点云超分辨率的模型搭建与特征补充  

Model Construction and Feature Adjunction in Point Cloud Super-resolution

作　　者：陆敬奔 李春国[1] 杨绿溪[1] LU Jingben;LI Chunguo;YANG Lüxi(School of Information Science and Engineering,Southeast University,Nanjing 211189,China)

机构地区：[1]东南大学信息科学与工程学院,江苏南京211189

出　　处：《无线电通信技术》2022年第6期1111-1116,共6页Radio Communications Technology

基　　金：国家自然科学基金(62171119);江苏省重点研发计划(BE2021013-3)。

摘　　要：点云数据在工程中有着广泛应用,由于设备精度不足、环境存在噪声等原因,实际场景获取的点云数据部分区域点数较少,会造成一定的信息丢失。在尽可能保留三维物体表面特征的前提下,点云超分辨率技术能对点云进行上采样,生成点数更多的高分辨率点云。基于MPU神经网络模型,重新建立了点云超分辨率的前向传播流程,包含特征提取、特征扩充、坐标重建三个步骤,通过在提取特征时添加特征补充模块,进一步提升了深层特征的表征能力。实验表明,所设计的模型与主流网络相比能取得较高的性能指标,添加的特征补充信息对模型超分辨率效果的提升有着积极作用。Point cloud data is widely used in engineering project.Affected by equipment precision and environmental noise,point cloud data obtained suffer partial point missing problems,causing information loss of the point cloud.With original 3D structure preserved,super-resolution technology generates high resolution point cloud through point up-sampling operation.On the basis of MPU model,we reconstruct the forward propagation of point super-resolution task,i.e.feature abstracting,feature expanding and coordinate reconstruction.By adding adjunctive features,the representational ability of deep point features is increased.Experiments show that the model we designed performs better than mainstream networks.The added adjunctive features have positive effects on the quality improvement of super-resolution outputs.

关 键 词：点云超分辨率 深度学习 神经网络 

分 类 号：TN919.23[电子电信—通信与信息系统]