TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.7, July 201
4, pp
. 5645 ~ 56
5
4
DOI: 10.115
9
1
/telkomni
ka.
v
12i7.585
5
5645
Re
cei
v
ed Fe
brua
ry 24, 20
13; Re
vised
Ma
rch 25, 20
14; Accepted
April 12, 201
4
Digital Image Stabilization Based on Improved Scale
Invarian
t
Feature Transform
Xiaoran Guo
*
, Shaohui Cui, Dan Fang
Ordnanc
e Eng
i
neer
ing C
o
ll
eg
e, Shiji
azhu
an
g 050
00
3, Chin
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: vip85
05
22@
163.com
A
b
st
r
a
ct
A nove
l
di
gital
imag
e stabi
li
z
a
tion a
ppro
a
ch
usin
g Harris
a
nd Sca
l
e Invar
i
ant F
eature T
r
ansfor
m
(SIF
T
)
w
a
s pre
s
ented i
n
this article. Usin
g SIF
T
in di
gital i
m
a
ge stab
ili
z
a
t
i
on, t
oo many feature p
o
ints a
n
d
match
e
s
w
e
re
extracted,
b
u
t some of
the
m
w
e
re
not
s
o
st
abl
e. Usi
ng t
h
ese fe
ature
po
ints a
nd
match
e
s
can n
o
t only i
n
creas
e the c
o
mputati
o
n
a
l e
ffort, but
also enh
anc
e the
w
r
ong matchi
n
g
prob
ab
ility. W
e
prop
osed to us
e SIF
T
to detect featur
e point
s and inc
o
rpor
ate the Harris
crit
erion to se
l
e
ct the most stabl
e
feature p
o
ints
in the vi
deo s
equ
enc
e w
here imag
e motio
n
w
a
s happ
en
ed du
e to veh
i
cle or p
l
atfo
r
m
vibrati
on. W
i
th these feat
ur
e p
o
ints, w
e
use
gen
eral fe
ature
descriptor
and
match
i
ng
alg
o
r
ithm to ac
hi
ev
e
the i
m
age
stab
i
l
i
z
a
t
io
n. Exp
e
ri
me
nt
al
resu
lts
show
that th
e p
r
opos
ed
al
gorit
hm ca
n n
o
t o
n
l
y
brin
g
dow
n th
e
prob
abi
lity of
w
r
ong
matchi
n
g
an
d g
e
t
mor
e
accur
a
te
ma
tches, but a
l
so
reduc
e the
co
mp
utatio
n b
u
rd
e
n
than SIF
T
effe
ctively.
Ke
y
w
ords
: dig
i
tal i
m
a
ge stabi
li
z
a
tio
n
, SIF
T
,
Harris
Copy
right
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
Video stabili
zation tech
niq
ues h
a
ve bee
n studi
e
d
for decade
s to improve visua
l
qualities
of image se
q
uen
ce
s captu
r
ed by digital
video came
ras which a
r
e
hand held o
r
mounte
d
on
unsta
ble pl
atform
or ve
hicl
e, the captu
r
ed video
ge
n
e
rally lo
oks
shaky b
e
cau
s
e of un
de
sire
d
came
ra motio
n
s. Un
wante
d
video vibrations w
ould le
ad to degrad
ed view expe
rien
ce an
d also
greatly affe
ct the pe
rform
ances
of ap
plicatio
ns li
ke video
co
di
ng, video
su
rveillan
c
e, o
b
j
ect
detectio
n
, target tra
cki
ng,
image
naviga
t
ion and
guid
ance. Video
stabili
zation i
s
be
co
ming
an
indispen
sabl
e
techniq
ue in
indu
strial, m
ili
tary and co
nsumer a
ppli
c
at
ions field
s
.
The video
st
abilization sy
stem
s can b
e
cla
s
si
fied i
n
to three
ma
jor types: el
ectro
n
ic
image
stabili
zation
(EIS), o
p
tional im
age
stabili
zati
o
n
(OIS),
a
nd di
gital
imag
e st
abilization (DI
S
)
[1]. The EIS stabili
ze
s the
image
se
qu
ence by em
p
l
oying motion
sen
s
o
r
s, su
ch as gyro
sco
pe
and accel
e
ro
meter, to detect the came
ra move
me
nt for compe
n
sation. The OIS adopts a prism
assembly whi
c
h move
s op
posite the
sh
akin
g of ca
m
e
ra for
stabili
zation [2]. Th
e appli
c
ation
s
of
EIS and OIS are rest
ricte
d
to device, be
cau
s
e
b
o
th are hardware
d
epen
dent. DI
S is the pro
c
ess
of removing t
he und
esi
r
ed
motion affect
s to gene
rate
a stable ima
ge se
que
nce by using di
gital
image processing
tech
nique
s with
out
any
m
e
ch
ani
cal d
e
vice
s
su
ch
as gyro
scope,
accele
rom
e
te
r, or a fluid
pri
s
m.
DIS system
e
s
sentially co
nsi
s
ts
of two
u
n
its: the m
o
tion
estimation u
n
i
t and the motion com
pen
sation units.
The motion e
s
timation unit
estimates th
e global moti
on paramete
r
s between ev
ery two
con
s
e
c
utive f
r
ame
s
of the
input vid
eo
seq
uen
ce
s.
With the
s
e
gl
obal m
o
tion
para
m
eters, t
he
motion
comp
ensation u
n
it then g
ene
rates th
e co
mpen
sating
motion p
a
ra
meters n
eed
ed to
comp
en
sate f
o
r the
jitter of
a fram
e a
n
d
create
s
a
m
o
re vi
sual
sta
b
le ima
ge
se
quen
ce.
Unli
ke
most motio
n
estimation te
chni
que
s, in
DIS the
ro
bu
stne
ss
of the
motion e
s
timation is
critical
asso
ciated
with the fact t
hat an i
m
proper e
s
ti
mat
e
will yiel
d
an a
b
ru
pt jitter in
the vid
eo
seq
uen
ce.
Variou
s alg
o
rithms have b
een devel
op
ed to es
timat
e
the local
motion vecto
r
s. Block
matchin
g
[3] is the co
nvent
ional way to detect
the glob
al motion vector betwe
en two con
s
e
c
uti
v
e
frame
s
. In this method, e
a
ch ima
ge is divided in
to
squa
re
s, re
ctangl
es o
r
circul
ars [1] of a
certai
n dime
n
s
ion, an
d the
n
motion sea
r
ch i
s
exe
c
ut
ed to find the
motion vecto
r
for ea
ch bl
o
c
k
in cu
rrent fra
m
e. With the
s
e lo
cal
moti
on vecto
r
s, cl
usteri
ng m
e
th
od [3] is
used
to find the m
a
in
motion ve
cto
r
which i
s
reco
gni
zed
a
s
the
glob
al
motion ve
ct
or. Be
cau
s
e
of the tedi
ous
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5645 – 56
54
5646
comp
utationa
l cost of blo
ck m
a
tchi
ng
with full
pixel information,
some
sche
mes h
a
ve b
een
prop
osed to
spe
ed
up the
pro
c
e
s
s
with only
partial
pixel info
rm
ation, such a
s
rep
r
e
s
entat
ive
point mat
c
hin
g
sch
e
me [4],
sel
e
cte
d
a
r
e
a
s m
a
tchi
ng
scheme
[5], b
i
t-plane
matching
schem
e
[6]
and gray-cod
ed bit-pla
ne matchin
g
sch
e
me [7].
These sch
e
me
s can redu
ce t
he com
putati
on
compl
e
xity, but also lo
we
r the motion e
s
timati
on accu
racy. Besi
de
s block matchi
ng method
wi
th
full or
partial
pixel info
rm
ation, opti
c
al
flow [8] b
a
sed meth
od
h
a
s
been
p
r
o
posed fo
r im
age
stabili
zation.
In optical flo
w
alg
o
rithm t
he con
s
tr
aint
is the first-o
r
de
r ap
proxi
m
ation of Ta
ylor
seri
es, which
means that the motion
s can not be ve
ry large. Othe
rwi
s
e, the hig
her orde
r terms
of Taylor
se
ri
es
whi
c
h a
r
e
igno
red,
can
lead to
sig
n
i
f
icant in
co
rre
ct estim
a
tion.
Some featu
r
e
matchin
g
ba
sed algo
rithm
s
, such
as p
h
a
se
co
rrel
a
tio
n
schem
e [9], and edge
p
a
ttern matchi
n
g
scheme
[10]
have al
so
be
en p
r
o
p
o
s
ed
in DIS. T
h
e
s
e
sch
e
me
s a
r
e lack of ad
aptation to affine
and pe
rspe
ctive motions.
The m
o
re ef
fective meth
ods for affin
e
an
d
persp
ective m
o
tio
n
e
s
timation
are lo
cal
feature
s
ba
se
d one
s. Ha
rri
s [11]
co
rne
r
feature
s
are extracted
and
tracked in o
r
der to e
s
tima
te
global motio
n
[12]. Scale
invariant fe
ature tra
n
sf
o
r
m (SIFT) [1
3] feature
s
, con
s
id
ere
d
to be
invariant to
i
m
age
rotatio
n
an
d
scaling
,
are
bein
g
widely u
s
ed
in
the late
st me
thods fo
r gl
o
bal
motion e
s
tim
a
tion [14], image mat
c
hin
g
[15] and ima
ge re
gist
ratio
n
[16]. Also
many wo
rks
are
prop
osed to
improve the
distinctivene
ss of SI
FT descri
p
tor, such a
s
pri
n
cipal com
pon
ent
analysi
s
(PCA)-SIFT [17] whi
c
h wa
s le
ss di
stin
ctive than the SIFT
descripto
r proved by [18],
and
spe
ede
d up
robu
st feature
s
(S
URF)
[19]
whi
c
h
is fast
in matching
a
nd
suf
f
i
cie
n
t
l
y
dist
in
ct
iv
e,
b
u
t
wea
k
e
r
than
SIFT in the invariant of scalin
g and ro
tation. Such above-mentio
ned app
ro
aches
mainly focu
s on the development of d
e
scripto
r
s,
but in this paper, we
care
more abo
ut the
extraction of
the most stable
feature points,
which
can not only
bring down t
he probability of
wro
ng m
a
tchi
ng an
d get m
o
re
accu
rate
matche
s,
b
u
t also
effectively redu
ce
th
e co
mputatio
nal
effort. By carefully reviewi
ng the su
rve
y
on feat
ure point detecto
rs [20], we n
o
te that Harris
corne
r
could
provide
sta
b
le dete
c
tion
perfo
rman
ce with hig
h
repe
atability and lo
cali
zati
on
accuracy un
der vario
u
s
distortio
n
s a
nd geom
et
ri
c tran
sform.
Therefo
r
e, we propo
se
to
inco
rpo
r
ate t
he
Harri
s
cri
t
erion i
n
SIF
T
to se
le
ct the mo
st
sta
b
le featu
r
e
p
o
ints fo
r im
a
g
e
stabili
zation.
With these most sta
b
le
featur
e p
o
int
s
, we u
s
e g
eneral feature descriptor
and
matchin
g
al
g
o
rithm to
ge
nerate
match
e
s, an
d
u
s
e
those
mat
c
hes to
calcul
ate geo
metri
c
al
transfo
rm p
a
rameters. Fin
a
lly, we tran
sfor
m
one im
age u
s
ing th
e geom
etri
ca
l transfo
rmati
o
n
matrix to align with the oth
e
r imag
e, and
accompli
sh i
m
age comp
e
n
satio
n
.
2. Robus
t Lo
cal Feature
Points Extr
a
c
tion
Local feature
s
, su
ch as
co
rne
r
s, blob
s,
and re
gion
s, have been
wi
dely used fo
r object
detectio
n
, re
cog
n
ition, an
d retrieval pu
rpo
s
e
s
in
co
mputer visio
n
.
The intrinsi
c advantag
es of
these lo
cal fe
ature
s
are their invaria
n
ce
under g
eom
etric tra
n
sfo
r
ms. A comp
rehen
sive revi
ew of
the state
-
of-t
he-a
r
t lo
cal f
eature
s
ca
n
be fou
nd
in
[18]. Among
variou
s lo
cal
feature
dete
c
tors
and de
script
ors,
SIFT wa
s sho
w
n
to be relative
ly
optimal
con
s
i
derin
g the
trade-off bet
ween
robu
stne
ss, distin
ctivene
ss,
an
d efficie
n
cy,
an
d
Ha
rris corn
er co
uld p
r
ovide
stable d
e
tecti
o
n
perfo
rman
ce
with high
re
peatability an
d locali
zatio
n
accuracy u
nder va
riou
s distortio
n
s
and
geomet
ric tra
n
sform.
The ori
g
inal
SIFT algorith
m
con
s
ist
s
of the followin
g
four maj
o
r ste
p
s:
(1) S
c
ale
-
spa
c
e extrem
e d
e
tection
(2) A
c
curate f
eature p
o
ints
locali
zation
(3) O
r
ientatio
n assignm
ent
(4) F
eature p
o
ints de
script
or
Whe
r
e
ste
p
s
1 an
d
2 a
r
e
the featu
r
e
de
tection
and
steps
3
and
4
are the
de
scription
and ge
neratio
n of the feature point
s.
Our
m
e
thod u
s
e
s
the step 1,
step
3 and step
4 of
the
origin
al SIFT
algorith
m
, but
in step
2, when the
potential feature poi
nts ha
ve all
found, and the point
s of the edg
e and the lo
wer
contrast
point
s
were elimi
n
ated. Yet there are
st
ill too many feature points,
and some
of them
are
not so
stable, if ad
opti
ng all th
e feat
ure
point
s to
execute th
e f
eature
matchi
ng, not o
n
ly the
computational effort
will be tr
emendous,
but al
so t
he wrong matchi
ng
probab
ility
will be high. To
further i
dentif
y the most
st
able f
eatu
r
e
points,
we in
corpo
r
ate the
Ha
rri
s
criteri
on to
sele
ct the
most sta
b
le
SIFT feature
points.
Th
e u
nderlyin
g re
a
s
on i
s
that such a
crite
r
io
n help
s
keep
the
most stable
l
o
cal pattern
s
with
hig
her gradi
ents di
st
ribution
but reject
s the fe
ature p
o
ints
with
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Digital Im
age Stabilization
Based o
n
Impro
v
ed S
c
al
e
Invari
ant Feat
ure T
r
an
sform
(Xiaoran G
uo)
5647
lowe
r g
r
adi
e
n
t distri
butio
n. Also, the
Ha
rri
s-ba
se
d criterio
n i
s
self-ada
ptive and
can
yield
relatively stab
le detectio
n
p
e
rform
a
n
c
e.
The ste
p
s of the pro
p
o
s
ed
algorith
m
are:
(1) S
c
ale
-
spa
c
e extrem
e d
e
tection
The first sta
g
e
of comput
ation sea
r
che
s
over all
scale
s
an
d ima
ge l
o
catio
n
s to
find lo
cal
extrema. It is
impleme
n
ted
efficiently by usin
g
a seri
e
s
of differe
nce-of-Gau
ssia
n (DOG
) ima
ges
in the scale
spa
c
e
to identify poten
tial feature p
o
ints
that are invariant to scale an
d
orientatio
n.
The im
age
scale
spa
c
e
is
expre
s
sed
as a fu
nction
)
,
,
(
y
x
L
, that is ge
ne
rat
ed from th
e
convol
ution o
f
a serie
s
of
Gau
ssi
an kernel functio
n
s
)
,
,
(
y
x
G
with con
s
e
c
utively incre
m
ental
scale
s
, with a
n
input image
)
,
(
y
x
I
:
)
,
(
)
,
,
(
)
,
,
(
y
x
I
y
x
G
y
x
L
(1)
Whe
r
e
denote the convol
u
t
ion operator,
and the Gau
ssi
an fun
c
tion
)
,
,
(
y
x
G
:
2
2
2
2
/
)
(
2
2
1
)
,
,
(
y
x
e
y
x
G
(2)
The scal
e sp
ace
)
,
,
(
y
x
D
is set u
p
from the di
fference of two ne
arby
scale
s
by a
constant mult
iplicative fact
or
k
.
)
,
(
))
,
,
(
)
,
,
(
(
)
,
,
(
)
,
,
(
)
,
,
(
y
x
I
y
x
G
k
y
x
G
y
x
L
k
y
x
L
y
x
D
(3)
2) Accu
rate feature p
o
ints
loca
li
zation with
Harris crit
erion
Whe
n
the po
tential feature points h
a
ve all
found, to further id
e
n
tify the more stable
feature poi
nts, we will eli
m
inate the unstabl
e poi
nt
s, such as the point
s of the edge and the
lowe
r contrast points. Only
the
stabl
e fe
ature p
o
ints
remain. Acco
rding to Taylo
r
expan
sio
n
, let
the derivative
of
)
(
x
D
equal zero after an off
s
et
x
ˆ
set can
be found. T
h
i
s
x
ˆ
can take a
pixe
l
locatio
n
with
true lo
cal
extreme valu
e, then sub
s
titute
x
ˆ
into the T
a
ylor exp
a
n
s
ion.
If a value of
|
)
x
ˆ
D(
|
is less than
03
.
0
,
we will
say the point
has low cont
rast. The
)
(
x
D
is co
mput
ed belo
w
.
x
x
D
x
x
x
D
D
x
D
2
2
2
1
)
(
T
T
(4)
x
D
x
D
x
2
1
2
ˆ
(5)
|
ˆ
2
1
|
)
ˆ
(
x
x
D
D
x
D
T
(6)
After eliminat
ing the point
s
with low cont
rast, we
will cancel the point
s of t
he edge
perh
a
p
s
. Here we use a
2
2
He
ssi
an
mat
r
ix
E
which ena
bles u
s
to co
mpute the scale an
d
locatio
n
of the feature poi
n
t. The matrix is given by:
yy
xy
xy
xx
D
D
D
D
E
(7)
Whe
r
e
xx
D
is the second de
riv
a
tive of t
he image in the
x-dire
ction, a
nd
xy
D
is the first
derivative of the image in the x-dire
ction and y-di
rectio
n re
sp
e
c
tively,
yy
D
is the se
co
nd
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5645 – 56
54
5648
derivative of t
he ima
ge i
n
t
he y-di
re
ction
.
Then,
E
is do
ne in
the
de
compo
s
ition
of eige
nvalue
s
to get the two eigenvalue
s:
and
.
We u
s
e the ratio of the eigenvalue
s a
s
follows:
2
2
)
1
(
)
(
)
(
)
(
E
E
Det
Tr
(8)
(9)
The followi
ng
inequality must be satisfi
ed, if not, we say that this point is pro
b
ably on
the edge, an
d
it can be eli
m
inated like this:
2
)
1
(
)
(
)
(
E
E
Det
Tr
(10)
So far,
we
ge
t the a
set of
SIFT point
s
)
,
,
(
n
n
n
y
x
P
, w
h
er
e
n
x
is the
hori
z
ontal
o
r
d
i
nate,
n
y
is the ve
rtical ordinate,
and
n
is the
scale p
a
ram
e
ters of the
SIFT feature
point
n
.
Suppo
se tha
t
there are t
o
tal
N
feature
points, so
N
n
,...
3
,
2
,
1
. Then, we a
dd the Ha
rri
s
crit
e
r
ion
t
o
s
e
lect
t
h
e mo
st
st
a
b
le S
I
F
T
f
eat
ur
e p
o
i
n
ts. The
Harris
criterio
n i
s
b
a
sed o
n
t
he
autocorrelatio
n
matrix, whi
c
h rep
r
e
s
ent
s the g
r
adi
e
n
t distrib
u
tion
within a l
o
cal regi
on of t
he
sele
cted
poin
t. We a
dopt t
he
n
x
,
n
y
,
n
to cal
c
ulate the
Ha
rris respon
se
)
,
(
n
n
y
x
R
n
, where
n
is the
stand
a
r
d deviatio
n
of the Gau
ssian kern
el fu
nction
as th
e
weig
hted
windo
w u
s
ed to
comp
ute the
Harri
s
auto
c
o
rrel
a
tion mat
r
i
x
M
, which
re
prese
n
ts the
gradi
ent di
strib
u
tion withi
n
a
local
regi
on
of the sele
ct
ed point. Th
e Gau
s
sian
kernel
wind
o
w
in this
pa
per i
s
3
3
. The
autocorrelatio
n
matrix
M
of at feature point
)
,
,
(
n
n
n
y
x
P
is
r
e
pr
es
e
n
t
ed
a
s
:
)
,
(
)
,
(
2
2
)
,
(
n
n
y
x
W
y
x
y
y
x
y
x
x
I
I
I
I
I
I
y
x
M
(11)
)
,
(
n
n
y
x
W
is th
e G
a
u
s
sian
kern
el
windo
w
with th
e
)
,
(
n
n
y
x
as th
e
ce
nter to
dete
r
mine
the accum
u
la
ted regi
on, a
nd
)
,
(
y
x
is the Ga
ussian
ke
rnel
function of t
he wei
ghted
wind
ow,
x
I
and
y
I
are
the
image
g
r
adi
ents i
n
x-dire
ction
and
y-d
i
rectio
n. So
2
x
I
is
th
e pr
o
d
u
c
t o
f
the
image
gradie
n
ts in
x-di
rection. We
u
s
e
the G
a
u
ssi
a
n
kern
el fun
c
tion
)
,
(
y
x
as th
e
weig
hted
wind
ow to m
a
ke the
matrix
isotro
pi
c. If b
o
th eige
nvalu
e
s of m
a
trix
M
,
1
and
2
are
suffi
ciently
large
po
sitive
value
s
, this
point i
s
a
co
rne
r
p
o
int. Harri
s
pro
p
o
s
e
d
a fo
rmula
to evaluate
th
e
corne
r
point
s instea
d of co
mputing the two eig
envalu
e
s a
s
:
)
(
)
det(
)
(
2
2
1
2
1
M
trace
M
R
(12)
Whe
r
e
is a coeffici
ent wit
h
value 0.04-0.06. We
set 0.06 as its d
e
fault value in this
pape
r.
)
det(
M
is th
e dete
r
mina
n
t
of matrix
M
, and
)
(
M
trace
is th
e trace
of matrix
M
. We
cal
c
ulate th
e
)
,
(
n
n
y
x
R
n
of ea
ch fea
t
ure p
o
int
)
,
,
(
n
n
n
y
x
P
, and calculate
the mea
n
val
ue of
them. Finally, we set the m
ean value a
s
the thre
shold
to sele
ct robu
st SIFT feature points:
N
n
n
n
y
x
R
N
Threhold
n
1
)
,
(
1
(13)
If the
)
,
(
n
n
y
x
R
n
of feature poi
nt is greate
r
than
Threhold
, we keep this
feature poi
nt
and ta
ke
it a
s
a mo
st
stabl
e feature p
o
i
n
t, otherwise
reje
ct this fea
t
ure p
o
int.
With this th
re
sh
old,
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Digital Im
age Stabilization
Based o
n
Impro
v
ed S
c
al
e
Invari
ant Feat
ure T
r
an
sform
(Xiaoran G
uo)
5649
we ca
n re
serve
the
m
o
st stable
f
eatu
r
e
point
s
with
h
i
gher g
r
adi
en
ts di
strib
u
tion
, and
rej
e
ct t
he
feature
point
s with
lower g
r
adient
dist
rib
u
tion, an
d
th
en we co
uld
extract stabl
e
feature
p
o
int
in
image
s
unde
r g
eomet
ric transfo
rm
s eve
n
un
de
r the
a
ttack of
blurring.
Du
e to t
h
is th
re
sh
old
is
based on
Harri
s alg
o
rith
m, it is self-ada
ptiv
e an
d can yield
relatively stable dete
c
tion
perfo
rman
ce.
3) Ori
entation
Assign
ment
One o
r
mo
re
orientatio
ns
are a
s
sign
ed
to each fe
ature p
o
int location ba
sed
on local
image g
r
a
d
ie
nt dire
ction
s
. All future o
peratio
ns
are
perfo
rmed
o
n
image
dat
a that ha
s b
een
transfo
rme
d
relative to the assig
ned o
r
i
entati
on, scal
e, and locati
on fo
r each feature, there
b
y
providin
g invarian
ce to the
s
e tran
sfo
r
ma
tions.
After finding
the featu
r
e
points, th
e
fo
llowin
g
step will
defin
e the ma
gni
tude an
d
orientatio
n fo
r ea
ch featu
r
e point in
order to
u
s
e i
m
age m
a
tchi
ng. The o
r
ie
ntation of ea
ch
feature
point
is d
e
termi
n
e
d
by the
pea
k of th
e o
r
ie
ntation hi
stog
ram fo
rme
d
by the g
r
adi
ent
orientatio
ns
within it
s nei
ghbo
rho
od.
After fi
nding
the majo
r o
r
i
entation fo
r
an ima
ge fe
ature
point, whe
n
doing the im
age mat
c
hin
g
again, we can
rotate to
the same o
r
ientation in t
w
o
image
s, and thus a
c
hieve rotation-inva
ri
ance. Here
we are u
s
ing the distri
butio
n of the gradi
ent
dire
ction
fo
r
feature point
s taken as a
co
nsi
s
tent
o
r
ien
t
ation to ea
ch
feature
point
. Therefo
r
e,
we
will compute
the gradient
magnitud
e
a
nd t
he o
r
ie
ntation for
ea
ch imag
e sam
p
le
)
,
(
y
x
L
. The
comp
utation
of the gradie
n
t
magnitude a
nd the
orie
nta
t
ion in pixels
are given a
s
belo
w
.
2
2
))
1
,
(
)
1
,
(
(
))
,
1
(
)
,
1
(
(
)
,
(
y
x
L
y
x
L
y
x
L
y
x
L
y
x
m
(14)
)))
,
1
(
)
,
1
(
/(
))
1
,
(
)
1
,
(
((
tan
)
,
(
1
y
x
L
y
x
L
y
x
L
y
x
L
y
x
(15)
After the g
r
a
d
ient di
re
ctio
ns
of all
pixe
ls
within a re
gion aro
und
the
featu
r
e p
o
int
a
r
e
comp
uted, we cal
c
ul
ate th
e dire
ction
of the most
p
o
i
n
ts supp
ortin
g
the pri
m
ary
dire
ction. Th
e
weig
ht of each poi
nt adja
c
ent to the central pixe
l
s
is de
cide
d u
s
ing the p
r
o
d
u
ct of Gau
ssian
distrib
u
tion
s and gradie
n
t magnitude o
f
the point,
whe
r
e Ga
ussian distri
butio
n is a setting
at
5
.
0
.
4) Featu
r
e Po
ints De
scripto
r
The local ima
ge gra
d
ient
s are mea
s
u
r
e
d
at
the selected scale in the regi
on aro
und ea
ch
feature p
o
int. These are transfo
rme
d
in
to a rep
r
e
s
e
n
tation that a
llows for si
gn
ificant levels
of
local
shap
e
distortio
n
a
n
d
chan
ge i
n
illuminat
io
n. He
re
we a
dopte
an
ap
proa
ch
which is
basi
c
ally ba
sed on the ori
entat
ion histo
g
ram. The
computati
on a
nd de
scriptio
n of the image
gradi
ent is d
one by usi
n
g
the feature
poi
nts a
nd a
d
jacent pixel
s
, we em
plo
y
a
16
16
local
neigh
borhoo
d
blo
ck
around
the featu
r
e
p
o
ints, the
n
di
vide it up i
n
to
4
4
su
b-blo
c
ks
again, a
nd
define the ei
ght dire
ction
s
,
therefo
r
e, we treat the
128
8
4
4
dimensio
na
l vector as t
he
feature point
s
descriptor.
The feature
vector
i
s
then normali
zed to decrease the illuminat
ion
alteration effec
t
s
.
3. Featur
e M
a
tching a
nd
Image Comp
ensa
te
Once the
mo
st sta
b
le fe
ature
point
s h
a
v
e bee
n d
e
te
cted i
n
the
im
age
s, the
nex
t step i
s
to matc
h them. In this
paper, f
eatu
r
e p
o
ints a
r
e m
a
tche
d by con
s
tructing
KD-t
ree with
be
st-bin-
first se
arch,
the advantag
e of KD-tre
e
is red
u
ci
ng
the numb
e
r
of feature th
at need
s to be
sea
r
ched. Th
e feature d
e
scripto
r
s of the refere
n
c
e f
r
ame a
r
e
clu
s
tere
d by bui
lding KD-tre
e
.
A
feature i
s
th
en rand
omly sel
e
cte
d
to
be the
pa
rtition n
ode. Ea
ch fe
ature
e
x
tracted f
r
om
the
curre
n
t fram
e traverse
s t
he built KD-t
ree to fi
nd it
s corre
s
po
nd
ing point
with be
st-bin
-first
sea
r
ch. The
pair of matchi
ng feature
s
in
two di
fferent
image
s is call
ed a co
rresp
onde
nce. Wh
e
n
all co
rresp
o
n
den
ce
s have
been fou
n
d
,
the tr
ansfo
rmation m
a
trix can be
computed
usi
n
g
rand
om
sa
m
p
le a
nd
co
nsensus (RANSAC). In t
h
is
study, th
e f
o
cu
s
will
be
on
stabili
zati
on of
aerial i
m
age
s whe
r
e the
di
stan
ce of the
came
ra to t
he sce
ne i
s
much l
a
rger t
han
cha
nge
s in
scene
depth,
or if depth
va
riation i
s
la
rg
e, the ca
mera
motion is sm
all eno
ugh
so
that the fram
es
being
re
gist
ered
have
negligibl
e
lo
cal
geom
etri
c diffe
ren
c
e.
We
will
use
proje
c
tive
transfo
rmatio
n to registe
r
the image fra
m
es. As
sum
e
that the number of mismatche
s
can
be
redu
ce
d by u
s
ing
our im
proved ap
pro
a
c
h a
nd KD-tr
ee with b
e
st
-bin-first search, small a
m
o
unt
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5645 – 56
54
5650
of mism
atch
e
s
still o
c
cur,
and thi
s
m
a
y
lead to
un
reli
able
pre
d
ictio
n
of m
o
tion e
s
timation. T
h
us,
a furthe
r
ch
e
c
k of mat
c
hin
g
erro
rs is a
signifi
c
ant p
a
r
t of the
algo
rithm. We
use
the
well
kno
w
n
RANSAC al
g
o
rithm for th
e purp
o
se of both elimina
t
ion the mismatche
s
an
d
estimation o
f
th
e
para
m
eters
o
f
the proje
c
tive tran
sform
model
. Ba
se
d on th
e
corresp
ond
en
ce
s, RANSAC can
iteratively co
mpute the
p
a
ram
e
ters of
the motio
n
model
by det
ecting
the inl
i
ers an
d o
u
tliers
among
the
s
e
matching
pa
irs. T
r
a
d
ition
a
lly, t
he ge
o
m
etric tra
n
sf
ormatio
n
b
e
twee
n two im
age
s
can b
e
de
scri
bed by a hom
ogra
phy whi
c
h is a 3D mo
del.
The procedu
re to calculate
the homog
ra
phy can b
e
e
x
presse
d as f
o
llows:
(1)
We b
egin
with a sim
p
l
e
liner al
gorit
hm for dete
r
mining h
o
mo
grap
hy
H
by giving a
set of fo
ur
point
corre
s
pond
en
ce
s,
i
i
P
P
, whi
c
h are not three
points collinear. T
h
e
transfo
rmatio
n is given by equatio
n
i
i
HP
P
.
(2)
Com
pute
a simila
rity transfo
rmati
on
T
for the
points in th
e
refere
nce i
m
age,
con
s
i
s
ting of a transl
a
tion
and scali
ng, that take
s poi
nts
i
P
to a new
set of points
i
P
~
, the points
are tran
slate
d
and
scale
d
so t
hat the
ce
ntroid of th
e
points
i
P
~
is at t
he coordinate
origin
T
)
0
,
0
(
,
and thei
r average di
stan
ce
from the ori
g
i
n
is equ
al to
2
. So coo
r
dinat
es
i
P
in refere
n
c
e ima
ge
are repla
c
ed
by
i
i
TP
P
~
. Compu
t
e a similar tran
sform
a
tion
T
for the points in the cu
rre
nt
image, tra
n
sf
ormin
g
poi
nts
i
P
to
i
P
~
, that is
coo
r
din
a
tes
i
P
in cu
rrent im
age a
r
e
re
pla
c
ed
b
y
i
i
P
T
P
~
.
T
and
T
are
3
3
matrixes. Su
bstituting in the equ
ation
i
i
HP
P
, we de
riv
e
t
h
e
equatio
n
i
i
P
HT
T
P
~
~
1
. This rel
a
tion im
plies that
1
~
HT
T
H
is
the tran
sform
a
tion matrix f
o
r
the poi
nt correspon
den
ce
s
i
i
P
P
~
~
. Note th
at the e
quation
i
i
P
H
P
~
~
~
involves ho
mogen
eo
u
s
v
e
ct
or
s,
t
h
u
s
t
he 3-v
e
ct
o
r
s
i
P
~
and
i
P
H
~
~
are
n
o
t equal, th
e
y
have the
same di
re
ction
but ma
y
differ in m
a
g
n
itude by
a n
onzero
scal
e
factor.
T
he e
quation
may
be exp
r
e
s
sed
in term
s
of the
vector cro
s
s prod
uct a
s
0
~
~
~
i
i
P
H
P
. This form
will
enable a si
mple linea
r solution for
H
~
to
be
derived. If the j-th row of th
e matrix
H
~
is denoted by
T
~
j
H
, th
en we may write:
i
i
i
i
P
H
P
H
P
H
P
H
~
~
~
~
~
~
~
~
T
3
T
2
T
1
(16)
Writing
T
)
~
,
~
,
~
(
~
i
i
i
i
y
x
P
, the cross p
r
od
uct
may then be given explicitl
y
as:
i
i
i
i
i
i
i
i
i
i
i
i
i
i
y
x
x
y
P
H
P
H
P
H
P
H
P
H
P
H
P
H
P
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
T
1
T
2
T
3
T
1
T
2
T
3
(17)
Since
j
i
i
j
H
P
P
H
~
~
~
~
T
T
, for
3
,
2
,
1
j
, this give
s a
set
of three e
q
u
a
tions i
n
the
entrie
s
of
H
~
,
whic
h may be written in the form:
0
H
H
H
0
P
P
P
0
P
P
P
0
3
2
1
T
'
T
'
T
'
T
'
T
'
T
'
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
T
i
i
i
i
i
i
T
i
i
i
i
i
i
T
x
y
x
y
(18)
That is,
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Digital Im
age Stabilization
Based o
n
Impro
v
ed S
c
al
e
Invari
ant Feat
ure T
r
an
sform
(Xiaoran G
uo)
5651
0
8
7
6
5
4
3
2
1
0
~
~
~
~
~
~
~
~
~
0
0
0
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
0
0
0
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
~
0
0
0
h
h
h
h
h
h
h
h
h
x
y
x
x
x
y
y
y
x
y
x
y
x
x
x
y
x
y
y
y
x
y
y
x
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
i
(19
)
These equ
ations have the
form
0
h
A
~
i
, where
i
A
is a
9
3
matrix, and
h
~
is a 9-
vector ma
de
up of the entri
es of the matrix
H
~
, with
i
h
~
the i-th element of
h
~
.
T
3
2
1
~
~
~
~
H
H
H
h
,
8
7
6
5
4
3
2
1
0
~
~
~
~
~
~
~
~
~
~
h
h
h
h
h
h
h
h
h
H
(20)
Although the
r
e are th
ree
e
quation
s
in (18),
only two
of them are
linearly in
dep
ende
nt
(sin
ce the third row is o
b
tai
ned,
up to scale, from the sum of
i
x
~
times the first row and
i
y
~
times
the second
).
Each
poi
nt correspon
den
ce give
s
ri
se t
o
two
ind
epe
ndent
equ
atio
ns i
n
th
e e
n
tries
of
H
~
. Given a set of four
su
ch p
o
int co
rres
p
ond
en
ce
s, we obtain
a
set of equ
ations
0
h
A
~
,
whe
r
e
A
is the
matrix of eq
uation
coeffi
cients b
u
ilt fro
m
the matrix
rows
i
A
cont
rib
u
ted from
each co
rresp
onde
nce, an
d
h
~
is the vector of un
kn
own e
n
trie
s of
H
~
. Executing the SVD
(Singul
ar Val
ue
De
comp
o
s
ition) of
A
, the unit
sing
ul
ar ve
ctor correspon
ding
to the
small
e
st
sing
ular valu
e is the sol
u
tion
h
~
, and the matrix
H
~
is determin
ed
from
h
~
. Then
we ca
n
cal
c
ulate h
o
m
ography
H
by
1
~
HT
T
H
.
(3)
We
cal
c
ulate the e
r
ror of a
corresp
ond
en
ce
from hom
og
raphy
H
using
the
s
y
mmetric
trans
f
er
erro
r, that is
2
2
1
-
2
)
(
)
(
i
i
i
i
transfer
d
d
d
HP
P
P
H
P
,
,
. If the
2
transfer
d
is less than
the given th
reshold
valu
e, then
confi
r
ming th
e
correspon
den
ce inlie
rs, oth
e
rwi
s
e
outlie
rs.
Rep
eat fo
r a
numbe
r
of sa
mples,
and
reco
rd
t
he l
a
rgest
num
ber
of inliers. Fi
n
a
lly, re-estim
ate
R
H
from
all
co
rrespon
den
ce
s
cla
ssifie
d
as i
n
lie
rs,
an
d u
s
ing
the
final ho
mog
r
aphy
R
H
to
impleme
n
t image compe
n
sate.
4. Results a
nd Analy
s
is
To test the
stabilization abi
lity of the pro
posed alg
o
rit
h
m and
com
pare it
with traditional
SIFT algorith
m
, some ima
ge pairs with
different ch
an
ges a
r
e u
s
ed.
(a) T
he refe
re
nce ima
g
e
(b) T
he curre
n
t image
Figure 1. Input Images (a) Refere
nce Image an
d (b
) Current Imag
e
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5645 – 56
54
5652
Figure 1 i
s
the inp
u
t ima
ges,
with sca
ling an
d rotat
i
onal vari
atio
ns. Figu
re
1(a) is th
e
referen
c
e im
age, an
d Fig
u
re
1(b
)
i
s
th
e cu
rrent im
a
ge. As
we
ca
n se
e, Figu
re
1(b
)
takes
great
scaling a
nd rotational cha
nge compa
r
e
d
to Figure 1
(
a).
Figure 2 sho
w
s the mat
c
h
i
ng points fro
m
SI
FT algori
t
hm and our
algorith
m
. As we can
see, SIFT ge
nerate
s
too many
feature
points and
matche
s. At
the same tim
e
, our algo
rithm
gene
rate
s le
ss featu
r
e
p
o
ints
and
m
a
tche
s
but
with hi
gh
co
rre
ctne
ss, which
ma
ke
s
the
sub
s
e
que
nt calcul
ation of t
r
an
sform
very promi
s
ing.
This allows us
to
gain better
accuracy
t
han
SIFT.
(a) SIFT ap
proa
ch
(b) Propo
se
d approa
ch
Figure 2. Matchin
g
Re
sult
s from (a) SIF
T
and (b
) the
Propo
se
d Approa
ch
Figure 3
(
a
)
shows th
e
co
mpen
sated
i
m
age
usi
ng
SIFT, and
Fi
gure
3
(
b)
sh
ows the
comp
en
sated
image usi
ng
our alg
o
rithm.
(a) SIFT ap
proa
ch
(b) Propo
se
d approa
ch
Figure 3. Stabilizatio
n Images fro
m
(a
) SIFT and (b
) the Propo
se
d
Approa
ch
We al
so test
the stabilization ability
of
the two
algorithm
s usi
ng the im
ages
with
illumination
chang
es, viewpoint ch
ang
e
s
, and blu
r
.
Nume
ri
cal ev
aluation
of the qu
ality of the
imag
e
stabili
zation i
s
fulfilled u
s
i
ng pe
ak
sign
al-to
-
noi
se ratio
(PS
N
R) as
e
rro
r
measure. PSNR bet
wee
n
frame
t
and
frame
1
t
is
defined a
s
:
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Digital Im
age Stabilization
Based o
n
Impro
v
ed S
c
al
e
Invari
ant Feat
ure T
r
an
sform
(Xiaoran G
uo)
5653
M
y
N
x
t
t
y
x
I
y
x
I
MN
t
MSE
11
2
1
)]
,
(
)
,
(
[
1
)
(
(21)
)
(
log
10
)
(
2
10
t
MSE
I
t
PSNR
MAX
(22)
W
h
er
e
)
(
t
MSE
is t
he m
ean
-sq
uare
e
rro
r
b
e
twee
n fra
m
es,
MAX
I
is the
maximum
intensity valu
e of a pixel,
and
M
and
N
are th
e fram
e dimen
s
io
ns.
PSNR
measure
s
the
simila
rity betwee
n
two
im
age
s, hen
ce,
it is u
s
eful
to
evaluate h
o
w much a im
ag
e is
stabili
ze
d
by
the algorith
m
by simply evaluating t
he si
milarity of the two image
s.
In order to prove the stabilization effe
ct of the algorithm, we
calcul
ated the
PSNR
betwe
en ea
ch frame of th
e su
ccessive
50 frame
s
u
n
stabl
e video
sequ
en
ce
s, the expe
rime
ntal
results
are
sh
own i
n
Fig
u
re
4. The
cu
rve
at the botto
m rep
r
e
s
e
n
t the
PSNR
of the ori
g
inal n
on-
comp
en
sated
video sequ
e
n
ce
s, the mid
d
le one is the
PSNR
of compen
sated video se
quen
ce
s
usin
g SIFT
method,
and
the ab
ove o
n
e
is the
PSNR
of compen
sate
d video seq
uen
ce
s
u
s
in
g
the prop
osed
approa
ch.
Figure 4. Experime
n
tal Re
sults of PSNR
At the same ti
me, we u
s
e t
he execution
time to
evalu
a
te the efficie
n
cy of the alg
o
rithm
s
.
In Table
1,
executio
n time comp
ari
s
ons amo
ng
variou
s vari
a
t
ions
betwe
e
n
SIFT an
d
our
approa
ch a
r
e
sho
w
n. As
can be
see
n
from Figu
re 4
and Ta
ble 1,
the
PSNR
of our a
p
proa
ch i
s
a little bigge
r
than SIFT un
der th
e vari
o
u
s va
riation
s
,
and th
e exe
c
ution time of
our
app
roa
c
h
is
much le
ss than SIFT, so o
u
r app
ro
ach outperfo
rm
s the SIFT approach.
Table 1. Execution Time Compa
r
ison
s
variation
execution time of
SIFT
execution time of
our app
roach
rotations and scale
6203.7ms
3394.9ms
illumination
2385.6ms
171
1.0ms
view
point
4153.5ms
2895.6ms
blur 3716.0ms
2868.1ms
5. Conclusio
n
In con
c
lu
sion
, we pro
p
o
s
e a novel ro
bust lo
cal fe
ature p
o
int e
x
traction a
p
p
r
oa
ch in
image sta
b
ili
zation. First, we use SIFT algorithm
de
tect the feature poi
nt of the image, then
adopt Harris
crite
r
ion to select the mo
st rob
u
st fea
t
ure point
s. The Harris
Threhold
is self
-
adaptive
and
ca
n yield
rel
a
tively stable
dete
c
tion p
e
r
forma
n
ce. E
x
perime
n
tal result
s
sho
w
t
he
prop
osed scheme yield
s
better perfo
rman
ce
s tha
n
SIFT, it h
a
s a little better stabili
zation
accuracy, an
d save
s a lot of computatio
n burd
en.
0
5
10
15
20
25
30
35
40
45
50
11
12
13
14
15
16
17
18
19
F
r
am
es
PS
N
R
(
d
B)
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 7, July 201
4: 5645 – 56
54
5654
Referen
ces
[1]
Xu
LD, Li
n
XG. Digital im
ag
e
stabiliz
atio
n Based o
n
Circ
u
l
a
r Block Matc
hin
g
.
IEEE Transactions on
Cons
u
m
er Ele
c
tronics
. 200
6; 52: 566-
57
4.
[2]
Sato K. Contr
o
l tech
niq
ues
for optica
l
ima
ge stab
iliz
in
g
s
y
stem.
IEEE
Transactions on Consum
er
Electron
ics.
19
93; 39: 46
1-46
6.
[3]
Vell
a F
,
Castor
ina A, M
ancus
o M. Dig
ital
im
age sta
b
i
lizati
o
n
b
y
ada
ptive block
m
o
tion
v
e
ctors filteri
ng.
IEEE Transactions on Cons
um
er Electronic
. 2002; 4
8
: 796-
800.
[4]
Uomori K. Aut
o
matic ima
ge
stabiliz
in
g s
y
stem b
y
full-
dig
i
tal sig
n
a
l
proc
e
ssing.
IEEE Transactions o
n
Cons
u
m
er Ele
c
tronics
. 199
0; 36: 510-
51
9.
[5]
Enge
lsber
g A, Schmidt G. A compar
ative rev
i
e
w
of
di
gita
l i
m
age stab
ilis
in
g alg
o
rithms fo
r mobil
e
vide
o
communications.
IEEE
Transactions on Cons
umer Electronics.
1999; 45: 5
92-5
97.
[6]
Ko SJ, L
ee S
H
, Le
e KH. D
i
g
ital
ima
ge st
abil
i
zi
ng
alg
o
ri
thms bas
ed
o
n
bit-p
l
a
ne m
a
tching.
IEEE
T
r
ansactio
n
s o
n
Cons
u
m
er El
ectronics.
19
98
; 44: 617-6
22.
[7]
Ko SJ, Lee SH
, Jeon SW
. Fa
st digital im
ag
e
stabiliz
er bas
e
d
on gra
y
-c
od
e
d
bit-pl
an
e matchin
g
.
IEEE
T
r
ansactio
n
s o
n
Cons
u
m
er El
ectronics.
19
99
; 45: 598-6
03.
[8]
Cha
ng J
Y
, Hu
W
F
, Cheng
M
H
. Di
gital
ima
g
e
trans
lati
on
a
nd r
o
tatio
n
mot
i
on
stabi
liz
atio
n us
ing
o
p
tica
l
fl
o
w
te
ch
ni
qu
e
.
IEEE Transactions on Consum
er Electronic
s.
2002; 48: 10
8-11
5.
[9]
Erturk S. Digita
l
imag
e stab
iliz
ation
w
i
th su
b-i
m
age
phas
e c
o
rrelati
on
bas
e
d
gl
oba
l moti
o
n
estimati
on,
IEEE Transactions on Cons
um
er Electronic
s.
2003; 49: 13
20-1
325.
[10]
Paik JK, Park
YC, Kim DW
.
An a
daptiv
e m
o
tion
dec
isio
n
s
y
stem for
di
gi
tal ima
ge st
abi
l
i
zer b
a
se
d o
n
edg
e pattern m
a
tchin
g
.
IEEE
Transactions on Cons
um
er Electronics
. 19
92
; 38: 607-6
16.
[11]
Harris C, Step
hens M.
A co
mb
in
ed corn
er
and e
dge
det
ection.
Proc. 4
t
h Alve
y
V
i
sio
n
Confer
enc
e.
198
8; 189-
192.
[12]
Z
hu JJ, Guo BL. Electronic
image sta
b
ili
zation s
y
stem
based
on gl
oba
l feature tr
ackin
g
.
IEEE
T
r
ansactio
n
s o
n
Cons
u
m
er El
ectronics
. 20
08
; 19: 228-2
33.
[13]
Lo
w
e
DG. Di
stinctive Imag
e F
eat
ures fr
om Scal
e-Inva
riant Ke
y P
o
i
n
ts.
Internatio
nal J
ourn
a
l
of
Co
mp
uter Visi
on.
200
4; 60: 9
1
-11
0
.
[14]
Yang J
L
, Scho
nfeld D, Mo
ha
m
ed M. Rob
u
s
t
Video Stab
iliz
ation B
a
sed
o
n
Particle Filt
e
r
T
r
acking of
Projecte
d C
a
m
e
ra Moti
on
. IE
EE T
r
ansacti
o
n
s o
n
Circ
u
its
and
Syste
m
s for Vid
eo T
e
ch
nol
ogy
. 20
09
;
19: 945-
95
4.
[15]
Can Su
n, Jin-g
e
W
ang, Z
a
ix
in
Liu, Junmin L
i
. W
e
ighted Mult
i-Scale Imag
e Matchin
g
Base
d on Harris-
SIFT
Descriptor.
T
E
LKOMNIKA Indon
esia
n
Journa
l of Electrical En
gin
e
e
rin
g
. 201
3; 11(10): 59
65-
597
2.
[16]
Quan Su
n, Jia
n
xun Z
h
an
g.
Parall
el
Rese
arc
h
an
d Impl
eme
n
tation
of SAR
Image R
egistr
a
tion B
a
se
d
on Optimiz
ed
SIFT
.
T
E
LKOMNIKA Indo
ne
sian J
our
nal
of
Electrica
l
En
g
i
ne
erin
g
. 20
14;
12(2):
112
5-
113
1.
[17]
Yan K, Sukth
ankar
R.
Pca-sift: A m
o
re
distinctive repres
entation for local image descriptors
. In
Proceedings of
the IEEE Com
puter
Soc
i
et
y
Confer
ence on Computer
Vision and P
a
tte
rn Recognition.
200
4; 506-
513.
[18]
Mikola
jcz
y
k K.
A p
e
rforma
nc
e ev
alu
a
tio
n
o
f
loca
l
descri
p
tors
. Journ
a
l
of the
Ch
ines
e Institute
of
Engi
neers
. 2
0
0
5
; 27: 161
5-16
30.
[19]
Ba
y
H.
Surf: Spee
de
d up robust features.
Computer Vis
i
on a
nd Image
Understa
ndi
n
g
. 2008; 11
0:
161
5-16
30.
[20]
T
u
y
t
el
aars T
.
Loca
l
invar
i
ant
feature detec
tors: A surve
y
.
F
oundatio
ns
and T
r
ends i
n
Co
mp
uter
Graphics a
nd
Visio
n
.
200
8; 3: 177-20
8.
Evaluation Warning : The document was created with Spire.PDF for Python.