TELKOM
NIKA
, Vol.11, No
.3, March 2
0
1
3
, pp. 1414 ~ 1421
ISSN: 2302-4
046
1414
Re
cei
v
ed O
c
t
ober 1
6
, 201
2; Revi
se
d Ja
nuary 16, 20
1
3
; Acce
pted Janua
ry 2
7
, 20
13
Gray-scale Edge Detection and Image Segmentation
Algorithm Based on Mean Shift
Li Zhengzho
u*
1
, Liu Mei
1
, Wang Huiga
i
1
, Yang Yan
g
1
, Chen Jin
1
, Jin Gang
2
1
Colle
ge of Co
mmunicati
on E
ngi
neer
in
g, Ch
ong
qin
g
Un
iver
sit
y
, Cho
n
g
q
in
g, 4000
44, Ch
i
n
a
2
Chin
a Aerod
y
namics R
e
sear
ch and D
e
ve
lo
pment
Ce
nter, Mian
ya
ng, Sic
hua
n, 621
00
0, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: lizhe
ngzh
o
u
@
cqu.e
du.cn
A
b
st
r
a
ct
T
o
solve the p
r
obl
em
of the i
naccur
a
te seg
m
e
n
ta
tio
n
for the gray i
m
ag
e
,
a mod
i
fie
d
al
gorith
m
base
d
on th
e me
an sh
ift is in
troduce
d
. T
he mo
difi
ed a
l
gor
it
hm c
onstructs
a nove
l
kern
el f
unctio
n
histo
g
r
a
m
by co
mbi
ng th
e pos
ition
infor
m
ati
on
and t
h
e
gray-scal
e
i
n
formatio
n
of a
pixel,
and th
en
mak
e
s us
e of
th
e
me
an sh
ift alg
o
rith
m w
i
th this
new
kernel fu
nction
hist
ogr
a
m
to a
u
to
matic
a
lly d
e
tect the
mo
des i
n
the g
r
ay-
scale i
m
a
ge, w
h
ich cou
l
d be c
onstructed ful
l
y
by the
kernel functio
n
defin
e
d
abov
e, filter
and se
gment the
gray i
m
a
ge. Ex
peri
m
e
n
ts bas
ed o
n
a gray
i
m
a
ge w
i
th
gro
und
backgr
o
u
n
d
are carr
ied
o
u
t by Can
n
y, Sobe
l
and the pr
opos
ed mea
n
shift meth
od, an
d the results show
that the mea
n
shift algor
ith
m
could effective
l
y
extract not onl
y bright o
b
ject
but also w
eak
object, an
d th
e result of the
introd
uced
alg
o
rith
m is
more
fi
t
factual scen
e
than that of the
usua
l seg
m
e
n
t
a
ti
on a
l
g
o
rith
m such as Can
n
y
and Sob
e
l a
l
gorith
m
.
Ke
y
w
ords
:
image segm
entation, mean sh
ift, kernel pr
obability dens
ity
function, gray-scale image
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Image se
gme
n
tation or obj
ect extractio
n
in natur
e sce
ne is the sub
s
tantial found
ation of
su
ch tasks a
s
feature extraction, pattern rec
ognition
and target tra
cki
ng, and it
has be
en wid
e
ly
applie
d in ma
ny fields.
The ima
ge
segmentatio
n
essentially i
s
the f
eatu
r
e
clu
s
terin
g
in
some
on
e sp
ace. T
he
aim of the image segme
n
tation is to se
parate
an im
age into som
e
indep
end
en
t, conne
ctive and
meanin
g
ful
p
a
rts, whi
c
h are co
rrespo
nding
to
the
s
e real obje
c
ts
a
nd ba
ckgroun
d.
Fo
r
the
objects are complicated themselv
es and are al
so affected by
nature scene, illumination and
shadow, image segmentation is
still an elementary
problem at present. O
ne serious
problem of
image
seg
m
entation i
s
h
o
w to
determine the
a
m
ount of o
b
j
ects
und
er t
he un
su
pervi
sed
clu
s
terin
g
scheme. Base
d
on the fact that objec
t
s
are uniform ge
nerally and t
here a
r
e obvi
ous
comm
on bo
unda
rie
s
am
ong differe
nt cluste
rs, th
e edge d
e
te
ction is o
n
e
of the primary
clu
s
terin
g
me
thods. Althou
gh many cla
ssi
cal ed
ge detectio
n
alg
o
rithm
s
, such
as Canny a
n
d
Sobel alg
o
rit
h
m, co
uld eff
e
ctively su
pp
ress
noi
se
a
nd lo
cate o
b
j
e
ct’s
edg
e, they co
uldn’t
yet
achi
eve gratif
ying result an
d perfo
rman
ce of e
liminating nature sce
ne, and over-segm
entation
or
lack-seg
ment
ation might h
appe
n if image info
rm
atio
n and p
r
io
r knowl
edge
co
uld not be fu
lly
utilized. Simu
ltaneou
sly, the bou
nda
rie
s
betwe
en
obj
ect an
d b
a
ckgrou
nd
are
b
l
urry a
nd eve
n
discontinuous for the
illumination’
s uniformity.
Th
is will resul
t
in
that some segmentation
algorith
m
s ba
sed upo
n
the
edge’
s conti
nuity
coul
dn’t
get sati
sfying re
sult [1]. At the sam
e
time,
some
se
gme
n
tation alg
o
rit
h
ms,
which a
r
e ba
se
d u
p
o
n
the hi
stogra
m
of t
he gray image, such
as
the maximum
entro
py met
hod, la
ck ne
cessary
ro
b
u
st
ness in
ap
plication
wh
en t
he obj
ect
s
h
a
v
e
multi-gray level usually [2,
3].
The mea
n
shi
ft [4] is a nonpara
m
etri
c st
atistical m
e
th
od for see
k
in
g the nea
re
st mode of
a point samp
le distrib
u
tion
by estimating
the density
gradie
n
t, and has be
en widely utilize
d
recently in pa
ttern analy
s
is, espe
cially in
the
col
o
r im
a
ge segme
n
tation [5, 6] and
target tra
c
kin
g
[7, 8]. The gray image
onl
y has
sp
ace i
n
formatio
n a
nd
inten
s
ity informatio
n, a
nd it is l
a
ck o
f
the
colo
r inform
ation. This woul
d indu
ce that the
origin
al m
ean shift method, whi
c
h m
a
ybe ha
s bee
n
applie
d successfully to se
gment the
co
lor ima
ge,
co
uldn’t effectiv
ely extract th
e obje
c
ts in t
h
e
gray-scale im
age [9].
In this pa
per,
an effective
algorith
m
ba
sed on m
ean
shift with a
n
o
vel ke
rnel f
unctio
n
histog
ram, which is con
s
tructed by com
b
ining t
he sp
ace informati
on and intensity information, is
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Gra
y
-scale E
dge Detectio
n and Im
age Segm
entation Algorithm
Base
d on … (L
i Zheng
zho
u
)
1415
introdu
ce
d. It coul
d autom
a
t
ically se
ek th
e mode
s, sup
p
re
ss
scen
e
and extra
c
t o
b
ject fro
m
gra
y
-
scale image.
Finally, the experime
n
tal result
s of
Can
n
y, Sobel and the propo
sed mean shift fo
r
the actu
al im
age
with the
compl
e
x ba
ckgroun
d sho
w
that the l
a
tter
could
effectively extract not
only bright ob
jects b
u
t also
wea
k
obj
ect
s
.
2. Mean Shift-Bas
e
d Image Segmentation
The me
an
shift is a sim
p
le an
d non
para
m
etri
c te
chni
que fo
r seeki
ng the m
ode
s by
estimating th
e gradie
n
t of the probabil
i
ty dens
ity fu
nction (P
DF). The image segmentatio
n
algorith
m
based on mea
n
shift is a
mathemati
c
mappin
g
, by
whi
c
h the gray-scale ima
ge is
cla
s
sed a
s
some definite
mode
s, and i
t
could be di
v
i
ded into four succe
s
sive steps, i.e. kernel
prob
ability density functi
on estimatio
n
, mode
de
tection, imag
e segm
entat
ion and regi
on
mergi
ng. No
w, let’s introd
uce the first p
a
rt, t
he kernel
proba
bility density functio
n
estimation.
2.1. Kernel P
r
obabilit
y
Densit
y
Function Estimation
The ke
rnel probability den
sity function estimati
on m
a
ke
s use of the pixels in the local
area to
esti
mate the probability den
sity function.
Given
n
pixels
i
x
,
1,
,
in
=
in
the
d
-
d
i
me
ns
io
na
l s
p
ac
e
d
R
, the m
u
ltivariate
kernel p
r
oba
bility density fun
c
tion e
s
timat
o
r
with kern
el
function
H
K
and a symmetric
positive defini
t
e
d
d
dimensio
n
a
l bandwi
d
th matrix
H
is
given
by
4
(
)
(
)
1
1
ˆ
n
i
i
fK
n
=
=-
å
H
xx
x
(1)
The ke
rnel fu
nction
H
K
is a nonneg
ative function with ze
ro as cente
r
and integral b
e
one in its
definitional re
gion. Usually,
the ker
nel function h
a
s
su
ch a form a
s
(
)
(
)
12
12
KK
-
-
=
H
xH
H
x
(2)
Whe
r
e
K
is a boun
ded fun
c
tion with co
mpact supp
o
r
t. The cho
o
s
ing of the kernel fun
c
tio
n
woul
d directly
affect the
re
sult of ima
g
e
seg
m
ent
atio
n. Accordi
ng
to the theo
re
m of the patt
e
rn
clu
s
terin
g
, the further the
data is from t
he ce
nt
er of the pattern, th
e litter the probability of being
this pattern would be
8
. So the functio
n
K
is commonly
symmetri
c
al
and regressiv
e
su
ch a
s
the
followin
g
equ
ation.
(
)
(
)
2
,
kd
Kc
k
=
xx
(3)
Whe
r
e
,
kd
c
is the normali
zatio
n
con
s
tant, and
(
)
k
x
, the profile of the kern
el, is consecu
t
ive an
d
differential wit
h
in the definit
ion regi
on.
To redu
ce th
e calculation
compl
e
xity, th
e band
width matrix
H
is usu
a
lly either diagonal
matrix
22
1
,,
d
di
ag
h
h
éù
=
ëû
H
or pro
portion to the identity ma
trix
2
h
=
HI
, where
I
is the identity
matrix. The supe
rio
r
ity of the later equ
at
ion
is that there is o
n
ly one band
width
para
m
eter
h
sh
ould be p
r
ovi
ded.
Whe
n
the band
width ma
trix
H
is the proportio
n
to the identity
matrix
2
h
=
HI
, th
e
kernel p
r
ob
ab
ility density function i
s
es
ti
mated as the
followin
g
expression.
(
)
2
,
11
1
ˆ
nn
kd
ii
dd
ii
c
fK
k
hh
nh
nh
==
æö
æö
--
ç÷
ç÷
==
ç÷
ç÷
ç÷
èø
èø
åå
xx
xx
x
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 3, March 20
13 : 1414 – 1
421
1416
The gray
-sca
le image only
has the posi
t
ion informati
on
(
)
,
x
y
and the intensity information
u
,
and these two feature
s
co
ns
tru
c
t
s
a feature vector
(
)
()
(
)
(
)
()
,,
,
,
ux
y
u
x
y
=
xx
. Usually, the pixel,
whi
c
h belon
g
s
to same one pattern, is
con
g
re
gative
in space, and
the
further the pixel is fr
om
the center of
the pattern, the litter the probability
of b
e
ing the pattern woul
d be. Accordi
ng to the
fact just ment
ioned a
bove, the profile ab
out positio
n
x
coul
d be expressed a
s
(
)
(
)
2
1
2
11
0
k
othe
rw
ise
p
ì
-£
ï
=
í
ï
î
xx
x
(5)
Dori
n
6
adopt
s the krone
cker d
e
lta fun
c
tion
to describ
e the pro
f
ile function
about
intensity feature.
(
)
()
(
)
()
2
ku
u
d
=
xx
(6)
In fact, the probability of th
e pixel, which
intensity is
e
qual to that of the cent
ral pixel, is
very little
within the
sp
atial band
wid
t
h
h
, and that
the data abi
des
by the Gau
ssi
on fun
c
tion
with the
intensity of th
e central pixe
l as center
is
more like. So, the profile functio
n
(
)
(
)
2
ku
x
abo
ut the
intensity co
ul
d be den
oted
as
(
)
()
(
)
(
)
(
)
1
2
2
2
1
2e
x
p
1
2
0
uu
ku
ot
he
rwi
s
e
p
-
ì
æö
ï
ç÷
-£
ï
ç÷
=
í
èø
ï
ï
î
xx
x
(7)
By combing
the formula
s
(5) a
nd (7
),
i.e.
the intensity and the
position, the
kernel
prob
ability de
nsity function
estimato
r for
gray-scale im
age could b
e
denote
d
as
(
)
(
)
(
)
2
2
,
12
1
ˆ
n
kd
i
i
d
i
c
uu
fk
k
nh
s
=
æö
æö
-
-
ç÷
ç÷
=
ç÷
ç÷
ç÷
ç÷
èø
èø
å
xx
xx
x
h
(8)
Whe
r
e
is the intensity ban
dwidth.
Once the kernel probabilit
y density fun
c
tion
(
)
ˆ
f
x
is estim
a
ted by the formul
a (8
), the
followin
g
step
is to find the mode
s in the gray imag
e.
2.2. Mode De
tec
t
ion
The modes l
o
cate
at the
posit
ion, where the value
of ker
nel probability density function
estimator is maximum
or minimum.
Therefor
e the pro
c
ed
ure of the mode
s detecti
on is
searching the zero point of
the gradient of the probability
density function estimat
o
r,
namely
(
)
0
c
f
Ñ=
x
, where
c
x
is the po
sition of the mod
e
.
As we kn
own
,
the gradient
estimation of t
he probabili
ty density fun
c
tion is the gradie
n
t
of the kernel den
sity estimator, i.e.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Gra
y
-scale E
dge Detectio
n and Im
age Segm
entation Algorithm
Base
d on … (L
i Zheng
zho
u
)
1417
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
(
)
2
2
,
12
2
1
2
2
12
2
2
1
,
12
2
2
2
1
12
ˆ
ˆ
2
2
n
kd
i
i
i
d
i
n
i
i
i
i
n
kd
i
i
d
i
i
i
ff
c
uu
kk
nh
uu
kk
c
uu
kk
nh
uu
kk
ss
s
ss
s
+
=
=
+
=
Ñ=
Ñ
æö
æö
-
-
ç÷
ç÷
¢
=-
ç÷
ç÷
ç÷
èø
èø
æö
æö
-
-
ç÷
ç÷
¢
ç÷
ç÷
æö
ç÷
æö
-
-
èø
ç÷
èø
ç÷
¢
=
ç÷
ç÷
æö
ç÷
æö
-
-
èø
ç÷
èø
ç÷
¢
ç
ç÷
ç
èø
èø
å
å
å
xx
xx
xx
xx
h
xx
xx
x
h
xx
xx
h
xx
xx
h
1
n
i
=
éù
êú
êú
êú
-
êú
êú
êú
÷
êú
÷
êú
ëû
å
x
(9)
The later te
rm of the formula (9
) is the
mean shift vector
4
,
and it could be exp
r
e
s
sed a
s
(
)
(
)
(
)
(
)
(
)
2
2
12
1
2
2
12
1
n
i
i
i
i
n
i
i
i
uu
kk
MG
uu
kk
s
s
=
=
æö
æö
-
-
ç÷
ç÷
¢
ç÷
ç÷
ç÷
èø
èø
=-
æö
æö
-
-
ç÷
ç÷
¢
ç÷
ç÷
ç÷
èø
èø
å
å
xx
xx
x
h
xx
xx
xx
h
(10
)
Let define
{
}
j
y
,
1,
2
,
j
=
, be the ce
nte
r
se
quen
ce
of the kernel
proba
bility den
sity
func
tion es
timator. If
x
is equal to
j
y
, the
next center
1
j
+
y
will be equal
to
j
y
plus the mean
shif
t
v
e
ct
or
(
)
MG
x
expresse
d as
(
)
1
jj
MG
+
=+
yy
x
(11
)
The mea
n
shifts alway
s
points to
wa
rd the dire
cti
on of maxim
u
m increa
se
in the
probability density function
at the ste
epe
st velocit
y
. In t
he low de
nsity areas
su
ch
as the
boun
dary of
an obj
ect, th
e velocity is
greate
r
tha
n
that in the hi
gh de
nsity area such a
s
the
interio
r
of the
obje
c
t. Whe
n
the velo
city is very
slo
w
or
clo
s
e to
zero at the l
o
catio
n
cj
x
, the
mean shift proce
dure co
ul
d stop and a
new mo
de
(
)
{}
,
cj
cj
u
xx
,
1,
2
,
j
=
of the image would be
detecte
d.
After these
mode
s in th
e gray-scal
e
image
have
been
dete
c
ted, the sub
s
eq
uent
pro
c
ed
ures a
r
e to segmen
t image and to merge the small mode i
n
to the
simila
r
an
d
g
r
eat
m
ode
r
e
spec
tively.
2.3. Image Segmentation
Let define
i
z
,
1,
2
,
i
=
,be the image pi
xel segm
ente
d
by the mea
n
shift proced
ure
s
. If a
pixel
i
x
belong
s to
the mode
(
)
{
}
,
cj
cj
u
xx
, the gray
value of
i
z
is equal to
the gray value of the
j
th mode ,i.e.
(
)
(
)
,
ii
c
j
u
=
zx
x
(12
)
2.4.
Region Merging
After the co
u
r
se
of imag
e segm
entation
,
ev
ery pixel of the gray
-scale im
age
h
a
s b
een
belon
ged to
one
co
rrespondi
ng mo
d
e
, and
ha
s
been
set
on
e co
rrespon
ding g
r
ay va
lue.
Actually, there are m
any si
milaritie
s
am
ong
some
m
o
des, a
nd so it is ne
ce
ssary
to merge th
e
s
e
spe
c
ial
mode
s, nam
ely to
inco
rpo
r
ate t
he pixel
s
of t
hese si
milar
mode
s into
o
ne mu
ch
big
ger
regio
n
, in ord
e
r to perfe
ctly describ
e the
s
e obj
ect
s
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 3, March 20
13 : 1414 – 1
421
1418
If the difference
s
between
two mode
s
(
)
(
)
,
cj
c
j
u
xx
and
(
)
()
,
ck
c
k
u
xx
are littler than not
only the
spa
c
e b
and
width
h
but al
so th
e
gray-scale
b
and
width
, the two m
ode
s coul
d b
e
inco
rpo
r
ate
d
into a new p
a
ttern. Usually
, these
mod
e
s
with ma
ny pixels could
be re
se
rved, but
the other mo
des with little pixels, su
ch a
s
noi
se, are filtered out wit
h
the spa
c
e b
and
width
h
and
the gray-scal
e
band
width
increa
sing.
More
over, if a mode
cont
ains le
ss tha
n
M
pixels, it could be m
e
rg
ed into the n
eare
s
t
pattern in spac
e.
3. Experiment Re
sult an
d Analy
s
is
One g
r
ay-sca
le image
with
gro
und
ba
ckgrou
nd i
s
ad
opted to eval
uate the p
e
rf
orma
nce
of edge dete
c
tion an
d image segme
n
tation betwee
n
two cla
s
sical segm
entat
ion algo
rithm
s
,
Can
n
y and Sobel, and the
mean shift algorithm introd
uce
d
in this p
aper.
Figure 1 is t
he ori
g
inal g
r
ay-scale im
age,
and th
e
r
e are sky, mountain, g
r
ass land,
bottomland (some little section in the
grass lan
d
),
a tree and an airpl
ane, whi
c
h is in the
comm
on bo
u
ndary bet
wee
n
the mountai
n and the tre
e
.
Figure 1. The
original g
r
ay-scale imag
e
Figure 2 is th
e edge im
ag
e extracte
d b
y
the
function
edge (’
ca
nn
y’), Canny al
gorithm,
with the aut
omatic p
a
ra
meter in MA
TLAB envir
o
n
ment. This
result sh
ows that the Ca
nny
algorith
m
cou
l
d effectively sup
p
re
ss the
stationa
ry
sky and mo
untai
n, but is very
sen
s
itive to the
non-station
a
ry and textured gra
ss la
nd,
where t
here
are many u
nord
e
rly edg
es. The ed
ge
s of
the airplan
e
are su
bme
r
g
ed in these unorde
rl
y edges, and this would indu
ce the succe
s
sive
target re
co
gni
tion and tra
c
king ba
sed o
n
the regio
n
fea
t
ure analy
s
is
very difficult.
Figure 3 i
s
the ed
ge ima
ge extra
c
ted
by t
he functi
on ed
ge (’
so
bel’), Sob
e
l a
l
gorithm,
with the auto
m
atic pa
rame
ter in MATLA
B
environme
n
t. From this
Figure, it is shown that Sobel
algorith
m
cou
l
d also effecti
v
ely suppre
s
sed the
statio
nary sky and
mountain, but isn’t sensit
ive
enou
gh to an
d coul
dn’t extract the
s
e lo
w co
ntra
st co
mmon bo
und
arie
s betwee
n
the sky an
d
the
mountain, the
bottomland a
nd the tree.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Gra
y
-scale E
dge Detectio
n and Im
age Segm
entation Algorithm
Base
d on … (L
i Zheng
zho
u
)
1419
Figure 2. The
edge imag
e extracted by
Can
n
y algorit
hm
Figure 3.The
edge ima
ge e
x
tracted by Sobel alg
o
rith
m
Figure 4 i
s
the se
gmente
d
result by the
mean
shift a
l
gorithm i
n
tro
duced in thi
s
pape
r,
and the m
a
in
para
m
eters
are
spatial
b
and
wi
dth be
five, gray ba
ndwi
d
th be
seven, and
M
be
ten. Figure (a) is the filtered image, Fi
gure (b
) is th
e segm
ented
image, and Figure (c) is the
edge imag
e. The Figure (a
) sho
w
s
that the non-statio
nary and textur
e grass lan
d
are effectively
smooth
ed a
s
con
n
e
c
tive re
gion
s. From the Figu
re
(b
), it is known t
hat
the sky a
nd the mount
ain
are also effectively suppre
s
sed, and the lo
w contra
st comm
on bou
ndari
e
s bet
ween the sky and
the mountain
,
the bottomland and the tree, could
b
e
extracted
well and truly
.
The Figure (c)
sho
w
s that the edge
s of these obj
ect
s
, su
ch as t
he sky, the mountain, the tree,
the airplan
e
, the
gra
ss la
nd an
d the bottomland, are effe
ctively ex
tracted, and all bou
ndari
e
s a
r
e
cl
ose a
nd full.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No
. 3, March 20
13 : 1414 – 1
421
1420
(a)The filtere
d
image
(b) T
he segm
ented imag
e
(c) The e
dge
image
Figure 4.The
edge ima
ge e
x
tracted by m
ean shift
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Gra
y
-scale E
dge Detectio
n and Im
age Segm
entation Algorithm
Base
d on … (L
i Zheng
zho
u
)
1421
As we
kno
w
n
,
there are m
a
inly three pa
ramete
rs a
d
o
p
ted to evalu
a
te the perfo
rmance
of image seg
m
entation alg
o
rithm, and they are t
he region co
nsi
s
t
ency, the regi
on contrast a
n
d
the regio
n
sh
ape [9-12]. For the re
gion
sha
pe ne
ed a
refere
nce se
gmentation i
m
age, the re
gion
con
s
i
s
ten
c
y and the regio
n
cont
ra
st are adopt
e
d
he
re. Fro
m
the
segm
ented i
m
age d
e
scri
b
ed
above, the re
gion co
nsi
s
te
ncy extracte
d
by the
introduce
d
mean shift is more consi
s
tent, and
the re
gion
co
ntrast i
s
lo
we
r than
th
at of the Ca
nny a
nd the So
bel
.
So the obj
ects extra
c
ted
by
the introdu
ce
d mean shift
fit actual scen
e than
that extracted
by the usual
segme
n
tatio
n
algorith
m
such as
Ca
nny a
nd Sobel
algo
rithm. So the
introdu
ce
d m
ean
shift algo
rithm is
more
fit
factual sce
n
e
than the usu
a
l segm
entati
on algo
rithm
su
ch a
s
Ca
nn
y and Sobel a
l
gorithm.
4. Discussio
n
and Con
c
lusion
The usual se
gmentation al
gorithm
s u
s
u
a
lly
have the fault of ov
er-segm
entation
and/or
lack-seg
ment
ation, and
co
uldn’t entirely
and a
c
cura
t
e
ly extract the obje
c
ts in
the gray-scal
e
image for it is lack of the color inform
ation. T
he introduced mean
shift con
s
tru
c
ts a novel kernel
probability density function by co
m
b
ing
the position i
n
formation a
nd the intensi
t
y information of
the pixels, a
nd then m
a
kes u
s
e of m
ean
shift to
automatically detect t
he
model
s, filter and
segm
ent the
gray-scale im
age. Experim
ents ba
se
d o
n
one g
r
ay image with
ground b
a
ckg
r
o
und
is carri
ed out by the Canny
, Sobel and the introd
uced mean shift, and the re
sults show that the
prop
osed al
g
o
rithm n
o
t on
ly supp
re
ss the st
ron
g
an
d textured
ob
ject, but al
so
extract the
weak
obje
c
t effectively. The introdu
ced mea
n
shift algor
it
hm is more fit factual sce
n
e
than the usual
segm
entation
algorithm
su
ch a
s
Ca
nny and Sobel al
g
o
rithm.
Otherwise, the mean shi
ft based seg
m
entation
al
gorithm is sensitive to the spatial
band
width, and that the
spatial ban
d
w
idth is ada
ptively chose
n
accordi
ng to the statistical
cha
r
a
c
ter i
s
o
ne of the re
se
arch emp
h
a
s
es.
Ackn
o
w
l
e
dg
ements
This
work
wa
s su
ppo
rted i
n
part by Nati
onal Natural
Scien
c
e Fo
un
dation of Chi
na und
er
grant No. 6107
1191, and Cho
ngq
ing
Nature and
Sci
e
n
c
e
Fund
u
n
d
e
r
G
r
ant
s No.
CSTC201
1BB2048.
Referen
ces
[1]
Han
d
a
y
a
n
i T
,
Ari W
,
Nanik S. Optic Nerve Head Segm
e
n
tation Usi
ng
Hou
gh T
r
ansform and Active
Conto
u
rs.
TEL
K
OMNIKA
Indones
ia Jo
urna
l of Electrical En
gin
eeri
n
g
. 20
1
2
; 10(3): 53
1-5
36.
[2]
Du F
,
Shi W
K
,
Che
n
LZ
, etc. Infrared Ima
ge
Segme
n
tatio
n
w
i
t
h
2-D Ma
xi
mum Entrop
y
Method B
a
se
d
on Particle S
w
arm Optimization (PSO).
Pattern Rec
ogn
itio
n Letters
, 200
5
;
26(5): 597-6
0
3
.
[3]
T
ao
W
B
,
T
i
an JW
, Liu J. Ima
ge Segme
n
tati
on b
y
T
h
ree-le
vel T
h
resholdi
n
g
Based on Ma
xim
u
m F
u
zz
y
Entrop
y an
d Genetic Al
gorith
m
.
Pattern Recogn
ition
Letter
s
. 2003; 24(
16)
: 3069-3
0
7
8
.
[4]
Dorin C, Peter M. Mean Shift: A Robus
t Approach to
w
a
r
d
F
eature Spac
e Anal
ys
is.
IEEE T
r
ansaction
on Pattern An
a
l
ysis an
d Mach
ine Intel
lig
enc
e
. 2002; 24(
5): 603-6
18.
[5]
Pan C, Z
hen
g
CX, W
a
n
g
HJ
. Robust C
o
lor
Image Se
gme
n
tation B
a
sed
on Mea
n
Shift
and Mark
ed-
control
l
ed W
a
t
e
rshe
d Algor
ithm.
Proceed
in
gs of the Second Intern
atio
nal Co
nfere
n
c
e
on Machi
n
e
Lear
nin
g
an
d Cyber
netics
, Xi
’an. 20
03: 27
5
2
-27
56.
[6]
Gong YY, Hua
ng H, Ma JP. T
e
xture Image
Segm
entati
o
n
Based o
n
Gaussia
n
Mi
xture
Model
w
i
t
h
Patch Information.
Journ
a
l of Informatio
n
an
d
Computati
o
n
a
l
Scienc
e. 20
12
; 9(4): 889-89
8
.
[7]
Dorin C, Visa
natha
n R, Peter
M. Kernel-
base
d
Object T
r
acking.
IEEE T
r
ansaction
s on Pattern
Analys
is and M
a
chi
ne Intel
lig
e
n
ce
. 200
3; 25(
5): 564-5
77.
[8]
Hali
l I Cuce,
Ahmet E Ceti
n. Mea
n
-shi
ft
T
r
acking of Moving Ob
je
cts Using Mu
ltidime
n
si
ona
l
Histogr
ams.
SPIE proce
edi
n
g
of Si
gna
l a
n
d
Data Pr
ocessi
ng of S
m
all T
a
rgets
, Bell
in
gh
am. 200
4; 7
0
-
77.
[9]
Hon
g
YP, Yi JQ, Z
hao DB. Improve
d
Mea
n
Shift Segme
n
tation A
ppro
a
ch
for Natural Im
ages.
Applied
Mathe
m
atics a
nd Co
mputati
o
n
. 2007; 1
85(2)
: 940-95
2.
[10]
Z
hang YJ. A Surve
y
on Eval
u
a
tion Metho
d
s for Image Segmentatio
n.
Pattern Recognition
. 1996; 29
(8): 1335-
13
46
.
[11]
Ibaa J, Akram
UM, Anam T
.
Retina
l Imag
e
Preproc
essi
ng: Backgr
oun
d an
d Nois
e
Segme
n
tatio
n
.
T
E
LKOMNIKA Indon
esi
a
Jour
nal of Electric
al
Engin
eeri
n
g
. 2
012; 10(
3): 537
-544.
[12]
Liu
X
Y
, Lu
L, Chen
X
G. Ear Image Edge
Detection Algorithm Bas
e
d on Mean Shift.
Journal o
f
Information a
n
d
Co
mp
utatio
n
a
l Scie
nce
. 20
08; 5(4): 17
65-
177
0.
Evaluation Warning : The document was created with Spire.PDF for Python.