结构计算机感知,结构感知深度学习的三维形状分类方法

韩丽, 朴京钰, 兰鹏燕, 王晓旻, 于冰, 佟宇宁, 徐圣斯

(辽宁师范大学计算机与信息技术学院 大连 116081)

分类号：

TP391.41 DOI: 10.3724/SP.J.1089.2021.18280

出版年,卷(期):页码:

2021

,

33

(

1

):

29-38

韩丽

摘要：

为了解决复杂、海量三维模型的形状识别问题, 提出了一种结构感知深度学习的三维形状分类方法. 通过联合学习三维模型的几何结构和空间结构, 生成具有结构感知的特征向量表示, 该特征向量具有更强的识别力与稳定性, 在三维形状分类中取得显著的效果. 首先, 提取优化的多尺度热核特征, 并通过CNN学习模型, 有效地获取三维形状的几何结构特征; 其次, 建立多视图卷积学习网络提取三维形状的空间结构特征; 最后, 通过联合优化学习生成具有结构感知的深度特征表示. 文中采用了C++, Matlab, TensorFlow框架实现, 并在公开的三维数据库中进行了大量实验, 实验结果表明, 文中方法获取的深层结构特征对于复杂拓扑结构、大尺度几何形变的三维形状具有稳定性; 与相关方法对比, 在三维形状分类中具有更高的分类精度.

关键词：

形状分类; 结构感知; 卷积神经网络; 深度学习

3D Shape Classification Method Based on Shape-Aware Deep Learning

Han Li, Piao Jingyu, Lan Pengyan, Wang Xiaomin, Yu Bing, Tong Yuning, and Xu Shengsi

(School of Computer Science and Information Technology, Liaoning Normal University, Dalian 116081)

abstract：

In order to solve the problem of complex and massive 3D shapes recognition, we propose a structure-aware deep learning model for 3D shape classification. By learning the geometric structure and spatial structure jointly, we generate a structure-aware deep feature vector, which has stronger discriminative ability and stability on 3D shape classification. Firstly, optimal multi-scale HKS features are extracted to construct discerning geometric shape descriptor based on a CNN learning model. Secondly, a multi-view based CNN learning framework is built to extract spatial structure feature. Finally, all the features are jointly learned and generate a structure-aware feature vector. We explore our method by using C++, Matlab, TensorFlow platform, a serial of experimental analysis are carried out in the public 3D databases, they have shown that our deep features are stable for complex topological structure and large-scale geometric deformation. Compared with related methods, this method has higher classification accuracy in 3D shape classification.

keyword：

shape classification; structure-aware; convolutional neural networks; deep learning