Professional Documents
Culture Documents
基于混合核wls Svr的古汉字识别 英文 胡根生
基于混合核wls Svr的古汉字识别 英文 胡根生
基于混合核wls Svr的古汉字识别 英文 胡根生
l.45,
No.
4
2
0 5年4月
1 JOURNAL
OF
UNI
VERS
ITY
OFSC
I
ENCE
AND
TECHNOLOGY
OF
CHINA Apr.2
0 1
5
文章编号:
0253-2778(
2015)
04-0321-08
c
Reon
git
ion
ofa
nc
ien
t Ch
ine
sec
ha
rac
ter
s b
ase
d o
n h
ybr
id
ker
nel
WLS-SVR
Gensheng1 ,SUN
HU Yingy
ing1 ,XU
Lingy
ing2 ,LIANG
Dong1 ,SUN
Xiaoq
i1
(
1.Schoo
l of
Ele
ctr
oni
c and
s Info
rma
ti Engi
on nee
ring,Anhu
i Un
ive
rsi
ty,Hefe
i 230601,Ch
ina;
2.Ed
ito
ria
l Depar
tmen
t of
Anhu
i Un
ive
rsi
ty,Hefe
i 230039,Ch
ina)
Abst
rat:The
c shapes of
ancien
t Chinese
cha
rac
ter
s are o
ften unce
r t
ain, whi
ch r educes
t he
ac
curac
y o
f r
ecogniti
on by many c
las
sif
ier
s.To so
lve
th
is pr
obl ,
em a new r
ecogni
tion algo
rithm
comb
ini
ng
adap
tive we
igh
ted
le
ast
squa
res
suppo
rt ve
cto
r r
egr
ess
ion(WLS-SVR)wi
th hybr
id
ke
rne
l f
unc
tion
wa pr
s opos
edt
or
ecogn
ize
anc
ien
t Ch
ine
secha
rac
ter
s.The
wei
ght
coe
ffi
cien
tso
f
WLS-SVR
dec
ayed
ata
r
ate
oft
he
exponen
tia
l f
unc
tion
ofpr
ed
ict
ion
err
ors.The
hybr
id
kerne
l
wascons
t
ructed
usingt
he
wavel
etke
rnel
func
tion with
local
proper
tiesand
RBF kernelf
uncti
on
wi
th gl
obalpr
opert
ies.For
fe
ature
extr
acton,g
i loba
l poin
t densi
ty and componen
t structur
e are
f
used
wit
h l
oca
l f
eat
ure
s o
f ps
eudo
2De
las
tic
mesh
and
loc
alpo
i
ntdens
ity.Expe
rimen
t r
esu
lts
show t
he good
robust
nessand
hi
gh re
cognit
ion a
ccur
ac o
y ft
pr
he oposed me
thod.
Key wors:
d anci
ent
Chine
se cha
racte
rs re
cognit
ion;WLS-SVR;hybr id ke l;f
rne ea
tur
e f
usi
on
CLC r:
numbe TP18 Do
cumen
t ode:
c i:
A do 3969/
10. j.i
s 0253-2778.
sn. 2015.
04.
010
i
Cta
ti Gensheng,SUN
on: HU Yingy
ing,XU
Lingy
ing,e
t a
l.Re
cogn
iti
on
ofanc
ien
t Ch
ine
secha
rac
ter
s ba
sed
on
hyb
rid
ker
ne WLS-SVR[
l J].J
our
nal
of
Uni
ver
sit
y o
f Sc
ienc
e and
Techno
l o
ogy f
Chna,2015,
i 45
(
4):
321-328.
胡 根生,孙莹莹,徐玲英,等 .基于混合核 WLS-SVR 的古汉字识别[
J].中国科学技术大学学报,
2015,
45(
4):
321-328.
摘要:针对现有多种分类器对具有不确定字形的古汉字识别精度不高的问题 ,提出了一种基于混合核加权最
小二乘支持向量回归(WLS-SVR)的 古 汉 字 识 别 算 法 .WLS-SVR 的 权 重 系 数 采 用 预 测 误 差 的 指 数 衰 减 函
数,混合核是由具有良好局域特性的小波核函数与具有良好全局特性的 RBF 核函数构成 .在特征提取阶段,
由于全局点密度与部件结构具有全局特征,而伪二维弹性网格与局部点密度具有局部特征,因此融合了古汉
字的全局和局部特征 .仿真实验表明,该算法具有较高的准确率与良好的鲁棒性 .
关键词:古汉字识别;WLS-SVR;混合核;特征融合
Received:
2014-06-10;
Rev
isd:
e 2014-12-29
Founda
tion
iem:
t Support
ed byt
he Na
tiona
l Na
tur
alSc
i
enc
e Founda
tion
of
Chna (
i 61172127),Na
tur
alSc
i
enc
e Founda
tion o
f Anhu
i
Provi
nc ( )
e 1408085MF121 .
Bi
ogaphy:
r Hu Gensheng(c
orrespondi
ng author),male,born n 1971.PhD/ a
i ss
oci
ate p
rof
ess
or.Re
sea
rch
fi
eld: Ma
chi
ne l
ear
ning,
r
emotes
ens
ing
image pr
ocess
ing and
int
ell
igen
t al
gor
ithm.E-mai :
lhugs 2906@s
ina.
com
322 中国科学技术大学学报 第 45 卷
i
n f
iel
ds f
o l
cas
sif
ica
tion,f
unc
tion e
stima
tion,
0 I
ntr
odu
cti
on dens
it est
ima
tion
and
so
on.SVM
needsto
sol
ve
y
Anc
ient Chinese char
acters reco
rded a large a
mode
l o
f quad
rat
ic
prog
rammi
ng.Suykens
et
al.
amounto
f pol
it
i c
al,e conomic and histor
ical
data propos
ed l
eas
t squa
res suppo
rt vecto
r ma chi
ne
so on,and
and thus have a very high his
tori
cal (LS-SVM)by us
ing
squar
e l
ossf
unc
tioni
ns
tead of
[
6-8]
va
lue. Anc
ien
t Ch
ine
se cha
rac
ter
s r
ae lways
a εi
nsens
iti
ve l
oss
func
tion .LS-SVM r
educ
es
appea
red i n
t he forms of insc
rit
p i
on and t
hecompu
tat
iona
l comp
lex
it by
y sol
ving
a mode
l o
f
handwri
ti ,
ng thes
e char
act
er ’
s st
rokes are ve
ry l
inear
pr
ogramming.Al
though LS-SVM
hasso
l
ved
di
ffe
r ent
from t
he cur
rent pri
n ted cha
rac
ter
s. the
comput
ati
ona
l comp
lexi
t o
y f SVM,it
los
est
he
Furthermo ,
e many ex
r is
ting ancien
t Ch
ines
e spa
rsi
t and
y robus
tne
sso
fSVM
[
9-12]
.Th
i pape
s r
cha
rac
ter
s a
re de
formed o
r i
ncomp
lee, mak
t ing pr
opos
es a
cl
ass
ifi
cat
ion a
lgo
rit
hm t
hat
comb
ine
s
t
hem
hard
to
re
ad
orr
ecogn
ize.The
ref
ore
it
is
of adap
tive we
igh
ted
le
ast
squa
res
suppo
rt ve
cto
r
r
gea
t s
ign
ifi
canc
e o
t e
rcogn
ize
anc
ien
t Ch
ine
se r
egr
ess
ion
wit
h hybr
id
kerne
l f
unc
tion
fo
r anc
ien
t
cha
rac
ter
s and r
eal
ize
the d
igi
tal managemen
t o
f Ch
ine
se cha
rac
ter
re
cogn
iti
on. The pr
opos
ed
c
las
sic
al Ch
ine
se i
lte
rat
ure by us
ing mode
rn algo
rit
hm ha
s the advant
ages of hi
gh robustness
[
1]
r
ecogn
iti
ont
echno
logy . andl
ow comput
a t
ionalcomp
l
ex i
tyt, hus
impr
ov i
ng
[]
e
Lv ta
l. 2 pr
opos
edt
he
Four
ier
de
scr
ipt
or- t
hea
ccur
ac o
y fr
ecogn
iti
on.
FDCH ba
sed on cur
vat
ure h
ist
ogr
am t
o c
las
sif
y
ins
cri
pti
ons on bones.Si
nce t
he i
nscr
ipti
ons on 1 Hyb
rid-ke
rne
l WLS-SVR
bones are pic
tograph
ic, t
he curva
ture based LS-SVR can be
expr
ess
ed a
s t
he f
oll
owi
ng
met
hod proved to be prac
tica
l. However,the op
timizat
i pr
on obem:
l
a
lgor
ithm i
nclude
s the se
arch for
a cha
ract
e ’
r s N
1 T 1
nJ(
mi ω,e)= ω ω + γ∑ei
2
c
ente
r of
gravi
ty.The r
ecogn
iti
on rat
e i
s high
for 2 2 i=1
s
ing
le s
tructur
e cha
rac
ter ,
s butf
or
the
le
ft-r
ighto
r t.yi = ωTφ(
s. xi)+b+ei, i =1,…, N ( 1)
up-down
struc
tur
e cha
rac
tes,t
r her
ecogn
iti
on
rat
e whereω i
s t
he weighted vec
tor,γis t
he ba
lance
i
s l
ow. cons
t t,
an φis a
map f
unc t
i andei(
on i=1,…, N)is
[]
e
Chen t l. 3 pr
a opos a me
ed thod
to ex
tra
ct t
hee
st
ima
tion
err
orf
or
the
it
h s
amp
le.Eq.(
1)
f
eat
ure
s o
f i
nsc
rit
pions
on
bone
s ba
sed
ont
he
cr
oss c
an be
conve
rted
in
to he
t o
fll
owi
ng f
orm by
poi
nts of s
troke
s and the r
elat
ive posi
tions
o f Lag
r ange mu
lti
pli
er and ma
tri
x r
tans
forma
tion
char
acte
rs.Howeve r,it
is d
iff
icul
t to r
ecogniz
e method :
cha
rac
ter
s wi
th a pa
rti
cul
arl
y l
arge numbe
r o
f 1vT
熿0 燄b 0
s
troke
s us
ing
such
fe
atur
es.
1
1
燀v Ω +γ 燅α
I
=
y
[ ] [] (
2)
Many chara
cter
isti
cs o f
anci
en t Ch
ines
e
y= [y1 ,…,
yN ] i r,
T
chara
cte
rs, such as i r
regu
lar str
oke s, var
iant where s
samp
le
out
put
ve
cto
forms,def
ormity or
incompl
etene
ss,i ndic
ate
that 1v = [
1,…,
1] T
,I = d ag{
i 1,…,1}, α =
t
he opti
cal OCR system for
the
recognit
ion of [a1 ,…,aN ]
T
is Lagrange mul
til
pir, and Ω i
e s
modern Chine
se cha
ra c
ter
s i ’
sn t suit
able fo
r kernel
matrix,Ωij = φ (
xi)φ(
T
xj)= K(xi,
xj)fo
r
[]
r
ecogn
izi
ng
anc
ien
t Ch
ine
secha
rac
tes4 ,wh
r ich
is i,
j =1,…,
N .By
sol
vi Eq.(
ng 2),t
he
pred
ict
ion
a d
iff
icu
lt pr
obl
em i
n he
t i
fel
d o
f pa
tte
rn f
unc
tion c
an be ob
tai
ned and ha
s t
he f
oll
owi
ng
r
ecogn
iti
on. f
orm:
Suppor
t vector machne(
i SVM ),pr opos
ed by N
[
5]
, y(
x)= ∑aK(x,x )+b
i i (
3)
Vapn
ik e
t al. has ood
g g ene
ral
iza
tion t
abi
liy i=1
[
13]
and
le
arn
ing
per
formanc
e.I
t ha
s be
en
wide
l us
y ed The ba
sic
ide
a o
f WLS-SVM i
s t
hat
a
第4期 Re
cogn
iti
on
ofanc
ien
t ch
ine
secha
rac
ter
s ba
sed
on
hyb
rid
ker
nel
WLS-SVR 323
N
we
igh
ted
fa
cto
r ωi wi
ll i
be gven
to
the xi -xi
′)
Kwav(
x,x′ )= ∏h( =
co
rre
spond
i e
ng rr
orva
r
iab
leei o
f e
ach
samp
lexi .
i=1 α
N
The
opt
imi
zat
ion
prob
l change
em s t
o (
xi -xi )
' 2 ' 2
‖xi -xi‖ ) ( )
N
∏ (1-
i=1
2
ai
)
exp(
- 2
2ai
8
1 1 烌
nJ (
mi ω* ,
*
e* )= 2
‖ ω* ‖ + C∑ ωiei*2 wh
i s good
ch ha loc
al pr
ope
rti
es. Wave
let
ke
rne
l
2 2 i=1 烍
f
unc
tion
is
sens
iti
vet
ol
oc
als
ingu
lar
iti
es.
t.yi = ω*Tφ(
s. xi)+b* +ei* ,
i =1,…,
N 烎
A
sing
le
kerne
l f
unc
tion
is
re
str
ict
edi
n
aspe
cts
(
4)
o
f pr
edi
cton a
i ccur
acy and gene
ral
iza
tion. Th
is
Usi
ng the Lag
range mu l
til
pier method and [ ]
pape
r cons
truc
tsa
hybr
id
ke l 14 by
rne comb
ini
ng
a
cco
rding
tot
he KKT cond
iti ,
ons t he
dua
l problem
wavel
et ke
rne
l f
unc
tion
wit
h RBF
kerne
l f
unc
tion
o
f Eq.(
4)c
an
beexpr
ess
ed
as:
a
s fo
llows:
T
r
(
5) K(
xi,
x)=βK wav(
xi,
x)+ (
1-β)
KRBF(
xi,
x)
(
9)
1 1
ag{ ,…,
Cω }
whe
re t
he mat r
i V =d
x r i .In whe
reβi
s a
wei
ght
edf
ac
tor.
Cω 1 N
t
his r,t
pape he
wei
ght
edf
ac
torωii
s cons
truc
ted
as:
e
2 Fe
atu
ree
xt
rac
tion
i
ωi =e -
s (
6)
Duet
ot
he
randomn
e ss
and i
rre
gular
it o
y fan
c
ien
t
IQR
whe
res=
.IQR i
s he qua
t rti
le o
f Ch
ine
se cha
rac
ter ,
s recogni
tion base
d on a si
ngl
e
2×0.6745
s
amp
lee
rr
or.The
propos we
ed i
ght
edf
ac
tor
de
cays f
eat
ure
le
ads
to
a h
igh
rat
e o
f mi
scl
ass
if
ica
tion
s.Mu
lti
-
wi
tht
he
exponen
tia
l f
unc
tion
ofpr
ed
ict
ion
err
ors. f
eat
ure
fu
sion
can
opt
imi
zef
ea
tur
e v
ect
ors
and
imp
rov
e
[ ]
We
can s
ee r
fom F
ig.
1 t
hat
la
rge
r e
rro
r r
ecogn
iti
on
rat
es 15 .
co
rre
sponds
to
sma
lle
r we
igh
t.Thus
the
ef
fec
t o
f Gl
oba
l f
eat
ure
s a
re
not
se
nsi
tiv
e t
o ima
ge
noi
se,
t
hee
rr
orand
noi
set
ot
he
mode
l c
an
ber
educ
ed. handwr
it
ing d
efo
rma
tion and s
cal
e v
ari
ati
on.Lo
cal
f
eat
uresc
an d
ist
ingu
ish
anc
ien
t Ch
ine
sec
ha
rac
ters wi
th
s
imil
ar str
ucture
s. There
fore,the c
ombina
tion of
l
gob
alf
ea
tur
es
wit
h l
oca
l f
eat
ure
s c
an
gua
ran
tee
the
r
obu
stn
ess
and a
ccu
rac
y o
f r
ecogn
iti
on o
f an
cie
nt
Ch
ine
sec
ha
rac
ter
s.
S
truc
tur
e f
eat
ure
is
one
oft
he
globa
l f
eat
ure
s.
S
troke
st
ruc
tur
e a
s a
st
ruc
tur
e f
eat
ure ha
s be
en
wi
del
y us
ed i
n mode
rn Ch
ine
se cha
rac
ter
r
ecogn
iti
on.Strokeso
f anc
ien
t Chi
nese cha
ract
ers
F
ig.
1 S
chema
tic
dia
gram
oft
he
re
lat
ion
shi
p a
re ma
inl cur
y ,
ved and t
hecur
va
tures
ar
e di
ffe
rent
b
etwe
en
wei
ght
edf
ac
tor
s and
pre
dic
ti e
on rr
ors evenf
or t
he str
okes o
f the same cha
rac
ter.The
The pe
rfo
rmanc f WLS-SVM i
e o s g
rea
tly dir
ect
ional
fea
tures
such as hori
zonta
l l
ine,t
op-
i
nfl
uenc
ed
byt
he
se
lec
tion
ofke
rne
l f
unc
tion.RBF down
ver
tic
al l
ine,l
eft-downwa
rd s
lope
li
ne and
ke
rne
l f
unc
tion
is
def
ined
asf
ol
lows: sho
rt paus
ing s
troke
ar
e no
t su
itab
le f
or
‖x -xi‖ )
2
recogn
iti
on of anci
ent Chine
se char
acters. This
KRBF(
xi,
x)=exp(
- 2
(
7)
2σ paperex
tra
cts
four
kinds
of
components
truc
turs,
e
RBF ke
rnel
func
tion has good g lobal name
ly l
eft-r
i t,
gh up-down, i
n-ou
t and
pr
ope
rti
es,so i
t has
a st
rong
lea
rni
ng ab
ili
ty fo
r i
ndependenc
e a
s s
truc
tur
e f
eat
ure
s. Componen
t
ad
jacent
s ample
s. Wave
let
ke
rne
l unc
f tion
is s
truc
tures o
f anc
ient Chines
e chara
cte
rs are not
de
fined
asf
olows:
l s
ensi
tive
to
noi
se and
sca
le vari
ati
on.Besi s,t
de he
324 中国科学技术大学学报 第 45 卷
l
goba
l po
int
dens
ity
fe
atur
e ha
s a
st
rong
ant
i-no
ise e
las
tic
mesh
can
wel
l adap
t t
o t
hel
oc
alde
f
orma
tion
capabi
li
ty and c
an adapt to handwri
ting and d
ist
ingu
ish anc
ien
t Ch
ine
se cha
rac
ter
s wi
th
deformat
ion.So,two
global
fea
tures,component s
imi
lar
st
ruc
tur
es.So
four
kinds
of
fe
atur
es
have
s
truc
tur
e f
eat
ure
and
globa
l po
int
dens
ity
fe
at e,
ur been extr
act
ed f
or r
ecogn
ition of
anc
ien
t Ch
ines
e
a
re s
ele
cted
to
re
ali
ze a
rough c
las
sif
ica
tion o
f chara
c t
ers:component
structur
e fea
t e,g
ur lobal
ancien
t Ch i
nes
e characte
rs.I n orde
r t o
real
ize po
int
dens
it f
y e
at e, ps
ur eudo 2D e
las
tic me
sh
accurat
e r
ecogni
tion
ofanc
ien
t Chi
nese cha
rac
ters, f
eat
ure
and
loc
alpo
i
ntdens
ity
fe
atur
e.
l
goba
l f
eat
ure
s a
re comb
ined wi
th l
oca
l f
eat
ure
s. The
s
chema
tic d
iag
ram o
f componen
t
The
loc
al po
int
dens
ity o
f e
ach p
ixe
l wi
thi
n t
he s
truc
tur
e f
eat
ure
ex
tra
cti
oni
sshown
in
Fig.
2.
F
ig.
2 S
chema
tic
dia
gram
ofc
ompon
ent
st
ruc
tur
e f
eat
ure
ex
tra
cti
on
l
The gobal poin
t density of
the cha
r acte
r whe
reσ2i
s t
he
var
ianc
e o
f Gaus
sian
fuz
zyf
unc
tion,
[16]
image
aft
er
bina
rizat
ion i
s de
fined
as f
olows :
l u,
vrepr
esen
tst
he
row
orco
l
umn
oft
he
pi l,u0 ,
xe
n n
v0i
s a
part
icula
r r
ow orco
umn,du (
l u,
v)or
dv(
u,
∑ ∑f(
i,j)
i=1 j=1 v)is the
i n
terva
l dens
it f
y unc
tion of
st
roke
s.
α= 2
,1 ≤ n ≤ 128 (
10)
n
Accumu
lati
ng loc
al f
uzz
y line
ar dens
it f
y unc
tions
wher
e f(
i,j)repre
sentst
he
pixel
val
ue a
t pos
iti
on
by
rowso
r co
lumns,we have
(
i,j),n repr
e s
ents t
he image si
ze andαi s
the
x
l
goba
l po
int
dens
ity. 烌
H(
x,v0)= ∑ρ (u,
v ),x =1,
x 0 2,…,
U
u=0
Pse
udo 2D el
asti
c mesh is sui
tab
le f
or a
la
rge
y
烍
numb
ero
fv
ari
ant
s cause
d by d
iffe
rent
writ
er ,
s wh
ich S(
u0 ,
y)= ∑ρy(u0,
v),y =1,
2,…,
V
u
ses
lo
cal
fu
zzy
l
ine
ard
en
sit
y f
unc
tion
in
ste
ad
of
glob
al v=0 烎
d
ens
ity
pro
jec
tion
fun
cti
ont
o
obt
ain
a ab
good s
ort
pion (
12)
c
apa
cit
y f
or l
oca
l e
dfo
rma
tion o
f an
cie
nt Ch
ine
se The genera
tion f
uncti
ons of t
he ps
eudo 2D
[ ]
c
haract
ers17 .Lo
cal
fu
zzy
li
nea
r d
ens
ity
fun
cti
ons
ar
e e
las
tic
mesh ar
e defi
ned a
s fo
llows:
de
fin
ed as
fol
lows: fh(
m/M, v0)= {
x|H(x, v0)= (
m/M)H( U,v0)}
烌
V 烍
1 -(
2
v-v0)/(
2烌
σ ) fs(
u0 ,
n/N)= {y|S(
u0 ,
y)= (n/N)
S(u0 ,
V)} 烎
ρ(
u,v0)= ∑d (u,
v) 2
e
u
v=1 槡2πσ (
13)
U 烍
1 -(
2
u-u0)/(
2
σ )
where U,Vrepres
entst
he width
and he
igh
t o
f an
ρ(
u0 ,
v)= ∑d (u,
v) 2
e
v
u=1 槡2πσ 烎 image o
f a anc
ient Chi
nes
e char
act
er,M and N
(11)
第4期 Re
cogn
iti
on
ofanc
ien
t ch
ine
secha
rac
ter
s ba
sed
on
hyb
rid
ker
nel
WLS-SVR 325
repre
sentt
he numbero
fr
ows and
col
umns
of
the St
ep 2 Pr eproces
sing.S i
nce t
he input
image
ps 2De
eudo la
stic
mesh,respe
cti
vel
y. i a who
s le page of
anc i
ent Chinese
cha r
act
ers,
The
loc
all
ine
ardens
ity
func
tion
ate
ach
row char
acter
segmen t
ati
on mus t
bec
arr
ied out.Then,
o
r co
lumn
is d
iff
eren
t.The me
sh l
ine
is
cur
ved in
orde
r t
o r
emove no
iseand
f
aci
li
tate
the f
eatur
e
ra
ther than str
aigh
t whi
ch can't
be de
s c
ribed by extr
act
i i
on n l
ate
r stage, we need
t o conduct
l
inea
r equation.He e,t
r he expr
ess
ion
of mesh(m,n) b
ina
riz
aton,
i no
rma
liz
ati
on and deno
isi
ng
c
an be conver
ted
to
the po
in coo
t rdi
nat
e s
et o
f ope
rat
ions
to
cha
rac
ter
s.
mesh
li
nes : S
te 3 Fe
p atur
e ex
tra
cti
on.S
inc
e i
t i
s d
iff
icu
lt
x1(
m)= {
(fh(
(m -1)
/M),
v0),
v0)|v0 =1,
2,…,
V}烌 fo
r t he
s i
ngle fea
t ur
e t o
f u
ll r
y ef
lect t he
m ), ) i
nformati
on of
anc
ient Chi
nese chara
cte
r ,
s g lobal
x2(
m)= {
(fh( ,
v0 v0 |v0 =1,
2,…,
V}
M f
eat
ure
s a
re us
f ed wi
th oc
l a
l e
fat
ure
s. Mu
lti
-
n 烍
y1(
n)= { u0, -1),
fs( u0)|u0 =1,
2,…,
U} f
eat
ure
fus
ion c
an r
efl
ect
anc
ien
t Ch
ine
se
N
cha
rac
ter
s f
rom
all
si
des.
n
y2(
n)= {
fs(
u0, ),
u0)|u0 =1,
2,…,
U} S
tep 4 Cl
ass
ifi
cat
ion.The
tr
ain
ing s
amp
le
N 烎
(
14) f
eat
ure
s a
re
used
to
tr
ain
the
WLS-SVM
cla
ssi
fie
r
ps
The eudo
2De
las
tic
mesh
can
be
obt
ained mode
l f
irs
tly.Then
the
te
st s
amp
le f
eat
ure
s a
re
by
conne
cti po
ng i
ntcoo
rdi
nat
eso
fe
ach
mesh
li
ne. i
npu
t t
o de
cide
the pa
rame
ter
s o
f a
cl
ass
ifi
er.
Local
poin
t dens
ity of each p
i l wi
xe thi
n t
he F
ina
ll anc
y ien
t Ch
ine
secha
rac
ter
s a
rer
ecogn
ized
as
e
las
tic
meshi
s defi
ned asf
olows:
l shown
in
Fig.3.
x+1 y+1
∑ ∑f(u,
u xv
v)
B = ( = =y)( (15)
ux+1 -ux vy+1 -vy )
whe
rex ∈ fh (
0,M ),y ∈ fs(
0, N),Br epr
esents
t
hel
oc
alpo
i
ntdens
it ma
y t
rix.
Th
is pape
r f
irs
tl f
y us
es l
oca
l po
int
dens
ity
f
eat
ure and
t he pseudo 2D el
ast
ic me
sh f
eat
ure.
Thef
us
ion express
ion is
as
fo
llows:
P =λ1BI0 +λ2BI90 +λ3BI45 +λ4BI135
(
16)
whe
re I0 repres
ents the
left-r
ight
direct
ion,I90
r
epre
sents the up-down di
rec t
ion,I45 r
epres
ents
the
l ower-l
eft
direc
tion and I135 r
epre
sents t
he
upper-r
ight di
rec
tion,λi(i = 1,…, 4)repr
esents
we
igh
t coe
ffi
ci t,name
en lyt
he
ra
tio
ofs
tr
oke
s o
f
e
ach
dir
ect
ion
to
anc
ien
t Ch
ine
secha
rac
ter
s.
Then
the
fus
ed e
fat
ure
is
comb
ined wi
th
componen
t t
sruc
tur
e e
fat
ure
and g
loba
l po
int
dens
it f
y e
atur
e t
o ob
tai
n t
he f
eat
ure ve
cto
r f
or
r
ecogn
izi anc
ng ien
t Ch
ine
secha
rac
ter
s.
3 S
tep
s o
f r
eco
gni
tion
F
ig.
3 Re
cogn
iti
on
pro
ces
s o
f anc
ien
t cha
rac
ter
S
tep
1 I
npu
t image.Pape
ryimage
s shou
ld
be
conve
rted
in
to
ele
ctr
oni
c one
s by
a c
ame
ra.
326 中国科学技术大学学报 第 45 卷
Tab.
1 Th
e f
eat
ure
va
lue
s c
orr
espond
ing
to
4 Expe
rimen
tsand
re
sul
tsana
lys
is t
hef
ou
r c
ompon
ent
st
ruc
tur
e f
eat
ure
s
Pi
ctures we
re taken and a
char
acter
li
bra
ry f
eat
ure
val
ue
wa
s bui
lt f
r om “Ancien
t Lao Tze Words”.The up-down l
eft-r
i
gh
t s
ing
le i
n-ou
t
sel
eced anc
t i
ent Chine
se char
act
ers
incl
ude four
10 20 30 40
component
struc
tur
e feat
ures,namel l
y eft-r
ight,
up-down,i n-out
and independenc
e. Cha
rac
ter
s Due
to
the
limi
ted numbe
r o
f s
amp
les,t
he
chosen
fo
r t
he expe
r
imen
ts are 國,及,久,仁,什, s
amp
les
et
of
ea
ch
componen
t s
truc
tur
e i
s d
ivi
ded
守,吾,也,大,各,多,
a t
ota
l o
f e
leven
cha
rac
ter i
nto
thr
ee c
atego
ris,name
e ly e
asy, med
ium and
s
ets,each
seti
ncl
uding several
va
riantsubs
e
ts.If comp
lex a
cco
rdi
ng t
o t
he g
loba
l po
in dens
t ity
e
ach va
rian
t subs
et i
s regarded
asa
separa
tec
la
ss, f
eat
ure.The d
ivi
sion pa
rame
ter
s a
re shown
in
t
her
e a
re a
to
tal
of
th
irt
y-e
igh
t c
las
ses. Four Tab.2,whe
reai (
i=1,
2)r
epr
esen
tsc
la
ssi
fic
ati
on
s
amp l
eso
fe
ach c
lass
in
the
thir
ty-e
ight
cl
asse
s a
re coe
ffi
cien
t t
hat
is
a s
imp
le d
ivi
sion o
f d
iff
icu
lty
s
ele
ctedf
or
tra
ining,two for
test
ing
and the
res
t deg
ree;Δi (
i=1,
2)i
s t
head
us
j tmen
t f
act
ort
ha
t i
s
f
orr
ecogn
iti
on.Fig.
4showst
hree
var
iant
s of“及” s
ett
o
avo
id
mis
cla
ssi
fic
ati
on
and
le
akage.
a
fte
r operat
ions
of bi
nar
iza
tion, deno
ising and Tab.2 Th
res
hol
dso
fd
iv
isi
on
par
ame
ter
s
co
rros
i expans
on ion. o
f l
goba
l po
int
den
sit
y
t
hre
sho
ld
s
truc
tur
e f
eat
ure
a1 Δ1 a2 Δ2
up-down 0.
019 0.
003 0.
028 0.
006
l
eft-r
igh
t 0.
015 0.
005 0.
032 0.
004
i
n-ou
t 0.
018 0.
004 0.
028 0.
008
s
ing
le 0.
017 0.
004 0.
031 0.
004
Afte
r obta
ining
t he component
s t
ruc
tur
e
f
eat
ure
and the gl
obal poi
nt dens
it f
y eat
ure,we
ex
tra
ctt
he
anc
ien
t Ch
ine
se cha
rac
ter
image
fr
om
F
ig.
4 Th
ree
va
rian
tso
f “及”
t
he
pseudo
2D e
las
tic me
sh f
eat
ure
s.The
si
ze o
f
Aft
er opera
tions f
o normal
iza
tion, r
ever
se
co
lorand t
hinni
n ,ima e
s of “及 ” ar
e shown e
ach
mesh
is
8×8,a
to
tal
of
64g
rids.The
dens
ity
g g
i
n F
ig.
5. of
each
gri
d isc
al
cula
ted
fo
r t
he l
ocalpo
i
nt dens
ity
f
eature.Me anwhie,t
l heima ei
g s s
canned t
at 0°a
l
eft-r
igh
t d
ire
cton,90
i °at
up-down
dir
ect
ion,45
°
a
t lower-l
eft
dire
cti
on and 135°a t uppe
r-r
igh
t
d
ire
cti ,
on r espe
cti
vely.The l
eft-r
i t up-down,
gh ,
l
owe
r-l
eft,uppe
r-r
igh
t d
ire
cti
ons
have
7s
can
li
nes
F
ig.
5 Ima
ges
of “及”
aft
erop
e
rat
ion
s o
f no
rma
liz
aton,
i r
espe
cti
vel
y.By c
alcu
lat
ing p
ixe
l va
lue
s o
f e
ach
r
eve
rse
co
lor
and
th
inn
ing s
c l
an ne,we
i can ge
t t
hef
ea
tur
e ma
tri
xI0 ,I90 ,
The
four
component st
ruct
ure
f ea
tur
es tha
t I45,
I135 .By
fus
ing
the
loc
alpo
i
ntdens
it ma
y t
rix
c
an
be r
ecogn
ized
are
le
ft-r
ight,up-down,in-ou
t wi
tht
he
fe
atur
e ma
tri
x ob
tai
ned be
foe,we
r can
and s
ingle.I n order t
o de s
cri
be t he
componen
t ge
t t
he column
vect
o r.
st
ructur
e mo
re c
lear
ly,t he
four
structur
e f
eat
ure
s Fo
r example,thes
econd
var
ian
t f “及 ”c
o an
a
re
def
ined
in
Tab.
1. be
fus
ed
acco
rdi
ngt
o
expr
ess
i Eq.(
on 16).S
inc
e
第4期 Re
cogn
iti
on
ofanc
ien
t ch
ine
secha
rac
ter
s ba
sed
on
hyb
rid
ker
nel
WLS-SVR 327
t
he
val
ue
ofλi(
i = 1,…,
4)i
n Eq.(
16)gene
rat
ed s
amp
les,so
it
take
s t
he s
ame va
lue.The
loc
al
sma
ll impa
ct on
the
expe
rimen
t f
or t
he l
imi
ted po
int
dens
it ma
y t
rix
is
shown
asf
ol
lows:
熿0.
0587 0.
1207 0.
1655 0.
1655 0.
1878 0.
2018 0.
0291燄
0.
1281 0.
2500 0.
2667 0.
2889 0.
2600 0.
2000 0.
1234
0.
2500 0.
2738 0.
2381 0.
2381 0.
3857 0.
3175 0.
2036
B = 0.
1927 0.
2917 0.
4074 0.
2593 0.
4333 0.
3889 0.
1986 (
17)
1696 0.
0. 2143 0.
2381 0.
3333 0.
3000 0.
2381 0.
1641
0.
1281 0.
1750 0.
2667 0.
2889 0.
2800 0.
2889 0.
1213
0521 0.
燀0. 1068 0.
1140 0.
1111 0.
1872 0.
1111 0.
0169燅
Gr
id
scann
ing
met
hod
is
used
to
get
co
lumn f
ina
l f
usi
on
mat
ri Pi
x s
shown
asf
ol
lows:
ve
cto
rsco
r
respond
ing
to
dif
fer
ent
di
rec
tions.The
I n o
rde
r t
o ve
rif
y the
val
idi
ty o
f the
propos
ed expe
rimen
t i
s conduc
ted us
ing
the
same
samp
le
a
lgo
rithm, th
is paper
compa res
i t with two da
tas
et
s con
tai
ning
266s
amp
les
of
38c
las
ses.The
t
rad
iti
ona
l me
thods, LIB-SVM and LS-SVM, op
tima
l pa
rame
ter
s and
the
indexe
s o
f quan
tit
ati
ve
wh
ich
have
been
wide
l us
y ed
in
recogn
iti
on.The compa
ri a
son re
shown
in
Tab.
3.
Tab.3 Th
e quan
tit
ati
vec
ompa
ris
on
oft
hr
eer
ec
ogn
iti
on
alo
gri
thms
t
rai
ning r
ecogn
iti
on r
ecogn
iti
on
a
lgo
rit
hm pa
rame
ter
s t
rai
ning
accu
racy(% )
ime(
t s) ime(
t s) a
ccu
racy (% )
γ =128,
2
LIB-SVM σ =0.
0078 0.
0390 92.
82 0.
0041 75
γ =2,
2
LS-SVM σ =2 0.
0837 94.
40 0.
0645 78
WLS-SVR 8,
γ =0.
2
σ =8 0.
4639 97.
46 0.
2125 82
Fr om Tab.3,we c
an s
ee tha
t t
he ef
fecto
f
,
5 Conc
lus
ion
WLS-SVR bas
ed a
lgo
rithm works best with
the
LS-SVM
and
LIB-SVM
occupy
ing
the
se
cond
and Be
caus
e o
f t
he
high
unc
ert
ain
tyi
nt
he
shape
s
t
hird p
lac
es. The
al
gor
ithm pr
opos
ed n
i h
tis o
f anc
ien
t Ch
ine
se cha
rac
tes, t
r he e
rcogn
iti
on
pape
r ha
s a
high
accur
acy.Al
though
it
take
s t
he a
ccur
acy o
f ex
ist
ing c
las
sif
ier
s i
s l
ow.Suppo
rt
l
onge
stt
ime
fo
r t
rai
ning
and
re
cogn
iti
on
bec
aus
e ve
cto
r r
egr
ess
ion ha
s advan
tage
s o
f good
of
it
s comp
lex
ity,it
is
st
il
l wi
thi
n t
he a
cc
eptab
le gene
ral
iza
tion ab
ili
ty and
le
arn
ing pe
rfo
rmanc
e.
r
ange. Overa
ll, t
he me thod
in the pape
r is Th
i pape
s r pr
opos
ed a
re
cogn
iti
on a
lgo
rit
hm t
hat
e
ffe
cti
ve. comb
ine
s a hybr
id-ke
rne
l f
unc
tion wi
th adap
tive
328 中国科学技术大学学报 第 45 卷
we
igh
ted
le
ast
squa
res
suppo
rt ve
cto
r r
egr
ess
ion Unmi
xing Mode
l Ba
sed on Le
ast
Squa
res Suppo
rt
f
or Vec
tor Mach
ine With Unmix
ing Residue Cons
trai
nts
anc
ien
t Ch
ine
se cha
rac
ter
re
cogn
iti
on. The
[ ]
J .IEEE Geosc
ience
and Remote Sensi
ng Lett
ers,
wave
let
ke
rne
l f
unc
tion wi
th l
oca
l pr
ope
rti
es and
2013,10(6):1592-1596.
RBF
kerne
l f
unc
tion wi
th good g
loba
l pr
ope
rti
es [9 ]Liu
B Y,YangR G.A nove l method bas
ed on PCA
a
recomb
ined
to
cons
truc
t t
he
hybr
id-ke
rne
l.The and LS-SVM f or powe r load
f oreca
sting [C ].
we
igh
t coe
ffi
cien
ts de
cay a
t a
ra
te o
f he
t I
nte
rna
tiona
l Con
fer
enc
e on El
ect
ric Ut
il
ity
exponen
tia
l f
unc
tion o
f pr
edi
cti
on e
rro
rs. De
regu
lati
on and Re struc
tur
ing and Powe
r
Te
chnol
ogies,NanJ
ing,2008:759-763.
Expe
rimen
t r
esu
lts
show
tha
t t
he
propos me
ed t
hod
[
10]Zhang H R,Wang X D,Zhang C Je
ta
l.Sof
t s
enso
r
ha
s good
robus
tne
ssand
high
recogn
iti
on
accur
acy. te
chni []
que
usi
ng
LS-SVM and
standa
rd SVM J .IEEE
f
utur
e r
ese
arch
is
to
fi
nd
a f
ast
erso
l
uti
onf
or
the I
nte
rna
ti l Con
ona fer
ence on I
nformati
on Ac
qui
sit
ion,
WLS-SVR
algo
rit
hm
tor
educ
e r
ecogn
iti
ont
ime. Hong
Kong and
Macau,2005:124-127.
[
11]Xi
e J
H.Pri
nted cha
rac
ter
recogn
iti
on us
ing Kerne
l
Re
fer
enc
es CCA wit
h LS-SVM me thod [C ]. Comput
er and
Automati
on Engine
ering,2010:284-287.
[1 ]Zhang P.Res
earch
ofd
ii
gta
l const
ruc
tion o
f Pape
r f
ile
s
[
12]Yin
D Y,Wu Y Q. De t
ecti
on of Small Ta
rge
t i
n
andr
el
ic [ ]
s J .Cult
ura
l Rel
ics of
Centra
l Pla
i ,
ns 2009
I
nfrar
ed Image Based on KFCM and LS-SVM [C].
(5):104-107.
In
terna
tiona
l Conf
erence
onI
nt
ell
igen
t Human-Ma
chi
ne
[
2]Lv X Q,L i
M N e
t al.An or
acl
e c
lass
ifi
cat
ion me
thod
Syst
ems and Cybe
rneti
cs,2010:309-312.
baed on
s fi
gur
e rec
ogni
ti [ ]
on J .J ournal
of Be
iji
ng
[
13]SuykensJ
A K,de Br
abant
er J,Luka L,Vande
s ewal
le
Unive
rsi
ty of In
forma
tion Sc
ienc
e and Te
chno
logy,
J.We ighted Lea
st squa
res suppo
rt ve
cto
r machi
nes:
2010(
25):92-96.
r
obustness
and spa rs
e app rox
ima
tion [J ].
[3 ]Chen D,Li N,Li L.Onli
ne
handwri
ting
re
cogn
iti
on
Neur
oc omput
i , ( ):
ng 2002 48 85-105.
re
seach o
r f anc
ien
t cha
rac
ter[J].Journa
l o
f Be
iji
ng
[
14]Smit
s F,J
G o
r da
an E M.Impr
oved SVM reg
res
sion
Ins
tit
ute
of
Mechani
cal
Industy,2008(
r 4):32-37.
us
ing mixt
ures of ke
rne
ls [C ]. tex
tit
in Neu ra
l
[4 ]Zang G Q.Expe
riment
and Improvemen
t o
f a
ccura
cy
of
Ne
two
rks,I
nte
rna
tiona
l J
oin
t Con
fer
enc
e on,2002,3:
OCR for
t ext-d
igi
ta image [J ]. I
l nt
ell
i e,
genc
2785-2790.
I
nfo
rma
tion
and Shar
i , ( ):
ng 2010 3 62-67.
15]温昌兵 .基于特征融合的脱机手写体汉字识别[
[ D].北
[5 ]Vapn
ik
V.The Na
ture
of Sta
tist
ica
l Le
arn
ing
The
ory
京,北京科技大学, 2005.
[M].New
York:Spr
inger-Ver
lag,1995.
[
16]Zhang
X B,Huang H,Zhang
S J.A
FCM
clus
ter
ing
[6 ]Suykens
J
A K,Vandewall
e J.Le
ast
Squa
res
Suppor
t
a
lgo
rit
hm bas
ed
on Semi-supe
rvi
sed and
Poin
t Densit
y
Vector Mach
ine Cl
ass
ifi
ers [J]. Neura
l Proce
ss
Wei
ghted[C].Int
ell
igent Computi
ng and Int
ell
igent
Le
tters,1999 (
3):
293-300.
Sy
stems,2010:710-713.
[7 ]Miran
ian A, Abdol
lahzade M. De
vel
opi
ng a Lo
cal
[ ]
17 Tu Y K,Chen Q H,Huang
L.Handwr
itt
en
Chi
nes
e
Le
ast-Squa
res
Suppo
rt
Vec
tor
Mach
ine
s-Ba
sed
Neu
ro-
chara
cte
r re
cogni
tion based on
pseudo two-d
imens
iona
l
Fuzz
y Model
for Nonl
inea
r and Chaot
ic Time Se
ries
ela
sti
c mesh [J].J our
na l
of Huazhong Unive
rsi
ty o
f
Pr
edict
ion[J].IEEE Transac
tions
on
Neural
Networks
Sc
ienc
e and
Techno
logy,2010(
38):
38-40.
and
Lear
ning Sy
stems,2013,24(
2):207-218.
[8 ]Wang L G L, i ,
u D F Wang Q M e t al.Spe
ctr
al