|
|
自动化学报 2009
一种基于自然图像统计模型的图像合成检测方法DOI: 10.3724/SP.J.1004.2009.01564, PP. 1564-1567 Keywords: Imagecompositing,generalizedGaussianmodel(GGD),maximum-likelihood(ML),supportvectormachine(SVM),imageforensics Abstract: ？Nowadays,digitalimagescanbeeasilytamperedduetotheavailabilityofpowerfulimageprocessingsoftware.Asdigitalcamerascontinuetoreplacetheiranalogcounterparts,theimportanceofauthenticatingdigitalimages,identifyingtheirsources,anddetectingforgeriesisincreasing.Blindimageforensicsisusedtoanalyzeanimageinthecompleteabsenceofanydigitalwatermarkorsignature.Imagecompositingisthemostcommonformofdigitaltampering.Assumingthatimagecompositingoperationsaffecttheinherentstatisticsoftheimage,weproposeanimagecompositingdetectionmethodonbasedonastatisticalmodelfornaturalimageinthewavelettransformdomain.ThegeneralizedGaussianmodel(GGD)isemployedtodescribethemarginaldistributionofwaveletcoefficientsofimages,andtheparametersofGGDareobtainedusingmaximum-likelihoodestimator.ThestatisticalfeaturesincludeGGDparameters,predictionerror,mean,variance,skewness,andkurtosisateachwaveletdetailsubband.Then,thesefeaturevectorsareusedtodiscriminatebetweennaturalimagesandcompositeimagesusingsupportvectormachine(SVM).Toevaluatetheperformanceofourproposedmethod,wecarriedouttestsontheColumbiaUncompressedImageSplicingDetectionDatasetandanotheradvanceddataset,andachievedadetectionaccuracyof92%and79%,respectively.Thedetectionperformanceofourmethodisbetterthanthatofthemethodusingcameraresponsefunctiononthesamedataset.
|
