基于弱监督深度学习的三维模型多特征自适应形状分析方法研究

With the wide usage of 3D models in intelligent manufacturing, 3D printing and computer aided design, how to effectively achieve the semantic shape analysis has been an important issue that must be resolved as soon as possible. However, the traditional methods still cannot comprehensively utilize multiple features of 3D models very well and the cost of training data preparation still remain high. The precision rates of the analysis results are not high enough. The project plans to propose weakly-supervised adaptive multi-feature 3D shape analysis methods by extending the deep learning theory, and solve the main difficulty of the traditional methods. The main research contents include: . Firstly, the project will establish a deep learning model for the multi-feature 3D shape analysis to obtain the nonlinear mapping relationship between the different kinds of feature vectors and memberships of 3D shapes or their parts. . Secondly, the project will construct the penalty term describing both the non-consistency and non-sparsity of the weighting coefficients, propose the mechanism for solving the problem of possible conflicts among different features in the 3D shape analysis, and overcome the shortcoming of the original deep learning theory. . Finally, the project will extend the deep learning theory further to greatly reduce the cost of training data preparation by constructing novel loss functions, which describe both the objective function and the consistency of geometric features on the surface of 3D shapes. It will realize the weakly-supervised learning on massive 3D models. . The research of the project will improve the effectiveness and efficiency of the current 3D shape analysis methods further. Therefore, it will effectively improve the efficiency of the related industries and has a strong theoretical significance and application prospects.

随着三维模型在智能制造、3D打印、计算机辅助设计等领域的广泛应用，如何有效实现其形状语义分析，已成为亟待解决的重要问题。但传统方法仍存在多种特征难以有效综合利用、训练模型标注成本过高等问题，准确率并不高。本项目扩展深度学习理论，提出基于弱监督的三维模型多特征自适应形状分析方法，解决传统方法存在的主要问题。研究内容包括：.①建立三维模型多特征形状分析的深度学习模型，获取三维模型不同特征与所属分类之间的非线性映射关系；.②构造刻画多特征下权重系数不一致性和非稀疏性的双重惩罚项，提出多种特征之间潜在冲突性的解决机制，改进原始深度学习理论不足；.③构造刻画优化目标函数与三维模型表面几何特征一致性的双重损失函数，进一步扩展深度学习理论模型，大幅降低训练模型标注成本，实现面向海量模型的弱监督学习。.本项目的研究将进一步提高三维形状分析方法的效能，提升相关产业的生产效率，具有较强的理论意义和应用前景。

随着三维模型在智能制造、3D打印、计算机辅助设计等领域的广泛应用，如何有效实现其形状语义分析，已成为亟待解决的重要问题。但传统方法仍存在多种特征难以有效综合利用、训练模型标注成本过高等问题，准确率并不高。本项目以前期预研成果为基础，以进一步提高三维模型形状分析方法的效能，适应智能制造、3D打印、计算机辅助设计、动漫制作和游戏设计等产业规模化和自动化过程中的新需求为目标，紧密围绕基于弱监督的三维模型多特征自适应形状分析开展研究工作，取得了一系列三维形状分析基础理论和关键技术领域的重要研究成果。.项目组共提出8种基于深度学习的三维形状分析新算法，包括：基于涂鸦的弱监督学习三维形状分割算法、基于软密度峰值聚类和半监督学习的三维形状分割算法、基于投影神经网络的三维兴趣点检测算法、基于交替优化的三维模型兴趣点提取算法、基于分层学习的三维模型兴趣点提取算法、基于权重能量自适应分布的三维形状分割算法、基于深度体卷积神经网络的三维形状分割算法、基于生成对抗网络的点云形状保结构补全算法等。项目共发表学术论文11篇，其中SCI期刊6篇，EI期刊3篇，CCF推荐C类会议1篇，核心期刊1篇。其中6篇论文以项目负责人为第一作者发表在《IEEE Transactions on Visualization and Computer Graphics》、《IEEE Transactions on Multimedia》、《Computer-Aided Design》、《计算机辅助设计与图形学学报》等国内外计算机辅助设计与图形学领域权威学术期刊上。同时，这些算法已经编制成程序，申请国家发明专利9项，其中7项已获得授权。.这些成果有望在计算机辅助设计和工业软件领域进行应用，对提升相关产业生产效率起到推动作用，具有较强的理论意义和应用前景。