融合图卷积残差网络与边收缩池化的VQ-VAE网格重建算法  

VQ-VAE mesh reconstruction algorithm integrating graph convolution residual networks and edge contraction pooling

作　　者：丁阳 杨华民[1] 韩成[1] 刘宇 卢时禹 DING Yang;YANG Huamin;HAN Cheng;LIU Yu;LU Shiyu(School of Computer Science and Technology,Changchun University of Science and Technology,Changchun 130022,China;Faculty of Computer Science and Information Technology,University Putra Malaysia,Serdang 43400,Malaysia)

机构地区：[1]长春理工大学计算机科学技术学院,长春130022 [2]马来西亚博特拉大学计算机科学与信息技术学院,沙登43400

出　　处：《重庆理工大学学报（自然科学）》2024年第10期112-121,共10页Journal of Chongqing University of Technology：Natural Science

基　　金：吉林省自然科学基金项目(20220101134JC)。

摘　　要：3D网格因其复杂性和不规则性使其有效表示成为挑战。为解决常规图卷积难以对3D网格有效传递和融合信息的问题,提出基于变分量化自编码器的3D网格模型,以探索其隐空间并用于3D网格的生成。提出带残差的图卷积模块,在处理3D网格这种复杂的图结构时,残差连接更有效地整合多层特征信息,支持更深的网络结构,显著提升模型的性能和泛化能力。在网格简化的边收缩算法基础上构建了可靠的多层池化和反池化操作,有效编码了层次结构中较粗糙和较稠密网格之间的对应关系。将3D网格形状投影到潜在空间的过程中,潜在特征被过度压缩导致信息损失,采用向量量化将潜在特征映射到预先定义的离散向量,在保持紧凑表示下更有效地编码和重建数据。实验结果表明,所提算法能够学习对可变形形状集合的紧凑表示,且在形状生成、形状插值等各种应用中表现出色。Challenges exist for the effective representation of 3D meshes due to their complexity and irregularity.To address the limitations of conventional graph convolution in propagating and integrating information across 3D meshes,this paper proposes a 3D mesh model based on variational autoencoders with vector quantization to explore their latent space for 3D mesh generation.The introduction of residual graph convolution modules,specifically designed for intricate graph structures like triangular meshes,enhances the integration of multi-layered feature information through residual connections,supporting deeper network architectures and significantly improving model performance and generalization.Building upon a reliable edge contraction algorithm for mesh simplification,a hierarchical structure is encoded through robust multi-level pooling and unpooling operations,effectively capturing correspondences between coarser and denser meshes.Meanwhile,in the process of projecting 3D mesh shapes into the latent space,potential feature compression leading to information loss is addressed by employing vector quantization to map latent features to predefined discrete vectors.Our experimental results demonstrate the proposed algorithm learns compact representations for deformable shape collections,delivering outstanding performances in various applications such as shape generation and interpolation.

关 键 词：网格生成 变分量化自编码器 网格插值 图卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]