TELKOM
NIKA
, Vol. 11, No. 2, Februa
ry 2013, pp. 761~773
ISSN: 2302-4
046
761
Re
cei
v
ed Au
gust 10, 20
12
; Revi
sed
De
cem
ber 2
9
, 2012; Accepte
d
Jan
uary 13,
2013
A Quick Image Registration Algorithm Based on
Delaunay Triangulation
Yongmei Zhang*
1,2
, Li M
a
1,2
, Rui Zhang
1,2
1
School of Infor
m
ation En
gi
ne
erin
g, North Ch
ina U
n
ivers
i
t
y
of
T
e
chnol
og
y,
Beiji
ng, Chi
na,
1001
44
2
Colle
ge of Info
rmation a
nd C
o
mmunic
a
tio
n
Engi
neer
in
g,
North Univ
ersit
y
of Chin
a, T
a
iyu
an, Chi
na,
030
05
1
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: zhang
ym
@n
cut.edu.cn
A
b
st
r
a
ct
T
he traditi
ona
l imag
e match
i
ng a
l
gor
ith
m
s
adopt
more c
o
mpl
e
x strateg
i
es w
hen de
al
i
ng w
i
th
mis
m
atch cau
s
ed by a l
o
t of noise. In th
is pap
er,
a si
mp
le, intu
itive
and effective
noise
proces
si
n
g
alg
o
rith
m
is pr
opos
ed
bas
ed
on D
e
l
a
u
nay tr
ian
gul
atio
n i
n
c
o
mputati
o
n
a
l g
e
o
m
etry.
T
h
e
alg
o
rith
m extra
c
ts
feature
po
ints
usin
g SIF
T
method,
re
sp
ecti
vely
establ
ish
e
s
De
lau
nay
tr
i
ang
ulati
o
n
in
mu
lti-spectr
al an
d
panc
hro
m
atic i
m
a
ges, an
d re
mov
e
s the f
eat
ure po
ints that
three po
ints ar
e coll
ine
a
r an
d
four poi
nts are
on
circle by
Del
a
u
nay tria
ngu
lati
on, obta
i
ns th
e
regi
strati
on i
m
ages thr
o
u
gh t
he corr
espo
nd
ence
betw
een
th
e
Dela
un
ay trian
gul
ations. T
h
e effect of image
regist
rati
on is
eval
uated
by
obj
ective
meth
od. In the i
m
a
g
e
match
i
n
g
, Del
aun
ay trian
g
u
l
atio
n is intro
duce
d
.
T
he establ
ish
m
e
n
t of Delau
nay
triangu
lati
on
i
s
ind
epe
nd
ent of
the selecti
on
of initia
l val
ues
. In
gener
al, a
uni
que
Del
a
u
n
a
y triang
ul
ation
can be
got w
h
e
n
a featur
e
poi
nt set is
giv
en, a
nd
it
can
pr
ovi
de th
e acc
u
rac
y
of the
al
gor
ithm. T
h
e
alg
o
rit
h
m is s
i
mpl
e
a
n
d
clear for c
onve
r
ting a
lot of
mi
smatch
no
ise t
o
the o
per
ation
of Del
a
u
nay tr
ian
gul
atio
n. Experi
m
e
n
t resu
lt
s
show
the alg
o
r
ithm c
an ke
e
p
the go
od rot
a
tion fe
at
ure a
nd transl
a
tio
n
invari
anc
e in
SIF
T
method,
the
nu
mb
er of extr
action f
eature
poi
nts
has
be
e
n
sig
n
ifica
n
tly
reduc
ed
in th
e
alg
o
rith
m c
o
mpare
d
w
i
th SIF
T
meth
od, r
egistr
a
tion
spe
e
d
an
d acc
u
racy
are
better th
an th
e reg
i
stratio
n
a
l
gorit
hm b
a
sed
on c
onv
entio
n
a
l
SIF
T
method.
Ke
y
w
ords
:
Image re
gistratio
n
, Dela
un
ay triang
ulati
on,
Mu
lti-spectral i
m
ag
e, Panchro
m
ati
c
ima
g
e
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 re
gistration is an i
m
porta
nt ste
p
in the app
lication
s
of image fusi
on,
target
recognitio
n
, chang
e det
ecti
on, imag
e
re
gistratio
n
i
s
t
he n
e
cessa
r
y preli
m
ina
r
y
work to
imp
r
ove
the a
c
cura
cy
and validity of
the ab
ove p
r
oblem
s [1
-4]. Image
regi
stration me
an
s
determi
ning t
he
transfo
rmatio
n para
m
eters between im
a
ges a
c
cordin
g
to the similarity measure
so that the two
or m
o
re
ima
g
e
s f
r
om
different sen
s
ors,
angle
s
, di
fferent
time with the
same sce
ne
tra
n
sfo
r
m to
the same
coo
r
dinate
syste
m
and obtain
the best mat
c
h in pixel layer [5,6].
Image re
gist
ration involve
s
a lot of techn
o
logy such as
com
p
le
x feature extractio
n
,
optimizatio
n algorith
m
s,
i
m
age se
gme
n
tation,
patte
rn
recognitio
n
and
matchi
ng, but it la
cks
of
system
atic th
eoreti
c
al gui
dan
ce in ma
ny ways. Th
erefo
r
e, it is the develop
ment dire
ctio
n to
improve the
automation, registra
tion a
c
curacy a
nd speed of the
registration al
gorithm
s [7-9
]. In
respon
se to the above difficultie
s, this paper p
r
e
s
ent
s a regi
stratio
n
algorithm
suitable for m
u
lti-
spe
c
tral a
nd
pan
chromati
c images to im
prove regi
stra
tion accuracy
and sp
eed.
2. Fea
t
ure
Extrac
tion o
f
Multi
-
Spec
tral an
d Pa
nchroma
t
ic
Images Thr
ough Tr
aditi
onal
SIFT Method
Remote
se
n
s
ing im
age registration i
s
the
best m
a
tchin
g
process for two
or mo
re
image
s, overl
appe
d a
r
ea
s
are
geom
etri
c di
stortio
n
o
r
in
con
s
i
s
ten
c
y in
spatial
coo
r
din
a
tes, i
t
is
necessa
ry to find the geom
etric tra
n
sfo
r
ma
tion pa
ram
e
ters fo
r imag
e regi
stratio
n
.
In the com
m
on characte
ri
stic info
rmati
on,
the poi
nt is the mo
st
usu
a
l feature. The
method
dete
c
ts fe
ature p
o
ints
of an
i
m
age, d
e
scri
bes the fe
ature
point
s a
c
cordi
ng to
t
he
neigh
borhoo
d
of the feature poi
nts,
a
nd finally cal
c
ulate
s
the
corre
s
p
ond
en
ce relation
sh
ip
betwe
en the f
eature
point
s
in the o
r
iginal
image
s.
Poin
t feature h
a
s
lowe
r
calculat
ion, and
doe
s
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No. 2, Februa
ry 2013 : 761 – 773
762
not dam
age t
he imp
o
rtant
gray info
rmat
ion of a
n
ima
ge, therefore,
it
can
greatl
y
improve th
e
regi
stratio
n
speed a
nd a
c
cura
cy. Harri
s
detectio
n
met
hod is
a tradit
i
onal featu
r
e
point extra
c
tion
method,
whi
c
h is very
se
n
s
itive to
cha
n
ges in i
m
ag
e
scale,
and
it
largely le
ad
s to limitation
o
f
the appli
c
ati
on sco
pe. Harri
s featu
r
e
points
ca
n
get
accu
rate
regi
stratio
n
results with only
rotation, tra
n
s
lation, a
nd
small-scale tra
n
sform of
an
image, but fo
r the la
rge
-
scale tran
sfo
r
m
of
an ima
ge, th
e meth
od
ca
n not
gua
ra
ntee the
corre
c
t
regi
stratio
n
results. Later,
the re
sea
r
ch
ers
have propo
se
d a larg
e nu
mber of featu
r
e poi
nt
dete
c
tion meth
od
s with
scale i
n
varian
ce, aff
i
ne
invariance, as well as local invariance.
David Lo
we
summ
ari
z
ed t
he existing i
n
variant
feature-ba
se
d tech
nique
s, and f
o
rmally
prop
osed a l
o
cal fe
ature
based o
n
scale spa
c
e, which
ca
n re
m
a
in invari
ant
whe
n
tran
slat
ion,
rotation,
scali
ng, o
r
affine
tran
sform
a
tion
, and
he
prop
ose
d
a
de
scri
ptor
ba
sed
o
n
this featu
r
e
in
2004. He na
med this Scal
e Invariant Fe
ature T
r
an
sfo
r
m, that is SIFT algo
rithm later [10].
SIFT method detect
s
a feature in the sca
l
e
spa
c
e an
d determi
ne
s the scale and l
o
catio
n
of the featu
r
e poi
nts, u
s
e
s
the
main
d
i
rectio
n of th
e nei
ghbo
rh
o
od g
r
adi
ent
as th
e di
re
ction
feature of th
e feature
poi
nt in ord
e
r t
o
enabl
e the
operator in
d
epen
dent of
the scale a
n
d
dire
ction. T
h
e featu
r
e
s
by
SIFT meth
o
d
can
be
u
s
e
d
for reliable
matchin
g
wit
h
same
obj
ect or
scene, th
ey a
r
e inva
ria
n
t to
image
scalin
g an
d
rotation
, and
have
go
od
robu
stne
ss in
llumin
a
tio
n
cha
nge
s, noi
se, and affine
transfo
rmatio
n. In additi
on, SIFT features are the lo
ca
l characte
ri
stics
of an image, they can
corre
c
tly match wit
h
high proba
b
ility.
This p
ape
r resp
ectively e
x
tracts featu
r
e points in
multi-spe
c
tral
and pan
ch
romatic
image
s u
s
in
g
the SIFT me
thod, featur
e
points incl
ud
e po
sition,
scale,
si
ze and
dire
ction.
Fig
u
re
1 gives th
e feature
point
s extracted
by the SI
FT method in m
u
lti-sp
ect
r
al a
n
d
pan
ch
rom
a
tic
image
s.
(a) A multi-sp
ectral im
age
(b) A pan
ch
ro
matic imag
e (c) Featu
r
e p
o
ints u
s
ing th
e SIFT method
Figure 1. Extract Fe
ature Points Using
The SIFT Method
3. Selection fea
t
ure poin
t
s throug
h Delauna
y
triangulation
Feature is
a stru
cture
prop
erty sh
o
w
n by one
or so
me
pixels relative to its
neigh
borhoo
d
,
it generally
better mai
n
tai
n
s inva
rian
ce
in tran
slation
and rotation.
Point feature
is
a basi
c
featu
r
e, but its scope is
the m
o
st extensive
,
and its calc
ulation and d
e
scriptio
n is
very
simple, so feature poi
nts a
r
e sel
e
cte
d
a
s
the re
gistration point
s in the pap
er.
3.1. The pro
p
ert
y
of Dela
una
y
triangulation net
w
o
r
k
In the fiel
d of
geo
sci
en
ce
s,
a la
rge
num
b
e
r
of di
screte
data often
n
e
ed to
be
p
r
o
c
essed
,
the data di
stribute uneve
n
, therefor
e, it need
s to con
s
ide
r
ho
w to rationally an
d
effectively use
the valuable
data. In 1908
, G.Voronoi firstly limited
the effective scop
e of each
discrete poi
n
t
in
mathemati
cs,
that is th
e
scope
of effe
ctively
refle
c
ting regio
nal i
n
formatio
n, a
nd d
e
fined t
he
Voron
o
i di
ag
ram
on t
w
o
-
dimen
s
ion
a
l
plane
(referred to
a
s
V
-g
raph
). In
191
1, A.H.Thie
ssen
applie
d V-gra
ph in
the
ave
r
age
rainfall
of big
re
gion
s. In
193
4, B
.
Delau
nay ev
olved
Dela
un
ay
triangul
ation
from
V
-gra
ph (referred
to as
D
-t
ri
angul
ation). Since
then,
V
-gra
ph an
d
D
-
triangul
ation
have be
come
universally accepted a
nd
widely u
s
ed t
o
analyse discrete d
a
ta.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Quick Im
age Regi
stratio
n
Algorithm
Base
d on Dela
una
y Tria
ngul
ation (Yon
gm
ei Zhang
)
763
Dela
unay tri
a
ngulatio
n i
s
t
he a
s
so
ciate
d
g
r
aph
of
V
-grap
h
[11]
(al
s
o
kn
own a
s
Thiesse
n
diagram, Di
ri
chlet m
ap,
Vigner-Seith
z gra
ph).
V
-g
raph
is
defi
ned a
s
foll
o
w
s. Su
ppo
si
n
g
V
={
v
1
,
v
2
,...,
v
N
},
N
≥
3,
V
is a
set of p
o
ints
in. Euclide
an
plane. And
th
ese
point
s a
r
e not colline
a
r
,
any four poi
n
t
s are n
o
t co
circula
r
.
d
(
v
i
,
v
j
) i
s
th
e Eu
cl
idean
dista
n
ce of
v
i
,
v
j
. Suppos
e
point
x
is
on the pla
ne,
V
(
i
)=
{
x
∈
E
2
︱
d
(
x
,
v
i
)
≤
d(
x,
v
j
),
j
=
1
,2,...,
N
,
j
≠
i
} is called V
o
ronoi Polygon (
V
-Poly
g
o
n
),
all
V
-polygon
s of each poi
nt jointly constitute
V
-graph
.
V
-gra
ph o
n
th
e plan
e can b
e
se
en a
s
e
a
c
h p
o
int in p
o
i
nt set
V
a
s
a
gro
w
th
core, expand
outwa
rd in th
e sam
e
rate,
until they meet eac
h oth
e
r and fo
rm
the grap
hs
o
n
the plane. In
addition to
th
e oute
r
mo
st
point of the
formatio
n
for
open
area
s,
the re
st p
o
int
s
form
convex
polygon
s [12]
.
D
-tri
ang
ulatio
n is
V
-polygo
n
s
with com
m
on ed
ge
s,
and i
s
called
adja
c
ent
V
-p
olygon
s.
Con
n
e
c
ting a
ll adjace
n
t
V
-polygon g
r
o
w
th center formed the trian
g
le netwo
rk is calle
d the
D
-
triangul
ation netwo
rk.
The oute
r
bo
unda
ry of
D
-triangul
ation is
a convex poly
gon,
whi
c
h
co
nsi
s
ts of conv
ex set
c
o
nn
ec
te
d
V
, often referred
to as convex
hull.
D
-tria
n
g
l
e has two very important p
r
ope
rtie
s.
Empty circle
prope
rty. Delaun
ay trian
gulat
ion i
s
u
n
ique (any four poi
nts cannot be
co
cir
c
ula
r),
in
D
-tri
angl
e formed by poi
nt set
V
, its ci
rcumci
rcl
e
of e
a
ch
tri
angl
e contain
s
no
othe
r
points in
V
, it is sh
own in Figure 2.
The large
s
t minimum a
n
g
l
e pro
perty. In the triang
ul
ar net
work fo
rmed by p
o
in
t set
V
,
the minimum
angle in th
e
triangle
of
D
-tri
ang
ulatio
n is the la
rg
est. In this
sense, Dela
un
ay
triangul
ation i
s
"the clo
s
e
s
t to the regula
r
izatio
n"
trian
gular n
e
t. Specifically refe
r to the conv
ex
quad
rilate
ral
diago
nal fo
rmed by
two
adja
c
ent tri
a
n
g
les,
wh
en
e
x
chan
ging, th
e minim
u
m a
ngle
of the six interior a
ngle
s
no
longer in
crea
se
s. It is sho
w
n in Figu
re
3.
Figure 2. Empty Circ
le
Figure 3. The Larges
t
Minimum Angle
These two p
r
opertie
s
dete
r
mine
D
-trian
gulation ha
s great appli
c
at
ion
value. Me
anwhile,
it is al
so
th
e uni
que
an
d be
st two-d
i
mensi
onal
p
l
ane tri
ang
ul
ation. Mile
s
proved
that
D
-
triangul
ation
wa
s "go
od" triangul
ation. S
i
bso
n
foun
d
"
i
n a finite
poi
nt set, the
r
e i
s
o
n
ly a p
a
rti
a
l
and i
s
om
etri
c triang
ula
r
n
e
t
work, which i
s
D
-tri
ang
ulat
ion". Ling
as furthe
r d
e
mon
s
trated
that "i
n
gene
ral,
D
-tri
angul
ation i
s
optimal". Tsai
thoug
ht that
"in Euclid
ean
plane, th
ere
is n
o
mo
re th
an
three adj
acen
t points in a ci
rcle,
D
-
t
r
i
an
gu
la
tio
n
is
un
iqu
e
"
.
These two
prop
ertie
s
ef
fectively ensure
Delau
n
a
y
triangulati
on is the
optimal
triangul
ation
clo
s
e
s
t to eq
uiang
ular
or
equilate
ral, a
nd ma
ke the
control poi
n
t
s distri
bute
as
evenly as po
ssible in e
a
ch small tria
ngle
.
Dela
unay tria
ngulatio
n hel
ps to avoi
d n
a
rrow
or
too
small a
c
ute t
r
iangle in th
e
ca
se of
points evenly
distri
buted. Trian
g
le
s
in t
r
iang
ulation
should
all be
a
c
ute tri
angl
es, or three e
d
g
e
s
roug
hly equal
to each
othe
r, the triangl
e
s
do n
o
t
cross and
overla
p. Delau
nay triang
ulation i
s
clo
s
e
s
t to e
quilateral tria
ngulatio
n. In variou
s t
w
o
-
dime
nsi
onal
triang
ulation
,
only Del
a
u
nay
triangul
ation
can b
o
th mee
t
the global a
nd local opti
m
ization.
3.2. Selectio
n of Fea
t
ure
Points
Select a
p
p
r
opriate
feat
ure i
s
parti
cula
rly impo
rtant in
re
g
i
stration. In
normal
circum
stan
ce
s, the
best
ch
oice
is i
n
the
absen
ce
of
d
e
formatio
n trend, a
nd
regi
stration
requi
res
enou
gh poi
nts, evenly distributed a
s
p
o
ssible in th
e whol
e ima
ge, in orde
r to ensu
r
e the
accuracy of registration. In this pap
er,
we ex
pe
rim
ented a vari
ety of feature point extra
c
tion
method
s, the
n
we respe
c
tively used th
e SIFT to
extract featu
r
e
points fo
r mu
lti-spe
c
tral an
d
pan
chromati
c image
s, wh
ich
can give
the coo
r
di
n
a
tes of feat
ure p
o
ints,
scale
size
a
n
d
D
C
B
A
D
C
B
D
C
B
A
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No. 2, Februa
ry 2013 : 761 – 773
764
orientatio
n fo
r multi-sp
ectral and
pan
ch
romati
c
ima
g
es. In thi
s
al
gorithm,
rem
o
ve dupli
c
ati
on
matche
s
and
multy to one
matchin
g
p
o
i
n
ts, then
e
s
ta
blish
Del
aun
a
y
triangul
atio
n, rem
o
ve thre
e
collinear or four
cocircul
ar feature poi
nts by Delaunay
trian
gulation.
Determin
e the homologous
control p
o
int
pairs by the
Euclide
an di
stance, in
ord
e
r to
redu
ce
the false mat
c
h rate. By u
s
i
n
g
the
hom
olog
ous co
ntrol point
pai
rs, make
affi
ne transfo
rmatio
n
between multi-spe
c
tral
an
d
pan
chromati
c image
s thou
gh the lea
s
t square met
h
o
d
, get tran
sfo
r
mation
para
m
eters, cal
c
u
l
ate
the locatio
n
betwe
en regi
stration im
ag
es. The va
lid
basi
s
of mul
t
i-spe
c
tral an
d pan
ch
roma
tic
image regi
stration usi
ng tri
angul
ation is l
i
sted a
s
follo
ws.
(1) T
he n
e
a
r
est. The n
e
a
r
est thre
e nei
ghbo
r poi
nts
form a tria
ngl
e, and e
a
ch segm
ent
(trian
gle si
de) does n
o
t interse
c
t to ea
ch
other.
(2)
Uniq
uene
ss. Eventuall
y
get consi
s
te
nt result
s no
matter wh
ere
to start.
(3) Optimality
.
If a co
nvex
quad
rilate
ral
diago
nal fo
rm
ed by
any two adj
acent tri
angle
s
i
s
interchan
gea
ble, the small
e
st inne
r angl
es
for two tria
ngle
s
do not
become big
g
e
r.
(4) T
he mo
st rule. If the minimum a
ngle of ea
ch
triangle in t
he trian
gulati
on is in
ascen
d
ing o
r
der, the mini
mum angl
e of the Delau
n
a
y
triangulatio
n is the maximum value.
(5)
Regi
onal.
Increm
ent, deletion, mov
e
ment
one v
e
rtex only affects the n
e
ig
hbori
ng
triangle
s
.
(6) Co
nvex
p
o
lygon sh
ell. T
he
outermo
st bo
und
ary
of the tri
ang
u
l
ar
netwo
rk fo
rms the
shell of a con
v
ex polygon.
Since
the fe
ature
point
s
use
d
in
this
pape
r
i
s
ado
pted a
s
co
ntrol poi
nts, the
feature
points are ge
nerally con
c
e
n
trated
in
th
e feature
lin
es
,
whi
c
h i
s
co
nsistent f
eature of
Trian
gulate
d
Irreg
u
lar Net
w
ork
(TIN). The
bigg
est advantag
e
of TIN is the accurate de
scriptio
n of the
compl
e
x terra
i
n. We
ca
n e
s
tablish
a tria
n
gular net
work for the
wh
ole
cont
rol p
o
int
s
in th
e ima
g
e
,
establi
s
h
den
se tri
ang
ulati
on in l
a
rg
er u
ndulatin
g regi
ons,
and
co
n
s
tru
c
t spa
r
se
triang
ulation
in
the flat region
s.
Comm
on con
s
tru
c
tion tria
n
gulation m
e
th
ods in
clu
de a
ngle jud
g
men
t
method, Thi
e
ssen
Polygon and
Delaun
ay triangul
ation. No matter
where to build
Delaun
ay triangul
ation n
e
t,
Dela
unay tria
ngulatio
n is u
n
ique a
s
lon
g
as
the feature points d
o
n
o
t chan
ge.
In this
pap
er, sele
ct
Dela
unay tria
ngul
ati
on in
the
experim
ents.
In the
pro
c
ess of
establi
s
hm
en
t netwo
rk, e
a
ch
co
ntrol
point pa
rtic
ip
ated in the
netwo
rk stru
cture, th
ere i
s
n
o
cross, cracks and other i
rregularities.
Construc
t procedure of Delaunay
tri
angul
ation is as
follows
.
(1) T
r
av
e
r
s
e
all
s
c
attered
points, find t
he incl
usive
ca
se of
the p
o
int set, get the initial
triangle
s
a
s
the point set convex hull, and pl
a
c
e it into the linke
d list of the trian
g
le.
(2) Ins
e
rt the s
c
attered points
for the
point
set in
orde
r, find o
u
t the trian
g
le wh
ose
circum
circle contai
ns
the inse
rtion
p
o
ints in linked li
st of the triangle
,
delete t
he publi
c
sid
e
s
affect the tria
ngle, co
nne
ct
the inse
rtion
points
with
all vertice
s
in
fluenci
ng the
triangle, thu
s
compl
e
te a p
o
int inse
rtion
in the linke
d list for Del
aun
ay.
(3) A
c
cording
to optimizati
on criteri
on, optimize th
e l
o
cally an
d ne
wly forme
d
tri
angle
s
,
put the forme
d
triangle into
the linked li
st for Delau
nay
triangle.
(4) Pe
rform
step (2
) and (3), until all scattered p
o
int insertion is ove
r
.
In this pape
r, respe
c
tively construct
De
la
unay tri
angul
ation for multi-spe
c
tral and
pan
chromati
c image
s
usi
ng the
del
a
unay fun
c
tio
n
in
Matlab.
Del
aun
ay triang
ulation
for
multispe
ctral and pa
nchro
m
atic imag
es ar
e co
nst
r
u
c
ted as
sho
w
n i
n
Figure 4.
In Figure 4
(c) i
s
to respe
c
tively extract featu
r
e point
s in
multi-sp
ect
r
al and
pan
chromati
c images u
s
i
ng tradition
a
l
SIFT al
gorithm, feature
points in
clu
de dupli
c
ation
matche
s, m
u
l
t
y to one
mat
c
hin
g
p
o
ints,
three
colli
nea
r o
r
fou
r
co
circula
r
featu
r
e
points, th
e tot
a
l
numbe
r is 11
2. The numb
e
r of matchin
g
points after removeme
nt duplicatio
n matche
s, mul
t
y to
one mat
c
hing
points is 9
7
. The numb
e
r of matching
points afte
r removeme
nt three
collin
ea
r or
four
coci
rcula
r
feature poi
n
t
s by
Del
aun
ay triangul
ation is
91. The
intersectio
n
o
f
straight lin
e
s
in
Figure 4 (d
) shows the mismatch by
u
s
ing tradition
al SIFT algorith
m
.
It can be see
n
in Figure 4,
Delaun
ay triangul
at
ion ca
n automatical
ly adjust the grid si
ze
according
to the g
r
ou
nd flu
c
tuation
and
compl
e
xity
in the overl
appi
ng a
r
ea
s. Sel
e
ct fewer
poi
nts
in flat region
s, form larg
er triangl
es.
In ground fluctuatio
n an
d
complex
i
ty, form s
m
all
e
r
triangles thro
ugh further
matching interpolat
io
n, all triangles are acute triangles.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Quick Im
age Regi
stratio
n
Algorithm
Base
d on Dela
una
y Tria
ngul
ation (Yon
gm
ei Zhang
)
765
(d) Mi
smat
ch
usin
g Tra
d
itio
nal SIFT Algorithm
(e)
Red Fe
ature poi
nts wit
h
in the Re
cta
ngle
s
for Fepe
ated
Matchin
g
, Mu
lty to one
(f) Image
s for Removem
e
n
t
Repeate
d
, Many-
to-one M
a
tchi
ng
(g)
Red Fe
ature Point
s
for Removem
e
n
t
three
Collinear or four Coci
rcular Poin
ts using
Dela
unay Tri
angul
ation
Figure 4. Con
s
tru
c
t Dela
un
ay Triang
ulati
on Thr
oug
h Feature Poi
n
ts Extracted by SIFT Method
4. A Ne
w
Re
gistra
tion Al
gorithm Ba
s
e
d on Delau
n
a
y
Triangulation
Matchin
g
me
thod ba
se
d
on SIFT de
scripto
r
h
a
s succe
ssfully a
pplied i
n
ma
ny fields,
su
ch
as targ
et identification,
pa
norami
c
ima
ge
mo
saicin
g. Howe
ver, the SIFT
algo
rithm i
s
still
less appli
ed i
n
remote
se
nsin
g image
regi
stratio
n
. The main
re
aso
n
s a
r
e th
at the traditio
nal
SIFT algo
rith
m ado
pts
Lo
we'
s
the
ne
a
r
est
neig
hbo
r and
next n
e
are
s
t nei
ghb
or di
stan
ce
ratio
method, thre
shol
d is al
ways ch
osen
empiri
cally
, accuracy of
remote
sen
s
i
ng imag
es f
o
r
(a) A multi-
sp
ectral
Image
(b) A Pan
c
hromatic
Image
(c) Extract Fe
ature u
s
ing T
r
adition
al SIFT
Algorithm
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No. 2, Februa
ry 2013 : 761 – 773
766
compl
e
x matching feature
points i
s
lower. The accu
racy of matching points will directly affect the
sub
s
e
que
nt image regi
stra
tion accuracy
.
In image regi
stration o
n
the basi
s
of feature
point
s, feature poi
nt extraction n
eed
s to be
con
s
id
ere
d
fi
rstly. When
extracting
fe
ature
poi
nt
s,
feature p
o
in
ts ne
ed to
b
e
si
gnificant
and
evenly dist
rib
u
ted, and
the
numb
e
r i
s
n
o
t too fewe
r.
Feature extra
c
tion i
s
mo
re
compl
e
x, and
it
need
s to find a sTabl
e, effective and sim
p
le feature ex
traction o
p
e
r
a
t
or.
SIFT features have many f
i
ne f
eatures,
but there
are still so
m
e
shortcomings li
sted as
following:
(1) Be
cau
s
e
feature dete
c
tion need
s to search fo
r
multi-scal
e space, which requi
re
s a lot of
convol
ution
operation, a
nd in
or
der
to pro
d
u
c
e f
eature
poi
nt de
scripto
r
s,
req
u
ire
mul
t
iple
weig
hted hi
st
ogra
m
op
erat
ions. All the
s
e ope
ration
s
inclu
de a l
a
rge nu
mbe
r
o
f
floating poi
nt
operation
s
, th
erefo
r
e, relati
ve to the
nu
mber of
its
fe
ature point
s, the
computati
onal com
p
lex
i
ty
of algorithm i
s
high
er, and
the comp
utation is large
r
, speed i
s
lower.
(2) SIFT feat
ure
s
are
firstly applie
d to t
a
rget
re
co
gnit
i
on, which n
e
eds to d
e
tect
feature
point
s as
many as possible. However, large num
ber of f
eatures will increase feature mat
c
hing time.
(3) SIFT feat
ure set is not
very significant, ther
e are
still some inst
ability points in the set. Many
feature
point
s det
ected
b
y
SIFT algori
t
hm ca
nnot
d
e
scrib
e
conto
u
r featu
r
e
s
, they are neith
er
edge p
o
ints n
o
r co
rn
er poi
nts.
For
remote
sensi
ng ima
g
e
regi
stratio
n
with differe
nt sea
s
o
n
s, resolution
s, and
sen
s
o
r
s,
the feature p
o
int matchi
ng
is more co
mplex
than g
eneral imag
e
s
. Tra
d
itional
SIFT algorit
hm
can
not get ri
d of som
e
e
rro
r mat
c
h p
o
int pairs for remote
sen
s
ing im
age
s.
Mean
while,
the
traditional
SIFT metho
d
d
oes not ta
ke
into a
c
count
repe
ated
mat
c
hin
g
, multy to on
e mat
c
hi
ng
points a
nd so
on, and it is with larg
er
sp
ace to optimi
z
e mat
c
hing
accuracy.
Whe
n
cal
c
ul
ating the dire
ction
s
of SIFT
key point
s, the same
key point may have a
prima
r
y dire
ction, one or
more a
u
xiliary direction.
In the prop
osed algo
rithm,
we sh
all cla
ssify
them into different featu
r
e
points.
Th
ese
feature p
o
int
s
may po
ssib
ly all or in pa
rt gene
rate th
e
corre
c
t match
points, but they are a
c
tu
ally the
same
points, it will gene
rate rep
eated mismat
ch.
SIFT feature
match
with
exhau
st
ive search
may a
l
so
pro
d
u
c
e
one to
many
, many to o
n
e
matchin
g
. Th
ese fal
s
e mat
c
he
s ne
eds t
o
be eliminat
ed one by on
e, otherwi
se i
t
will affect th
e
sub
s
e
que
nt image
regi
stration a
c
curacy. Lei Xiaoqu
n and
etc.
an
alysed th
e m
a
in e
rro
r
sou
r
ce
s
in SIFT feature point m
a
tch
i
ng, po
ssi
ble
error
pai
rs
were
eliminate
d
gradually, a
nd p
r
eci
s
e
pa
irs
were
extract
ed a
s
many
as
po
ssi
ble.
Then
u
s
e th
ese
mat
c
he
d
point
s a
s
re
gistratio
n
con
t
rol
points, affine
transfo
rmatio
n and tiny fa
cet pri
m
it
ive rectifying were tested
on d
i
fferent time
and
resolution
im
age
s. Experi
m
ental
re
sult
s
sho
w
e
d
thi
s
al
go
rithm
could
get hi
g
her a
c
cura
cy
in
regis
t
ration.
While the fea
t
ure point
s are descri
bed
by t
he SIFT
algorith
m
, the descri
p
tor i
s
a 128
-
dimen
s
ion
a
l
vector,
whi
c
h cau
s
ed
ex
ce
ssive
com
putation i
n
t
he
cou
r
se
of de
scripto
r
a
nd
sub
s
e
que
nt regi
stratio
n
. Many schola
r
s h
a
ve pro
posed vario
u
s
improved
SIFT metho
d
s,
prin
cipal
com
pone
nt analy
s
is
(PCA
) ca
n red
u
ce th
e
descri
p
tor di
mensi
onality, thus, it red
u
c
e
s
the computati
onal
com
p
lex
i
ty. Re and
Suktha
nkar
etc. use
d
PCA
method to
re
duce de
scri
ptor
dimension in
the norm
a
lized gradi
ent field, the st
ability was higher,
but t
he preci
s
ion
was lower.
Speede
d Up
Rob
u
st F
eat
ure
s
(S
URF) method
re
d
u
ce
d
the co
mputational complexity,
which
wa
s suitable
for large
r
ch
ange
s in
ima
ge resolution
, but und
er t
he affine t
r
a
n
sformation
and
illumination
chang
es, th
e
effect was u
n
sati
sfac
to
ry. Mikeolaj
czy
k p
r
o
posed
expan
sion
SIFT
descri
p
tor G
r
adient
Lo
cati
on-O
r
ie
ntatio
n Hi
stog
ra
m
(GL
O
H),
it strength
ene
d
d
i
stinctive qua
lity
for the
c
h
arac
teris
t
ic
des
c
riptor, but the effec
t
on the
fast imag
e
matchin
g
wa
s not
satisfa
c
tory.
Wan
g
Ti
anji
a
redu
ce
d
dimen
s
ion
a
lity of hig
h
-di
m
ensi
onal
d
e
scripto
r
fo
r SIFT al
gori
t
hm,
simultan
eou
sl
y stren
g
thene
d the nei
ghb
o
r
hoo
d pixel
in
formation
of the key point
s, and a
c
hieve
d
a fas
t
matc
h
effec
t.
In the traditi
onal alg
o
rith
m for the fe
ature
p
o
ints
extracted f
r
o
m
two ima
g
e
s, firstly
cal
c
ulate
the
nearest
neig
h
bor
matching
of ea
ch
f
eat
ure
point fo
r t
he first ima
g
e
to the
seco
nd
image,
which
is the
key p
o
int de
scripto
r
vecto
r
of mi
nimum Eu
clid
ean di
stan
ce.
The tradition
al
SIFT algo
rith
m u
s
ed
Lo
we
's
nea
re
st n
e
i
ghbo
r
and
su
b-ne
arest
nei
ghbo
r
distan
ce ratio to
mat
c
h.
Whe
n
the
rat
i
o is le
ss tha
n
a threshold
co
rre
sp
ondi
ng to the p
o
i
n
t, it is as th
e co
rrect m
a
tch,
otherwise it woul
d be a
b
ando
ned, ratio usually
uses L
o
we
's
re
comm
end
ed
value of 0.8. For
remote
sen
s
i
ng image
re
gistratio
n
wit
h
different re
solutio
n
s o
r
sen
s
o
r
s, the
matchin
g
fea
t
ure
points a
r
e m
o
re complex than the gen
e
r
al image
s. Although the traditional SIFT algorithm u
s
e
s
a smalle
r
rati
o value
to o
b
t
ain bette
r m
a
tch
re
sult
s
with hi
ghe
r a
c
cura
cy, te
sting
re
sults for a
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Quick Im
age Regi
stratio
n
Algorithm
Base
d on Dela
una
y Tria
ngul
ation (Yon
gm
ei Zhang
)
767
large
number of remote
sensing im
ages show that ev
en with a
sm
a
ller ratio val
ue, it still cannot
get rid of som
e
errors in re
mote sen
s
in
g
images.
This pa
per p
r
ese
n
ts an im
proved regi
strati
on algo
rith
m base
d
on Dela
unay tria
ngulatio
n
to achieve
multi-spe
c
tral
and
p
anchromatic im
ag
e regi
stration
, whi
c
h
is o
n
the
groun
d of
con
s
trai
ned
Dela
unay tria
ngulatio
n feature poi
nt regi
stration. Th
e spe
c
ific
step
s are as follo
ws.
(1)
Re
spe
c
tively extract feature poi
nts i
n
multi-spe
c
tral and pa
nchromatic ima
g
e
s
.
(2)
Even with
a small ratio
value,
it
still cannot
com
p
le
tely get rid
of
the erro
r p
o
in
ts, and
the numb
e
r
of feature p
o
i
nts extra
c
ted
is very li
ttle at this time, whi
c
h is
unf
avorabl
e in t
he
followin
g
ima
ge regi
stratio
n
. In ord
e
r t
o
extr
a
c
t ma
ny matchi
ng
points, thi
s
p
aper u
s
e
s
th
e
prop
osed ratio of 0.8 by Lo
we, ado
pting
the nea
re
st n
e
ighb
or a
nd sub-n
e
a
r
e
s
t neighb
or meth
od
for the initial match.
(3) Ba
sed on
the initial match point
s, the erro
rs a
r
e eli
m
inated on
e by one. Aiming at the
traditional
SIFT metho
d
d
oes not ta
ke
into a
c
count
repe
ated
mat
c
hin
g
, multy to on
e mat
c
hi
ng
points a
nd so
on, the pape
r com
p
a
r
e
s
the pixel
co
ordinate
s
for m
a
tch poi
nts a
nd traverse
s
the
corre
s
p
ondin
g
points to remove dupli
c
ation mat
c
h
e
s an
d multy to one matching poi
nts in
traditional SI
FT method, a
nd en
sures th
e matchin
g
p
o
ints uni
que
and on
e-to
-o
ne.
(4)In thi
s
pap
er
, re
spe
c
tive
ly construct
De
la
unay tria
ngulatio
n of the multi-spe
c
tral and
pan
chromati
c images. Re
mote sen
s
in
g image
s ca
n be divided
into some grid area
s by
Dela
unay tria
ngulatio
n in orde
r to guara
n
tee that
feature poi
nts are examined i
n
each g
r
id.
The
improve
d
re
gi
stration
algo
ri
thm based o
n
SIFT
can
m
a
ke the
extract
i
on feature p
o
ints u
n
iformly
distrib
u
ted, range feature
point
s in each small triangle, and it can ef
fectively improve the
regi
stratio
n
preci
s
ion.
(5)
Revise the feature poi
nts by the Delaun
ay
trian
gulation, re
m
o
ve three col
linear o
r
four co
ci
rcula
r
feature p
o
in
ts obtaine
d b
y
SI
FT descri
p
tor usi
ng the
Delau
nay triangul
ation.
(6) T
o
redu
ce the false m
a
tch rate, Eu
clide
an di
sta
n
ce i
s
introd
uce
d
to dete
r
mine the
homolo
gou
s
control poi
nt pairs
in this
pape
r. Cal
c
ul
ate the Eucli
dean di
stan
ce betwe
en t
w
o
feature
poi
nts de
scripto
r
, t
hat is to fin
d
out
the
ne
are
s
t
an
d su
b-n
eare
s
t neig
h
b
o
r
fe
ature
p
o
i
n
t
descri
p
tors
and
with th
e feature
poi
nts
de
scripto
r
by Euclid
ean di
stan
ce
. Then
cal
c
ulate
the
ratio of E
u
cli
d
ean
dista
n
ce
of
de
scripto
r
s
an
d
,
and
. If the ratio
r
is
less than
the
spe
c
ified th
re
shol
d (Usuall
y
, here
), th
en con
s
ide
r
succe
ssful
match, poi
nt pairs (
,
) are
a pair of mat
c
hi
n
g
point
s
of image seq
uen
ce, othe
rwise, the
matc
h fails
.
(7)
Using th
e
homolo
gou
s cont
rol poi
nt pairs, obtain
the affine tra
n
sformation
b
e
twee
n
multi-spe
c
tral
and
pa
nch
r
omati
c
im
a
ges by a
d
o
p
ting the
le
ast
squ
a
re
method,
obt
ain
transfo
rmatio
n param
eters, cal
c
ulate
the loca
tion
of the regist
ration imag
e
s
, and get the
regi
st
rat
i
o
n
re
sult
s.
SIFT algo
rith
m can d
e
tect
a lot of fe
ature poi
nt
s, the
numbe
r
of m
a
tchin
g
i
s
mo
re, thi
s
i
s
one featu
r
e
of SIFT algo
rithm. At first, SIFT featur
es
was
used
in target re
cog
n
ition, target
recognitio
n
n
eed
s to match smalle
r targets fr
om a l
a
rge n
u
mbe
r
of image databa
se, theref
ore
the small
e
r t
a
rget
s al
so n
eed ri
ch
er fe
ature info
rma
t
ion. This is
also ve
ry important to im
age
regi
stratio
n
with small
e
r
proportio
n
of
o
v
erlap, b
e
cau
s
e i
n
this case we ne
ed to
en
sure that t
i
ny
image overl
a
p still has en
ough featu
r
e
points. Of course, image
r
egist
ration i
s
different from
obje
c
t re
co
gn
ition, the re
gi
stration
for t
w
o im
age
s d
oes
not n
eed
too many m
a
tch p
o
ints,
and
this is the imp
r
oveme
n
t in the pap
er.
Table 1. The Test Results
'
i
q
"
i
q
i
p
i
p
'
i
q
i
p
T
h
e
i
m
a
g
e
p
a
rt
w
i
t
h
re
l
a
t
i
o
n
s
h
i
p
I
D
rI
d2
0
w
a
s
no
t
f
o
und i
n
t
h
e
f
i
l
e
.
T
h
e im
ag
e p
a
r
t
w
i
t
h
r
e
lat
i
o
n
s
h
ip
I
D
r
I
d
2
1
w
a
s
n
o
t
f
o
u
n
d
in
t
h
e f
ile.
T
h
e im
ag
e p
a
r
t
w
i
t
h
r
e
lat
i
o
n
s
h
ip
I
D
r
I
d
2
2
w
a
s
n
o
t
f
o
u
n
d
in
t
h
e f
i
le.
T
h
e im
ag
e p
a
r
t
w
i
t
h
r
e
lat
i
o
n
s
h
ip
I
D
r
I
d
2
3
w
a
s
no
t
f
o
und i
n
t
h
e
f
i
l
e
.
T
h
e im
ag
e p
a
r
t
w
i
t
h
r
e
lat
i
o
n
s
h
ip
I
D
rI
d2
4
w
a
s
no
t
f
o
und i
n
t
h
e
f
i
l
e
.
Performance
Traditional
SIFT
algorithm
The algorithm ba
sed on Delauna
y
triangulation
The numbe
r of fe
ature points
296
251
Correct rate
64%
72%
Registration time(
s
) 1.5058
1.3852
RM
SE
0.9665
0.6254
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No. 2, Februa
ry 2013 : 761 – 773
768
In this paper
,
a lot of multi-sp
ect
r
al and
panch
r
om
atic image
s hav
e been re
spe
c
tively
tested by the
traditional SI
FT
algo
rithm
and the p
r
op
ose
d
algo
rith
m in the pape
r
,
the test re
sults
are sh
own in
T
abl
e 1. Expe
riment re
sults show
the pro
posed algo
rithm has sig
n
ificantly redu
ced
the numb
e
r
of extraction
feature
poi
nts, im
prove
d
the mat
c
h
i
ng spee
d a
nd re
du
ced
the
mismat
ch rat
e
, and improved the regi
stration accu
ra
cy
.
As an examp
l
e, a regist
rat
i
on re
sult for t
he propo
se
d algorith
m
is sh
own in Figure 5.
The ru
nnin
g
time is rel
a
tively slower th
an onl
y usi
n
g
SIFT control
points in im
age regist
rati
on
without
Dela
u
nay trian
gulat
ion. For multi
-
sp
ect
r
al a
nd
pan
chromati
c image
s illu
st
rated i
n
Figu
re
4, the registration run
n
ing
time after Delaun
ay correction is 0.2
0
98
s
, the regi
stration
RMS
E
is
0.6524,
whil
e the
re
gistration time
b
y
dire
ctly u
s
ing SIFT
co
ntrol p
o
ints
is 0.2
1
5
s
, t
he
regi
stratio
n
RMSE
is 0.680
8
(d) Mi
smat
ch
usin
g Tra
d
itio
nal SIFT Algorithm
(e)
Regi
strati
on Image
s using SIFT Algorithm
(f) Re
d point
s for repe
ated,
multy to one
matchin
g
extracted by tradi
tional SIFT
algorith
m
(g) Ima
ges fo
r remove
men
t
repeated m
any-
to-one m
a
tchi
ng
(a) A multi-
sp
ectral
Image
(b) A pan
ch
ro
matic
Image
(c) Extract Fe
ature Point
s
usin
g
Traditio
nal SIFT Algorithm
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
A Quick Im
age Regi
stratio
n
Algorithm
Base
d on Dela
una
y Tria
ngul
ation (Yon
gm
ei Zhang
)
769
(h) Red featur
e
points for rem
o
veme
nt three
colli
ne
ar or fou
r
cocircular fe
a
t
ure poi
nts usin
g
Dela
un
a
y
tri
a
n
gul
ation
(i) Registrati
on
imag
es
Figure 5. Re
gistratio
n
Re
sults
5. Registr
a
ti
on Effe
ct an
d Ev
aluation
There are two regi
stratio
n
accuracy a
nalysi
s
meth
ods, on
e is
statistical an
alysis of
control point
s, Root Mean
Square Erro
r
(RMSE). It
is assume
d that the matching
control p
o
int
s
can
re
pr
ese
n
t
t
he simil
a
r
cha
r
a
c
t
e
ri
st
ic
s,
an
d dete
r
mine the
tra
n
sformation
relation b
e
twe
e
n
image
s. The
other i
s
direct compa
r
ison
betwee
n
the
gray of ima
ges. Hi
stog
ra
m matchin
g
of
regi
stratio
n
image
s is re
q
u
ired. Thi
s
p
aper u
s
e
s
RMSE
as the quantitative
evaluation of image
regi
st
rat
i
o
n
re
sult
s.
The
RMSE
b
e
twee
n refe
rence image
S
(
x
,
y
) an
d re
gistratio
n
ima
g
e
T
1
(
x
,
y
) i
s
denote
d
as
E
1
,
RMSE
is defined a
s
follows.
N
M
y
x
T
y
x
S
E
M
x
N
y
11
2
1
1
)
,
(
)
,
(
whe
r
e
E
1
reflects the differen
c
e b
e
tween
S
(
x
,
y
) a
nd
T
1
(
x
,
y
), the small
e
r
E
1
is, the smaller
differen
c
e i
s
. The algo
rith
m still has
b
e
tter regi
stration effects
when the ima
g
e
exists rotation
and tran
slati
on.
The expe
rim
ent image
s with 0
0
~3
60
0
rotation
can
accurately re
gister. Fig
u
re
6 sho
w
s
the regi
strati
on
re
sults in
the case of
rotation 3
6
0
for the
multi-sp
ectral
ima
ge.
The regi
stration
time is
0.149
8
s
, a
nd
RMSE
is
re
spe
c
ti
vely 0.6524
and 0.7
452
b
e
fore
and
after
rotation. A
fter
rotation, the
registration ti
me is 0.16
96
s
,
RMSE
i
s
separately
0.6
808 and 0.87
33
in Figu
re 6
by
the tradition
al
SIFT algorith
m
, the re
sult
sho
w
s
the p
r
opo
sed
regi
st
ration al
go
rith
m has
strong
er
adapta
b
ility to rotation.
(a) A pan
ch
ro
matic
image
(b) Rotation
3
6
0
for a
multi-spectral
image
(c) Mismat
ch
by using ditio
nal SIFT
algorith
m
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NIKA
Vol. 11, No. 2, Februa
ry 2013 : 761 – 773
770
(d) Regi
strati
on image
s by
using tra
d
itio
nal
SIFT algorith
m
e) Extract fea
t
ure point
s using tradition
al
SIFT algorith
m
(f) Re
d feature points
for re
peated
mat
c
h
i
ng,
multy to one
(g) Ima
ges fo
r remove
men
t
repeated m
any to
one matvhing
(h)
Red featu
r
e point
s for removeme
nt three
collinear or four coci
rc
ular f
eature points
usin
g Del
aun
ay triangulati
o
n
(i) R
egi
stratio
n
image
s
Figure 6. Re
gistratio
n
Re
sults
In Figure 6 (e), the total numbe
r of points is
69 respectively extracted in multi
-
sp
ect
r
al
and p
a
n
c
h
r
o
m
atic ima
g
e
s
usin
g tra
d
itio
nal SIFT alg
o
r
ithm, the fea
t
ure p
o
ints i
n
clud
e du
plication
matche
s, m
u
l
t
y to one
mat
c
hin
g
p
o
ints,
three
colli
nea
r o
r
fou
r
co
circula
r
featu
r
e
points, th
e tot
a
l
matchin
g
nu
mber
of point
s is
60. Th
e
red
poi
nts within
the re
ctan
gles sh
ow re
peated
m
a
tch
i
ng,
multy to one f
eature
point
matchin
g
in F
i
gure
6 (f).
T
h
e red
point
s
within the
re
ctangle
s
in Fig
u
re
6 (h) are removement three collinear or four
cocircul
ar feature point
s using Delaunay
triangul
ation,
the total num
ber of m
a
tchi
ng poi
nts i
s
60 after
rem
o
vement rep
eated mat
c
hi
ng,
Evaluation Warning : The document was created with Spire.PDF for Python.