TELKOM
NIKA
, Vol.12, No
.4, Dece
mbe
r
2014, pp. 10
73~107
8
ISSN: 1693-6
930,
accredited
A
by DIKTI, De
cree No: 58/DIK
T
I/Kep/2013
DOI
:
10.12928/TELKOMNIKA.v12i4.788
1073
Re
cei
v
ed Se
ptem
ber 4, 2014; Re
vi
sed
Octob
e
r 15, 2
014; Accepte
d
No
vem
ber
4, 2014
An Algorithm Based on
Wavelet Neural Network for
Garment Size Selection
Luo Ron
g
lei*
1
,
He Wenjie
2
,
Li Cheng
y
i
2
1
School of F
a
s
h
io
n
,
Z
heji
ang Sci-T
e
ch
Unive
r
sit
y
,
Ha
ngzh
o
u
, Z
hejia
ng 3
1
001
8, Chi
n
a
2
Engine
erin
g
Rese
arch Ce
nter of Clothi
ng o
f
Z
hejian
g
Prov
ince, Ha
ngz
ho
u, Z
hejia
ng 3
1
001
8, Chi
n
a
T
e
l. 057186
84
348
0, fax: 05
7
186
84
348
1
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: luoro
ngl
ei@
1
63.com
A
b
st
r
a
ct
Si
z
e
fitting problem
is a main
ob
stacle to large scale
online
garm
e
nt sales.It is the difficult t
o
customers to fi
nd the fit gar
ments w
hen they
could
n
’
t try
on.
In this pap
er, w
e
present a
n
alg
o
rith
m b
a
se
o
n
w
a
velet ne
ural
netw
o
rk to help cust
omer
choos
ing th
ei
r
clothin
g
sp
eci
f
ications
auto
m
atic
ally. After
the
reaso
nab
le w
a
velet functio
n
i
s
selected, w
e
establ
is
he
d the
mod
e
l structur
e and t
he i
n
itia
l
para
m
eters. T
he
w
a
velet ne
ura
l
netw
o
rk is t
r
ain
ed by
the
body
m
eas
u
r
es an
d the
result of AHP
alg
o
rith
m aft
e
r
nor
mal
i
z
a
ti
on.
T
he n
e
w
data
are
use
d
to t
e
st the
net
w
o
rk. As a res
u
lt,
the
error fro
m
w
a
ve
let
ne
ura
l
netw
o
rk is sma
ller, an
d the pr
edicti
on accur
a
cy is pr
oved th
an that from th
e alg
o
rith
m ba
sed on trad
itio
na
l
BP network.
Ke
y
w
ords
: W
a
vel
e
t neur
al n
e
tw
ork, w
a
velet transform, Garm
ent si
z
e
se
lectio
n, match
i
ng al
gorit
hm,
F
i
t
gar
me
nt match
i
ng
1. Introduc
tion
With the developme
n
t of shoppi
ng onlin
e, the st
atus of garment n
e
twork sale
s is risi
n
g
grad
ually. According
to th
e stati
s
tics,
50% re
turn
s are b
e
ca
use of
u
n
fitted problem
s
si
nce
cu
stome
r
s ca
n’t try ga
rme
n
t
s on
by
onlin
e sale
s [1
]. M
any research
es
devote th
e
m
selve
s
to
fin
d
a method
ca
n ke
ep cust
omer’
s
me
asurem
ents m
a
tch with g
a
rment si
ze. In garment si
ze
sele
ction alg
o
rithm, least
-
squ
a
re meth
od has b
een
applie
d on garme
nt size
classificatio
n
for a
long time [2]
-
[4], it can de
velop the
rati
onality by u
s
i
ng weightin
g
method. Yu [
5
] put forwa
r
d a
garm
ent si
ze
sele
ction
m
e
thod by
co
mpari
ng t
he
matchin
g
rel
a
tionship b
e
twee
n figures and
garm
ent size. Throu
gh the comp
ari
s
o
n
betwee
n
che
s
t
、、
wai
s
t
hip width of figures and the o
n
e
s
on garment
s, define the matchin
g
relati
onship bet
we
en figure
s
an
d garm
ent si
ze. He [6] made
resea
r
ch on
Yu’s metho
d
, and propo
se
d modified
a
d
v
ice. Xu put forward usi
ng
AHP method
to
s
e
lec
t
garment s
i
z
e
[7],the method
makes
s
i
z
e
s
e
lec
t
ion into quant
ific
ation,
which is suitable for
garm
ent
sale
s o
n
line.
On
the
ba
sis o
f
AHP,
Di
ng
and
Xu
also devel
ope
d
a
garment
size
sele
ction met
hod com
b
ine
d
immun
e
al
gorithm
s
a
n
d
AHP [8], an
d su
ppo
rt th
at the predi
ction
accuracy
of it is hi
ghe
r tha
n
usi
ng A
H
P
alone.
T
he m
e
thod from
Chen, ma
de u
s
e of BP net
work
to pro
c
e
ss
si
ze cl
assification [9], can a
pply on
ga
rm
ent size sele
ction the
sam
e
. Some garment
size matchi
n
g
algorith
m
s are co
mplet
ed in the fo
rm of virtual fitting. In
this case, garm
ent
pattern
ha
s b
een
reg
a
rd
ed
as
a pu
re
g
eometry p
r
o
b
l
em tran
sfe
r
red from
two
dimen
s
ion
s
i
n
to
three
dimen
s
i
ons. Structu
r
e gri
d
was m
ade o
n
the
p
a
tterns,
and
then finite
ele
m
ent meth
od
is
use
d
to
cal
c
ulate the
su
rf
ace
curvatu
r
e of g
a
rm
e
n
t model
s i
n
o
r
de
r to
pro
c
e
s
s salient p
o
i
n
ts
matchin
g
a
u
tomatically [1
0]-[11]. Prop
ose
a
co-evolutiona
ry im
mune
algo
rithm for th
e
multi-
crite
r
ia d
e
ci
si
on ma
kin
g
(MCDM)
mod
e
l, and
us
e t
he mo
del to
solve the
large
scal
e
g
a
rment
matchin
g
problem
[12]. The
ap
plication of
ne
ural
net
work on
si
ze
selectio
n u
s
e
th
e
measurement
points a
s
i
nput and th
e
corre
s
po
ndi
ng ga
rment
sizes a
s
o
u
tput to train the
netwo
rk [13]
-[14] .
In this pape
r, we develop
ed the meth
od of garme
nt size mat
c
hing on the
basi
s
of
wavelet ne
ural network. It coul
d input
new
som
a
tic
data an
d act
ual suitable
garm
ent si
ze
by
trying on in t
he follo
w-u
p
step, by mea
n
s of tr
ai
ning
redu
ce th
e
discre
pan
cy
betwe
en n
e
u
r
al
netwo
rk o
u
tp
ut and actu
al results.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 107
3 – 1078
1074
2. Wav
e
let neural net
w
o
r
k size match
i
ng algorith
m
2.1. The sele
c
tion of
w
a
v
e
let fun
c
tion
We
cho
o
sed
several wavelet functio
n
s as the b
a
si
c fun
c
tion of
the neu
ral
netwo
rk,
inclu
d
ing Mo
rlet’s wavel
e
t ,which is u
s
e
d
to establi
s
h wavelet ne
ural n
e
two
r
k
and ha
s be
e
n
applie
d o
n
ki
nds of field
s
widely. T
he
wavelet fu
nct
i
on i
s
u
s
e
d
a
s
hi
dde
n lay
e
r
nod
es of t
he
netwo
rk.
Hid
den laye
r n
o
des
num
ber
is 10
0. The
error
of mea
n
sq
ua
re
()
e
f
or trai
ning
s i
s
defined a
s
fol
l
ow:
2
1
1
()
1
iL
ii
i
ed
y
L
(1)
In the formul
a: L is a
s
hid
den laye
r no
des, d i
s
a
s
t
he ide
a
l re
sul
t
s. (Thi
s pa
p
e
r results
from AHP), y is as n
eural n
e
twork outp
u
t.
After 100
tim
e
s t
r
aini
ng, t
he n
e
two
r
k e
rro
r
use Little
w
oo
d-p
a
ley
wavelet fu
nct
i
on
can
de
crease
by about
0.35
,while the
on
e use Mo
rlet
wavele
t i
s
a
b
out 0.4.Mad
e
experim
ents
on Mexi
can
cap
wav
e
let
、
Sh
a
nnon
wavelet
to p
r
o
c
e
s
s te
st, the
re
sults is
still n
o
t b
e
tter tha
n
Little
w
oo
d-p
a
ley.
So
this pap
er u
s
es Littlewood
-paley wavele
t f
unction as
basi
c
fun
c
tion
of hidden lay
e
r.
2.2. Confirm
hidden la
y
e
r
nodes num
ber
It’s very imp
o
rtant fo
r wa
velet neu
ral
netwo
rk train
i
ng stu
d
y to
confirm hid
d
en laye
r
node
s n
u
mb
er. if the hid
den laye
r no
des
numb
e
r
is too little, netwo
rk
will
not po
ssess
the
essential
abili
ty of processi
ng dat
a.
On t
he othe
r
side,
if there
a
r
e t
oo ma
ny nod
es, the
netwo
rk
compl
e
xity will be increa
se
greatly, whi
c
h will slo
w
d
o
w
n the n
e
two
r
k le
arni
ng
speed. p
r
o
c
e
s
s.
The
comm
on
app
roa
c
h i
s
cut-a
n
d
-
try, that is, hid
den
layer n
ode
s
numbe
r i
s
co
nfirmed
by tri
a
ls.
The test sta
r
t
s
with a min
o
r
node. As th
e node
s
nu
m
ber i
s
increa
sed step by
step, the re
sult
s
were compa
r
ed. We g
e
t the test on th
e sel
e
cted
hi
dden laye
r n
ode
s num
ber. The re
sult
s are
sho
w
e
d
as b
e
low, in which the errors a
r
e the data of
100 times tra
i
ned.
Table 1. No
d
e
s num
be
r an
d Traini
ng error
Nodes
number
10 20
40 80
120
160
Error
0.30
1
0.27
5
0.21
7
0.10
6
0.09
9
0.08
0
The re
sult
s show that, the training e
r
ror
will be getting
smaller a
s
th
e node
s num
ber get
bigge
r
within
a certai
n ra
nge.
We
use the n
u
mbe
r
10
0 a
s
o
u
r hidde
n laye
r node
s
num
ber
con
s
id
erin
g that the bi
gge
r numb
e
r
woul
d lead
to
the
over-fitting of
neural net
wo
rk an
d influe
n
c
e
on trainin
g
sp
eed.
2.3. The sele
c
tion of initi
a
l parameter
s
The
sel
e
ctio
n
of initial
pa
rameters
ma
kes
effect
on l
o
cal
minim
u
m poi
nt an
d
netwo
rk
conve
r
ge
nce
sp
eed. If ini
t
ial wei
ght va
lue i
s
n
o
t
sui
t
able, lea
r
nin
g
process
wil
l
get into
lo
cal
minimum p
o
i
n
t, and mig
h
t even sto
p
worki
ng. In the
wavelet n
e
u
r
al network, we co
uld u
s
e t
he
rand
om valu
e between
(0
, 1) a
s
the in
itializat
ion of
con
n
e
c
tion weights betwe
en
hid
den
l
a
yer
and inp
u
t layer, and the o
ne between o
u
tput layer
an
d hidde
n laye
r, as we used
in BP network.
2.4. The confirmation of learning ra
te
Since the
sel
e
ctive wavelet function m
o
stly
ha
s limi
t
ed su
ppo
rt, the big lea
r
ni
ng rat
e
might lea
d
th
e network p
a
rameters to
b
e
out of
suita
b
le sco
pe, an
d then th
e ou
tput value mi
ght
be to ze
ro. So the wavelet
learnin
g
rate
sho
u
ld be a p
o
sitive numb
e
r as
small a
s
po
ssi
ble.
Experiment
s sh
ows th
at the bi
g le
arnin
g
rate
s (i.e. le
arni
n
g
rate of
conne
ction
coeffici
ent µ1
=0.1, lea
r
nin
g
rate of
scali
ng pa
ramete
rs an
d tran
slat
ion pa
ramete
rs µ1
=0.1
) le
ad
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
An Algorithm
Based o
n
Wa
velet Neural Networ
k for
Garm
ent Size Selection (Luo Ronglei)
1075
the network f
a
ll into lo
cal
minimum, th
us the
traini
n
g
failed.
Whil
e µ1=0.1 an
d µ1=0.001
are
sele
cted, the
training
re
sult
is better.
Figure 1 sho
w
s the
esta
bl
ishme
n
t pro
c
ess of
neu
ral
netwo
rk
stru
cture above. Figure
2
sho
w
s the ne
ural net
wo
rk t
r
ainin
g
proce
ss.
3. Simulation
3.1. The sele
c
tion for
trai
ning sample set
637
sets
of b
ody data i
s
selecte
d
from
age 1
8
-3
5, which i
s
m
e
a
s
ured
by BoS
S
-21 3
D
body me
asurement
syste
m
from
Zheji
ang S
c
ien
c
e
and T
e
chnol
o
g
y Unive
r
sity
fashi
on in
stitute.
600
set
s
of d
a
ta are a
pplie
d for
neu
ral
n
e
twork
t
r
ainin
g
,
and othe
r 37 sets are
a
pplied
fo
r neu
ral
network
tes
t.
Neu
r
al net
wo
rk st
ru
cture i
s
5×1
0
0
×
1
(
in
put
layer: 5 node
s, hidde
n layer: 100 node
s,
output: 1 no
de).In the tra
i
ning process,
5
c
r
itic
a
l
pa
r
t
me
as
ur
eme
n
t
s
(c
he
s
t
、
wai
s
t
、
s
h
ould
e
r
width
、
sle
e
ve
length
、
cent
re back len
g
th
)are putted o
n
input layer
node
s The id
eal re
sults, d,
is
got from the output re
sult of algorithm b
a
se
d on AH
P
.
-1 is used t
o
rep
r
e
s
ent small size, an
d 0
is u
s
ed to repre
s
e
n
t mid
d
le si
ze, an
d
1 is u
s
ed t
o
rep
r
e
s
ent l
a
rge
si
ze, a
nd 2 is
used
to
rep
r
e
s
ent extra large si
ze.
Data is n
o
rm
alize
d
to the interval betwe
en [-1,1] befo
r
e traini
ng.
Figure 1. wav
e
let neural ne
twork mod
e
l flow chart
Co
nfirm
activ
atio
n
fun
c
tion
C
o
n
f
i
r
m
hi
dde
n l
a
y
e
r
n
odes
num
ber
C
o
n
f
i
r
m
hi
dde
n l
a
y
e
r
num
ber
Co
nf
ir
m
o
u
t
put layer
no
d
e
s
C
o
n
f
i
r
m
i
nput
l
a
y
e
r n
odes
Th
eselection
of
in
itial
param
e
ters
Co
nfirm
learn
i
n
g
rate
Select the rele
vant m
easurem
ent pa
rts m
a
tch with
g
a
rm
en
t typ
e
s,
su
ch
as: ch
est,
waist, etc
B
a
sed
on
rec
o
m
m
e
ndat
i
on si
ze w
o
r
k
e
d
by
AH
P:
Use
-
1
,
0
,
1
,
2 t
o
represe
n
t sm
all,middle,large,e
t
ra large
size
sep
a
rately.
One
lay
e
r
100 nodes
H
i
dde
n la
y
e
r func
ti
on
:
sin(
2
)
sin(
)
()
x
x
x
x
Outp
ut la
y
e
r
:
pie
c
e
w
i
se
func
ti
on
2
1
0
1
y
R
a
nd
om
val
u
e
bet
w
ee
n
[
0
,
1]
µ
1
=0.01
,
µ
2
=0.001
,
x
≥
1.5
,
0.5
≤
x<1.5
,
-0.5
≤
x<0.5
,
x<-0
.5
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 107
3 – 1078
1076
Figure 2. Neu
r
al network training p
r
o
c
e
ss
Defi
n
itio
n
:
N
:
Inpu
t layer nodes nu
m
b
er ;l:ou
t
pu
t layer
n
odes nu
m
b
er; L:hid
d
e
n
layer no
d
e
s num
b
e
r; wln
:
th
e
co
nn
ection
co
efficien
t
b
e
tween
th
e n
t
h
inpu
t
layer
no
de a
n
d t
h
e
1
s
t
hi
d
d
e
n
l
a
y
e
r
no
de;
wl
:
t
h
e
o
n
e
bet
w
ee
n t
h
e
1st
hi
d
d
e
n
l
a
y
e
r
no
de
and
o
u
t
p
ut
l
a
y
e
r
no
des;
a:
t
h
e
expa
nsi
o
n
coe
f
fi
ci
ent
o
f
wav
e
l
e
t
fu
nct
i
on;
b:
he
tran
slatio
n co
efficien
t
o
f
wavelet fun
c
tio
n ,
µ:learn
i
ng
rate; k
:iteratio
ns.
In
itialize co
efficien
t
No
rm
al
i
ze i
nput
an
d
o
u
t
p
ut
d
a
t
a
:
tn
:v
alu
e
fo
r the n
t
h nod
e after
n
o
rm
al
izin
g
so
m
a
t
i
c d
a
ta on
in
pu
t
layer
Ŷ
:
id
eal resu
lts,
d
’
s
:
no
rm
alize
d
v
a
lu
e
()
ln
n
l
n
l
l
l
wt
b
yw
a
dw
l
=-(
ŷ
-
y
)*
ψ
l
dw
ln
=-(
ŷ
-
y
)*
w
l
*
ψ
l
*
t
n
/
a
db
l
=(
ŷ
-
y
)*
w
l
*
ψ
l
’
/
a
a
l
v
a
riation
da
l
=
(
ŷ
-
y
)*
w
l
*
ψ
l
’*
(
ln
n
l
n
wt
b
)/
a
2
w
l
=
w
l
–
µ
1
*
d_w
l
w
ln
=
w
ln
-
µ
1
*
d_w
ln
b
l
=
b
l
-
µ
2
*
d_
b
l
a
l
=
a
l
-
µ
2
*
d_
a
l
Iterations
<
k
Y
Pu
t
d
a
ta to datab
a
se
b
a
se
N
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
1693-6
930
An Algorithm
Based o
n
Wa
velet Neural Networ
k for
Garm
ent Size Selection (Luo Ronglei)
1077
3.2. Net
w
o
r
k
training
After about 1
00 times trai
ning, the error
do
es n
o
t decrea
s
e
ob
viously anym
o
re. Th
e
minimal traini
ng error valu
e is 0.38, sho
w
n as Fi
g. 3.
The re
com
m
ende
d size of 37 sets of bo
dy
data by neu
ral netwo
rk i
s
sho
w
n a
s
Fig
.
4, and t
he accuracy i
s
9
5
%, compa
r
e
d
with trying
on
person.
Figure 3. Traini
ng erro
r of wavelet
Figure 4. wavelet neural n
e
two
r
k testing
neural n
e
two
r
k
3.3. Contr
a
s
t
experiment
W
h
en
fu
nc
tion
1
()
1
x
fx
e
is u
s
e
d
a
s
neural net
wo
rk hid
den l
a
yer prim
ary fun
c
tion a
nd
other
set
s
a
r
e put a
s
the
same a
s
the
wavelet neu
ral
netwo
rk, th
e recom
m
en
ded
re
sult from
the
BP algo
rithm
is
sho
w
n
a
s
Fig. 5. T
he t
r
aining
er
ro
r
d
e
crea
se
s i
n
to
0.4 i
n
the
en
d. The
a
c
curacy
of 37 sets of
data test sa
m
p
les i
s
89%.
Figure 5. Trainin
g
error of Back
Figure 6. Back Prop
agation n
e
twork
Propagati
on network
testing
4. Conclusio
n
The pa
per
extends the
ap
plicatio
n of wavele
t neu
ral
netwo
rk to g
a
rme
n
t size selectio
n
field. The tra
i
ning e
r
ror
a
nd p
r
edi
ction
accu
ra
cy h
a
ve bee
n de
veloped
com
pare
d
with
the
algorith
m
ba
sed on
traditio
nal BP net
wo
rk. Ind
eed, th
ere
are
still
some p
r
obl
em
s ne
ed to
sol
v
e
in future. F
o
r example, th
e net
work m
a
y fall into lo
cal mi
nimum
point, and
re
que
sts
huge
data
prep
ari
ng for models train
i
ng at early times. A
nd the
network ne
e
d
s traini
ng d
a
ta as mo
re
as
possibl
e. If th
e sampl
e
s a
r
e not enou
gh,
training resul
t
s will not be
better a
s
we
expect.
Ackn
o
w
l
e
dg
ements
The resea
r
ch wo
rk was
sup
porte
d by
the Prog
ra
m for Zh
ejia
ng Le
adin
g
Team of
Scien
c
e an
d
Techn
o
logy
Innovation No.20
1
1
R
50
004 and
Na
tural Scien
c
e Found
atio
n of
Zhejian
g
Pro
v
incial un
der
Grant
No LQ
12F02
018.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 16
93-6
930
TELKOM
NIKA
Vol. 12, No. 4, Dece
mb
er 201
4: 107
3 – 1078
1078
R
e
fe
re
nc
es
[1]
Xu Yic
h
a
o
, Din
g Yongs
hen
g. Rese
arch an
d app
licati
on of g
a
rment fit assessments.
Journ
a
l of T
e
xtile
Research.
20
0
7
;28(1
0
): 127-
130
.
[2]
W
ang Ji
an
pin
g
,
Li Yu
eli, Y
u
F
ang. D
i
gitiz
ed
cat
egor
izatio
n f
o
r garm
ent siz
e
des
ig
natio
n
base
d
o
n
th
e
appr
oach
i
n
g
pr
incip
l
e.
Jour
na
l
of T
e
xtile Res
earch.
20
07; 2
8
(11): 10
6-1
1
0
.
[3]
Zhang H
e
n
g
, Zhan
g Xin. T
he t
y
p
e
filin
g for g
a
rment size t
y
p
e
base
d
on th
e
lot custom-ma
de.
Journ
a
l
of Dong
hu
a Un
iversity: Natura
l Editio
n.
200
9; 35(4): 436-
44
0
.
[4]
X
u
J
i
h
ong, Z
han
g
X
i
a
n
g
hui
, Zhang Wen
b
in. Stud
y
on
the di
gital
i
zat
i
on
of filin
g c
u
stom-mad
e
garme
nt’t
ypes
and cuttin
g
pl
a
n
.
Journa
l of Dong
hu
a Univ
er
sity: Natural Ed
ition.
20
03; 29(
1): 37-42.
[5]
Yu Xi
aok
un,
W
ang Ji
anp
in
g
.
T
he matchin
g
princi
pl
e of bo
dil
y
form a
nd
garme
nt size s
e
ries a
nd
its
app
licati
on in garme
nt
electr
onic
c
o
mmerc
e.
Journ
a
l
of Don
ghu
a U
n
iv
ersity: Natura
l
Editio
n.
20
03;
29(1): 43-
47.
[6] He
Li.
Study
o
n
a
n
thro
po
metr
y an
d si
z
e
rec
o
mme
ndi
ng
of cl
othin
g
mark
eti
ng
onl
in
e.
W
u
Xi:
J
i
a
ngn
a
n
Univers
i
t
y
. 2
0
0
8
.
[7]
Xu Y
i
-Ch
ao, Qi Jie, Din
g Yo
ng-Sh
en
g. AHP Ba
sed Opti
mal Sel
e
ction
of Garment Sizes for Onlin
e
Shop
pin
g
.
Jour
nal of Do
ng
hua
University.
20
07; 24(2): 2
22-
225
.
[8]
Ding
Yo
ngs
he
ng,
Xu
Yic
h
a
o
. Intell
ig
ent
optima
l
se
lecti
on
of g
a
rmen
t
sizes by
using immun
e
alg
o
rithm an
d AHP metho.
Jo
urna
l of T
he T
e
xtile Institute.
2008; 99(
3): 281
-286
.
[9] Chen
Li
w
e
n.
T
he rese
arch & app
licati
on of data min
i
ng
mode
l base
d
on
neura
l
netw
o
rk theory
. Da
Lia
n
: Dali
an M
a
ritime Un
ivers
i
t
y
. 20
04: 46-
5
7
.
[10]
Deca
udi
n Ph
ili
ppe, J
u
li
us Da
n,
W
i
ther Jami
e, et al. Virt
ual
Ga
rments: A full
y
geom
etric
appr
oach
f
o
r
clothi
ng d
e
sig
n
.
Comp
uter Graph
ics F
o
ru
m.
200
6; 25(3): 62
5-63
4.
[11]
Magn
enat-T
hal
mann N
a
d
i
a,
Seo H
y
e
w
o
n
,
Cordi
e
r
F
r
ed
er
ic. Automatic
mode
lin
g of vir
t
ual h
u
mans
and b
o
d
y
cl
othi
ng.
Journ
a
l of
Co
mp
uter Scie
nce an
d T
e
chn
o
lo
gy.
200
4; 19(5): pp.57
5-5
84.
[12]
Don
g
Yo
ngsh
e
ng, Hu
Z
h
i-H
u
a, Z
han
g W
e
n
b
in.
Mu
lti-criter
ia d
e
cisi
on m
a
king
ap
proac
h
base
d
o
n
immune
c
o
-ev
o
luti
onar
y alg
o
r
ithm w
i
t
h
a
p
p
licatio
n to gar
ment matchin
g
probl
em.
Expert System
s
w
i
th Applicati
o
ns
. 2011; 3
8
(8)
:
10377-
10
383
[13]
Don
g
Mi
ao, H
ao K
u
a
ngro
ng,
Din
g Y
i
on
gsh
eng.
E
x
pert s
ystem for siz
e
recomme
ndi
ng
of cl
othin
g
base
d
on fuzz
y neura
l
net
w
o
r
k
.
Microcomput
er Appl
icatio
ns
.
2010; 26(
3): 21-23.
[14] Z
heng
A
i
hu
a.
Study on s
i
z
e
reco
mme
ndi
ng
of clothi
ng
sy
stem b
a
se
d o
n
back
prop
ag
ation
neur
a
l
netw
o
rk.
Hang
Z
hou: Z
heji
a
n
g
Sci-T
e
ch Universit
y
.2
01
0.
[15]
Hu J
uel
ian
g
, D
ong
Jia
n
min
g
,
et al. Garm
ent
pa
ttern des
ign
base
d
o
n
artific
i
al ne
ural
n
e
t
w
ork.
Jour
nal
of Textile Research.
200
6; 7(2
)
: 49-52.
Evaluation Warning : The document was created with Spire.PDF for Python.