基于纹理分析的板带材表面缺陷分类与检测识别方法研究

IUDC20041220051200412IIADissertationinMechanicalDesignandTheoryInspectionandClassificationofRolledSteelStripsSurfaceDefectsBasedonTextureAnalysisByFuHaiyanSupervisor:Prof.YanYunhuiNortheasternUniversityDecember2004III*FBP90%*(2003CCA03900)AbstractIIIInspectionandClassificationofRolledSteelStripsSurfaceDefectsBasedonTextureAnalysis*AbstractTherolledstripisoneoftheimportantproductsinthesteelindustryusedinautomobiles,electricalhomeappliances,shipbuildingandministryofaeronauticsandastronauticsindustry,whosesurfacequalitycanhasadirecteffectontheproductquality.Withthemarketcompetitionbecomingacutedaybyday,thesurfacequalityhasbecomeaverykeytargetinthedomesticandinternationalcompetition.Themainsteelcountriesallhavespendsomuchmoneyandvigoronsearchingforstripssurface’sinspectingandcontrollingmethods.Toimprovethestripssurfacequality,thefirstquestiontodealistoinspectandclassifythedefects,thentheanalysisofthedefectscause,andproposethemethodofeliminatethedefects.Theon-linesurfacedefectsinspectionandclassificationcandedicatenotonlytothetheorybutalsotothepracticalvalues.Firstly,aprimaryinspectionmethodusedthevarianceofthegreyvalueorgreyabsolutevalue,isputforwardhere,followingthefeatureofgreyhistogramisproposed.Ithasbeenprovedpracticalindistinguishingthedefectexistornot.Thencombiningwavelettransformmethodandgraylevelco-occurrencematrix,afeatureselectionmethodcalledsequentialforwardselection(SFS)wasputforward,whichcanimprovetheclassificationgreatly.Basedonthetheoryabove,takingtheadvantagesofthefuzzypatternrecognitioninrejectingnoiseandoftheBPrecognitioninmodalfittingandnon-linearrecognition,anewFuzzy-BPnetworkrecognitionsystemisproposed.AndtheexperimenthasproveditismoreeffectivethanfuzzypatternrecognitionorBPrecognition.Inthispaper,wejustusetheabovetheoryandmethodinthosesixtypedefects(edgesawtooth,weldingseam,mixedmaterial,wrinkles,theholesandtherollmarks),AbstractIVwhichhasabetterrecognitionresultbothinback-recognitionandrecognitioninnewones.Allresearchworksarebasedonanindependentlydevelopmentsoftwareenvironmentofclassificationandrecognitionofrolledstripdefects.Itasestablishedsolidfoundationforthefurtheron-lineinspectionandclassificationresearchofrolledstripdefects.Keyword:rolledstrips,greyhistogram,wavelettransform,graylevelco-occurrencematrix,featureextraction,featureselection,sequentialforwardselection(SFS),fuzzy-Bppatternrecognition*Thestudyisfundedby:ChinaMajorFormerBasicResearchProjectofTheMinistryofScienceandTechnologyofthePeople'sRepublic(No.2003CCA03900)-1-1.1-2-1.2[1](50m/min)100%,2060CCD()[2]ASIC()1.2.1702070,[3]70,,,()Honeywell1982,CCD[4]WestinghouseCCD0.7mm2.3mmCentroSviluppoMateriali[4],Sick-3-()30[4]Fractal,,,X,90,()RautaruukkiNewTechnology[4]SmartvisCognex1994iS-2000iLearn80GOPSParsytec[4]1997HTS-2,(ANN)EuropeanElectronicSystem(EES)1.2.21994CCFPCDSP[56][47],-4-[7]90%[29]LVQ[10]1.3(AGC)(AWC)(ASC)1.2.3.4.90VLSI(DSP)1.4-5-15%2[6]12343(ART)ARTSOM[11]BP-6--7-[12]123128*128128*128[12]2.12.1.1-8-M*N[13][14]A/DConverter.CCDCCDCCDRS170/CCIRPAL/NTSCS-VideoUSBMatroxMeteroMrtroxGenesisDSPNOAAV8MpegMpeg-1VCDEPSONSMARTPANEL256400*400-9-2.1.212.12.2Fig.2.1EdgeSawtoothFig.2.2WeldingSeam2.32.4-10-Fig.2.3TheWrinklesFig.2.4MixedMaterial2.52.6Fig.2.5TheHolesFig.2.6RollMark2.2[12]Fourier-2.2.1256255,......3,2,1,0=i,i)(iNNi-11-NiNip/)()(=)(ip[12])(ip256256∑−=×=10)(Liipiµ2.1∑−=×−=1022)()(Liipiµσ2.2∑−=×−=1033)()(1LiipiSµσ2.3∑−=×−=1044)()(1LiipiKµσ2.4∑−==102))((Liipenergy2.5)(log)(10ipipentropyLi∑−=×−=2.6LL=256)(ipiNiNip/)()(=-12-2.2.2J.Morlet1984[16-18]80AT&TFourier2.2.2.1()0=∫dxxRψψba,ψ()0,,2/1,≠∈−=−aRbaabxaxbaψψ2.7()tf()()0,,2/1≠−=∫adttfabtabaWfψ2.8ψmotherwavelet()()()dttftfcnmnm∫∞∞−=,,ψ2.9()()0,10,00002/0,≠−=−−banbtaatmmnmψψ2.10Fourierψ-13-()()0,122==∫∫dtttdttRRψψ2.11()tψ()0=∫dttRψψ()tabtaba,2/1ψψ=−−()tba,ψa/1abψψFourierZooming()0=∫dttRψ()ωψˆ0=ωψ),(±ψψωtψtψ()ωωψωωωψd20ˆ∫∞±±=2.12()2/122ˆˆ=∆∫±ωωψωψd2.13ψ0,0ω±ψ∆±∆ψˆba,ψ),(0abω±ψ∆a±∆ψˆ1aa12.16-14-()()=∫∫∞±±−Rnmnmmmddtttaanbωωωψωψω02,2,0000ˆ,),(2.14m100at2.2.2.2[19,20])(2RLZjjV∈}{2)(2RLjjjVVV2222×=)()(),(yxyxϕϕϕ=2.15)(xϕ)(yϕZjjV∈}{222jVZnmZjmynxjjj∈∈∀−−−−−,,).2,2(2ϕ2.1622jO22jV221+jV2222221+=+jjjVOV2.17)()(),(yxyxLHψϕ=Ψ,)()(),(yxyxHLϕψ=Ψ,2.18)()(),(yxyxHHψψ=Ψ.},,{HHHLLH=Ω)2,2(2),2,2(2),2,2(2mynxmynxmynxjjHHjjjHLjjjLHj−−−−−−−−−−−Ψ−−Ψ−−Ψ2.1922jO)2,2(2mynxjjj−−−−−ϕ)(2RL-15-2.8Fig.2.8TheSketchMapof2DWaveletRegenerationh~2h~2g~22h~2g~2g~fDjHLfAjfDjHHfDjLHAj+1ffDj2fDfDfDjHHjHLjLH,,12+jj22.712+jfAj1+j2fAjfDfDfDjHHjHLjLH,,2.82.72.8fAjh~g~2.7Fig.2.7TheSketchMapof2DWaveletDecompositionh2g2h2g2h2g2fDjLHAj+1ffAjfDjHLfDjHH-16-gghh==~,~)(xϕ)(xψfAjfAj1+fDjLHfAj1+fDjHLfAj1+fDjHHfAj1+NN22×NN22×1L1L∑==MnmnmxMe1,2),(12.20MMnmxEnergymn×=∑∑),(22.21MMnmxnmxEntropymn×=∑∑)),(log(),(222.22MM×mnxK3*K+12.2.322Haralick1973[12]-17-2.2.3.1iDxDyj2()()()(){}1-N,0,1,2,;,,,|,,Λ==++==ΡyxjDyyDxxfiyxfyxj,i,θδ(3.23)1-,,2,1,0,LjiΛ=xyL2xyijδ()DyDx,2.9x+Dx,yx,y+Dyx+Dx,y+DyDx1Dy00oDx=1,Dy=-1,45oDx=0,Dy=1,90oDx=1,Dy=-1,135o()[]θδ,,,jipijδij0o45o,90o,135o4[21]256xoyDxDyij2.9Fig.2.9Pixe