[CAIP 2009] Face Detection Using GPU-Based Convolutional Neural Networks

《[CAIP 2009] Face Detection Using GPU-Based Convolutional Neural Networks》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、FaceDetectionUsingGPU-BasedConvolutionalNeuralNetworksFabianNasse1,ChristianThurau2,andGernotA.Fink11TUDortmundUniversity,DepartmentofComputerScience,Dortmund,Germany2FraunhoferIAIS,SanktAugustin,GermanyAbstract.Inthispaper,weconsidertheproblemoffacedetect

2、ionun-derposevariations.Unlikeothercontributions,afocusofthisworkre-sideswithinefficientimplementationutilizingthecomputationalpowersofmoderngraphicscards.Theproposedsystemconsistsofaparallelizedimplementationofconvolutionalneuralnetworks(CNNs)withaspecialem

3、phasizeonalsoparallelizingthedetectionprocess.Experimentalvali-dationinasmartconferenceroomwith4activeceiling-mountedcamerasshowsadramaticspeed-gainunderreal-lifeconditions.1IntroductionThepastyearsyieldedincreasinginterestintransferingcostlycomputationsto

4、GraphicsProcessingUnits(GPUs).Duetoparallelexecutionofcommandsthisoftenresultsinamassivespeedup.However,italsorequiresacarefulladaptionandparallelizationofthealgorithmtobeimplemented.Asnotedin[1],convo-lutionalneuralnetworks(CNNs)[8,2,4,3]offerstateoftheart

5、recognizersforavarietyofproblems.However,theycanbedifficulttoimplementandcanbeslowerthanotherclassifiers,e.g.traditionalmulti-layerperceptrons.ThefocusofthispaperresideswithintheimplementationdetailsofparallelizingCNNsforthetaskoffacedetectionandposeestimatio

6、nandevaluatingitsrun-timeperformance.Incontrastto[1],whereGPUoptimizedCNNswereadaptedfordocumentprocessing,theconsideredfacedetectiontaskrequiresadditionalcon-siderations:Indetail,thecontributionsofthispaperare(a)extensionofthefacerecognition(andposeestima

7、tion)systemin[3]byparallelizingimportantpartsofthecomputationalprocessandimplementingitonagraphicscard,and(b)furtherenhancingthesystembyanoptimizeddetector.Experimentalvalidationtakesplaceinamulti-cameraenvironmentandshowsaccuratedetectionandhighperformanc

8、eunderreal-lifeconditions.Theremainderofthispaperisorganizedasfollows:InSection2wegiveabriefintroductiontoconvolutionalnetworksandhowtheyareusedinthiswork.Section3explainsindetailtheprocessoffacedetectionusin