Multi-resolution Analysis based Salient Contour Extraction JingWANG*,KazuoKUNIEDAt, MakotoIWATAt, Hirokazu KOIZUMIt,
Takeshi Ikenaga, Satoshi Goto · 2006
Detecting salient contours incomplex backgrounds isimportant inimageanalysis andsceneunderstanding. The II.RELATEDWORKS local context ofanedgeorline segment feature iscommonlyused tomeasure itssaliency degree asapart oftheobject boundary. Significant amountofresearch hasbeenmadeinsalient However, traditionally thecontext information iscaptured by contour extraction inthepast. Alsoseveral saliency measures studying several features inthepredefined neighborhood. Inthis havebeendefined underthecommonGestalt psychological paper, anovel salient contour extraction algorithm based onthe principles(1) like continuity andproximity. Thefeatures that multi-resolution analysis isproposed andanewsaliency measurecanformlongandsmooth contours arefavored toassign high isdefined tocharacterize thesignificance offeature. Relation of features thatarecorresponding tothesamepartoftheobject lenve. boundary acrossresolutions isutilized toestimate thecontext Theter ofsaliency was fired b Shashua and information andfeature significance value. Experimental resultsUllman (2)intheliterature. Theydefined aparticular measure showthat theproposed method canextract salient contours more quantified theproperties ofsmoothness andlength. The efficiently thancenter-surround interaction based methods and saliency ofanelement isiteratively updated bythesaliencies still provide robust results. ofallneighboring elements initslocal network. Later this methodwasextended byAlter andBasri(3) toaddress the I. INTRODUCTION problem thatthemaximalsaliency measure maynotbe consistent tothemostvisually salient curve because ofthe Iuntaerptatiponbandldemscriptn cofpimageisione ofthelocal andgreedy nature ofthemeasure. GuyandMedioni (4) fundaentalprobems i comuter isioand mage proposed anon-iterative methoddenoted astensor voting. processing. However, real worldimages ofnatural scene are Arothea oneighborhod determineneoring abundant withdetail, texture andclutters that makeitdifficult .elemensco ne byrvtin fodeachmoter andithenral to~~~~ exlrth img.otn.Acrdn ohmnvso ~~~~~~~All theabovemethodsworkundera center-surround carry outsalient contour extraction online fitting results of . The a ffinity mer o everysparof ~~~~~~~~~many pairs andinstead computes onlyoneacross-resolution pair ofline segments that arecorresponding toeachother.