I
n
t
e
r
n
at
ion
al
Jou
r
n
al
of
E
lec
t
r
ical
an
d
Com
p
u
t
e
r
E
n
gin
e
e
r
in
g
(
I
JE
CE
)
Vol.
16
,
No.
5
,
Oc
tober
20
26
,
pp
.
2622
~
2640
I
S
S
N:
2088
-
8708
,
DO
I
:
10
.
11591/i
jec
e
.
v
16
i
5
.
pp
2
622
-
2640
2622
Jou
r
n
al
h
omepage
:
ht
tp:
//
ij
e
c
e
.
iaes
c
or
e
.
c
om
C
on
t
o
u
r
-
g
u
id
e
d
c
on
v
e
xi
t
y d
e
f
e
c
t
g
e
om
e
t
r
y f
or
p
r
e
c
is
e
R
OI
e
xt
r
a
c
t
io
n
i
n
d
o
r
sal
h
an
d
ve
i
n
r
e
c
og
n
ition
Hab
ib
Kad
e
m
,
M
os
t
e
f
a
M
e
r
ah
D
e
pa
r
tm
e
nt
of
E
le
c
tr
ic
a
l
E
ngi
ne
e
r
in
g a
nd L
a
bor
a
to
r
y o
f
S
ig
na
l
s
a
nd S
ys
te
ms
(
L
S
S
)
, F
a
c
ul
ty
of
S
c
ie
nc
e
s
a
nd T
e
c
hnol
ogy,
U
ni
ve
r
s
it
y A
bde
lh
a
mi
d I
bn B
a
di
s
of
M
o
s
ta
ga
ne
m, M
o
s
ta
ga
n
e
m, Alge
r
ia
Ar
t
icle
I
n
f
o
AB
S
T
RA
CT
A
r
ti
c
le
h
is
tor
y
:
R
e
c
e
ived
J
a
n
31,
2026
R
e
vis
e
d
Apr
28,
2026
Ac
c
e
pted
J
ul
22,
2026
D
o
r
s
al
h
an
d
v
e
i
n
b
i
o
met
r
i
cs
d
ep
e
n
d
s
t
r
o
n
g
l
y
o
n
accu
r
at
e
reg
i
o
n
-
of
-
i
n
t
eres
t
(RO
I)
ex
t
ract
i
o
n
,
y
e
t
c
o
n
v
en
t
i
o
n
a
l
co
n
v
e
x
i
t
y
-
d
efect
-
b
a
s
ed
me
t
h
o
d
s
rema
i
n
v
u
l
n
era
b
l
e
t
o
co
n
t
o
u
r
n
o
i
s
e,
s
k
i
n
t
ex
t
u
re,
an
d
s
h
a
d
o
w
-
i
n
d
u
ce
d
art
i
fact
s
.
T
h
i
s
p
ap
er
i
n
t
r
o
d
u
ces
co
n
t
o
u
r
-
g
u
i
d
e
d
co
n
v
ex
i
t
y
d
efect
g
e
o
met
ry
(
CG
-
CD
G
),
a
g
eo
me
t
ry
-
d
ri
v
en
R
O
I
ex
t
ract
i
o
n
framew
o
rk
i
n
w
h
i
ch
a
can
d
i
d
a
t
e
co
n
v
e
x
i
t
y
d
efect
i
s
accep
t
ed
as
a
n
an
a
t
o
m
i
cal
v
a
l
l
e
y
o
n
l
y
i
f
i
t
j
o
i
n
t
l
y
s
at
i
s
f
i
es
t
w
o
i
n
d
ep
e
n
d
e
n
t
cri
t
eri
a
:
ch
o
rd
-
c
o
n
t
o
u
r
t
o
p
o
l
o
g
i
ca
l
co
n
s
i
s
t
e
n
c
y
,
en
s
u
r
i
n
g
an
at
o
mi
ca
l
rel
e
v
an
ce,
a
n
d
a
m
i
n
i
mu
m
e
n
cl
o
s
e
d
-
area
t
h
re
s
h
o
l
d
,
en
s
u
ri
n
g
p
h
y
s
i
cal
p
l
au
s
i
b
i
l
i
t
y
.
U
n
l
i
k
e
p
ri
o
r
co
n
v
ex
i
t
y
-
d
efect
-
b
a
s
ed
met
h
o
d
s
,
w
h
i
ch
accep
t
a
ca
n
d
i
d
a
t
e
v
al
l
ey
o
n
t
h
e
b
a
s
i
s
o
f
a
s
i
n
g
l
e
d
e
p
t
h
-
o
r
a
n
g
l
e
-
b
a
s
ed
cri
t
er
i
o
n
ev
a
l
u
a
t
ed
i
n
i
s
o
l
at
i
o
n
,
CG
-
CD
G
ap
p
l
i
e
s
t
h
i
s
j
o
i
n
t
p
l
a
u
s
i
b
i
l
i
t
y
t
e
s
t
b
efo
re
d
ef
i
n
i
n
g
t
h
e
reg
i
o
n
o
f
i
n
t
eres
t
,
a
s
q
u
are
RO
I
c
en
t
ere
d
o
n
t
h
e
h
a
n
d
cen
t
r
o
i
d
an
d
a
l
i
g
n
e
d
w
i
t
h
t
h
e
d
o
m
i
n
a
n
t
v
al
l
ey
o
ri
e
n
t
a
t
i
o
n
.
E
x
p
er
i
men
t
s
o
n
a
p
u
b
l
i
c
d
o
rs
a
l
h
a
n
d
v
ei
n
d
a
t
ab
a
s
e
o
f
1
,
0
2
4
i
mag
e
s
fro
m
1
3
8
s
u
b
j
ect
s
s
h
o
w
t
h
a
t
,
co
mb
i
n
e
d
w
i
t
h
h
i
s
t
o
g
ram
o
f
o
r
i
en
t
ed
g
ra
d
i
e
n
t
s
(
H
O
G
)
d
es
cr
i
p
t
o
r
s
an
d
Ci
t
y
B
l
o
c
k
d
i
s
t
a
n
ce,
CG
-
CD
G
ach
i
ev
es
an
eq
u
al
err
o
r
rat
e
(E
E
R)
o
f
0
.
0
2
6
6
,
w
i
t
h
a
9
5
%
b
o
o
t
s
t
ra
p
co
n
f
i
d
e
n
ce
i
n
t
er
v
al
o
f
[0
.
0
2
4
1
,
0
.
0
2
8
9
]
o
v
er
t
h
e
mat
ch
i
n
g
s
co
re
s
(B
=
1
,
0
0
0
res
amp
l
es
),
t
h
e
n
arr
o
w
e
s
t
i
n
t
er
v
al
amo
n
g
t
h
e
fo
u
r
RO
I
e
x
t
rac
t
i
o
n
me
t
h
o
d
s
co
m
p
ared
u
n
d
er
t
h
e
p
ri
ma
ry
ev
al
u
at
i
o
n
s
e
t
t
i
n
g
.
A
p
o
s
t
-
h
o
c
W
i
l
c
o
x
o
n
s
i
g
n
e
d
-
ran
k
t
es
t
w
i
t
h
Bo
n
ferr
o
n
i
co
rrect
i
o
n
c
o
n
f
i
rms
t
h
a
t
t
h
i
s
i
m
p
ro
v
emen
t
i
s
s
t
at
i
s
t
i
ca
l
l
y
s
i
g
n
i
f
i
ca
n
t
o
v
er
a
l
l
co
mp
e
t
i
n
g
me
t
h
o
d
s
(p
<
0
.
0
0
1
).
O
n
t
h
i
s
d
at
a
s
et
,
g
e
o
met
r
i
c
v
a
l
i
d
at
i
o
n
s
u
b
s
t
a
n
t
i
al
l
y
red
u
ces
i
n
t
ra
-
cl
as
s
v
ari
an
ce
an
d
y
i
e
l
d
s
s
t
at
i
s
t
i
ca
l
l
y
v
al
i
d
a
t
e
d
p
erfo
rman
ce
w
i
t
h
o
u
t
req
u
i
ri
n
g
an
y
t
ra
i
n
i
n
g
d
at
a,
mak
i
n
g
CG
-
CD
G
a
s
t
ro
n
g
can
d
i
d
a
t
e
b
o
t
h
as
a
s
t
a
n
d
a
l
o
n
e,
t
rai
n
i
n
g
-
free
b
i
o
met
r
i
c
s
o
l
u
t
i
o
n
an
d
as
a
n
an
at
o
m
i
cal
l
y
g
r
o
u
n
d
ed
p
rep
r
o
ces
s
i
n
g
s
t
ag
e
f
o
r
h
y
b
ri
d
d
ee
p
-
l
ear
n
i
n
g
p
i
p
el
i
n
e
s
.
K
e
y
w
o
r
d
s
:
B
iom
e
tr
ics
Dor
s
a
l
ha
nd
ve
in
r
e
c
ognit
ion
R
OI
e
xtr
a
c
ti
on
C
onve
xit
y
de
f
e
c
t
Ge
ometr
ic
va
li
da
ti
on
H
is
togr
a
m
of
o
r
iente
d
gr
a
dients
E
qua
l
e
r
r
o
r
r
a
te
Th
i
s
i
s
a
n
o
p
en
a
c
ces
s
a
r
t
i
c
l
e
u
n
d
e
r
t
h
e
CC
B
Y
-
SA
l
i
ce
n
s
e
.
C
or
r
e
s
pon
din
g
A
u
th
or
:
Ha
bib
Ka
de
m
De
pa
r
tm
e
nt
of
E
lec
tr
ica
l
E
nginee
r
ing
a
nd
L
a
bor
a
tor
y
of
S
ignals
a
nd
S
ys
tems
(
L
S
S
)
,
F
a
c
ult
y
of
S
c
ie
nc
e
s
a
nd
T
e
c
hnology,
Unive
r
s
it
y
Abde
lhamid
I
bn
B
a
di
s
of
M
os
taga
ne
m
R
oute
B
e
lahc
e
ne
.
B
p
277,
M
os
taga
ne
m,
Alge
r
ia
E
mail:
ha
bib
.
ka
de
m@uni
v
-
mos
ta.
dz
1.
I
NT
RODU
C
T
I
ON
B
iom
e
tr
ic
r
e
c
ognit
ion
s
ys
tems
a
r
e
c
r
it
ica
l
c
ompo
ne
nts
of
s
e
c
ur
e
identi
f
ica
ti
on
in
f
r
a
s
tr
uc
tur
e
s
,
with
dor
s
a
l
ha
nd
ve
in
r
e
c
ognit
ion
of
f
e
r
ing
d
is
ti
nc
t
a
dva
ntage
s
s
uc
h
a
s
high
dis
ti
nc
ti
ve
ne
s
s
a
nd
s
poof
ing
r
e
s
is
tanc
e
[
1]
.
How
e
ve
r
,
unli
ke
pa
lm
-
ve
in
r
e
c
ognit
ion
[
2]
,
dor
s
a
l
ve
in
pa
tt
e
r
ns
f
a
c
e
s
e
ve
r
e
c
ha
ll
e
nge
s
in
r
e
gion
-
of
-
int
e
r
e
s
t
(
R
OI
)
e
xtr
a
c
ti
on
due
to
low
c
ont
r
a
s
t
a
nd
a
lac
k
of
c
lea
r
a
na
tom
ica
l
r
e
f
e
r
e
nc
e
s
.
As
e
mphas
ize
d
by
J
a
in
e
t
al.
[
3
]
,
s
ys
tem
pe
r
f
or
manc
e
de
pe
nds
he
a
vil
y
on
r
obus
t
p
r
e
pr
oc
e
s
s
ing
to
mi
ti
ga
te
s
e
ns
or
nois
e
a
nd
int
r
a
-
c
las
s
va
r
iations
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
C
ontour
-
guided
c
onv
e
x
it
y
de
fec
t
ge
ome
tr
y
for
pr
e
c
is
e
R
OI
e
x
tr
ac
ti
on
in
…
(
Habib
K
ade
m
)
2623
I
n
dor
s
a
l
ha
nd
ve
in
biom
e
tr
ics
,
r
e
c
ognit
ion
s
tr
ug
gles
mor
e
with
int
r
a
-
c
las
s
va
r
iabili
ty
than
f
e
a
tur
e
qua
li
ty.
C
ha
nge
s
in
ha
nd
pos
e
,
wr
is
t
a
ngle,
f
ing
e
r
pos
it
ion,
a
nd
c
a
ptur
e
c
ondit
ions
de
s
tabili
z
e
a
na
tom
ica
l
loca
li
z
a
ti
on,
dr
ivi
ng
up
f
a
ls
e
r
e
jec
ti
on
r
a
te
(
F
R
R
)
.
C
ons
is
tent
R
OI
e
xtr
a
c
ti
on
r
e
mains
a
c
r
it
ica
l
pr
e
r
e
quis
it
e
f
or
r
e
li
a
ble
ve
in
-
ba
s
e
d
identif
ica
ti
on
s
ys
tems
.
P
r
e
c
is
e
R
OI
loca
li
z
a
ti
on
is
pa
r
a
mount
,
a
s
e
xt
r
a
c
ti
on
e
r
r
o
r
s
p
r
opa
ga
te
ir
r
e
ve
r
s
ibl
y
th
r
ough
matc
hing
s
tage
s
[
4]
.
E
xis
ti
ng
a
ppr
oa
c
he
s
e
xhibi
t
dis
ti
nc
t
li
mi
tations
in
thi
s
r
e
ga
r
d:
dis
tanc
e
-
tr
a
ns
f
or
m
-
ba
s
e
d
methods
[
5]
de
tec
t
f
inger
ti
ps
a
s
e
xtr
e
mal
point
s
but
lac
k
r
obus
tnes
s
to
pos
e
va
r
iation,
c
ontour
-
ba
s
e
d
a
ppr
oa
c
he
s
[
6]
a
r
e
s
us
c
e
pti
ble
to
s
pur
ious
de
f
e
c
t
s
c
a
u
s
e
d
by
v
e
in
s
ha
dows
,
a
nd
c
onve
xit
y
de
f
e
c
t
methods
[
7]
r
e
main
s
e
ns
it
ive
to
s
e
gmenta
ti
on
nois
e
.
F
ur
ther
mo
r
e
,
t
r
a
dit
ional
r
e
c
tangula
r
R
OI
s
of
ten
tr
unc
a
te
va
lua
ble
ve
in
pa
tt
e
r
ns
a
long
the
ha
nd
bor
de
r
s
[
8]
,
[
9]
,
li
mi
ti
ng
dis
c
r
im
inative
c
a
pa
c
it
y.
Ada
pti
ve
r
e
gion
-
gr
owing
[
10]
a
nd
de
e
p
lea
r
ning
a
ppr
oa
c
he
s
[
11]
,
[
12]
ha
ve
a
dva
nc
e
d
the
f
ield,
ye
t
a
c
r
it
ica
l
ga
p
pe
r
s
is
ts
a
c
r
os
s
thes
e
pa
r
a
digm
s
(
s
e
e
s
e
c
ti
on
2
f
or
de
tailed
a
na
lys
is
)
:
they
typi
c
a
ll
y
ope
r
a
te
withou
t
r
igor
ous
va
li
da
ti
on
o
f
a
na
tom
ica
l
landma
r
ks
pr
ior
to
s
e
gmenta
ti
on,
lea
ving
e
xtr
a
c
ti
ons
vulner
a
ble
to
no
is
e
,
s
pur
ious
de
f
e
c
ts
,
a
nd
f
a
ls
e
pos
it
ives
.
Unde
r
da
ta
-
s
c
a
r
c
e
c
ondit
ions
c
omm
on
to
dor
s
a
l
ha
nd
ve
in
r
e
c
ognit
ion,
whe
r
e
publi
c
ly
a
va
il
a
ble
da
tas
e
ts
a
r
e
typi
c
a
ll
y
ins
uf
f
icie
nt
f
o
r
c
omp
r
e
he
ns
ive
de
e
p
ne
twor
k
t
r
a
ini
ng
[
13]
a
nd
li
mi
ted
da
ta
ha
s
be
e
n
s
hown
to
de
gr
a
de
e
ve
n
s
ophis
ti
c
a
ted
a
r
c
hit
e
c
tur
e
s
[
14]
,
ge
ometr
y
-
dr
iven
pr
e
pr
oc
e
s
s
ing
that
e
nc
ode
s
a
na
tom
ica
l
pr
ior
s
of
f
e
r
s
a
mo
r
e
r
obus
t
a
l
ter
na
ti
ve
t
o
pur
e
ly
da
ta
-
dr
iven
a
ppr
oa
c
he
s
.
W
e
pr
opos
e
c
ontour
-
guided
c
onve
xit
y
de
f
e
c
t
ge
ometr
y
R
OI
(
C
G
-
C
DG
R
OI
)
,
whic
h
int
r
oduc
e
s
two
ge
ometr
ica
ll
y
-
gr
ounde
d
va
li
da
ti
on
c
r
it
e
r
ia:
i)
c
hor
d
-
c
ontour
topol
ogica
l
c
ons
is
tenc
y
e
ns
ur
ing
va
ll
e
ys
c
or
r
e
s
pond
to
s
im
ple
int
e
r
-
f
inger
indenta
ti
ons
a
nd
ii
)
mi
nim
u
m
e
nc
los
e
d
a
r
e
a
thr
e
s
hold
gua
r
a
ntee
ing
phys
ica
l
plaus
ibi
li
ty
by
dis
c
a
r
ding
s
ha
ll
ow
a
r
t
if
a
c
ts
T
he
va
li
da
ted
va
ll
e
ys
de
f
ine
a
s
qua
r
e
R
OI
c
e
nt
e
r
e
d
on
the
ha
nd
c
e
ntr
oid
a
nd
a
li
gne
d
with
the
domi
na
nt
va
ll
e
y
or
ienta
ti
on
.
C
ompr
e
he
ns
ive
e
va
l
ua
ti
on
on
a
publ
ic
da
taba
s
e
de
mons
tr
a
tes
that
our
method
a
c
hieve
s
0.
0266
e
qua
l
e
r
r
or
r
a
te
(
E
E
R
)
[
0
.
024
1
-
0.
0289]
with
his
togr
a
m
of
or
iente
d
g
r
a
dients
(
HO
G)
de
s
c
r
ipt
or
s
,
s
igni
f
ica
ntl
y
outper
f
or
mi
ng
f
our
s
tate
-
of
-
the
-
a
r
t
tec
hniques
(
W
il
c
oxon
s
igned
-
r
a
nk
pos
t
-
hoc
tes
t
with
B
onf
e
r
r
oni
c
or
r
e
c
ti
on
,
a
ll
p
<
0
.
001
ve
r
s
us
C
G
-
C
DG
;
boots
tr
a
p
95
%
C
I
f
or
the
p
r
opos
e
d
method:
[
0.
0241
-
0.
0289]
)
.
Our
method
pr
oc
e
s
s
e
s
im
a
ge
s
in
44
ms
on
s
tanda
r
d
ha
r
dwa
r
e
,
of
f
e
r
ing
r
e
a
l
-
ti
me
c
a
pa
bil
it
y
without
tr
a
ini
ng
da
ta.
W
e
e
mphas
ize
that
the
c
ontr
ibut
ion
of
C
G
-
C
DG
is
not
a
ne
w
pr
oc
e
dur
e
f
or
c
omput
ing
c
onve
xit
y
de
f
e
c
ts
,
the
unde
r
lyi
ng
hull
a
nd
de
f
e
c
t
c
ons
tr
uc
ti
on
f
oll
ow
the
s
tanda
r
d
f
or
mul
a
ti
on
,
but
a
va
li
da
ti
on
laye
r
a
ppli
e
d
to
c
a
ndidate
de
f
e
c
ts
be
f
or
e
they
a
r
e
tr
us
t
e
d
a
s
a
na
tom
ica
l
landma
r
ks
.
P
r
ior
c
onve
xit
y
-
de
f
e
c
t
-
ba
s
e
d
methods
[
7]
,
[
14]
–
[
16]
us
e
the
de
f
e
c
t's
de
pth
(
a
nd
,
in
s
ome
c
a
s
e
s
,
it
s
int
e
r
io
r
a
ngle)
d
ir
e
c
tl
y
a
s
the
l
a
ndmar
k
s
e
lec
ti
on
c
r
it
e
r
ion;
none
of
them
ve
r
if
ies
that
the
de
f
e
c
t
c
hor
d
c
los
e
s
int
o
a
topol
ogica
ll
y
c
o
ns
is
tent,
phys
ica
ll
y
plaus
ibl
e
va
ll
e
y
be
f
or
e
it
is
us
e
d
to
de
f
i
ne
the
R
OI
.
2.
RE
L
AT
E
D
WORK
R
OI
e
xtr
a
c
ti
on
r
e
li
e
s
he
a
vil
y
on
the
a
c
c
ur
a
te
loca
li
z
a
ti
on
of
a
na
tom
ica
l
r
e
f
e
r
e
nc
e
point
s
.
Ove
r
the
ye
a
r
s
,
a
ppr
oa
c
he
s
ha
ve
e
volved
f
r
om
t
r
a
dit
ional
im
a
ge
pr
oc
e
s
s
ing
a
nd
manua
l
s
e
gmenta
ti
on
to
a
da
pti
ve
ge
ometr
ic
methods
a
nd,
mor
e
r
e
c
e
ntl
y,
de
e
p
lea
r
n
ing
-
ba
s
e
d
s
olut
ions
.
De
s
pit
e
the
s
e
a
dva
nc
e
ments
,
e
xis
ti
ng
methodologi
e
s
e
xhibi
t
dis
ti
nc
t
li
mi
tations
r
e
ga
r
ding
ge
ometr
ic
va
li
da
ti
on,
r
obus
tnes
s
to
nois
e
,
a
nd
ge
ne
r
a
li
z
a
ti
on
a
c
r
os
s
diver
s
e
im
a
ging
c
ondit
ions
.
2.
1.
Color
an
d
in
t
e
n
s
it
y
-
b
as
e
d
m
e
t
h
od
s
E
a
r
ly
a
ppr
oa
c
he
s
e
xploi
ted
s
kin
c
hr
omaticity
a
nd
pixel
int
e
ns
it
y
f
or
s
e
gmenta
ti
on
[
17
]
.
W
hil
e
e
f
f
e
c
ti
ve
in
vis
ibl
e
li
ght
,
thes
e
methods
a
r
e
uns
uit
a
ble
f
or
dor
s
a
l
ha
nd
ve
in
r
e
c
ognit
ion,
whic
h
r
e
li
e
s
on
ne
a
r
-
inf
r
a
r
e
d
(
NI
R
)
im
a
ge
r
y
whe
r
e
c
olor
c
ue
s
a
r
e
a
bs
e
nt.
F
ur
ther
mo
r
e
,
they
de
pe
nd
on
global
th
r
e
s
holdi
ng
a
nd
s
table
a
na
tom
ica
l
pr
opo
r
ti
ons
,
lac
king
loca
l
g
e
ometr
ic
va
li
da
ti
on
.
T
his
r
e
nde
r
s
them
s
us
c
e
pti
ble
to
il
lum
ination
a
nis
otr
opy
a
nd
s
kin
tone
diver
s
it
y,
of
ten
f
a
il
ing
to
ve
r
i
f
y
the
a
na
tom
ica
l
plaus
ibi
li
ty
of
de
tec
ted
r
e
gions
.
2.
2.
Geom
e
t
r
ic
lan
d
m
ar
k
an
d
c
on
t
ou
r
an
alys
is
Ge
ometr
ic
a
ppr
oa
c
he
s
loca
te
int
e
r
-
f
inger
va
ll
e
ys
us
ing
c
onve
xit
y
de
f
e
c
ts
,
dis
tanc
e
t
r
a
ns
f
or
ms
,
or
c
ur
va
tur
e
a
na
lys
is
.
C
ha
i
e
t
al.
[
14
]
a
nd
Z
ha
ng
e
t
al
.
[
15]
uti
li
z
e
d
c
onve
x
hull
s
a
nd
tange
nt
-
ba
s
e
d
d
e
tec
ti
on,
while
other
s
[
6]
,
[
16]
a
ppli
e
d
f
il
ter
s
or
E
uc
li
de
a
n
metr
ics
to
identif
y
va
ll
e
y
mi
nim
a
.
De
s
pit
e
their
c
omput
a
ti
ona
l
e
f
f
icie
nc
y,
thes
e
methods
s
ha
r
e
a
f
unda
menta
l
we
a
kne
s
s
:
they
d
e
tec
t
loca
l
e
xtr
e
ma
(
de
pth,
a
ngle,
or
c
ur
va
tu
r
e
)
without
va
li
da
ti
ng
the
a
na
tom
ica
l
f
e
a
s
ibi
li
ty
of
the
unde
r
lyi
ng
indenta
ti
on
.
S
pe
c
if
ica
ll
y,
pr
ior
c
onve
xit
y
-
de
f
e
c
t
-
ba
s
e
d
methods
[
7]
,
[
14]
–
[
1
6]
,
us
e
the
de
f
e
c
t's
de
pth
(
a
nd,
in
s
ome
c
a
s
e
s
,
it
s
int
e
r
io
r
a
ngle)
dir
e
c
tl
y
a
s
the
landma
r
k
s
e
lec
ti
on
c
r
it
e
r
ion
;
none
of
them
ve
r
if
ies
that
the
de
f
e
c
t
c
hor
d
c
los
e
s
int
o
a
topol
ogica
ll
y
c
ons
is
tent,
phys
ica
ll
y
plaus
ibl
e
va
l
l
e
y
be
f
or
e
it
is
us
e
d
to
de
f
ine
the
R
OI
.
L
a
c
king
c
ons
tr
a
int
s
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
16
,
No.
5
,
Oc
tober
20
26
:
2622
-
2640
2624
s
uc
h
a
s
c
hor
d
–
c
ontour
c
ons
i
s
tenc
y
or
e
nc
los
e
d
-
a
r
e
a
c
r
it
e
r
ia,
they
a
r
e
highl
y
vulne
r
a
ble
to
s
pur
ious
de
f
e
c
ts
c
a
us
e
d
by
c
ontour
nois
e
,
s
ha
dows
,
a
nd
s
e
gmenta
ti
on
e
r
r
or
s
.
T
r
a
dit
ional
manua
l
a
ppr
oa
c
he
s
[
18]
,
[
19
]
f
ur
the
r
highl
ight
thes
e
li
mi
tations
,
be
ing
ti
me
-
c
ons
umi
ng,
p
r
one
to
human
e
r
r
or
,
a
nd
incons
is
tent
whe
n
a
p
pli
e
d
to
s
mall
da
tas
e
ts
.
2.
3.
Adap
t
ive,
h
ar
d
war
e
-
b
as
e
d
,
a
n
d
p
os
t
-
p
r
oc
e
s
s
in
g
m
e
t
h
od
s
T
o
ha
ndle
pos
e
va
r
iations
,
a
da
pti
ve
methods
dyn
a
mi
c
a
ll
y
a
djus
t
R
OI
bounda
r
ies
.
Da
mak
e
t
al.
[
9]
a
nd
Noz
a
r
i
P
ou
r
e
t
al
.
[
20]
pr
opos
e
d
w
r
is
t
-
dis
tan
c
e
pr
of
il
e
s
a
nd
f
loating
R
OI
s
to
maximi
z
e
ve
in
c
ove
r
a
ge
.
W
hil
e
f
lexible,
thes
e
tec
hniques
r
e
ly
on
unva
li
d
a
ted
e
xtr
e
mal
point
s
a
nd
lac
k
dua
l
ge
ometr
ic
pl
a
us
ibi
li
ty
tes
ts
.
Ha
r
dwa
r
e
-
ba
s
e
d
s
olut
ions
s
u
c
h
a
s
mul
ti
s
pe
c
tr
a
l
im
a
ging
[
21]
of
f
e
r
pr
e
c
is
ion
but
im
pos
e
pr
ohibi
ti
ve
c
os
ts
.
P
os
t
-
s
e
gment
a
ti
on
a
li
gnment
methods
[
11
]
f
oc
us
on
nor
malizing
pos
e
a
f
ter
R
OI
e
xt
r
a
c
ti
on
b
ut
do
not
a
ddr
e
s
s
the
pr
im
a
r
y
c
ha
ll
e
nge
of
dis
ti
nguis
hing
va
li
d
int
e
r
-
f
inger
va
ll
e
ys
f
r
om
s
pur
ious
de
f
e
c
ts
dur
in
g
ini
ti
a
l
R
OI
c
ons
tr
uc
ti
on.
2.
4.
De
e
p
lear
n
in
g
an
d
d
at
a
-
d
r
iven
ap
p
r
oac
h
e
s
T
he
tr
a
ns
it
ion
to
de
e
p
lea
r
ning
ha
s
ma
r
ke
d
a
s
igni
f
ica
nt
a
dva
nc
e
ment
in
R
OI
e
xtr
a
c
ti
on
methodologi
e
s
.
C
onvolut
ional
ne
ur
a
l
ne
twor
k
(
C
NN
)
-
ba
s
e
d
s
olut
ions
s
uc
h
a
s
F
a
s
ter
R
-
C
N
N
[
10]
,
[
22]
a
nd
pa
lm
pr
int
C
NN
s
[
18
]
e
li
m
inate
manua
l
s
e
gmenta
ti
on,
p
r
ovidi
ng
a
utom
a
ti
c
R
OI
loca
li
z
a
ti
on
with
i
mpr
ove
d
a
c
c
ur
a
c
y
a
nd
r
e
duc
e
d
human
e
r
r
or
.
Da
i
e
t
al
.
[
23]
f
ur
the
r
int
r
oduc
e
d
a
li
ghtwe
ight
HR
Ne
t
f
or
ke
ypoint
loca
li
z
a
ti
on
in
nonc
ontac
t
pa
lm
ve
in
R
OI
e
xtr
a
c
ti
on,
a
c
hieving
97.
36%
a
c
c
ur
a
c
y
with
c
ompac
t
model
s
ize
a
nd
f
a
s
t
inf
e
r
e
nc
e
.
R
e
c
e
nt
wor
k
c
onti
nue
s
to
a
dva
nc
e
the
f
ield:
Z
ha
ng
e
t
al
.
[
2
4]
pr
opos
e
d
a
n
im
pr
o
v
e
d
U
-
Ne
t
f
or
ke
ypoint
ba
s
e
d,
non
-
c
ontac
t
dor
s
a
l
-
ha
nd
R
OI
e
xtr
a
c
ti
on
unde
r
c
ompl
e
x
ba
c
kgr
ounds
;
Z
ha
o
e
t
al.
[
25
]
int
r
oduc
e
d
VPC
F
or
mer
,
a
tr
a
ns
f
or
me
r
-
ba
s
e
d
mul
ti
-
view
f
inger
-
ve
in
r
e
c
ognit
ion
model;
a
nd
L
i
e
t
al.
[
26
]
de
mons
tr
a
ted
f
us
ion
of
R
e
s
Ne
t
a
nd
HO
G
f
e
a
t
ur
e
s
f
or
s
mall
-
s
c
a
le
dor
s
a
l
ha
nd
ve
in
da
taba
s
e
s
.
W
he
n
s
uf
f
icie
nt
a
nnotate
d,
in
-
domain
da
ta
a
r
e
a
va
il
a
ble
,
s
uc
h
lea
r
ne
d
loca
li
z
e
r
s
c
a
n
matc
h
o
r
e
xc
e
e
d
r
ul
e
-
ba
s
e
d
ge
ometr
ic
methods
;
thi
s
is
a
n
e
mpi
r
ica
l,
da
ta
-
a
va
il
a
bil
it
y
-
c
ondit
ioned
li
mi
tation
of
the
lea
r
ni
ng
-
ba
s
e
d
pa
r
a
digm
,
not
a
s
tr
uc
tur
a
l
one
.
How
e
ve
r
,
lea
r
ni
ng
-
ba
s
e
d
method
s
r
e
main
"
blac
k
boxe
s
,
"
lac
king
e
xpli
c
it
ge
ometr
ic
r
e
a
s
oning.
B
ounding
box
r
e
gr
e
s
s
ion
doe
s
not
gua
r
a
ntee
a
na
tom
ica
l
va
li
dit
y,
lea
ving
models
vulner
a
ble
to
out
-
of
-
dis
tr
ibut
ion
pos
e
s
.
M
or
e
ove
r
,
their
r
e
li
a
nc
e
on
lar
ge
a
nnotate
d
da
tas
e
ts
li
mi
ts
their
pr
a
c
ti
c
a
li
ty,
a
s
publi
c
ly
a
va
il
a
ble
ve
in
da
tas
e
ts
a
r
e
typi
c
a
ll
y
not
lar
ge
e
nough
to
pe
r
f
or
m
c
ompr
e
he
ns
ive
tr
a
ini
ng
of
de
dica
ted
ne
twor
ks
[
12]
.
B
a
gc
hi
e
t
al.
[
13]
f
ur
ther
de
mons
tr
a
te
thi
s
on
dor
s
a
l
ha
nd
ve
ins
s
pe
c
if
ica
ll
y,
s
howing
that
li
mi
ted
da
tas
e
t
a
va
il
a
bil
it
y
c
a
us
e
s
mor
e
c
ompl
e
x
de
e
p
a
r
c
hit
e
c
tur
e
s
(
S
iame
s
e
,
T
r
ipl
e
t
ne
twor
ks
)
to
unde
r
pe
r
f
or
m
s
im
pler
C
NN
s
.
C
G
-
C
DG
is
s
c
ope
d
to
c
ompl
e
ment
,
r
a
ther
than
r
e
plac
e
,
thes
e
a
ppr
oa
c
he
s
:
it
tar
ge
ts
the
no
-
tr
a
ini
ng
-
da
ta,
l
ow
-
da
ta,
a
nd
c
r
os
s
-
domain
r
e
gim
e
,
whe
r
e
,
a
s
B
a
gc
hi
e
t
a
l.
[
13]
a
nd,
mor
e
br
oa
dly
,
He
mi
s
e
t
al.
[
27]
,
c
ompl
e
x
lea
r
ne
d
a
r
c
hit
e
c
tur
e
s
a
r
e
known
to
unde
r
pe
r
f
o
r
m
s
im
pler
models
whe
n
a
nnotate
d
da
ta
a
r
e
s
c
a
r
c
e
.
2.
5.
M
u
lt
im
od
al
,
h
y
b
r
id
,
an
d
ad
van
c
e
d
s
e
gm
e
n
t
at
ion
m
e
t
h
od
s
B
y
c
ombi
ning
the
ve
in
pa
tt
e
r
n
with
pa
lm
pr
int
or
ge
ometr
y
f
e
a
tur
e
s
,
hybr
id
s
ys
tems
im
pr
ove
r
e
c
ognit
ion
[
1]
,
[
5]
,
[
28]
,
[
29]
.
T
hough
f
us
ion
e
nh
a
nc
e
s
matc
hing
pe
r
f
o
r
manc
e
,
thes
e
s
ys
tems
inher
it
the
R
OI
e
xtr
a
c
ti
on
li
mi
tations
o
f
thei
r
ge
ometr
ic
modul
e
s
.
I
n
ge
ne
r
a
l,
they
a
pply
a
s
tr
a
ight
f
or
wa
r
d
c
e
ntr
oi
d
-
ba
s
e
d
wa
y,
whe
r
e
no
va
li
da
ti
on
wa
s
done
to
c
he
c
k
it
s
e
xtr
a
c
ti
on
s
tr
uc
tur
a
l
in
tegr
it
y,
thus
e
r
r
or
s
in
the
f
ir
s
t
R
OI
s
tage
will
pr
opa
ga
te
thr
oughout
the
matc
hing
pipel
ine.
R
e
c
e
nt
r
e
s
e
a
r
c
h
s
tudi
e
d
a
dva
n
c
e
d
s
e
gmenta
ti
on
tec
hniques
in
hybr
id
f
r
a
mew
or
ks
.
Ac
c
or
ding
to
L
a
gha
r
i
e
t
al
.
[
19]
,
the
pr
opos
e
d
hyb
r
id
a
uto
matic
s
e
gmenta
ti
on
method
(
HH
M
)
ut
il
ize
s
hi
s
togr
a
m
e
qua
li
z
a
ti
on
a
long
with
mo
r
phologi
c
a
l
a
nd
thr
e
s
holdi
ng
-
ba
s
e
d
a
lgor
it
hms
to
a
c
hieve
84%
a
c
c
ur
a
c
y
whe
n
a
utom
a
ti
c
s
e
gmente
d
a
nd
a
ugmente
d
the
i
mage
s
.
Ndu
e
t
al
.
[
30]
pr
opos
e
d
a
n
a
ppr
oa
c
h
whic
h
us
e
s
a
c
ombi
na
ti
on
of
d
im
e
ns
ionalit
y
r
e
duc
ti
on
methods
P
C
A,
F
olded
-
P
C
A,
K
-
mea
ns
c
lus
ter
ing
to
3D
-
C
NN
-
ba
s
e
d
ve
in
de
tec
ti
on
on
hype
r
s
pe
c
tr
a
l
im
a
ge
s
.
E
ve
n
though
thes
e
methods
im
pr
ove
s
e
gmenta
ti
on
qua
li
ty,
they
do
not
im
pos
e
ge
ometr
ic
c
ons
tr
a
int
s
on
the
e
xtr
a
c
ted
l
a
ndmar
ks
.
2.
6.
M
ot
iva
t
ion
a
n
d
r
e
s
e
ar
c
h
gap
A
r
e
c
ur
r
in
g
li
mi
t
a
ti
on
a
c
r
o
s
s
the
g
e
ome
tr
ic,
a
d
a
pt
ive,
a
nd
l
e
a
r
ni
ng
-
ba
s
e
d
li
ter
a
tur
e
r
e
vi
e
we
d
a
bo
ve
is
th
a
t
a
na
tom
i
c
a
l
l
a
nd
mar
k
s
a
r
e
v
a
li
d
a
te
d
a
g
a
i
ns
t
a
s
ingl
e
,
lo
c
a
l
c
r
it
e
r
ion
(
a
d
e
f
e
c
t
-
de
pt
h
thr
e
s
hol
d,
a
n
int
e
r
ior
a
ngle,
or
a
le
a
r
ne
d
c
onf
i
de
n
c
e
s
c
or
e
)
e
va
lu
a
ted
in
i
s
ol
a
ti
o
n,
with
no
s
ub
s
e
q
ue
n
t
c
he
c
k
that
the
c
a
nd
idat
e
i
s
ge
ometr
ica
ll
y
c
on
s
i
s
tent
with
a
g
e
nu
ine
int
e
r
-
f
ing
e
r
va
l
ley.
T
hi
s
li
mi
tat
ion
i
s
mo
s
t
e
x
pli
c
it
i
n
c
onv
e
xit
y
-
de
f
e
c
t
-
b
a
s
e
d
meth
od
s
:
S
a
le
hin
e
t
al.
[
7
]
,
C
ha
i
e
t
al.
[
14]
,
Z
ha
ng
e
t
al.
[
15]
,
a
nd
Ke
kr
e
e
t
al.
[
16]
a
ll
r
e
ta
in
a
c
a
ndi
da
t
e
onc
e
it
s
d
e
f
e
c
t
de
pth
(
a
n
d,
in
s
ome
v
a
r
ia
nts
,
it
s
int
e
r
ior
a
n
gle)
e
xc
e
e
d
s
a
f
ixed
thr
e
s
hold,
witho
ut
v
e
r
if
yin
g
th
a
t
th
e
c
hor
d
s
p
a
nn
ing
th
e
d
e
f
e
c
t
a
c
tua
ll
y
c
lo
s
e
s
i
nto
a
bo
unde
d,
ph
ys
i
c
a
ll
y
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
C
ontour
-
guided
c
onv
e
x
it
y
de
fec
t
ge
ome
tr
y
for
pr
e
c
is
e
R
OI
e
x
tr
ac
ti
on
in
…
(
Habib
K
ade
m
)
2625
plau
s
ibl
e
va
ll
e
y.
Ada
pti
ve
va
r
i
a
nt
s
[
9]
,
[
20]
r
e
l
a
x
thi
s
t
hr
e
s
hol
d
to
a
c
c
omm
od
a
te
po
s
e
v
a
r
iat
ion
b
u
t
inher
it
the
s
a
m
e
we
a
kn
e
s
s
,
s
inc
e
e
x
tr
e
ma
l
poin
ts
a
r
e
s
t
il
l
a
c
c
e
p
ted
witho
ut
a
n
y
ind
e
pe
nde
nt
pla
us
ibi
li
t
y
c
he
c
k.
C
on
s
e
que
n
tl
y,
a
ll
the
s
e
meth
od
s
r
e
main
s
e
n
s
it
iv
e
to
s
pur
iou
s
de
f
e
c
t
s
pr
oduc
e
d
by
c
onto
ur
ir
r
e
g
ular
it
i
e
s
,
s
kin
f
ol
ds
,
or
s
e
g
ment
a
ti
o
n
noi
s
e
,
wh
ich
c
a
n
loca
ll
y
mi
mi
c
a
ge
nuin
e
int
e
r
-
f
i
nge
r
in
de
nt
a
t
ion.
T
he
c
ontr
ib
uti
on
of
the
pr
e
s
e
nt
wor
k
i
s
pr
e
c
i
s
e
ly
t
his
mi
s
s
ing
ve
r
if
i
c
a
ti
on
s
tep,
r
a
t
he
r
th
a
n
a
ne
w
wa
y
of
de
te
c
ti
n
g
c
o
nve
xit
y
de
f
e
c
t
s
.
W
e
pr
op
os
e
CG
-
C
D
G
in
whic
h
a
c
a
ndid
a
te
va
ll
e
y
iden
ti
f
ied
by
the
c
a
ndid
a
te
-
s
e
l
e
c
ti
on
s
t
a
ge
C
S
-
C
DG
i
s
r
e
tain
e
d
only
if
it
joi
ntl
y
s
a
ti
s
f
ie
s
i)
a
c
hor
d
-
c
on
tour
topol
o
gic
a
l
c
on
s
i
s
te
nc
y
te
s
t,
ve
r
if
yin
g
that
t
he
c
hor
d
a
s
s
o
c
iat
e
d
wi
th
the
de
f
e
c
t
int
e
r
s
e
c
ts
th
e
h
a
nd
c
onto
ur
in
a
ma
nne
r
c
omp
a
ti
b
le
with
a
ge
nui
ne
int
e
r
-
f
in
ge
r
va
ll
e
y,
a
nd
ii
)
a
mi
nim
um
e
nc
l
os
e
d
-
a
r
e
a
te
s
t,
dis
c
a
r
di
ng
c
a
n
dida
te
s
who
s
e
bound
e
d
r
e
gion
i
s
t
oo
s
mall
to
c
or
r
e
s
pon
d
to
a
n
a
na
to
mi
c
a
ll
y
pl
a
u
s
ibl
e
no
tch
(
s
e
c
ti
o
n
3.
6,
e
q
ua
ti
on
(
6)
)
.
U
nli
ke
t
he
s
ingl
e
-
c
r
it
e
r
ion
a
c
c
e
ptan
c
e
r
ule
s
of
[
7]
,
[
9]
,
[
14
]
,
[
16]
,
[
20]
,
thi
s
joi
nt
te
s
t
d
oe
s
not
mer
e
ly
ti
g
hten
t
he
de
te
c
ti
o
n
thr
e
s
hol
d;
it
a
d
ds
a
n
ind
e
pe
nde
nt
ge
o
metr
ic
c
on
s
i
s
ten
c
y
r
e
q
uir
e
m
e
nt
th
a
t
a
s
pur
io
us
de
f
e
c
t
i
s
unli
ke
ly
to
s
a
ti
s
f
y
by
c
ha
nc
e
.
T
hi
s
dis
ti
nc
ti
on
(
a
ve
r
if
iabl
e
dua
l
te
s
t
r
a
th
e
r
than
a
s
ingl
e
de
pt
h
-
or
a
ngle
-
ba
s
e
d
s
c
or
e
)
i
s
the
pr
inci
pa
l
m
e
th
odolo
gic
a
l
no
ve
lt
y
of
thi
s
s
tudy,
a
c
hi
e
v
e
d
wi
thout
tr
a
ini
ng
d
a
ta
a
nd
with
f
ull
int
e
r
pr
e
ta
bil
it
y
of
th
e
de
c
i
s
ion
c
r
it
e
r
i
a
.
R
e
c
e
nt
de
e
p
-
le
a
r
ning
-
ba
s
e
d
loc
a
li
z
e
r
s
,
incl
uding
U
-
N
e
t
-
ba
s
e
d
ke
ypo
int
de
t
e
c
ti
on
[
24]
,
tr
a
n
s
f
or
m
e
r
-
ba
s
e
d
m
ult
i
-
view
a
r
c
hit
e
c
tur
e
s
[
25]
,
a
nd
R
e
s
Ne
t
-
H
OG
f
u
s
ion
s
c
h
e
me
s
[
26]
,
c
a
n
matc
h
or
e
xc
e
e
d
the
a
c
c
ur
a
c
y
of
ge
o
metr
y
-
ba
s
e
d
meth
od
s
onc
e
tr
a
ine
d
on
s
uf
f
ic
ient
ly
lar
ge
,
r
e
pr
e
s
e
nt
a
ti
v
e
d
a
ta,
a
nd,
unli
ke
f
ix
e
d
g
e
ome
tr
ic
r
ules
,
c
a
n
i
n
pr
in
c
ipl
e
be
r
e
tr
a
in
e
d
t
o
a
c
c
omm
od
a
te
n
e
w
a
c
qu
is
i
ti
on
c
ondit
ion
s
;
e
m
e
r
ging
e
xpl
a
in
a
bil
it
y
to
ols
(
e
.
g.
,
a
tt
e
nti
o
n
a
nd
s
a
li
e
n
c
y
map
s
)
a
r
e
a
l
s
o
na
r
r
owing
their
his
tor
ic
a
l
int
e
r
pr
e
ta
bil
it
y
ga
p.
T
he
ir
a
ppli
c
a
bil
it
y
to
dor
s
a
l
ha
nd
ve
in
im
a
ge
r
y,
how
e
v
e
r
,
r
e
main
s
c
on
s
tr
a
i
ne
d
by
the
s
c
a
r
c
it
y
of
a
nnota
ted
tr
a
in
ing
d
a
ta
f
or
thi
s
modal
it
y
[
12]
,
a
nd,
a
s
B
a
g
c
hi
e
t
al.
[
13]
r
e
por
t
s
pe
c
if
ic
a
ll
y
f
or
thi
s
moda
li
ty,
mor
e
c
o
mpl
e
x
lea
r
ne
d
a
r
c
hit
e
c
tur
e
s
c
a
n
und
e
r
pe
r
f
or
m
s
im
pl
e
r
mod
e
ls
wh
e
n
da
ta
i
s
li
mi
ted.
C
G
-
C
D
G
is
a
c
c
or
di
ngly
pr
op
o
s
e
d
not
a
s
a
c
om
pe
ti
t
or
to
thes
e
a
ppr
o
a
c
he
s
but
a
s
a
c
ompl
e
me
ntar
y,
z
e
r
o
-
s
h
ot
s
olu
ti
on
f
or
the
low
-
d
a
ta
a
nd
c
r
o
s
s
-
dom
a
in
r
e
gim
e
in
whic
h
th
e
y
r
e
m
a
in
lea
s
t
r
e
li
a
bl
e
,
a
nd
a
s
a
c
a
n
did
a
te
a
n
a
tom
i
c
a
ll
y
gr
oun
de
d
pr
e
pr
o
c
e
s
s
ing
s
t
a
ge
f
or
hybr
id
pipe
li
ne
s
on
c
e
lar
ge
r
a
nnot
a
te
d
da
t
a
s
e
t
s
be
c
ome
a
va
i
labl
e
.
R
e
lative
to
thi
s
r
e
gim
e
,
C
G
-
C
DG
of
f
e
r
s
z
e
r
o
-
s
hot
ge
ne
r
a
li
z
a
ti
on,
lowe
r
c
omput
a
ti
ona
l
c
os
t,
a
nd
f
u
ll
int
e
r
pr
e
tabili
ty
of
it
s
va
li
da
ti
on
c
r
it
e
r
ia
.
T
a
ble
1
s
ys
tema
ti
c
a
ll
y
c
ompar
e
s
c
ur
r
e
nt
R
OI
e
xtr
a
c
ti
on
methods
with
r
e
s
pe
c
t
to
ge
ometr
ic
va
li
da
ti
on
c
r
it
e
r
ia,
a
na
to
mi
c
a
l
va
li
da
ti
on,
tr
a
ini
ng
de
pe
nde
nc
y,
int
e
r
p
r
e
tabili
ty,
a
nd
r
obus
tnes
s
,
s
umm
a
r
izing
the
s
tr
uc
tur
a
l
a
dva
ntage
s
of
the
p
r
opos
e
d
a
ppr
oa
c
h.
T
a
ble
1.
C
ompar
is
on
of
R
OI
e
xtr
a
c
ti
on
methods
f
o
r
dor
s
a
l
ha
nd
ve
in
r
e
c
ognit
ion
M
e
th
od
G
e
ome
tr
ic
va
li
da
ti
on c
r
it
e
r
ia
A
na
to
mi
c
a
l
va
li
da
ti
on
T
r
a
in
in
g
I
nt
e
r
pr
e
ta
bi
li
ty
R
obus
tn
e
s
s
K
e
kr
e
e
t
al
.
[
16]
N
one
No
No
H
ig
h
M
e
di
um
D
a
ma
k
e
t
al
.
[
9]
E
xt
r
e
ma
l
poi
nt
s
onl
y
No
No
H
ig
h
H
ig
h
H
a
o
e
t
al
.
[
21]
N
one
No
No
M
e
di
um
M
e
di
um
I
z
a
dpa
na
h
e
t
al
.
[
10]
N
one
No
Y
e
s
*
M
e
di
um
H
ig
h
P
r
opos
e
d
C
hor
d
-
c
ont
our
i
nt
e
r
s
e
c
ti
on +
e
nc
lo
s
e
d a
r
e
a
t
hr
e
s
hol
d
Y
e
s
No
V
e
r
y H
ig
h
H
ig
h
* T
r
a
ns
f
e
r
l
e
a
r
ni
ng on pr
e
tr
a
in
e
d I
ma
ge
N
e
t
w
e
ig
ht
s
.
N
ot
e
:
N
o
in
di
c
a
te
s
a
b
s
e
nc
e
of
e
xpl
ic
it
ge
ome
tr
ic
pl
a
u
s
ib
il
it
y
te
s
t
pr
io
r
to
R
O
I
c
ons
tr
uc
ti
on.
T
he
pr
opos
e
d
me
th
od
is
th
e
onl
y
a
ppr
oa
c
h
e
nf
or
c
in
g
dua
l
ge
ome
tr
ic
c
ons
tr
a
in
ts
(
to
pol
ogi
c
a
l
c
ons
is
te
nc
y
a
nd
phy
s
ic
a
l
a
r
e
a
th
r
e
s
hol
d)
be
f
or
e
va
ll
e
y
-
to
-
R
O
I
ma
ppi
ng
3.
P
ROP
OS
E
D
M
E
T
HO
DOL
OG
Y
W
e
pr
e
s
e
nt
C
G
-
C
DG
R
OI
,
a
ge
ometr
y
-
dr
iven
pip
e
li
ne
f
or
r
obus
t
R
OI
e
xtr
a
c
ti
on
f
r
om
dor
s
a
l
ha
nd
ve
in
im
a
ge
s
.
F
igur
e
1
il
lus
tr
a
tes
the
ove
r
a
ll
wor
k
f
low
o
f
the
p
r
opos
e
d
method
,
whic
h
ope
r
a
tes
on
s
tanda
r
d
gr
a
ys
c
a
le
input
s
without
r
e
quir
ing
tr
a
ini
ng
da
ta
or
s
pe
c
ialize
d
ha
r
dwa
r
e
,
a
nd
c
ompr
is
e
s
the
f
oll
owing
s
e
que
nti
a
l
s
teps
:
S
tep
(
1)
input
gr
a
ys
c
a
le
do
r
s
a
l
ha
nd
im
a
ge
;
S
tep
(
2)
bina
r
y
s
e
gmenta
ti
on
us
ing
Ga
us
s
ian
s
moot
hing
f
oll
owe
d
by
Ots
u
thr
e
s
holdi
ng;
S
tep
(
3)
ha
nd
c
ontour
(
gr
e
e
n)
a
nd
c
onve
x
hull
(
o
r
a
nge
)
e
xt
r
a
c
ti
on;
S
tep
(
4)
r
a
w
c
onve
xit
y
de
f
e
c
ts
de
tec
ti
on
with
c
ho
r
d
e
ndpoint
s
(
r
e
d
point
s
)
;
S
tep
(
5)
ge
omet
r
ica
ll
y
va
li
da
ted
va
ll
e
ys
(
V1,
V2,
V3)
us
ing
int
e
r
s
e
c
ti
on
a
nd
a
r
e
a
-
ba
s
e
d
c
r
it
e
r
ia;
S
tep
(
6)
domi
na
nt
or
ienta
ti
on
e
s
ti
mate
d
via
lea
s
t
-
s
qua
r
e
s
f
it
ti
ng
(
θ
=
-
14.
81°
)
;
S
tep
(
7)
R
O
I
c
ons
tr
uc
ted
a
r
ound
the
ha
nd
c
e
ntr
o
id
a
li
gne
d
with
the
e
s
ti
mate
d
or
ienta
ti
on;
a
nd
S
tep
(
8
)
f
inal
R
OI
f
or
s
u
bs
e
que
nt
f
e
a
tur
e
e
xtr
a
c
ti
on.
A
high
-
leve
l
ove
r
view
of
the
pr
opos
e
d
pipeline
is
s
hown
in
F
igur
e
2.
T
he
pr
oc
e
s
s
c
ons
is
ts
of
f
ive
main
s
tage
s
:
pr
e
pr
oc
e
s
s
ing
to
e
xtr
a
c
t
the
ha
nd
c
ontour
,
de
f
e
c
t
a
na
lys
is
to
loca
te
va
ll
e
y
point
s
,
s
ur
f
a
c
e
de
tec
ti
on
to
f
or
m
polygonal
r
e
gions
,
gr
oup
s
e
lec
ti
on
to
identif
y
the
opti
mal
2
-
3
a
li
gne
d
va
ll
e
ys
,
a
nd
R
OI
e
xtr
a
c
ti
on
to
ge
ne
r
a
te
a
r
otate
d
s
qua
r
e
r
e
gion
c
e
nte
r
e
d
a
t
the
ha
nd
c
e
ntr
oid.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
16
,
No.
5
,
Oc
tober
20
26
:
2622
-
2640
2626
F
igur
e
1.
W
or
kf
low
of
the
C
G
C
DG
method
f
or
R
OI
e
xtr
a
c
ti
on
F
igur
e
2.
W
or
kf
low
of
the
C
G
C
DG
method
f
or
R
OI
e
xtr
a
c
ti
on
3.
1.
P
r
e
p
r
oc
e
s
s
in
g
a
n
d
h
an
d
s
e
gm
e
n
t
at
ion
T
he
pr
opos
e
d
f
r
a
mew
or
k
be
gins
by
pr
e
pr
oc
e
s
s
ing
the
input
gr
a
ys
c
a
le
im
a
ge
s
(
,
)
.
A
Ga
us
s
ian
s
moot
hing
f
il
ter
wi
th
a
5
×
5
ke
r
ne
l
is
a
ppli
e
d.
T
his
c
onf
igur
a
ti
on
wa
s
s
e
lec
ted
a
s
it
is
opti
mal
f
or
a
t
t
e
nua
ti
ng
high
-
f
r
e
que
nc
y
nois
e
(
e
.
g.
,
s
kin
textur
e
a
nd
ha
ir
)
while
pr
e
s
e
r
ving
the
s
tr
uc
tur
a
ll
y
wide
r
dor
s
a
l
ve
in
bounda
r
ies
,
whic
h
ult
im
a
tely
f
a
c
il
it
a
tes
the
e
xtr
a
c
ti
on
of
a
s
ha
r
pe
r
c
ontour
a
f
ter
s
e
gmenta
ti
on
[
2
8]
,
[
31
]
.
S
ubs
e
que
ntl
y,
the
ha
nd
r
e
gion
is
s
e
gmente
d
f
r
om
t
he
ba
c
kgr
ound
us
ing
Ots
u's
thr
e
s
holdi
ng
method
[
32]
,
a
nd
the
r
e
s
ult
ing
binar
y
mas
k
I
bin
is
r
e
f
ined
th
r
ough
mor
phologi
c
a
l
c
los
ing
a
nd
ope
ning
ope
r
a
ti
ons
with
a
7x7
e
ll
ipt
ica
l
s
tr
uc
tur
ing
e
leme
nt
to
e
li
m
inate
a
r
ti
f
a
c
ts
a
nd
e
ns
ur
e
a
c
onti
guous
r
e
gion
[
33]
,
[
34]
.
_
=
(
_
ℎ
ℎ
•
)
◦
(
1)
I
mpl
e
menta
ti
on
de
tails
.
T
he
r
e
f
e
r
e
nc
e
pipeline
wa
s
im
pleme
nted
in
P
ython
with
Ope
nC
V
a
nd
NumP
y.
C
ontour
s
we
r
e
e
xtr
a
c
ted
with
2
.
(
_
,
_
_
)
;
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
C
ontour
-
guided
c
onv
e
x
it
y
de
fec
t
ge
ome
tr
y
for
pr
e
c
is
e
R
OI
e
x
tr
ac
ti
on
in
…
(
Habib
K
ade
m
)
2627
the
c
onve
x
hull
with
2
.
;
a
nd
c
onve
xit
y
de
f
e
c
ts
with
2
.
.
Ope
nC
V
r
e
por
ts
de
f
e
c
t
de
pth
a
s
a
f
ixed
-
point
va
lue
s
c
a
led
by
256;
a
ll
c
ompar
is
ons
a
ga
ins
t
m
i
n
(
2)
a
r
e
pe
r
f
o
r
med
a
f
ter
div
idi
ng
the
r
e
por
ted
va
lue
by
256
.
0.
T
he
pa
r
a
mete
r
s
α
de
pth,
β
,
δ
,
a
nd
α
(
2)
-
(
4)
a
nd
(
7)
a
r
e
li
s
ted
with
their
numer
ica
l
va
lues
a
nd
r
ole
in
T
a
ble
2;
ℓ
(
3)
is
u
s
e
d
only
to
de
r
ive
m
i
n
a
na
lyt
ica
ll
y
a
nd
is
not
a
ppli
e
d
a
s
a
n
indepe
nde
nt
f
il
ter
.
T
a
ble
2.
P
a
r
a
mete
r
r
e
gis
tr
y
f
or
the
C
G
-
C
DG
pipeline.
All
thr
e
s
holds
a
r
e
e
xpr
e
s
s
e
d
a
s
s
c
a
le
-
nor
malize
d
ge
ometr
ic
invar
iants
r
e
lative
to
the
ha
nd
bounding
-
box
he
ight
(
H
bbox
)
a
nd
pa
lm
width
(
L
p
a
l
m
)
S
ymbol
V
a
lu
e
R
ol
e
E
qua
ti
on
α
d
e
p
th
0.02
M
in
im
um va
ll
e
y de
pt
h f
a
c
to
r
(
2)
β
0.20
M
in
im
um va
ll
e
y w
id
th
f
a
c
to
r
(
a
na
ly
ti
c
a
l
onl
y)
(
3)
δ
0.002
C
ompos
it
e
a
r
e
a
f
a
c
to
r
(
αde
pt
h
·
β
/
2)
(
4)
α
30°
A
ngul
a
r
t
ol
e
r
a
nc
e
f
or
va
ll
e
y
-
gr
oup pa
r
a
ll
e
li
s
m
(
7)
Algor
it
hm
1
.
CG
-
C
DG
R
OI
–
c
ontour
-
guided
c
on
ve
xit
y
de
f
e
c
t
ge
ometr
y
f
or
R
OI
Input
: Grayscale dorsal hand vein image
I(x, y)
Output
: Extracted square ROI
R
1
I
blur
←
GaussianBlur
(I, (5, 5));
2
B ←
OtsuThreshold
(I
blur
);
3
I
bin
←
MorphClose/Open
(B,
elliptical
kernel
7×7);
4
C ←
LargestExternalContour
(Ibin);
5
H ←
ConvexHull
(C);
6
D ←
ConvexityDefects
(C, H);
7
F
il
te
r
D:
k
ee
p
de
fe
ct
s
sa
ti
sf
yi
ng
t
he
de
pt
h
co
ns
tr
ai
nt
δ
i
≥
δ
m
in
(
Eq
.
(2
))
;
th
e
ang
ul
ar
t
ol
er
an
ce
α
i
s
ap
pl
ie
d
la
te
r,
d
ur
in
g
va
ll
ey
-
gr
ou
p
se
le
ct
io
n
(A
lg
or
it
hm
2
,
Eq
.
(7
))
;
8
V ←
∅
;
9
foreach
defect
d
i
∈
D
do
10
Let
s
i
, e
i
=
endpoints
, v
i
=
valley
point
;
11
Compute
chord
ℓ
i
=
̅
̅
̅
̅
̅
;
12
Compute
contour
segment
C
[si,ei]
;
13
Compute
enclosed
area
A
i
=
Area
(ℓ
i
, C
[si,ei]
);
14
Find
intersections
P = ℓ
i
∩ C;
15
if
|P | ≥ 2
and
A
i
≥ area_threshold
then
16
V ← V
∪
{v
i
};
17
V
∗
←
SelectBestGroup
(V ) //
Algorithm
2
;
18
(c
x
, c
y
) ← Centroid(C);
19
Fit
line
L
through
V
∗
;
compute
orientation
θ;
20
L ← 0.95 ·
max
v
∈
V
∗
∥
v − (c
x
, c
y
)
∥
∞;
21
Rotate
I
by
−θ
around
(c
x
, c
y
);
22
Crop
square
of
size
L × L → R;
23
return
R;
3.
2.
Cont
o
u
r
e
xt
r
ac
t
ion
an
d
c
on
ve
x
h
u
ll
c
om
p
u
t
at
ion
F
r
o
m
the
r
e
f
ined
bina
r
y
mas
k
I
bin
,
th
e
e
xte
r
na
l
ha
nd
c
on
tou
r
is
e
x
tr
a
c
ted
us
i
ng
bo
r
de
r
f
oll
owin
g.
T
h
e
la
r
g
e
s
t
c
on
to
u
r
b
y
a
r
e
a
is
s
e
lec
te
d
a
s
t
he
ha
nd
b
ou
nd
a
r
y
.
T
he
c
o
nv
e
x
h
ul
l
=
(
)
i
s
t
he
n
c
om
pu
te
d
f
r
o
m
t
he
c
on
to
u
r
=
{
1
,
2
,
.
.
.
,
}
,
f
o
r
mi
ng
th
e
m
in
i
ma
l
c
o
n
ve
x
po
ly
go
n
e
nc
los
in
g
t
he
ha
nd
s
ha
p
e
.
3.
3.
Convexit
y
d
e
f
e
c
t
d
e
t
e
c
t
ion
an
d
in
it
ial
f
il
t
e
r
in
g
L
e
t
=
{
1
,
2
,
.
.
.
,
}
r
e
pr
e
s
e
nt
the
ha
nd
c
ontour
poin
ts
a
nd
=
(
)
it
s
c
onve
x
hull
.
T
he
c
onve
xit
y
de
f
e
c
ts
=
{
1
,
2
,
.
.
.
,
}
a
r
e
c
omput
e
d
a
s
the
s
e
t
of
loca
l
maximum
de
viations
be
twe
e
n
C
a
nd
H
.
E
a
c
h
de
f
e
c
t
d
i
∈
D
is
de
f
ined
a
s
:
=
(
,
,
,
)
whe
r
e
:
a.
,
∈
:
s
tar
t
a
nd
e
nd
point
s
o
f
de
f
e
c
t
i
on
the
c
onve
x
hul
l.
b.
∈
:
de
e
pe
s
t
point
(
f
a
r
thes
t
f
r
om
H
withi
n
de
f
e
c
t
i
)
.
c.
=
(
,
)
:
pe
r
pe
ndicula
r
dis
tanc
e
f
r
om
to
the
li
ne
s
e
gment
T
he
ini
ti
a
l
f
il
ter
ing
f
unc
ti
on
:
→
′
,
(
whe
r
e
′
⊆
)
a
ppli
e
s
two
ge
ometr
ic
c
ons
tr
a
int
s
De
pth
a
nd
Angula
r
C
ons
tr
a
int
:
a.
De
pth
C
ons
tr
a
int
:
δ
i
≥
δ
m
i
n
whe
r
e
δ
m
i
n
is
the
mi
nim
um
de
pth
thr
e
s
hold
b.
Angula
r
C
ons
tr
a
int
:
≤
≤
whe
r
e
=
∠
(
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
,
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
⃗
)
is
the
int
e
r
ior
a
ngle
a
t
.
T
he
f
il
ter
e
d
de
f
e
c
t
s
e
t
′
c
ontains
only
de
f
e
c
ts
s
a
ti
s
f
ying
both
c
ons
tr
a
int
s
,
e
li
mi
na
ti
ng
s
ha
ll
ow
c
ontour
nois
e
a
nd
e
xtr
e
me
indenta
ti
ons
unli
ke
ly
to
r
e
pr
e
s
e
nt
va
ll
e
ys
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
16
,
No.
5
,
Oc
tober
20
26
:
2622
-
2640
2628
Algor
it
hm
2
.
B
e
s
t
-
va
ll
e
y
gr
oup
s
e
lec
ti
on
Input
: All Valid Inter
-
Finger Valleys
V
Output
: Best Valley Combination
V*
1
SelectBestGroup(V)
2
V* ←
∅
; d
min
← +∞;
3
Step 1
: Try all 3
-
valley combinations
4
foreach
triplet I = {i, j, k}
⊆
V, |I| = 3
do
5
{
v
i
, v
j
, v
k
}
← the valley points in I;
6
if
3
points are collinear
then
7
Verify: all valley parallel to
L (angle ≤ α);
8
Compute mean pairwise distance
=
1
3
∑
|
|
−
|
|
ℓ
<
9
if
d
pair
< d
min
then
10
V* ← I, d
min
← d
pair
;
11
Step 2
: If no triplet is found, try all pairs
12
if
V* =
∅
then
13
foreach
pair I = {i, j}
⊆
V, |I| = 2
do
14
{
v
i
, v
j
}
←
valley point in
I;
15
Fit line
L
through
{
v
i
, v
j
};
16
Verify: both start
–
end segments parallel to
L (angle ≤ α);
17
d
pair
← ||v
i
− v
j
||;
18
if
d
pair
< d
min
then
19
V* ← I, d
min
← d
pair
;
20
Step 3
:
If
|V*| = 2,
try adding a third valley
21
if
|V*| = 2
then
22
foreach
candidate
k
∉
V*
do
23
Merge
V*
with candidate
k
to form a triplet
;
24
Fit line
L
through the triplet; Check collinearity;
25
if
collinear, parallelism is valid,
and
compactness is improved
then
26
V* ← V*
∪
{
k
}
, d
min
← d
pair
;
27
break
;
28
return
V*;
3.
4.
Adap
t
ive
ge
om
e
t
r
ic
t
h
r
e
s
h
old
s
f
or
va
ll
e
y
f
i
l
t
e
r
in
g
T
o
r
e
jec
t
non
-
s
igni
f
ica
nt
va
ll
e
ys
(
nois
e
or
s
mall
c
ontour
ir
r
e
gular
it
ies
)
,
we
int
r
oduc
e
ge
ometr
ic
thr
e
s
holds
de
f
ined
r
e
lative
to
the
ha
nd’
s
a
na
tom
ica
l
s
ize
,
e
ns
ur
ing
s
c
a
le
-
invar
ianc
e
a
nd
a
z
e
r
o
-
s
hot
na
tur
e
of
the
method.
T
he
s
e
thr
e
s
holds
a
r
e
c
omput
e
d
f
r
om
t
he
global
ha
nd
c
ontour
dim
e
ns
ions
a
nd
:
a.
:
He
ight
of
the
ha
nd’
s
bound
ing
box
a
f
ter
s
e
gmenta
ti
on.
b.
:
P
a
lm
width,
de
f
ined
a
s
the
hor
izonta
l
dis
tanc
e
be
twe
e
n
the
two
c
ontour
point
s
a
t
the
meta
c
a
r
popha
lange
a
l
joi
nts
(
typi
c
a
ll
y
the
maximu
m
loca
l
width
in
the
uppe
r
ha
lf
of
the
ha
nd
c
ontour
)
.
T
he
mi
nim
um
a
dmi
s
s
ibl
e
va
ll
e
y
de
pth
is
de
f
ined
a
s
:
a.
De
pth
thr
e
s
hold
(
mi
ni
mum
a
dmi
s
s
ibl
e
va
ll
e
y
de
pt
h)
:
δ
m
i
n
=
α
d
e
pt
h
.
H
b
b
ox
,
with
α
d
e
pt
h
=
0
.
02
.
(
2)
F
or
the
mea
n
im
a
ge
r
e
s
olut
ion
of
the
da
tas
e
t
us
e
d
(
640×
480
pixels
[
35]
)
,
thi
s
c
or
r
e
s
ponds
e
mpi
r
ica
ll
y
to
a
ppr
oxim
a
tely
15
pixels
.
b.
W
idt
h
thr
e
s
hold
(
e
s
ti
mating
the
mi
nim
um
va
ll
e
y
width)
:
ℓ
m
i
n
=
β
.
L
pa
lm
,
e
with
β
=
0
.
20
.
(
3)
T
r
e
a
ti
ng
the
va
ll
e
y
c
r
os
s
-
s
e
c
ti
on
a
s
a
tr
iangula
r
r
e
gion
bounde
d
by
the
c
hor
d
a
nd
the
c
ontour
,
the
m
ini
mum
a
dmi
s
s
ibl
e
a
r
e
a
is
:
c.
Ar
e
a
thr
e
s
hold:
A
m
i
n
=
δ
m
i
n
·
ℓ
m
i
n
2
=
δ
.
H
b
b
ox
.
L
pa
lm
,
whe
r
e
δ
=
α
d
ep
th
·
β
2
=
0
.
002
.
(
4)
T
hus
,
A
m
i
n
=
0.
002
H
b
b
o
x
·
L
p
a
l
m
,
whic
h
is
≈
300
pi
xe
ls
2
f
or
a
typi
c
a
l
ha
nd
s
ize
.
T
he
s
e
c
ons
tants
(
α
d
e
p
t
h
a
nd
β
,
he
nc
e
δ
)
we
r
e
f
ixed
onc
e
a
nd
f
or
a
ll
by
de
s
c
r
ipt
i
ve
s
tatis
ti
c
a
l
a
na
lys
is
of
our
da
ta
[
35]
.
T
he
r
a
w
c
onve
xit
y
de
f
e
c
t
de
pths
s
how
a
bim
oda
l
dis
tr
ibu
ti
on,
with
c
ontour
a
r
ti
f
a
c
ts
c
e
nter
e
d
be
low
1%
of
H
bbox
a
nd
tr
ue
int
e
r
-
f
inger
va
ll
e
ys
a
bove
5%
.
W
e
c
hos
e
δ
d
e
p
t
h
=
2%
a
s
a
r
obus
t
s
e
pa
r
a
tor
.
S
im
il
a
r
ly,
a
na
lys
is
of
the
c
los
e
d
-
polygon
a
r
e
a
s
(
via
the
s
hoe
lac
e
f
or
mul
a
)
r
e
ve
a
ls
a
c
lea
r
ga
p
be
twe
e
n
non
-
phys
ica
l
indenta
ti
on
s
(
<
0.
1%
of
H
b
b
o
x
.
L
p
a
l
m
)
a
nd
plaus
ibl
e
va
ll
e
ys
(
>
0.
4%
)
.
S
e
tt
in
g
δ
=
0.
2%
e
li
mi
na
tes
ove
r
90%
of
s
pur
ious
de
f
e
c
ts
while
r
e
taining
≈
98%
of
va
li
d
va
ll
e
ys
.
T
he
s
e
thr
e
s
holds
r
e
main
c
ons
tant
a
c
r
os
s
a
ll
e
xp
e
r
im
e
nts
(
no
pe
r
-
s
u
bjec
t
or
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
C
ontour
-
guided
c
onv
e
x
it
y
de
fec
t
ge
ome
tr
y
for
pr
e
c
is
e
R
OI
e
x
tr
ac
ti
on
in
…
(
Habib
K
ade
m
)
2629
pe
r
-
s
e
s
s
ion
tuni
ng)
,
pr
e
s
e
r
ving
the
f
ull
y
ge
ometr
ic
,
z
e
r
o
-
s
hot
na
tur
e
of
C
G
-
C
DG
.
Ac
c
or
dingl
y,
a
ny
c
a
ndidate
va
ll
e
y
of
de
pth
a
nd
a
r
e
a
is
r
e
taine
d
if
:
≥
(
0
.
02
.
≈
15
on
da
tas
e
t
[
35]
)
a
nd
≥
(
0
.
0
0
2
⋅
⋅
≈
300
2
)
on
the
da
tas
e
t
[
35
]
(
5)
other
wis
e
,
it
is
d
is
c
a
r
de
d.
T
his
a
da
pti
ve
c
r
it
e
r
ion
e
f
f
e
c
ti
ve
ly
r
e
moves
s
ha
ll
ow
c
ontour
a
r
ti
f
a
c
ts
(
s
kin
textur
e
,
ve
in
s
ha
dows
)
while
pr
e
s
e
r
ving
a
ll
a
na
tom
ica
ll
y
va
li
d
int
e
r
-
f
inge
r
va
ll
e
ys
,
e
ve
n
f
or
s
mall
ha
nds
.
W
hil
e
thes
e
thr
e
s
holds
a
r
e
gr
ounde
d
in
a
na
tom
ica
l
invar
iants
a
nd
nor
malize
d
by
ha
nd
s
c
a
le,
we
a
c
knowle
dge
that
their
numer
ica
l
va
lues
we
r
e
de
r
ived
f
r
om
the
s
a
me
da
tas
e
t
us
e
d
f
or
pe
r
f
o
r
manc
e
e
va
luation.
T
his
c
ons
ti
tut
e
s
a
li
mi
tat
ion
in
he
r
e
nt
to
the
a
bs
e
nc
e
of
a
s
e
pa
r
a
te
c
a
li
br
a
ti
on
c
or
pus
.
F
ut
ur
e
wor
k
will
va
li
da
te
thes
e
c
ons
tants
on
a
n
ind
e
pe
nde
nt
be
nc
hmar
k
to
r
ule
out
a
ny
da
ta
-
s
nooping
bias
.
3.
5.
P
ar
a
m
e
t
e
r
s
e
lec
t
ion
an
d
s
e
n
s
it
ivi
t
y
an
alys
i
s
T
he
a
ngular
tol
e
r
a
nc
e
pa
r
a
mete
r
α
plays
a
c
r
i
ti
c
a
l
r
ole
in
s
tabili
z
ing
va
ll
e
y
gr
oup
s
e
lec
ti
on
unde
r
pos
e
va
r
iations
.
I
n
dor
s
a
l
ha
nd
ve
in
a
c
quis
it
ion,
wr
is
t
r
otation
a
nd
f
inger
s
pr
e
a
d
int
r
oduc
e
or
ienta
t
ion
nois
e
that
c
a
n
lea
d
to
a
na
tom
ica
ll
y
i
mpl
a
us
ibl
e
c
onve
xit
y
de
f
e
c
ts
be
ing
s
e
lec
ted.
B
y
c
ons
tr
a
ini
ng
va
ll
e
y
or
ienta
ti
on
withi
n
a
bounde
d
a
ngular
r
a
nge
,
C
G
-
C
DG
e
xpli
c
it
ly
f
il
ter
s
pos
e
-
induce
d
a
r
ti
f
a
c
ts
r
a
ther
than
r
e
lyi
ng
on
loca
l
e
xtr
e
ma
a
lone.
T
he
a
ng
ula
r
to
ler
a
nc
e
α
wa
s
e
xpe
r
im
e
ntal
ly
de
te
r
mi
ne
d
thr
ough
g
r
i
d
s
e
a
r
c
h
o
ve
r
[
20
°,
6
0°
]
wit
h
5°
incr
e
ments
.
F
igu
r
e
3
s
hows
that
α
=
3
0°
o
pti
mi
z
e
s
the
t
r
a
de
-
o
f
f
be
t
we
e
n
va
ll
e
y
r
e
jec
t
ion
(
t
oo
s
t
r
ic
t)
a
nd
f
a
ls
e
pos
it
i
ve
incl
us
ion
(
to
o
pe
r
mi
s
s
ive
)
,
a
c
hiev
ing
s
ta
ble
E
E
R
a
c
r
os
s
99
.
1
%
o
f
tes
t
c
a
s
e
s
.
S
e
ns
it
ivi
ty
a
na
lys
is
c
onf
ir
ms
that
the
p
r
op
os
e
d
me
thod
r
e
ma
ins
s
tab
le
f
or
α
va
l
ue
s
wit
hin
thi
s
r
a
nge
,
de
mons
t
r
a
t
ing
that
C
G
-
C
DG
doe
s
n
ot
de
pe
nd
on
f
ine
pa
r
a
me
ter
t
unin
g
a
nd
e
xh
i
bit
s
r
o
bus
t
be
ha
vio
r
a
c
r
os
s
va
r
yi
ng
a
c
qu
is
it
i
on
c
on
dit
i
ons
.
F
igur
e
3.
P
a
r
a
mete
r
s
e
ns
it
ivi
ty
a
na
lys
is
of
the
p
r
op
os
e
d
method
3.
6.
Geom
e
t
r
ic
vali
d
at
ion
an
d
vall
e
y
gr
ou
p
s
e
le
c
t
ion
F
oll
owing
ini
ti
a
l
c
onve
xit
y
de
f
e
c
t
de
tec
ti
on,
our
pipeline
a
ppli
e
s
dua
l
ge
ometr
ic
c
r
it
e
r
ia
to
f
il
ter
a
na
tom
ica
ll
y
plaus
ibl
e
va
ll
e
ys
f
r
om
s
pur
ious
c
ontour
a
r
t
if
a
c
ts
a
nd
then
s
e
lec
ts
the
opt
im
a
l
va
ll
e
y
c
onf
igur
a
ti
on
f
o
r
r
obus
t
R
OI
a
li
gn
ment.
F
igur
e
4
p
r
e
s
e
nts
the
ove
r
a
ll
pipeline,
a
s
il
lus
tr
a
ted
in
F
igur
e
4(
a
)
f
o
r
the
va
ll
e
y
va
li
da
ti
on
modul
e
,
F
igu
r
e
4
(
b)
f
or
th
e
gr
oup
s
e
lec
ti
on
modul
e
,
a
nd
F
igur
e
4(
c
)
f
or
t
he
R
OI
e
xtr
a
c
ti
on
modul
e
.
E
a
c
h
c
a
ndidate
de
f
e
c
t
unde
r
go
e
s
two
va
li
da
ti
on
tes
ts
:
a.
C
hor
d
-
c
ontour
int
e
r
s
e
c
ti
on
:
T
he
ba
s
e
li
ne
s
e
gment
c
onne
c
ti
ng
de
f
e
c
t
e
ndpoint
s
mus
t
int
e
r
s
e
c
t
the
ha
nd
c
ontour
a
t
e
xa
c
tl
y
two
point
s
(
int
e
r
s
e
c
ti
on
pa
r
a
mete
r
0
≤
t
≤
1)
,
e
ns
ur
ing
a
tr
ue
int
e
r
-
knuc
kle
va
ll
e
y
r
a
ther
than
s
upe
r
f
icia
l
c
ontour
nois
e
.
b.
M
ini
mum
a
r
e
a
thr
e
s
hold
:
T
he
polygon
f
o
r
med
by
the
c
hor
d
a
nd
int
e
r
ve
ning
c
ontou
r
mus
t
e
nc
los
e
a
n
a
r
e
a
≥
m
i
n
(
e
mpi
r
ica
ll
y
s
e
t
to
a
ppr
oxim
a
tely
5
%
–
10%
o
f
t
he
pa
lm
width)
,
ther
e
by
e
li
m
inating
non
-
phys
ica
l
indenta
ti
ons
.
T
he
a
r
e
a
is
c
omput
e
d
us
ing
the
S
hoe
l
a
c
e
f
or
mul
a
:
=
1
2
|
∑
+
1
−
1
=
1
+
1
−
∑
+
1
−
1
=
1
−
1
|
(
6)
Only
de
f
e
c
ts
s
a
ti
s
f
ying
both
c
r
it
e
r
ia
a
r
e
r
e
taine
d
a
s
va
li
da
ted
va
ll
e
ys
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
S
N
:
2088
-
8708
I
nt
J
E
lec
&
C
omp
E
ng
,
Vol
.
16
,
No.
5
,
Oc
tober
20
26
:
2622
-
2640
2630
Va
ll
e
y
gr
oup
s
e
lec
ti
on:
va
li
da
ted
va
ll
e
ys
a
r
e
a
na
lyze
d
us
ing
ge
ometr
ic
c
ohe
r
e
nc
e
c
r
i
ter
ia
a
nd
pa
r
a
ll
e
li
s
m
thr
ough
a
hier
a
r
c
hica
l
thr
e
e
s
tage
s
e
a
r
c
h
s
tr
a
tegy.
T
he
a
lgor
i
thm
p
r
ior
it
ize
s
3
va
ll
e
y
g
r
oups
f
or
a
na
tom
ica
l
c
ompl
e
tene
s
s
,
f
a
ll
s
ba
c
k
to
2
va
ll
e
y
p
a
ir
s
whe
n
ne
c
e
s
s
a
r
y,
a
nd
a
tt
e
mpt
s
int
e
ll
igent
pa
ir
-
to
-
tr
ipl
e
t
e
xpa
ns
ion.
T
he
gr
oups
mus
t
s
a
ti
s
f
y
a
ge
ometr
ic
c
ompac
tnes
s
thr
e
s
hold.
˗
Va
ll
e
y
s
ur
f
a
c
e
va
li
da
ti
on
modul
e
,
a
s
s
hown
in
F
ig
ur
e
4(
a
)
:
thi
s
c
omponent
pr
oc
e
s
s
e
s
c
onve
xit
y
de
f
e
c
ts
to
identif
y
a
na
tom
ica
ll
y
va
li
d
int
e
r
-
f
inger
va
l
leys
.
F
or
e
a
c
h
de
f
e
c
t,
it
(
a
)
e
xtr
a
c
ts
the
c
ontour
s
e
gment
be
twe
e
n
s
tar
t
a
nd
e
nd
point
s
to
f
or
m
a
polygon
a
l
s
ur
f
a
c
e
,
(
b
)
c
omput
e
s
the
e
nc
los
e
d
a
r
e
a
us
in
g
the
s
hoe
lac
e
f
or
mul
a
,
a
nd
(
c
)
a
ppli
e
s
a
mi
ni
mum
a
r
e
a
thr
e
s
hold
f
il
ter
(
a
r
e
a
≥
thr
e
s
hold
px2)
to
r
e
tain
only
phys
ica
ll
y
plaus
ibl
e
va
ll
e
y
s
ur
f
a
c
e
s
while
r
e
jec
ti
ng
s
ha
ll
ow
c
ontour
a
r
t
if
a
c
ts
.
˗
Va
ll
e
y
gr
oup
s
e
lec
ti
on
modul
e
a
s
s
hown
in
F
igur
e
4(
b)
:
the
a
lgo
r
it
hm
e
mpl
oys
a
hier
a
r
c
hica
l
th
r
e
e
-
s
tage
s
tr
a
tegy
to
identif
y
the
opti
mal
va
ll
e
y
c
onf
igur
a
ti
on:
(
1)
pr
i
mar
y
s
e
a
r
c
h
f
or
qua
li
f
ying
3
-
va
ll
e
y
gr
oups
s
a
ti
s
f
ying
ge
ometr
ic
c
ons
tr
a
int
s
of
a
li
gnment,
pa
r
a
ll
e
li
s
m,
a
nd
c
ompac
tnes
s
;
(
2)
f
a
ll
ba
c
k
to
2
-
va
ll
e
y
pa
ir
s
if
no
va
li
d
tr
ipl
e
t
is
f
ound;
(
3)
e
xpa
ns
ion
pha
s
e
that
a
tt
e
mpt
s
to
upg
r
a
de
va
li
d
pa
ir
s
to
tr
ipl
e
t
s
by
incor
por
a
ti
ng
a
ddit
ional
c
oll
inea
r
c
a
ndidate
s
.
S
e
le
c
ti
on
pr
io
r
it
ize
s
gr
oups
with
mi
ni
mi
z
e
d
mea
n
pa
i
r
wis
e
dis
tanc
e
,
f
a
vor
ing
t
ight
ly
c
lus
ter
e
d
va
ll
e
y
f
or
matio
ns
f
or
s
table
R
OI
or
ienta
ti
on
e
s
ti
mation
.
˗
R
oi
e
xt
r
a
c
t
i
on
m
od
ul
e
a
s
s
ho
wn
in
F
ig
u
r
e
4
(
c
)
:
t
his
m
od
ul
e
t
r
a
ns
f
o
r
ms
va
li
da
ted
va
l
le
ys
i
nt
o
a
s
ta
nd
a
r
d
iz
e
d
R
OI
t
h
r
o
ug
h
g
e
o
me
t
r
ic
n
o
r
ma
l
iza
ti
on
:
c
o
mp
ut
in
g
t
h
e
h
a
n
d
c
e
n
t
r
o
id
,
de
te
r
mi
ni
ng
o
r
i
e
n
ta
ti
on
v
ia
le
a
s
t
-
s
qua
r
e
s
va
l
ley
f
it
t
in
g
,
r
ot
a
t
i
ng
to
a
l
ig
n
w
it
h
c
o
o
r
d
in
a
t
e
a
x
e
s
,
a
nd
c
r
op
pi
ng
a
s
q
ua
r
e
r
e
gi
on
c
e
n
te
r
e
d
a
t
t
he
c
e
n
t
r
o
id
.
T
he
o
u
tp
ut
p
r
o
vi
de
s
c
ons
is
ten
t
,
or
ie
nt
a
t
io
n
-
n
o
r
m
a
l
ize
d
r
o
is
f
o
r
r
e
l
iab
le
f
e
a
t
u
r
e
e
xt
r
a
c
t
i
on
.
(
a
)
(
b)
(
c
)
F
igur
e
4.
T
he
C
G
-
C
DG
pipeline:
modul
e
s
f
or
do
r
s
a
l
ha
nd
ve
in
va
ll
e
y
s
e
lec
ti
on
a
nd
R
OI
e
xtr
a
c
ti
on
(
a
)
va
ll
e
y
va
li
da
ti
on
modul
e
,
(
b)
gr
oup
s
e
lec
ti
on
modul
e
,
a
n
d
(
c
)
R
OI
e
xtr
a
c
ti
on
modul
e
.
3.
7.
Valley
gr
ou
p
s
e
lec
t
ion
via
ge
om
e
t
r
ic
c
oh
e
r
e
n
c
e
T
o
is
olate
the
s
pe
c
if
ic
s
e
t
of
knuc
kles
,
we
e
mpl
oy
a
n
int
e
ll
igent
gr
oup
s
e
lec
ti
on
a
lgor
it
hm
that
e
va
luate
s
c
ombi
na
ti
ons
of
va
li
d
va
ll
e
ys
ba
s
e
d
on
t
wo
pr
inciples
of
ge
ometr
ic
c
ohe
r
e
nc
e
:
a.
C
oll
inea
r
it
y:
Knuc
kles
a
r
e
e
xpe
c
ted
to
be
a
ppr
ox
im
a
tely
a
li
gne
d.
T
his
is
ve
r
if
ied
by
f
it
ti
ng
a
li
ne
to
a
c
a
ndidate
gr
oup
o
f
va
ll
e
ys
a
nd
e
ns
ur
ing
the
pe
r
pe
ndicula
r
dis
tanc
e
of
e
a
c
h
va
ll
e
y
to
the
li
ne
is
wit
hin
a
s
mall
tol
e
r
a
nc
e
.
b.
P
a
r
a
ll
e
l
is
m:
T
he
o
r
ien
tati
on
o
f
the
va
ll
e
ys
mus
t
be
c
ons
is
ten
t.
F
or
a
ny
two
va
ll
e
ys
,
th
e
a
lgo
r
i
thm
c
o
m
pa
r
e
s
the
a
ngle
o
f
the
ir
ba
s
e
li
ne
s
e
g
ments
.
T
he
mi
n
im
u
m
a
ng
ula
r
di
f
f
e
r
e
nc
e
.
=
(
|
1
−
2
|
,
180°
−
|
1
−
2
|
)
≤
(
7)
mus
t
not
e
xc
e
e
d
a
pr
e
de
f
ined
tol
e
r
a
nc
e
α
(
s
e
t
to
30
°)
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
nt
J
E
lec
&
C
omp
E
ng
I
S
S
N:
2088
-
8708
C
ontour
-
guided
c
onv
e
x
it
y
de
fec
t
ge
ome
tr
y
for
pr
e
c
is
e
R
OI
e
x
tr
ac
ti
on
in
…
(
Habib
K
ade
m
)
2631
3.
8.
ROI
p
ar
am
e
t
e
r
d
e
t
e
r
m
in
a
t
ion
T
he
ge
ometr
ica
ll
y
va
li
da
ted
va
ll
e
y
gr
oup
e
s
tabli
s
he
s
the
pa
r
a
mete
r
s
f
or
pr
e
c
is
e
R
OI
e
xtr
a
c
ti
on
thr
ough
thr
e
e
c
omput
a
ti
ona
l
s
teps
:
a.
Or
ienta
ti
on
e
s
ti
mation:
A
lea
s
t
-
s
qua
r
e
s
li
ne
f
it
ted
to
the
s
e
lec
ted
va
ll
e
ys
yields
the
o
r
ienta
ti
on
a
ngle
ϕ
,
f
r
om
i
ts
dir
e
c
ti
ona
l
ve
c
tor
(
dx
,
dy
):
=
2
(
,
)
(
8)
b.
I
mage
nor
maliza
ti
on:
T
he
i
mage
is
r
otate
d
by
−
ϕ
a
bout
the
ha
nd
c
e
ntr
oid
(
C
x
,
C
y
)
via
the
a
f
f
ine
matr
ix
R
,
a
li
gning
the
va
ll
e
y
a
xis
ho
r
izonta
ll
y:
=
g
e
tRo
ta
ti
o
nM
a
tr
ix2
D
(
(
,
)
,
−
,
1
.
0
)
(
9)
c.
R
e
gion
e
xtr
a
c
ti
on:
A
s
qua
r
e
R
OI
is
c
r
oppe
d,
c
e
nter
e
d
on
the
c
e
ntr
oid
(
C
x
,
C
y
)
o
f
the
ha
nd
mas
k,
with
s
ide
length:
L
=
0
.
95
×
m
in
(
W
b
b
ox
,
H
b
b
ox
)
(
10)
W
b
b
o
x
a
nd
H
b
b
o
x
a
r
e
the
width
a
nd
he
ight
,
r
e
s
pe
c
ti
ve
ly,
o
f
the
bounding
box
(
a
xis
-
a
li
gne
d
r
e
c
tangle
)
a
r
ound
the
s
e
gmente
d
ha
nd.
T
he
95%
f
a
c
tor
wa
s
s
e
t
e
mpi
r
ica
ll
y
to
r
e
tain
pe
r
ipher
a
l
va
s
c
ular
de
tail
while
e
xc
ludi
ng
ba
c
kgr
ound
nois
e
a
nd
s
e
gmenta
ti
on
a
r
ti
f
a
c
ts
;
the
5%
mar
gin
a
bs
or
bs
r
e
s
idual
c
ontour
ir
r
e
g
ular
it
ies
.
C
ombi
ne
d
with
the
a
da
pti
ve
de
f
e
c
t
f
il
ter
ing
(
s
e
c
ti
on
3.
4
)
a
nd
ge
ometr
ic
va
ll
e
y
gr
oup
s
e
lec
ti
on
(
s
e
c
t
ion
3
.
7)
,
thi
s
tr
a
ini
ng
-
f
r
e
e
,
thr
e
e
-
s
tep
pipeline
yields
c
on
s
is
t
e
nt,
or
ienta
ti
on
-
nor
malize
d
R
OI
s
r
obus
t
to
pos
e
,
s
c
a
le,
a
nd
plac
e
ment
va
r
iation.
3.
9.
Com
p
u
t
at
io
n
al
c
om
p
lexit
y
L
e
t
W
×
H
de
note
the
input
im
a
ge
r
e
s
olut
ion,
N
the
number
of
e
xtr
a
c
ted
c
ontour
point
s
,
M
the
number
of
r
a
w
c
onve
xit
y
de
f
e
c
ts
pa
s
s
ing
the
de
pth
f
il
ter
,
k
the
a
ve
r
a
ge
loca
l
c
ontour
-
s
e
gment
length
e
xa
mi
ne
d
pe
r
c
a
ndidate
de
f
e
c
t
dur
ing
ge
ometr
ic
va
li
da
ti
on,
a
nd
|
V
|
the
number
o
f
ge
ometr
ica
ll
y
v
a
li
da
ted
va
ll
e
ys
.
P
r
e
pr
oc
e
s
s
ing
(
Ga
us
s
ian
s
moot
hing,
Ots
u
thr
e
s
holdi
ng,
mo
r
phologi
c
a
l
c
los
ing/
ope
ning)
is
O
(
W
·
H
)
;
c
ontour
e
xtr
a
c
ti
on
is
O
(
N
)
;
c
onve
x
-
hull
c
omput
a
ti
on
is
O
(
N
log
N
)
;
c
onve
xit
y
-
de
f
e
c
t
de
tec
ti
on
is
O
(
N
)
;
c
hor
d
–
c
ontour
int
e
r
s
e
c
ti
on
a
nd
s
hoe
lac
e
-
a
r
e
a
va
li
da
ti
on
(
6
)
is
O
(
M
·
k
)
wi
th
k
≪
N
;
a
nd
hier
a
r
c
hica
l
va
ll
e
y
-
gr
oup
s
e
lec
ti
on
(
Algor
it
hm
2)
is
O
(|
V
|
3
)
in
the
wo
r
s
t
c
a
s
e
,
with
|
V
|
a
na
tom
ica
ll
y
bounde
d
to
a
s
mall
c
ons
tant
(
typi
c
a
ll
y
≤
10)
.
T
he
ov
e
r
all
c
ompl
e
xit
y
is
ther
e
f
or
e
,
(
·
+
+
·
+
|
|
3
)
(
11)
domi
na
ted
in
pr
a
c
ti
c
e
by
the
O(
W
·
H)
pr
e
pr
oc
e
s
s
ing
ter
m,
c
ons
is
tent
with
the
ne
a
r
-
r
e
a
l
-
ti
me
r
u
nti
me
of
44
3
.
61
pe
r
i
mage
mea
s
ur
e
d
on
De
ll
L
a
ti
tude
5520
laptop
r
unning
Ubuntu
24
.
04
L
T
S
,
e
quipped
with
a
n
I
ntel®
C
or
e
™
i7
-
1185G7
pr
oc
e
s
s
or
a
nd
16
GB
of
R
AM
.
4.
E
XP
E
RI
M
E
NT
AL
RE
S
U
L
T
S
AN
D
DI
S
CU
S
S
I
ON
T
his
s
e
c
ti
on
pr
e
s
e
nts
a
c
ompr
e
he
ns
ive
e
va
luation
of
R
OI
e
xtr
a
c
ti
on
qua
li
ty
us
ing
f
e
a
tu
r
e
dis
tanc
e
a
na
lys
is
a
s
a
n
objec
ti
ve
pr
oxy
f
or
R
OI
c
ons
is
tenc
y
a
nd
dis
c
r
im
inabili
ty
.
4.
1.
E
xp
e
r
im
e
n
t
al
S
e
t
u
p
T
he
pr
opos
e
d
C
G
-
C
D
G
R
OI
method
wa
s
e
va
l
ua
ted
on
a
publi
c
ly
a
va
il
a
ble
dor
s
a
l
ha
nd
ve
in
da
taba
s
e
[
35]
c
ompr
is
ing
1,
024
im
a
ge
s
f
r
om
1
38
s
ubjec
ts
(
r
ight
a
nd
lef
t
ha
nds
)
.
T
he
da
tas
e
t
pr
e
s
e
nts
s
igni
f
ica
nt
c
ha
ll
e
nge
s
,
including
va
r
iations
in
il
lu
mi
na
ti
on,
pos
e
,
a
nd
s
kin
textur
e
.
E
a
c
h
s
ubjec
t
c
o
ntr
ibut
e
s
f
our
im
a
ge
s
of
the
lef
t
ha
nd
a
nd
f
our
i
mage
s
of
the
r
ight
ha
nd.
Due
to
the
inher
e
nt
a
na
tom
i
c
a
l
a
nd
mor
phologi
c
a
l
dif
f
e
r
e
nc
e
s
be
twe
e
n
the
two
ha
nds
of
the
s
a
me
indi
vidual,
lef
t
a
nd
r
igh
t
ha
nds
a
r
e
tr
e
a
ted
a
s
indepe
nde
nt
biom
e
tr
ic
ins
tanc
e
s
.
Ac
c
or
dingl
y,
a
ll
e
xpe
r
im
e
nts
a
r
e
c
onduc
ted
in
a
s
ubje
c
t
-
wi
s
e
mann
e
r
,
with
matc
hing
a
nd
dis
tanc
e
c
omput
a
ti
ons
pe
r
f
or
med
s
e
pa
r
a
tely
f
or
lef
t
-
ha
nd
a
nd
r
ight
-
ha
nd
im
a
ge
s
.
W
it
h
f
our
im
a
ge
s
pe
r
ha
nd
pe
r
s
ubjec
t,
138
s
ubjec
t
s
a
nd
2
ha
nds
(
tr
e
a
ted
a
s
indepe
nde
nt
in
s
tanc
e
s
)
yield
C
(
4,
2)
=
6
ge
nuine
c
ompar
is
ons
pe
r
s
ubjec
t
-
ha
nd,
i.
e
.
,
1
,
656
ge
nuine
c
ompar
is
ons
in
tot
a
l;
im
pos
tor
c
ompar
i
s
ons
a
r
e
dr
a
wn
be
twe
e
n
dif
f
e
r
e
nt
s
ubjec
ts
withi
n
the
s
a
me
ha
nd
s
ide.
Evaluation Warning : The document was created with Spire.PDF for Python.