基于三维网格卷积的布料仿真模拟  

Fabric simulation based on 3D convolutional neural network

作　　者：靳雁霞[1] 刘亚变 杨晶[1] 史志儒 张翎 乔星宇 JIN Yan-xia;LIU Ya-bian;YANG Jing;SHI Zhi-ru;ZHANG Ling;QIAO Xing-yu(School of Data Science and Technology,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学计算机科学与技术大数据学院,山西太原030051

出　　处：《计算机工程与设计》2024年第4期1264-1271,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(62071281);山西省自然科学基金项目(202103021224218)。

摘　　要：针对当前布料仿真中模拟成本高且耗时长的问题,提出一种直接对三维布料网格卷积的布料模拟方法。对三维布料网格进行螺旋卷积,使用点填充方法增加采样点,存储局部域;采用顶点抽取算法对局部块进行池化,使用基于Gauss-Bonnet定理曲率的惩罚因子对顶点抽取算法进行增强,当局部块数量小于开始时数量的60%时停止顶点抽取;最后训练神经细分网络对抽取后的整体网格进行上采样,其中使用改进的蝶形细分算法生成新顶点。实验结果表明,与目前已有方法相比,该布料模拟方法能够保留丰富真实的褶皱,减少模拟成本和时间,是一种高效的布料模拟方法。For the costly and time-consuming simulation problem in the current fabric simulation,a cloth simulation method for the direct convolution of 3D cloth mesh was proposed.The 3D cloth mesh was spiral convolved,the point filling method was used to increase the sampling point,and the local domain was stored.The vertex extraction algorithm was used to pool the local blocks,and the penalty factor based on the curvature of the Gauss-Bonnet theorem was used to enhance the vertex extraction algorithm,and the vertex extraction was stopped when the number of local blocks was less than 60%of the initial number.Final neural segmentation network was trained to upsample the extracted overall grid,in which new vertices were generated using the improved butterfly segmentation algorithm.Experimental results show that the cloth simulation method not only retains the rich real folds,but reduces the simulation cost and time.

关 键 词：布料仿真 螺旋卷积 局部域 惩罚因子 顶点抽取 神经细分网络 蝶形细分算法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]