1
MixedNoiseRemovalbyWeightedEncodingwith
SparseNonlocalRegularization
JielinJiang,LeiZhang,andJianYang
Abstract—Mixednoiseremovalfromnaturalimagesisachallengingtasksincethenoisedistributionusuallydoesnothaveaparametricmodelandhasaheavytail.OnetypicalkindofmixednoiseisadditivewhiteGaussiannoise(AWGN)coupledwithimpulsenoise(IN).Manymixednoiseremovalmethodsaredetectionbasedmethods.They rstdetectthelocationsofimpulsenoisepixelsandthenremovethemixednoise.However,suchmethodstendtogeneratemanyartifactswhenthemixednoiseisstrong.Inthispaper,weproposeasimpleyete ectivemethod,namelyweightedencodingwithsparsenonlocalregularization(WESNR),formixednoiseremoval.InWESNR,thereisnotanexplicitstepofimpulsepixeldetection;instead,softimpulsepixeldetectionviaweightedencodingisusedtodealwithINandAWGNsimultaneously.Meanwhile,theimagesparsitypriorandnonlocalself-similaritypriorareintegratedintoaregularizationtermandintroducedintothevariationalencodingframework.ExperimentalresultsshowthattheproposedWESNRmethodachievesleadingmixednoiseremovalperformanceintermsofbothquantitativemeasuresandvisualquality.
IndexTerms—Mixednoiseremoval,weightedencoding,non-local,sparserepresentation.
I.Introduction
D
URINGimageacquisitionand/ortransmission,noisewillbemoreorlessintroduced.Denoising(ornoiseremoval)isafundamentalprobleminimageprocessing,aimingtoestimatetheoriginalimagefromitsnoise-corruptedobservationwhilepreservingasmuchaspossibletheimageedges,texturesand nescaledetails.Thepriorknowledgeofnoisedistributionplaysanimportantroleinnoiseremoval.TwotypesofcommonlyencounterednoiseareadditivewhiteGaussiannoise(AWGN)andimpulsenoise(IN).AWGNisoftenintroducedduetothethermalmotionofelectronincamerasensorsandcircuits[22].INisoftenintroducedbymalfunctioningpixelsincamerasensors,faultymemorylocationsinhardware,orbiterrorsintransmission[23].ManypapershavebeenpublishedonremovingeitherAWGN[12]-[20]orIN[2]-[11].ThemixtureofAWGNandIN,however,isalsocommonlyencounteredinpracticeduetothemultiplesourcesofnoise.Avarietyofmixednoiseremovalmethodshavebeenproposedinpastdecades[22]-[36].
AnimagecorruptedbyINwillhaveaportionofitspixelsreplacedbyrandomnoisevalueswiththeremainingpixelsunchanged.TwotypesofwidelyencounteredINaresalt-and-pepperimpulsenoise(SPIN)andrandom-valuedimpulsenoise
J.JiangandJ.YangarewiththeDepartmentofComputerScience,NanjingUniversityofScienceandTechnology,Nanjing,China
L.ZhangiswiththeDepartmentofComputing,TheHongKongPolytech-nicUniversity,HongKong.
Correspondingauthor.Email:cslzhang@comp.polyu.edu.hk.
(RVIN).AnimagecorruptedbySPINshowsdarkpixelsinbrightregionsandbrightpixelsindarkregions.Nonlinear lterssuchasmedian lters[1]havebeendominantlyusedtoremoveIN.However,oneshortcomingofmedian ltersisthattheimagelocalstructurescanbedestroyed,makingthedenoisedimageslookunnatural.ThisproblembecomesseriouswhentheINdensityishigh.Variousimprovementsofmedian ltershavebeenproposedtobetterpreservetheimagelocalstructures[2]-[10].Amongthem,theweightedmedian lter[2],thecenter-weightedmedian lter[3]andthemultistatemedian lter[4]donotdistinguishwhetherthecurrentpixelisanoisepixelornot,andtheytendtoover-smooththe nescaleimagedetails.AnalternativewayistodetectandprocessthecorruptedINpixels,andleavetheuncorruptedpixelsunchanged.Therepresentativemethodsalongthislineincludeswitchingmedian lter[5],adaptivemedian lter(AMF)[6],tristatemedian lter[7],adaptivecenter-weightedmedian lter[8],conditionalsignal-adaptivemedian lter[9],anddirectionalweightedmedian lter[10],etc.Thegeneticprogramming lter[11]byswitchingbetweentwoINdetectorsandtheirassociatedestimatorswasalsodevelopedforINremoval.
AWGNisthemostwidelystudiednoisemodelinimagedenoisingliterature[12]-[20].AteachpixelofanimagecorruptedbyAWGN,avalueindependentlysampledfromazero-meanGaussiandistributionisaddedtothepixelgraylevel.Traditionallinear lteringmethodssuchasGaussian lteringcansmoothnoisee cientlybuttheywillover-smooththeimageedgesatthesametime.Tosolvethisproblem,nonlinear lteringmethodshavebeendeveloped.Thewell-knownbilateral lter(BF)[12]isgoodatedgepreservation.Itestimateseachpixelastheweightedaverageoftheneighbor-ingpixelsbuttheweightsaredeterminedbyboththeintensitysimilarityandspatialsimilarity.Thenonlocalmeans(NLM) lteringmethod[15]canbeviewedasasigni cantextensionofBFbasedonthefactthatsimilarpixelsinanimagecanbespatiallyfarfromeachother.InNLM,eachpixelisestimatedastheweightedaverageofallitssimilarpixelsintheimage,andtheweightsaredeterminedbythesimilaritybetweenthem.Bygroupingthenonlocalsimilarpatchesintoa3Dcubeandapplyingtransformbasedshrinkage,theBM3Dmethod[14]hasbecomeabenchmarkforAWGNremoval.Zhangetal.[16]groupedthesimilarpatchesintoamatrixandappliedprincipalcomponentanalysis(PCA)toremoveAWGN.Theso-calledLPG-PCAalgorithmachievesverygoodedgepreser-vationperformance.Inrecentyears,thesparserepresentationanddictionarylearningbasedmethodshavebeenattractingsigni cantattentioninimagerestoration.Theseminalworkof
REGULARIZATION BY NONLOCAL CONDITIONS OF THE INCORRECT PROBLEMS FOR DIFFERENTIAL-OPERATOR E We consider the differential-operator equation du(t) dt + Au(t...
Edge-preserving tomographic reconstruction with nonlocal regularization We propose a new objective function for the image reconstruction problem, where the image...
Nonlocal regularization and High order TV regularization_工学_高等教育_教育专区 暂无评价0人阅读0次下载举报文档 Nonlocal regularization and High order TV ...
1 Nonlocal Discrete Regularization on Weighted Graphs: a Framework for Image and Manifold Processing Abderrahim Elmoataz, Olivier Lezoray, S? bastien Bougleux...
Super-resolution with Nonlocal TV Regularization ...In our method, a weighted fidelity term is ...details are enhanced simultaneously with noise ...
Nonlocal Regularization and Spontaneously Broken Abelian Gauge Theories for an Arbitrary Ga We study the non-local regularization for the case of a ...
Nonlocal Regularization and Spontaneously Broken Abelian Gauge Theories for an Arbitrary Ga We study the non-local regularization for the case of a ...
热门文档