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
216
~122
6
I
S
SN
: 208
8-8
7
0
8
1
216
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
Image Mos
a
ici
n
g f
o
r W
i
de Angl
e P
a
norama
G. Div
y
a
,
Ch.
Ch
andr
a Sek
h
ar
Department o
f
I
n
formation Tech
nolog
y
,
AITAM, Tekk
ali, A
.
P., I
ndia
Article Info
A
B
STRAC
T
Article histo
r
y:
Received
J
u
n 12, 2015
Rev
i
sed
Ju
l 15
,
20
15
Accepte
d Aug 3, 2015
Im
ages
are inte
gral part in our dail
y liv
es
. W
ith a norm
a
l camera it is
not
possible to get
a
wide angle pano
rama
with high r
e
solution
.
Im
age
Mosaicing
is one of the n
ovel techniqu
es, for co
mbining
two or more images of the
s
a
m
e
s
cene take
n in differen
t
vi
ews
into one im
age. In th
e dark
areas
, th
e
obtain
e
d image is a panoramic image w
ith high
resolution witho
u
t m
a
sk. But
in the c
a
se of lig
hting are
a
s, th
e r
e
sultant
im
age is
genera
ting m
a
sk. In order
to gets
wide angle panoram
a,
in the exis
ting
s
y
s
t
em
, extra
c
ting featur
e
points, find
ing the best stitching
line,
Cluster Analy
s
is (CA) an
d D
y
namic
Programming (
D
P)
methods are used. Al
so used Weighted Average (WA)
m
e
thod for smooth stitch
i
ng results and also
elim
inat
e int
e
nsit
y
seam
effectively
.
In the proposed sy
stem,
to get feature extr
action and featu
r
e
m
a
tching S
I
F
T
(S
caled Invar
i
ant
F
eature Trans
f
o
r
m
)
algorithm
us
ed. In this
process, outliers
can b
e
g
e
nerated
.
R
ANSAC (Random Sample Consensus) is
used for dete
ct
ing the out
lie
r
s
from
the resultant
im
age.
Masking is
significantly
red
u
ced
b
y
using Algebraic R
econstr
uction
Techniqu
es (ART).
Keyword:
ART
Hom
o
g
r
ap
hy
Im
age Mosaicing
Panoram
a
RANS
AC
SIFT
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
:
G. Divy
a,
Depa
rt
m
e
nt
of
In
fo
rm
at
i
on Te
chn
o
l
o
gy
,
AIT
A
M
,
Tek
k
a
li,
A.P
., I
ndia.
Em
a
il: d
i
v
y
acse.aita
m
@
g
m
a
i
l
.
co
m
1.
INTRODUCTION
Im
ag
e Mo
saicin
g is a tech
n
i
q
u
e
wh
ich en
ab
les
u
s
to jo
in to
g
e
t
h
er m
a
n
y
sm
a
ll i
m
ag
es in
to a
o
n
e
larg
e im
ag
e, fro
m
wh
ich
m
o
re in
fo
rm
atio
n
can
b
e
g
a
th
e
r
e
d
easily.
T
h
is technique has
been
us
ed sinc
e
the
devel
opm
ent
o
f
p
h
o
t
o
gra
phy
t
o
i
n
crea
se t
h
e
fi
el
d o
f
vi
ew
b
y
past
i
ng t
oget
h
er m
a
ny
sm
all
im
ages t
a
ken
fr
om
the cam
era. For
gene
rating a com
p
rehensible m
o
saic, th
ese im
ages have to be
first
aligned t
o
a certain
referen
ce im
ag
e an
d
th
en
p
a
sted
to
g
e
th
er.
Based
on
th
e p
a
th
in
wh
ich
th
e ca
m
e
ra
m
o
v
e
s wh
ile tak
i
n
g
th
e
im
ages, al
i
gni
ng t
h
ese i
m
ages t
oget
h
e
r
m
a
y pro
d
u
ce di
st
o
r
t
i
ons suc
h
as el
on
gat
i
o
n
,
co
nt
ract
i
o
n
,
an
d vi
ewi
n
g
at an a
n
gle, in them
. In t
h
is
pro
ces
s, c
o
mbines
seve
ral i
m
ages with
o
v
erl
a
ppi
ng
fi
el
d
of
vi
e
w
(F
OV
) t
o
pr
o
duce
a
pa
no
ram
i
c im
age or
a
hi
g
h
-
r
esol
ut
i
o
n
i
m
age [1]
,
as show
n in Figu
r
e
1.
I
m
ag
e
mo
saic techno
log
y
i
s
wid
e
ly used in mili
tary an
d civ
ilian
field
s
, su
ch as
satellite rem
o
te sen
s
ing
,
UAV surv
eillan
ce and
search
ing
,
robo
t v
i
sion
, med
i
cal prob
e, electron
i
c im
ag
e
stab
ilizatio
n
an
d v
i
rtu
a
l reality, etc.
Fig
u
re
1
.
Sequen
tial Im
ag
es to
b
e
M
o
saiced
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
:
121
6
–
12
26
1
217
There are 2 types of Im
age
Mosaicin
g, suc
h
as Direct, Fe
ature base
d
[2]. Direct approa
ch com
p
ares
all pixel intensities of i
m
ages with
each other. It can provi
d
e ve
ry accura
te registration
and it has a li
mited
range
of c
o
nverge
n
ce. T
h
es
e approac
h
es
include
Four
ier
an
a
l
ys
is
te
c
h
n
i
q
u
e
s
an
d als
o
co
a
r
s
e
to f
i
ne
opt
i
m
i
zati
on of
cost
or
o
b
j
ect
i
v
e fu
nct
i
o
ns.
Feat
ure
-
ba
sed
app
r
oaches
ha
ve t
h
e ad
va
nt
a
g
e o
f
bei
ng m
o
re r
o
b
u
st
agai
n
s
t
scene m
ovem
e
nt
and are
p
o
t
en
tially fast
er, if im
p
l
e
m
e
n
ted
th
e ri
g
h
t
way. Th
e
b
i
ggest ad
v
a
n
t
ag
e,
h
o
wev
e
r, is th
e ab
ility
to
“reco
gn
ize
pan
o
r
am
as,” i
.
e., t
o
a
u
t
o
m
a
t
i
cal
l
y
di
scove
r
t
h
e a
d
jace
ncy
(o
verl
a
p
)
rel
a
t
i
ons
hi
ps am
ong
an
u
n
o
r
d
ere
d
set
o
f
i
m
ag
es, wh
ich
mak
e
s th
em
id
eally su
ited
fo
r fu
lly au
to
m
a
te
d
stitch
i
n
g
o
f
pan
o
ram
a
s tak
e
n
b
y
casu
a
l
u
s
ers. It
is su
itab
l
e fo
r
fu
lly au
to
m
a
ti
c Mo
saicin
g
.
Featu
r
e d
e
tection
and
m
a
tch
i
n
g
are u
s
ed
for i
m
ag
e stitch
i
n
g
fro
m
th
e in
pu
t im
ag
es wh
ich
are
hav
i
ng
th
e feat
ures in both t
h
e
im
ages whic
h
are ha
vi
n
g
co
r
r
esp
o
ndi
ng
p
o
i
n
t
s
t
o
fi
n
d
, an
d ha
ve
t
o
m
a
t
c
h t
hose
corre
sp
o
ndi
ng
poi
nt
s t
o
get
a si
ngl
e pa
no
ra
m
i
c im
age. Th
e gr
ou
ps o
f
fe
at
ures
co
nsid
er po
in
t
features.
Whi
c
h
in
cl
u
d
e
s meth
od
s
u
s
ing
li
n
e
in
tersectio
ns, h
i
g
h
v
a
rian
ce po
in
ts, m
a
x
i
mall
y
d
i
stin
ct po
in
ts
with
resp
ect t
o
a sp
eci
fied m
e
asure
of sim
ila
rity, and corne
r
s.
With
reg
a
rd
to
feat
u
r
e d
e
tectio
n
,
in
m
o
st in
stances th
e co
re alg
o
rith
m
s
fo
llow th
e d
e
fin
ition
o
f
a po
in
t as a lin
e in
tersectio
n
o
r
as th
e cen
t
ro
id
of a cl
ose
d
-b
o
u
n
d
a
r
y
regi
on
. It
has bee
n
fo
u
nd t
h
at
cor
n
e
r
s
fo
rm
their o
w
n class o
f
feature as the prope
rty of
b
e
ing
a corn
er
is h
a
rd
t
o
d
e
fi
ne m
a
th
em
a
tical
ly.
There
are
two
main appro
aches to
finding feature
points a
n
d th
eir corres
p
ondence
s. T
h
e first is
t
o
find features in one im
age tha
t
can
be accura
tely
tracked
us
ing a local sear
ch techni
que s
u
ch
as correlation
or
l
east
square
s.
The sec
o
n
d
i
s
t
o
i
nde
pe
nde
nt
l
y
det
ect
feat
ures i
n
al
l
t
h
e i
m
ages un
de
r consi
d
erat
i
o
n a
nd t
h
en
match features
base
d on thei
r local appea
r
a
n
ce. T
h
e
fo
rmer approac
h
is
to com
b
in
e tho
s
e im
ag
es to
g
e
t a
si
ngl
e
pa
no
ra
m
i
c im
age. T
o
get
a
hi
g
h
resol
u
t
i
o
n
pa
n
o
ram
a
, t
h
e Im
age M
o
sai
c
i
n
g
t
echni
que
us
es t
h
e
fo
llowing
t
w
o alg
o
rith
m
s
. Fo
r th
e
f
eature
extraction a
n
d feature m
a
tchi
ng SI
FT (Sca
l
e
d
I
n
vari
a
n
t
Feat
ur
e
Transfo
r
m
)
alg
o
r
ith
m
is u
s
ed
. Fo
r d
e
tecting
th
e ou
tliers
from th
e resu
ltan
t
i
m
ag
e RANSAC (Rand
o
m
Sam
p
l
e
Co
n
s
en
su
s) algo
rith
m
is u
s
ed
.
2.
LITERATU
R
E
SU
RVE
Y
Rafel C.
Gon
z
alez et al [3
], in
th
ei
r
work imag
e acqu
i
sitio
n pro
cedu
r
e an
d im
ag
e restoratio
n fo
r th
e
feature m
a
tching
points to
get a panoram
i
c im
age. Li
Ji
n et
al
[4]
,
p
r
op
ose
d
in thei
r work how feature
match
i
n
g
is u
s
ed
to
ex
tract featu
r
e po
in
ts effectiv
ely,
b
y
usin
g
SI
FT algor
ith
m
.
Th
is
m
e
th
od
also
pro
p
o
s
es a
reliab
l
e p
a
rameter estim
a
tio
n
m
e
th
o
d
, an
d t
h
e
resu
lt is
reliab
l
e to
stitch
i
ng
a larg
e im
ag
e
.
Patil et a
l
[5
], p
r
op
o
s
ed
in
their work
th
e Scale
In
v
a
rian
t Featu
r
e Tran
sfor
m
(SIFT) algorith
m
can
b
e
appl
i
e
d t
o
p
e
r
f
o
r
m
t
h
e det
ect
i
on an
d m
a
tchi
n
g
co
nt
r
o
l
poi
nt
s, f
o
r t
h
e im
age
m
o
sai
c
. SIFT al
go
ri
t
h
m
p
r
ov
id
ing
m
o
re reliab
l
e featu
r
e m
a
tch
i
n
g
for th
e pu
rpose of
o
b
j
ect
reco
gn
itio
n wit
h
in
a sing
le
v
i
ew.
R
ANS
AC
i
s
a
re-s
am
pl
i
ng t
echni
que
t
h
at
gene
rat
e
s ca
n
d
i
dat
e
sol
u
t
i
o
n
s
by
usi
n
g t
h
e
m
i
nim
u
m
nu
m
b
er
obs
er
vat
i
o
n
s
(
d
at
a p
o
i
n
t
s
)
r
e
qui
red
t
o
est
i
m
a
t
e
t
h
e un
d
e
rl
y
i
ng m
odel
pa
ram
e
t
e
rs. unl
i
k
e
co
n
v
e
n
t
i
onal
sam
p
lin
g
tech
niq
u
e
s th
at u
s
e as
m
u
ch
o
f
the d
a
ta as p
o
ssi
ble to
o
b
t
ain
an
in
itial so
lu
tio
n
an
d
th
en
pro
c
eed
to
p
r
un
e ou
tliers
RANSAC
u
s
es th
e sm
al
lest s
e
t p
o
ssi
b
l
e and p
r
o
ceed
s
to
en
larg
e th
is set
with
con
s
isten
t
d
a
ta
p
o
i
n
t
s. Mallick
[6
] propo
sed
featu
r
e
b
a
sed
tech
n
i
q
u
e
i
n
th
ei
r wo
rk
for th
e
Im
ag
e Mo
saicin
g. Firstly select th
e
corres
ponding corners
in
th
e two
im
ag
es. R
A
NSAC is u
s
ed
to
esti
m
a
te
th
e h
o
m
o
g
r
aph
y
relatin
g
th
e two
im
ages. The e
s
t
i
m
a
t
e
d hom
o
g
ra
p
h
y
i
s
refi
n
e
d usi
ng
Newt
on'
s n
o
n
-
l
i
n
ear
m
e
t
hod.
A dy
nam
i
c prog
ra
m
m
i
ng
base
d
bl
en
di
n
g
al
go
ri
t
h
m
was use
d
t
o
seam
l
e
ssl
y
bl
en
d t
h
e
t
w
o
i
m
ages.
M
a
l
a
vi
ka B
h
as
kara
na
nd a
n
d San
d
eep B
h
at
[7]
p
r
o
p
o
se
d i
n
t
h
ei
r
wo
r
k
, r
e
gi
st
rat
i
on a
n
d
m
o
sai
c
ki
n
g
of im
ages. Fea
t
ures in im
ages are detected
using th
e Scal
e- Inva
riant Fe
ature Trans
f
orm (SIFT
)
. A
nearest
n
e
igh
bor algo
rith
m
with
Euclid
ean
distance
m
easure is
use
d
for esta
b
lis
h
i
ng
co
rr
es
p
ond
en
c
e
s
b
e
tw
e
e
n
im
ages. T
h
e
norm
alized Dire
ct Lin
ear
T
r
an
sfo
r
m
a
t
i
on (
D
LT) t
oget
h
er
wi
t
h
t
h
e R
a
nd
om
Sam
p
l
e
C
o
nse
n
sus
(RANSAC) alg
o
rith
m
is
u
s
ed
to
esti
m
a
te
t
h
e ho
m
o
g
r
a
p
hy between the
im
ages. The
images are then warpe
d
t
o
a com
m
on coo
r
di
nat
e
sy
st
em
usi
ng t
h
e e
s
t
i
m
a
t
e
d hom
ogra
p
hy
. Al
p
h
a
bl
en
di
n
g
base
d o
n
di
st
a
n
ce
of t
h
e
p
i
x
e
l
fro
m
th
e i
m
ag
e b
o
rd
er i
s
u
s
ed
t
o
stitch th
e im
ag
es in
to
a
sm
o
o
t
h
m
o
saic.
Gui
d
o
B
a
rt
ol
i
[8]
pr
op
ose
d
t
h
e
pr
ocess
of
i
m
age re
gi
st
rat
i
o
n
as a
n
a
u
t
o
m
a
t
i
c
or m
a
nual
pr
oce
d
u
r
e
whi
c
h t
r
i
e
s c
o
r
r
esp
o
ndi
ng
p
o
i
n
t
s
bet
w
een t
w
o i
m
ages an
d
sp
atially alig
n
th
em
to
m
i
n
i
mize a d
e
sired error,
i.e. a c
o
nsistent
distance m
easure
betwee
n two im
ages
. Features
should be
dis
tinct
ive objects which
are
pos
sibly uniform
ly spread over the im
age
s
and easily
det
ect
abl
e
by
usi
ng S
I
FT
algorithm
.
The detected
feature
s
in the
refe
rence a
n
d sense
d
im
ages
can be m
a
tch
e
d
b
y
m
ean
s o
f
th
e i
m
ag
e in
ten
s
ity v
a
lu
es in th
ei
r
cl
ose n
e
i
g
hb
or
ho
o
d
s, t
h
e
feat
ure
spat
i
a
l
di
st
ri
b
u
t
i
o
n
,
or
t
h
e feature sym
b
olic d
e
scrip
tion
.
For
re-sam
p
lin
g th
e
m
a
ppi
n
g
f
u
nct
i
ons c
o
nst
r
uct
e
d d
u
ri
n
g
t
h
e
p
r
evi
ous st
e
p
ar
e use
d
t
o
t
r
a
n
s
f
o
r
m
t
h
e sensed i
m
age and t
hus
t
o
regi
st
er t
h
e i
m
ages. T
h
e t
r
a
n
s
f
o
r
m
a
ti
on ca
n
be real
i
zed i
n
a
fo
rwa
r
d o
r
ba
ckwa
r
d
m
a
nner.
Davi
d G
.
Lo
we [
9
]
pr
o
pose
d
i
n
t
h
ei
r pa
pe
rf
or
i
m
age m
a
t
c
hi
ng a
n
d
rec
o
gni
t
i
on,
S
I
FT
feat
ures
are
fi
r
s
t
e
x
t
r
act
ed
f
r
om
a set
o
f
refe
rence im
ages and stored i
n
a
database.
A ne
w im
ag
e is m
a
tched by i
ndi
vidua
lly com
p
aring each feature
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
I
m
a
g
e
Mo
sa
ic
in
g fo
r W
i
d
e
A
n
g
l
e
Pa
no
ra
ma
(G
.
D
i
vya
)
1
218
fr
om
t
h
e new im
age t
o
t
h
i
s
pre
v
i
o
us
dat
a
b
a
se and
fi
n
d
i
n
g can
di
dat
e
m
a
t
c
hi
ng
feat
u
r
e
s
base
d o
n
E
u
cl
i
d
ean
distance
of thei
r feat
ure
vect
ors.
3.
METHO
D
OL
OGY
For
Im
age Mosai
c
i
n
g t
h
e f
o
l
l
o
wi
ng m
e
t
hods a
r
e use
d
, suc
h
as SIF
T
(Scal
ed
In
vari
ant
Feat
u
r
e
Transform
)
algorithm
and RANSAC (Ra
n
dom
Sa
m
p
le Co
n
s
en
su
s) algo
r
i
t
h
m
.
3.
1. SIFT
Al
g
o
ri
thm
a)
Sc
al
e-sp
ace
E
x
tre
m
e
Dete
cti
o
n
SIFT al
go
ri
t
h
m
uses Di
ffe
r
e
nce o
f
Ga
uss
i
an (D
o
G
)
wh
i
c
h i
s
an ap
pr
oxi
m
a
t
i
on of
Lapl
aci
an
of
Gaus
si
an
(LoG). Diffe
re
nce of Ga
ussi
an
is obtai
ned
as
the diff
ere
n
ce of Ga
ussi
an blurring
of an
im
ag
e with
two
di
ffe
re
nt
, l
e
t
i
t
be
and
. T
h
e L
a
pl
aci
an
of
Ga
ussi
an
i
s
(
1
)
Thi
s
pr
ocess
i
s
d
one
f
o
r
di
ffe
r
e
nt
oct
a
ves
of
t
h
e i
m
age i
n
Ga
ussi
an
Py
ram
i
d as s
h
o
w
n i
n
Fi
gu
re
2.
Fi
gu
re 2.
Gau
ssi
an
Py
r
a
m
i
d
The Ga
ussi
a
n
bl
u
rri
n
g
can
b
e
cal
cul
a
t
e
d by
usi
n
g eq
uat
i
o
n (2
).
Di
ffe
re
n
ce of Ga
ussi
a
n
can be cal
cul
a
t
e
d by
usi
n
g
e
q
uat
i
o
n
(3
).
(2
)
(3
)
Once t
h
is DoG is found, im
ag
es are sea
r
che
d
for local
e
x
tre
m
a over
scale and
space
. For eg,
one pi
xel in a
n
im
age i
s
com
p
ared
wi
t
h
i
t
s
8
nei
g
hb
o
u
rs
as
wel
l
as
9 pi
xel
s
i
n
next
scal
e
and
9
pi
xel
s
i
n
pre
v
i
ous
scal
e
s
. I
f
i
t
is a lo
cal ex
trema, it is a po
ten
tial k
e
y
p
o
i
n
t
. It
b
a
sically
means t
h
at key
point is
best
re
presente
d i
n
tha
t
scale
as show
n in
Fig
u
r
e
3
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
JECE Vo
l. 5
,
N
o
. 5
,
O
c
tob
e
r
20
15
:
121
6
–
12
26
1
219
Fi
gu
re
3.
Scal
i
n
g
o
f
Im
age pi
xel
s
R
e
gar
d
i
n
g di
ff
erent
param
e
t
e
rs of
em
pi
ri
cal
dat
a
w
h
i
c
h
can be
s
u
m
m
ari
zed
as
, num
ber o
f
oct
a
ves
=
4
,
n
u
m
b
e
r
o
f
scal
e lev
e
ls =
5
,
initial,
etc as op
timal v
a
lu
es
b)
Key
-
po
int Lo
ca
liza
t
io
n
Once
pote
n
tial key points locations are f
ound, they have
to be refi
ned t
o
get m
o
re accurate res
u
lts. Taylor
series expansion as s
o
wn in the equations
(4),(5) a
n
d (6) of scale s
p
ac
e to get m
o
re accurate locati
on
of
ex
trem
a, an
d
if th
e in
ten
s
ity at th
is ex
trem
a
i
s
less th
an
a thresho
l
d
v
a
lu
e
0
.
0
3
, it is rej
e
cted
. So
it eli
m
i
n
ates
any
l
o
w-c
o
nt
ra
st
key
poi
nt
s a
n
d
ed
ge
key
po
i
n
t
s
an
d
rem
a
ins st
r
o
ng
i
n
t
e
r
e
st
p
o
i
n
t
s
.
(
4
)
(5
)
(6
)
c)
O
r
ient
at
ion A
s
s
i
g
n
ment
Now an
o
r
ien
t
atio
n
is assi
g
n
ed
to
each
k
e
y
po
in
t to
ach
i
ev
e inv
a
rian
ce t
o
im
ag
e ro
tation
.
A
n
e
igh
borho
od
is
t
a
ken ar
o
u
n
d
t
h
e key
p
o
i
n
t
l
o
cat
i
on
de
pen
d
i
n
g o
n
t
h
e sc
al
e, and t
h
e g
r
adi
e
nt
m
a
gni
t
ude can
be cal
cul
a
t
e
d
usi
n
g eq
uat
i
o
n
(7) an
d di
rect
i
on i
s
cal
cul
a
t
e
d i
n
t
h
at
regi
o
n
by
usi
n
g eq
u
a
t
i
on (
8
).
An o
r
i
e
nt
at
i
on
hi
st
o
g
ram
wi
t
h
36
bi
ns c
o
veri
ng
3
6
0
deg
r
ees i
s
c
r
eat
ed
.
m
x,
y
√
Lx
1
,
y)-
L(
x,y-1
)
)
2
+ (L(
x
,y
+
1
)
–
L
(x
-
1
,y
)
)
2
(7
)
,
t
a
n
,
1
,
1
/
1
,
1
,
(8
)
d) Key-p
o
int Descriptor
N
o
w
k
e
y po
in
t
d
e
scr
i
p
t
or
is cr
eated
.
A
1
6x16
n
e
i
g
hbo
rho
o
d
arou
nd
th
e
key p
o
i
n
t
is taken
.
I
t
is d
i
v
i
sed
in
to
16 s
u
b-bloc
ks of 4x4 sizes. For each s
u
b-block, 8 bi
n orie
ntati
on hist
ogra
m
is
created. So a total of 128
bin
val
u
es
are a
v
ai
l
a
bl
e. It
i
s
re
pr
esent
e
d
as a
ve
ct
or t
o
f
o
rm
key
poi
nt
de
scri
p
t
or.
e) Ke
y-
poin
t
Ma
tchin
g
Key
poi
nt
s bet
w
een t
w
o i
m
ages are m
a
t
c
hed by
i
d
e
n
t
i
f
y
i
n
g
t
h
ei
r
nearest
nei
g
hb
or
s. B
u
t
i
n
som
e
cases, t
h
e
seco
nd cl
osest
-
m
a
t
c
h
m
a
y
be very
near t
o
t
h
e fi
rst
.
It
m
a
y
hap
p
e
n
due t
o
noi
se
or
som
e
ot
he
r reas
o
n
s.
In t
h
at
case, ratio of c
l
osest-distance
to sec
o
nd
-clo
sest d
i
stan
ce is
tak
e
n
.
If it is
gr
eater tha
n
0.8, they are
re
ject
ed.
It
eliminates around 90%
of
false m
a
tches wh
ile discards
only 5%
correct m
a
tches.
3
.
2
.
RA
NSAC
A
l
go
rithm [
1
0
]
Steps
:
a)
Sel
ect
ra
nd
o
m
l
y
t
h
e
m
i
nim
u
m
num
ber o
f
poi
nt
s re
q
u
i
r
e
d
t
o
det
e
rm
i
n
e t
h
e m
odel
pa
ra
m
e
t
e
rs.
b)
So
lve for the p
a
ram
e
ters o
f
th
e m
o
d
e
l.
c)
Det
e
rm
i
n
e h
o
w
m
a
ny
poi
nt
s fr
om
t
h
e set
of
all po
in
ts
fit with
a
pre
d
efi
n
ed tolera
nce
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
I
m
a
g
e
Mo
sa
ic
in
g fo
r W
i
d
e
A
n
g
l
e
Pa
no
ra
ma
(G
.
D
i
vya
)
1
220
d)
If the fraction
of the
num
b
er of
inliers ove
r the tot
a
l num
ber points
in the set exceeds a pre
d
efi
n
e
d
th
resh
o
l
d
,
re-esti
m
ate th
e
m
o
d
e
l p
a
ram
e
ters u
s
ing
all th
e iden
tified
i
n
liers
an
d term
in
ate.
e)
Ot
her
w
i
s
e,
r
e
peat
st
eps
a)
t
h
r
o
ug
h
d)
m
a
xim
u
m
of
N t
i
m
es
.
Th
e
n
u
m
b
e
r of iteratio
n
s
,
N, i
s
cho
s
en
h
i
gh
en
oug
h
t
o
ensu
re t
h
at th
e pro
b
a
b
ility p
(usu
ally set to
0
.
99
) t
h
at
at least o
n
e
of
th
e sets
o
f
ran
d
o
m
sa
m
p
les d
o
es no
t in
clud
e
an
o
u
tlier. Let
u
represen
t t
h
e p
r
ob
ab
ility that an
y
sel
ect
ed dat
a
poi
nt
i
s
a
n
i
n
l
i
e
r an
d
v =
1
−
u th
e
p
r
ob
ab
ility o
f
o
b
serv
ing
an
ou
tlier.
N iteration
s
of th
e
m
i
nim
u
m
num
ber
o
f
poi
nt
s
d
e
not
e
d
m
are r
e
qui
red
,
whe
r
e
(
9
)
and
t
h
us
wi
t
h
s
o
m
e
m
a
ni
pul
at
i
on,
(
1
0
)
a.
Abs
t
rac
t
Rec
o
nstruc
tion
Te
chniques (ART)
[11]
In
AR
T
t
h
e i
m
age s
o
l
u
t
i
o
n
i
s
app
r
oxi
m
a
t
e
d by
a
wei
g
ht
ed
sum
of
basi
c
f
unct
i
o
ns
, eac
h
of
w
h
i
c
h
i
s
a
sh
ifted b
a
sic basis fu
n
c
tion
is
,
)
(
1
1
)
The m
o
st
com
m
on basi
c fu
nc
t
i
on i
s
a pi
xel
t
h
at
has a s
q
uar
e
sup
p
o
rt
a
nd i
s
val
u
e
d
one i
n
si
de t
h
e s
u
p
p
o
r
t
and
zero
ot
her
w
i
s
e
.
T
h
e
wei
g
ht
s
c
ij
o
f
t
h
e su
mmatio
n
are the co
efficien
ts
o
f
th
e
resu
lting
ap
pro
x
i
m
a
ted
imag
e
vector.
In
real
a
p
pl
i
cat
i
ons
, m
easure i
ndi
rect
l
y
a fi
ni
t
e
num
ber
o
f
l
i
n
e i
n
t
e
gral
s
.
L
e
t
w
l
be a m
easure
d
i
n
tegrals
along
a line
l
, t
h
en in the a
b
se
nce
of noise for a
n
i
m
age in the form
of
equ
a
tio
n (1
1)
t
h
at
(
1
2
)
Whe
r
e r
ij
(
l
) is th
e in
teg
r
al o
f
b(x
-
x
i ,
y-
y
j
)
al
ong
l
, w
h
i
c
h can be cal
c
u
l
a
t
e
d fr
om
t
h
e geom
et
ry
of dat
a
co
llectio
n
.
When
n
o
i
se is presen
t, as it is the case in
real ap
p
lication
s
, t
h
e equ
a
lity in
eq
u
a
tion
(1
2)
beco
m
e
s
an approxim
at
i
o
n. In
ART, a
syste
m
of equalities, each of whic
h in t
h
e fo
rm
of e
quation (12) is
formed
by
con
s
i
d
eri
ng al
l
t
h
e l
i
n
es al
on
g w
h
i
c
h
dat
a
h
a
ve bee
n
co
llected
. An
ART-typ
e
alg
o
rith
m is essen
tially
b
a
sed
o
n
th
e
fo
llo
wi
ng
relax
a
tion
meth
od
for so
lv
i
n
g a co
nsisten
t
syste
m
o
f
lin
ear eq
u
a
lities. Let th
e system
b
e
(
1
3
)
Wi
t
h
S
unk
now
n
s
and
K
equatio
n
s
.
For 1
≤
k
≤
K
, let
R
k
be t
h
e t
r
ans
pose
of t
h
e
k
th
row
of
R
a
nd
w
k
be t
h
e
k
th
com
pone
nt
of
w
; th
en
c
(0)
is th
e
S
-
di
m
e
nsi
onal
co
l
u
m
n
vect
or
o
f
zeros
(1
4)
(
1
5
)
Wi
t
h
(
1
6
)
Whe
r
e
1
a
n
d
is a relax
a
tion p
a
ram
e
ter th
at co
n
t
ro
ls
ho
w well th
e cu
rren
t
equation is sat
i
sfied.
The
se
quence
c
(
m
)
c
o
n
v
er
ges t
o
t
h
e
uni
que
m
i
nim
u
m
Eucl
i
d
ean
no
rm
sol
u
t
i
o
n
of
t
h
e
syste
m
in
eq
uatio
n
(13
)
. In th
e in
con
s
isten
t
case,
t
h
ere is no
co
nv
er
g
e
n
ce
p
r
oo
f.
H
o
w
e
v
e
r, A
R
T-
type
al
go
ri
t
h
m
s
are wi
des
p
rea
d
,
d
u
e
t
o
i
t
s
rel
a
t
i
v
el
y
superi
or
per
f
o
r
m
a
nce for
v
a
ri
o
u
s rec
o
nst
r
uct
i
on t
a
s
k
s [
1
2]
. I
n
practice
λ
is k
e
p
t
lo
w
(abou
t 0
.
1
-
0
.
01
) to
red
u
ce
no
ise fittin
g
and
/
o
r
t
o
im
p
r
o
v
e
conv
erg
e
n
ce sp
eed
.
Ano
t
h
e
r
im
portant
para
meter is the order in
which t
h
e data is
acce
ssed.
It has
be
en obse
rve
d
that faster conve
rge
n
ce
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
JECE Vo
l. 5
,
N
o
. 5
,
O
c
tob
e
r
20
15
:
121
6
–
12
26
1
221
can
be ac
hieve
d
if
a linear order (i.e
.,
in
cr
easin
g
or
d
ecr
easin
g)
is g
i
v
e
n
up
in
f
a
vo
r
of
a n
on-
sequ
en
tial
o
r
d
e
r
(e.
g
.,
di
rect
i
o
n
s
t
h
at
are
as
ort
h
o
g
onal
a
s
pos
si
bl
e t
o
t
h
e
pre
v
i
o
us
o
n
es)
.
Ano
t
h
e
r
reason
for its
po
pu
l
a
rity is th
e
possib
ility o
f
i
n
co
rpo
r
ating
p
r
i
o
r kno
wled
g
e
in
th
e recon
s
tru
c
tio
n
process
.
For e
x
am
ple, if the im
age is known t
o
be
non-negative
,
the
n
after each iteration one ca
n set the
n
e
g
a
tiv
e
v
a
lues to
zero
prior to
th
e n
e
x
t
iterativ
e step
. Su
ch
adju
stm
e
n
t
h
a
s b
een
shown to
i
m
p
r
ov
e the sp
eed
of c
o
nve
rge
n
c
e
t
o
a desi
rabl
e reco
nst
r
uct
i
o
n. T
h
e a
ppl
i
c
a
t
i
on o
f
a t
r
a
n
s
f
o
r
m
a
ti
on
of t
h
e i
m
age i
n
b
e
t
w
een
two
iterativ
e st
ep
s
.
To
b
e
m
a
t
h
em
at
ically p
r
ecise, in
an
ART-typ
e
algo
ri
th
m
with
trick
,
th
e equ
a
tio
n
(1
5) is
repl
ace
d by
(
1
7
)
=
T
(
(
1
8
)
Wh
ere T is th
e tran
sfo
r
m
a
tio
n
th
at d
e
fin
e
s the trick
.
4.
RESULTS
Case
S
t
ud
y
1:
For
t
h
e
da
rk
ar
eas t
h
e sam
e
i
m
age t
a
ke
n f
r
o
m
t
w
o
di
ffe
re
n
t
angl
es
are
gi
v
e
n as
i
n
put
,
as
sho
w
n i
n
Fi
gu
r
e
4.
Fig
u
re
4
.
Two in
pu
t im
ag
es
The i
n
put im
ages after
h
o
m
o
g
r
ap
hy
war
p
i
n
g,
as s
h
o
w
n i
n
Fi
gu
re
5.
Fi
gu
re 5.
H
o
m
o
g
r
a
phy
war
p
i
n
g
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
I
m
a
g
e
Mo
sa
ic
in
g fo
r W
i
d
e
A
n
g
l
e
Pa
no
ra
ma
(G
.
D
i
vya
)
1
222
For each
desc
riptor in t
h
e first im
age, select its
m
a
tch
to second im
a
g
e wit
h
lines
joi
n
ing the ac
cept
e
d
m
a
t
c
hes, i
.
e., t
h
e feat
u
r
e m
a
tchi
n
g
as sh
ow
n i
n
Fi
g
u
re 6
.
For i
n
p
u
t
im
age 1, t
o
t
a
l
17
38
key
-
poi
nt
s fo
u
nd a
n
d
fo
r i
n
p
u
t
im
age 2, 14
2
8
key
-
poi
nt
s fo
u
n
d
.
Out
o
f
t
h
ese k
e
y
poi
nt
s, t
h
e
num
ber o
f
m
a
tches i
n
b
o
t
h
i
m
ages i
s
13
3.
Fi
gu
re
6.
Feat
ure
m
a
t
c
hi
ng
Th
ese m
a
tch
e
s are
sub
m
it
ted
to
RANSAC
wh
ich
calcu
l
ates
a tran
sform
a
t
i
o
n th
at align
s
th
e
p
o
i
n
t
s in imag
e1
and
i
m
age2 an
d al
s
o
ret
u
r
n
t
h
e i
n
l
i
e
rs a
n
d
be
st
m
a
t
c
hed
poi
nt
s, as
sh
o
w
n
i
n
Fi
gu
re
7.
Fi
gu
re
7.
B
e
st
m
a
t
c
hed p
o
i
n
t
s
Finally using the best m
a
tched points the s
titch
m
odul
e s
titches the bot
h im
ages gives an output which i
s
pan
o
r
am
i
c
vi
ew
of
b
o
t
h
i
m
ag
es as, s
h
ow
n i
n
Fi
g
u
re
8
.
Evaluation Warning : The document was created with Spire.PDF for Python.
I
S
SN
:
2
088
-87
08
I
JECE Vo
l. 5
,
N
o
. 5
,
O
c
tob
e
r
20
15
:
121
6
–
12
26
1
223
.
Fig
u
re
8
.
Fi
n
a
l stitch
e
d
im
ag
e
From
Fi
gu
re
8
,
fi
n
d
t
h
at
t
h
e
out
put
i
m
age i
s
ha
vi
n
g
m
o
re resol
u
t
i
o
n an
d m
o
re cl
ari
t
y
t
h
an t
h
e gi
ve
n i
n
put
im
ages.
Case
S
t
ud
y
2:
For t
h
e l
i
g
ht
i
n
g areas t
h
e sa
m
e
im
age t
a
ken fr
om
t
w
o di
ffe
re
nt
angl
e
s
are gi
ve
n as i
nput
, as s
h
o
w
n i
n
Fi
gu
re 9.
Fi
gu
re 9.
Tw
o i
n
p
u
t
i
m
ages
The i
n
put im
ages after
h
o
m
o
g
r
ap
hy
war
p
i
n
g,
as s
h
o
w
n i
n
Fi
gu
re
1
0
.
Evaluation Warning : The document was created with Spire.PDF for Python.
IJECE
ISS
N
:
2088-8708
I
m
a
g
e
Mo
sa
ic
in
g fo
r W
i
d
e
A
n
g
l
e
Pa
no
ra
ma
(G
.
D
i
vya
)
1
224
Fi
gu
re 1
0
. H
o
m
ograp
hy
wa
r
p
i
n
g
For each
desc
riptor in t
h
e first im
age, select its
m
a
tch
to second im
a
g
e wit
h
lines
joi
n
ing the ac
cept
e
d
m
a
t
c
hes. The f
eat
ure m
a
t
c
hi
ng, as s
h
o
w
n i
n
Fi
gu
re 1
1
. F
o
r i
n
p
u
t
i
m
age
1, t
o
t
a
l
1
6
8
2
0
key
-
poi
nt
s fo
u
nd a
n
d
fo
r i
n
p
u
t
i
m
age 2,
4
2
1
0
key
-
poi
nt
s f
o
un
d.
Out
o
f
t
h
es
e
k
e
y
-
p
o
i
n
t
s
, t
h
e
num
ber
of
m
a
tches i
n
bot
h i
m
ages i
s
:
72
3 a
n
d
uni
qu
e m
a
t
c
hes are:
62
9.
Fi
gu
re 1
1
.
Fea
t
ure
m
a
t
c
hi
ng.
Th
ese m
a
tch
e
s are
sub
m
it
ted
to
RANSAC
wh
ich
calcu
l
ates
a tran
sform
a
t
i
o
n th
at align
s
th
e
p
o
i
n
t
s in imag
e1
and
i
m
age2 an
d al
s
o
ret
u
r
n
t
h
e i
n
l
i
e
rs a
n
d
be
st
m
a
t
c
hed
poi
nt
s, as
sh
o
w
n
i
n
Fi
gu
re
1
2
.
Fi
gure
12.
Best m
a
tched
points.
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
:
121
6
–
12
26
1
225
Finally using the best m
a
tched points the s
titch
m
odul
e s
titches the bot
h im
ages gives an output which i
s
pan
o
r
am
i
c
vi
ew
of
b
o
t
h
i
m
ag
es, as s
h
ow
n i
n
Fi
g
u
re
1
3
.
Fig
u
re
13
.
Final stitch
e
d
im
a
g
e
Fro
m
Fig
u
r
e
13
, fi
n
d
th
at th
e o
u
t
pu
t im
ag
e
is h
a
v
i
ng
m
o
re reso
lu
tion
an
d
m
o
re clarity t
h
an
th
e
g
i
v
e
n
in
pu
t
im
ages.
5.
CO
NCL
USI
O
N
The res
u
l
t
s
obt
ai
ned we
re rea
s
on
abl
y
go
o
d
f
o
r t
h
e da
rk a
r
e
a
s. B
u
t
for t
h
e
l
i
ght
i
ng a
r
eas,
t
h
e
m
a
sk is
gene
rat
i
n
g
fo
r
t
h
e pa
n
o
ram
i
c im
age. Thi
s
m
a
y
be
du
e t
o
n
o
n
-i
deal
t
r
a
n
sf
o
r
m
a
ti
ons a
n
d t
h
res
h
ol
d
val
u
e
s
set
i
n
th
e ho
m
o
g
r
aphy alg
o
r
ith
m
.
In
fu
t
u
re
wo
rk
, th
e m
a
sk
will b
e
redu
ced b
y
u
s
i
n
g
Ab
stract Reco
n
s
t
r
u
c
tio
n
Techn
i
q
u
e
s fo
r t
h
e
lig
h
tin
g
areas an
d
also
lik
e to
p
e
rform
a
m
o
r
e
d
e
taile
d
,
qu
an
titativ
e an
alysis o
f
feature match
i
n
g
p
e
rfo
rman
ce
on
l
a
r
g
e
dat
a
ba
ses o
f
pa
no
ram
i
c im
ages.
REFERE
NC
ES
[1]
Yan Gong, Hon
g
Xie and Lei Y
u
, “Research an
d Anal
y
s
is of Key
Techno
logies
in Image Mosaic”,
Internationa
l
Journal of
Signa
l Processing, Im
age
Processing
and Pat
t
ern R
e
cognition
, vol. 6
No. 5, 2013, pp.
237-243.
[2]
Harshal JPRET, “Image Mosaicing A
pproach and
Evalu
a
tion Methodolog
y
”
,
I
n
ternational Jou
r
nal of Advan
c
ed
Technology &
E
ngineerin
g
Res
e
arch (
I
JATER)
,
Volume 1 (2013
), pp
.
576-585.
[3]
Rafael C
.
Gonzalez, Richard
E.W
oods, “Digital
Image Processin
g
”.
[4]
Li Jin, Wang
Yanwei,
Lia
ng
Hng, “Image Mosaic Ba
se
d on Simplifie
d SIFT
”
,
Internationa
l Conferen
ce o
n
Mechanical Eng
i
neering
and Automation
, vo
l. 6,
2012, pp
. 90-95
.
[5]
Tej
a
sha Pati
l,
Shweta Mishra, Poor
va Chaudhari, Shalaka
Khandale,
“Im
a
ge stitch
i
ng
using m
a
t lab”,
International Jo
urnal of
E
ngineering Trends and
Technology
, vo
l. 4, Issue
3, 2013, pp. 302-306.
[6]
S
a
t
y
a P
r
ak
as
h M
a
lli
ck,
“
F
eature
Bas
e
d Im
age
M
o
s
a
icing
”
,
Unive
rsity of
Cali
forni
a
, San
Diego.
[7]
M
a
lavik
a
Bhas
k
a
ranand
and
S
a
n
d
eep Bh
at
,
“Ima
ge
Re
gi
st
ra
ti
on and
Mosaicking
”,
Universit
y
o
f
C
a
lifornia.
[8]
Guido Barto
li, “Image
Registr
a
tio
n Techniques
”
,
Universit_adeg
l
i Studi di S
i
ena
, J
une 2007.
[9]
David G. Lowe
,
“
D
istinctive Im
age Fea
t
ures fro
m Scale-Invariant Key
points”,
I
n
ternational
Jou
r
nal of Computer
Vision
, 2004.
[10]
Konstantinos G.
Derpanis, “overview of
the
RANSAC algorithm”, V
e
rsion 1
.
2,
May
13, 2010.
[11]
HstauY.Liao
,
”A
Gradually
Unmasking
Method for Limited Data Tomograph
y
,”
Institute for Mat
h
ematics and its
Applica
tions.
Universit
y
of Minn
esota Minneapo
lis,
MN 55455, U
S
A.
[12]
S
j
ors
H.W
.
S
c
he
res
,
Roberto ar
a
b
ini, S
a
lv
atore L
a
nzav
ec
ch
ia
, Rances
c
a
Cant
ele
,
Twan
Rutten, tephen D. Fuller
,
Jose
´ M.
Ca
ra
z
o
,
Roge
r M.
Burnett, Carmen San Mart
ı
´
n
, “
C
lassifica
tion of single-
projection reconstructions for
cr
y
o
-electron icr
o
scop
y
data
of
icosahedral viruses”,
Journal of Structural
Biology 151 (
2005)
Elsevier
, pp. 79–91
Evaluation Warning : The document was created with Spire.PDF for Python.