NII-Electronic Library Service
A
Method
to
Construct
a
Form
from
lX-Y"C[g5<MF.uli.Ooseenl
J7
)iRLJ7X
7-Ftf t-)iAndreas NORDGREN
KejoUniversity
fieescpt#<\
lmages
This
paper
describesa method torelate formimpressionstodesignparameters;a mapping of image-space to
design-space v[a a neural network solution, A survey designed by
thelaguchimethod was used tocollect dataon how altering
various designparametersof an automobile affects itsimage.
Principal
Component
Analysiswas performedtoextractcor-related factorsfromthisdataand each sample includedinthe
survey was
described
inthisnew spaceby
calculatingits
fac-torscore.
A
neuralnetwork was constructed and trainedwith the$e samples, which a[lowed aCAD-system,
based
on thisneural network, toautomatically
generate
3D-models
corre-sponding toanyform
impression
presented
toit.
These
results show thatit
is
possib[etogiveadesign
system a `sense' otshapes,
prev[ously
restrictedonlytothedesignen1.
Introduction
Withthe fiercecompetition inthe automotive industry,itis absolutelyessential
to
have
the
access tomodernCADICAM-systems inorder todevelop new cars ina short amount of time.A new model needs not onlytofulfilltheoftenconflicting
demands of thepotentialcustomers, butalso tobring some-thingnew
-
functions,features,outstandlng qualityorinnova-tivedesign
-
tostand out among thecompetitors. Therearemethods toincludethe customeris voice invarious stages of
thedevelopment of a new mode], butwhen itcomes tothe
de$iredimage,theysuffer fromthesubtleties of form
impres-sions, Whereas customer demands for
properties
such aspassenger
space, fueleconomy, performanceand safety are easy toquantifyand consider inthedesignprocess,theformimpressionisa subtle association or a feeling,and thusfar
lesseasy torelate todesignparameters.Such a form
impres-sionisoften expressed innatural Ianguage,which israther subjectiveand
differs
from
customers tocustomers.CADICAM-systems are very helpfulinthedesign and
manufacturing process,
but
theyhave
no sense ofcreativity orknowledge about whatkind
offormimpressions
certain shapes wil]give.Ithasalways beentheroleofthedesignertointerpretsubtleties incustomer demands fora design,transfer
itintosurfaces and curves, and create aproductthatwil[ ap-pealtothe targeted groupof customers. Howevec thisisnot an easy thingtodo
-
itrequire$ experience and $ensitivitytocurrent trends,or betteryet,thetrend forthe coming years. For
instance,
what shapes and curves willgiveanimpres-sion of a sporty and powerfulcar,
yet
with a classicIookand elements offormality
init,toa young woman? lstheresome set ofproportionsthatwillyieldaform
impression
that coin-cides wi±h
what thetargetedcustomerdesires?
Knowing
such a set ofproportionswou[dbe
very valuablein
creating suc-cessful cars.Agood
designerprobably
hasafeelingfortheseproportions,
but
it
would certainlybe
valuable toalsohave
thistypeofsensitivitybuiltintoa CAD-system, and beable
tovisualizehow various form impressionswill affectthe
de-sign parameters.By
learning
tomapimage-space
todesign-space, thistype ofsystem could be used intheeariystages ofthe
design
processtosuggestdesign
parameters,orby
the reverse mapping, validate how the form impressionchangesafterare-design,
The
methoddescribed
in
thispapercanbe
used as afoundationtodevelopa designsupport system with these
features.
Thus
this
methodhas
the
potentialto
shortenthe designproces$,become avaluable too[ inthe creative
process,as we]1 as validate thatthe car reflects customer
re-quirementson theformimpression.
2.
Method
The method presentedinthispaperuses several techniques fromstatistics, multivarjate analysis and artificial neural
net-works, and thesections belowdescribehow thesetechniques were appl[ed and what assumptions were made.
2.1. DesignParameters
The
first
step of thismethod istodecidea set ofdesignpa-rameters thatcan adequately describetheshape in
question.
For
thiswork, a sedan typeofcar was considered and12
parame
±ersfor
thebasic
proportions
of thesideprofile
shapewere decidedas shown inFigure1
.
The angle-andlength-pa-38Tnyif(y#nvveksce
$pecialissueefjapanesesecietyferthescieneeetdesLgnyel.15-4 ne.6e 200B
t" ts op...,,,,...e7e6...tj op
-・・・・--.・
---
va ls..e.f
,,.Fig.1.12DesignParametersforSideShape
9.,.ig
/ al,
i'
・・v・・・l'
u'1
Fig.2.IWoDesignParametersforCrossSectionProfileShape rameters of thissetup provideda basiccontrol poiygonwhich laterwas used to placethecurves and surfaces forthecar's
body.
Differenttypes of vehicles have some fundamental
differ-ences
in
their
basic
shapes, andtherefore
the
design
param-etersforone type of car do not necessarily correspond well to
thatofanother type.
Furthermore,the
decision
ofdesign
parametersisnotlim-itedtobasicproportionsofthecar
-
dependingon the type ofshape thedesigner
is
working on, itispossibleto definedesignparametersfreely,as longas theyhave a significant
effect on theshape. Inmost cases the
designer
has
many constraintstofollow,and theseconstraintsshou[d becaretully considered whendeciding
the
design
parametersand theirrange.
Only
two parameterswere used forthecross section shape of the car inorder to minimize the samples needed forthe survey(Figure
2).The focuswas on provingthatthe method works rather than tomake a production-qualitymodel, andthereforethissimplification was made. ai sets the tumble-home,theangle ofglassfromthe
beltline
totheroof as viewed fromthe frontor the rear of thevehicle, while tvi controls the shoutdeltwidth ofthecar,2.2. Form lmpressionAttributes
lnorder todescrlbean image innatural language,aset of tenattributes were chesen as
parameters
torepresent theform impressionof a shape, as perceivedby a customer.
A 5-pointscale was used to
put
weights on each attribute,therebyyieldinga form impressionvector Fliketheone in
lable1
.
Ahigherweight signifiesahighercorrelationtotheat-tribute.
Ttable.1
.
Weighted SetofAttributes CuteSportyCIassicFermalPewerful2 s 3 3 4
ModernRebustSpaciousSleekLuxurious
4 4 1 4 3
2.3.
Survey
and faguchi MethodA survey was designedand conducted inorder tocollect
dataon the
form
impressions
of various shapes, The aim forthissurvey was to provethe methodology, not togatherdata foran in-depthanalysis offormimpressionsforcars, and the results should
be
vlewed with consideratlon tothescarce amount ofdatacol]ected.Everydesign
parameter
wilfigivea contribution tothe
form
impression,and thereforetheymust all
be
accounted forandchanged uniformly torthesamples
included
in
thesurvey,fofacilitatethecreation of samples and ensure a minimum of samp[es with uniformly changed parameters,thelaguchi
Anethodwas used fordesigningthesurvey, Each
parameter
fortheside profileshape was
given
threelevelstochangebe-tween,and using theorthogonal array developedbylaguchi
[1],
27 differentsamples weregiven.
Eachsample consistedof a si[houette of the side profileof thecar, as displayedin
Figure3.The frontparameter$were restrictedtotwo levels
in
order to minimize the contribution tothesurvey with foursamples
(omitted
from
Figure
3),Furthermore, thesurvey was dividedintoseparate partsfor thecross section and side profileshapes, inorder tosimplity
thesurvey bypresenting2D-silhouettesinsteadofmore
com-plex3D-models,Viaa web-browser interface,thesesamples
were presentedto21 individualsconsisting ofmen and
wom-en intheages between
20
and30
yearsold.The
`average'model inFigure3 was created withthe
parameters
given
by
thesecond level,and displayedas an example toeach indi-vldualbeforethesurvey was
given.
The
form
impressions
for
thesamples were collected and saved ina database.When thesurvey was completed bythe
group,
theaverage response foreach model was calculated and used as a value to de-scribe theformimpression.Forthiswork, the
permissible
range ofthe[evels
were ratherlarge
in
order tocreate siightlyexaggerated samples whichwouid be easierto separate forthe peopletakingthe survey.
For
example, thehood
was givenlevelsof30-, 50-,and70-]ength
unitstochange between.[na productionenviron-ment, thedesignteam willaiready have an ideaor concept
7'ifly7mvenstg
specia[issueofjapanesesocietytorthescienceefdesign
NII-Electronic Library Service
Aeeercae
Arrpde-5
AmotieLle
Artnden5
ArredetaO
n
rr"demsFig.3, A#ndemm rrmde-2A
A
mode-coAmedelll
AnoSe"E
"pdeRl mhd}6 "mde1ew i,itMttmdeE12Arrnden?
modaaenAmodeza1
27SamplesforSurvey(Side
Shape)Aptodems
Amodems
vaodeL13Amdens
..d-bmodeuaAmodeM4
AmodeR9
Amoden4
,,A,motiengAmedee4
of what typeof cartheywjll develop,and ±heywillalso have toconsider many engineering and manufacturing constraints. Thesefactorswili[eadtoa much smaller range of permissible values
for
thelevels,
and thuscreate more realistic samples which show lessvariation.2.4.
DataProcessing
by
FactorAnalysis
The datafromthesurvey
provided
amultivariatedatasetof12parametersmeasured over
27
samples, yieldinga 27×12primarydatamatrix, However,thisdatasetshowed thatsome attributes hadsimilarities and were correlatedtoeach othen fo eliminate
thls
overlap ofmeaning,Principal
Component
Analy-sis
{PCA)
was used toextractcorrelatedfactors.The goalof PCA is
to,
via analysjs of eigenvectors andeigenvalues, finda transformation matrix thatwill
provide
anew set of coordinate axes where the
data
canbe
projected insuch away thatthevariance ismaximized alongsubse-quent,orthogonal axes
(Principal
Component
Axes).Aseachextracted factoraccounts forlessand iessvariance inthe data,itispossjb[eto obtain a reduction ofparameterswhile
preserving
most oftheinformationin
thedataset
{Gorsuch
[2]).
Thisreduction inparametersisimportantinorder tolowerthedimensionalityof theproblem
-
according toFriedman
[3],
a40T-ff()\ffvetskg
speclalissueetjapanesesocietyforthescienceefdesjgn vol.15-4 no.60 2008 4.S ScrecTest 4S,53 2.1.sI"21.51e.5o'
x
x
1x1×
'xx・xx-x-x.
'
xu.7-M--.ts
'
12S Fig,4,Screele$t 45e NumberotFaETors7B910
function
defined
in
high
dimensional
space jslikelytobe more complex thanonein
alower
dimensionalspace, and theretore harderfor
thenetwork tosolve.The number of
factors
toextract israther subjective andthereare many techniquesavailabletoajd inthisdecision,
Generally,
asufficientnumber of factorsmust beextractedtoaccurately reproduce thedatamatrix fromfactorloadings
and
factor
scores, butthegoalotthisanalysjsis
stilltoextracta limitednumber of factorsthatwill contain themaximum amount of
information.
A
Scree
lest,developedbyCattell[4],
was performed
but
provided
inconclusiveresults.Accordingto Figure4,the number ofextracted factorsshould beeither
fouror six.The Kaiser
Criterion
[5],
which statesthatonly fac-torswhose eigenvalues aregreaterthan 1.0should bekept, suggested thatthefirst
threefactors
would sufficetoexpressmo$t of the informationcontained inthedataset,Therefore,
with theresult$ fromthesetwotests,we
decided
toretainfour factors,which accounted for90.3%
ofthevariance,lnaddition toextracting theprincipalcomponent axes, the
PCA also providedthe factorloadingmatrix which contains
thevariable loadingson each oftheretained axes, thus
dis-playing
thecorrelationsbetweentheattributes and thefac±ors.Inthenext step of thefactoranaiysis the principalaxes were
Varimax-rotated
{algorithm
fromHarman[6])
in
order toforce
them toalignas closely as possiblewith strongly correlated
subsets among theformimpressionattributes,
The
idea
ofthisprocedureistoobtain a simpler structure which heipinthe
interpretationand labelingofthefactors.Inasimple structure each factorloadshighlyon afewvariables, and each variable
loads
highly
on on]y one factor.Furthermore,variables thatlable.2,Varimax RotatedFactorLoadingsforAttribute$ Factor1Factor2Factor3Factor4 Cute
.O.3503ffO,0882.o.o"giil・lll,iE.lsiSlllll$/E'li
Sortl/{/l///,t,ff'"i'/i""'t'tth'.O.l311O.lle7O.l958 Classie-O.2124-O.]413l・ll・l-l・"・llasSiee;,l-O.O180 Format-O.18I2-O.3151Lll//es,ime.'#i・i'-e.o2og PowerfulO.0667l;・/Y,i',.・esi.pa.l・ll-O,O122-e.e154 Medern11illliilg'li'・"・ss"x-e.2499l・111・ltl,#///ewl.lll'O.0994 Rebust.O,1126'Ei,sw"'lif'i'.'iifgl'-O.4i31-O.1333 Sacieusl・l・lt,l・llE'l,eswag・ 1-e,3621ffO.1074-O.1452
Sleek1il,IIIIict"i・・Sgewl'i'-O,1451O.216902035 LuxuriousO.3602l\,l.waee.,gl・i''g/l'll.'os'"S.,l'.'I"ll・O.1100loadheavilyon thesame factorare related, whereas unrelated
variables would loadon differentfactors,Forthesignificance of theseloadings,Hairet al.
[7]
suggest a guideiinewhere loadingslangerthan ±O.5can beconsidered as practically sig-nificant, thatis,they haveameaningful effectonthe varlables, The rotatedfactor
loadings
aredisplayed
in
lable
2
with the significant loadingshigh[ighted,An
analysis ofthistablecangivean
interpretation
ofthefactors,
but
for
thefunction
ofthe systemdescribed
in
thispaper,aformal
labe[ingofthefactors was not necessary, and thereforeomitted.Withthisfactorloadingmatrixitwas
possible
toproject
thesamples, expressed with theiroriginalform impression at-tributes,intothenew space spanned bytherotated principal cornponent axes bycomputing thefactorscores
(coordinates
in
PC
space)for
each sample. Thefactor
score matrix Swasgiven
byi
S-XB
(1)
where X isthe
data
matrix, standardized with meanO
andstandard deviation1, and B isa score coefficient matrix such
±hat:
B-ACi
{2)
A istherotated factorloadingmatrix, and C represents the variance-covariance derivedfromthe retained factors,given by:
C-ATA
(3)
The factorscores could be calcula ±ed with
(l)-(3),
and everysample was projectedintothenew fouFdimensional
image-space.2.5.
System
Construction
UsingNeuralNetworkThe most importantcomponent ofthissystem istheneural
network, Itisresponsible foraccurately mapping the
image-space tothedesign-spacespanned bythedesign
parameters.
A properfiyset up and trainedneura] network hastheability
to
generalize,
thatis,produce
accurate outputs forinputsnotencountered duringtraining.HoweveL itisdjfficulttoachieve good generalization,and inorder tomake a reliable system itisessential to make sure thatthe network solution is
ac-curate
by
validatingit.
Generalization
is
influenced
by
the
size and qua]ityofthe trainingset,the architecture of the neuralnetwork, and thephysicalcomplexity of theproblem,
There
isa tradeoffto beconsidered regard[ng the tra[ningset-
thelaguchi
method willminimize thesamples and make thesur-vey easier toperform,butat thesame time theneural network
trainingwill
benefit
from
aiarger
dataset,
With
afixed,
small trainingsetand no way tocontrolthecomplexity oftheprob-Iem,thenetwork architecture was carefullychosen toable to
represent theunderlying problemand achieve a
good
general-ization.
A
feed-forward,
back-propagating
neural network(Haykin
[8])
was constructed withfournodes intheinputlayer(image-space), and
1
2
nodesin
the outputlayer
(design-space),
corresponding tothedesignparameters
forthesideprofile
shape. Thisnetwork usesgradient
descenton theerrorsto traintheweights, with adifferentiable activation functionoftheweighted sum ofinputsv,definedby:
1
ep(v)=
(4)
1+expGvi
As
thissigmoid bounds theoutputin
theinterval
O
to1,
it
was necessary tonormalize thedesign
parametersfor
the samples inorder to use them as the desiredoutput, or target values,forthetraining.Witha hiddenlayerof nine nodes, a 4-9-12 structure of theneuralnetwork was used
for
thetrainingofthesideprofileshapes.
fo
avoid overtrainingandlosing
theabilityto general-ize,errordecay was implemented intothe network, while a $malllearning
rateand momentum providedstable andef-ficientlearning.The number of hiddenunits was a critical pa-rameter ofthenetwork
-
toofew unitsin
thehiddenlayerwillnot givethe network enough flexibilityto properlyrepresent
theunknown underlying function,whereas too many units may leadtoa network thatalso fitsthenoise, not
iust
thesignal, leadingtooverfitting.Neuralnetworks trainedwith a scarce amount of case$
are pronetooverfitting, and validation must beperformed
T-iftyvmxrvsee
speclalissueotjapanesesocietyforthescienceofdesLgn
NII-Electronic Library Service
totesttheperformance of the network. The cross section
profile
$hapes were trainedseparately on a differentnetworkarchitecture,with a datasetof fourcases, compared tothe
27 cases available fortheside profile.Due tothelimited
data
available, theuse of a testset would waste a lotof datafor
the trainingand therefore multifold cross-validation methods
seemed appropriate forthissystem, as allthecases can
be
used inthe training.Leave-one-outcross-validation was
per-formed toestimate theperformanceof a number ofnetwork
models, and aid inthe selection ofthe
best
model,This
meth-od was alsoimplemented
toincorporate
earlystoppingin
thetraining.Withthese measures taken,
it
waspossibly
toachieve a neural network with good generalizationdespitethesmall training$et.2.6.
AutomaticGeneration
of3D-model
With
thedesign
parametersgiven
bythe neural network solution,a3D-mode[
ofthedesiredcarcould beconstructed,The
aim ofthissystem, programmed inOpenGLTM,
was togenerate
a simple model and displaytheproportions
suggest-edby
theneural network so]ution,A series of seven Beziercurves
(Farin
[9])
oforderfourwerejoined
togethertoformthemidline,side
profiie
curve inFigure5.This curve was duplicated,transposeda]ong thez-axis to formtheshouider-Iine curve. Withthesetwo curves laidout,
B6zier
surfaces were used toconnect thecurves and create thesurfaces forthemodel.
Bezier
curves passthroughtheirendpojnts, which allowed the endpoints forthesjdeprofilecurve tobe
placed
accordingtothesolution presented
by
the
neural network, and therebyyieldthe desiredproportions.Furthermore,one desirable
featureof thiscurve was to allowforlocalchanges ofcontrol
pointswjthout affecting the shape of the whole curve.
There-forethe
jolnts
were restricted totangent continuity{C1
}.
The interiorcontrol pointsalso
influence
theshape ofthe curve, but as thiswork focuseson the basicproportions, thesepolntswhere onlyusedto
givethecurve somesmooth-ness, and considered tohave no effect on the form impres-sion,
3.
Resultsand DiscussionWiththesimple CAD-system outlined intheprevious sec-tion,itwas possibletovisualize how changes of theattributes
weights, controlled by thedesigner,affecttheshape ofthe
can Figure6shows two models with formimpressionvectors
42fifty#ewvek:e
specialissueefjapanesesocietyferthesclenceofde$ignvol.15-4 no.6e 2008
Roqf
Fig.5.SideProfi]eShapebyB6zierCurves
i''" I・
t't't
'tttt
t't'tttt'ttttt
t'tt'
tttti
.t
'./-...t・..1.ni.・...----,.-
---
-- --- ---
--i]l
t.t
/
Fig.6.Example of3D-models
R=l3333333333]andFle-[3333335533].
Thatis,the
parameters
forRobust and Spacioushave beenincreased,
resultingina model which seems tocarry thosefeatures
bytheincreasedsize ofitsglasshouse,
Thisshowsthatthjssystem
ha$
theabilitytoconstruct a numeric3D-model tofitan image thatiseasilyalteredbymanipulating the
weights ofthe
form
impressionvectot Adesignercanget
aninstantfeedbackon theshape
given
by
any formjmpression,and this
feedback
isstillcarryjng theinformationcollected in thesurvey,and thusincorporating
thevoice ofthecustomerin
the creativeprocess.fo
conclude, a3D-model
of acar was createdfrom
design
parameters,suggested bya neural network trained to relate
subtle form impressionsto
basic
proportions.Validation
oftheneural network performanceand avisual inspectionof the
created models indicatedthatthe system could produce
ac-curate results. Thjsshows thatitispossibietocreate a design support system with sensibility to shapes, which can aid the
designerinthecreativeprocess
by
v[sualizingthe
relationshipbetween imageand design,and presentanumeric 3D-model.
Theseresults are
promising
fortutureresearchin
thismeth-od, which byno means isrestricted toonly
proposing
basic
proportionsof a car, ltcould beused inany situationwhere
the designparameters,relating toa speciflc image,are
de-sired, One caveat of thissystem isthefactthatform impres-sions are very complex, and not only associated witha
few
el-ements ofaproduct,Foracarferinstance,thecoloc material, sound,details
inthe designand even themarketing effortto promoteaspecificimage
willallcontributetotheimpressionspeople
get
when theysee it,Thus theextraction a fewparam-eters
from
such a complex pictureisa simplificationof thisproblem.
However,a largeset of parameterswould leadtoasurvey withso many samples thatitwould bevirtually
impos-sibletocarryoutinan efficient mannen
References
1
.
laguchi,Genichi,2005,
laguchi's
Quality
Engineering
Handbook, John Wiley&
Sons,
inc.,Hoboken,
Newsey.2.
Gorsuch,RichardL, 1983, Factor
Analysis,
2"d
Edition,
LawrenceErlbaumAssociates,lnc.,Hillsdaie,
New
Jersey.
3. Friedman,J.H.,1995, "An overview of
prediction
ingand functionapproximation," lnXL
Cherkassky,
J.H.
Friedman,and H.Wechsle4eds.,FromStatistics
toNeural Networks:Theoryand PatternRecognitionApplications, SpringeFVerlag,New Nbrk.4.
Cattell,
R.B.,1966, The scree ±estforthenumber of tors,MultivariateBehavioralResearch,1{2),
245-276,5. Kaiser,H.E, 1960, The application of electronic
puters
tofactoranalysis,Educationaland PsychologicalMeasurement,20:14I-51.
6.
Harman, H.H.,1976,
Modern
Factor
Analysis,
3rd
Edition,Universityof
Chicago
Press,
Chicago,
lllinois.7. HainyJ.EJr.,Anderson,R.E.,
latham,
R.L,&BIack,W.C.,1998,
Multivariate
Data
Analysis,
5'h
Edition,PrenticeHail,
Upper
Saddte
Rive4New Jersey.8. Haykin,
Simon,
1999,
Neural
Network$:
Asive
Foundation,
2"d
edition,Prentice-Hall,
lnc.,Upper
Saddle
River,
New
Jersey.
9.
Farin,G.E,
1997,
Curves
and surfaces forcomputeraided geometric
design
:A practicalguide,4thEdition, Academic Press,San Diego,California.ittF-y\mmaIg・
specialissueotjapanesesocletyferthescienceofdesign