数字图像复原算法研究

华侨大学硕士学位论文数字图像复原算法研究姓名：陈华玲申请学位级别：硕士专业：通信与信息系统指导教师：冯桂20100501I[0]houghIIhoughIIIABSTRACTThetechnologyofdigitalimagerestorationisoneofthemostimportantresearchareasinthedigitalimageprocessingfield.Duringgeneration,transportationandrecording,theimagequalitywillbedecreasedduetotheinfluences,suchasthedeficiencyofimagerysystems,theinfluenceoftransmissionmedium,therelativemotionbetweencameraandobject,randomnoisesandsoonthuscauseformationofdistortedimage,defocus-blurredimageandmotion-blurredimage.Theaimofrestorationisimprovementofimagequalityandstudyinghowtogetrealimagefromrestorationofdistortedimage.Thatistosay,theaimofimagerestorationisstudyinghowtogettrueinformationfromgiveninformation.Basedonthetraditionaldigitalimagerestorationalgorithm,thisthesisdiscussednoiseremoval,parameterdeterminationfordefocusblurandmotionblur,andimagerestorationforunknowndegradationtype.Thoseimagesdegradationarecausedbyfactorssuchasnoise,defocusblurandmotionblur.Wehopethatafterrestoration,imagecanremainedgedetailsasfaraspossible.Inthemoduleofimagenoiseremoval,amixturefiltermethodisproposed.TheproposedmethoddetectedthepixelswhichwerecorruptedIVbynoisefirst,thenreplacedthepixelwiththevalueofthepixelthathadaminimalabsolutevalueinaspecificwindow.Inthisway,imageblurringcanbereduced.Theexperimentalresultsshowthattheproposedmethodeliminatesnoiseeffectivelywhilemaintainingthedetailinformation,especiallytheboundaryinformation.Startedfromthecauseofdefocus-blurringandmotion-blurring,thepapergivessomeimprovementbasedontheoriginalinverse-filter.ThencombiningwiththerelationshipbetweenblurredimageandFourierspectrum,wecanworkoutthePSFbyHoughtransform.Inthisway,originalimagecanbeeducedwithinverse-filter.Theexperimentalresultshowsthatrestorationoftheblurredimageisgood,andthemethodiseasytoimplement.Unfortunately,theaccuracyofHoughtransformparameter’sestimationneedsfurtherimprovement.Inordertoovercometherestrainedproblemsoftypicalimagerestorationmethods',thelargeestimationerrorofPSFandthecomplexcomputationinblindimagerestorationmethods,anadaptiveblindimagerestorationalgorithmisproposed.AccordingtothefeaturethatthecertainblurmayleadtothespecificfrequencycomponentdistortionoftheimageFourierspectrum,thedegradedimagesareclassifiedfirst:fortheusualtypesofblur,suchasmotionblur,defocusblur.WeestimatethePSFwithcorrespondingmethodfirst.Thenextstep,workoutthePSFofdefocusblurbyHoughtransform,andthePSFofmotionblurbyprincipleofVdirectionalderivativeandscaleidentification.Atlast,restoretheimagewithrestorationmethodofWienerfilter.KeyWord:imagerestoration,mixednoiseelimination,inverse-filteringrestoration,identificationofdegradedimage’sblurtype11.120X21.2Turky[1]Wend[2]Serra[3][4][5][6,7]Helstrom[8]HuntB.R.[9,10,11]Philips1998Kundur[12]SNRLagendijkR.L.[13]FriedenB.R.[14]1998ChanT.F.[15][16](xh)3xhAyersC.A.DaintyJ.C.[17]DaveyB.L.k.SeldinJ.H.[18]Wiener1995WienerWiener[19,20,21]2000N.X.Nguyen2002SendurLBivaShrink[22]Rudin-Osher-Fatemi(TotalVariation,TV)[23]2008[24]LYiCao[25]2009AihongLiuLSF[26]1.345[27][28]PSF2.12.1.1(,)fxy2()(,)(,)juxvyFuvfxyedxdyπ−+∞−∞=∫(2-1)(,)Fuv(,)(,)(,)FuvRuvjIuv=+(2-2)2()(,)(,)juxvyfxyFuvdudvπ+∞−∞=∫(2-3)221/2|(,)|[(,)(,)]FuvRuvIuv=+(2-4)61(,)(,)tan(,)IuvuvRuvϕ−=(2-5)222(,)|(,)|(,)(,)EuvFuvRuvIuv==+(2-6)(,)fxy|(,)|fxydxdy∞∞−∞−∞∞∫∫(2-7)2.1.2(,)fxy(0,1,...,1;0,1,...,1)xMyN=−=−M*N112()00(,)(,)vyuxMNjMNxyFuvfxyeπ−−−+===∑∑(2-8)0,1,...,1uM=−0,1,...,1vN=−112()001(,)(,)vyuxMNjMNuvfxyFuveMNπ−−+===∑∑(2-9)0,1,...,1xM=−0,1,...,1yN=−(,)fxy(,)Fuv(,)(,)(,)FuvRuvjIuv=+(2-10)(,)Ruv(,)Iuv(,)Fuv(,)(,)|(,)|juvFuvFuveφ=(2-11)1(,)(,)tan[](,)IuvuvRuvφ−=(2-12)7221/2|(,)|[(,)(,)]FuvRuvIuv=+|(,)|Fuv(,)fxy(,)uvφ(,)fxy(,)fxy222(,)|(,)|(,)(,)PuvFuvRuvIuv==+(2-13)M=N112()/00(,)(,)NNjuxvyNxyFuvfxyeπ−−−+===∑∑(2-14),0,1,...,1uvN=−112()/2001(,)(,)NNjuxvyNuvfxyFuveNπ−−+===∑∑(2-15),0,1,...,1xyN=−2.1.31(2-14)(2-15)112/2/00(,)[(,)]NNjvyNjuxNxyFuvfxyeeππ−−−−===∑∑,0,1,...,1uvN=−112/2/2001(,)[(,)]NNjvyNjuxNuvfxyFuveeNππ−−===∑∑,0,1,...,1xyN=−281122(,)(,),(,)(,)fxyFuvfxyFuv⇔⇔1212(,)(,)(,)(,)afxybfxyaFuvbFuv+⇔+3(,)(,)fxyFuv⇔(,)(,)FuvFumNvnN=++**(,)(,)fxyFuv⇔−−4(,)(,)fxyFuv⇔002()/00(,)(,)juxvyNfxxyyFuveπ−+−−⇔002()/00(,)(,)juxvyNfxyeFuuvvπ+⇔−−5(,)(,)fxyFuv⇔1(,)(,)||uvfaxbyFabab⇔6(,)(,)frFθωφ⇔00(,)(,)frFθθωφθ+⇔+9(,)frθ(,)Fωφ(,)fxy(,)Fuv(,)fxy0θ(,)Fuv0θ7112001(,)(,)NNuvfxyfxyN−−===∑∑0uv==(2-15)112001(0,0)(,)NNuvFfxyN−−===∑∑(,)(0,0)fxyF=8(,)(,),(,)(,)fxyFuvgxyGuv⇔⇔(,)*(,)(,)(,)fxygxyFuvGuv⇔•1(,)(,)(,)*(,)2fxygxyFuvGuvπ•⇔2.2()Hf(x,y)g(x,y)n(x,y)2.1(,)(,)(,)gxyHfxynxy=+(2-16)2.1g(x,y)n(x,y)f(x,y)h(x,y)10[29]1H11221122[(,)(,)](,)(,)HkfxykfxykHfxykHfxy+=+2H(,)(,)gxyHfxy=(,)fxyαβ(,)(,)Hfxygxyαβαβ−−=−−3H(PointSpreadFunctionPSF)2.2.1xy(,)fxyxy(,)fxy(,)(,)(,)fxyfxyddαβδαβαβ∞∞−∞−∞=−−∫∫(2-17)(,)xyδαβ−−δ(,)0xyδαβ∞−−=,,xyxyαβαβ==≠≠(,)1xydxdyδαβ∞∞−∞−∞−−=∫∫(2-18)(2-17)(2-16)11(,)(,)(,)gxyHfxynxy=+(,)(,)(,)Hfxyddnxyαβδαβαβ∞∞−∞−∞=−−+∫∫(,)(,)(,)fHxyddnxyαβδαβαβ∞∞−∞−∞=−−+∫∫2-19(,,,)(,)hxyHxyαβδαβ=−−H(,)αβ(2-19)(,)(,)(,)(,)gxyfhxyddnxyαβαβαβ∞∞−∞−∞=−−+∫∫(2-20)(,)(,)Hxyhxyδαβαβ−−=−−(2-20)(,)(,)(,)(,)gxyfhxyddnxyαβαβα