Internati
o
nal
Journal of Ele
c
trical
and Computer
Engineering
(IJE
CE)
V
o
l.
5, N
o
. 5
,
O
c
tob
e
r
201
5, p
p
. 1
202
~121
5
I
S
SN
: 208
8-8
7
0
8
1
202
Jo
urn
a
l
h
o
me
pa
ge
: h
ttp
://iaesjo
u
r
na
l.com/
o
n
lin
e/ind
e
x.ph
p
/
IJECE
Segm
entation of
Fingerprint Image B
a
s
e
d on Gradient
Magnitude and Coherence
Sap
a
ru
din
1
an
d Gh
az
ali Sul
on
g
2
1
Faculty
of Computer Scie
nce, Sr
iwijay
a
University
, South
Sumatera, Indonesia
2
Faculty
of Computing, University
Tec
hnolog
y
M
a
lay
s
ia, Johor B
a
hru, M
a
lay
s
ia
saparudin1204@
y
a
hoo
.com
1
,
gha
zal
i@s
p
aceu
tm
.edu.m
y
2
Article Info
A
B
STRAC
T
Article histo
r
y:
Received
J
u
n 12, 2015
Rev
i
sed
Au
g
20
, 20
15
Accepted Aug 26, 2015
Fingerprint image segmentatio
n is an
import
a
nt pre-processing step in
automatic fingerprint recognition s
y
stem. A
well-design
ed
fingerprin
t
segmentation technique
can improve the
a
c
c
u
ra
cy
i
n
c
o
ll
ec
t
i
ng cl
ea
r
fingerprin
t
ar
ea
and m
a
rk no
is
e are
a
s.
The
tradi
tiona
l gre
y
v
a
rian
c
e
segmentation m
e
thod is widely
and eas
ily
used, but it can hard
ly
segmen
t
fingerprin
t
s with low contrast of
high noise. To overcome th
e low image
contrast,
combining two-blo
c
k
feature;
me
a
n
o
f
g
r
a
d
i
e
nt
ma
g
n
i
t
u
de
a
nd
coheren
c
e, where the fingerpr
i
nt image is se
gmented into background
,
foreground or n
o
is
y
r
e
gions, h
a
s been done. Ex
cept for
the nois
y
r
e
gions in
the for
e
ground,
there are still su
ch noi
ses ex
isted in th
e backgro
und whose
coheren
ces
are
low, and
ar
e mistaken
ly
assign
ed as for
e
groun
d. A novel
segmentation method based on combinati
on lo
cal mean of grey
-scale and
local variance of
gradient magnitude is
presented
in this paper
.
Th
e proposed
extra
c
tion beg
i
n
s
with norm
a
lization of the fing
erprint
.
Then
, it
is
followed
b
y
foreground r
e
gion separation
from the background. Finally
,
the gradien
t
coheren
ce
appr
oach is
us
ed t
o
dete
ct th
e n
o
is
e regions
ex
is
ted in th
e
foreground.
Exp
e
rimental results on NI
ST-Database14 f
i
ngerp
rint images
indicate
th
at the proposed
metho
d
gives
the impr
essive results.
Keyword:
Fi
nge
r
p
ri
nt
seg
m
ent
a
t
i
on
Gra
d
i
e
nt
m
a
gn
i
t
ude c
ohe
re
nc
e
Grey varia
n
ce method
Copyright ©
201
5 Institut
e
o
f
Ad
vanced
Engin
eer
ing and S
c
i
e
nce.
All rights re
se
rve
d
.
Co
rresp
ond
i
ng
Autho
r
:
Sapa
ru
di
n
,
Depa
rtem
ent of Inform
atic Engi
neeri
n
g,
Uni
v
ersitas Sri
w
ijaya,
Jalan
Raya Pal
e
m
b
an
g-
Pr
abum
u
l
ih
, K
m
. 32
, In
dr
alaya, Ogan
I
lir
.
Em
a
il: sap
a
rudin
1
2
04@yahoo
.co
m
1.
INTRODUCTION
Fi
nge
r
p
ri
nt
se
gm
ent
a
t
i
on i
s
a t
echni
q
u
e
i
n
w
h
i
c
h
fe
at
ures s
h
ari
n
g
or
regi
on
s
wi
t
h
si
m
i
l
a
r
ch
aracteristics are id
en
tified an
d gr
oup
ed co
llectiv
ely. In
o
t
h
e
r
words, th
e seg
m
en
tatio
n
is sp
littin
g
t
h
e
f
i
ng
erp
r
i
n
t imag
e in
t
o
t
w
o reg
i
on
s, wh
ich
ar
e called
fo
r
e
g
r
ou
nd
and
b
a
ck
gro
und
r
e
g
i
o
n
s
. Th
e
fo
r
e
gr
oun
d
regi
ons c
o
r
r
es
po
n
d
t
o
cl
ear
fi
nge
rp
ri
nt
ar
eas cont
ai
ni
ng
ri
dges a
n
d v
a
l
l
e
y
s
, whi
l
e
back
g
r
o
u
nd
re
gi
o
n
s
cor
r
es
po
n
d
t
o
regi
ons
o
u
t
s
i
d
e b
o
r
de
rs o
f
fi
n
g
er
pri
n
t
ar
ea, w
h
i
c
h
do
not
c
ont
ai
n
any
val
i
d
fi
n
g
e
rp
ri
nt
i
n
f
o
rm
at
i
on. I
f
bac
k
g
r
o
u
n
d
regi
o
n
s ha
ve
uni
f
o
rm
grey
-l
evel
an
d are
l
i
ght
er t
h
an
fo
reg
r
ou
n
d
, t
h
en a
n
app
r
oach
base
d o
n
l
o
cal
i
n
t
e
nsi
t
y
coul
d be
effect
i
v
e
fo
r se
parat
i
n
g t
h
e
fo
reg
r
o
u
nd
fr
om
t
h
e bac
k
g
r
ou
n
d
,
bu
t
in
p
r
actice,
fing
erp
r
i
n
t seg
m
en
tatio
n
is sen
s
itiv
e to
th
e qu
al
ity o
f
fing
erp
r
i
n
t i
m
ag
e. Thu
s
, th
e low qu
ality o
f
fi
n
g
er
pri
n
t
i
m
age i
s
a
p
r
o
b
l
e
m
,
w
h
i
c
h
re
q
u
i
r
e
s
m
o
re r
o
b
u
st
s
e
gm
ent
a
t
i
on t
echni
que
s [
1
1]
[
18]
.
The first proble
m
is
the prese
n
ce of
noise
t
h
at
resulte
d from
dust and
gre
a
se on t
h
e s
u
rface of live
-
scan fingerpri
n
t
scanne
rs or
ink-on-p
ape
r
rol
l
ed fi
nge
r
p
ri
nt
.
The sec
o
nd
p
r
oblem
is false tracings in the i
m
age
acq
u
i
sition
.
The th
ird
p
r
o
b
l
em is th
e d
r
yn
ess o
r
wetn
ess t
h
at can
influ
e
nce th
e q
u
a
lity o
f
ridg
es and
v
a
lleys
structure. T
h
e
last one is the prese
n
ce of an indistinct
bo
u
nda
ry
w
h
en t
h
e feat
ures i
n
w
i
nd
ows
of
fi
xe
d si
ze
are use
d
. D
u
e t
o
t
h
ese p
r
o
b
l
e
m
s
, separat
i
ng noi
se re
gi
o
n
s f
r
om
t
h
e fore
gr
ou
n
d
re
gi
o
n
s are nee
d
ed
. Det
ect
i
o
n
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 5
,
N
o
. 5
,
O
c
tob
e
r
20
15
:
120
2
–
12
15
1
203
of noise re
gions in the fore
ground
regi
ons is a challenging problem
in fi
ngerpri
nt segm
e
n
tation,
becaus
e
the
q
u
a
lity of th
e fin
g
e
rprin
t
area
g
r
eatly con
t
ri
bu
tes to im
p
r
o
v
e th
e
qu
ality o
f
ridg
es
or
v
a
lleys d
i
rection
s
.
Fig
u
re
1 sh
o
w
s a fi
n
g
er
pri
n
t
im
age and i
t
s
segm
ent
a
t
i
on
resul
t
t
h
at
consi
s
t
s
of
back
g
r
o
u
n
d
regi
o
n
s
,
f
o
re
gr
o
u
n
d
regi
ons
, a
n
d
n
o
i
se regi
on
s.
There
are
t
w
o gene
ral
t
y
pes
o
f
feat
ure
s
used
fo
r fi
n
g
er
pri
n
t
segm
ent
a
t
i
on,
i
.
e., bl
ock
-
wi
se
and
pi
xel
-
wi
se feat
u
r
es.
Gene
ral
l
y
, pi
x
e
l
-
wi
se feat
ure
of
fi
n
g
er
pri
n
t
segm
ent
a
t
i
on pr
ovi
des acc
urat
e re
sul
t
s
,
but
i
t
s
com
put
at
i
onal
com
p
l
e
xi
t
y
and t
i
m
e consum
i
ng a
r
e m
a
rked
l
y
hi
ghe
r t
h
an
m
o
st
of bl
ock
-
wi
se feat
ures
.
Si
nce
pi
xel
-
wi
se
bas
e
d se
gm
ent
a
t
i
on m
e
t
hod i
s
t
e
di
o
u
s a
nd t
i
m
e co
nsum
i
ng,
bl
ock
-
wi
se feat
u
r
es are
m
o
re w
i
del
y
u
s
ed
in au
t
o
m
a
tic fin
g
e
rprin
t
reco
gn
itio
n syste
m
s [2
] [18
]
.
(a)
(b)
Fi
gu
re 1.
A sam
p
le o
f
fing
erp
r
i
n
t seg
m
en
tatio
n
resu
lts; (a) orig
in
al
fing
erp
r
i
n
t im
ag
e,
(
b
)
b
ackgr
oun
d, fo
r
e
gr
oun
d and
n
o
i
se p
a
tch
e
s of
th
e seg
m
en
ted
f
i
ng
erp
r
i
n
t imag
e
2.
RELATED WORK
Sev
e
ral features u
s
ed
in
fi
ng
erp
r
i
n
t seg
m
en
tati
o
n
are kno
wn
fro
m
liter
a
tu
res, su
ch
as g
r
ey-scale
statistical,
lo
cal d
i
rectio
n
a
lity, an
d
co
nsistency o
f
o
r
ien
t
atio
n. Detailed
ex
p
l
an
atio
ns
o
f
th
ese features are as
fo
llows.
2.
1
Grey-Sc
a
le Statis
tical Fe
atures
Grey-s
cale statistical features
in fingerpri
n
t se
gm
ent
a
t
i
on i
n
cl
u
d
e gl
obal
m
ean, local mean,
global
vari
a
n
ce, l
o
cal
vari
a
n
ce, a
n
d
hi
st
o
g
ram
.
In
gene
ral
,
t
h
e
mean of
grey-le
v
el val
u
es in t
h
e fore
ground
is lowe
r
t
h
an t
h
e
val
u
e
s
i
n
t
h
e bac
k
gr
ou
n
d
.
On t
h
e o
t
her
han
d
,
t
h
e
varia
n
ce of gre
y
-level va
lu
es
in
th
e fo
regroun
d
is
hi
g
h
er t
h
a
n
t
h
e
back
gr
ou
n
d
.
Ho
we
ver
,
m
ean an
d vari
a
n
ce
grey
-scal
e-
bas
e
d al
go
ri
t
h
m
does n
o
t
wo
r
k
wel
l
on
lo
w
qu
ality fing
erp
r
i
n
t im
ag
e [2
]
[9
]
[11
]
.
Meh
t
r
e
et al.
c
o
m
put
ed
di
rect
i
onal
i
m
age, re
prese
n
t
i
n
g l
o
c
a
l
ri
d
g
e
ori
e
nt
at
i
on al
on
g ei
g
h
t
di
ffe
rent
di
rect
i
o
ns i
n
bl
ock
s
of si
ze
16
16
pi
xel
s
and se
gm
ent
e
d i
t
usi
n
g t
h
e bl
oc
k
-
wi
se
hi
st
o
g
ram
of the di
rect
i
o
nal
im
age val
u
es
[
12]
.
The
hi
st
o
g
ram
of t
h
e
di
r
ect
i
onal
i
m
age t
echni
qu
e
gi
v
e
s a
go
o
d
re
sul
t
fo
r l
o
w c
o
nt
r
a
st
an
d
noi
sy
i
m
ages, but
i
t
fai
l
s
f
o
r
im
ages wi
t
h
u
n
i
f
orm
regi
o
n
s
.
C
o
nt
rari
l
y
, t
h
e vari
a
n
ce
gre
y
-l
evel
m
e
t
hod
i
s
not
pr
o
duci
ng
g
o
o
d
res
u
l
t
s
f
o
r l
o
w c
o
nt
rast
i
m
age and i
n
st
ead i
t
does
p
r
od
uce
go
o
d
re
sul
t
s
fo
r i
m
ages wi
t
h
uni
fo
rm
regi
on
s. Al
so t
h
i
s
m
e
t
h
o
d
has
no s
e
nse f
o
r cl
ari
t
y
of ri
d
g
es an
d t
h
ei
r di
rect
i
o
ns, a
nd
hence
,
cann
o
t
det
ect
noi
sy
r
e
gi
o
n
s as
bac
k
g
r
ou
n
d
. T
h
i
s
l
eads M
e
ht
re
and C
h
at
t
e
rj
ee
uses t
h
e c
o
m
posi
t
e
m
e
tho
d
by
com
b
ining hist
ogram
and variance
m
e
t
hods
[13]. T
h
is m
e
thod is reporte
d
to give
good results for uniform
regi
ons
, e
nha
n
ced i
n
p
u
t
i
m
ages, a
n
d
also
poo
r fo
r low
co
n
t
r
a
st im
ag
es.
Rath
a
et al
.
pr
op
ose
d
a
fi
nge
rp
ri
nt
se
gm
ent
a
t
i
on
t
o
se
par
a
t
e
t
h
e fi
nger
p
ri
nt
area t
o
avoi
d ext
r
act
i
o
n
o
f
f
eat
u
r
e in
n
o
i
sy and
b
a
ck
gro
und
ar
eas u
s
in
g
v
a
r
i
an
ce o
f
gr
ey-
l
ev
el in
a d
i
r
ectio
n
o
r
t
h
ogon
al to
th
e
ori
e
nt
at
i
on
fi
el
d i
n
eac
h
bl
oc
k o
f
si
ze
16
16
. The
angl
e
of
o
r
i
e
nt
at
i
on fi
el
d
i
s
q
u
ant
i
zed
i
n
t
o
16
di
rect
i
o
n
s
[1
5]
. I
n
[
13]
, t
h
e va
ri
ance at
every
pi
xel
i
n
a set
of k
n
o
w
n
di
rect
i
o
ns use
d
t
o
deci
de
whet
her t
h
e
pi
xel
i
s
i
n
t
h
e
foregro
und
. In
ad
d
ition
,
Rath
a u
s
ed
th
e
v
a
rian
ce to
d
ecide th
e qu
ality o
f
th
e fing
erprin
t imag
e in
ter
m
s
o
f
th
e
im
age cont
rast
of t
h
e bl
oc
k u
nde
r co
nsi
d
e
r
a
t
i
on. The
un
de
rl
y
i
ng assum
p
t
i
on i
s
t
h
at
t
h
e
noi
se re
gi
o
n
s h
a
ve n
o
B
ackground
Foreground
Noise
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Seg
m
e
n
t
a
t
i
o
n of
Fi
n
g
er
pri
n
t
Im
ag
e
B
a
se
d o
n
Gra
d
i
e
nt
Ma
gni
t
u
de a
n
d
C
ohe
rence
(Sapa
rud
i
n
)
1
204
di
rect
i
o
nal
de
pen
d
e
n
ce,
w
h
ereas
regi
ons
of i
n
t
e
rest
e
x
hi
bi
t
ha
ve
a
very
hi
g
h
var
i
ance
i
n
a
di
r
ect
i
on
ort
h
o
g
onal
t
o
t
h
e ori
e
nt
at
i
o
n
of t
h
e pat
t
e
r
n
and a very
l
o
w vari
a
n
ce al
o
ng t
h
e ri
d
g
es.
In ot
her
wo
rd
s
,
t
h
e
back
g
r
o
u
nd
ha
s l
o
w va
ri
anc
e
i
n
al
l
t
h
e direct
i
ons
. I
n
[
4
]
used t
h
e sa
m
e
m
e
t
hods a
s
[1
5]
, b
u
t
pri
o
r t
o
segm
ent
a
t
i
on p
r
oces
s, fi
nge
r
p
ri
nt
i
m
age i
s
cro
p
p
ed
usi
n
g
C
a
ndel
a
’
s
ap
pr
oach
[
3
]
and i
s
m
a
nual
l
y
al
i
gned i
n
u
p
righ
t p
o
s
ition
.
Ch
en
et al.
p
r
o
p
o
sed
fi
nge
r
p
ri
nt
segm
ent
a
t
i
on
usi
n
g
12
12
pi
xel
bl
oc
ks a
n
d
t
r
ai
ne
d a
l
i
n
ea
r
cl
assi
fi
er t
o
separat
e
f
o
re
gr
ou
n
d
base
d o
n
t
h
ree feat
ur
es, nam
e
ly
:
(i) t
h
e bl
oc
k cl
ust
e
r de
gree
, (i
i
)
t
h
e
d
i
fferen
ce b
e
t
w
een
l
o
cal m
e
an
and
g
l
ob
al mean
o
f
g
r
ey
-l
ev
el, and
(iii) th
e b
l
o
c
k
v
a
rian
ce of
g
r
ey-level. Th
is
sel
f
-
d
efi
n
e
d
bl
ock cl
ust
e
r de
gree i
s
a
m
easure o
n
h
o
w
w
e
l
l
t
h
e pi
xel
s
of t
h
e bl
ock c
o
n
g
r
egat
e wi
t
h
i
n
t
h
e
bloc
k by com
p
aring each pixel’s intensity
with the global
m
ean
intensity. A
m
o
rphologi
cal operation is the
n
appl
i
e
d
d
u
ri
ng
po
st
-p
r
o
cessi
n
g
t
o
n
o
rm
al
i
z
e t
h
e res
u
l
t
s
[
5
]
.
Thi
s
m
e
t
hod
i
s
cl
aim
e
d t
o
pr
o
v
i
d
e sat
i
s
fa
ct
or
y
results
for
high quality im
age but
has
hi
ghe
r
com
putati
onal
com
p
lexity than m
o
st unsupe
rvised m
e
thods.
Feng
et al.
i
m
prove
d t
h
e
grey
-
v
a
r
i
a
nce
-
base
d fi
nge
r
p
r
i
nt
segm
ent
a
t
i
on
al
g
o
ri
t
h
m
by
com
b
i
n
i
n
g
grey-le
v
el bloc
k m
ean and va
riance.
T
h
ey reported that this algorithm
ac
hieve
d
m
o
re accurate and rel
i
able
segm
ent
e
d fi
n
g
er
pri
n
t
re
sul
t
s
[
6
]
.
Ho
we
ver,
t
h
e
r
o
b
u
st
ne
ss
of their m
e
th
od
yet to
b
e
test
ed
on
m
o
re com
p
le
x
fingerpri
n
ts such as t
h
e
ones
that co
n
t
ain
no
t on
ly r
e
gu
lar sp
ik
es
bu
t also
fo
r
e
i
g
n
obj
ects
su
ch
as ar
tef
a
cts and
h
a
ndwritten
ann
o
t
ation
s
t
h
at
n
o
rm
all
y
fo
und
in raw
fing
erp
r
i
n
ts.
2.
2
Local Directi
o
nality
Fe
ature
s
Fi
nge
r
p
ri
nt
i
m
age can
be vi
e
w
ed as t
w
o
di
s
t
i
n
ct
regi
o
n
s i
.
e., ri
d
g
es
’ re
gi
on a
n
d n
o
n
-ri
d
g
es’
regi
on
.
The m
a
i
n
pur
p
o
se o
f
t
h
e
fi
n
g
e
rp
ri
nt
segm
en
t
a
t
i
on i
s
m
a
i
n
l
y
t
o
ext
r
act
t
h
e
ri
dge
s’ r
e
gi
on
fr
om
t
h
e fi
nge
rp
ri
nt
im
age.
W
i
t
h
r
e
gar
d
s t
o
t
h
at
,
som
e
st
udi
es em
pl
oy
ed ri
d
g
e
s ori
e
nt
at
i
on
t
o
segm
ent
t
h
e im
age. Gene
r
a
l
l
y
, a
ridg
e
d
i
rection can
b
e
esti
m
a
ted
b
y
calcu
latin
g
its
grad
ient, wh
ich no
rm
ally p
e
rfo
r
m
e
d
in
a p
i
x
e
l wi
se or
bl
oc
k
wi
se
ope
rat
i
ons
. T
h
e est
i
m
a
t
e
d gra
d
i
e
n
t
of
a ri
dge
i
s
t
e
rm
ed as o
r
i
e
n
t
at
i
on fi
el
d
.
M
a
i
o
an
d M
a
l
t
oni
use
d
m
ean of
gra
d
i
e
nt
m
a
gni
t
u
de
of
o
r
i
e
nt
at
i
on
fi
el
d i
n
i
m
age bl
ock
s
t
o
sepa
rat
e
f
o
r
e
gro
und
f
r
om b
ack
gr
oun
d. Th
ey o
b
s
er
v
e
d
th
at th
e g
r
ad
ien
t
r
e
spon
se is h
i
g
h
e
r
in
th
e fo
r
e
g
r
ou
nd
co
m
p
ar
ed
to
th
at in
th
e b
ackg
r
o
und
.
By ex
p
l
o
itin
g
th
is in
fo
rm
a
tio
n
,
th
ey su
ccessfu
lly ex
tracted
th
e foregroun
d;
ho
we
ver
,
t
h
ey
fai
l
e
d t
o
i
d
e
n
t
i
f
y
t
h
e noi
se
pat
c
hes
[1
0]
.
Zha
ng a
n
d Ya
n
i
m
prove
d t
h
e abo
v
e m
e
t
hod
by
com
b
ining m
e
an of gradie
nt
m
a
gnitude a
nd c
ohe
re
nce
value [21].
They success
f
ully segm
ente
d the
fi
n
g
er
pri
n
t
i
m
age i
n
t
o
t
h
ree
part
s
nam
e
l
y
, bac
k
gr
ou
n
d
,
fo
re
gr
ou
n
d
a
nd
n
o
i
s
y
re
gi
ons;
h
o
we
ve
r,
t
h
ei
r
assu
m
p
tio
n
s
that all n
o
i
se ar
eas ar
e irr
e
lev
a
n
t
and
d
o
no
t
co
n
t
ain im
p
o
r
tan
t
in
for
m
at
io
n
ar
e pr
ov
en
w
r
ong.
Later, Qi
an
d Xie
propo
sed
a m
o
re creative seg
m
en
tatio
n alg
o
rith
m
u
s
in
g th
e sam
e
mean
m
a
g
n
itu
d
e
o
f
th
e
g
r
ad
ien
t
bu
t this ti
me th
ey co
m
b
in
ed
it with
th
e v
a
rian
ce of the gradie
nt vector’s
directional im
age, instead
[1
4]
. T
h
i
s
m
e
tho
d
i
s
re
po
rt
ed
has
achi
e
ved
go
o
d
resul
t
s
; ho
wev
e
r, no
ise p
a
tch
e
s
in
foreg
r
ou
nd
area
are
failed
to detect a
n
d are treated a
s
part of the
bac
k
ground.
Zhu
et al.
pro
p
o
se
d a no
vel
segm
ent
a
t
i
on t
e
chni
que t
h
at
g
r
a
dual
l
y
ext
r
act
t
h
e fo
re
gr
ou
n
d
is
based o
n
cor
r
ect
ori
e
nt
at
i
on fi
el
ds
. Th
e corre
ctness
of an orientation field is ob
tain
ed
b
y
train
i
ng
th
e Neu
r
al Network
(N
N)
. The t
r
a
i
ned N
N
cl
assi
fi
er i
s
use
d
t
o
di
st
i
n
g
u
i
s
h
bet
w
een t
h
e co
rre
ct
and i
n
c
o
r
r
ec
t
ori
e
nt
at
i
on
fi
el
ds.
The
n
, a f
o
re
gr
ou
n
d
i
s
gra
d
u
a
l
l
y
form
ed by
addi
ng
on
bl
oc
k by
bl
o
c
k
of t
h
e co
rrect
e
d
or
i
e
nt
at
i
on fi
el
ds
, a
n
d
i
t
s
fo
rm
ati
on i
s
exact
l
y
resem
b
l
a
nce
t
h
e
fam
ous
re
gi
o
n
g
r
o
w
i
n
g c
once
p
t
[
22]
[2
3]
. T
h
e
t
echni
que
i
s
e
v
i
d
ent
l
y
tedious and time consum
ing b
ecause every
single orientation field
has to
be analysed and co
rrected s
h
ould its
b
eari
n
g
is off
directio
n
.
Worse still, in
real li
fe app
licatio
n
,
fing
erp
r
i
n
ts i
m
ag
es m
a
y co
me in
v
a
ri
o
u
s
qualitie
s
and in act
ual fa
ct som
e
ar
e b
e
yo
nd
o
u
r im
ag
in
atio
n.
Later, Yu
et al.
ad
opt
e
d
a
g
r
a
d
i
e
nt
pr
o
j
ect
i
o
n m
e
t
hod
t
o
e
x
cl
ude
bac
k
gr
o
u
n
d
re
gi
o
n
c
h
a
r
act
eri
zed
by
low
grey-scale variation, a
n
d coarsely
ob
tai
n
ed
t
h
e fo
r
e
gro
und
r
e
g
i
on
o
f
th
e f
i
ng
erpr
int i
m
ag
e. I
n
add
itio
n
,
noi
se
re
gi
o
n
s,
w
h
i
c
h c
ont
ai
n sm
udge
s an
d st
ai
ns
, are
excl
u
d
ed
by
u
s
i
ng
g
r
adi
e
nt
cohe
re
nce a
p
p
r
oac
h
.
Fi
nal
l
y
, m
o
rph
o
l
o
gi
cal
o
p
erat
i
ons i
n
cl
u
d
i
n
g
edge
det
ec
tion
are ap
p
lied on
th
e edg
e
s of the fing
erprin
t i
m
ag
e
t
o
obt
ai
n a sm
oot
h b
o
u
n
d
ary
of t
h
e f
o
re
g
r
o
u
n
d
[
2
0]
. Al
t
h
ou
g
h
t
h
e p
ubl
i
s
he
d res
u
l
t
s
evi
d
ent
l
y
reveal
e
d
t
h
at
t
h
e ext
r
act
ed
f
o
re
gr
o
u
n
d
s a
r
e
sm
oot
h b
u
t
s
o
m
e
part
s
of t
h
e
fo
re
gr
ou
n
d
a
r
e u
n
i
n
t
e
nt
i
o
nal
l
y
bl
ackene
d
.
Hence
,
im
port
a
nt
i
n
f
o
r
m
at
i
on m
i
ght
b
e
va
ni
she
d
t
h
at
m
a
y
l
ead t
o
di
sap
p
eara
n
ce
of
si
n
gul
ar
p
o
i
n
t
s
.
Fin
a
lly, Teix
eira and
Leite
pr
op
ose
d
a rat
h
e
r
com
p
l
e
x fi
ng
erp
r
i
n
t
segm
ent
a
t
i
on al
go
ri
t
h
m
based o
n
ori
e
nt
at
i
on
fi
el
ds,
whi
c
h e
x
pl
oi
t
s
m
o
rp
hol
ogi
cal
m
a
t
h
ematical
tran
sfo
r
m
a
tio
n
s
th
at
in
clud
e d
ilatio
n and
erosi
on
[
17]
.
Al
t
h
o
u
gh t
h
ei
r
expe
ri
m
e
nt
s yi
el
d pr
om
i
s
i
ng resul
t
s
, t
h
e
pr
ocesses i
n
v
o
l
v
ed are
rat
h
e
r
t
e
di
o
u
s
with
h
i
gh
com
p
lex
i
t
y
. Wo
rse still,
th
e en
tire o
r
ien
t
atio
n
fiel
d
s
are
co
m
p
u
t
ed
prior to
th
e foregroun
d
ex
traction
,
wh
ich
is con
s
id
ered
ag
ain
s
t t
h
e
no
rm
al p
r
actices, and is
re
garde
d
as
counter
productive.
2.
3
Coherence Fe
atures
A c
ohe
rence
f
eat
ure
rep
r
ese
n
t
s
t
h
e st
re
ngt
h
of l
o
cal
g
r
a
d
i
e
nt
s cent
r
ed
at
t
h
e t
a
r
g
et
pi
xel
,
w
h
i
c
h
has
d
o
m
in
an
t rep
r
esen
tatio
n. Gen
e
rally, th
e coh
e
ren
ce is al
s
o
hi
ghe
r in t
h
e foreground,
whe
r
e the
gre
y
-level
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 5
,
N
o
. 5
,
O
c
tob
e
r
20
15
:
120
2
–
12
15
1
205
val
u
es a
r
e m
u
ch sm
oot
he
r al
on
g t
h
e
di
rect
i
on
o
f
t
h
e
ri
dge
.
On the c
o
ntrary, the c
ohe
re
nce is com
p
aratively
lo
wer at th
e reg
i
on
wh
ere th
ere is a lo
t o
f
sp
ik
es ex
is
t, which are em
anated from
noises such as stains a
nd
sm
udges [
20]
.
It
seem
s t
h
at
the co
here
nce i
s
very
p
r
om
i
s
ing t
o
be us
ed
as a si
ngl
e fea
t
ure t
o
se
gm
ent
t
h
e
foregro
und
; howev
er it is no
t su
fficie
n
t
fo
r
robu
stin
g
segmen
tatio
n
.
Th
erefo
r
e, a systematic co
m
b
in
atio
n
of
several
feature
s
is nece
ssary
[2] [18].
Hi
st
ori
cal
l
y
, t
h
e w
o
r
d
c
ohe
re
nce
was fi
rst
p
r
op
ose
d
by
Kas
s
an
d
W
i
t
k
i
n
w
h
o
de
fi
ne i
t
as
t
h
e n
o
rm
of
t
h
e sum
of
o
r
i
e
nt
at
i
on
vect
o
r
s di
vi
ded
by
t
h
e s
u
m
of t
h
ei
r i
n
di
vi
dual
no
rm
s;
t
h
i
s
scal
ar al
way
s
l
i
e
s i
n
t
h
e
ran
g
e
of
[
0
,
1]
.
Ori
e
nt
at
i
ons
wi
t
h
paral
l
e
l
di
rect
i
ons
p
r
o
v
i
d
e m
a
xim
u
m
co
here
nce
(i
.e. i
t
s
val
u
e i
s
1
)
,
w
h
ereas
ori
e
nt
at
i
ons
wi
t
h
o
p
p
o
s
i
t
e
di
r
ect
i
ons
gi
ve
m
i
ni
m
u
m
cohere
nce
(i
.e. i
t
s
val
u
e i
s
0)
[
2
4]
.
Bazen and Ge
rez
pr
o
p
o
s
ed
a pi
xel
-
wi
se segm
ent
a
t
i
on t
echni
que
base
d
on t
h
e c
ohe
r
e
nce, w
h
i
l
e
m
o
rph
o
l
o
gy
i
s
use
d
t
o
o
b
t
a
i
n
sm
oot
h r
e
gi
o
n
s. T
h
e s
e
gm
ent
a
t
i
on m
e
t
hod i
s
ca
pabl
e
of s
u
cc
essful
l
y
id
en
tifying
v
e
ry n
o
i
sy reg
i
o
n
in
th
e fing
erprin
t [1
]. Later in
20
01
, th
ey i
m
p
r
o
v
e
d
th
e tech
n
i
q
u
e
u
s
ing th
ree
diffe
re
nt features nam
e
ly; c
ohe
re
nce, loca
l
m
ean
, and
local variance
, whic
h are
com
puted for each
ove
rl
ap
pe
d bl
o
c
k o
f
pi
xel
[
2
]
.
The segm
ent
a
t
i
on p
r
oces
s i
s
carri
ed
out
o
n
pi
xel
-
by
-
p
i
x
el
basi
s i
n
whi
c
h t
h
e
f
o
r
e
gro
und
separ
a
tio
n is
p
e
rfo
r
m
ed
u
s
i
n
g a lin
ear
classi
f
i
er
. Th
en
,
a m
o
r
pho
log
i
cal oper
a
tio
n is ap
p
l
ied
as
p
o
s
t p
r
o
cessing
to
ob
tain
p
e
rfect
clu
s
ters an
d
to
re
duce
categorization errors
. T
h
eir
expe
rim
e
ntal results
showe
d
that t
h
e m
e
thod provi
d
es acc
u
r
ate
resu
lts; ho
wev
e
r, its co
m
p
u
t
atio
n
a
l co
m
p
lex
ity is m
a
rk
ed
ly h
i
g
h
e
r
t
h
an m
o
st
of
t
h
e desc
ri
be
d
bl
oc
k-
wi
se ap
pr
oac
h
es. M
o
r
e
ove
r,
Yi
n
,
et al.
p
r
o
p
o
se
d
a no
vel
pi
xe
l
-
wi
se
fingerpri
n
t segmentation approac
h
base
d on quadric surface
m
odel [19]
.
They claime
d that their propos
e
d
m
e
t
hod
has
si
g
n
i
f
i
cant
l
y
re
d
u
ced se
gm
ent
a
t
i
on
er
ro
rs a
s
op
pos
ed
t
o
t
h
at
o
f
t
h
e
l
i
n
ear
cl
assi
fi
er.
Meanwhile, Kl
ein
et al.
ado
p
t
e
d f
o
u
r
di
f
f
ere
n
t
pi
xel
base
d
feat
ure
s
nam
e
ly
, grey
-scal
e
m
ean, grey
-
scal
e vari
a
n
ce,
gra
d
i
e
nt
c
o
he
rence a
n
d
Ga
bo
r
resp
o
n
se
f
o
r t
h
e fi
n
g
er
p
r
i
n
t
se
gm
ent
a
ti
on.
The
se
gm
ent
e
d
fing
erp
r
i
n
t i
m
ag
e is d
eco
m
p
o
s
ed
in
to
three p
a
rts v
i
z.
foregroun
d, b
a
ckg
r
ou
nd
and
low-q
u
ality reg
i
o
n
s
. In
ad
d
ition
,
a h
i
dd
en
Marko
v
m
o
d
e
l (HMM) is ap
p
lied
to
reso
l
v
e th
e frag
m
en
ted
fo
regroun
d
s
in
stead o
f
a
com
m
on
m
o
rp
hol
ogi
cal
o
p
er
at
i
on. T
h
e pi
x
e
l
feat
ures are
m
odel
l
e
d as t
h
e out
put
o
f
a hi
dde
n M
a
r
k
ov
p
r
oces
s
[8]
.
T
h
e pe
rf
o
r
m
a
nce of H
M
M
-
base
d seg
m
ent
a
t
i
on hi
g
h
l
y
depe
n
d
s o
n
t
h
e ch
oi
ce o
f
pi
xel
feat
ure
s
. Thei
r
expe
rim
e
ntal r
e
sults reveale
d
that the outcomes are
very
enco
u
r
agi
ng w
i
t
h
l
e
ss fragm
ent
e
d f
o
re
gr
o
u
n
d
. I
n
ad
d
ition
,
th
e categ
o
rizatio
n
of low-qu
ality reg
i
on
p
r
ov
id
es an ex
tra ad
v
a
n
t
ag
e th
at th
e
in
fo
rm
atio
n
in th
i
s
regi
on
i
s
n
o
t
t
o
t
a
l
l
y
di
scar
de
d.
H
o
we
ve
r, t
h
e
per
f
o
r
m
a
n
ce of th
e HMM
is greatly relied
o
n
th
e cho
i
ces of
featu
r
es u
s
ed
an
d
nu
m
b
er o
f
state assig
n
e
d
fo
r th
e
b
a
ck
grou
nd
,
foreg
r
o
und
and
low-qu
ality reg
i
o
n
s
. In
actu
a
l
fact, h
a
v
i
ng
to
o
m
a
n
y
featu
r
es an
d
states will in
crease th
e co
m
p
u
t
atio
n
a
l co
m
p
lex
ity an
d
as well as
com
put
i
ng t
i
m
e.
Zha
ng a
n
d Ya
n p
r
op
ose
d
t
h
e
fi
n
g
er
pri
n
t
se
gm
ent
a
t
i
on usi
ng t
w
o
-
bl
oc
k f
eat
ures;
m
ean of
gra
d
i
e
nt
mag
n
itu
d
e
an
d coh
e
ren
ce. The fing
erprin
t i
m
ag
e is firs
t
co
nvo
lv
ed
with
a
2D Gau
ssian filter.
Th
e grad
ien
t
s
i
n
h
o
ri
z
o
nt
al
and
ve
rt
i
cal
di
r
ect
i
ons a
r
e est
i
m
a
t
e
d usi
n
g S
obel
o
p
erat
or.
The m
ean o
f
g
r
adi
e
nt
m
a
gni
t
ude
i
s
th
en
co
m
p
ared with
a th
reshold
v
a
lu
e of th
e g
r
ad
ien
t
. Th
ey
defi
ned “i
n
v
al
i
d
regi
o
n
s” i
n
t
h
e fo
re
gr
o
u
n
d
as t
h
e
sets of c
o
nnect
ed elem
ents with lo
w cohe
re
nce
value. T
h
e
fingerprint im
age is se
gm
ented int
o
background,
foregro
und
or
n
o
i
sy reg
i
on
s
[2
1
]
. Ex
cep
t
for th
e no
is
y regio
n
s
in
t
h
e fo
reg
r
ou
nd
, t
h
ere are still su
ch
n
o
i
ses
existed i
n
the
background
whose c
o
here
nces
are
l
o
w, a
n
d a
r
e m
i
stakenly assigne
d a
s
fore
ground.
3.
PROP
OSE
D
METHO
D
For
e
g
r
o
u
nd e
x
t
r
act
i
on i
s
act
ual
l
y
part
of
f
i
nge
rp
ri
nt
i
m
age segm
ent
a
t
i
on
, w
h
i
c
h ai
m
s
t
o
separat
e
foregro
und
from
its b
ackg
r
oun
d and o
t
h
e
r
fo
rei
g
n obj
ects
lik
e artefacts an
d h
a
nd
written ann
o
t
ation
s
,
wh
ich
are c
o
m
m
on i
n
i
nke
d
fi
n
g
er
p
r
i
n
t
s
. It
al
s
o
t
a
s
k
s
fo
r
det
ect
i
n
g
noi
se
re
gi
o
n
s
f
o
u
n
d
i
n
t
h
e
f
o
re
gr
o
u
n
d
.
The p
r
o
p
o
se
d ext
r
act
i
o
n be
gi
ns wi
t
h
n
o
r
m
a
li
zat
i
on of t
h
e f
i
nge
rp
ri
nt
’s i
n
t
e
nsi
t
y
val
u
es b
y
adopt
i
n
g
Ho
n
g
’s
n
o
rm
al
i
zat
i
on ap
pr
o
ach
[7]
.
The
n
, i
t
i
s
f
o
l
l
o
w
e
d
by
f
o
re
g
r
o
u
n
d
re
gi
o
n
se
parat
i
o
n
fr
om
t
h
e
back
g
r
o
u
nd
us
i
ng t
h
e
pr
o
p
o
s
ed segm
ent
a
t
i
on t
ech
ni
q
u
e. F
i
nal
l
y
, t
h
e gra
d
i
e
nt
co
he
renc
e app
r
oach
, w
h
i
c
h i
s
pi
o
n
eere
d
,
by
Zhan
g an
d
Yan
is adop
ted
to
d
e
tect
th
e n
o
i
se reg
i
on
s ex
isted in
th
e foreg
r
ou
nd
.
Di
ag
ram
m
a
t
i
c
al
l
y
, t
h
e m
e
t
hod i
s
di
spl
a
y
e
d
i
n
Fi
gu
re
2.
Fi
gu
re
2.
The
f
i
nge
rp
ri
nt
se
g
m
ent
a
t
i
on sc
he
m
e
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Seg
m
e
n
t
a
t
i
o
n of
Fi
n
g
er
pri
n
t
Im
ag
e
B
a
se
d o
n
Gra
d
i
e
nt
Ma
gni
t
u
de a
n
d
C
ohe
rence
(Sapa
rud
i
n
)
1
206
3.
1
Grey-Sca
le No
rma
liza
t
io
n
Norm
all
y
, th
e
in
ten
s
ity v
a
lu
e o
f
fing
erprin
t i
m
ag
es is g
r
eatly v
a
ried
fro
m
o
n
e
prin
t to
an
o
t
h
e
r
o
v
e
r
ti
m
e
o
f
cap
t
u
ri
n
g
.
As a resu
lt, th
ere are prin
t
s
who
s
e in
te
nsi
t
y values c
o
nc
entrated in the
uppe
r-range
of grey-
l
e
vel
s
, fo
r i
n
st
ance 1
28
– 25
5, w
h
i
c
h i
n
di
c
a
t
e
s bri
g
ht
im
ages o
r
o
v
er
-e
xp
os
ure
.
O
n
t
h
e co
nt
rary
, t
h
ere are
fi
n
g
er
pri
n
t
s
w
hos
e grey
-l
e
v
e
l
s rangi
ng f
r
o
m
0 – 128 o
r
l
o
we
r-
ra
nge
, w
h
i
c
h i
n
di
cat
es dar
k
im
ages or
un
der
-
ex
po
su
re. Th
e
u
n
e
v
e
n
or irreg
u
l
ar d
i
stribu
tio
n
o
f
ligh
t
in
ten
s
ities
m
a
y affect th
e statist
i
c
a
l in
fo
rm
atio
n
o
f
the
im
age such as
mean and
va
riance of
grey-l
evels, a
nd
t
h
erefore norm
a
l
i
z
a
tio
n
is n
e
ed
ed
. Th
is no
rm
alizatio
n
pr
ocess ai
m
s
at
reduci
n
g va
ri
at
i
on i
n
grey
-l
evel
val
u
es al
o
ng ri
dge
s and
val
l
e
y
s
wi
t
h
o
u
t
changi
ng t
h
e
cl
ari
t
y
of t
h
eir structures. T
h
ere
f
ore, the in
put fi
ngerprint im
age is standa
rdized
t
o
a desi
re
d m
ean an
d
vari
a
n
c
e
. Th
e
No
rm
al
i
z
at
i
on
m
e
t
hod
pr
o
p
o
s
ed i
n
[
7
]
co
nsi
s
t
s
of
t
h
ree
st
e
p
s:
Fi
rst
l
y
,
gl
o
b
al
m
ean val
u
e
of
fi
n
g
e
r
p
r
i
n
t
im
ag
e
is d
e
term
in
ed
. Second
ly,
g
l
o
b
a
l
v
a
rian
ce
v
a
lu
e of
fing
erprin
t im
ag
e is co
m
p
u
t
ed. Fin
a
lly, n
e
w inten
s
ity
values
are c
a
lculated.
Detailed
pro
c
ess of the
n
o
rm
a
lizatio
n
is p
e
rfo
r
m
e
d
as
fo
llows:
1.
Let
)
,
(
n
m
I
d
e
no
te the g
r
ey-lev
el o
r
in
ten
s
ity v
a
lue o
f
th
e
p
i
x
e
l
at th
e
m
-th ro
w
an
d
n
-t
h c
o
l
u
m
n
of
H
W
pi
xel
s
of fi
n
g
e
rp
ri
nt
i
m
age si
ze. Let
Mg
and
Vg
den
o
t
e
t
h
e
gl
obal
m
ean and
gl
o
b
al
va
ri
anc
e
v
a
lu
es of
fing
erprin
t im
ag
e, resp
ectiv
ely.
2.
Calculate the
norm
alized gr
ey-level value
at pixel
)
,
(
n
m
of f
i
nge
rp
ri
nt
im
age, w
h
i
c
h i
s
d
e
not
e
d
by
)
,
(
n
m
N
, an
d i
s
de
fi
ne
d
as f
o
l
l
o
ws:
otherwise
)
)
)
,
(
(
(
)
,
(
if
)
)
)
,
(
(
(
)
,
(
1
0
1
0
2
0
0
1
0
1
0
2
0
0
W
m
H
n
W
m
H
n
Vg
Mg
n
m
I
Vg
Mg
Mg
n
m
I
Vg
Mg
n
m
I
Vg
Mg
n
m
N
(1
)
whe
r
e
0
Mg
and
0
Vg
are the desired mean and va
riance values, re
spectively. Ideally, the recommende
d val
u
e
fo
r bot
h
0
Mg
and
0
Vg
is 10
0.
3.
2
Prop
osed
F
o
r
e
gro
und
E
x
tr
acti
on
Met
h
o
d
Once the normalised grey-le
v
el valu
es of the finge
rprint i
m
age are ob
tain
ed, th
e n
e
x
t
pro
cess is to
ext
r
act
t
h
e
fo
re
gr
o
u
n
d
f
r
o
m
t
h
e fi
n
g
er
pri
n
t
i
m
age. The
p
r
o
cess i
s
d
one
ba
sed
on
bl
oc
k-
b
y
-
bl
oc
k
basi
s s
t
art
i
n
g
fr
om
t
op l
e
ft
cor
n
er a
n
d e
n
d
e
d at
b
o
t
t
o
m
ri
ght
c
o
r
n
er
.
Wi
t
h
rega
r
d
t
o
t
h
at
, a
new s
e
g
m
ent
a
t
i
on ap
p
r
oac
h
,
whic
h com
b
ines local
m
ean
value
of t
h
e
norm
alised grey-level
and local va
riance
value
of t
h
e gradie
nt
m
a
gni
t
ude
, i
s
pr
o
pose
d
. T
h
i
s
m
e
t
hod co
nsi
s
t
s
of t
h
ree m
a
in st
eps:
Fi
rst
,
gl
o
b
al
m
ean (i
.e.
Mn
in
sh
ort) an
d
local m
ean (i.e.
)
,
(
j
i
Mb
in
sh
ort) valu
es o
f
no
rm
alized
fing
er
print im
age are
calcula
ted. Se
cond,
local
vari
a
n
ce (i
.e
.
den
o
t
e
d
by
)
,
(
j
i
Vgr
)
an
d th
resho
l
d (i.e.
th
G
i
n
sh
o
r
t
)
val
u
e
s
o
f
gr
adi
e
nt
m
a
gni
t
ude
ar
e
co
m
p
u
t
ed
. Finally, th
e targ
et b
l
o
c
k
is assig
n
e
d
as a
p
a
rt o
f
foreg
r
o
und
if th
e fo
llowing
con
d
ition
s
are
fu
lfilled
:
(i
) if th
e lo
cal m
e
an
is sm
aller
th
an
g
l
ob
al mean
,
o
r
(ii) the lo
cal v
a
riance is g
r
eater t
h
an
t
h
e
t
h
res
hol
d. Di
a
g
ram
m
ati
cal
ly
,
t
h
i
s
pr
ocess
i
s
di
spl
a
y
e
d
i
n
Fi
gu
re 3.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 5
,
N
o
. 5
,
O
c
tob
e
r
20
15
:
120
2
–
12
15
1
207
Mn
)
,
(
j
i
Mb
)
,
(
j
i
Vgr
th
G
Mn
j
i
Mb
)
,
(
th
G
j
i
Vgr
)
,
(
Fi
gu
re
3.
Fl
o
w
chart
of
f
o
re
g
r
ou
n
d
e
x
t
r
act
i
o
n
[1
6]
a.
C
o
m
put
at
i
on o
f
gl
o
b
al
m
ean
and local m
ean value
s
Gl
o
b
al
m
ean is obt
ai
ne
d by
com
put
i
ng t
h
e
avera
g
e of
gr
ey
-scal
e val
u
e
s
of t
h
e w
hol
e
norm
a
l
i
zed
im
age, w
h
erea
s l
o
cal
m
ean val
u
e i
s
c
o
m
put
ed
base
d
on
bl
oc
k
of
pi
xe
l
s
. The c
a
l
c
ul
a
t
i
on i
s
per
f
o
r
m
e
d as
fo
llows.
1.
Let
H
W
be
the
size of the
no
rm
alized im
age. Let
B
B
pi
xel
s
be
a n
o
n
-
o
verl
a
p
pi
n
g
bl
oc
k
of
t
h
e
norm
alized image.
In this cas
e
16
B
. Let
)
,
(
v
u
N
b
e
th
e i
n
ten
s
ity v
a
lu
e
of th
e
p
i
x
e
l at t
h
e
u
-t
h r
o
w a
n
d
v
-th
co
lu
m
n
of
th
e
B
B
bl
oc
k.
Let
P
be t
h
e
num
ber
o
f
bl
oc
ks i
n
t
h
e ent
i
r
e i
m
age.
2.
Calculate the global m
ean val
u
e,
Mn
. Calculate
the local m
ean
of each
bl
ock using followi
ng
equation.
,
)
,
(
)
,
(
1
1
B
B
v
u
N
j
i
Mb
B
i
i
u
B
j
j
v
(2
)
whe
r
e
)
,
(
j
i
is first p
i
x
e
l at
i
-th ro
w an
d
j
-t
h col
u
m
n
of t
h
e
B
B
bl
ock
,
16
...,
32,
16,
,
0
W
i
, and
16
...,
32,
16,
,
0
H
j
.
b.
C
o
m
put
at
i
on o
f
l
o
cal
vari
a
n
c
e
o
f
gra
d
i
e
nt
m
a
gni
t
u
de
The l
o
cal
vari
ance o
f
gra
d
i
e
nt
m
a
gni
t
ude
of eac
h
bl
oc
k
i
s
com
put
ed a
ccor
d
i
n
g t
o
t
h
e fol
l
o
wi
n
g
steps:
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Seg
m
e
n
t
a
t
i
o
n of
Fi
n
g
er
pri
n
t
Im
ag
e
B
a
se
d o
n
Gra
d
i
e
nt
Ma
gni
t
u
de a
n
d
C
ohe
rence
(Sapa
rud
i
n
)
1
208
1.
For eac
h pi
xel
)
,
(
n
m
o
f
th
e nor
m
a
l
i
zed
f
i
ng
er
pr
in
t
i
m
ag
e
)
,
(
n
m
N
; estim
ated gradie
nts in horizontal and
vert
i
cal
di
rect
i
ons
, w
h
i
c
h are
sym
bol
i
zed b
y
)
,
(
n
m
G
x
and
)
,
(
n
m
G
y
, res
p
ec
tively are compute
d
using
t
h
e fol
l
owi
ng
So
bel
m
a
sk
3
3
o
p
erat
ors
.
H
o
ri
z
ont
al
S
obel
m
a
sk o
p
e
r
at
o
r
)
,
(
q
p
S
x
and
vert
i
cal
S
o
bel
mask operat
or
)
,
(
q
p
S
y
)
)
,
(
)
,
(
(
)
,
(
1
1
1
1
pq
x
x
q
n
p
m
N
q
p
S
n
m
G
(3
)
)
)
,
(
)
,
(
(
)
,
(
1
1
1
1
pq
y
y
q
n
p
m
N
q
p
S
n
m
G
(4
)
2.
C
a
l
c
ul
at
e t
h
e g
r
adi
e
nt
m
a
gni
t
ude
)
,
(
n
m
Gr
fo
r eac
h
p
i
xel
)
,
(
n
m
as
foll
ows
.
))
,
(
)
,
(
(
)
,
(
2
2
n
m
G
n
m
G
n
m
Gr
y
x
(5
)
3.
Determ
in
e
th
e th
resh
o
l
d
v
a
lu
e
th
G
o
f
t
h
e
g
r
adi
e
nt
m
a
gni
t
udes
us
i
ng
Zha
n
g a
n
d
Yan
’
s
m
e
th
o
d
as
fo
llo
ws:
3.
1
Let
)
,
(
n
m
Gr
de
n
o
t
e
t
h
e
g
r
adi
e
nt
m
a
gni
t
ude at
eac
h
pi
xel
)
,
(
n
m
of
th
e
H
W
im
ag
e size.
3.
2
Det
e
rm
i
n
e t
h
e
m
a
xim
u
m
and t
h
e m
i
nim
u
m
of t
h
e g
r
adi
e
nt
m
a
gni
t
ude
s,
max
)
,
(
n
m
Gr
and
min
)
,
(
n
m
Gr
, res
p
ectively.
3.
3
C
a
l
c
ul
at
e t
h
res
hol
d
val
u
e
usi
n
g t
h
e
f
o
l
l
o
wi
n
g
eq
uat
i
o
n
.
min
min
max
)
,
(
)
)
,
(
)
,
(
(
n
m
Gr
n
m
Gr
n
m
Gr
c
G
th
(6
)
whe
r
e
c
i
s
t
h
e t
h
res
h
ol
d
fact
o
r
t
h
at
ca
n
be c
hos
en
wi
t
h
i
n
a
ra
nge
o
f
]
3
.
0
,
05
.
0
[
de
p
e
ndi
ng
o
n
i
m
age
cont
rast
. A
sm
al
l
e
r val
u
e
of
c
will en
cou
r
ag
e th
e b
l
o
c
k
t
o
beco
m
e
fo
reg
r
ou
nd
,
wh
ile larg
er
v
a
lu
e
will
tr
an
sf
or
m
th
e blo
c
k
t
o
b
ackgro
und
.
Em
p
i
r
i
cally,
1
.
0
c
is chosen.
4.
Local
vari
a
n
ce
o
f
gra
d
i
e
nt
m
a
gni
t
u
de
ca
n be
det
e
rm
i
n
ed
as fol
l
o
ws.
4.
1
Let
)
,
(
v
u
Gr
be t
h
e
gr
adi
e
nt
m
a
gni
t
u
de o
f
t
h
e
pi
x
e
l
at
u
-t
h r
o
w
a
nd
v
-th
co
lu
m
n
in
t
h
e
B
B
bl
oc
k.
4.
2
Calculate the local m
ean values of
gra
d
ient
magnitude
)
,
(
j
i
Mgr
of each bloc
k
)
,
(
j
i
as
fo
llows:
,
)
,
(
)
,
(
1
1
B
B
v
u
Gr
j
i
Mgr
B
i
i
u
B
j
j
v
(7
)
whe
r
e
16
...,
32,
16,
,
0
W
i
, a
n
d
16
...,
32,
16,
,
0
H
j
.
4.
3
Calculate the l
o
cal va
riance
values
of
gra
d
ient m
a
gnitude
)
,
(
j
i
Vgr
o
f
eac
h
bl
oc
k
)
,
(
j
i
that are
defi
ned
as
fol
l
ows:
B
B
j
i
Mgr
v
u
Gr
j
i
Vgr
B
i
i
u
B
j
j
v
1
1
2
))
,
(
)
,
(
(
)
,
(
(8
)
whe
r
e
1
...,
32,
16,
,
0
W
i
, and
1
...,
32,
16,
,
0
H
j
.
c.
If (
Mn
j
i
Mb
)
,
(
and
th
G
j
i
Vgr
)
,
(
), t
h
en
t
h
e t
a
r
g
et
bl
oc
k i
s
assi
gne
d a
s
a pa
rt
o
f
bac
k
g
r
ou
n
d
re
gi
o
n
,
ot
he
rwi
s
e i
s
de
si
gnat
e
d as
f
o
r
e
gr
o
u
n
d
regi
on
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 5
,
N
o
. 5
,
O
c
tob
e
r
20
15
:
120
2
–
12
15
1
209
3.
3
Noise
Are
a
s I
d
entific
a
ti
on and Mar
k
ing
Using Gradient Coherence
Gene
ral
l
y
, gra
d
i
e
nt
co
her
e
nc
e i
s
used t
o
de
scri
be t
h
e
vari
at
i
on o
f
g
r
ey
-l
evel
val
u
es i
n
an i
m
age. It
can also be applied to invest
igate on
how
doe
s each pi
xe
ls-bloc
k
beha
ves in term
s of its gradie
nt va
lue in
relation
t
o
fin
g
e
rp
rint rid
g
e flows
.
T
h
e
large
r
value i
ndicat
es that e
v
er
y
pix
e
l of
the
b
l
ock
sh
ar
es a commo
n
di
rect
i
o
n,
w
h
i
c
h i
s
i
n
acc
or
da
nce t
o
ri
dge
di
rect
i
o
n
.
On
th
e con
t
rary, th
e
smaller v
a
lu
e si
g
n
i
fies th
at m
a
j
o
rity
of t
h
e pi
xel
s
h
a
ve n
o
n
-
u
n
i
f
or
m
di
rect
i
ons,
and
does
not
r
e
sem
b
l
e
t
r
ue ri
dge fl
o
w
. T
h
e
gradi
e
nt
cohe
renc
e
val
u
e i
s
us
ual
l
y
l
a
rger
i
n
fo
re
gr
o
u
n
d
of t
h
e
f
i
nge
rp
ri
nt
im
a
g
e,
where t
h
e
grey values are
m
u
ch s
m
oothe
r along
the direction
of the
ridge than t
h
at
at
t
h
e per
p
en
di
c
u
l
a
r di
rect
i
o
n
of
the ridge. T
h
e gra
d
ient coherenc
e
measures range in
]
1
,
0
[
.
Gra
d
i
e
nt
cohe
re
nce val
u
e of
0 i
n
di
cat
e
s
t
h
at
t
h
e
gra
d
i
e
nt
s i
n
t
h
e
bl
o
c
k are e
q
ual
l
y
di
st
ri
b
u
t
e
d
o
v
e
r
al
l
di
rect
i
o
n
s
. O
n
t
h
e
ot
he
r
ha
nd
,
gra
d
i
e
n
t
co
here
nce
va
l
u
e
of
1
i
n
di
cat
es al
l
pi
xel
s
o
f
t
h
e
bl
oc
k sha
r
e t
h
e sam
e
ori
e
nt
at
i
on. Si
nce
gra
d
i
e
nt
co
he
r
e
nce i
s
base
d
on t
h
e
bl
oc
k
i
n
fo
rm
at
i
on of t
h
e
fi
n
g
er
pri
n
t
im
age, t
h
e fi
n
g
e
r
pri
n
t
im
age i
s
di
vi
de
d i
n
t
o
no
n
-
o
v
erl
a
ppi
n
g
bl
oc
ks o
f
B
B
sized
, in
th
is case
16
B
. Fo
r a gi
ven
n
o
rm
al
i
zed fi
ng
erp
r
i
n
t
i
m
age, gra
d
i
e
nt
c
ohe
r
e
nce
)
,
(
j
i
Coh
of eac
h b
l
ock at
pi
xel
)
,
(
j
i
is calcu
lated
as fo
llo
ws:
1.
Let
)
,
(
v
u
G
x
and
)
,
(
v
u
G
y
d
e
no
t
e
th
e grad
ien
t
s in
x
an
d
y
di
r
ect
i
ons of
t
h
e pi
xel
at
u
-th
r
o
w
an
d
v
-th
co
lu
m
n
in
t
h
e
B
B
bl
oc
k.
2.
Calculate the gradie
nt cohe
re
nce
)
,
(
j
i
Coh
as fo
llo
ws
,
)
,
(
)
,
(
)
,
(
)
,
(
2
2
j
i
V
j
i
V
j
i
V
j
i
Coh
z
y
x
(9
)
w
h
er
e
))
,
(
)
,
(
(
)
,
(
2
2
1
1
v
u
G
v
u
G
j
i
V
y
x
B
i
i
u
B
j
j
v
x
(1
0)
)
,
(
)
,
(
2
)
,
(
1
1
v
u
G
v
u
G
j
i
V
y
x
B
i
i
u
B
j
j
v
y
(1
1)
))
,
(
)
,
(
(
)
,
(
2
2
1
1
v
u
G
v
u
G
j
i
V
y
x
B
i
i
u
B
j
j
v
z
(1
2)
whe
r
e
16
...,
32,
16,
,
0
W
i
, a
n
d
16
...,
32,
16,
,
0
H
j
.
An e
x
am
pl
e of
t
h
e res
u
l
t
a
nt
i
m
age aft
e
r u
n
d
er
g
one t
h
e ab
ove
pr
ocess i
s
gi
ve
n i
n
Fi
gu
r
e
4. I
n
t
h
i
s
case,
5
.
0
)
,
(
j
i
Coh
.
Figure
4. Nois
e areas
of the
fore
g
r
ou
nd
are i
d
en
tified
an
d lab
e
lled
The w
h
i
t
e
col
o
ure
d
bl
oc
ks
i
n
dicate the
noise areas
Evaluation Warning : The document was created with Spire.PDF for Python.
I
J
ECE
I
S
SN
:
208
8-8
7
0
8
Seg
m
e
n
t
a
t
i
o
n of
Fi
n
g
er
pri
n
t
Im
ag
e
B
a
se
d o
n
Gra
d
i
e
nt
Ma
gni
t
u
de a
n
d
C
ohe
rence
(Sapa
rud
i
n
)
1
210
(
N
o
t
e: th
e i
n
put i
m
ag
e is r
e
f
e
rr
ed to
t
h
e ex
tr
acted
fo
r
e
gr
ound
im
ag
e in
Fi
gu
r
e
1(
a) abov
e)
Aft
e
r t
h
e
bac
k
g
r
ou
n
d
, f
o
re
gr
o
u
n
d
, a
n
d
noi
se
re
gi
o
n
s
are det
ect
e
d
,
fu
rt
he
r n
o
i
s
e
regi
ons
are
enha
nce
n
d
u
s
i
n
g
m
e
t
hods a
d
opt
e
d
fr
om
[16
]
.
4.
E
X
PERI
MEN
T
RES
U
LT A
N
D
DI
SC
US
S
I
ON
As m
e
n
tio
n
e
d
in
th
e
p
r
ev
iou
s
ch
ap
ter,
fing
erprin
t is seg
m
e
n
ted
u
s
ing
a co
m
b
in
atio
n
o
f
lo
cal m
ean
val
u
e a
nd l
o
cal
vari
ance
of
ori
e
nt
at
i
on fi
el
d
s
’
gra
d
i
e
nt
m
a
gni
t
ude. I
n
o
r
der
t
o
m
easure t
h
e
per
f
o
r
m
a
nce of t
h
e
p
r
op
o
s
ed
seg
m
en
tatio
n
tech
n
i
q
u
e
i
n
term
s o
f
v
i
su
al in
sp
ectio
n
;
fiv
e
d
i
fferen
t fi
n
g
e
rp
rint q
u
a
lities v
i
z.
g
ood
,
dry,
wet, low
cont
rast, and s
t
ain are used.
Th
is sim
i
lar criterio
n
was also
used
i
n
[1
1
]
.
Fi
gu
re
5 t
o
Fi
gu
re
9
sho
w
fi
ve set
s
of fi
n
g
e
r
p
r
i
n
t
s
t
a
ken be
f
o
re
and aft
e
r t
h
e
segm
ent
a
t
i
on pr
ocess t
h
at
rep
r
ese
n
t
t
h
e abo
v
e
situ
atio
n
.
Fig
u
re
5
.
Resu
l
t
o
f
th
e seg
m
en
tatio
n
pro
cess
o
f
goo
d qu
ality fing
erprin
t
Fi
gu
re
6.
R
e
sul
t
of
t
h
e se
gm
ent
a
t
i
on p
r
ocess
of
d
r
y
fi
nge
rp
r
i
nt
Fi
gu
re
7.
R
e
sul
t
of
t
h
e se
gm
ent
a
t
i
on p
r
ocess
of
wet
fi
n
g
er
pr
i
n
t
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
J
ECE Vo
l. 5
,
N
o
. 5
,
O
c
tob
e
r
20
15
:
120
2
–
12
15
1
211
Fi
gu
re
8.
R
e
sul
t
of
t
h
e se
gm
ent
a
t
i
on p
r
ocess
of
l
o
w c
ont
ra
st
fi
n
g
er
p
r
i
n
t
Fig
u
re
9
.
Resu
l
t
o
f
th
e seg
m
en
tatio
n
pro
cess
o
f
fing
erprin
t wh
ich
co
n
t
ai
n
in
k
stain
s
an
d h
a
ndwritten
characte
r
s
Ove
r
al
l
,
base
d on
t
h
e
a
b
ove
fi
g
u
r
es,
t
h
e pr
op
ose
d
segm
ent
a
t
i
on t
echni
q
u
e has
per
f
o
rm
ed
ex
cep
tion
a
lly
well in
m
o
st c
a
ses esp
ecially fo
r
g
ood
, dry an
d
wet p
r
i
n
ts. Th
e foreg
r
ou
nd
s are well separated
fr
om
t
h
e back
gr
o
u
n
d
. H
o
we
ver
,
i
n
som
e
l
o
w co
nt
rast
pr
i
n
ts, there are certain areas
o
f
t
h
e fo
re
gr
ou
n
d
s
have
b
een wr
ong
ly mar
k
ed
as b
ack
gro
und
s. Li
kew
i
se,
f
o
r
st
ai
n
pri
n
t
s
, t
h
e
r
e
are s
o
m
e
areas of
bac
k
gr
ou
n
d
have
b
een
falsely lab
e
lled
as
foreg
r
ou
nd
.
Desp
ite th
e im
p
e
rfections, t
h
e se
gmented im
age
s
or
fore
grounds are
d
e
fi
n
itely well su
ited
t
o
facilitate th
e sub
s
equen
ce
p
o
s
t
-
pro
c
essin
g
in
clud
ing
ri
d
g
e
orien
t
at
io
n
field
estimatio
n
an
d
singu
lar po
in
t d
e
tectio
n.
Besid
e
th
e
h
u
man
in
sp
ecti
o
n
,
wh
ich
is co
n
s
i
d
ered
as
a qu
alitativ
e measu
r
e, altern
ativ
ely, th
e
assessm
en
t can b
e
carried
ou
t qu
an
titativ
ely su
ch
as b
y
coun
tin
g th
e nu
mb
er of
false and
m
i
ssed
fing
erprin
t
feat
ure
s
l
i
k
e
m
i
nut
i
ae or si
ng
ul
ar
p
o
i
n
t
s
[2
].
As for the sin
g
u
l
ar
po
in
ts; th
e p
e
rforman
ce is
m
e
a
s
u
r
ed
according to t
h
e ratio
of
num
ber of true s
i
ngular
poi
nt
s that have
bee
n
discarded t
o
the total num
b
er of
g
e
nu
in
e
sing
u
l
ar po
in
ts th
at
ex
isted
in
t
h
e p
r
i
n
t. In
ot
he
r w
o
r
d
s, t
h
i
s
m
easurem
ent
is eq
ui
val
e
nt
o
f
t
h
e
perce
n
t
a
ge
of
t
h
e di
scar
de
d
gen
u
i
n
e si
n
g
u
l
a
r
poi
nt
s.
Al
t
e
rnat
i
v
el
y
,
t
h
e assessm
ent
can al
so b
e
do
ne
according to t
h
e ratio
betwe
e
n num
b
er of falsely accepted singular poi
nts and total num
ber
of
ge
nui
ne
si
ng
ul
ar
p
o
i
n
t
s
.
Th
us, i
n
or
de
r
t
o
eval
uat
e
t
h
e
pe
rf
orm
a
nce
of
t
h
e
p
r
o
p
o
se
d se
gm
ent
a
t
i
on t
ech
ni
que
i
n
t
e
rm
s of t
h
e
ab
ov
e
qu
an
titativ
e m
easu
r
e,
an
ex
p
e
rim
e
n
t
is set up
u
s
i
n
g
5
0
0
p
r
i
n
ts
of th
e NIST-DB1
4 (i.e. f00
0
0
001
to
f000
050
0). In ad
d
ition
to
th
at, th
e techn
i
qu
e is also b
e
n
c
h
m
ark
e
d ag
ain
s
t sev
e
ral well estab
lish
e
d
segm
ent
a
t
i
on
m
e
t
hods i
n
cl
u
d
i
n
g l
o
cal
m
ean
of
grey
-sc
a
l
e
based
t
echni
que
, l
o
cal
vari
a
n
ce
of
g
r
adi
e
nt
m
a
gni
t
ude
, a
n
d a
com
b
i
n
at
i
o
n
of
l
o
cal
m
ean
of
g
r
a
d
i
e
nt
m
a
gni
t
ude
a
n
d
bl
oc
k
co
here
n
ce ap
pr
oac
h
by
Zha
n
g
and
Ya
n.
T
h
e c
o
r
r
es
po
n
d
i
n
g r
e
sul
t
s
are
gi
ve
n i
n
Ta
bl
e 1
.
Evaluation Warning : The document was created with Spire.PDF for Python.