基于通道增益的可变比特率点云压缩  

Variable Rate Compression of Point Cloud Based on Channel Gain

作　　者：江照意[1] 邹文钦 宋超[1] 杨柏林[1] Jiang Zhaoyi;Zou Wenqin;Song Chao;Yang Bailin(School of Computer Science and Technology,Zhejiang Gongshang University,Hangzhou 310018)

机构地区：[1]浙江工商大学计算机科学与技术学院,杭州310018

出　　处：《计算机辅助设计与图形学学报》2024年第11期1816-1824,共9页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(62172366);浙江省自然科学基金(LY21F020013,LY22F020013)。

摘　　要：针对现有基于深度学习的点云压缩方法需要训练多个网络,耗费大量的时间和空间资源的缺陷,提出一种基于通道增益的可变比特率点云压缩方法.首先在网络的编码端利用层次化结构,通过每个层级提取点云特征和应用偏移注意力机制,有效地捕捉输入点云的关键特征信息;然后引入增益单元评估和缩放各个隐向量通道的重要度,消除向量通道间的信息冗余,仅需训练单个网络即可实现可变比特率压缩;为了预测特征向量的概率分布,对特征向量进行超先验编码,构造高斯熵模型,通过熵编码进一步降低编码量;最后在解码端采用子点卷积进行上采样重构原始点云,避免顶点的局部聚集,提高点云的重构质量.实验结果表明,在ShapeNet数据集上,以率失真曲线的BD-rate作为性能评价指标,与VRR和Draco方法相比,平均比特率分别降低48.66%和63.56%;压缩性能得到了显著的提升.Existing point cloud compression methods based on deep learning require training multiple networks,which consumes a large amount of time and space resources.To address this issue,a variable rate point cloud compression network model based on channel gain is proposed.First,the encoding end of the network uses a hi-erarchical structure to extract point cloud features and apply offset attention mechanisms at each level,effectively capturing the key feature information of the input point cloud.Secondly,gain units are introduced to evaluate and scale the channel of each hidden vector,eliminating information redundancy between vector channels,and re-quiring only one network to achieve variable rate compression.Then,to predict the probability distribution of feature vectors,a hyperprior encoder is performed on the feature vectors to construct a Gaussian entropy model,further reducing the coding volume through entropy coding.Finally,the decoder uses sub-point convolutions for upsampling to reconstruct the original point cloud,avoiding local point aggregation and improving the recon-struction quality.For the ShapeNet dataset,with BD-rate on the rate-distortion curve as the performance evalua-tion index,our method reduces the average bitrate by 48.66%and 63.56%compared with VRR and Draco meth-ods,respectively.The experimental results show that the compression performance has been significantly im-proved compared to current methods.

关 键 词：可变比特率 点云压缩 注意力机制 子点卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]