摘要
针对现有的基于深度学习的三维模型识别与分割方法忽略了三维模型高级全局单点特征和低级局部几何特征之间的关系而导致识别效果不佳的问题,提出了一种多特征融合的三维模型识别与分割方法。首先,通过加深卷积核的宽度和网络的深度构建全局单点网络以提取具有高级语义识别能力的全局单点特征;其次,通过构建注意力融合层学习全局单点特征和局部几何特征的隐含关系来充分挖掘更能表征模型类别的细粒度几何特征;最后,将全局单点特征和细粒度几何特征进一步融合,达到优势互补、增强特征丰富性的目的。分别在三维模型识别数据集ModelNet40、ModelNet10和分割数据集ShapeNet Parts、S3DIS、vKITTI上进行了实验验证,并与当前主流识别算法进行了对比,表明该算法不仅有更高的识别和分割准确率,而且具有更强的鲁棒性。
Current methods focusing on 3D model recognition and segmentation have to some extent ignored the relationship between the high-level global single-point features and the low-level local geometric features of those models,resulting in poor recognition results.A multi-feature fusion approach which takes into consideration the aforementioned ignored relationship is proposed.First,a global single-point network is established to extract the global single-point features with high-level semantic recognition ability by increasing both the width of convolution kernel and the depth of the network.Second,an attentional fusion layer is constructed to learn the implicit relationship between global single-point features and local geometric features to fully explore the fine-grained geometric features that can better represent model categories.Finally,the global single-point features and fine-grained geometric features are further fused to achieve the complementation of advantages and enhance the feature richness.Experimental verification is carried out on the 3D model recognition datasets ModelNet40,ModelNet10 and segmentation datasets ShapeNet Parts,S3DIS,vKITTI,respectively,and comparison with current mainstream recognition algorithms shows that the proposed algorithm not only has higher recognition and segmentation accuracy,but also has stronger robustness.
作者
党吉圣
杨军
DANG Jisheng;YANG Jun(School of Electronic and Information Engineering,Lanzhou Jiaotong University,Lanzhou 730070,China)
出处
《西安电子科技大学学报》
EI
CAS
CSCD
北大核心
2020年第4期149-157,共9页
Journal of Xidian University
基金
国家自然科学基金(61862039,61462059)。
关键词
三维点云
目标识别
语义分割
注意力融合
深度学习
卷积神经网络
3D point cloud
object recognition
semantic segmentation
attention fusion
deep learning
convolutional neural networks
