《电视技术》第46卷第12期(总第565期)67PARTS&DESIGN器件与设计文献引用格式：黄炜，朱映韬，陈冬杰，等.基于稀疏卷积的非对称特征聚合点云压缩算法[J].电视技术，2022，46（12）：67-71，76.HUANGW，ZHUYT，CHENDJ，etal.Sparseconvolutionbasedasymmetricfeatureaggregationalgorithmforpointcloudcompression[J].VideoEngineering，2022，46（12）：67-71，76.中图分类号：TP311.1文献标识码：ADOI：10.16280/j.videoe.2022.12.015基于稀疏卷积的非对称特征聚合点云压缩算法黄炜1，朱映韬1，陈冬杰2，王宝土2*，陈建1，2（1.福州大学先进制造学院，福建泉州362200；2.福州大学物理与信息工程学院，福建福州350108）摘要：当前基于深度学习的点云压缩算法存在局部特征学习不足的问题，点云庞大的数据量也限制了网络规模。为了保障重建质量的同时合理控制计算复杂度，提出一种基于稀疏卷积的非对称特征聚合点云压缩算法，设计非对称特征聚合编解码网络、逐通道稀疏残差卷积提升率失真性能。经实验验证，相较于现有的G-PCC、V-PCC和Learned-PCGC算法，所提算法的BD-Rate分别减少88%，46%，40%以上，BD-PSNR分别增加8.9dB，2.4dB，1.8dB以上。关键词：点云压缩；自编码器；稀疏卷积；非对称特征聚合SparseConvolutionBasedAsymmetricFeatureAggregationAlgorithmforPointCloudCompressionHUANGWei1,ZHUYingtao1,CHENDongjie2,WANGBaotu2*,CHENJian1,2(1.CollegeofAdvancedManufacturing,FuzhouUniversity,Quanzhou362200,China;2.CollegeofPhysicsandInformationEngineering,FuzhouUniversity,Fuzhou350108,China)Abstract:Currentdeeplearning-basedpointcloudcompressionalgorithmssufferfrominsufficientlocalfeaturelearning,andthehugedatavolumeofpointcloudsalsolimitsthenetworksize.Inordertoguaranteethereconstructionqualitywhilereasonablycontrollingthecomputationalcomplexity,asparseconvolutionbasedasymmetricfeatureaggregationpointcloudcompressionalgorithmisproposed,andasymmetricfeatureaggregationcodecnetworkandchannel-wisesparseresidualconvolutionaredesignedtoimprovetheratedistortionperformance.ItisexperimentallyverifiedthatcomparedwiththeexistingG-PCC,V-PCCandLearned-PCGC,theBD-Rateisreducedbymorethan88.7%,46.7%and40.9%,respectively,andtheBD-PSNRisincreasedbymorethan8.99dB,2.42dBand2.36dB,...
