TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol.12, No.4, April 201
4, pp. 3105 ~ 3
1
1
7
DOI: http://dx.doi.org/10.11591/telkomni
ka.v12i4.4781
3105
Re
cei
v
ed Au
gust 28, 20
13
; Revi
sed
No
vem
ber 1
3
, 2013; Accepte
d
De
cem
ber
6, 2013
Adaptive Wallis Filter via Sparse Recognition for
Automatic Control Points Extraction
Leilei Geng*
1
, Deshen Xia
1
, Quansen S
u
n
1
, Kai Yuan
2
1
Nanji
ng U
n
ive
r
sit
y
of Scie
nce
and T
e
chnol
o
g
y
, N
anj
ing, C
h
in
a
2
Head
quarters
of Jinan Mi
litar
y Are
a
Comma
nd, Jina
n, Chi
n
a
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: leile
ig
eng
@g
mail.com
A
b
st
r
a
ct
With the r
api
d
deve
l
op
ment
o
f
the re
mote se
nsin
g
sate
llite,
the si
z
e
a
nd t
h
e res
o
luti
on
of
satellit
e
imag
es grow
increas
ing
l
y. T
he eva
l
u
a
tion
of remote
sens
ing i
m
age
qua
l
i
ty requir
e
s pre
c
ise infor
m
atio
n of
control p
o
i
n
ts extracted fro
m
uneva
l
u
a
ted i
m
a
ges
a
nd re
ference i
m
ag
e
s
. T
herefore, w
e
propos
e a
n
ada
ptive W
a
l
l
i
s
filter meth
od
based
on sp
a
r
se recog
n
it
i
o
n
to increas
e th
e nu
mb
er of control p
o
i
n
ts an
d
improve
the
matchin
g
pr
ecisi
on. F
i
rstly, feat
ure v
e
ct
ors of
i
m
a
ges
are
con
s
tructed by
co
mp
utin
g the
i
m
ag
e
radi
ation-
par
a
m
eters. S
e
con
d
ly, the cl
assif
i
catio
n
of
su
b-regi
on terra
in i
n
the i
m
age
c
an b
e
d
e
ter
m
i
n
e
d
usin
g spars
e
r
e
cog
n
itio
n. F
i
n
a
lly, acc
o
rdi
ng
to specif
ic typ
e
of sub-re
gio
n
terrain, w
e
en
h
ance th
e re
gio
n
s
by the W
a
ll
is fil
t
er base
d
on c
o
rresp
ond
in
g filter para
m
eters
and extract co
ntrol p
o
ints w
h
i
c
h w
ould l
e
a
d
t
o
the a
u
to
matic
eval
uatio
n for
geo
metric pr
ec
ision.
T
h
e
exp
e
r
iments s
how
t
hat the
pr
op
os
ed
metho
d
ca
n
get
better res
u
lts e
s
peci
a
lly
in th
e
detai
l o
n
th
e i
m
a
ges
of
Res
o
urse-3 s
a
tel
lite,
henc
e c
an i
n
cr
ease t
he
nu
mb
er
and i
m
prove a
ccuracy of cont
rol po
ints.
Ke
y
w
ords
:
sp
arse reco
gniti
o
n
, radiati
on-
par
ameters, ada
pt
ive Wallis e
n
h
a
n
ce
me
nt, extract control poi
nts
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
Duri
ng
rem
o
te sensi
ng im
age p
r
o
c
e
s
si
ng, ther
e i
s
geomet
ric
de
formation
rel
a
tive to
actual targets due to the chang
e of image proj
ecti
o
n
pattern, the alteration of sen
s
o
r
exteri
or
orientatio
n, u
neven sen
s
o
r
media
and
terrain
, rota
tion and
curvature of ea
rth, and so
on.
Therefore,
re
mote se
nsi
n
g image
ne
eds to b
e
g
eometri
cally corre
c
ted be
fore
ap
plication.
Ho
wever, a
c
curacy of g
eometri
c correctio
n
is
often limited, whi
c
h lead
s to existence of
unpredi
cted resid
ual defo
r
mation in the
remote se
n
s
ing image aft
e
r sy
stematic corre
c
tion. In
fact, it is ne
ce
ssary to
carry
out
geo
metric
a
c
curacy eval
uatio
n for ge
omet
rically
correct
e
d
image to
gui
ding g
eom
etric p
r
e
c
isi
on
corre
c
tion
[1, 2]. Geom
etric a
c
curacy
evaluation
m
u
st
depe
nd o
n
g
r
oun
d
control
points (GCP
s)
whi
c
h
are
usu
a
lly marked bet
wee
n
remote
sen
s
in
g
image a
nd re
feren
c
e ima
g
e
by engine
er. The accu
ra
cy of GCPs rel
i
es on
engin
e
e
r‘s
kn
owl
edg
e
and skill, so
that GCPs colle
ction is
time-c
on
sumi
ng and a
c
cu
racy of GCP
s
is lo
w. The
traditional m
e
thod se
rio
u
s
ly affects the effici
en
cy of geometric accu
ra
cy evaluation [3].
Therefore,
ho
w to in
crea
se
the num
be
r
and to im
pro
v
e accu
ra
cy of GCP
s
i
s
a
key p
r
obl
em
of
remote
sen
s
i
ng image eva
l
uation.
In recent years, ma
ny re
sea
r
che
s
of aut
omatic
extraction
of GCPs h
ad be
en don
e,
whi
c
h ca
n be
divided into two cla
s
se
s: pixel-
ba
sed m
e
thod
s and feature
-
b
a
sed
methods. In the
first metho
d
, they match
G
C
Ps th
rou
gh
comp
uting
co
rrel
a
tion b
a
se
d on the
gray
of pixels, whi
c
h
is ea
sy to im
plement. But the first cl
ass need
s la
rge
amount of
ca
lculatio
n to g
e
t the matchi
ng
result and
is
easily influ
e
n
c
ed
by light
and di
st
ortio
n
of remote
sensi
ng ima
g
e
.
In the se
co
nd
method
s, the
y
match
GCP
s
throug
h
si
milarity
mea
s
ure
ba
sed
on
feature
poi
nts extra
c
ted
from
image,
whi
c
h
ope
rate
s
si
mply and
ma
tche
s fa
st wi
t
h
hig
h
a
c
curacy. Recently, the re
se
arches
focu
s on Harris op
erator [
4
], Forstne
r
o
perato
r
[5], scale inva
riant
feature
s
tran
sform al
gorit
hm
[6] and
spe
e
d
ed up
robu
st
feature
s
(Surf
)
alg
o
rithm
[7, 8]. Wh
en ev
aluating
geo
metric a
c
cura
cy,
we
sh
ould
u
s
e multi-so
urce refere
nce i
m
age
whi
c
h
i
s
a
mo
sai
c
i
m
age
co
mpo
s
ed
of diffe
re
nt
spe
c
tra,
pha
se, resolution
and
se
nsor
due to th
e m
ap of
remote
se
nsin
g ima
ge g
r
o
w
ing [
9
].
Becau
s
e th
ere is a
great d
i
fference of g
r
ay in mult
i
-
source
referen
c
e ima
ge, it i
s
impo
ssible
to
extract an
d
match
GCP
s
usin
g pixel-b
a
se
d metho
d
and it is
difficulty to do th
at usin
g feat
ur
e-
based m
e
tho
d
. Therefore, this p
ape
r p
r
esents a
n
e
w
meth
od of
two level
m
a
tchin
g
st
rat
egy
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 3105 – 3
117
3106
based on gl
o
bal optimization to extract
and matc
h G
C
Ps fro
m
re
mote sen
s
in
g
image and
multi-
sou
r
ce refe
re
nce ima
ge.
We u
s
ually carry out imag
e pro
c
e
ssi
ng
before
extracting GCPs in
orde
r to incre
a
se the
numbe
r a
nd t
o
improve ma
tching
accu
ra
cy of G
C
Ps.
Re
cently the
method
s of i
m
age
pro
c
e
s
sin
g
are
divided
in
to two
cl
asse
s. O
ne i
s
hist
ogra
m
-b
ased
enh
an
ceme
n
t
algo
rithms.
For
exampl
e, in
orde
r to eli
m
inate gray di
stortion, Xia
o
c
hu
n Li
u ta
kes hi
stog
ram
simila
rity tra
n
sformation f
o
r
remote
se
nsing imag
e a
nd the
reference im
ag
e
[10]. The
other o
ne i
s
texture-b
a
sed
enha
ncement
algo
rithm
s
.
For
example,
Li Z
hang
en
han
ce
s
textures of
ima
ge usin
g Walli
s
f
ilter
to increa
se
th
e num
ber an
d to imp
r
ove
matchin
g
a
ccura
cy of feat
ure
point
s [1
1]. With the
rapid
developm
ent of remote sensing sa
tellite, the resolution of remo
te sensi
ng image becom
es more
and m
o
re a
c
curate. T
he
re
solutio
n
of
CCD came
ra i
s
30 m
e
ters
an
d re
sol
u
tion
o
f
hyper-spe
ctral
came
ra
is 10
0 mete
rs in
Environme
n
t
and
Di
sa
st
er
Monitori
ng
sa
tellite (HJ) la
unched
in
20
08.
Then the resolution of CCD ca
mera is
decrea
s
e
d
to 2.1 meters a
nd re
solutio
n
of multi-sp
ect
r
al
came
ra i
s
re
duced to 5
meters in
Re
sou
r
se-3 satellite (ZY-3) l
aun
che
d
in 2
012 [12]. We
can
con
c
lu
de th
at the
regi
onal
area
i
s
1
4
tim
e
s
bigg
er
in
HJ than
that i
n
ZY-3
whe
n
the two
ima
g
e
s
are in
the
sa
me si
ze. So t
here
are vari
ous textu
r
e
s
i
n
HJ ima
ge a
nd the
r
e a
r
e
simple f
eatures in
ZY-3 im
age.
Traditio
nal
Wallis filter with
glob
al p
a
ra
meters i
s
suit
able fo
r lo
w
resol
u
tion im
a
ge
su
ch as
HJ i
m
age. But when it is used
for high resol
u
tion image, su
ch as ZY
-3
image, there
will
be ma
ny pixels
with
satu
rated
gray. T
o
solve
this
probl
em, this pape
r
pre
s
e
n
ts a
n
ad
apti
v
e
Walli
s filter
b
a
se
d o
n
sub
-
regio
n
: recog
n
ize
sub-re
gi
on u
s
in
g
spa
r
se
recognitio
n
alg
o
rithm
a
n
d
enha
nce su
b-regio
n
usi
ng
Walli
s filter wi
th local pa
ra
meters.
Re
cently, the
r
e a
r
e m
any
traditional
me
thods fo
r
re
mote sen
s
ing
image
cla
s
si
fication,
su
ch a
s
prin
cipal com
pon
e
n
t analysi
s
(PCA) [
13] an
d indep
ende
nt compo
nen
ts analysi
s
(I
CA)
[14]. Recentl
y
some
ne
w mea
s
ures
were involv
e
d
,
su
ch as artificial neu
ral
network
[
15],
deci
s
io
n tre
e
[16], suppo
rt vector ma
chine [1
7]
an
d expe
rt
syst
em [18], a
n
d
so
on.
A lot
of
spatial i
n
form
ation in
high
re
solutio
n
remote
sen
s
in
g imag
e will
be waste
d
, if we
use
sing
le
traditional
cl
assificatio
n
method
s. Th
at’s
be
cau
s
e
we do
not
make
goo
d
use of
con
t
ext
informatio
n, shape info
rma
t
ion and feature
s
of tar
get
, which can b
e
got form sp
atial informati
on
of high re
sol
u
tion remote
sen
s
ing im
age. The
r
efo
r
e, we con
s
truct recogniti
on vector
wi
th
radiatio
n-parameters in
clu
d
ing
statistical inform
ation
of stru
cture
and g
r
ay, texture an
d gray in
orde
r to do b
e
tter wo
rk fo
r high re
sol
u
tio
n
remote
sen
s
ing ima
ge.
This pa
pe
r u
s
e sparse
re
cog
n
ition alg
o
rithm with
better cla
s
sification p
r
e
c
isi
on and
robu
stne
ss fo
r hig
h
re
sol
u
tion remote
sensi
ng ima
g
e
.
This al
go
rithm is
ba
sed
on comp
re
ssed
sen
s
in
g (CS)
prop
osed by
Allen Y. Yang
and Yi
M
a
[1
9]. If each te
st sampl
e
can
be represent
ed
by
spa
r
se polynomial, we can
re
cogni
ze su
b
-
region te
rrain
,
whic
h co
mputes
sparse
rep
r
e
s
entatio
n of training
sample a
ccordin
g to
test sam
p
les
and cl
assifie
s
test sa
mpl
e
s.
Assu
ming th
at all sampl
e
s bel
ong t
o
the sa
me
low dime
n
s
ion
spa
c
e
and the lin
e
a
r
rep
r
e
s
entatio
n of ea
ch te
st sampl
e
can
be g
o
t acco
rding to traini
ng sampl
e
s
whi
c
h a
r
e i
n
the
same
cla
s
s
with test
sam
p
le, pape
r [2
1] comp
ut
es
spa
r
se repre
s
entatio
n of trainin
g
sampl
e
s
and re
co
gni
zes tho
s
e sa
mples u
s
in
g spa
r
se re
cog
n
ition algo
rith
m. Allen Y.
Yang and Yi
Ma
solve face re
cog
n
ition [21
]
and Fei Yin recogni
ze
s remote sen
s
ing ima
ge target [22] using
spa
r
se re
cog
n
ition algo
rith
m. Therefo
r
e,
we cla
ssify sub-regi
on terrain of remote
sen
s
ing ima
ge
usin
g sp
arse
recognitio
n
al
gorithm to get
better re
cog
n
ition rate.
This p
ape
r p
r
esents
a ne
w metho
d
to
increa
se the
numbe
r a
n
d
to improve
matchin
g
accuracy of GCP
s
for hig
h
resolution remote se
nsi
n
g image that there i
s
simpl
e
x feature in sub
-
regio
n
. The a
l
gorithm i
s
as follows: 1
)
Divide image i
n
to numbe
rs of sub
-re
gion
s whi
c
h a
r
e i
n
the same
size; 2) Con
s
tru
c
t re
cognitio
n
vector
with radiation
-
pa
ra
meters co
mp
uted from su
b-
regio
n
; 3
)
Reco
gni
ze
su
b
-re
gion
u
s
ing
sp
arse
reco
gnition
algo
ri
thm; 4) Enh
ance
sub
-
reg
i
on
usin
g a
daptiv
e Wallis filter with l
o
cal p
a
ram
e
te
rs. T
he exp
e
rim
e
nts
sho
w
th
a
t
com
pared
with
existing featu
r
e extra
c
tion
and re
co
gniti
on method,
o
u
r metho
d
ca
n get better result
s. Adaptive
Walli
s filter
eliminates the number
of pixels with
saturated gray, whi
c
h traditional
Wallis filter
can
not do for high re
sol
u
tion remote
sensi
ng image
s
.
In this
way, we c
an effec
t
ively extrac
t
more
an
d hi
g
her a
ccu
rate
GCP
s
to
achi
eve the
evalu
a
tion of
geo
metric p
r
e
c
isi
on a
u
tomatically
and a
c
curatel
y
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Adaptive
Wall
is Filter Via Sparse Recog
n
iti
on for Autom
a
tic Control Points… (L
eilei Gen
g
)
3107
2. T
w
o Lev
e
l
s
Matching Based on Wallis Enhancement
There are so
me probl
em
s about geom
etric a
c
cu
ra
cy evaluation with the map of remote
sen
s
in
g ima
ge g
r
o
w
ing.
Whe
n
evalu
a
t
ing geo
metri
c
a
c
curacy,
we
sho
u
ld u
s
e m
u
lti-sou
r
ce
referen
c
e im
a
ge, which i
s
a
mo
saic ima
g
e
compo
s
e
d
of differe
nt sp
ectra,
ph
ase,
resolution
an
d
s
e
ns
or
du
e to
th
e
ma
p o
f
r
e
mo
te se
ns
in
g
imag
e g
r
ow
in
g
.
T
h
er
e
is
a gr
e
a
t
d
i
ffe
r
e
nc
e o
f
g
r
ay in
multi-source
referen
c
e i
m
age, a
s
sh
o
w
n i
n
Fi
gure
1(b),
so
it is difficult to
e
x
tract a
nd
m
a
tch
GCP
s
from re
mote sen
s
in
g
image and m
u
lti-so
urce referen
c
e im
age
using exi
s
tin
g
method
s.
(a)
Remote S
ensi
ng Image
(b) Multi
-
s
our
ce R
e
fere
nc
e
-
imag
e
Figure 1. Re
mote Sensi
n
g
Image and Its Refe
ren
c
e
-
i
m
age
Paper [9] prese
n
ts t
w
o l
e
vels m
a
tchi
ng b
a
se
d o
n
tradition
al
Walli
s en
han
ceme
nt to
extract G
C
Ps. The main al
gorithm i
s
followin
g
:
(1)
Take th
e first
level matchi
ng: First, take down sa
m
p
ling for rem
o
te sen
s
in
g image
and m
u
lti-so
urce
referen
c
e im
age
wi
th the
same
paramete
r
s.
Seco
nd, ex
tract a
nd m
a
tch
origin
al G
C
P
s
u
s
ing S
u
rf
algorith
m
, wh
ich h
a
s
many
advantag
es,
su
ch
as
scal
ing an
d rotation
invarian
ce, a
n
ti-illuminatio
n cha
nge
an
d anti-vie
w
po
int transfo
rm
ation, and
so
on. Third, g
e
t
rid
of wro
ng m
a
tchin
g
G
C
Ps t
houg
h e
s
tima
tion epip
o
lar
geomet
ry co
nstrai
nt [19, 20] incl
uding
M-
Estimation
al
gorithm
[21
-
2
3
] and
rand
o
m
sample
co
nse
n
sus alg
o
r
ithm
(RA
N
S
A
C) [2
4-27].
Su
r
f
algorith
m
in
cl
ude
s fou
r
ste
p
s: extract fe
ature
poi
nt
s,
determi
ne th
eir m
a
in
dire
ction, g
ene
ra
te
their de
script
ion and mat
c
h the
s
e poi
nts. Whe
n
image contai
ns a lot of similar structu
r
es,
feature
point
s extracted from th
ese structures
will easily mism
atch
due to simil
a
rly l
o
cal
neigh
borhoo
d
information
containe
d in
d
e
scriptio
n of feature p
o
ints.
(2)
Comp
en
sate
geom
etri
cal
inform
ation
of remote
se
nsi
ng i
m
age,
whi
c
h are
gene
rated f
r
o
m
origi
nal G
C
Ps
extracte
d in the
first
level matchi
n
g
. We d
e
scri
be ge
ometri
cal
relation
between remote
sen
s
in
g imag
e and m
u
lti-
reso
urce reference
imag
e usin
g
first
-
orde
r
polynomial, a
s
sh
own in Equation (1) a
nd Equation
(2):
12
3
4
X
=a
a
x
a
y
ax
y
(
1
)
12
3
4
Y=
b
b
x
b
y
bx
y
(
2
)
Whe
r
e,
(x
,y)
is the featu
r
e
point of rem
o
te se
nsi
ng i
m
age;
(X,Y)
i
s
the featu
r
e
point of mult
i-
res
o
u
r
c
e
refe
ren
c
e ima
ge.
(3)
Enhan
ce the
textures of th
e above imag
es by traditio
nal Walli
s filter.
(4)
Take
the
se
cond l
e
vel mat
c
hin
g
: Extract
and
match
G
C
Ps usi
ng S
u
rf algo
rithm
a
n
d
then get rid of
wron
g match
i
ng GCP
s
tho
ugh e
s
timatio
n
epipol
ar ge
ometry co
nst
r
aint.
Take th
e su
b-regio
n
ima
ge of HJ for ex
ample, which its
si
ze
is 400
×4
00
and its
resolution
is 30 m
e
ters. The
su
b-re
gion
cove
rs
12km
×
1
2
km
area
an
d co
ntains
abu
nd
ant
terrai
n
s an
d
compli
cate
d t
e
xtures.
The
experim
ents
sho
w
th
at th
e sub-re
gion
of low preci
s
i
o
n
can
be
well
e
nhan
ce
d by t
r
adition
al
Wal
lis filter,
a
s
shown in
Figu
re 2.
With the
same
si
ze, th
e
area
of su
b-region too
k
b
y
image is 1
4
times bi
gg
er than th
at took by ZY-3
, which cove
rs
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 3105 – 3
1
17
3108
820m
×82
0
m area
an
d
the textures and terrai
n
s are
si
mplex. The t
r
adition
al
Wal
lis filter
can
n
o
t
enha
nce texture
s
of hig
h
resolution i
m
age
well d
u
e
to contai
nin
g
many pixel
s
with
satu
ra
ted
gray, a
s
sh
o
w
n i
n
Fi
gure
3. The
r
efo
r
e,
we
propo
se
a
ne
w m
e
thod
of two
levels
matchin
g
ba
sed
on ada
ptive Walli
s enh
an
ceme
nt.
(a) HJ
Ima
g
e
(b) Walli
s
Enhan
ceme
nt
Figure 2. Co
mpari
s
o
n
of Walli
s Enhan
ceme
nt on Lo
w Pixel Image
(a) ZY
-3 Ima
g
e
(b) Walli
s
Enhan
ceme
nt
Figure 3. Co
mpari
s
o
n
of Walli
s Enhan
ceme
nt on Hi
gh Pixel Image
3. Adaptiv
e
Wallis E
nhancement based on Ra
diation-parameters
Whe
n
we en
han
ce textures of hig
h
re
solutio
n
rem
o
te se
nsin
g i
m
age u
s
in
g tradition
al
Walli
s filter, t
here
are ma
n
y
pixels
with
saturat
ed
gra
y
so that th
e
numbe
r of
G
C
Ps i
s
le
ss a
n
d
the accuracy
of those is low. So we propo
se
a ne
w method of two levels m
a
tching b
a
se
d
on
adaptive Wal
lis enha
ncem
ent to incre
a
s
e the num
b
e
r and to im
prove a
c
cura
cy of GCPs, as
sho
w
n in Fig
u
re 4. The m
a
in step
s are followin
g
:
(1)
Tak
e
the firs
t level mat
c
hing to ex
tract an
d mat
c
h o
r
igin
al
GCP
s
after
down
sampli
ng
of remote
se
nsi
ng ima
ge a
n
d
multi-re
sou
r
ce
refe
re
nce imag
e. Th
en comp
en
sate
geomet
rical informatio
n of remote
sen
s
i
ng im
age, wh
ich are gen
erated from o
r
iginal G
C
Ps.
(2)
Con
s
tru
c
t re
cognition ve
ctor of su
b-regi
on with ra
diat
ion-p
a
ramete
rs.
(3)
Re
cog
n
ize su
b-regio
n
usin
g spa
r
se re
co
gnition algo
rit
h
m.
(4)
Enhan
ce
su
b-regio
n
ima
ge usi
ng a
d
aptive Walli
s filter with local p
a
rame
ters
according to the cla
s
s of su
b-regio
n
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Adaptive
Wall
is Filter Via Sparse Recog
n
iti
on for Autom
a
tic Control Points… (L
eilei Gen
g
)
3109
(5)
Take
the
second l
e
vel ma
tching
to extract a
nd
match G
C
Ps u
s
in
g Surf
algo
rithm
and then g
e
t rid of wrong
matchin
g
GCPs though e
s
t
i
mation epip
o
l
ar geo
metry con
s
trai
nt.
Figure 4. Two
level Matchin
g
Based o
n
Adaptive Walli
s Enhan
ce
me
nt
3.1. Cons
tru
c
t Re
cogni
tion Vector
Based on
Rad
i
ation-p
a
ram
e
ter
s
(a) Wate
r
(b) Mo
untain
(c
) Subu
rb
s with mino
rity
buildings
(d)
City with some buildings
(e)
City with mass
buildings
Figure 5. Re
mote Sensi
n
g
Sub-re
gion
Remote
se
nsing su
b-regi
o
n
s a
r
e cl
assif
i
ed into five sets: (1)
Wate
r set: Becau
s
e most
area
of ea
rth
is water, we
con
s
tru
c
t
wa
ter set, as
sh
own i
n
Figu
re 5(a
)
. (2
) M
ountain
set:
One
third of
land
i
s
m
ountain
a
nd its texture
s
a
r
e
pa
rt
icul
ar, a
s
sh
own
in Figu
re
5(b). (3) Sub
u
rb
s
set
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 3105 – 3
117
3110
with mino
rity building
s
: Su
b-regio
n
co
ntains fe
w co
n
s
tru
c
tion
s an
d its textures are sim
p
le, as
sho
w
n in
Fig
u
re 5
(
c). (4
)
City set with
some
buildi
n
gs: Sub
-re
gio
n
co
ntains
so
me co
nst
r
u
c
tions,
as sho
w
n in Figure 5(d). (5)
City set with
mass buildin
g
s
: Sub-regio
n
contai
ns
mass
con
s
tru
c
tion
s and its texture
s
a
r
e co
mplex, as
shown in Fig
u
re 5
(
e
)
. Th
en we
co
nst
r
uct
recognitio
n
vector of re
mo
te sen
s
in
g su
b-regio
n
u
s
in
g twelve
radi
ation-p
a
rame
ters. T
r
aini
ng
set
and test set are co
nstituted
by those re
cognition ve
ctors.
The m
a
in
re
sea
r
ch of
fe
ature
extra
c
tion for remot
e
sen
s
ing
im
age fo
cu
se
s
on pixel
-
based meth
o
d
su
ch
as P
C
A and I
C
A
and
so o
n
. F
o
r the
past f
e
w yea
r
s, m
any re
sea
r
ch
ers
made g
r
eat e
fforts to improve cla
ssifi
ca
tion pre
c
i
s
ion
.
Howeve
r, cl
assificatio
n
p
r
eci
s
io
n cann
ot
be effectively improved
du
e to the limitation of pixel
-
ba
sed m
e
th
ods. Be
cau
s
e high resolu
tion
remote
sen
s
i
ng image
co
ntains a
bun
d
ant of inform
ation and
co
mplex textures, this pa
p
e
r
con
s
tru
c
t
s
re
cog
n
ition vector of remote
sen
s
ing
sub
-re
gion
with twelve ra
diati
on-p
a
ramete
rs,
inclu
d
ing
col
u
mn of
sig
n
a
l
to noi
se
rati
o,
detail
ene
rgy, gray m
e
a
n
, edg
e e
nergy, gene
rali
zed
noise, gradi
e
n
t, angular
seco
nd mom
e
nt, gray varia
n
ce, ent
ropy,
definition, co
ntrast a
nd si
g
n
a
l
to noise
ratio
(SNR). Refle
c
ting compli
cation and
di
rection of texture, tho
s
e twelve para
m
et
ers
descri
be det
ails and edg
e
features
from
bot
h
spa
t
ial domai
n
and frequ
en
cy domain.
T
h
e
recognitio
n
of
rem
o
te sen
s
ing
sub
-re
gio
n
s m
a
inly
reli
es
on
statisti
cal info
rmatio
n of st
ru
cture
s
su
ch a
s
imag
e texture and
edge info
rm
ation. Due to
both image t
e
xture an
d e
dge info
rmati
on
belon
ging to
high fre
que
ncy information,
the noi
se
m
a
y be mista
k
e
n
for ima
ge te
xture, if we o
n
ly
con
s
id
er im
a
ge texture. T
herefo
r
e, in
o
r
de
r to
imp
r
o
v
e re
cognitio
n
accu
ra
cy, this p
ape
r sel
e
cts
SNR, gen
era
lized noi
se,
colum
n
of signal to noi
se ratio and
other ra
diatio
n-pa
ram
e
ters as
recognitio
n
vector of remo
te sen
s
ing
su
b-regio
n
s.
3.2. Recog
n
ition of Sub-r
e
gion Terrai
n
v
i
a Sparse Recog
n
ition
Sparse re
co
gnition algo
ri
thm
can get
high
er
re
co
gnition rate
and better
ro
bustn
ess,
comp
ared
with tradition
al reco
gnition
al
gorithm
s
su
ch as th
e ne
a
r
est
neigh
bo
r cla
ssifie
r
. T
h
is
method
effectively avoids over
-fitting and und
er-fitting,
be
ca
use
linea
r rep
r
e
s
entatio
n
of
test
sampl
e
ca
n b
e
got by some large
weigh
t
training
sam
p
les. The ve
ctor of linear repre
s
e
n
tation
is
spa
r
se, whi
c
h
reflect
s
the d
i
fference of sample
s amo
n
g cla
s
ses, a
s
sho
w
n in Equ
a
tion (3
):
0
0
ˆ
s.
t
.
x
argm
i
n
x
x
Xr
(
3
)
Whe
r
e,
x
is
t
e
st
sa
mple
;
r
is the v
e
ctor of line
a
r repres
ent
ation an
d re
cog
n
ition ve
ctor;
12
1,
1
1
,
,
1
,
,,
,,
,,
mn
nm
m
n
XR
xx
x
x
, it is
training
s
a
mples matrix;
x
i,j
is the
j
th
s
a
m
p
l
e
i
n
t
h
e
i
th
cla
ss;
m
is the numbe
r of sample cl
asse
s;
n
is the nu
mber of samp
les in pe
r cla
s
s.
Equation
(3
)
belon
gs to
th
e NP
-ha
r
d
problem
and
ca
nnot b
e
solve
d
in p
o
lynomi
a
l time.
Fortun
ately, compresse
d
sensi
ng d
r
a
w
s the con
c
lu
si
on as foll
owi
ng [32]: if the optimal sol
u
tion
(
0
ˆ
x
) of Equatio
n (3
) is fully
spa
r
se,
0
ˆ
x
ca
n
be solve
d
which Eq
uatio
n (3
) is tran
sformed i
n
to
minimal
l
1
norm, as sh
own i
n
Equation (4
):
0
1
ˆ
s.
t
.
l
xa
r
g
m
i
n
x
x
Xr
(
4
)
There is a
n
effective method for Eq
u
a
tion (4
), be
cau
s
e it is
a
convex opti
m
ization
probl
em. The
l
1
norm
can
be solved u
s
i
ng Basi
s Pu
rsuit (BP)
algo
rithm [33, 34]
, which is
sta
b
le
and well ro
b
u
st. Sparse
repre
s
e
n
tation
can b
e
tran
sform
ed into
linear
pro
g
ra
mming p
r
obl
em
though BP al
gorithm, be
cause minimi
zation of
l
1
norm and line
a
r
prog
ram
m
ing
can be
defin
ed
as con
s
train
e
d
optimizatio
n probl
em
s, as sh
own in Equation (5):
1
0
,,
ˆ
s.t
.
l
ij
ij
xa
r
g
m
i
n
x
x
x
r
(
5
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Adaptive
Wall
is Filter Via Sparse Recog
n
iti
on for Autom
a
tic Control Points… (L
eilei Gen
g
)
3111
3.3. Adaptive Wallis Enhancement
Walli
s filter can map th
e
mean a
nd va
rian
ce
wi
thin
wind
ows to g
i
ven values l
eadin
g
to
approximate
mean an
d varian
ce in i
m
age. The
r
e
f
ore, it can
stren
g
t
hen
contra
st whe
n
th
e
contrast is lo
w and al
so can wea
k
e
n
contra
st whe
n
the contrast i
s
high, so tha
t
the tiny cha
nge
of gray can b
e
transfo
rme
d
into visual textures
. In this way, Walli
s filter can grea
tly enhance t
h
e
texture an
d contra
st of ima
ge while effe
ctiv
ely supp
re
ssing
noi
se. Al
though th
e im
age e
nha
nce
d
by Walli
s filter is
similar to
noi
se imag
e,
we
can
match GCP
s
in l
o
wer un
ce
rtain
t
ies an
d in fal
s
e
prob
ability to improve mat
c
hing a
c
cura
cy. The Wa
llis
filter can be e
x
presse
d as
Equation (6):
10
(,
)
(
,
)
c
g
x
yg
x
y
rr
(
6
)
Whe
r
e,
1
()
fg
f
r
cs
c
s
s
c
, it is
multi-factor.
0
(1
)
f
g
rm
m
bb
, it is add
-facto
r.
m
g
and
s
g
a
r
e th
e m
e
an a
nd va
ria
n
ce
withi
n
wi
ndo
ws.
m
f
an
d
s
f
are the expected values
of Wallis filter
for imag
e an
d they are pa
ramete
rs of
Walli
s filter.
s
f
s
h
ou
ld
be
a
s
m
a
ller
va
lue w
h
en
th
e
s
i
ze
o
f
wind
ows d
e
crea
se
s.
b
i
s
a con
s
tant which
de
scrib
e
s
the
bri
ghtn
e
ss of im
age
and
c
is
a
l
so
a
con
s
tant whi
c
h de
scribe
s the
cont
rast
of image.
Th
e pa
ram
e
ter
of
c
sh
ould
b
e
a
bigg
er va
lue
whe
n
the si
ze
of windo
ws incr
ea
se
s. Th
e multi-fa
ctor
of
r
1
determines the performance of
Wallis
filter, and the relation
shi
p
with other pa
ra
meters ca
n b
e
descri
bed a
s
Equation
(7
):
2
1
11
g
f
rc
s
s
(
7
)
We
enh
an
ce
different textu
r
es u
s
ing
dyn
a
mic
si
ze of wind
ows with
co
rrespon
din
g
lo
cal
parameters t
o
make great use of Wallis filter,
as
shown in T
a
ble 1. We
set small si
ze of
wind
ows,
s
f
a
nd
c
to e
nha
nce
sub-regi
o
n
with
co
mpl
e
x textures,
such
a
s
city a
r
ea, and
set b
i
g
para
m
eters t
o
enh
an
ce
sub-regi
on
with sim
p
le text
ure
s
, such a
s
lake, mou
n
ta
in and
farm.
The
experim
ents sho
w
that a
daptive
Wa
lli
s filter
with
dynamic si
ze
of win
d
o
w
s eliminate
s
t
he
numbe
r
of pix
e
ls
with
satu
rated g
r
ay, wh
ich
gets a
bet
ter en
han
ce
m
ent for hug
e
remote
sen
s
ing
image
co
ntai
ning va
riou
s
terrai
n
s. In
th
is
way,
we
can g
e
t more
and
high
accura
cy of
cont
rol
points.
Table 1. Parameters of Wallis
Filter for S
ub-regi
on Terrains
Sub-region
size of w
i
ndo
w
s
s
f
c
Water
33
139
0.9
Mountain 31
135
0.87
Suburbs
w
i
th min
o
rit
y
buildings
25
131
0.85
Cit
y
w
i
th some b
u
ildings
21
127
0.83
Cit
y
w
i
th mass b
u
ildings
17
121
0.8
4. Results a
nd Analy
s
is
W
e
c
o
n
s
tr
uct
r
e
c
o
gn
itio
n
ve
c
t
or
o
f
hig
h
s
o
lu
tion
r
e
mo
te
se
ns
in
g
s
u
b-
re
g
i
on
w
i
th
radiatio
n-parameters a
nd
recogni
ze
su
b-regio
n
u
s
in
g sp
arse
re
cognition
algo
rithm. In ord
e
r to
increa
se
the
num
ber an
d
to imp
r
ove
the a
c
cura
cy
of G
C
Ps,
this
pap
er en
han
ce
s diffe
rent
remote
sensing sub
-re
gi
on terrain
s
usin
g ad
apti
v
e Walli
s filter
with
corresp
ondi
ng lo
cal
para
m
eters.
4.1. Discuss
Reco
gnition
Precision of
Remote Sen
s
ing Sub-reg
i
on Terrain
We
con
s
t
r
uct
t
r
ain
sam
p
le
set
a
nd t
e
st
samp
l
e
set with feature ve
ctors
com
posed fro
m
radiatio
n-parameters of
remote
sen
s
i
ng
sub
-re
gio
n
terrain
s
: water, mo
untai
n, su
burbs
with
minority build
ings, city wit
h
som
e
build
ings, an
d
city with mass buildin
gs
. In fac
t, 30 samples
belon
g to t
r
ai
n sample
set
whi
c
h
a
r
e
ra
ndomly
sele
cted from
ea
ch cl
ass i
n
cl
ud
ing 4
0
sam
p
les,
and othe
rs o
f
each
cla
s
s belong to t
e
st sample
set. Take me
an and
stan
dard
deviatio
n
of
recognitio
n
rate as
mea
s
ure of e
a
ch algo
rithm,
whi
c
h a
r
e
com
puted f
r
om 1
000 ti
mes
recognitio
n
s.
In orde
r to prov
e the effectiveness of ou
r method,
we desi
gn thre
e experim
ents
as
follows
:
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 3105 – 3
117
3112
(1) Di
scu
s
s reco
gnition
rat
e
of
remote
sens
i
ng su
b-re
gion
te
rrain u
nder different
feature
vectors and t
hat unde
r different cl
assifi
cation
st
rateg
i
es. Accordin
g to the mean and stan
da
rd
deviation of
recognitio
n
rate, we
cho
o
s
e the
mo
st suitabl
e feat
ure ve
ctor
a
nd cl
assification
strategy
so t
hat we
can
get the
high
est
re
co
gn
itio
n
pr
ec
is
io
n
.
T
h
e fe
a
t
ur
e r
e
p
r
es
en
ta
tio
n
s
inclu
de
gray
inform
ation,
feature ve
ctor b
a
sed
on
PCA a
nd
radiation
-
pa
ra
meters. And
the
recognitio
n
a
l
gorithm
s
co
ntain the
ne
are
s
t n
e
ighb
or
cla
s
sifier
and th
e
spa
r
se
recogniti
o
n
algorith
m
.
We
ca
n see
the differe
nce of recogniti
on rate
b
e
tween the
ne
arest n
e
ighb
or
cla
ssifie
r
and th
e
spa
r
se
re
co
gnitio
n
alg
o
rithm
u
nder three
fe
ature
vecto
r
s in Fi
gure
6:
Und
e
r differe
nt
dimen
s
ion
of
feature
vect
or, the
cha
n
g
e
of re
co
gniti
on ave
r
ag
e i
s
de
scri
bed
by the solid l
i
ne;
unde
r th
e giv
en di
men
s
ion
of featu
r
e
ve
ctor, th
e
cha
nge
of
stand
ard
deviation
is de
scrib
e
d
by
the dashed li
ne, whi
c
h reflects
stability of algorit
hm.
The expe
rim
ents sho
w
th
at the recogn
ition
rate i
s
betwe
en 7
0
% a
nd
80%, a
s
d
e
scribed
by
gree
n line
an
d
re
d line
in
Figu
re 6
(
a),
when
we
extract featu
r
e vector
ba
se
d on g
r
ay inf
o
rmatio
n
and
PCA usi
ng t
he ne
are
s
t n
e
ighb
or
cla
s
sifier;
then the reco
gnition rate is alway
s
mo
re
than 80%, th
e high
est bei
ng 95%, a
s
d
e
scrib
ed by b
l
ue
line in Fig
u
re 6(a
)
, wh
en
we
comp
ute radi
ati
on-p
a
ram
e
ters a
s
recognitio
n
vector. Th
us as
descri
bed
by
green
line
i
n
Fig
u
re
6
(
b), the rec
ogni
tion rate of f
eature
ve
ctor ba
sed
on
g
r
ay
informatio
n i
s
app
roximate
37% u
s
in
g
sparse
re
co
gni
tion alg
o
rithm
.
From
the
re
d line
in Fi
gu
re
6(b
)
, we
can
se
e that
hig
her re
co
gniti
on
rate
ca
n
be g
o
t with
t
he di
men
s
ion
of featu
r
e v
e
ctor
based on PCA increa
sin
g
. The recognit
i
on rate is
ab
out 86%, when the dimen
s
ion of feature
vector is ove
r
20. This pap
er ca
n get the
best re
cog
n
ition rate which the highe
st is 99.64% u
s
i
n
g
radiatio
n-parameters, as
descri
bed by
blue line in Figure 6(b
)
. It can be co
n
c
lud
ed that we
sho
u
ld com
p
ute
ra
diation
-
para
m
eters a
s
recogni
tio
n
vector
usi
n
g
spa
r
se
re
co
gnition al
gorit
h
m
to get the best result.
(a)
Re
cog
n
ition Rate of th
e Nea
r
e
s
t Nei
ghbo
r
Cla
ssif
i
e
r
(b)
Re
cog
n
ition Rate of Sp
arse Re
co
gni
tion
algorith
m
Figure 6. Re
cognition
Rate
for Variou
s F
eature T
r
an
sf
ormatio
n
s a
n
d Cla
ssifie
r
s
(2)
Di
scuss t
he impo
rtan
ce of ra
diation
-
pa
ramete
rs. Removin
g
ra
diation-pa
ram
e
ter o
ne
by one
from
re
co
gnition
vector,
we
comput
e
re
co
gnition rate
based on sp
arse re
cog
n
ition
algorith
m
. T
hen thi
s
pa
per
so
rts twelve
radi
ation-p
a
ramete
rs (RP) a
c
cordin
g to th
ose
recognitio
n
ra
tes. The hi
gh
er the
spa
r
se
misrecogni
tio
n
rate i
s
, the
more im
po
rta
n
t the radi
atio
n
-
para
m
eter is.
For
example
,
the sp
arse
recognitio
n
rate of feat
ure vecto
r
excl
uding
col
u
mn
of
sign
al to noise ratio is 98.
64% and spa
r
se mi
sr
ecog
nition rate is
1.36%, which
is the highe
st
spa
r
se mi
sre
c
og
nition rate
amon
g twelv
e
ra
diation
-
pa
ramete
rs. It can be
co
ncl
u
ded that
colu
mn
of sig
nal to
n
o
ise
ratio
g
r
e
a
t influent
sp
arse
re
cog
n
ition a
nd it i
s
t
he mo
st im
p
o
rtant
radi
ation-
para
m
eter. A
s
sho
w
n i
n
Figure 7, the
experim
ent
s sho
w
th
at the imp
o
rtan
ce of radiatio
n-
para
m
eters i
s
following:
co
lumn of si
gna
l to noise
rati
o (RP
-A
), d
e
tail ene
rgy
(
R
P
-B
), gray me
an
(RP
-C
), edg
e energy
(RP
-D
), ge
neralized
noi
se
(RP
-E
), gradient
(RP
-F
),
ang
ular seco
nd mome
nt
(RP
-G
), gray varian
ce (RP
-H
), entropy (RP
-I
), definition (RP
-J
), contr
a
st (RP
-K
) a
n
d
signal to no
ise
ratio (
R
P
-L
).
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Adaptive
Wall
is Filter Via Sparse Recog
n
iti
on for Autom
a
tic Control Points… (L
eilei Gen
g
)
3113
Figure 7. Importan
c
e of Ra
diation-Featu
r
es
(3) Di
scu
ss
the effectiveness and ro
bustn
ess of
different cl
assificatio
n
strategie
s
inclu
d
ing
the
nea
re
st n
e
ighbo
r
cla
ssif
i
er a
n
d
the
spa
r
se
re
co
gnition
algo
ri
thm ba
sed
on
radiatio
n-parameters. Th
e exper
i
m
ent
s sho
w
that
the sp
arse
reco
gnition
ra
te of these t
w
o
recognitio
n
al
gorithm
s a
r
e
improve
d
wit
h
the incr
ea
se
of
training sampl
e
s, whi
c
h
is rob
u
st, as
sho
w
n in
Fig
u
re 8.
Whe
n
the num
ber
of training
sam
p
les i
s
bet
we
en 20 a
nd 3
0
,
the recogniti
on
rate of spa
r
se recognitio
n
algorithm in
cre
a
ses
stea
dily and that
of neare
s
t ne
ighbo
r cla
s
sifier
increa
se
s
after
slig
htly de
cre
a
si
ng. T
h
e me
an
of
re
cog
n
ition
rat
e
is hig
h
e
r
a
nd the
stand
ard
deviation of t
hat is
sm
alle
r ba
se
d o
n
sparse
re
cog
n
i
tion algo
rith
m than th
ose
of the n
eare
s
t
neigh
bor
cla
s
sifier. So, the spa
r
se recognition
al
go
rithm can get
highe
r re
co
gnition p
r
e
c
ision
and better
ro
bustn
ess tha
n
those of oth
e
r traditio
nal
method
s.
Figure 8. Re
cognition
Rate
base
d
on Ra
diate-Pa
ram
e
ter
Therefore, thi
s
pap
er
com
putes
radiati
on-p
a
ramete
rs as featu
r
e
vector a
nd re
cog
n
ize
remote
sen
s
i
ng su
b-regi
o
n
terrai
n
usi
n
g spa
r
se re
cognition alg
o
rithm to get higher
re
cog
n
ition
pre
c
isi
on an
d
better rob
u
st
ness.
4.2. Discuss
Adaptiv
e
Wallis Enhancement
We enh
an
ce
textures
of
hi
gh solutio
n
re
mote sen
s
in
g
imag
e
u
s
in
g traditional
Wa
llis
filter
and
adaptive
Walli
s filter.
In ord
e
r t
o
p
r
ove the
effe
ctiveness of a
daptive
Walli
s e
nha
nce
m
e
n
t,
we de
sig
n
two experim
ent
s as follo
ws:
(1) Di
scu
s
s t
he e
nha
ncem
ent of traditio
nal
Wa
lli
s filter
and
ad
apti
v
e Walli
s filter for hi
gh
solutio
n
rem
o
te sen
s
ing
su
b-regio
n
terra
i
n.
Figure 9 and
Figure 10
show t
he difference of enhancem
ent bet
ween traditional Wallis
filter and ad
aptive Walli
s filter for water-regi
on an
d middle
-
de
n
s
ity region.
The expe
rim
ents
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 4, April 2014: 3105 – 3
1
17
3114
sho
w
that th
ere a
r
e ma
n
y
pixels with
saturated g
r
a
y
when
we e
n
han
ce
remo
te sen
s
ing
sub-
region terrain with simpl
e
textures
usi
n
g Wallis filter
with sm
all wi
ndow,
s
f
and
c
, a
s
sh
ow
n in
Figure 9
(
b
)
.
There a
r
e ve
ry few pixel
s
with
satu
rate
d gray u
s
ing
adaptive
Wall
is filter,
whe
n
we
enha
nce wat
e
r-regi
on and
mountai
n-re
gion with sim
p
le texture
s
.
That is be
ca
use
the
gray
and
dynamic ran
ge image (gray range of image
) of re
mote sen
s
in
g sub
-re
gion
terrain is
small.
Ho
wever,
ou
r metho
d
g
r
ea
t red
u
ces the
num
ber of p
i
xels
with
sat
u
rated
g
r
ay,
comp
ared
wit
h
traditional
Wallis filter,
as
sho
w
n
in Fi
g
u
re
9(c).
Thu
s
the
expe
ri
ments
sh
ow that a l
o
t of pi
xels
with satu
rate
d gray also e
x
ist when
we
enha
nce
rem
o
te sen
s
in
g sub-regi
on terrain with comp
le
x
textures using Wallis filter
with bi
g
window,
s
f
a
nd
c
, as sh
own
in
Figure
1
0
(b
). Our
metho
d
also
great redu
ce
s the nu
mbe
r
of pixels wit
h
satu
rated g
r
ay, whe
n
we enha
nce ci
ty-regio
n
, whi
c
h
contai
ns va
ri
ous texture
s
with big
dyna
mic
rang
e im
age, a
s
sho
w
n in Fi
gure 1
0
(c). Th
erefo
r
e,
we
sh
ould
en
han
ce diffe
re
nt rem
o
te se
nsin
g
sub
-r
e
g
i
on terrain
s
u
s
ing
ada
ptive Walli
s filter
with
different si
ze
of windo
w an
d other p
a
ra
meters so
th
at we can effectively avoid
the existen
c
e of
pixels
with saturated
gray.
In
this way,
remote
sensi
ng su
b-region
terrain
s
can
b
e
b
e
tter
enha
nced, so that the nu
mber of G
C
Ps ca
n
be in
cre
a
sed a
n
d
the accu
racy of that can
be
improve
d
.
(a) Wate
r-
reg
i
on
(b) Walli
s
enh
ancement
(c) Adaptive Walli
s
enha
ncement
Figure 9. Co
mpari
s
o
n
of Enhan
c
em
en
t
(a) Mid
d
le
-de
n
sity
regio
n
(b) Walli
s
enh
ancement
(c) Adaptive Walli
s
enha
ncement
Figure 10. Co
mpari
s
o
n
of Enhan
cem
e
n
t
(2)
Di
scuss the pe
rform
a
n
c
e
of ad
aptive Walli
s en
h
ancement fo
r high solution
remote
sen
s
in
g sub-regio
n
terrain
.
Enhance te
xtures of
re
mote se
nsi
n
g su
b-regio
n
terrai
n
s
of ZY-3
who
s
e
size is 400×400 u
s
i
ng tradition
al
Wallis f
ilter
and ad
aptive
Wallis filter.
Then extra
c
t the
GCP
s
from
e
nhan
ce
d re
m
o
te se
nsi
ng
sub
-
regio
n
te
rrai
n
s
usi
n
g t
w
o level
s
ma
tching m
e
thod
descri
bed in
cha
p
ter 3. We comp
are t
he numb
e
r
a
n
d the accu
racy of GCPs using differe
n
t
filters to prov
e the perfo
rm
ance of adapt
ive Wallis e
n
han
ceme
nt.
Evaluation Warning : The document was created with Spire.PDF for Python.