基于可逆神经网络的点云几何有损编码

打开文本图片集

中图分类号：TP391.4 文献标志码：A文章编号：1003-3114（2025）06-1351-08

Abstract：Withtherapiddevelopmentof wirelesssensing technologyandlaserdigitalacquisitiontechnology，3Dpointclouddata hasbeenwidelyusedinmultiplefields.However，thelargescaleandhighredundancyofpointclouddatamakegreatchalengestoits application，andthidustryurgentlydsefiientpointcloudgeometriclossycoingagorits.Traditioalpointcloudgeoetrc losyncodingalgorithmshaveloweficiencyandpoorencodingperformance，hiledep-learning-basedpointcloudgeometricncoding algorithmsmostlyuseAutoEncoder（AE）neuralnetworkarcitectures，hichsuferfromacertaindegreeoffeatureinfomationloss.In addition，inecentyars，ostsudieshavefousedonimproingtetropycodingileglectingtetizatioofttasfor mationbetweenpointcloudgeometryspaceanditspotentialfeaturespace.Inresponsetothaboveissues，thispaperproosesalossy pointcloudgeometryencodingalgorithmbasedonInvertibleNeuralNetwork（INN）.ThisalgorithusesanINNwithmathematicalyrig orousreversiblepropertiesforfeatureextractionofpointcloudgeometryinformation，avoidinginformationlossduringthencoding processandensuringthestabilityofreconstructedpointcloudsduring thedecodingproce.Thispaperdesignsa3D-Dense-Blockmoduleandchaelsueeeouletoanefatureifoatioeaseoearxpreionbilityoftealgritok， andavoidtheoueeofsuboptialsutiosdurintrngExperimentalsultssothatealgriceesbeteatedisto tionperformancethanthe Moving Picture Experts Group（MPEG）benchmarkalgorithmonMicrosoft Voxelized Upper Bodies（MVUB） and MPEG 8i datasets.

Keywords：wireless sensing technology; point cloud data; lossy point cloud encoding; INN

0引言

3D点云数据是一种用于表示3D空间中物体形状和结构的数据形式。（剩余10505字）