UNIVERSITY
OF TRENTO
DIPARTIMENTO DI INGEGNERIA E SCIENZA DELL’INFORMAZIONE
38050 Povo – Trento (Italy), Via Sommarive 14
http://www.disi.unitn.it
SVM-
BASED
C
LASSIFICATION
A
PPROACH FOR
S
YNTHETIC
-I
MPULSE
M
ICROWAVE
I
MAGING
–
SVM
I
NPUT
D
ATA
F. Viani, M. Benedetti, M. Donelli, L. Lizzi, A. Cresp, I. Aliferis,
C. Pichot, and A. Massa
October 2008
InformationandCommuni ation Te hnology Dept. Universityof Trento
ViaSommarive14,38050Trento,ITALY Phone+390461882057Fax +390461882093
E-mail: andrea.massaing.unitn.it
Contra t No. DIT-PRJ-08-005
Te hni al Report N. 3
SVM-based Classi ation Approa h for Syntheti -Impulse
Mi rowave Imaging
SVM Input Data F.Viani
(1)
,M.Benedetti(1)
,M.Donelli(1)
,L.Lizzi(1)
, A.Cresp(2)
,I.Aliferis(2)
,C.Pi hot(2)
, andA.Massa(1)
(1)
DepartmentofInformationEngineering andComputerS ien e ELEDIAResear hGroup
UniversityofTrento,ViaSommarive14,38050Trento-Italy Tel.+39 0461882057,Fax. +390461882093
E-mail: andrea.massaing.unitn.it
(2)
Laboratoired'Ele tronique,AntennesetTélé ommuni ations(LEAT) UniversitédeNi e-SophiaAntipolis
250rueAlbertEinstein,06560Valbonne
Tél: +33(0)492942802,Fax: +33(0)492942812 E-mail: hristian.pi hotuni e.fr
1. Introdu tion 3
2. SVM InputData 3
Inthelast years,thelearningmethodologyhasbeeninspiredbytheoryofstatisti allearningleadingupto
solutionswithgoodperforman eandrmmathemati alproperties. Inthisframework,thetheoryofsupport
ve torma hine (SVM)isbasedontheintera tionbetweenoptimizationtheoryand kerneltheory[1℄.
Re ently, widely used ma hine learning algorithms have been su essfully applied in the framework of
wireless ommuni ationproblems[2℄ andinverses atteringproblems[3℄[4℄inorderto exploittheir
general-ization apabilities and real-time hara teristi s. Moreover,when a losesolution to the problem at hand
doesnot exist, SVM appears to be a good andidate to solve the optimization problem with a trial and
errorapproa h. Asfor theSyntheti -Impulse Mi rowaveImagingSystem (SIMIS)developed atLEAT, the
learning methodologyadopted for the dete tionof target position anbe onsidered asupervisedlearning
sin eitexploitsinput/outputexamplesthatarereferredtoasthetrainingdata. Whenanunderlying
fun -tionfrominputsto outputsexists,itisreferredtoasthetargetfun tion. Intheframeworkof lassi ation
theory,thisfun tion is alledde isionfun tion andgivesbinaryoutputsifabinary lassi ationproblemis
dealt with, otherwiseit givesanitenumberof ategories formulti- lass lassi ation. The omputational
timesavingprovidedbyanonlinebinary lassi ationapproa hjustiessomelimitationslikethequalitative
re onstru tionoftheobje tpositioninsteadofthequantitativeestimationoftheele tromagneti properties.
Within theintegrationof aSVM lassierand theSIMISforobje tsdete tionandmorein generalforthe
re onstru tionoftheinvetigationarea,themaingoal onsistsinthedenitionofariskmapofthepresen e
ofthetargets.
2 SVM Input Data
The lassi ationpro edureisbasedupontrainingandtestingdata. Inthetrainingsetea hinstan e onsists
inatargetvalue,thatisthelabelofthe lass,towhi hdatabelongs,andseveralattributes,namedfeatures.
Thelearningma hinerequiresthatalldatainstan esarerepresentedasave torofrealnumbers. Inorderto
denetheformatofthepro esseddata,letusbrieyintrodu etheSVMbasi s. Thetrainingsetis omposed
by
L
observations,ea hofthem onsistsofpairs(x
i
, y
i
) , i = 1, ..., L
,wherex
i
∈ R
n
istheve torofattributes that determines theinput spa e andy
i
∈ {1, −1}
is theasso iatedtarget value givenby atrusted sour e. Theve torsx
i
, i
= 1, ..., L
aremappedinahigherdimensionalfeaturespa ewhereaseparatinghyperplane hastobefoundinorder tomaximizethemarginbetweenthetrainingdatathat belongto dierent lasses.AsfarastheSIMISdatahavebeen on erned,theelds olle tedbythemeasurementsystembe omea
partoftheinputve tors
x
i
, i
= 1, ..., L
. InordertoexploittheUWBpropertiesoftheM
exponentialtapered slots (ETS)antenna,theinputdataI
i
arethetime-gateddierentialtime-domainS
21
data[5℄obtainedby the appli ation of the inverse fourier transform. The position of then
-obje t, identied by its bary enterX
obj
n
, Y
n
obj
, is also afeature that belongs to theinput ve tor. The adopted binary lassi ationrequires
that inputdata
x
i
arelabeledbothto thepositive lassy
i
= +1
andnegative lassy
i
= −1
. Towardsthis end, for ea h input dataI
i
,K
training data are labeled with negative lass and, instead of the features related to the obje tposition,random spatial pointsX
emty
k
, Y
empty
k
, k = 1, ..., K
where the obje t does notresideareusedasshownin Tab. I.[1℄ V.N.Vapnik,Thenatureofstatisti allearningtheory,statisti sforengineeringandinformations ien e.
SpringerVerlag,2ndedition,1999.
[2℄ M.J. Fernandez-Getino Gar ia, J. L.Rojo-Alvarez, F. Alonso-Atienza, M. Martinez-Ramon, Support
ve torma hinesforrobust hannelestimationin OFDM,IEEE SignalPro ess.Lett.,vol.13,no.7,pp.
397-400,Jul. 2006.
[3℄ I.T.Rekanos,Neural-networkbasedinverse-s atteringte hniqueforonlinemi rowavemedi alimaging,
IEEETrans.Magneti s,vol.38,no.2,Mar.2002.
[4℄ A.Massa,A.Boni,andM.Donelli,A lassi ationapproa hbasedonSVMforele tromagneti
subsur-fa esensing, IEEETrans. Geos i.RemoteSens.,vol.43,no.9,Sept.2005.
[5℄ V. Chatelee, A. Dubois, I. Aliferis, J.-Y. Dauvigna , C. Pi hot, M. J. Yedlin, Real data mi rowave
imaging and time reversal, IEEE Antennas Propagat. So . Int. Symp., Homolulu, Hawaii, USA, pp.
y
i
x
i
+1
X
1
obj
, Y
1
obj
I
i
+1
X
2
obj
, Y
2
obj
I
i
. . . . . . . . .+1
X
N
obj
, Y
N
obj
I
i
−1
X
1
emty
, Y
1
empty
I
i
−1
X
2
emty
, Y
2
empty
I
i
. . . . . . . . .−1
X
K
emty
, Y
K
empty
I
i
Tab. I-Inputdataformatofa