基于特征融合与软阈值残差的稠密点云几何压缩网络  

Dense Point Cloud Geometry Compression Network Based on Multiscale Feature Fusion and Soft Threshold Residual Structure

作　　者：朱威[1,2] 施海东 汪宵 郑雅羽 何德峰[1,2] ZHU Wei;SHI Haidong;WANG Xiao;ZHENG Yayu;HE Defeng(College of Information Engineering,Zhejiang University of Technology,Hangzhou 310023,China;United Key Laboratory of Embedded System of Zhejiang Province,Hangzhou 310023,China)

机构地区：[1]浙江工业大学信息工程学院,杭州310023 [2]浙江省嵌入式系统联合重点实验室,杭州310023

出　　处：《小型微型计算机系统》2025年第3期662-671,共10页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(62173303)资助;浙江省自然科学基金项目(LY21F010009)资助.

摘　　要：点云是一种重要的三维数据表示形式,但其巨大的原始数据量阻碍了它在网络传输和存储记录等方面的应用.因此,本文提出了一种基于多尺度特征融合与软阈值残差结构的点云几何压缩网络,实现了对三维稠密点云的高效压缩.首先通过逐步融合多尺度特征和构建软阈值注意力机制,实现特征加强和冗杂特征的消除,以解决体素化过程中特征丢失等问题.此外,采用构建特征掩膜层的方法,加速模型收敛.最后,引入动态非等比例损失函数提高网络的学习效果.实验结果表明,该方法在MVUB、8iVFB和Owlii数据集上相较于现有方法同样的点云分辨率下,具有更高的点云重建质量和较快的编解码速度.Point cloud is an important form of three-dimensional data representation,but its huge amount of raw data hinders its application in network transmission and storage recording.Therefore,this article proposes a point cloud geometric compression network based on multi-scale feature fusion and soft threshold residual structure,which achieves efficient compression of three-dimensional dense point clouds.Firstly,by gradually integrating multi-scale features and constructing a soft threshold attention mechanism,feature enhancement and elimination of redundant features are achieved to address issues such as feature loss during voxelization.In addition,the method of constructing feature mask layers is adopted to accelerate the convergence of the model.Finally,introducing a dynamic non proportional loss function improves the learning performance of the network.The experimental results show that this method has higher point cloud reconstruction quality and faster encoding and decoding speed compared to existing methods at the same point cloud resolution on MVUB,8iVFB,and Owlii datasets.

关 键 词：稠密点云压缩 多尺度特征 软阈值残差结构 特征掩膜 动态损失函数 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]