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8.9 dB,2.4 dB,1.8 dB以上。
Current deep learning-based point cloud compression algorithms suffer from insufficient local feature learning, and the huge data volume of point clouds also limits the network size. In order to guarantee the reconstruction quality while reasonably controlling the computational complexity, a sparse convolution based asymmetric feature aggregation point cloud compression algorithm is proposed,and asymmetric feature aggregation codec network and channel-wise sparse residual convolution are designed to improve the rate distortion performance. It is experimentally verified that compared with the existing G-PCC, V-PCC and Learned-PCGC, the BD-Rate is reduced by more than 88.7%, 46.7% and 40.9%, respectively, and the BD-PSNR is increased by more than 8.99 dB, 2.42 dB and 2.36 dB, respectively.
作者
黄炜
朱映韬
陈冬杰
王宝土
陈建
HUANG Wei;ZHU Yingtao;CHEN Dongjie;WANG Baotu;CHEN Jian(College of Advanced Manufacturing,Fuzhou University,Quanzhou 362200,China;College of Physics and Information Engineering,Fuzhou University,Fuzhou 350108,China)
出处
《电视技术》
2022年第12期67-71,76,共6页
Video Engineering
基金
国家自然科学基金(62001117)
国家级创新训练项目(202210386028)。
关键词
点云压缩
自编码器
稀疏卷积
非对称特征聚合
point cloud compression
auto-encoder
sparse convolution
asymmetric feature aggregation
