Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
gitalimagecapturingdevicesenabletocollectalar
genumberofdigitalinformationandstorethe
minicomputerreadableformats.Thelargenumbe
Toretrieveanyimage,wehavetosearchforitamo
rsofimageshasposedincreasingchallengestocom
ngthedatabaseusingsomesearchengine.Then,t
putersystemstostoreandmanagedataeffectivelya
hissearchenginewillretrievemanyofimagesrela
ndefficiently.Althoughthisareahasbeenexplore
tedtothesearchdone.Themainproblemfo r t he
dfordecadesandmanyresearchershavebeendone
user
todevelopsomealgorithmsthatsolvesomeofitspr
isthedifficultyoflocatinghisrelevantimageinth
oblems,notechniquehasachievedtheaccuracyof
islargeandvariedcollectionofresultedimages.T
humanvisualperceptionindistinguishingimages.
osolvethisproblem,text-basedandcontent-
To
basedarethetwotechniquesadoptedforsearcha
fulfill
the
latest
technological
ndretrieval.
The
tobuildmoregeneralizedCBIRsystemwhichincr
8
main
objective
of
this
paper
is
provide
more
Toimprovetheretrieval
approach
and
heuristic
approach
are
accuracy
thesystemhastakenthefeedbackfromtheuser
automatically.
retrieval.
Nowadays
evaluation
are
precision,recallandretrievaltime.
The
qualitative
national
and
geographic
quantitative
imagery
image
metricsproves togood.
Keywords:
CBIR, Genetic Algorithm, HARP Algorithm,
Precision, Recall.
1. Introduction
known
have
image
largeimage
to
search
various
built-in
as
querybyimagecontent
database
based
(QBIC),
on
which
content
low-
drawback.
Dissimilarity
is
represented
is
also
features,
by
their
quantitative
10
Clustering
patternsintogroups.Itsmaintaskistoassigningase
tofobjectsintogroupssothat
theobjectsinthesameclusteraremoresimilartoea
chotherthantothoseinotherclusters.
Inthis
paper,Color,TextureandShapefeatureswereex
tractedandcombinedtoformfeaturevectorofim
age.Forcolorfeatures,themomentsofthecolordi
stributionwerecalculatedfromtheimagesandus
cluster.
edascolordescriptor.Fortexturefeatures,
The
subsequent
process
is
istheunsupervisedclassificationof
weusedGaborfilter,whichisapowerfultexturee
xtractiontechniqueindescribingthecontentofi
on the implementation.
mage.Forshapefeatures,edgehistogramfeature
InsteadofRelevance
sthatincludefivecategorieswereusedasshapede
Feedbackwecanuseanyclusteringalgorithmtha
scriptor.Thesethreedescriptorswerecombined
tbasedonthefeaturesextractedfromtheimagest
andoptimizedusingGAwith
hemselves,andallocatesthoseimagesintothene
clu st ering
arestcluster.Thealgorithmcalculatesandallocat
accuracyasafitnessfunctiontoselectoptimumw
eightsoffeatures.WeperformedGAwithH A R
of featurepoints ineachcluster.
P
onthedatabaseasanofflinestep,
HARP
clu st er ing
and
thesystemdoesnotneedtosearchtheentiredatab
11
gesarerequiredtobesearchedforimage
similarity.
2. Previous Study
discussed below.
d) SharadhRamaswamyet.al.have published a
are
have
also
used
till
the
exact
nearest
retrieved
texture
histogram.
Image
Input:
Thelearningphasetellsaboutthetrainingproc
the user.
esswhichahugeamountsampleimagesareinp
utinthefirststep. The geneticalgorithm is
used
to
trainthefeatureswithdifferentweights.For
optimizing
thefeatureweights
fitnessfunction,
13
HARP
and
algorithm
for
is
histogram,
color
momentsand
esultandstoresitinthefeaturedatabase.Allthe
sestepsperformedofflineandeachclasswillb
retrieving
eindexingalongwithitsassociatedclassIDint
heindexfiles.
features.
Thefeatureofeachimageisverymuchsmaller
insizecomparedtotheimagedata,so
thefeaturedatabasecontainsanabstractionoft
heimagesintheimagedatabase.
audiovisual
data.
color
The
or
descriptors
in
CBIR
systems
color space.
14
needed.
wavelet
transform,
co-occurrence
iteration to another.
HARP-a
Clusteringalgorithm:
algorithm
is
approach.Initiallypick
based
on
up
The
bottom-up
each
element
used
complicated
following operators:
to
find
solution
to
15
4. PerformanceEvaluation
Here
system.This
weintroducethedatabasethatweselecttotestour
phasedescribestheimagessearchingprocess.
system,andwealsocomparethe
Theuserentersaqueryimageforwhichthesyste
systemresultswithsome
mextractscolor,textureandshapefeaturesthefea
CBIRsystems.
turesvectorsofdatabaseimagesarepreviouslyex
Theimagesdatabasethatweusedinourevaluatio
tractedandstored.
nisWANGdatabase.ItisasubsetoftheCoreldata
Similarity
baseof1,000imagesinJPEGformat.
Matching:Usingthesimilaritymetricsdefined
1,000imagedatabasewentthroughourimpleme
forcolor,textureandshape,thesimilaritydistan
ntedsystemtoextractthefeaturesandstoredthe
cesbetweenthequeryimageandthecentroidim
m.TheextractedfeaturesareweightedbyGAan
ageofeachclassarecalculated.
dtheyareusedforclassificationb y
Thesmallestdistance(mostsimilar)willdeterm
HARP
inetowhichtheimagebelongs.Theclasswithth
algorithm.Thelevelofretrievalaccuracyis
esmallestdistanceisreturnedandtheimagesint
fa ct o r
hisclasswillbecomparedwiththequeryimage.
performance.InCBIR,the
Retrieval:The
used
mostmatchingimageswillberetrievedand
performancemeasuresarePrecisionandRecall.
thentheyare
Precisionisdefinedastheratioofthenumberofr
ThefirstN
sorted
in
ascendingorder.
similartarget
to
new
other
inf lu e nc e
most
existing
usingthe
a
t he
commonly
images
etrievedrelevantimagestothetotalnumberofret
withsmallestdistancevaluetothequeryareretri
rievedimages.Thismeansthatprecisionmeasur
evedandshowntotheuser.
estheaccuracyofthe
16
retrieval.Recall
t h e accuracy.
evantimagestothetotal
Here
numberofrelevant
by
usingclusteringpre-
processofthedatabaseimagevia
therobustnessoftheretrieval.
algorithmdecreasestheaveragequeryresponset
InCBIR,iftheprecisionscoreis1.0
then
H AR P
ime,
everyimageretrievedbyasearchisRelevant.
thesimilaritysearchtimeforimagematching
are
robust.
Comparisonof
u s in g
otherexisting systems
retrievedbythe
search
is
Weevaluatethen e w systemb y
the
new
systemwith
twometricsviz:theRetrievalEffectivenessandth
Foreachclassinthedatabase,werandomlyselect
eRetrievalEfficiency.
ed20imagesasqueries.Sincewehave
a. RetrievalEffectiveness:Aretrievedimageis
3 classesinthedatabase,wehave60queryimages
consideredamatchifitisinthesameclassasth
.Foreachquery,wecalculatetheprecisionandrec
equeryimage.Thesystemworkswellanditre
alloftheretrieval.Theaverageprecisionsandthe
trievesb e t t e r
averagerecallforeachclassbasedonthereturned
resultsovertherandomlyselectedimagesasq
top20imageswererecorded.Moreover thenew
ueries
systemresultiscomparedagainsttheperformanc
by
using
GAand
HARP
a lgo r it h m.
weightstoeachfeature
CF
Average Precision
b. RetrievalEfficiency:Byassigningdifferent
to
CTF
0.8
improvetheefficiencywe
GAwitha
HARP
have
a lg o r it hm
used
CSF
0.6
to
CTSF
0.4
elephantClass ID horse
selectoptimumweights offeaturest o g e t
17
dinosaurs
1.1
Improved K-means
above
Multi-round RF
figureshowsthatthenew
Average Precision
0.9
0.8
classes
Interactive GA
except
elephant
clustering-based indexing
technique
New system
0.7
class.Thisresultconfirmsthatafusionofmultipl
efeaturescanincreasetheperformance
0.6
of
0.5
thesystem.
The
below
figure
showsthatthe
0.4
new
Elephant
systemperformssignificantlybetterthanothere
Horse
dinosaurs
Class ID
system.Thereasonbehindthelimitationintwocl
Interactive GA
Improved K-means
assesisthatthoseclasses'imagesareverysimilari
0.8
Average Recall
Multi-round RF
new
0.6
systemmayconfusedbetween
clustering-based indexing
technique
New system
0.4
them.
0.2
0
Elephant
Horse
dinosaurs
Class ID
The
above
figure
showsthatthenew
systemperformssignificantlybetterthanothers
ystems,forallclasses.Thismeansthatthenew
systemcanretrievemostofdatabaseimagesthat
matchqueryimage.The
new
systemworkswellintheclassificationpartofusin
gGAwith
HARP
algorithm.Theaverageprecisiona n d
theaveragerecallincreasedfrom78.1%to88.2%
results.
and
Theefficiencyof
50.4%to69.9%respectively
and
the
new
system
is
improvedbyconsideringcandidateimages
for
similaritycomputation
not
purpose
i.e.
candidateimageliesinthesameclusterwiththequ
past
decade
to
design
Content-
efficient
basedImageRetrievalSystems,IEEE
Feng,FundamentalsofContentBased
5. R.Chang,J.Ho,S.Lin,C.Fannand
ImageRetrieval,MultimediaSignalP
Y.Wang,ANovelContentBasedIma
rocessingBook,Chapter1,SpringergeRetrievalSystemusingKVerlag,BerlinHeidelbergNew York,
meanswithFeatureExtraction,Intern
2003.
ationalConferenceonSystems
and
Retrieval, published
International
6. I.El-
Journal of
Naqa,Y.Yang,N.Galatsanos,R.Nishi
arningApproachtoContentBasedImageRetrieval:Applicationto
DigitalMammography,IEEE
2009.
7. B.WANG,X.ZHANG,andN.LI,
RelevanceFeedbackTechniqueForC
ontent-
BasedImageRetrievalUsingNeuralN
etworkLearning,ProceedingsoftheF
ifth
2006.
8. R.Datta,J.Li,andJ.Wang,Content20
ceptsandTechniques,2ndEd.,Morgan
ApproachesandTrends of theNew
KaufmannPublisher, 2006.
Age,ACMComputing Surveys,
10. P.JeyanthiandV.JawaharSenthil
Kumar,ImageClassificationbyK-
meansClustering.Advances
9. J.HanandM.Kambr,DataMiningCon
inComputational
Sciences and Technology, 2010.
Prof. S.SELVAM M.Sc., M.Phil, has been working as Assistant Professor and Head,
Department of Computer Application, N.M.S.S.Vellaichamy Nadar College, Nagamalai, Madurai19, Tamilnadu, India. He has more than 13 years of teaching experience. He has published three
research papers in various refereed International/National Level Journals/Proceedings and
Conference. His research paper also published in the book titled Research and Trends in Data
mining and Image Processing Technologies and Applications, Bloomsbury publishing India PVT,
LD, London, New Delhi, New York, Sydney pp-1-11, October 2014, ISBN: 978-93-84052-11-9.
Under his guidance two M.Phil scholar were awarded. His area of interest is Digital Image
Processing.
21
22