【作者】 刘伟；

【导师】 华顺刚； 马赛；

【作者基本信息】 大连理工大学 ， 机械工程， 2025， 硕士

【摘要】 随着数智化设计、互联网技术、AI技术的快速发展,工业产品数字化模型数量激增,在海量的三维模型中快速准确地对零件模型进行归类、检索成为计算辅助设计及图形学领域的研究热点。本文基于深度学习网络从零件模型的点云和多视图数据中提取模态特征,实现零件模型的识别与检索。研究了零件模型的预处理方法。针对STL格式的零件模型以面积大小为概率选择三角面片,并在选中的三角面片上等位置概率采样出三维点,如此循环采样出固定点数且分布均匀的零件模型点云;为了获得零件模型的多个视图,基于Blender渲染工具,利用多线程技术批量导入零件模型,自动调整虚拟相机视角,并按指定角度旋转模型,生成零件模型多视角灰度图,该方法适用于大规模数据集的处理。研究了一种基于深度学习的零件模型识别方法。针对Point Net++网络的结构进行了改进,在每个集合抽象层内部和不同集合抽象层之间引入自注意力机制,这种设计使网络能够自动判别不同层次的点云特征的重要性,增强了对零件模型的识别性能。改进后的Point Net++网络在零件数据集上的测试实例准确率和类别准确率分别为92.6%和91.9%。结合GVCNN网络和改进Point Net++网络提出了一种模态特征融合网络,使多视图和点云特征在公共特征空间上更好地融合,再利用全连接层实现零件模型识别。零件模型识别实验结果表明:该融合网络在零件数据集上的测试实例准确率和类别准确率分别为96.3%和95.8%。研究了一种基于深度学习的零件模型检索方法。将模态特征融合网络提取到的特征向量作为零件模型的特征描述符,选择皮尔逊相关系数作为零件模型的相似性度量方法。开发了基于深度学习的零件模型检索系统,利用深度神经网络提取查询零件模型的特征描述符,与数据库中的零件模型特征进行相似度计算,实现了零件模型检索。零件模型检索实验结果表明:本文提出的模态特征融合网络在零件数据集上的平均精度均值为92.8%,验证了该检索方法的有效性。更多还原

【Abstract】 With the rapid development of digital and intelligent design,Internet technology,and AI technology,the number of digital models of industrial products has surged.Quickly and accurately classifying and retrieving part models from a massive number of 3D models has become a research hotspot in the community of computer-aided design and graphics.In this thesis,modal features are extracted from the point cloud and multi-view of part models based on a deep learning network to achieve the recognition and retrieval for part models.The preprocessing method of part models is studied.For part models with STL format,triangular patches are selected according to the probability corresponding to their area size,and three-dimensional points are sampled with the equal position probability on the selected triangular patches.In this way,the point cloud for a part model with a fixed number of points and uniform distribution is sampled cyclically.By adjusting automatically the perspective of the virtual camera and rotating the model,several views of a part model are obtained using multi-threading technology with Blender rendering tool.This method is suitable for the processing of large-scale datasets.A part model recognition method based on deep learning is studied.By improving the structure of the Point Net++network,the self-attention mechanism is introduced into each Set Abstraction layer and between different Set Abstraction layers.This design enables the network to automatically determine the importance of point cloud features at different levels and enhances the recognition performance for part models.The improved Point Net++network achieved test instance accuracy and class accuracy of 92.6%and 91.9%respectively on the part dataset.A modal feature fusion network is proposed by combining the GVCNN network and the improved Point Net++network,which enables better feature fusion between multi-view and point cloud in the same feature space,and achieves part model recognition with the fully connected layer.Recognition experiments with our network shown that:the test instance accuracy and class accuracy of this fusion network on the part dataset are 96.3%and 95.8%respectively.A part model retrieval method was studied based on deep learning.The feature vectors extracted by the modal feature fusion network were used as the feature descriptors for part models.The Pearson correlation coefficient was selected to measure the similarity among part models.With our part model retrieval system,the deep neural network was used to extract the feature descriptor for a query,then calculate the feature similarity among part models in the database to realize part model retrieval.Retrieval experiments shown that:the mean average precision of our modal feature fusion network on the part dataset is 92.8%,verifying the effectiveness of our retrieval method.更多还原