
El
e
c
tro
n
i
c
s
a
n
d
In
s
tru
m
e
n
t
a
ti
o
n
E
n
g
i
n
e
e
ri
n
g
.
2014
;
3
z
2
):
7
5
0
1
-
7
5
0
8
.
[2
3
]
Bh
o
k
a
r
e
R
,
Sa
l
e
D
,
J
o
s
h
i
MA
,
G
a
i
k
w
a
d
M
S.
M
u
l
ti
s
p
e
c
tra
l
Pa
l
m
I
m
a
g
e
Fu
s
i
o
n
:
A
Cri
ti
c
a
l
Rev
i
e
w
.
In
te
rn
a
ti
o
n
a
l
J
o
u
rn
a
l
o
f
A
d
v
a
n
c
e
d
Re
s
e
a
rc
h
i
n
C
o
m
p
u
te
r
E
n
g
i
n
e
e
r
i
n
g
&
Te
c
h
n
o
l
o
g
y
(
IJ
AR
CET)
.
2
0
1
3
;
2
(6
):
2
1
5
9
-
21
64
.
[2
4
]
M
e
rk
l
e
J
,
Ke
v
e
n
a
a
r
T
,
Ko
rte
U
.
M
u
l
ti
-
m
o
d
a
l
a
n
d
m
u
l
t
i
-
i
n
s
ta
n
c
e
fu
s
i
o
n
fo
r
b
i
o
m
e
tri
c
c
r
y
p
t
o
s
y
s
t
e
m
s
.
2
0
1
2
BIO
SIG
-
Pro
c
e
e
d
i
n
g
s
o
f
t
h
e
In
te
rn
a
ti
o
n
a
l
Co
n
fe
re
n
c
e
o
f
th
e
Bi
o
m
e
tr
i
c
s
Sp
e
c
i
a
l
In
te
r
e
s
t
G
ro
u
p
(BIOS
IG)
.
2012
:
1
-
6.
[2
5
]
Rei
n
h
a
r
d
t
K,
W
e
i
s
s
S,
Ros
e
n
b
a
u
e
r
J
,
G
ä
rtn
e
r
J
,
Vo
n
Kri
e
s
R
.
M
u
l
ti
p
l
e
s
c
l
e
ro
s
i
s
i
n
c
h
i
l
d
re
n
a
n
d
a
d
o
l
e
s
c
e
n
t
s
:
i
n
c
i
d
e
n
c
e
a
n
d
c
l
i
n
i
c
a
l
p
i
c
tu
re
-
n
e
w
i
n
s
i
g
h
t
s
fr
o
m
th
e
n
a
ti
o
n
w
i
d
e
G
e
rm
a
n
s
u
rv
e
i
l
l
a
n
c
e
(2
0
0
9
-
2
0
1
1
)
.
Eu
ro
p
e
a
n
J
o
u
rn
a
l
o
f
N
e
u
ro
l
o
g
y
.
2
0
1
4
;
21
(4
)
:
6
5
4
-
9.
[2
6
]
Fa
u
s
e
tt
L.
Fu
n
d
a
m
e
n
t
a
l
o
f
Ne
u
ra
l
Ne
tw
o
rk
s
,
Ar
c
h
i
t
e
c
t
u
re
s
,
Al
g
o
ri
t
h
m
s
a
n
d
Ap
p
l
i
c
a
t
i
o
n
s
.
Pr
e
n
ti
c
e
Ha
l
l
In
t.
Sn
c
.
1
9
9
4
.
[2
7
]
L
i
J
,
Y
in
G
,
Z
h
a
n
g
G
.
Ev
a
l
u
a
t
i
o
n
o
f
to
b
a
c
c
o
m
i
x
i
n
g
u
n
i
fo
rm
i
t
y
b
a
s
e
d
o
n
c
h
e
m
i
c
a
l
c
o
m
p
o
s
i
ti
o
n
.
2
0
1
2
3
1
s
t
C
h
i
n
e
s
e
Co
n
tro
l
Co
n
f
e
re
n
c
e
(CCC
)
.
2
0
1
2
:
7
5
5
2
-
7
5
5
5
.
[2
8
]
L
u
z
AG
,
O
s
i
s
M
J
,
Rib
e
i
ro
M
,
Cec
a
t
ti
J
G
,
Am
a
ra
l
E
.
Pe
r
s
p
e
c
ti
v
e
s
o
f
p
ro
f
e
s
s
i
o
n
a
l
s
p
a
rti
c
i
p
a
ti
n
g
i
n
th
e
Bra
z
i
l
i
a
n
Ne
tw
o
rk
fo
r
th
e
S
u
rv
e
i
l
l
a
n
c
e
o
f
Se
v
e
re
M
a
te
rn
a
l
M
o
rb
i
d
i
ty
re
g
a
rd
i
n
g
t
h
e
i
m
p
l
e
m
e
n
t
a
ti
o
n
o
f
ro
u
t
i
n
e
s
u
rv
e
i
l
l
a
n
c
e
:
a
q
u
a
l
i
t
a
ti
v
e
s
t
u
d
y
.
Rep
ro
d
u
c
ti
v
e
He
a
l
th
.
2
0
1
4
;
11
(1
)
:
29.
[2
9
]
M
øl
l
e
r M
F
.
A
s
c
a
l
e
d
c
o
n
j
u
g
a
te
g
ra
d
i
e
n
t
a
l
g
o
r
i
th
m
fo
r
fa
s
t
s
u
p
e
rv
i
s
e
d
l
e
a
r
n
i
n
g
.
Neu
ra
l
Net
wo
rk
s
.
1993
;
6
(4
)
:
5
2
5
-
5
3
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.