Saliency-Detection-via-Dense-and-Sparse-Reconstruc

SaliencyDetectionviaDenseandSparseReconstructionXiaohuiLi1,HuchuanLu1,LiheZhang1,XiangRuan2,andMing-HsuanYang31DalianUniversityofTechnology2OMRONCorporation3UniversityofCaliforniaatMercedAbstractInthispaper,weproposeavisualsaliencydetectional-gorithmfromtheperspectiveofreconstructionerrors.Theimageboundariesarefirstextractedviasuperpixelsaslike-lycuesforbackgroundtemplates,fromwhichdenseandsparseappearancemodelsareconstructed.Foreachim-ageregion,wefirstcomputedenseandsparsereconstruc-tionerrors.Second,thereconstructionerrorsarepropa-gatedbasedonthecontextsobtainedfromK-meanscluster-ing.Third,pixel-levelsaliencyiscomputedbyanintegra-tionofmulti-scalereconstructionerrorsandrefinedbyanobject-biasedGaussianmodel.WeapplytheBayesformulatointegratesaliencymeasuresbasedondenseandsparsereconstructionerrors.Experimentalresultsshowthattheproposedalgorithmperformsfavorablyagainstseventeenstate-of-the-artmethodsintermsofprecisionandrecall.Inaddition,theproposedalgorithmisdemonstratedtobemoreeffectiveinhighlightingsalientobjectsuniformlyandrobusttobackgroundnoise.1.IntroductionVisualsaliencyisconcernedwiththedistinctperceptualqualityofbiologicalsystemswhichmakescertainregion-sofascenestandoutfromtheirneighborsandcatchim-mediateattention.Numerousbiologicallyplausiblemodel-shavebeendevelopedtoexplainthecognitiveprocessofhumansandanimals[12].Incomputervision,moreem-phasisispaidtodetectsalientobjectsinimagesbasedonfeatureswithgenerativeanddiscriminativealgorithms.Ef-ficientsaliencydetectionplaysanimportantpreprocessingroleinmanycomputervisiontasks,includingsegmentation,detection,recognitionandcompression,tonameafew.Motivatedbytheneuronalarchitectureoftheearlypri-matevisionsystem,Ittietal.[13]definevisualattentionasthelocalcenter-surrounddifferenceandproposeasaliencymodelbasedonmulti-scaleimagefeatures.Rahtuetal.[18]proposeasaliencydetectionalgorithmbymeasuringthecenter-surroundcontrastofaslidingwindowovertheentireimage.Whilecenter-surroundcontrast-basedmeasuresareabletodetectsalientobjects,existingbottom-upapproach-esarelesseffectiveinsuppressingbackgroundpixels.D-ifferentfromthecenter-surroundcontrast,localcontrastismeasuredbycomparingaregiononlywithitsrelevantcon-texts(definedasasetofregionneighborsinthespatialorfeaturespace)[9,14,4].Despitelocalcontrastaccordswiththeneuroscienceprinciplethatneuronsintheretinaaresensitivetoregion-swhichlocallystandoutfromtheirsurroundings,globalcontrastshouldalsobetakenintoaccountwhenonere-gionissimilartoitssurroundsbutstilldistinctinthew-holescene.Inotherwords,globalcontrastaimstocapturetheholisticrarityfromanimage.Recentmethods[7,8]measureglobalcontrast-basedsaliencybasedonspatiallyweightedfeaturedissimilarities.Perazzietal.[17]formu-latesaliencyestimationusingtwoGaussianfiltersbywhichcolorandpositionarerespectivelyexploitedtomeasurere-gionuniquenessanddistribution.In[4],globalsaliencyiscomputedinverseproportionallytotheprobabilityofapatchappearingintheentirescene.However,globalcon-trasthasitsinherentdrawbacks.Whenaforegroundre-gionisgloballycomparedwiththeremainingportionofthescene(whichinevitablyincludestheotherforegroundregionsunlesstheobjectboundaryisknown),itscontrastwiththebackgroundislessdistinctandthesalientobjectisunlikelytobeuniformlyhighlighted.Inaddition,priorsorheuristicsregardingthelikelypositionsofforeground(e.g.,neartheimagecenter)andbackground(e.g.,neartheimageboundary)havebeenshowntobeeffectiveinrecentmeth-ods[5,21,23].Inthispaper,weexploitimageboundariesasthelike-lybackgroundregionsfromwhichtemplatesareextract-ed.Basedonthebackgroundtemplates,wereconstructtheentireimagebydenseandsparseappearancemodelsfromwhicherrorsareusedasindicationofsaliency.Whiledenseorsparserepresentationshavebeenseparatelyappliedtosaliencydetectionrecently[8,4],thesemethodsaredevel-opedfordescribinggenericscenes.Inaddition,eachim-agepatchisrepresentedbythebaseslearnedfromasetofnaturalimagepatchesratherthanotheronesdirectlyfromthescene,whichmeansthatthemostrelevantvisualinfor-mationisnotfullyextractedforsaliencydetection.There-fore,thesemethodsdonotuniformlydetectsalientobjects2013IEEEInternationalConferenceonComputerVision1550-5499/13$31.00©2013IEEEDOI10.1109/ICCV.2013.3702976ĂĂĂĂ              !ĂĂ  (a)BackgroundTemplates(b)SaliencyviaReconstructionError(c)BayesianIntegrationFigure1.Mainstepsoftheproposedsaliencydetectionalgorithm.orsuppressthebackgroundinascene.Toaddresstheabovementionedissues,wemakefulluseofthevisualinformationbyusingbackgroundtemplatesfromeachindividualimageandreconstructimageregionswithbothdenseandsparserepresentations.Inthiswork,thesaliencyofeachimageregionismeasuredbytherecon-structionerrorsusingbackgroundtemplates.Weexploitacontext-basedpropagationmechanismtoobtainmoreuni-formreconstructionerrorsovertheimage.Thesaliencyofeachpixelisthenassignedbyanintegrationofmulti-scalereconstructionerrorsfollowedbyanobject-biasedG