基于混合上下文熵模型的点云几何编码算法

Point cloud geometric coding algorithm based on hybrid context entropy model

针对三维点云存在的大量空域冗余信息,提出一种基于混合上下文熵模型的点云几何编码算法框架.通过多层感知机与Resnet网络分别对基于八叉树结构的点云和基于体素结构的点云特征进行上下文特征提取,并使用选择单元对上下文信息进行裁剪、选择和融合,使网络能够针对当前编码体素建立更加准确的概率模型,从而提高三维点云的压缩效果.同时,针对模型复杂度高的问题提出并行多尺度自回归进行概率估计的方案,大大降低了编解码时间.实验结果表明:点云几何编码算法能够有效降低每个体素所占的比特数并且整个编码过程无损;与G-PCC编码算法相比,压缩后比特率下降了14.27%.

Aiming at the large amount of spatial redundant information in 3D point cloud,a geometric coding algorithm framework of point cloud based on hybrid context entropy model is proposed in this paper.Through multi-layer perceptron(MLP)and Resnet network,the context features of point cloud based on octree structure and point cloud based on voxel structure are extracted respectively,and the selection unit is used to cut,select and fuse the context information,so that the network can establish a more accurate probability model for the current coding voxel,so as to improve the compression effect of three-dimensional point cloud.At the same time,aiming at the problem of high model complexity,this paper proposes a parallel multi-scale autoregressive probability estimation scheme,which greatly reduces the encoding and decoding time.Experimental results show that the proposed method can effectivel y reduce the number of bits per voxel,and the whole coding process is lossless.Compared with G-PC C coding algorithm,the bit rate after compression is decreases by 14.27%.