融合卷积神经网络和Transformer的人脸欺骗检测模型.pdfVIP

学术论文

DOl:10.12379/j.issn.2096-1057.2024.01.05ResearchPapers

融合卷积神经网络和Transformer的

人脸欺骗检测模型

黄灵何希平1.21旷奇弦

贺丹杨楚天

400067)

1（重庆工商大学人工智能学院重庆

2（检测控制集成系统重庆市工程实验室（重庆工商大学）重庆400067）

(2207346808@)

FaceSpoofingDetectionModelwithFusionofConvolutional

NeuralNetworkandTransformer

HuangLing,HeXipingl-2,HeDan,YangChutian,andKuangQixian

I(SchoolofArtificialIntelligence,ChongqingTechnologyandBusinessUniversity,Chongqing400067)

2(ChongqingEngineeringLaboratoryforDetection,ControlandIntegratedSystem(ChongqingTechnologyand

BusinessUniversity),Chongqing400067)

AbstractInthefieldoffaceanti-spoofing,themethodsbasedonConvolutionalNeuralNetwork

(CNN)canlearnfeaturerepresentationwithfewerparameters,yettheirreceptivefieldsremain

local.Incontrast,Transformer-basedmethodsofferglobalperceptionbutentailanimpractical

volumeofparametersandcomputations,hinderingwidespreaddeploymentonmobileoredge

devices.Toaddressthesechallenges,thispaperproposedafacespoofingdetectionmodelthat

integratesCNNandTransformer,aimingtoachieveabalancebetweentheamountofparameters

andaccuracywhilemaintainingtheabilitytoextractglobalandlocalfeatures.Firstly,localface

imagesarecroppedandselectedasinputtoeffectivelyavoidoverfitting.Secondly,thefeature

extractionmodulebasedoncoordinateattentionisdesigned.Finally,thefusionofCNNand

Transformermodulesaredesignedtoextractlocalandglobalfeaturesofimagesthroughlocal-

global-localinformationexchange.Theexperimentalresultsshowthatthemodelachievedan

accuracyof99.31%andanaverageerrorrateof0.54%ontheCASIA-SURF(Depthmodality)

dataset;Mo