第29卷第1期2024年1月doi:10.13682/j.issn.2095-6533.2024.01.010西安邮电大学学报JOURNALOFXI'ANUNIVERSITYOFPOSTSANDTELECOMMUNICATIONS基于混合反向注意力机制的息肉分割网络Vol.29No.1Jan.2024兰蓉，孙宇浩，赵凤，郭迪（西安邮电大学通信与信息工程学院，陕西西安710121)摘要：针对肠息肉图像分割结果存在伪影、边界模糊及精度低等问题，提出一种基于混合反向注意力机制的息肉分割网络。该网络使用U型网络结构，设计多尺度并行空洞卷积注意力模块，以更加细粒度的多尺度特征减少下采样细节的损失。采用密集连接和特征融合的方式设计跨阶段局部模块，减少上下文之间的语义差异，补充细节特征。利用位置注意力和通道注意力同反向注意力相结合策略，构建区域与边界关系的同时学习位置和通道的特征，进而清晰分割出息肉与正常粘膜。实验结果表明，该网络提高了分割精度，消除了边界外部的部分伪影，在一定程度上改善了边界模糊的问题。关键词：图像分割；息肉分割；注意力；空洞卷积；U-Net中图分类号：TN911.73；R735.34PolypsegmentationnetworkbasedonhybridreverseLANRong,SUNYuhao,ZHAOFeng,GUODi(SchoolofCommunicationsandInformationEngineering,Xi'anUniversityofPostsandTelecommunications,Abstract:Toaddressthechallengesofartifacts,blurredboundaries,andlowaccuracyintheseg-mentationresultsofpolypimages,apolypsegmentationnetworkbasedonhybridreverseattentionmechanismisproposed.ThenetworkincorporatesaU-shapedstructureandintroducesamulti-scaleparalleldilatedconvolutionattentionmodule.Thismodulehelpspreservefiner-grainedmulti-scalefeaturesduringdownsampling,reducingthelossofimportantdetails.Additionally,densecon-nectivityandfeaturefusionareemployedtocrossstagepartialmoduletobridgesemanticdiffer-encesbetweencontextsandenhancedetailedfeatures.Furthermore,acombinationofpositionalat-tentionandchannelattention,integratedwiththeinverseattentionstrategy,isemployedtolearnlo-cationandchannelfeatureswhileestablishingtheregion-boundaryrelationshipforaccuratepolypandnormalmucosasegmentation.Experimentalresultsdemonstratethatthepolypsegmentationnetworkbasedonhybridreverseattentionmechanismimprovesthesegmentationaccuracy,reducesartifactsoutsidetheboundary,andmiti...
