葛仁彥-機器視覺的自適應算法在工業(yè)自動化檢測中的應用

AutomaticSLInspectionEquipment1Nanovision?Contents目錄1Background背景資料2AdaptiveAlgorithmIntroduction自適應算法簡介3AdaptiveFilter自適應濾波4AdaptivePatternRestoration自適應模式恢復5AdaptivePatternMatching自適應模式匹配6TypicalApplicationCase典型應用案例7FutureWork將來的工作Nanovision?1.Background背景資料IndustrialvisionsystemsrequireexpertiseinmultiplefieldsastheyarecompositionsofdifferenttechnologiesArchitecturesSoftware?PC+CameraSensors?EmbeddedVision?Identification?SmartCameras?CMOS?Measurement?PC+SmartCamera?CCD?Guidance?CMOS/CCD?Other?OtherOptics?LensesMachineVisionDataCapture?Prisms&BeamSplittersRobiticGuidance?Analog?ColorFiltersFactoryAutomation?Digital?PolarizersProcessorsSpectrum/Illumination?FPGAInterfaces?Visible?LED(FieldProgrammableGatrArray)?GigabitEthernet(GigE)?DSP?Infrared(IR)?Halogen?Firewire(1394b)(DigitalSignalProcessor)?Ultraviolet(UV)?Other?CameraLink?RISC?USB2.0-3.0(ReconfigurableInstructionSetComputer)?Other?FrameGrabbersSource：Frost&Sullivan，2010Nanovision?Background背景資料CurrentAlgorithmsmethodforIndustrialParts/SurfaceInspection目前在工業(yè)部件和表面檢測中的常用算法CADBasedPatternMatching與CAD設計圖案匹配GoldenTemplateMatchingDesignRuleCheckDRC金色模板匹配設計規(guī)則檢查ArtificialNeuralNetworkFeaturebasedCompare人工神經網路基于特征的比較Nanovision?Background背景資料目前算法存在的問題CADBasedPatternMatching（與CAD設計圖案匹配）CAD模板圖像化過程比較繁瑣CAD模板與實際采集圖像很難配準，影像采集根據應用不同有不同程度的變形GoldenTemplateMatching（金色模板匹配）需要很多良品制作金色模板，沒有統(tǒng)一性受影像采集過程中運動等影響，模板匹配會有局部偏差DesignRuleCheck–DRC（設計規(guī)則檢查）DRC需要知道被檢測物品的設計規(guī)則，需要經驗值DRC只能檢測非常微小的局部缺陷，有局限性ArtificialNeuralNetwork（人工神經網絡）學習矩陣需要大量的樣本計算量比較大，計算收斂慢，對在線實時檢測需要FPGA或DSP的加速計算FeaturebasedCompare（基于特征的比較）特征點的提取計算量比較大被檢測物品的特征很難量化，有的物品特征不是很明顯Nanovision?AdaptiveAlgorithmIntroduction自適應算法簡介AdaptiveBackgroundnormalization自適應背景歸一化AdaptiveFilter自適應濾波

AdaptiveAlgorithmsMachineVision

AdaptivePatternRestoration自適應模式恢復AdaptivePatternMatching自適應模式匹配Nanovision?AdaptiveFilter自適應濾波自適應濾波存在于信號處理、控制、圖像處理等許多不同領域，它是一種智能更有針對性的濾波方法，通常用于去噪。自適應濾波是近年以來發(fā)展起來的一種最佳濾波方法。它是在維納濾波（wienerfiltering),Kalman濾波等線性濾波基礎上發(fā)展起來的一種最佳濾波方法。由于它具有更強的適應性和更優(yōu)的濾波性能。算法包括:LMS自適應濾波算法RLS自適應濾波算法變換域自適應濾波算法仿射投影算法共扼梯度算法基于子帶分解的自適應濾波算法基于QR分解的自適應濾波算法在圖像處理中的應用它是指在一張圖像的不同區(qū)域具有各自的圖像特性時，分別針對這些不同特性選取最優(yōu)的、各不相同的參數，濾波器或濾波方法進行濾波。Nanovision?AdaptivePatternRestoration自適應模式恢復自適應模式恢復的主要技術要點：AutomaticNoiseRemove自動噪聲去除AutomaticBinarization（multi-levelsegmentation）自動二值化或多值分割FindtheregulationofthepatternbasedonFrequencydomainandSpacedomain根據頻率域和空間域的信息發(fā)現模式的規(guī)律性Patternrestorationbasedonregularfrequencyandtones根據影像的頻率和色調恢復重建具有規(guī)律的影像Nanovision?AdaptivePatternMatching自適應模式匹配DigitalCameras

AdaptiveAdaptiveTemplateFilterBufferRegistrationActiveAdaptivePatternPipelineRegistrationRestorationObject

DynamicAdaptiveAutomaticDefectPatternDefectEnhancementMatchingClassificationHostClassifiedDefectsNanovision?TypicalApplicationCase典型應用案例TypicalApplicationCase典型應用案例：前端市場：AutomaticLCD/FTFInspectionSystem(LCD/TFTAOI)AutomaticScreenInspectionSystem(ScreenAOI)AutomaticshadowmaskInspectionSystem(ShadowMaskAOI)后端市場：導光板質量檢測手機玻璃蓋板質量檢測手機模組質量檢測工業(yè)部件檢測：電阻應變片檢測被動元器件材料檢測Nanovision?TypicalApplicationCase典型應用案例AutomaticLCD/FTFInspectionSystem(LCD/TFTAOI)Nanovision?TypicalApplicationCase典型應用案例AutomaticScreenInspectionSystem(ScreenAOI)Nanovision?TypicalApplicationCase典型應用案例AutomaticshadowmaskInspectionSystem(ShadowMaskAOI)Nanovision?TypicalApplicationCase典型應用案例導光板質量檢測Nanovision?TypicalApplicationCase典型應用案例手機玻璃蓋板質量檢測：Nanovision?TypicalApplicationCase典型應用案例手機模組質量檢測（A）flexible 電路板ControlPWBSource信號GateGate信號PWBASICConnecter

SourcePWB間隔橡膠液晶面板 注入口封口處Nanovision?TypicalApplicationCase典型應用案例手機模組質量檢測（B）MuraInspectionisthekeyTwodimensionWaveletTransform,f(x,y)L2(R),cwt(s,a,b)1f(x,y)(xa,yb)dxdy(timedomain)sssCWT (s,w1,w2)sF(w1,w2)(sw1,sw2) (frequencydomain) Nanovision?TypicalApplicationCase典型應用案例電阻應變片檢測Nanovision?TypicalApplicationCase典型應用案例被動元器件材料檢測斷線 缺損圖案缺損 網版劃傷Nanovision?FutureWork將來的工作自適應算法將來要解決的主要問題：對規(guī)律性不強的影像處理影像檢測缺陷的自動分類問題現有算法在SmartCamera或FPGA/DSP上實現的可能性人工神經網計算的快速收斂和加速問題人工智能（AI）在模式識別和工業(yè)自動化檢測中的應用前景Nanovision?FutureWork將來的工作20092014201520202021&BeyondComponents

DynamicallyReconfigurableProcessorswilldisplaceFPGAsMulti-spectral(UVtoshortIR)sensorswillgainindustrialinterestandapplicationsAnaloguecamerasystemswillbereplacedwithdigitalcamerasTheneedforframegrabberswilldecreaseCMOSsensorswillgainadvantageoverCCDsduetolowerpowerconsumption,lowerpriceandhigherspeed

Old-to-newtechnologyreplacements(highersensorresolutionandcapturespeed)Embedded,industry-orientedsystemswillreplacePC-basedsystemsinthereplacementwaveMulti-spectralsensorswillgoclosertolongIRwavelengths(closeto8microns)andwillbelimitedbylensmaterialonmanufacturingInfluence

?Highaccuracymeasurementsin3Dandreal?Industrialvisionsystemswillbecometimewillbepossibleindispensablepartofflexibleautomationsystems?Machinevisionwillaidstatisticalanalysis,?Machinevisionsystemswillaidrapiddiagnosticsandmaintenanceinfactoriesmanufacturing?Remotecontrolandmonitoringofmachine?VisionsystemswillfindsafetyapplicationinvisionsystemswillbecommonfactoriesMachinevisionwillbeanintegralcontrolandguidancesolutionofrapidmanufacturingsystemsOtherRobotics

?Visionguidedrobotswilldecreasecycletime?Cycletimeofvisionguidedrobotsshallfall?Autonomousrobotswillusevisionsystemsbelow1secondenablingflexibleandfastrobotforhumanandobstacledetection,objectoperationtracking,navigationandlocalization?Visualservoingwillenablerobotstoaidflexible?Robotswillbereadytointelligentlypickprecision