基于特征调制的轻量化MVSNet研究

打开文本图片集

中图分类号：TP183 文献标识码：A文章编号：1006-8228（2025）12-83-06

Research on LightweightMVSNetBased on Feature Modulation

Liu Zhen，Cheng Yang，Chen Haoxian，Liu Junbiao，Chen Ren，Huang Yanping （SchoolofPhysicsandOptoelectronicEngineering，FoshanUniversityFoshan，Guangdong528225，China）

Abstract：Multi-viewStereo（MVS）isacrucialtechniquefor3Dscenereconstruction.Amongvariousapproaches，deepleaing basedmult-view3DreconstructionhasbeenextensivelystudiedduetoitspromisingperformanceHoweveritspractical applicationisoftenconstrainedbythememorybotteneck，makingitdificulttobaancereconstructioneffciencyandauracy. Toadressthisissue，thispaperproposesalightweightnetworknamed SSC-MVSNetformulti-viewdense3Dreconstruction. ThenetworkemploysSpatiall-AdaptiveFeatureModulation（SAFM）toenhancemulti-scalefeaturerepresentation，introducesa SemanticDifferenceModule（SDM）tostrengthenboundaryresponse，andutilizesCrossNormandSelfNormtooptimizedepth propagation，therebyconvertingsparsedepthmapsintodenseresultswithlowcomputationaloverhead.Experimentsdemonstrate that on the DTU dataset，the overall quality（Overall） and completeness（Comp） reach 0.359mm and 0.313mm ，respectively. Compared to Fast-MVSNet，the completenessmetricisimproved by approximately 22.4% ，while thememory consumption increases byonly 5% .Thispaper provides a high-performance solution for lightweight 3D reconstruction.

Keywords：Multi-View;3DReconstruction;MVSNet;Lightweight

0引言

多视图立体视觉（Multi-ViewStereo，MVS）作为多视图三维重建的核心技术，旨在从已知相机参数的多视角图像中恢复密集三维场景，其作为计算机视觉领域的一个重要问题而被广泛研究。（剩余8839字）