TELKOM
NIKA
, Vol. 11, No. 6, June 20
13, pp. 3012
~ 301
9
e-ISSN: 2087
-278X
3012
Re
cei
v
ed
De
cem
ber 2
7
, 2012; Re
vi
sed
March 15, 20
13; Accepted
April 11, 201
3
A Mixed Two-dimensional Linear Discriminate Method
Shuang Xu*,
Min Li, Yan
q
iu Cui
Information &
Commun
i
cati
o
n
Engi
ne
erin
g, Dali
an N
a
tion
al
ities Univ
ersit
y
,
Dalia
n, Chi
na
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: xus
hua
ngc
on
g@1
63.com
A
b
st
r
a
ct
F
eature
extra
c
tion is
on
e of
key techn
o
l
o
g
i
es of
the
pa
l
m
pr
int i
dentific
ation. In th
e li
ght of th
e
character
i
stics
subspac
e pal
mpr
i
nt ide
n
tific
a
tion te
ch
nol
o
g
y, the tw
o-dime
nsi
ona
l prin
cipal co
mpo
n
e
n
t
ana
lysis, tw
o-di
me
nsio
nal fi
sher li
ne
ar dis
c
rimina
nt an
d
tw
o-w
a
y tw
o-dimensi
o
n
a
l pr
i
n
cip
a
l co
mpon
ent
ana
lysis al
gori
t
hm is d
eep
ly
analy
z
e
d
. Ba
sed on tw
o-di
me
nsi
ona
l sub
s
pace p
a
l
m
pri
n
t identific
atio
n
alg
o
rith
m is a direct projecti
on
of the pal
mpr
i
nt imag
e matrix an
d is achi
eved very
good results for
di
me
nsio
n red
u
ction. T
h
is p
aper pro
pos
e
d
a mi
xe
d tw
o-dimens
ion
a
l
line
a
r discri
m
i
n
a
n
t dimen
s
ion
reducti
on al
gor
ithm w
h
ich ca
n eli
m
i
nate th
e releva
nc
e of
row
s
and columns to get the best proj
ec
tion
vector an
d extr
act opti
m
a
l
d
i
s
c
rimina
nt infor
m
ati
on.
Exper
i
m
e
n
tal r
e
sults
show
t
hat
the prop
osed
meth
o
d
has faster extraction sp
ee
d, high
er re
cog
n
iti
on rate an
d bet
ter robustness.
Ke
y
w
ords
: fea
t
ure extraction,
pal
mpri
nt ide
n
t
ificat
ion, tw
o-dimensi
o
n
a
l li
ne
ar discri
m
i
n
a
n
t
Copy
right
©
2013 Un
ive
r
sita
s Ah
mad
Dah
l
an
. All rig
h
t
s r
ese
rved
.
1. Introduc
tion
Feature extra
c
tion is one o
f
key technol
ogie
s
for palmprint re
cog
n
i
tion [1]. The measure
of a good or bad feature
extraction
method
s sh
o
u
ld take into
account the
best de
scrip
t
ion
feature a
nd the be
st cla
s
sificati
on featu
r
e, but also should
con
s
id
er the meth
o
d
to extract l
o
w-
dimen
s
ion
a
l feature. T
he
most typical f
eature
extraction method i
s
subs
pa
ce f
eature
extraction
method of princip
a
l com
p
onent an
alysi
s
(PCA
)
and
linear di
scri
minant analy
s
is (LDA
), which
conve
r
t the image matrix i
n
to vector
to realize feature
extraction [2]
.
Extractting th
e main fe
atu
r
es to PCA
method i
s
th
e de
scription
of the ima
g
e
feture,
while extra
c
ting the main feature
s
to LDA met
hod is the image cla
ssifi
cation fea
t
ure. Howeve
r,
PCA metho
d
exists ove
r
fitting p
r
oble
m
, and L
D
A
m
e
thod
will cau
s
e within
-class scatter m
a
tri
x
sing
ularity problem wh
en the number
of training
sa
mples is le
ss than the dimensi
on of the
feature ve
cto
r
, whi
c
h i
s
the so-calle
d
small
sam
p
le si
ze p
r
obl
em. In ord
e
r to solve
su
ch
probl
em
s, the first PCA is u
s
ed to dime
n
s
ion
a
lit
y redu
ction, and the
n
LDA tech
ni
que
s is ad
opt
ed
in the main l
o
w-dime
nsio
nal mole
cula
r spa
c
e.
Thi
s
is a PCA
+
LDA framework
unde
r sub
s
p
a
ce
linear di
scrimi
nant analy
s
is
techni
que
s of
dimensi
on re
ductio
n
[3] .
Ho
wever,
wh
en the n
u
mb
er of trai
ning
sampl
e
s
N
is
too exce
ssive, PCA and
LDA
need
co
nvert palmp
rint
image ma
trix into vector, resultin
g in solving larg
e ma
trix
comp
utationa
lly intensive.
So re
cog
n
itio
n nee
d to
con
s
ume
a lot of
comp
uting
re
sou
r
ces to tra
i
n
discrimi
nant
vectors, whi
c
h affe
ct the speed of cl
assification an
d identificatio
n efficien
cy. Two-
dimen
s
ion
a
l
princi
pal compon
ent a
nalysi
s
(2
DPCA) and t
w
o-dime
nsio
nal Fish
er l
i
near
discrimi
nant (2DFL
D
) [4,5] have solve
d
the problem
o
f
image matri
x
proje
c
tion v
e
ctor. T
hey a
r
e
dire
ctly proje
c
tion to palm
p
rint imag
e
matrix,
theref
ore, computi
ng re
so
urce
s con
s
um
ption
and
feature lo
ss a
r
e
small. 2
D
F
L
D i
s
al
so
kn
own
as
th
e two-dime
nsio
nal line
a
r di
scrimi
nant a
n
a
l
ysis
(2DLDA
) [6].
2DFL
D meth
od is directly
obtained fro
m
the im
age
matrix within-cl
a
ss
scatt
er matrix
and bet
wee
n
-cla
ss
scatter matrix image
matrix, wh
ich doe
s not nee
d to be conve
r
ted into a on
e-
dimen
s
ion
a
l vector. It ca
n be better t
han the L
D
A
method cl
a
ssifi
cation
ch
ara
c
teri
stics [6].
2DPCA meth
od is to extract the image comp
re
ssi
on
feature
s
, eliminate the co
rrelation of the
image col
u
m
n
s. But, there is still a hi
gh featur
e di
mensi
on with
out taking into account the
redu
nda
ncy row. In orde
r to solve the above m
entio
ned problem
s, it can eliminate the image
colum
n
co
rrel
a
tion while el
iminating the correl
at
ion of the image lines, i.e. bidirection
a
l 2DP
C
A
method [7, 8]. This metho
d
can
spee
d u
p
the re
cog
n
ition sp
eed, re
duce the feature dim
e
n
s
io
n ,
improve g
ene
ralization abili
ty and redu
ce
the computat
ional sto
r
ag
e resou
r
ces.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Mixed T
w
o-dim
e
sional Li
near
Discrim
inate Method
(Shua
ng Xu)
3013
PCA method, 2DPCA method or deriv
ed bi
dire
ction
a
l 2DPCA method are ba
sed o
n
prin
cipal
com
pone
nt analy
s
is m
e
thod. E
x
tracting
feat
ure b
a
sed on
prin
cipal
co
mpone
nt anal
ysis
is mainly d
e
scriptive featu
r
e of the imag
es, while
cla
s
sificatio
n
feat
ure i
s
u
s
ed i
n
the re
cogi
ntion
pha
se. Thu
s
on the basi
s
of the in-dept
h study
and solve the smal
l sample si
ze
problem PCA +
LDA, this pa
per p
r
op
so
ed
A Mixed Two-dim
e
n
s
iona
l Linear
Discriminate Meth
od (T
DL
D) T
he
conte
n
t can b
e
summ
ari
z
e
d
as the follo
wing two asp
e
cts:
(1) To
con
s
truct the overal
l scatter mat
r
i
x
of
the colu
mns an
d the overall scatte
r matrix
of the rows to solve ro
w proje
c
tion
matrix
and a
column p
r
oj
ection mat
r
ix, eliminate the
correl
ation o
f
the col
u
m
n
s a
nd
ro
ws, and extr
a
c
t bidire
ction
a
l 2DP
C
A su
bsp
a
ce featu
r
e
information.
(2) To
con
s
truct a ne
w within-cla
ss
sca
tter
matrix and betwe
en-cl
ass scatter
matrix in
bidire
ction
a
l
2DPCA
su
bspace to,defin
e a ne
w
cr
ite
r
ion fun
c
tion b
i
axial com
p
re
ssi
on o
n
the row
and column d
i
rectio
ns, an
d
extract discri
minant
features of bidirecti
onal 2
D
PCA
sub
s
p
a
ce .
2. Opera
t
ion Mode and E
v
a
l
uation Performan
ce of Palmprint Recognitio
n
2.1. Opera
t
ion Mode of P
a
lmprint Re
c
ognition
Figure 1
.
Stru
ctural
Diag
ra
m of Palmpri
n
t Recognitio
n
System
Structu
r
al dia
g
ram of palm
p
rint re
cog
n
ition sy
stem is sho
w
n in Fig
u
re 1. Any biometri
c
identificatio
n system
s incl
ude regi
stration and re
g
n
i
t
ion stage
s, and palmp
rin
t
identification
system
also i
n
clu
d
e
s
two
stages. In th
e
regi
stratio
n
st
age, first u
s
er name i
s
regi
stere
d
to o
b
tain
the palmpri
nt image throu
gh the palmp
rint acq
u
isit
io
n. Then it is to prep
ro
ce
ssi
ng to obtain the
effective ROI
of palmpri
nt image. Fina
lly, a us
er te
mplate is
created in the
databa
se af
te
r
extracting
the
feature
of the user. In th
e re
cog
n
ition
stage, firstly, palmp
rint feat
ure info
rmati
on
of the use
r
'
s
is accq
uired
,
and vario
u
s
palm
p
rint
f
eature i
n
form
ation is extra
c
ted. The
n
the
extracted fe
a
t
ure informati
on is mat
c
he
d with the
fe
ature info
rma
t
ion in the template library
to
identify user i
dentity. Finally, a fi
nal reco
gnition re
sult
is obtain
ed.
2.2. Ev
aluation Performa
nce of Palm
print Re
cog
n
ition
Palmprint acquisition process will be subject
to ma
ny uncertai
n
ties, so the palm of
the
same p
e
rson
at different times, in different
location
s which
colle
cted palmpri
nt image is no
t
exactly the
same. The
r
ef
ore th
ere
sh
ould b
e
an
evaluation
m
e
trics of p
a
l
m
print
re
cog
n
itio
n
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No. 6, June 20
13 : 3012 – 3
019
3014
algorith
m
whi
c
h is goo
d or bad. It usual
ly incl
ude
s three argume
n
ts indicators whi
c
h are Fal
s
e
Reje
ct Rate (FRR), False Acce
pt Rate (FAR) and G
enuin
e
Acce
p
t
Rate (GAR) to evaluate the
variou
s algo
ri
thms.
Input palmp
ri
nt feature
s
will do n
o
t exactl
y match
the regi
stered palm
p
rint
feature
s
durin
g the p
a
lmpri
n
t reco
gnition. If the matchi
ng d
e
g
ree i
s
gre
a
ter than a sel
e
cted thresh
old
value, the user will be seemed legitimat
e
user wh
i
c
h is accepted. When the matching degree is
smalle
r than
the sele
cted
threshold, th
e use
r
w
ill
b
e
con
s
id
ered
to be imperson
a
tor
whi
c
h is
rejected. FRR is the prob
ability which legitimate user is reje
cted as an impostor. FAR
is the
probability ref
e
rs to an impostor
as the l
egitima
te user is accepted. GA
R is the
percentage of
the corre
c
t numbe
r of samples an
d the total test
s
a
mple in the match. They can be calcul
ated
usin
g the followin
g
formul
a:
100%
NFR
FR
R
NG
A
(1)
1
00%
NF
A
FA
R
NI
A
(2)
10
0%
NCR
GAR
NTA
(3)
NGA re
prese
n
ts the Num
ber of Ge
nui
ne Acce
sses which is the
total numbe
r of the
legitimate user. NIA rep
r
e
s
ent
s the Nu
mber of Im
po
stor Acce
sse
s
whi
c
h i
s
the total numb
e
r of
the impo
ster
use
r
. NTA
repre
s
e
n
ts the
Numb
er
of T
e
st Acce
sse
s
whi
c
h i
s
the
total numbe
r
of
the test sam
p
les. NF
R st
and
s for the Numb
er of
False Reje
ctio
n. NFA stand
s for Numb
er
o
f
False A
c
cesses Numbe
r
of
False
Reje
ction. NCR
sta
n
ds for the
Nu
mber of Correct Re
co
gniti
on.
FRR a
nd FAR in the palmprint identifi
c
ati
on alg
o
rit
h
m can be a
s
small a
s
possible,
while GAR sh
ould be as large as possibl
e. FAR is
lower and GAR is highe
r, so the safety of
the
system is th
e higher. FRR is lowe
r and GAR is
hi
gher, so the
usa
b
ility of
the system is the
better. Ho
we
ver, the de
crease of
the
FRR i
n
the
actual
appli
c
ation sy
stem, will lead to
the
increase of t
he FAR, to the
contrary, the decrease
of the FAR will also accompanied the
increa
se in the FRR. Thu
s
FRR a
nd FAR ca
n
not be simultan
eou
sl
y reduced an
d will be alwa
ys
mutually re
st
ricting. T
h
e
r
efore, the
system
is
de
sign
ed to b
e
trad
e-offs FAR a
nd
FRR
relation
shi
p
. For sy
stems with high secu
rity r
equi
rements, such as military
systems, sh
ould
usu
a
lly redu
ce the FAR, and ease of use is more
importa
nt systems, ac
ce
ss control syste
m
s
s
u
c
h
as
c
i
vil s
h
ould be appropriate to reduc
e
the FRR.
In this paper for a comp
reh
ensive evalu
a
t
ion of
the overall pe
rform
ance of the p
a
lmpri
n
t
identificatio
n algorith
m
s, u
s
ing the follo
wing me
asure:
(1)
To
comp
are
different
palmp
rint ide
n
tification al
g
o
rithm
s
corre
c
t re
co
gnitio
n
rate, that t
he
accuracy of the re
cog
n
itio
n algorith
m
;
(2) Sy
stem resp
on
se time
of a variety
of algo
rithms,
which
the e
a
se uses of the
re
co
gnition
algorith
m
;
(3) ROC (Re
c
eiver O
pera
t
ing Cha
r
a
c
te
ristic) cu
rv
e of one of the criteria for t
he evaluation
o
f
biometri
c te
chnolo
g
y whi
c
h the curve
can refle
c
t the
GAR a
nd FA
R relation
s.
RO
C curve
of
the GAR a
n
d
FAR is
nea
rer the top
of the figure,
th
e system
pe
rforman
c
e i
s
t
he better. In
RO
C
curve
s
of the GA
R-FA
R, ab
scissa coo
r
din
a
tes of th
e
FAR ge
ne
ral
l
y are u
s
e
d
logarith
m
ic to
rep
r
e
s
ent, which
ca
n more sig
n
ificantly
reflect FA
R
accepta
b
le p
e
rform
a
n
c
e
near
ze
ro. Th
e indicator reflects the a
c
ce
ptability of th
e recognitio
n
algorith
m
.
The pro
p
o
s
ed algorith
m
mainly does perf
o
rma
n
ce evalu
a
tion arou
nd palmp
rint
recognitio
n
system accu
ra
cy, ease of u
s
e an
d acce
p
t
ability.
3. The Propo
sed Me
thod
2D dimen
s
io
n redu
ction method incl
u
des 2DP
C
A, 2DFLD and
bi-dimen
sio
nal PCA.
These metho
d
s are direct
ly to extract feat
ure to the image m
a
trix, which
can h
a
ve go
od
dimen
s
ion
a
lity reductio
n
effect.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Mixed T
w
o-dim
e
sional Li
near
Discrim
inate Method
(Shua
ng Xu)
3015
3.1. Bi-direc
tional 2DPCA Method
Bi-dire
c
tion
al
2DPCA m
e
thod i
s
to sol
v
e row
proje
c
tion mat
r
ix and
colum
n
proje
c
tion
matrix by ro
w total scatte
r matrix
and
column total scatter m
a
trix. Let
(1
,
2
,
,
)
i
X
iN
denote
s
the image feature matrix
()
mn
, where
m
is the numbe
r of pixels in the images afte
r down
sampli
ng an
d
n
is the numb
e
r of Gab
o
r
wavelet
s
ma
sks
whi
c
h i
s
the produ
ct of the numbe
r
of
orientatio
ns a
nd scale
s
.
The row total s
c
atter matrix
ro
w
i
C
is defined a
s
follows:
1
1
()
(
)
N
ro
w
T
ii
i
i
CX
X
X
X
Nm
(4)
Whe
r
e
,,
,
1
i
iv
X
X
Nn
is the glob
al me
an.
The
vect
or that maximize
s (4) is the
eigenve
c
tor
w
of
ro
w
i
C
corre
s
po
ndi
ng to the l
a
rgest ei
genval
ue. We
ne
e
d
to sele
ct a
set of
proje
c
tion
direction
s
,
12
[,
,
,
]
row
ro
w
r
o
w
ro
w
row
d
Ww
w
w
, whi
c
h a
r
e
orth
o
norm
a
l. In fa
ct, the p
r
oje
c
tion
dire
ction
s
,
12
[,
,
,
]
row
ro
w
r
o
w
ro
w
row
d
Ww
w
w
, are the orhto
n
o
rmal eig
env
ectors of
ro
w
i
C
co
rrespon
ding
to
the firs
t
row
d
larg
e
s
t eigenval
ue
s
ro
w
dn
.
The c
o
lumn t
o
tal s
c
atter matrix
co
l
i
C
is defined as follo
ws:
1
1
()
()
N
co
l
T
ii
i
i
CX
X
X
X
Nn
(5)
The o
r
thon
ormal projectio
n
matrix is
12
[,
,
,
]
co
l
co
l
c
o
l
co
l
co
l
d
Ww
w
w
and
col
d
is the nu
mber
of
the large
s
t ei
genvalu
e
to
co
l
i
C
.
From whi
c
h the image
s can be efficie
n
tly repre
s
en
ted along the
rows u
s
ing the
ro
w
W
eigenve
c
tor a
nd the col
u
m
n
s u
s
ing the
col
W
eigenve
c
tor a
s
:
T
co
l
r
o
w
YW
X
W
(6)
Whe
r
e
Y
is
a
row
c
o
l
dd
feature m
a
trix. Therefo
r
e, th
e bi-di
r
e
c
tion
al comp
re
sse
d
2DPCA
is a prin
cip
a
l comp
one
nt of row p
r
oje
c
tio
n
and column
proje
c
tion at the same tim
e
.
3.2. MTDL
D Metho
d
There are
L
known pattern
classe
s of trainin
g
sam
p
le set and
M
training sampl
e
s,
whe
r
e
i
M
the n
u
mbe
r
of tra
i
ning sampl
e
is set in
cl
asse
s
i
. Let the trainin
g
sampl
e
set
is
{
1
,
2
,
.
..
,
;
1
,
2
,
..
.,
}
,
mn
ij
i
i
j
X
xi
L
j
M
x
R
, and
ij
x
is the
jth training
sa
mple in cl
asses
i
. Let bi-
dire
ctional 2
D
PCA subspa
ce mapping se
t is
{
1
,
2
,
.
.
.
,
;
1
,
2
,
...
,
}
,
col
r
ow
dd
ij
i
i
j
YY
i
L
j
M
Y
R
, where
T
co
l
r
o
w
YW
X
W
.
T
h
e
be
tw
ee
n-
c
l
ass
s
c
a
tte
r
ma
tr
ix
B
S
and the within-class
sc
atter
matrix
W
S
in the
dire
ctional 2
D
PCA sub
s
p
a
ce are de
scri
b
ed by
1
1
1
()
(
)
1
()
()
L
T
Bi
i
i
i
L
TT
T
i
c
ol
r
o
w
i
i
c
ol
r
o
w
i
TT
col
r
ow
B
c
o
l
ro
w
SM
M
MW
W
W
W
M
WW
S
W
W
(7)
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No. 6, June 20
13 : 3012 – 3
019
3016
11
11
1
()
()
1
()
()
i
i
M
L
T
Wi
j
i
i
j
i
ij
M
L
TT
T
co
l
r
o
w
i
j
i
i
j
i
co
l
r
o
w
ij
TT
col
r
ow
W
c
ol
ro
w
SY
Y
M
WW
x
x
WW
M
WW
S
W
W
(8)
The optimal p
r
oje
c
tion vect
ors
()
J
of 2DFL
D shoul
d meet the followin
g
Fishe
r
criteri
o
n.
()
T
B
T
W
S
J
S
(9)
Substituting i
n
to the above
equation by (7 ) and (8 ):
()
TT
T
TT
c
o
l
r
ow
B
c
ol
r
o
w
B
TT
T
T
T
c
o
l
r
ow
W
c
ol
r
o
w
W
WW
S
W
W
S
J
WW
S
W
W
S
(10
)
In above form
ula
T
co
l
r
o
w
WW
, the formula (9) i
s
de
scribed a
s
:
()
T
B
T
W
S
J
S
(11
)
Therefore, th
ere is m
app
e
d
to:
T
co
l
r
o
w
WW
(12
)
Assu
med to
be the optim
al proj
ectio
n
matrix
12
[,
,
,
]
k
, the
proje
c
to
r feat
ure
matrix
for a gi
ven palmp
rint
image
is followen as:
12
1
2
[,
,
,
]
[
,
,
,
]
T
kk
c
o
l
r
o
w
WW
(13
)
4.2. Featur
e Classific
a
tio
n
In the testing
phase, the test sa
mple
s is pr
oj
ecte
d in the feature
sub
s
pa
ce, which the
proje
c
tion
of the test sa
m
p
les
X
is nam
e
d
as
, and th
e proj
ectio
n
of the trainin
g
sam
p
le
points
i
X
in the
feature sub
s
pa
ce i
s
re
corde
d
as
i
, then the no
rmalize
d
dist
ance
i
d
is
cal
c
ulate
d
be
tween the te
st sample
s an
d all the traini
ng sam
p
le po
ints,
i
d
can b
e
e
x
presse
d as:
1
2
,1
,
2
,
,
i
k
mm
i
m
m
di
M
(14
)
Find the
mini
mum di
stan
ce of the n
o
rmalize
d
di
sta
n
ce
i
d
between
the test
sam
p
le
X
and all of the training
sam
p
le points
i
X
is e
x
presse
d as:
mi
n
1
()
,
1
,
2
,
,
i
DM
i
n
d
i
M
(15)
S
i
milarly
,
t
he se
con
d
cat
e
g
o
ry norm
a
lize
d
distan
ce
j
d
be
tween the te
st sample an
d
th
e
training
sam
p
les an
d the same sa
mple
point ca
n be
expre
s
sed a
s
:
1
2
,1
,
2
,
,
j
k
mm
ji
m
m
dj
M
(16
)
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Mixed T
w
o-dim
e
sional Li
near
Discrim
inate Method
(Shua
ng Xu)
3017
Obtain
the
minimum dist
ance
j
d
betwee
n
the te
st sa
mple a
nd th
e traini
ng
sa
mples
from the sa
m
e
sampl
e
poi
nt is expre
s
sed as:
min
2
()
,
1
,
2
,
,
j
i
DM
i
n
d
j
M
(17
)
Comp
ari
ng two types of minimum dist
ance, if
mi
n
1
mi
n
2
DD
, the
res
u
lt is
c
o
rrec
t for
palmp
rint re
cognition; if
mi
n
1
mi
n
2
DD
, the re
sult is fa
lse for pal
mprint reco
gnitio
n
.
4. Results a
nd Analy
s
is
To redu
ce the computatio
nal co
st and the co
mputati
onal com
p
lex
i
ty, the resol
u
tion of
the palm
p
rint
image
s
crop
ped i
s
d
e
cre
a
se
d to 1
2
8
128. Du
rin
g
th
e expe
riment
s, the featu
r
e
s
are extracte
d
by usin
g th
e pro
p
o
s
ed
method
wi
th
different len
g
t
hs. The
wei
ghted Eu
clid
ean
distan
ce is e
m
ployed to cluster tho
s
e
featur
e
s
. All
experim
ents
were ca
rri
ed
out under t
he
Matlab7.0/PIV2.80GHz/1.
99GBRAM ex
perim
ental pl
atform.
Table 1. Co
m
pari
s
on
Re
sul
t
of Chara
c
teristics
Dime
nsi
on, Re
cog
n
ition Rate a
nd
Testing Ti
me
to a Varity of
Methods
Method
Feature
demision
Number of
cor
r
e
ct
Number of
error
Recognition
rate
Test time(s)
PCA 120×1
221
79
73.67
13.26
2DPCA
128×12
285
15
95.00
18.23
Bi-directional 2DPCA
10×10
290
10
96.67
5.25
PCA+LDA 96×1
283
17
94.33
8.64
2DPCA+LDA
256×1
292
8
97.33
15.11
TDLD
12×5
297
3
99.00
3.42
In PolyU-I pa
lmprint d
a
tab
a
se, u
s
in
g th
e first sta
ge
300 p
a
lmpri
n
t image a
s
a
training
sampl
e
set and the secon
d
stage
300
palmp
rint ima
ge as th
e test sample
s set, the reco
gniti
on
rate, test time and featu
r
e dimen
s
ion
s
to ac
hieve
high re
co
gn
ition rate a
r
e
obtained. T
he
experim
ental
results a
r
e sh
own in Ta
ble
1. Simula
tion results sho
w
that the feature dimen
s
ion
o
f
the prop
osed
method is lo
we
st relative
to other meth
od, whi
c
h ha
s only
1
2
5=
60
d
i
me
ns
io
ns
;
highe
st re
cog
n
ition rate rea
c
he
d 99.00%,
and sh
orte
st test time is ju
st 3.42 se
co
n
d
s.
Table 2. Co
m
pari
s
on
Re
sul
t
of the Reco
gnition Rate to Different T
r
aining Sam
p
les
in PolyU-I Da
tabase
Method
Training samples
1 2
3
4
5
Average
PCA 53.33
59.67
73.67
78.67
81.00
69.27
2DPCA
77.33
84.67
90.33
93.33
95.00
88.13
2DFLD
76.00
80.00
90.33
93.00
94.67
86.80
Bi-directional 2DPCA
87.33
90.67
95.00
97.33
98.67
93.80
PCA+LDA 83.33
89.00
94.33
96.00
96.67
91.87
2DPCA+LDA
89.00
92.67
96.67
98.33
99.00
95.13
TDLD
90.33
94.00
99.00
99.00
99.33
96.33
Evaluation Warning : The document was created with Spire.PDF for Python.
e-ISSN: 2
087-278X
TELKOM
NIKA
Vol. 11, No. 6, June 20
13 : 3012 – 3
019
3018
Tabble 3. Co
mpari
s
o
n
Re
sult of the Re
cog
n
ition Rate to Different
Traini
ng Sam
p
les
in PolyU-II Databas
e
Method
Training samples
1 2
3
4
5
Average
PCA 65.40
74.25
78.45
81.60
85.00
76.94
2DPCA
80.00
85.25
91.65
94.20
96.00
89.42
2DFLD
79.85
84.60
90.95
94.00
95.35
88.95
Bi-directional 2DPCA
89.65
91.60
95.35
97.00
98.15
94.35
PCA+LDA 82.80
90.75
93.85
96.25
97.40
92.21
2DPCA+LDA
91.00
93.40
95.65
97.80
99.00
95.37
TDLD
93.95
95.40
99.00
99.50
99.65
97.50
Palmprint
re
cog
n
ition i
s
a sm
all sam
p
le p
r
oble
m
. A different
algorith
m
to
sele
ct a
different num
ber of trainin
g
sampl
e
s in
the Po
lyU-I p
a
lmpri
n
t database i
dentification re
sults
are
given in Ta
b
l
e 2. Table
3 sh
ows recognition
re
su
lt to the different al
go
rith
ms 20
00 im
age
sampl
e
s to select a different numbe
r of traini
ng sa
mples in the
PolyU-II palmprint datab
ase.
Simulation re
sults
sh
ow th
at the more the num
be
r
of
training
sa
m
p
les, the
re
co
gnition rate is the
highe
r. Com
p
ared
with oth
e
r algo
rithm
s
, the propo
se
d
algorithm h
a
s
the high
est
recognitio
n
ra
te
in different nu
mber of traini
ng sam
p
le
s.
10
-2
10
-1
10
0
10
1
10
2
70
75
80
85
90
95
100
F
a
l
s
e A
c
c
ept
R
a
t
e
(
%
)
G
e
n
u
i
n
e A
c
c
e
pt
R
a
t
e
(
%
)
PC
A
+
L
D
A
2D
P
C
A
+
LD
A
TD
L
D
Figure 2. RO
C Dia
g
ra
m of TDLD M
e
tho
d
The experi
m
ent further compa
r
ed the
PCA+
L
D
A, 2DPCA
+
LDA, TDLD to identify
perfo
rman
ce.
Figure 2 reflects the RO
C cha
r
a
c
te
ri
stic curve
s
of GAR and FA
R relation
s. The
comp
ari
s
o
n
i
s
sele
cted 5
training
sa
mples i
n
the
PolyU-II dat
aba
se. Durin
g
the mat
c
hi
ng
process, each sample in the test sample se
t will be compared
with all samples so for each
sampl
e
for test sample
set
ther
e
will be
5 correctly m
a
tched and
495 mismatched. There
will
be
2500
corre
c
t matche
s an
d 2475
00 false
matche
s for
all the test sa
mples to com
p
lete the enti
r
e
experim
ent.
Corre
c
t match is correctly received GAR,
while error match is false received FAR. It
can be seen
from Figure 2, in
the same
false acc
ept
ance rate co
rrect re
ceive
d
rate of TDLD is
alway
s
highe
r than PCA+L
D
A, 2DPCA
+
LDA. Therefo
r
e the cha
r
a
c
tertics to rece
ive for TDLD
is
better.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
e-ISSN:
2087
-278X
A Mixed T
w
o-dim
e
sional Li
near
Discrim
inate Method
(Shua
ng Xu)
3019
In summa
ry, the experimental re
sult
s sh
ow
that
the propo
sed method
has fa
ster
recognitio
n
speed,
can
red
u
ce th
e featu
r
e di
men
s
io
n requireme
nts
and a
c
hi
eve
high recogniti
on
rate und
er lo
w feature di
m
ensi
on.
5. Conclusio
n
In terms of small sampl
e
probl
em
s whi
c
h exist in the sub
s
pa
ce p
a
lmpri
n
t identification
techn
o
logy a
nd the di
re
ct
irrel
e
van
c
e to
optimiz
e line
a
r di
scrimin
a
nt analysi
s
criterion fu
nctio
n
and the max
i
mization of the re
cog
n
itio
n rate,
this pape
r propo
sed a two-di
mensi
onal lin
ear
discrimi
nant
dimen
s
ion
re
ductio
n
algo
ri
thm whi
c
h m
e
rge
s
two di
mision
al pri
n
cipal
com
pon
ent
analysi
s
e
s
. This algo
rithm
is a direct projecti
o
n
of th
e palmpri
nt image matrix and is achiev
ed
very good re
sults fo
r dim
ensi
on re
du
ction whi
c
h can eliminate
the relevan
c
e of ro
ws and
colum
n
s to g
e
t the best projectio
n
vecto
r
and extra
c
t optimal discri
m
inant inform
ation.
Ackn
o
w
l
e
dg
ment
This wo
rks a
r
e partly sup
ported by the Sc
ience an
d
Technol
ogy Re
sea
r
ch Project of
Liaoni
ng Pro
v
incial De
partment of Ed
ucatio
n (Gra
nt No.L201
0
094 and L2
0
1247
9) and
the
Funda
mental
Re
sea
r
ch Fu
nds for th
e Centra
l Universities (Grant No. DC12
010
1132
).
Referen
ces
[1]
X
i
ao Q.
T
e
chnology
review
-b
iom
e
trics-te
chno
log
y
, a
p
p
licati
on,
cha
l
len
ge, an
d computati
o
n
a
l
intell
ig
ence s
o
l
u
tions.
C
o
mp
u
t
a
t
i
o
na
l
In
te
ll
i
gen
ce
Ma
ga
z
i
ne
. 200
7; 2(2): 5-2
5
.
[2]
Lu G,
Z
hang D, W
ang K. Palmpri
n
t recog
n
itio
n using e
i
gen
palms feat
ures.
Pattern Recognit
. 200
3;
24(1
0
): 146
3–1
467.
[3]
Mutelo RM, Khor LC, Woo WL, Dlay
SS.
A Novel Fish
e
r
Discrimin
ant for Biometrics Reco
gniti
o
n
:
2DPCA plus 2DFLD
. IEEE Internatio
nal S
y
mposi
u
m on Ci
rcuits and S
y
st
ems. 2006: 4
2
35-4
238.
[4]
Xi
on
g H, S
w
a
m
y
MNS, Ahm
ad MO. T
w
o-
di
mensi
ona
l F
L
D
for F
a
ce Rec
o
gniti
on.
Pattern Recognition.
200
5; 38(9): 11
21-1
128.
[5]
Z
uo W
M
, Z
h
a
ng D, W
ang K
Q. Bidirection
a
l
PCA
w
i
t
h
as
sembl
ed matri
x
dista
n
ce met
r
ic for image
recog
n
itio
n.
IEEE Transactio
n
s on Syste
m
s
,
Man,
and Cy
bern
e
tics, Part B:Cybern
e
tics
. 2006; 3
6
(4)
:
863-
872.
[6]
Cai D, He X, Han J, Z
hang HJ. Ort
hogona
l Lapl
acia
n faces for face recogn
ition.
IEEE Trans.Im
age
Processi
ng
. 20
06; 15(1
1
): 360
8-36
14.
[7]
Ming L, Yuan
B. 2D-LDA:
a statistical
line
a
r discrimi
nant an
al
ysis
for image matrix.
Pattern
Reco
gniti
on L
e
tter
. 2005; 26(5
)
: 527-53
2.
[8]
G. Lu, D.
Z
han
g and K.W
ang,
“Palmprint rec
ogn
iti
on us
in
g eig
enp
alms fe
atures,” Pattern Reco
gnit.,
vol.24, no.
10, pp. 146
3–
14
67
, 2003.
[9]
Lu JW
, Z
hang
EH, Kang
XB, Xu
e Y
X
, Che
n
YJ.
Palmpri
n
t recogniti
on u
s
ing w
a
vel
e
t d
e
co
mp
ositio
n
and 2
D
princ
i
p
a
l co
mpo
n
e
n
t ana
lysis
. Proc
eed
ings of Inte
rnatio
nal C
onf
erenc
e on Co
mmunicati
ons,
Circuits a
nd S
ystems. Guilin,
Chin
a: IEEE, 2006: 21
33-
213
6.
Evaluation Warning : The document was created with Spire.PDF for Python.