基于多特征融合的三维形状分割方法  

A 3D Shape Segmentation Method Based on Multi-feature Fusion

作　　者：赵天宇[1,2] 李海生 吴晓群[1,2] 蔡强 Zhao Tianyu;Li Haisheng;Wu Xiaoqun;Cai Qiang(School of Computer and Information Engineering,Beijing Technology and Business University,Beijing 100048;Beijing Key Laboratory of Big Data Technology for Food Safety,Beijing 100048)

机构地区：[1]北京工商大学计算机与信息工程学院,北京100048 [2]食品安全大数据技术北京市重点实验室,北京100048

出　　处：《计算机辅助设计与图形学学报》2018年第11期2011-2017,共7页Journal of Computer-Aided Design & Computer Graphics

基　　金：国家自然科学基金(61602015);北京市自然科学基金(4162019);北京航空航天大学虚拟现实技术与系统国家重点实验室开放课题(BUAAVR-16KF-06)

摘　　要：三维形状分割是三维形状分析中的一个重要问题.针对单一特征对同一类模型分割结果存在较大差异的问题,提出一种基于学习的多特征融合的三维形状分割方法.首先利用过分割方法将三维模型分割成多个子面片,分别对每个子面片提取多种几何特征;然后将几何特征作为低层特征输入深度神经网络模型,通过学习生成高层特征;最后基于该高层特征用高斯混合模型的方法得到聚类中心,利用图割得到最后分割结果.在普林斯顿标准数据集和COSEG数据集上的实验结果表明,与传统分割方法相比,该方法具有较好的一致性分割结果.3D shape segmentation is an important issue in shape analysis.This paper proposed a segmentation method based on multi-feature fusion to solve the consistency problem.A 3D model is first divided into multiple sub-patches by using over-segmentation.Then we use the geometric features extracted from each sub-patch as low-level features input for depth neural network model to generate high-level features.Finally,based on these high-level features,Gaussian mixture model is employed to get the clustering centers and graph-cut is adapted for the final segmentation.Experiments on PSB and COSEG datasets show that the proposed method outperforms the traditional geometric feature method,and can get good consistency results for the same kind of 3D shapes.

关 键 词：三维形状分割 过分割 深度学习 高斯混合模型 图割 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]