基于注意力机制和ACT网络的人脸图像修复

打开文本图片集

Face image inpainting based on attention mechanism and ACT network

TENGLin1，²，ZHANGQian1²，XU Kaili² （1.SchoolofData Scienceand InformationEngineering，Guizhou Minzu University，Guiyang 55oo25，China; 2.KeyLaboratoryofPatternRecognitionand InteligentSystemsofGuizhouProvince，Guiyang55O025，China）

Abstract：A faceimage inpainting method basedonconvolutional block atentionmodule（CBAM）andaggregated contextualtransformation（ACT）network isproposed tomakethecompletionofmissingsemanticfeaturesinfacialimages more realisticandtoenhancetherecoveryofdetailedinformation.Inthismethod，thetwobranchesofthebaselinemodelareretained. Inthesemanticandimagfilteringbranches，theCBAMlayersareaddedtocapturethecriticaldetailinformationforfilingthe mising areas intheimages.Thebaseline residual blocksarereplaced withACTresidual blocks，which can preserve therich detailsoutsidethemissingareasandcaptureabundantcontextualinformation.Thismadethesemanticinformationfilinginthis branchmoreaccurate，efectivelyremovedartifactsandenrichedimagedetails.Inthekernelpredictionbranch，thetwomodules areadded toenhancethereceptivefieldandcontextualreasoning perception when extracting image features，makingthedynamic predictionof filtering kernelsmore precise.Thismethod wasvalidatedontheCelebA-HQdataset，showing improvements in quantitative metrics such as PSNR （peak signal-to-noise ratio），SSIM （structural similarity index measure）and L1 . The qualitative repairresultsarealsoclearerandmorenatural.Thestudyconfirmsthattheproposed methodhavegoodeffectivenessforfacial image inpainting.

Keywords：image inpainting; CBAMatentionmechanism;ACTnetwork;encoder-decoder;facial image inpainting;mage filtering

0 引言

图像修复是指重建图像中缺失区域的任务。（剩余10343字）