基于图的稀疏注意力实例分割算法

打开文本图片集

关键词：视觉图神经网络；稀疏矩阵；注意力机制；实例分割；消融实验；全局稀疏视觉图注意力中图分类号：TN911.73-34；TP391.41 文献标识码：A文章编号：1004-373X（2026）01-0086-09

Graphbasedsparseattentioninstancesegmentationalgorithm

LI Sen，WANG Xiaoming （School ofComputer and SoftwareEngineering，Xihua University，Chengdu 61Oo39，China）

Abstract：The vision graph neural network （vision GNN）provides anew paradigm forcomputer vision.However，the foundationalmodelViGfaceshighcomputationalcostsduetoitsrelianceontheK-nearestneighbor（KNN）algorithmforgraph construction.InthesubsequentlyproposedMobile ViG，sparsevisual graphatention （SVGA）isutilizedtoimproveefficiency， but itsstaticgaphconstruction limitstheacquisitionof gloalfeatures.Toaddresstheseisses，thispaperproposesaglobal sparsevisual graphatention（GSVGA）model.IntheGSVGAmodel，anewsparsematrix isconstructedbyconnectingpixel nodesacrossrows，columns，anddiagonalsoftheimage，soastoeliminatethe featurelossin SVGAandavoid the timeconsumingtensorreshapingoperations.Amax-relativegraphconvolution（MRConv）isintroduced，implementing feature aggregationbyrollngoperationsacrossleft，right，anddownwarddimensions.Thereceptivefieldisexpandedcombinedwith dilatedconvolution，soastoobtainmorefeatureinformationInadition，themoduleintegratestheupdatedfeed-forwardnetwork andselectstheGELUactivationfunctiontoimprovethestabilityoffeaturetransformationunderlarge-scalecalculation. ExperimentalresultsshowthatGSVGAperformswellonCOCO2017dataset，iSAIDdatasetandBDD10OKdataset.Intermsof reasoningeficiencythetimeconsumptionofGSVGAdecreasessignificantlywiththeincreaseof iteration.Itsaccuracyon ImageNet-1K is88.12%，itsmAP @ 0.5：0.95 of the segmentation of medium object instance is O.675 1，and its test set error is thesmalest inthecomparisonalgorithms.Thevisualizationresultsshow thattheGSVGAcansegment singleand multiple instances more accurately，and has strong generalization ability and small object detection ability.

Keywords：vision GNN; sparse matrix;atention mechanism; instance segmentation; ablation experiment; GSVGA

0 引言

近年来随着图神经网络（GNN）的兴起[，GNN这种基于图结构数据的深度学习模型也在计算机视觉中开始应用。（剩余12375字）