3D模型一般由若干部件构成，神经生物学研究表明部件是人类感知几何形状的基本单元，基于部件的形状分析和3D模型合成已成为当前计算机图形学研究的热点和难点。目前部件感知的几何造型还存在着部件检测鲁棒性差、分解层次性不强以及合成的合理性和可用性不足等问题亟需解决。本项目拟以低秩表示、谱分析和深度学习理论为基础，以解决基于部件的3D建模中自动的多层次部件提取等关键问题为目的，研究：1)部件的非局部特征表示；2）利用部件结构引导的低秩子空间聚类，自动融合多种部件特征；3）部件感知的谱网格分析和有层次结构的部件分解；4）基于深度学习的部件变异和交叉，得到针对复杂3D模型快速建模方法。对于上述问题的研究，将有助于提高几何形状分析和视觉感知的一致性；加强用户控制和生成模型的合理性，得到更多有创意的3D模型，为相关应用问题提供理论和算法支持。

Researchers pay more attentions on part-aware shape analysis and modeling, since it plays an important role in human recognition. There are still some challenges in this field such as the robustness of part detection, part decomposition with deep hierarchy, balancing the fitness and diversity of generated models. Based on the local rank representation, mesh spectral theory and deep learning theory, this project aims to study several primary problems in rapid 3D model generation. We intend to explore: 1) non-local feature expressions of different principles for defining parts; 2) the fusion of multiple part features by low rank subspace clustering with proper structural guidance; 3) part-aware spectral analysis method and hierarchical part decomposition based on it; 4)rapid 3D modeling by part mutation and crossover based on the deep learning theory. The research may benefit the existed shape analysis approaches by making the results respecting the human recognition more and increase the fitness of generated models, enhance the control of from the user and generate more creative 3D models.Meanwhile, we wish to offer theory and algorithm support for relative applications.