TELKOM
NIKA
, Vol.12, No
.3, Septembe
r 2014, pp. 6
05~612
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v12i3.81
605
Re
cei
v
ed Ap
ril 2, 2014; Re
vised July 1
0
, 2014; Accept
ed Jul
y
26, 2
014
Combine Target Extraction and Enhancement Methods
to Fuse Infrared and LLL Images
Yong Chen
*
1
, Jie Xiong
1
, Huan-lin
Liu
2
, Qiang Fan
1
1
Ke
y
La
bor
ato
r
y
of Industri
a
l
Internet of T
h
ings& Ne
t
w
ork C
ontrol, MOE, Chon
gqi
ng U
n
iv
ersit
y
of Posts
and T
e
lecomm
unic
a
tions, Ch
ong
qin
g
, Chin
a, 4000
65
2
Ke
y
La
bor
ato
r
y
of Optical F
i
ber Comm
unic
a
tion T
e
chno
lo
g
y
, C
hon
gq
ing,
Chin
a, 400
065
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: chen
yo
ng@c
qupt.ed
u
.cn
A
b
st
r
a
ct
F
o
r gettin
g
the
usefu
l
o
b
ject
i
n
formatio
n
fro
m
infrare
d
i
m
a
ge
and
min
i
n
g
more
deta
il
of low
li
ght
level
(LL
L
) i
m
a
ge, w
e
pro
pos
e a n
e
w
fusio
n
meth
od
bas
ed
on se
g
m
e
n
tati
on a
nd
en
hanc
ement
meth
od
s in
the pap
er. F
i
rst, using 2D ma
ximu
m entropy
metho
d
to
segment the ori
g
i
nal infrar
ed i
m
age for extracti
ng
infrare
d
target,
enh
anci
ng
orig
i
nal L
LL
i
m
ag
e
by Z
ade
h trans
form for
mi
nin
g
mor
e
deta
il i
n
formatio
n
, on th
e
basis of th
e se
gmente
d
map
to fuse the e
n
hanc
ed L
LL i
m
age
and
orig
in
al infrar
ed i
m
a
ge. T
hen, or
igi
n
a
l
infrare
d
i
m
ag
e
,
the enha
nce
d
LLL
i
m
ag
e
and th
e first fused i
m
ag
e ar
e used to r
eal
i
z
e
fusio
n
in
n
on-
subsa
m
ple
d
contour
let trans
form (NSCT
)
do
ma
in, w
e
get the secon
d
fused i
m
ag
e. By contrast of
exper
iments, the fuse
d i
m
a
g
e
of the sec
o
n
d
fused
meth
o
d
’
s visu
al effec
t
is better than
other
meth
ods
’ fro
m
the literat
ure. F
i
nally, Obj
e
cti
v
e ev
al
uati
on i
s
used to ev
al
uate the fus
e
d
images
’
qu
alit
y, its results al
so
show
that the propos
ed
met
h
o
d
can po
p targ
et informatio
n
, improve fus
ed i
m
a
g
e
’
s resol
u
ti
on an
d contras
t
.
Ke
y
w
ords
: 2D
maxi
mu
m entr
opy, Z
adeh tra
n
sform,
e
nha
n
c
ement, the se
cond fuse
d, N
S
C
1. Introduc
tion
Image fu
sion
nam
ely u
s
e
s
re
dund
ant
d
a
ta an
d
com
p
lementa
r
y in
formation
fro
m
multi-
sen
s
o
r
s fo
r o
b
taining a im
age which ha
s accu
rate ta
rget, own
s
g
ood visu
al effects [1]. As a
n
importa
nt bra
n
ch of ima
ge
fusion, infrare
d
and lo
w lig
ht level (LLL
) image
s fusio
n
’s mai
n
task is
to achi
eve th
e rea
s
o
nabl
e
and
comp
re
h
ensive
de
sc
ri
ption of targe
t
and sce
ne,
on the
con
d
ition
to retain th
e
origin
al dat
a inform
ation
as m
u
ch a
s
possibl
e an
d avoid fal
s
e
informatio
n. At
pre
s
ent, the techn
o
logy h
a
s be
en wid
e
ly used in
i
n
telligent tra
n
sp
ortation,
safety monito
ring,
human visual
auxiliary fields and
so on [
2
]-[3].
A kind of sen
s
ors ha
s its o
w
n f
eature, so they can ca
pture so
me p
a
rt informatio
n of the
scene, so we
combi
nes th
e infrared an
d LLL sen
s
o
r
s to de
scribe
the whole
scen
e better. In
orde
r to bette
r com
b
ine im
aging a
d
vant
age
s of t
hese
two se
nsors, many schola
r
s h
a
ve don
e
a
lot resea
r
ch,
and put fo
rward a
se
rie
s
of fusion
m
e
thod
s, inclu
d
i
ng differe
nt kinds
of pyra
mid
fusion meth
o
d
s [5], wavel
e
t transfo
rm
methods
[6]
-
[7], curvelet
transfo
rm
method
s [8], th
e
conto
u
rlet ta
nsform meth
ods [9], the non-sub
s
a
m
pled conto
u
rlet tran
sform methods
[10],
she
a
rlet t
r
an
sform
metho
d
s
and
so
on [11
]. All
these met
hod
s b
a
ses on m
u
lti-scale
decompo
sitio
n
ap
pro
a
ch, f
i
rst, o
r
igin
al i
m
age
s
are
d
e
com
p
o
s
ed
i
n
to lo
w frequ
ency
co
effici
ents
and hi
gh fre
quen
cy coefficient
s, then,
different
fu
sion
rule
s a
r
e used to p
r
ocess the l
o
w
freque
ncy
co
efficients
and
the high fre
q
uen
cy coe
ffi
cients, re
sp
ect
i
vely. All these method
s
can
achi
eve a go
od fusi
on visual effect
s, b
u
t some
sh
ortcoming
s
in
p
r
eservin
g
ori
g
inal info
rmat
ion
from the ori
g
i
nal image
s,
esp
e
ci
ally, for the in
suffici
ent sun
s
hi
ne
or target
s conceale
d
and
so
on, easy to l
ead targets l
o
se o
r
un
ob
vious, so
we
can n
o
t ea
sy to underst
and the
sce
ne.
Therefore, i
n
re
cent ye
ars, the sch
o
lars h
ad
put forward
som
e
o
t
her fu
sion
m
e
thod
s [12]-[
15],
for
extra
c
ting
the
targ
et
o
r
mining more depth
detail b
e
tter. Literatu
re[12]
com
b
in
ed
comp
re
ssed
sen
s
in
g prin
ciple to fuse image
s, the fuse
d image
could de
crea
se fusion time
and had a b
e
tte
r
visual effe
cts, but its targ
et highlig
hts
unobvio
usly.
Li Shutao et
al. [13] pro
p
o
se
d a meth
od
whi
c
h b
a
sed
on ave
r
ag
e filter to d
e
comp
ose
the o
r
igin
al imag
es int
o
ba
se
and
d
e
tail layer, th
en,
use
d
guid
ed
filter stru
cture weig
ht map to fuse
the
original im
a
ges a
c
cordin
g to the wei
ght
grap
h. Thi
s
method q
u
ickly reali
z
e
s
f
u
sio
n
, it
also
had b
e
tter
perfo
rman
ce
in detail, but
the
contrast
of th
e fuse
d ima
g
e
wa
s
poo
r.
Literatu
re
[14
]
introdu
ce
d l
o
cal
histo
g
ra
m equ
alization to
enha
nce bot
h the infra
r
e
d
and
LLL i
m
age
s, then
denoi
se
d wit
h
the medi
an
filter, this m
e
thod
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 3, September 20
14: 60
5 – 612
606
fused i
m
age
s fast and th
e
fused i
m
age
s had
clea
r de
tail, but its target co
uldn’t show obviou
s
l
y
.
Xing Suxia et al propo
sed
a target extraction
meth
o
d
[15] which based on
Re
nyi entropy to
segm
ent the infrared imag
e for extracti
ng therm
a
l targets, an
d en
han
ce the LL
L image’
s det
ail
in NS
CT d
o
m
ain. The
fu
sed i
m
ag
e h
ad a
high
bri
ghtne
ss,
so l
o
st its
detail
informatio
n a
t
the
same time.
All those
m
e
thod
s ab
ove ca
n a
c
hie
v
e com
p
lem
entary a
nd
redu
nda
ncy
betwe
en
different ima
ges’ i
n
form
ation, co
ntra
st
or resolu
tion
has
bee
n im
proved
to a
certain
extent, but
contrast
and
definition
ca
n’t com
patibl
e
. In
order to enh
an
ce t
he
contrast
and im
prove
th
e
resolution
of
the fused im
age, the
pap
er p
r
e
s
ent
s a
ne
w meth
od
whi
c
h
se
gm
ents th
e o
r
igi
nal
infrared imag
e and enh
an
ce LL
L imag
e, then to
fuse the segm
e
n
ted image a
nd the enha
n
c
ed
image, the fu
sed i
m
ag
e is
calle
d the first fused i
m
age
. Aiming to p
r
event the in
complete im
ag
e
segm
entation
and
over e
nhan
cem
ent
LLL im
age
e
ffectively, the pap
er
uses the first fused
image
and
th
e ori
g
inal i
n
frared
an
d the
enh
anced
L
LL ima
ge to
fuse fo
r mo
re ori
g
inal
im
age
informatio
n, the fuse
d ima
ge is called th
e se
con
d
fused image.
The rest of
the pap
er i
s
o
r
gani
ze
d
as follo
ws: In Section
2, it mains
introdu
ce
pretreatment,
infrare
d
ima
ge se
gmentat
ion and
LL
L image en
han
cement metho
d
are de
scrib
ed.
In Section 3, fusion strate
gy is discussed.
In Section 4, experim
ents re
sult
s and analy
s
is
are
put. Finally, in Section 5, concl
u
si
on
s of the work a
r
e
made.
2 Pretre
atme
nt
Ho
w to re
se
rve importa
nt informatio
n of
the
scen
e be
tter has
bee
n
the main di
re
ction of
our
resea
r
ch. For this
purpose, the pa
per u
s
e
s
the
segm
entation
method o
n
infrared ima
g
e
to
extract the m
a
jor ta
rget inf
o
rmatio
n, in
additi
on, al
so
apply a
ce
rtain en
han
ce
ment metho
d
to
excavate mo
re details d
e
e
p
ly. Based o
n
the anal
ysi
s
above,
we
put forward
a
n
idea of
fusion
treatment to fuse infrared a
nd LLL ima
g
e
s
. We di
scuss the detail a
s
follows.
2.1 2D maxi
mum entrop
y
Therm
a
l targ
et information
is important
info
rmatio
n for the infrared
image, we u
s
e
s
the
2D maximu
m entropy thre
shol
d [16] to segment
the infrared
image, for extracting ta
rget
informatio
n to better poppi
n
g
the infrared
target.
Due to
probability distribution of the
s
egm
ented i
n
frared image’s target region and
backg
rou
nd region a
r
e diff
erent, so we prop
oses
po
sterior
pro
babi
lity of the gray and mea
n
of
gray
regi
on t
o
no
rmali
z
e
e
a
ch
region’
s
occur p
r
ob
abi
lity
ij
p
. Supp
ose
the segm
enta
t
ion value
of
the image is
,
s
t
, backgroun
d regio
n
’s p
r
ob
ability and
target regio
n
’s
prob
ability are
B
p
and
O
p
,
r
e
spec
tively.
,
B
ij
pp
i
j
(1)
whe
r
e,
1
,
2,
..
.,
is
,
1
,
2,
..
.,
jt
,
O
ij
pp
i
j
(2)
whe
r
e,
1
,
2,
...
,
1
is
s
L
,
1
,
2,
...,
1
jt
t
L
.
The discrete
2D entropy is defined a
s
:
lg
ij
ij
ij
Hp
p
(3)
The 2D e
n
tro
p
y of objective regio
n
:
lg
l
g
ij
ij
O
O
ij
OO
O
pp
H
HO
p
p
pP
(4)
whe
r
e,
1
,
2,
..
.,
is
,
1
,
2,
...,
jt
.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
9
30
Com
b
ine Ta
rget Extra
c
ion
and Enha
nce
m
ent Me
thods to Fuse Infrared …. (Y
on
g Che
n
)
607
The 2D e
n
tro
p
y of backgro
und re
gion:
lg
lg
ij
ij
B
B
ij
BB
B
pp
H
HB
p
pp
P
(5)
Whe
r
e,
1
,
2,
..
.,
1
is
s
L
,
1
,
2
,
..
.,
1
jt
t
L
.
1
BO
pp
(6)
1
BO
HH
(7)
The discrimi
nant functio
n
of image entropy is define
d
as:
,l
g
l
g
1
lg
1
OO
OO
OO
OO
OO
OO
HH
H
fs
t
H
O
H
B
p
p
Pp
p
HH
H
pp
Pp
p
(8)
Then, the opti
m
al thre
shol
d
vector
**
,
s
t
must satisfy the following
con
d
ition:
**
1
,a
r
g
m
a
x
,
sL
s
tf
s
t
(9)
Shown in Fi
g
u
re 1, dem
on
strated th
e se
gmentation ef
fect, Figure 1
(
b)
sho
w
s the
whole
and complete
person targe
t
and some h
o
t region
s.
(a
) th
e infrared image
(b) the se
gmented image
Figure 1. The
chart of seg
m
entation effection
2.2 Zadeh tr
ansform
A LLL
Imag
e contain
s
more
detail
i
n
format
io
n, its
scene
i
s
dark
and
co
ntrast
is
unobvio
us. I
n
order to
mine mo
re
details i
n
fo
rmation for the fused im
age, we u
s
e
s
th
e
enha
ncement
method to intensify the LL
L image.
The tradition
al meth
ods i
n
clu
d
ing
hi
stogra
m
e
qual
i
z
ation,
gray
stret
c
hing
a
n
d
so o
n
, the
s
e
method
s can not
process
d
y
namic effect
well
and
enh
ance the im
a
ge noi
s
e
more, we i
n
trod
u
c
e
the Za
deh
transfo
rm th
eo
ry and
p
r
in
ci
ple for ima
g
e
inform
ation
mined
metho
d
p
r
esented
f
r
om
our lab in lite
r
ature [17] to intensify the o
r
iginal L
LL im
age.
The ima
ge’
s
uncertainty p
r
ope
rty is th
e
rea
s
o
n
for i
m
age
blur
proce
s
sing. T
h
e imag
e
enha
ncement
algorithm is
desi
gne
d
in consi
deration
of human’
s subj
e
c
tive sen
s
e. If this method
can
com
b
ine
some vi
sual
chara
c
te
risti
c
s of the huma
n
, image an
d video quality visual effect
will
be signifi
cant
ly
increa
sed. Here,
the un
derlying
im
ag
e minin
g
i
s
t
o
obtai
n
the
highe
st p
o
ssi
b
le
contrast
re
sol
u
tion, so th
e
origin
al Za
de
h tran
sform [0,1] interval
extended to [
0
,255] to defi
ne
the s
patial domain, whic
h c
o
ns
titutes
the Z
ade
h tra
n
sform enh
a
n
cem
ent met
h
od, the pap
e
r
calle
d the me
thod Zad
eh transfo
rm. In o
r
de
r to mine
more im
age i
n
formatio
n of
the LLL im
a
ge,
the pape
r set
s
25
5
k
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
9
30
TELKOM
NIKA
Vol. 12, No. 3, September 20
14: 60
5 – 612
608
The co
ncrete
descri
p
tion is
as follo
ws
:
,,
Tx
y
k
O
x
y
(1
0)
whe
r
e,
,
Tx
y
and
,
Ox
y
re
p
r
es
en
t th
e
lo
c
a
tion
,
xy
of the targ
et an
d o
r
iginal im
age,
respe
c
tively.
0,
25
5
and
1
,
255
de
note
the
startin
g
point
and
g
oal of
co
nve
r
ting the
g
r
a
y
level imag
e,
r
e
spec
tively.
1,
2
5
5
k
repre
s
ent
s the co
efficient
of the space
expan
sion. When k i
s
gre
a
ter tha
n
the limit of
h
u
man
vision
contrast
re
sol
u
tion,
Eq.(1
0
) sh
ows th
e
p
e
rform
a
n
c
e
o
f
the u
nde
rlying
image
minin
g
functio
n
s;
wh
en
k i
s
le
ss th
an the
lim
it of
hum
an vi
sio
n
contrast
re
solution, Eq.(1
0
)
achi
eves
the image hidd
en
feature
s
.
1
, Eq.(10
) i
s
to
a
c
hieve
imag
e
bina
rization
function.
and
are minin
g
para
m
eters.
After Zadeh tran
sform, Eq.
(
11
) may occur
,2
5
5
Tx
y
or
,0
Tx
y
which
beyond the
domain, this
will cause
confusion.
Therefore, this need set
s
som
e
const
r
aint
s, namely, while the
gray value is
over 255, the
value is set to 255; while
th
e gray value i
s
low 0, the value is
set to 0.
255
,
,
255
,
0,
,
0
Tx
y
Tx
y
Tx
y
(11)
We illu
strate
the importa
nt of the enhan
ceme
nt method a
s
follows, one
pair fo
r
displ
a
ying th
e gray ima
g
e
,
sho
w
n a
s
F
i
gure
2, Figu
re 2(a
1
) i
s
th
e origi
n
al g
r
a
y
image, Fig
u
re
2(a2
) i
s
the
g
r
ay
spe
c
tru
m
of Fig
u
re 2
(
a
1
),
Fig
u
re 2
(
b
1
) i
s
th
e
co
rresp
ondi
ng
en
han
ced
imag
e,
Figure 2(b
2
) i
s
the gray sp
ectru
m
of Figur
e 2
(
b1
). We set the initial param
e
ters
0
and
50
.
In Figu
re 2
(
a
1
), altho
ugh
we
hardly se
e anything
in
it, when
u
s
e
d
zadeh
tra
n
s
form
to inte
nsify
the imag
e, its e
nhan
ce
d
effect sho
w
n
as Figu
re
2(b1).
Cont
ra
st the g
r
ay
sp
ectru
m
of Fi
g
u
re
2(a2
)
and
Fig
u
re
2(b2), th
e initial
spe
c
t
r
um i
n
Fig
u
re
2(a
1
)
had
b
een
expand
e
d
a lot. T
h
ro
ugh
this exampl
e
to demon
stra
te the enha
n
c
eme
n
t meth
od, we
can
realize that it is ne
ce
ssary
to
intensify the origin
al LLL i
m
age.
(a1) g
r
a
y
original image
(a2)
gra
y
spe
c
trum
(b
1
)
enhanced imag
e (b
2) g
r
a
y
spectr
um
Figure 2. Gra
y
image test
3 Fusion str
a
tegy
First, to segm
ent the ori
g
in
al infra
r
ed im
age a
n
d inten
s
ify the origi
n
al LLL im
age,
based
on the
segm
ented bi
nary
image to g
u
i
de its fu
sion,
its co
ncrete
method a
s
fo
llows. The ta
rget
regio
n
s’
pixel
s
a
r
e
con
s
titu
tes by ta
king
the pixels f
r
o
m
the corre
s
p
ondin
g
lo
cati
on in the
ori
g
i
n
al
infrared imag
e. The other
regio
n
s a
r
e p
i
cki
ng pixe
ls
from co
rresp
ondin
g
to the binary imag
e’s
non-ta
rg
et area of the loca
tion in the en
han
ced L
LL i
m
age, we o
b
tains the first fuse
d image F
1
.
Then, origi
nal
infrared im
a
ge(IR), the enhan
ce
d ima
ge(LE
) and th
e first fused i
m
age F1
are u
s
e
d
to NSCT tra
n
sform re
sp
ecti
vely, get the corre
s
po
ndi
ng low frequ
ency sub
-
ba
nd
coeffici
ents
,
IR
Cx
y
,
,
LE
Cx
y
,
1
,
F
Cx
y
; different
sca
l
e an
d di
re
ctional
high
fre
quen
cy
sub
-
band
coeffici
ents
,
,
IR
jl
Cx
y
,
,
,
LE
jl
Cx
y
and
1
,
,
F
jl
Cx
y
.
Low
freque
ncy sub-b
and NS
CT coefficie
n
ts
is
cal
c
ulate
d
by Eq.(12):
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Com
b
ine Ta
rget Extra
c
ion
and Enha
nce
m
ent Methods to Fuse Infrared …. (Y
on
g Che
n
)
609
21
,,
,
,
3
FI
R
L
E
F
Cx
y
C
x
y
C
x
y
C
x
y
(12
)
Usi
ng the ab
solute value of
the high freq
uen
cy sub
-
ba
nd NSCT co
e
fficients take
greate
s
t sh
o
w
n a
s
Eq.(13
):
2
1
,,
,
,
,m
a
x
,
,
,
,
,
FI
R
L
E
F
jl
jl
jl
jl
C
x
y
C
xy
C
x
y
C
xy
(13)
After a seri
es process ab
o
v
e, we get
fused ima
ge (F
2)’s
NSCT
co
efficients {
2
,
F
Cx
y
,
2
,
,
F
jl
Cx
y
}, finally, the
NSCT inve
rse transf
o
rm,
we get the fu
sed ima
ge F2
.
4 Experimen
t
results and
analy
s
is
The
experi
m
ents
ba
sed
o
n
the
MATLA
B
platform, i
n
frared
and
LLL i
m
age
of
the t
w
o
grou
ps f
r
om
Holla
nd T
N
O
institute’s pi
cture li
br
a
r
y
have bee
n re
gistratio
n
co
mpletely. Fig
u
re
s
3(a
)
-(b) a
s
th
e first grou
p sho
w
n, Figu
re 3(a
)
can id
entify a perso
n clea
rly, Figure 3
(
b)
can
see
road,
slop
e, shrub
and fe
n
c
e a
nd ot
h
e
r
scene
ry, but its co
ntra
st fe
els n
o
t well;
shown in Fig
u
res
4(a
)
-(b) a
s
th
e second
gro
up, Figu
re
4(a)
can
ide
n
tify person
s
,
ship an
d oth
e
r
small ta
rget
s,
Figure 4(b
)
can se
e the sh
ip, the sky cle
a
rly.
4.1 The exp
e
r
imental res
u
lts and su
b
j
ectiv
e e
v
a
l
uation
For the
pu
rp
ose
of comfirming th
e
co
rre
ct
ne
ss an
d effectivene
ss
of the p
r
opo
se
d
algorith
m
, we
use
s
the
oth
e
r fou
r
meth
o
d
s
co
mp
ared
with the first
fused im
age
and the
se
co
nd
fused im
age
to test. Figure
s
3(c) an
d Figures
4
(
c) a
r
e fu
sed
base
d
on t
he bior97
wavelet
transform method, the source
code is
publi
c
ally available (h
ttp://www.metpix.de/toolbox.ht
m
).
Figures 3(d) and
Fi
gures 4(d
)
a
r
e
the
fused
i
m
ag
es whi
c
h
are
fuse
d by the
NSCT
metho
d
.
Figures
3(c)-(d), Figu
re
s 4
(
c)-(
d) both choo
se
the
fu
sion rule
s
by
averagi
ng th
e low frequ
e
n
cy
coeffici
ents
a
nd getting th
e maximum
absolute val
ues
of high
freque
ncy
co
efficients, th
eir
decompo
sitio
n
level all are 4 layers, called wavelet
method a
n
d
NSCT m
e
th
od, re
spe
c
tively;
Figure 3(e
)
and Figu
re 4
(
e) a
r
e fu
se
d by the me
thod in litera
t
ure [15], its sou
r
ce co
d
e
is
publi
c
ally available (http://www.x
udongkang.weekly.
c
om
) called as
GF; Figure 3(f) and Fi
gure
4(f) a
r
e fused
by the method in literatu
r
e [15],
called
as Renyi entropy
method;
Figure 3(g) a
nd
Figure 4(g)
a
nd are the f
u
se
d
imag
es by the meth
ods
whi
c
h fu
se
s the
seg
m
ented inf
r
ared
image
and
th
e en
han
ce
d i
m
age, a
s
the
first fu
se
d i
m
age; b
a
sed
on th
e first f
u
se
d ima
ge,
we
fuse the
first
fused
imag
e
and th
e o
r
ig
inal infrare
d
i
m
age
and
e
nhan
ce
d LL
L
image
with t
h
e
se
con
d
fuse
d method, i
t
s visual effect sh
own as Fig
u
re 3
(
h) a
nd Fig
u
re 4
(
h). F
o
u
r
decompo
sitio
n
level
s
with
2,8,8,16 di
re
ctions f
r
om
co
arse
scale
to
finer
scale
are ad
opted
by
the
method
s whi
c
h use
s
the NSCT method.
Subjective ef
fects a
r
e a
n
a
l
yzed a
s
follo
ws. G
r
ou
p o
ne’s
effection
sho
w
n in
Fi
gure
3,
Figure 3(c)’s
fence
se
ems very fu
ssy, its co
ntra
st fe
els n
o
t cle
a
r
and d
e
tail lo
st se
riou
sly; we
can
see cl
ea
r detail from Figure 3(d
)
, but its c
ontra
st is very poor, this state
may hide so
me
importa
nt information; Fig
u
r
e 3(e)’
s
cont
rast i
s
so lo
w that we ca
n hard i
dentify some im
po
rta
n
t
informatio
n; Figure 3(f)’
s
contrast
see
m
s very
obvi
ously, and
we can
see
ob
vious target, due to
its over en
ha
nce
d
, its
sma
ll detail info
rmation d
o
e
s
not
very clea
r,
the whol
e visual see
m
s
not
very goo
d; F
i
gure
3(g)
ha
s go
od
definit
ion, its ta
rget
see
m
s obvio
usly, its q
ualit
y is bette
r th
an
the forme
r
m
e
thod
s; wh
en
com
pared
wi
th the Figu
re
3(h
)
, it has l
e
ss
cle
a
r
detai
l; so throug
h i
t
s
comp
ari
ng, Fi
gure
3
(
h
)
’s visual
effect
is
the be
st. Th
e
othe
r p
a
ir ca
n
be
analy
z
ed in
the
sa
me
way
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 3, September 20
14: 60
5 – 612
610
(a)
Original IR im
age
(b)
Original LLL i
m
age
(c) Wavelet method
(d) NS
CT meth
o
d
(e)
GF met
hod
(f) Ren
y
i ent
rop
y
method
(g) T
he first fuse
d
(h) T
he second f
u
sed
Figure 3. The first gro
up e
x
perime
n
ts of
infrare
d
and
LLL imag
e fusion
(a)
Original IR im
age
(b)
Original LLL i
m
age
(c) Wavelet method
(d) NS
CT meth
o
d
(e)
GF met
hod
(f) Ren
y
i ent
rop
y
method
(g) T
he first fuse
d
(h) T
he second f
u
sed
Figure 4. The
second g
r
ou
p experim
ent
s of infrared a
nd LLL ima
g
e
fusion
4
.3 Objectiv
e
e
v
aluation
In order to fu
rther
asse
ss t
he fu
sion
pe
rf
or
ma
nce of
di
fferent meth
o
d
s
obje
c
tively. Thre
e
fusion
quality
metrics
are
applie
d, i.e., averag
e g
r
ad
ient (AG
)
[18]
, spatia
l freq
uen
cy (SF) [19],
the edg
e rete
ntion metri
c
s
(
AB
F
Q
) [13] are
ad
opted in th
e
pape
r.
All the
evaluation
m
e
trics u
s
e
d
in the pape
r a
r
e define
d
as
follows:
1) Averag
e gradient(A
G)
Improveme
n
t of an image’
s quality ca
n
be expre
s
se
d by the average g
r
a
d
ient
,
whi
c
h
reflect
s
the clarity of the image, refle
c
ti
ng t
he small
details contra
st in the image and texture
variation, the greate
r
the a
v
erage g
r
a
d
ie
nt, the
image has b
e
tter integratio
n, defined a
s
follows:
12
22
11
1
,,
MN
ij
Gx
f
x
y
y
f
x
y
MN
(14)
whe
r
e,
,
x
fx
y
and
,
yf
x
y
repre
s
e
n
t pixels in the x
and y directi
ons’ first-o
r
d
e
r varian
ce
,
r
e
spec
tively.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
Com
b
ine Ta
rget Extra
c
ion
and Enha
nce
m
ent Methods to Fuse Infrared …. (Y
on
g Che
n
)
611
2) Spatial fre
quen
cy(SF
)
Ro
w frequ
en
cy of an imag
e is define
d
a
s
follows:
11
2
00
1
,,
1
MN
ij
RF
f
i
j
f
i
j
MN
(15)
Colum frequ
e
n
cy of an ima
ge is defin
ed
as follo
ws:
11
2
00
1
,1
,
MN
ij
CF
f
i
j
f
i
j
MN
(16)
whe
r
e, M and
N denote the
numbe
r of ro
ws a
nd colu
mns in the im
age. The spa
t
ial frequen
cy
of
the image is
defined a
s
fol
l
ows:
22
SF
R
F
C
F
(17
)
Spatial freq
u
ency
refle
c
ts
the overall le
vel ac
tivity of an ima
ge in t
he spatial d
o
m
ain. The
greate
r
the value, the visu
al effection is
better.
3)
AB
F
Q
The
gra
d
i
ent ba
se
d in
dex Q
evalu
a
tes th
e
eff
e
ctive kee
p
edge
inform
ation whi
c
h
transfe
rred from the sou
r
ce image
s to the
fuse
d ima
ge. It is calcul
ated as follo
ws:
11
11
,,
,,
,,
MN
AF
A
B
F
B
ij
AB
F
MN
AB
ij
Q
i
j
i
j
Q
ij
ij
Q
ij
ij
(18)
whe
r
e
AF
AF
AF
gO
QQ
Q
.
AF
g
Q
and
A
F
O
Q
are defined a
s
edge
stren
g
th
and orie
ntation pre
s
e
r
vati
on value
s
at location
,
ij
, resp
ectively. N and M a
r
e
as the wi
dth and hei
ght of the image
s.
B
F
Q
is similar
to
AF
Q
.
,
A
ij
and
,
B
ij
refle
c
t the importa
nce of
,
AF
Qi
j
and
,
BF
Qi
j
, res
p
ec
tively.
Tabel
1 sho
w
s that the
pro
posed m
e
tho
d
ha
s
the
gre
a
test ave
r
ag
e
gra
d
ient val
ue, the
greate
s
t spat
ial frequ
en
cy value an
d the great
e
s
t edge
retentio
n value, it shows that th
e
prop
osed me
thod ha
s the
best visu
al effect, this
consi
s
tent wit
h
the su
bje
c
tive evaluation
res
u
lts.
Tabel 1. Obje
ctive evaluati
on
The first grou
p e
x
periments
The second gro
u
p
expe
riments
AG SF
AB
F
Q
AG SF
AB
F
Q
Wavelet
method
5.4488
12.4351
0.3437
2.7144
7.9473
0.5774
NSCT
meth
od
5.0837
11.9110
0.4389
2.6694
7.6754
0.6627
GF met
hod
2.902
6.6384
0.3157
1.4837
4.8016
0.4359
Ren
y
i ent
rop
y
m
e
thod
4.3435
11.2408
0.3809
1.1292
5.8289
0.3102
The first fused
6.1873
15.2202
0.4321
1.2114
6.1556
0.4122
The second fuse
d
6.5524
15.8767
0.4398
2.7409
8.2393
0.6727
5 Conclu
sions
On the
ba
si
s
of origi
nal i
n
frared
an
d L
L
L
im
age
contai
ning
different
feature
info
rmation,
we uses
se
g
m
entation me
thod
to nat
urally extract
h
eat targ
et inf
o
rmatio
n in i
n
frare
d
ima
ge
and
mines mo
re d
epth detail
fro
m
LLL i
m
age
by enha
ncem
ent metho
d
. Acco
rdi
ng to
the se
gme
n
te
d
image, we fu
se the enh
an
ced LL
L ima
ge and the o
r
iginal infrare
d
image. In orde
r to optimize
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 3, September 20
14: 60
5 – 612
612
incom
p
lete
a
nd laye
ring
d
ue to the
in
complete
se
g
m
entation
or
over e
nha
ncement in t
he
first
fused
imag
e, the pa
pe
r u
s
e
s
the
se
co
nd fu
sed
me
thod to fu
se
the ori
g
inal i
n
frared im
ag
e ,
enha
nced L
L
L
imag
e an
d
the first fu
sed
image i
n
NS
CT d
o
main,
so the fu
sed i
m
age
ca
n bet
ter
kee
p
the main target information of the in
frared ima
ge and mine
more detail from LLL imag
e,
and
also
ha
s a b
e
tter visual effe
ct. Th
e expe
ri
ment
al re
sult
s
sh
ow th
at the
prop
osed
se
con
d
fused
imag
e
can
imp
r
ove
the re
sol
u
tio
n
an
d
cont
ra
st of the
ima
ge a
nd id
enti
f
y targets
bet
ter
comp
ared wit
h
other meth
ods.
Ackn
o
w
l
e
dg
ement
Authors woul
d like to than
k the Cho
ngq
ing
Educatio
n Committee
Scien
c
e of China for
sup
portin
g
the Found
ation
of prog
ram, No. KJ1305
29,
and
KJ140
04
34.
Referen
ces
[1]
Pan Y., Z
heng
Y., Sun Hua., et al. An ima
g
e
fusion b
a
se
d on pri
n
ci
pal c
o
mpon
ent an
al
ysis and tota
l
variati
on mod
e
l
.
Journal of Co
mp
uter-Ai
ded
De
sig
n
& Co
mputer Graph
ics
.
2011; 23(
7): 1200-
121
0.
[2]
Saee
di J., F
aez K.. Infrared and vis
i
bl
e imag
e fusion
usin
g fuzz
y
l
o
gic an
d po
pul
ation-
base
d
optimiz
ation.
A
ppli
ed Soft Co
mp
utin
g.
201
2; 12: 104
1-10
54
.
[3]
Brand
on M., Ismail B., Ma
x W
., et al. Spine im
ag
e fusi
on vi
a gra
ph
cuts.
IEEE Tr
ansaction on
Bio
m
e
d
ica
l
En
gin
eeri
ng.
20
1
3
; 60(7): 18
41-
185
0.
[4]
A.
T
oet. Image
fusion b
y
a rati
o of lo
w
-
p
a
ss p
y
ram
i
d.
Pattern Recognition Letters.
1993; 9:
245-2
53.
[5]
Jeese S., Michael A.
Lap
laci
a
n
base
d
imag
e
fusion
. Appl
ie
d Imager
y
P
a
ttern Reco
gn
itio
n W
o
rkshop
(AIPR), IEEE 3
9
th
. 2010; 23(7)
: 1-7.
[6]
Yuel
in Z
., Xiao
qia
ng L., T
ongming W
,
et al.
Visibl
e an
d infr
ared im
age fus
i
on us
ing th
e lif
ting
w
a
v
e
let.
T
E
LKOMNIKA Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
ng.
2013; 1
1
(11):
629
0-62
95.
[7] Z
heng
J.
,
Gong J., R
e
se
arch o
n
a
p
p
licat
ion
of
w
a
vel
e
t
transform
in
imag
e fusi
on.
Jour
nal
of
Cho
ngq
in
g Uni
v
ersity of Posts and T
e
lec
o
mmu
n
ic
ations.
2
008; 20(
5): 511
-567.
[8]
Liu K., Guo L
., Chen J. Co
ntourl
e
t transform for imag
e
fusion us
in
g c
y
cl
e spi
nni
ng.
Journa
l of
Systems En
gin
eeri
ng an
d Ele
c
tronics.
201
1, 22(2): 35
3-3
5
7
.
[9]
Yang
Y., Z
hu
M., He B, et
al. F
u
sio
n
alg
o
rithm b
a
se
d
on im
prov
ed
p
r
ojecte
d gr
adi
e
n
t NMF
an
d
NSCT
.
Optics
and Prec
isio
n Engi
neer
in
g.
2011, 19(
5): 114
3-11
50.
[10]
W
.
Kong, Y.
Lei, X. Ni. F
u
s
i
on tech
niq
ue fo
r gra
y
-scal
e
li
ght and i
n
frare
d
images b
a
se
d on no
n-
subsam
ple
d
c
ontour
let trans
form and
int
e
nsif
y
-
hu
e-satur
a
tion tr
ansfor
m
.
Signal Processing, IET.
201
3; 5(1): 75-
80.
[11]
Jian
hua L., Jia
ngu
o Y. Multifocus im
ag
e fu
sion b
y
SML i
n
the shear
let
subba
nds.
TEL
K
OMNIKA
Indon
esi
an Jou
r
nal of Electric
al Eng
i
ne
eri
ng.
2014; 12(
1): 618-6
26.
[12]
Meng
D., Li
W
., Bangfeng
W
.
Researc
h
on f
u
sio
n
m
e
thod
for i
n
fra
r
ed
and
visi
bl
e ima
ges
via
compressiv
e
s
ensi
ng.
Infrare
d
Physics&T
e
c
hno
logy
. 2
013;
57: 56-67.
[13]
Shutao L., Xud
ong K., Jian
w
e
n H.
Image fusion
w
i
th g
u
i
ded
filter.
T
r
ansaction o
n
Ima
ge Processi
ng.
201
3; 22(7): 28
64-2
875.
[14]
Qiao
X
., Han L., Wang B.
A fast fus
i
on
al
gorithm
of
visibl
e
an
d i
n
frared
ima
ges.
Jour
nal
o
f
Desig
n
&C
o
m
p
u
ter Graphics.
201
1; 23(7): 12
11-1
216.
[15]
Xi
ng S.,
Xi
ao
H., Chen
T
., e
t
al. Stud
y
of i
m
age f
u
sio
n
te
chno
log
y
bas
e
d
o
n
o
b
ject
e
x
traction
and
NSCT
.
Journal
of Optoeletron
i
cs·Las
er.
201
3; 24(3): 58
3-5
88.
[16]
Lia
n
fa B., Ying
bin
L., Jiang Y.
, et al.
Nig
ht vision i
m
ag
e fusi
on for targ
et d
e
tection w
i
th i
m
pr
ove
d
2
D
max
i
mu
m entr
opy
s
e
g
m
e
n
tati
on
. Intern
atio
n
a
l S
y
mp
osi
u
m
on P
hoto
e
lectr
onic
Detecti
on
and
Imagi
ng
201
3: Lo
w
-
Lig
h
t-Leve
l
T
e
chnolo
g
y
an
d Appl
icatio
ns. 201
3; 891
2: 891
20
x-
1-89
120
x-
8.
[17]
Che
n
Y., Ai A., W
ang J., et al. Automatic
parameter
op
timizatio
n
for lo
w
e
r l
a
yer im
age mi
nin
g
techni
qu
e.
Jou
r
nal of Optoel
e
c
tronics·L
a
ser.
2009; 2
0
(7): 9
50-9
53.
[18]
W
e
i
w
ei K., Jia
npi
ng L. T
e
chniqu
e for image
fusion b
a
se
d o
n
nons
ubs
amp
l
ed sh
earl
e
t transform and
improve
d
pu
lse
-
coup
led n
eur
a
l
net
w
o
rk.
Opti
cal Eng
i
ne
eri
n
g.
2013; 5
2
(1): 170
01-1-
01
70
0
1
-12.
[19]
L
u
Y., Gu
o
L
., L
i
H
.
R
e
mo
te se
n
s
in
g
i
m
a
g
e
fu
si
on
ed
ge
i
n
fo
rma
tio
n
a
n
d
fe
a
t
u
r
e
s
o
f
SAR
i
m
age
base
d
on curv
elet transform.
Acta Photon
ica
Sinica.
20
12;
41(9): 11
18-
11
23.
Evaluation Warning : The document was created with Spire.PDF for Python.