卷积神经网络的兴起产生了大量基于视图的三维模型识别方法,不同的视图融合方式影响了网络模型的特征提取性能.本文提出了一种自适应视图融合方法,将视图的动态奇异值信息作为三维模型的特征描述符,获得三维模型全局特征的方式由区域化...卷积神经网络的兴起产生了大量基于视图的三维模型识别方法,不同的视图融合方式影响了网络模型的特征提取性能.本文提出了一种自适应视图融合方法,将视图的动态奇异值信息作为三维模型的特征描述符,获得三维模型全局特征的方式由区域化分块、自适应SVD(Singular Value Decomposition)分解和维度压缩三部分组成,通过分块后的子区域极大地关注三维模型的局部特征,并用自适应的方法判断每个局部特征的影响大小,最后维度压缩去除较小影响的数值.动态奇异值网络是将这三部分作为卷积神经网络的后端,形成一个端对端(end to end)可训练的三维模型特征提取框架.与当今先进方法相比,在ModelNet40数据集上的分类和检索结果分别提升了1. 2%和0. 8%,在ModelNet10和ModelNet40的Top-10平均检索精度分别提高了3. 7%和4%.展开更多
Faults and fractures of multiple scales are frequently induced and generated in compressional structural system. Comprehensive identification of these potential faults and fractures that cannot be distinguished direct...Faults and fractures of multiple scales are frequently induced and generated in compressional structural system. Comprehensive identification of these potential faults and fractures that cannot be distinguished directly from seismic profile of the complex structures is still an unanswered problem. Based on the compressional structural geometry and kinematics theories as well as the structural interpretation from seismic data, a set of techniques is established for the identification of potential faults and fractures in compressional structures. Firstly, three-dimensional(3D) patterns and characteristics of the faults directly interpreted from seismic profile were illustrated by 3D structural model. Then, the unfolding index maps, the principal structural curvature maps, and tectonic stress field maps were obtained from structural restoration. Moreover, potential faults and fractures in compressional structures were quantitatively identified relying on comprehensive analysis of these three maps. Successful identification of the potential faults and fractures in Mishrif limestone formation and in Asmari dolomite formation of Buzurgan anticline in Iraq demonstrates the applicability and reliability of these techniques.展开更多
文摘卷积神经网络的兴起产生了大量基于视图的三维模型识别方法,不同的视图融合方式影响了网络模型的特征提取性能.本文提出了一种自适应视图融合方法,将视图的动态奇异值信息作为三维模型的特征描述符,获得三维模型全局特征的方式由区域化分块、自适应SVD(Singular Value Decomposition)分解和维度压缩三部分组成,通过分块后的子区域极大地关注三维模型的局部特征,并用自适应的方法判断每个局部特征的影响大小,最后维度压缩去除较小影响的数值.动态奇异值网络是将这三部分作为卷积神经网络的后端,形成一个端对端(end to end)可训练的三维模型特征提取框架.与当今先进方法相比,在ModelNet40数据集上的分类和检索结果分别提升了1. 2%和0. 8%,在ModelNet10和ModelNet40的Top-10平均检索精度分别提高了3. 7%和4%.
基金Project(2014CB239205)supported by the National Basic Research Program of ChinaProject(20011ZX05030-005-003)supported by the National Science and Technology Major Project of China
文摘Faults and fractures of multiple scales are frequently induced and generated in compressional structural system. Comprehensive identification of these potential faults and fractures that cannot be distinguished directly from seismic profile of the complex structures is still an unanswered problem. Based on the compressional structural geometry and kinematics theories as well as the structural interpretation from seismic data, a set of techniques is established for the identification of potential faults and fractures in compressional structures. Firstly, three-dimensional(3D) patterns and characteristics of the faults directly interpreted from seismic profile were illustrated by 3D structural model. Then, the unfolding index maps, the principal structural curvature maps, and tectonic stress field maps were obtained from structural restoration. Moreover, potential faults and fractures in compressional structures were quantitatively identified relying on comprehensive analysis of these three maps. Successful identification of the potential faults and fractures in Mishrif limestone formation and in Asmari dolomite formation of Buzurgan anticline in Iraq demonstrates the applicability and reliability of these techniques.