资源描述：
《1.1 Image Classification》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。
1、CS231nConvolutionalNeuralNetworksforVisualRecognitionThisisanintroductorylecturedesignedtointroducepeoplefromoutsideofComputerVisiontotheImageClassificationproblem,andthedata-drivenapproach.TheTableofContents:•IntrotoImageClassification,data-drivenapproach,pipeline•Near
2、estNeighborClassifier◦k-NearestNeighbor•Validationsets,Cross-validation,hyperparametertuning•Pros/ConsofNearestNeighbor•Summary•Summary:ApplyingkNNinpractice•FurtherReadingImageClassificationMotivation.InthissectionwewillintroducetheImageClassificationproblem,whichisthe
3、taskofassigninganinputimageonelabelfromafixedsetofcategories.ThisisoneofthecoreproblemsinComputerVisionthat,despiteitssimplicity,hasalargevarietyofpracticalapplications.Moreover,aswewillseelaterinthecourse,manyotherseeminglydistinctComputerVisiontasks(suchasobjectdetect
4、ion,segmentation)canbereducedtoimageclassification.Example.Forexample,intheimagebelowanimageclassificationmodeltakesasingleimageandassignsprobabilitiesto4labels,{cat,dog,hat,mug}.Asshownintheimage,keepinmindthattoacomputeranimageisrepresentedasonelarge3-dimensionalarray
5、ofnumbers.Inthisexample,thecatimageis248pixelswide,400pixelstall,andhasthreecolorchannelsRed,Green,Blue(orRGBforshort).Therefore,theimageconsistsof248x400x3numbers,oratotalof297,600numbers.Eachnumberisanintegerthatrangesfrom0(black)to255(white).Ourtaskistoturnthisquarte
6、rofamillionnumbersintoasinglelabel,suchas"cat".ThetaskinImageClassificationistopredictasinglelabel(oradistributionoverlabelsasshownheretoindicateourconfidence)foragivenimage.Imagesare3-dimensionalarraysofintegersfrom0to255,ofsizeWidthxHeightx3.The3representsthethreecolo
7、rchannelsRed,Green,Blue.Challenges.Sincethistaskofrecognizingavisualconcept(e.g.cat)isrelativelytrivialforahumantoperform,itisworthconsideringthechallengesinvolvedfromtheperspectiveofaComputerVisionalgorithm.Aswepresent(aninexhaustive)listofchallengesbelow,keepinmindthe
8、rawrepresentationofimagesasa3-Darrayofbrightnessvalues:•Viewpointvariation.Asingleinstanceofanobjectcanbeorien
