TELKOM
NIKA
, Vol.12, No
.4, Dece
mbe
r
2014, pp. 83
9~8
4
6
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v12i4.303
839
Re
cei
v
ed Au
gust 27, 20
14
; Revi
sed O
c
t
ober 1
4
, 201
4; Acce
pted
Octob
e
r 29, 2
014
Resear
ch on Image S
p
licing and Fusion Processing
Algorithm in Large Visual Field
Wu Jie*
1
, Fe
ng Zuren
1
,
Song Xiaoru
3
1,2
State Key
L
abor
ator
y
for M
anufactur
i
n
g
Syst
ems En
gin
e
e
rin
g
, Xi'
a
n Jia
o
tong U
n
iv
ersity
Xi'
a
n 71
00
49, PR Chi
n
a
1,3
School of Electronic Inform
ation En
gi
neer
i
ng, Xi'
an T
e
chnol
ogic
a
l Un
ive
r
sit
y
Xi'
a
n 710
03
2, PR Chi
n
a
*Corres
p
o
ndi
n
g
author, em
ail
:
xait_
b
s@1
63.
com
A
b
st
r
a
ct
T
o
obtai
n w
i
de
area
and
a l
a
rge fie
l
d of vi
e
w
im
ag
e, spl
i
ci
ng a
nd fusi
on
alg
o
rith
m is
pr
esente
d
.
Sing
le i
m
a
ge i
s
preprocess
e
d
by
uti
l
i
z
i
n
g
ro
ugh
match
i
ng
alg
o
rith
m,
w
h
ic
h
ca
n narrow
d
o
w
n
the
matchi
ng
rang
e to i
m
pr
o
v
e the sp
ee
d a
nd pr
ecis
i
on of
image stitc
h
in
g an
d fusio
n
, a
t
the same ti
me, Sing
le i
m
ag
e is
prepr
ocesse
d
by filter
proc
e
ssing
al
gorit
h
m
, w
h
ic
h w
ill
reduc
e i
n
terfer
ence
no
ise,
i
m
pr
ove
SNR
and
enh
anc
e the e
ffective charac
ter informatio
n
of imag
e;
the
best matchi
n
g
positi
on is foun
d by usin
g
a
combi
ned s
p
li
cing
alg
o
rith
m, w
h
ich are r
a
tio te
mp
late
match
i
n
g
al
g
o
rith
m a
nd te
mp
late
matchi
ng
alg
o
rith
m, and
the imag
es are
splice
d
at the best matc
hin
g
positi
on; w
e
take the nei
gh
bo
rhoo
d w
e
ighte
d
avera
ge fusio
n
algorit
hm to e
l
i
m
in
ate the di
stinct sp
licin
g trace. T
he capt
ured
i
m
a
ges a
r
e process
ed
by
usin
g corre
lati
on a
l
g
o
rith
m,
a lar
ge fi
eld
o
f
view
and
hi
g
h
qu
ality
i
m
ag
e is o
b
tai
n
e
d
. T
he exp
e
ri
me
nta
l
results verify the vali
dity of the
algor
ith
m
.
Ke
y
w
ords
:
image sp
lici
ng a
l
gorith
m
, i
m
a
g
e
fusion al
gorit
h
m
, templat
e
matchin
g
, imag
e process
i
ng
1. Introduc
tion
In the process of obtai
nin
g
info
rm
ation
,
vision is
a reflex whi
c
h i
s
mo
st intuiti
v
e, most
compl
e
te an
d most a
c
curate about th
e su
rro
undi
n
g
[1]. The informatio
ns o
b
tained by vision
occupy
mo
st of the
capt
uring
info
rma
t
ion. Ho
weve
r, in
many
case
s, h
u
ma
n
bein
g
s can
not
dire
ctly have
the acce
ss to the
inform
ation of the
surro
undi
ngs
by themselves. As
a ki
n
d
of
choi
ce, it can
be achi
eved
by capturin
g
image.
The
acq
u
isitio
n of image nee
d
s
the help fro
m
optical
sy
ste
m
. In order
to gain
high
er q
uality im
age
and
mo
re a
c
tual
su
rro
undi
ng
s, the
captu
r
ing
im
age
nee
d to
be
processed by
so
me
imag
e p
r
o
c
essing.
The
image
proce
ssi
ng
techn
o
logy
h
a
s
bee
n u
s
e
d
in m
any field
s
, ju
st a
s
,
we
can
test
the l
eakage
of pi
p
e
line
and
fix the
positio
n of le
aka
ge by im
age p
r
o
c
e
ssi
ng technolo
g
y
; and usin
g
image p
r
o
c
e
s
sing te
chn
o
lo
gy,
the plate nu
mber of the runnin
g
ca
r o
n
the road
wi
ll be get by;
and some tin
y
compon
ent
s ca
n
als
o
be rec
o
gniz
ed, ec
t [2].
In orde
r to g
e
t
the optimal
effect of visio
n
, image
act
s
as the t
r
an
si
tion media
of vision,
we
nee
d ima
ge
whi
c
h
can
refle
c
t a
c
tua
l
su
rroun
di
ng
s. Unde
r a
serie
s
of
pract
i
cal
engin
e
e
r
i
n
g
backg
rou
nd, one imag
e whi
c
h ca
n re
flect act
ual surroun
ding
s not only bases on the u
s
e
d
optical
sy
ste
m
, but al
so t
he
sub
s
e
que
nt image
pro
c
e
ssi
ng te
ch
nology fo
r o
r
iginal im
age
[3].
Acting a
s
th
e tran
sition
media of visi
on, the
imag
e sh
ould h
a
ve rich conten
t and large fi
eld
properties, only the iamge meet
s these demand, we ca
n get more view,
whi
c
h will hel
p us to
obtain i
n
tere
sting thing
s
. B
u
t due
to th
e
singl
e
came
ra defi
c
ien
c
y,
singl
e
came
ra ha
s li
mited
field
of view [4]-[5
], we ca
nnot
see
overall
view from
on
e
cam
e
ra. Pa
nora
m
ic im
a
ge, just li
ke
our
visual effe
ct, ca
nnot b
e
obtaine
d thro
ugh
one
s
h
o
o
t. To ove
r
come the
sho
r
tage, m
u
ltiple
came
ra
s
are
pla
c
ed
at d
i
fferent lo
cati
ons cove
rin
g
the e
n
tire fi
eld
contai
nin
g
the
effecti
v
e
informatio
n to obtain a l
a
rge field
of view ima
ge [6
]. The all ca
mera
s
can
g
a
in a g
r
ou
p
of
image
s, and t
here i
s
ove
r
la
p amon
g the
s
e image
s.
Via
image spli
cin
g
,
we can stitch
the overla
p
among
the
s
e
imag
es an
d
finally fuse
in
to a l
a
rg
e fiel
d of vie
w
and
co
mplete
ne
w im
age,
whi
c
h
contai
ns all i
n
formatio
n from the grou
p
of im
ages. Image fusi
on enha
nces the
visual asp
e
ct of
image which lay the founda
tion for huma
n
being
s to a
c
cess mo
re e
ffective information.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 839
– 846
840
2. Image preproces
sing
algorithm
2.1 Rough m
a
tching pr
oc
essing
In the ima
g
e
acqui
sition
pro
c
e
ssi
ng, t
here
will
be
dislo
c
atio
n p
henom
eno
n i
n
many
image
s [7]-[8
], so it is ne
ce
ssary to
calibrate
imag
e. We
sh
oul
d identify the
locatio
n
of t
h
e
feature
re
gio
n
of the
two i
m
age
s a
nd
calcul
ate
the
coordi
nate
differen
c
e
in th
e
feature
regio
n
between two images
. Acc
o
rding to coordinate
differenc
e, we adjus
t c
aptured images
and
prep
are for the followin
g
image splici
ng a
nd fusio
n
[9].
Becau
s
e th
e
optical
param
eters of the
camera
are co
nstant, the
r
e
are
som
e
correlation
on the overla
p prop
ortion
among the capture
d
imag
es and the si
milarity amon
g the pixels [10]-
[11]. Based o
n
the co
rrelat
ion, we
can
estimate
a
n
d
locate roug
h
l
y the locatio
n
of the overl
a
p
among th
e serie
s
of imag
es, whi
c
h
ca
n narro
w do
wn the matchi
n
g
ran
ge to im
prove the
sp
eed
and p
r
e
c
i
s
io
n of ima
ge
stitchin
g an
d
fusion.
The
schem
atic
d
i
agra
m
of
ro
ugh m
a
tchi
n
g
a
s
s
h
ow
n
in
F
i
gu
r
e
1
.
the
irrelevant area
of every image
o
v
erlaps of the
fou
r
imag
es
Figure 1. Sch
e
matic dia
g
ra
m of rough m
a
tchin
g
2.2 Filter pro
cessing algo
rithm
In the image
acquisition proces
s, It is inevitable that t
here
will
be
many different sort
s of
noise and di
stortion and
also
accompanying with redu
ction
of contrast, whi
c
h will
lead to
effective information in th
e image
were red
u
ced.
The imag
e q
uality will directly affect the
algorith
m
ic v
a
lidity and accuracy of sub
s
eq
uent im
a
g
e
stitchin
g an
d fusion [1
3]. If we stitch a
nd
fuse the ima
g
e
s which we
re dire
ct
ly coll
ected by
cam
e
ra, we often
can
not get th
e desi
r
e
d
re
sult.
So we
ne
ed t
o
prep
ro
cess the
colle
cted
image
s
by re
duci
ng inte
rfe
r
en
ce
noi
se, i
m
provin
g SNR
and en
han
cin
g
the effective cha
r
a
c
ter in
formation of i
m
age.
The me
dian
filtering is
a nonlin
ear
sig
nal
processin
g
tech
nology
based on th
e ord
e
r
s
t
atis
tical theory, it c
an effec
t
ively s
u
ppres
s
th
e no
ise, which i
s
a typical no
nlinea
r spati
a
l
filtering te
ch
n
o
logy [14]. T
he m
edian
fil
t
er
can
p
r
ote
c
t well th
e
si
gnal
detail
s
while
removi
ng
noise. It also
can
rem
o
ve singula
r
ities i
n
gray ima
ge.
Whe
n
the i
s
o
l
ated noi
se
e
x
ists in ima
g
e
s,
whi
c
h mea
n
s the number
of pixels occupied by t
he isolate
d
noise is small an
d the numbe
r of
pixels o
c
cupi
ed by target
is large. Un
d
e
r the
ci
rcum
stan
ce
s, it is suitable to
use the
medi
an
filtering
[15]. This pap
er u
s
e
s
the medi
an
filt
erin
g to
su
ppress
ba
ckgro
und
an
d enh
an
ce th
e
target, maki
n
g
the outline
of the image clea
re
r to
improve the qu
ality of the image. That will
be
con
d
u
c
ive to further
sea
r
ch for the matching lo
cation.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Re
sea
r
ch on
Im
age Splicin
g and Fu
sion
Processin
g
Algorithm
in Large Visual Fi
el
d (Wu Jie
)
841
3. The Image Splicing Algorithm Results and
Anal
y
s
is
3.1 Ratio T
e
mplate Ma
tc
hing Algorithm
In orde
r to ge
t a large field
of view image
, we need to
establi
s
h visu
al co
rrel
a
tion
among
the image
s
whi
c
h were
acq
u
ire
d
by came
ra
s pl
a
c
ed at different locatio
n
s. Via matchi
ng
prin
ciple, we
sho
u
ld com
-
p
a
re
the
l
e
vel of
simila
rity b
e
twee
n the
ta
rget
are
a
a
n
d
the
sam
e
size
area
fro
m
the
differe
nt search
area i
n
im
age
s, then
we ne
ed to
ide
n
tify the po
sition, where i
s
t
h
e
highe
st level
of simila
rity.
The po
sition
viewed a
s
th
e be
st spli
cin
g
po
sition [8]. Selecting t
w
o
image
s from t
he ima
g
e
s
ca
ptured
by
ca
mera
s, the
n
t
w
o set
s
of pixels we
re sel
e
cted at
interv
als
of a ce
rtain d
i
stan
ce in h
o
rizont
al di
re
cti
on, whi
c
h i
s
i
n
the ov
erl
a
p
of the first i
m
age. Th
e gray
ratio value
s
of the two sets of pixels w
ill be use
d
as the reference templat
e
, and the best
matchin
g
position from
the
overlap
of th
e
se
co
nd i
m
ag
e was sea
r
ch
ed. Be
cau
s
e
the g
r
ay val
u
e
s
of some
are
a
in different i
m
age
s are cl
ose, if t
he ref
e
ren
c
e tem
p
l
a
te wa
s sele
cted impro
p
e
r
l
y
,
it
is e
a
sy
to m
a
tch
wron
gly. The
key p
o
i
n
t to avoi
d
wrong
mat
c
hin
g
is h
o
w to
sele
ct effe
ctive
informatio
n, in the mea
n
time, red
u
ce the interfe
r
en
ce. The
co
ncrete imple
m
e
n
tation ste
p
s a
s
s
h
ow
n
in
F
i
gu
r
e
2
.
As sh
own in
Figure 2,
Im
agea
an
d
Im
aged
a
r
e the i
ndep
ende
nt area
s
without
overlap.
Imageb an
d Image
c are the overlap b
e
twee
n
Im
ageA
and
Im
ageB.
Step 1, We select two
set
s
of co
ntinuo
us fou
r
col
u
m
n
s of pixel
s
in the
Im
ageb
, and the
interval value
between the
two set
s
of pixels is
D.
The
gray ratio values of the two sets of pixe
ls
is
a unit, we
mak
e
the unit to be template
B
.
Step 2, Begi
nning th
e left
sid
e
of the
Image
c,
si
mil
a
r t
o
st
ep
1,
we
sele
ct
t
w
o s
e
t
s
of
contin
uou
s fo
ur column
s of
pixels, the int
e
rval value
b
e
twee
n the two
sets of pix
e
ls al
so i
s
D
. we
make the n
e
w
gray ratio
values
unit to be template
C1
.
(a)
(b)
Figure 2. Sch
e
matic dia
g
ra
m of ratio template matchi
n
g
algorith
m
Step 3, the
a
b
sol
u
te difference valu
e a
bout templ
a
te
B
an
d tem
p
late
C1
i
s
calcul
ated,
whi
c
h i
s
vie
w
ed a
s
tem
p
la
te
M1
. We
can obtai
n the
colu
mn ve
ct
or
summ
atio
n of the tem
p
late
M1
. And th
e
n
we
can
obt
ain the
evalu
a
tion fun
c
tion
XS1
which i
s
g
r
ay
simila
rity betwee
n
t
h
e
templates
B
a
nd
C1.
Step 4, the
two sets
of the pixels f
r
om
the
step
2 were
mo
ved re
sp
ecti
vely, the
movement is
from left to right a colu
mn
in hori
z
ontal
dire
ction. We
obtain the n
e
w pixel
s
, si
milar
to step 2, a n
e
w template
C2
was o
b
tai
ned. We
can
cal
c
ulate tem
p
late
M2
an
d
the evaluatio
n
function
XS2
whi
c
h i
s
gra
y
similarity b
e
twee
n the t
e
mplate
s
B
a
nd
C2
. In the s
a
me way, t
h
e
pixels were kept moving from left to rig
h
t in hor
i
z
o
n
tal dire
ction, t
hen we obtai
n two sets of
new
templates a
n
d
the evaluati
on functio
n
, whi
c
h are
Ci
,
Mi
and
XSi
(i=3,4,...).
Step 5, If the gray similarit
y
level of
the
two image
s is high, the value of the evaluation
function
is small. We
ne
ed to fin
d
th
e minim
u
m
of the eval
u
a
tion fun
c
tio
n
XSmin
an
d the
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 839
– 846
842
coo
r
din
a
tes o
f
the minim
u
m (x,y). Th
e
coo
r
din
a
tes p
o
sition
is the
best
po
sition
to spli
ce
Im
ageA
and
Im
ageB
.
The adva
n
ta
ge of the rati
o template m
a
tchin
g
algo
ri
thm is sim
p
le
and practi
ca
ble. The
spe
ed of the
algo
rithm is
fast. But the
algorith
m
will
lose
effica
cy
whe
n
the
bl
ack pixel
s
was
found in the template. Ho
wever, the template match
i
ng algo
rithm has wi
de u
s
a
ge. We ca
n find
the mat
c
hing
point
exactl
y by this
alg
o
rithm. In thi
s
p
ape
r,
we
co
mbine
th
e ratio templ
a
te
matchin
g
alg
o
rithm a
nd th
e template m
a
tchin
g
alg
o
ri
thm to accom
p
lish th
e ima
ge mat
c
hing.
We
use the templ
a
te matchin
g
algorith
m
wh
en there a
r
e
some bl
ack pi
xels in the image.
3.2 Template
Matching
Algorithm
In template
matchin
g
alg
o
rithm,
we
ne
ed to
select
the d
e
si
gnate
d
requi
reme
n
t
effective
informatio
n in
the first im
a
ge, these info
rmation
co
nta
i
n typical
regi
onal
cha
r
a
c
te
ristics, it is j
u
st
like
sh
ape,
te
xture, color a
nd
so
on.
We
get th
e m
a
th
ematical
d
e
scriptio
n
of the
s
e
pa
ramete
rs,
usin
g the
co
rrelation
theo
ry to se
arch
a
nd
co
mp
are
different a
r
e
a
s
in
the
se
co
nd ima
ge. Th
e
pixels co
rrela
t
ion of the two image
s is low, The
lowe
r the pixels correlation of the two image
s,
the lowe
r lev
e
l of simila
rity of the two image
s.
The
schemati
c
di
agra
m
of template mat
c
hing
algorith
m
as
sho
w
n in Fig
u
re 3.
(a)
(b)
Figure 3. Sch
e
matic dia
g
ra
m of template matching al
g
o
rithm
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Re
sea
r
ch on
Im
age Splicin
g and Fu
sion
Processin
g
Algorithm
in Large Visual Fi
el
d (Wu Jie
)
843
The co
ncrete
impleme
n
tation step
s de
scrib
ed bel
ow:
First, the sui
t
able templat
e
m(x,y) is
sele
cted in I
m
age
C. The
suitable
ref
e
ren
c
e
template is
condu
cive to redu
ce
cal
c
ul
ation of
the t
e
mplate m
a
tching al
gorith
m
[16], whi
c
h
can
improve th
e
matchin
g
spe
ed. The
sele
cted templ
a
te
must be vali
dity and diffe
ren
c
e.
We m
u
st
ensure
a
c
curacy a
nd th
e
feature info
rmation
fr
om
ori
g
inal
ima
ge i
s
p
r
e
s
e
n
t
ed reliably,
in
addition, th
e t
e
mplate
featu
r
e m
u
st
be
di
stinct
differe
n
t
from th
e oth
e
r
are
a
s of th
e imag
e [1
7], it
is better that the co
rrelation
betwee
n
the temp
late and
the rest p
a
rt of the image is low.
Sec
o
nd, template parameter was
s
e
t. We dete
rmi
n
e the si
ze of
the template
m(x,y),
whi
c
h is a
×
b
（：
）
unit
p
ixel
. And the first pixel coordi
nate of the template is
(x,y).
At last, the proce
s
s to sea
r
ch an
d match. T
he simila
rity is only part of the two image
s,
so the
r
e are
four dire
ctio
ns to use th
e te
mplate to match the
second im
a
ge, the moving
direc
t
ions
of matc
hing
template s
t
art at the
top-left, top-right, bottom-
left and bottom-right. The
template a
p
p
r
oa
che
s
th
e
se
co
nd im
a
ge by th
e f
our
direction
s
. Usin
g the
s
e
app
roa
c
hi
ng
dire
ction
s
is
condu
cive to redu
ce the
probabilit
y of wrong
and
omi
ssive m
a
tch
e
s
[18]-[19]. T
a
ke
starting
at bot
tom-left matching a
s
an
example, a
s
sh
own i
n
Figu
re
3,
im
ageD
is motionle
ss
a
nd
its size is A×B. The template approa
ch
to the
im
ageD
startin
g
at the bottom-le
ft and move to
right in hori
z
ontal dire
ctio
n. We cal
c
ul
ate the corr
el
ation of the different
overla
p, the function is
sho
w
n:
y
x
m
y
y
x
x
f
y
x
H
y
x
,
,
)
,
(
(1)
In formula (1),
a
A
x
,...,
2
,
1
,
b
B
y
,...,
2
,
1
. We
can
obtai
n t
he fun
c
tion
y
x
H
,
value by the
formula
(1
)
whe
n
any p
o
s
ition in th
e i
m
age
D
i
s
gi
ven. Whe
n
th
e value
of
x
and
y
have been
cha
n
g
ed, we
can
o
b
tain the who
l
e
y
x
H
,
value by the moving
of
the template
m
(
x,
y)
on the
im
ageD.
4. The Image fusion algor
ithm
In the image
spli
cing
pro
c
essing, b
a
se
d on a
bov
e
splicin
g algo
rit
h
m, we o
b
tai
n
the be
st
matching position. If we only overlap t
w
o images
at the best m
a
tching position, we
will find
distin
ct trail in two image
s across po
siti
o
n
. This will aff
e
ct the image
qaulity.
To elimin
ate
this
kind
p
henom
eno
n, we u
s
e
sm
o
o
thing al
gorit
hm to di
spo
s
e it. In
referen
c
e [18
], Shum.H Put forwa
r
d th
at the ce
ntra
l regio
n
of o
v
erlap i
s
vie
w
ed
as
smo
o
th
transitio
n reg
i
on. Weight i
s
viewed a
s
the di
stan
ce
betwe
en the
overlappi
ng
pixels and the
boun
dary. Th
en the gray value is
dispo
s
ed sm
oothly
by mathemati
c
al op
erat
ion,
which al
so can
solve the pixe
ls incohe
ren
c
e probl
em effectively.
The pixel values i
n
spli
cing p
o
sition are di
stinct
different fro
m
aroun
d pi
xel values. To prot
e
c
t most image
details, this paper u
s
e
s
the
neigh
borhoo
d
weighte
d
averag
e fusio
n
method to
dispose the disti
n
ct spli
cin
g
trail.
We sel
e
ct on
e
pixel
Q1(x,y)
, whi
c
h
is lo
cated in
Image
1 an
d i
s
not
locate
d in
the
overla
p
betwe
en Ima
ge1 a
nd Ima
ge2, ju
st like
Imagea i
n
fig
u
re
2. Keepin
g
the pixel va
lue
Q1
(x,y)
, we
can g
e
t formu
l
a (2):
y
x
Q
y
x
Q
,
,
1
(2)
Like
wi
se, a
n
o
ther pixel i
s
located i
n
I
m
age2
an
d i
s
n
o
t lo
cated
in the
ove
r
la
p bet
wee
n
Im
age1
and
Im
age2
, just like Image
d in
figure 2. Keeping the pix
e
l value
Q2(x,y)
, we
c
an get
formula (3):
y
x
Q
y
x
Q
,
,
2
(3)
Whe
n
the pixel is located
in the overla
p betwe
en
Im
age1
and
I
m
age2
, the formul
a is
sho
w
n a
s
bel
ow:
y
x
Q
y
x
Q
y
x
Q
,
-
1
,
,
2
1
(4)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 839
– 846
844
In formula
(4),
is the gradi
ent facto
r
, in the overla
p, according to
the dire
ction
fro
m
Image1 a
nd Image2, the
value ch
ang
e
s
from 1 to 0.
We can a
c
complish sm
o
o
th pro
c
e
s
sin
g
of overlap b
e
twee
n two image
s by using t
he ab
o
v
e algorithm.
To achieve
ideal smoot
h
pro
c
e
ssi
ng ef
fect, three
co
ndition
s: (1).
The intermed
iate gray stat
e sho
u
ld
kee
p
dra
b
sm
oot
h
cha
nge i
n
the
gra
dual
pro
c
ess. (2
). The
bord
e
r
cu
rve
of the interm
ediate g
r
ay
state sh
ould
try to
maintain s
m
ooth. (3). Charac
teris
t
ics
of target
sh
ould
be kept in the grad
ual pro
c
e
ss, no othe
r
extraneo
us fe
ature
s
are found.
5. Image processing re
su
lts and analy
s
is
This
pa
per
u
s
e
s
M
a
tlab software a
s
the im
age pro
c
e
ssi
ng platform. Firstly,
the four
captu
r
ed ima
ges a
r
e dispo
s
ed respe
c
tively As sho
w
n
in Figure 4, due to the limited field of view
of the cam
e
ra, The whole
scene
ca
nnot
be obtai
n
ed
by only one i
m
age. We ne
ed to splice a
n
d
fuse fou
r
im
a
ges to
get the
whol
e sce
ne.
We
use a pix
e
l as a u
n
it, the o
r
iginal
im
age
size a
r
e
all
756
×45
8
.
Firstly, doing
image prep
ro
ce
ssi
ng for fo
ur origi
nal im
age
s to redu
ce noise a
nd improve
SNR. We u
s
e median filtering to re
strain bac
kg
ro
u
nd and en
ha
nce the effective information
cha
r
a
c
teri
stics in the ima
ges, which can im
prove the quality of image and
prep
are for the
followin
g
ima
ge splicin
g a
nd fusio
n
. In the image
spl
i
cing p
r
o
c
e
s
si
ng, ba
sed o
n
above splici
ng
algorith
m
, we
obtain the be
st matchi
ng p
o
sition.
Figure 4. Fou
r
origi
nal ima
ges
As
s
h
ow
n in F
i
g
u
r
e
5
,
th
e
Im
age5
is
sp
lic
ed
b
y
p
r
oc
es
se
d
Im
ag
e1
an
d
Im
age2
, the
Im
age6
is spl
i
ced by
proce
s
sed
Im
age3
and
Im
age4.
We
can
see t
he di
stinct tra
il in two im
ag
es
across po
siti
on. In order t
o
eliminate the distin
ct splicing trace, it is
nece
s
sa
ry to smooth the
overlap
of image. The im
a
ge re
sult u
s
in
g the sp
li
cing
and fusi
on al
gorithm i
s
sh
own a
s
Imag
e7
in Figure 6, the size of the Image7 i
s
925
×52
2
.
(a)
(b
)
Figure 5. Splicing ima
g
e
s
by processe
d
original im
ag
es
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Re
sea
r
ch on
Im
age Splicin
g and Fu
sion
Processin
g
Algorithm
in Large Visual Fi
el
d (Wu Jie
)
845
Figure 6. Panorami
c
imag
e
without spli
ci
ng trail
In the Fi
gure
5, we
can
se
e that
anyon
e of
t
w
o im
ag
es
co
ntain
s
more
info
rma
t
ion than
anyone f
r
om f
our
origi
nal i
m
age
s, the
splicin
g al
g
o
rit
h
m ma
ke
s th
e image
ri
che
r
, but the
spli
cing
trail on
the i
m
age
s i
s
ve
ry distin
ct. To
eliminat
e
spl
i
cing
trail,
we
use
smoothi
ng al
gorith
m
to
disp
ose it in this pap
er. Th
e result is sh
own
a
s
figure
6
. After a seri
es of pro
c
e
ssing, the
Im
ag
e7
has b
een
en
han
ced in th
e visual. The
more
com
p
lete and
rich
conte
n
t can
be refle
c
ted i
n
th
e
Image7, whi
c
h is co
ndu
civ
e
to extract the effective
informatio
n in the large field o
f
view image.
6 Conclu
sions
This pa
pe
r u
s
e
s
spli
cin
g
a
nd fusio
n
alg
o
rithm, whi
c
h
provide
s
the
theoreti
c
al b
a
si
s and
impleme
n
tation method to obtain a large field of
view image
. In
the spli
cing an
d fusing
pro
c
e
ssi
ng f
o
r ma
ny ima
ges, firstly, we e
s
tima
te
roug
hly the range
of the overlap,
whi
c
h is
based o
n
the
physi
cal
stru
cture of
a
c
q
u
isition syste
m
. The roug
h
e
s
timation can redu
ce the tim
e
for su
bse
que
nt matching.
We u
s
e a co
mbined
spli
ci
ng algo
rithm
of ratio templ
a
te matchin
g
and
template matchin
g
in this paper, whi
c
h effectiv
ely
avoids the failure of the
algorithm. The
amount of eff
e
ctive feature
inform
ation
of template is increa
sed; we can
get the
best mat
c
hin
g
positio
n by the introdu
cin
g
simila
rity evaluation f
unctio
n
. Finally, we dispo
s
e the
spli
cing trail
by
usin
g the wei
ghted ave
r
ag
e method of f
u
sio
n
algo
ri
th
m, the image
transitio
n is
nature,
whi
c
h
is
fused
into a
large
field of
view ima
ge,
and th
e ima
ge contain
s
all inform
atio
n from th
e f
our
image
s. Th
e
experim
ental
re
sults
sho
w
that th
e al
gorithm
is re
aso
nabl
e a
n
d
fea
s
ible. I
m
age
fusion e
nha
n
c
e
s
the visual
aspe
ct of image;
the expe
riment ha
s a
c
hieved go
od result
s.
Ackn
o
w
l
e
dg
ements
This work was
partially supported
by
Ph.
D. Pro
g
ram
s
F
oun
dation of Mi
nistry of
Educatio
n of
Chin
a (2
0
1002
0111
003
1), Natio
nal
Natural Science Fou
n
d
a
tion of Chi
na
(611
051
26 &
6087
5043
), a
nd Spe
c
ial
F
ound
ation of
Pr
esi
dent of
Xi’an Te
chn
o
l
ogical University
(XGYXJ
J05
2
4
).
Referen
ces
[1]
W
u
fei. Res
ear
ch of
Larg
e
F
i
e
l
d of
Vie
w
Vi
de
o Pa
noram
a.
J
ourn
a
l
of Z
hej
i
ang
Un
iversity.
20
06; 4
0
(3):
395-
398.
[2]
Gaoai
hu
a, Z
huchu
an
gui. Si
mulati
on stu
d
i
e
s of comp
ou
nd e
y
e
ima
g
i
ng e
x
perim
ent
.
Journa
l o
f
Northw
estern
Univers
i
ty (nat
ural sci
ence e
d
i
tion)
. 19
98; 28
(2): 117-1
20.
[3]
Shum H Y, Szeliski R. S
y
ste
m
s and e
x
p
e
ri
ment
pap
er co
nstruction of p
anor
amic ima
g
e
mosaics
w
i
th
glo
bal a
nd l
o
ca
l alig
nme
n
t.
Internati
ona
l Jour
nal of Co
mpute
r
Vision
. 20
00;
36(2): 10
1-1
3
0
.
[4]
Z
hong
li, H
u
xia
o
feng. M
o
saic
Alg
o
rithm R
e
search
of Ove
r
lap
p
in
g Imag
e
.
Chin
ese
Jour
nal
of Imag
e
and Grap
hics
. 199
8; 25(5): 36
5-36
9.
[5]
Hans
han L
i
. Researc
h
on p
h
o
toel
ectronics
prop
erties
of arra
y
emittin
g
dio
de a
nd its light e
ner
g
y
distrib
u
tion i
n
d
e
tection scre
e
n
.
Optik
. 2014; 125(3): 10
96-
11
00.
[6]
Peters W
H
, Rans
on W
F
. Digita
l
im
agi
ng tech
ni
ques
ine
x
perim
ent
al stress a
n
a
l
y
s
is.
Optical
Engi
neer
in
g
. 1982; 21(
3): 427
-432.
[7]
Rafae
l
C. Go
nzal
ez. Dig
ital
Image Proc
e
ssing
Usin
g
MAT
L
AB.
Publishi
ng
Hous
e
of Electron
ic
s
Industry
. Beiji
n
g
. 2005.
[8]
Hans
han
L
i
, Z
h
i
y
on
g
L
e
i. Calcu
l
ati
on R
e
searc
h
o
n
In
frared
Rad
i
ati
on
Char
acteri
stic on
F
l
yin
g
Projectil
e
in Sk
y-Scre
en.
IEEE sensors jour
nal.
20
13; 13(
5
)
: 1959-1
9
6
4
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 839
– 846
846
[9]
W
angh
on
gmei.
Researc
h
Progress o
n
Image Match
i
n
g
.
Computer e
n
g
in
eeri
ng a
n
d
Applic
ations
.
200
4; 19(3): 42
-44.
[10]
Z
hang
ho
ngmi
n
. Color Ima
ge
Automatic Mos
a
ic Meth
od B
a
sed o
n
T
e
mpla
te Matchin
g
.
Mi
c
r
o
c
om
pu
te
r
Devel
o
p
m
ent.
200
3; 13(7): 24
-28.
[11] W
ang
w
e
i.
Dig
ital
Image Mos
a
i
c
T
e
chnol
og
y.
Micro Co
mp
ute
r
System
. 200
6
;
26(7): 134
7-1
351.
[12]
Hans
han Li,
Z
h
i
y
on
g
Lei. Stud
y
and an
al
ysis
o
n
a n
e
w
optic
al d
e
t
ection des
ig
n
method for
photo
e
l
e
ctric d
e
tection targ
et.
Sensor Rev
i
e
w
. 2013; 33(4):
315-3
22.
[13] Qichi,Li
uqi
an
g,Sunj
ia
g
u
a
ng. Panor
ama of Stit
ching Se
qu
ence vi
deo Im
age.
Jour
na
l of Comput
e
r
Aide
d Desi
gn
& Computer Gr
aph
ics
. 200
1; 13(7): 605-
60
9.
[14] Caoj
un
jie,
F
e
n
g
jin
gb
o,Suzh
i
x
un.
Pan
o
ram
i
c
Image Mos
a
ic
Algor
ithm.
Jo
urna
l of D
a
li
an
Univ
ersity of
T
e
chno
logy
. 2
003; 43(
1): 180
-182.
[15]
Barne
a
DI, Si
l
vverman
HF
. A class
of a
l
g
o
rithms for f
a
s
t
digit
a
l
imag
e
registrati
on.
I
EEE Trans
on
Co
mp
uters.
19
72; 21(1): 1
79-
186.
[16]
Hans
han
L
i
, Z
e
min
W
ang,
Junch
a
i G
ao,
Z
h
i
y
on
g
L
e
i. Anal
ys
is
a
n
d
calcul
atio
n obj
ect
detecti
o
n
capture rate i
n
multi-sk
y-scree
ns across mea
s
ureme
n
t s
y
ste
m
.
Optik
. 2013; 124(2
0
): 436
9
-
437
3.
[17] R.
Pless.
D
i
scr
ete a
nd
Differe
ntial T
w
o-Vi
ew
Ge
o
m
etry for
Genera
l
I
m
ag
i
ng Syste
m
s
. P
r
ocee
din
g
s
of
T
h
ird W
o
rkshop on Omnid
i
rec
t
iona
l Visio
n
. 2
002; 20(
4): 53-
59.
[18] Horridge
GA.
The evolution of
vi
sual proc
ess
i
ng
an
d the c
o
nstruction
of s
eeing syst
em
s
.
Proce
edi
ng
s
of the Ro
yal S
o
ciet
y of Lo
ndo
n B. 1987; 23
0
(
6): 279-2
92.
[19]
Yong
xin Z
han
g
,
Li Chen, Z
h
ih
ua Z
hao, Jia
n
Jia.
Multi-focus
Image F
u
sion
w
i
t
h
Spars
e
F
eature Base
d
Pulse C
oup
le
d Neur
al Net
w
o
r
k.
TELKOMNIKA.
2014; 12(
2
)
: 357-36
6.
Evaluation Warning : The document was created with Spire.PDF for Python.