面向三维形变目标的矢量型深度卷积网络研究

传统深度卷积网络不擅长处理视点变化引起的目标旋转和尺寸变化，难以直接应用于Mesh网格描述的三维目标，本项研究拟借鉴胶囊网络的矢量化优势，设计实现一种全新的矢量型三维Mesh网格深度卷积网络，将卷积操作扩展到三维不规则曲面，将传统池化操作扩展为新型的矢量加权组合池化；全新设计多层堆叠型三维网格卷积自编码器和生成对抗网络，通过三维Mesh金字塔保留目标几何结构，采用无监督方法学习三维目标的多层次结构特征；针对三维目标形变难点，引入等距变换不变的内蕴特征反映三维目标内在固有结构信息，有机结合几何和表观特征，最大限度地消除形变对目标分析与识别的影响；针对无监督学习的模型选择难题，采用先进的Dirichlet过程混合模型实现三维形变目标的流形聚类分析，自动确定流形类别数；最后提取深度学习的训练成果，实现一个三维形变目标分析应用原型系统，取得三维图像分析技术的新突破。

Traditional deep convolutional networks are not good with changes in object orientation and scale caused by the different viewpoint. It is difficult to directly apply the convolutional networks to the 3d objects in mesh representation. This project is aimed at the development of a novel mesh convolution vector network for 3D deformable objects. The mesh convolution is operated on the irregular surfaces of 3d object. Commonly used pooling is extended to a new type of weighted vector combination. New design of stacked 3d mesh autoencoder and generative adversarial network is carried out. Unsupervised deep learning is used for the feature hierarchies with a mesh pyramid to preserve the geometry of the 3d object. Geodesic features are utilized to reflect the intrinsic properties of the deformable objects, which offer a robust and isometric deformation invariant of the object. The project also intends to develop an advanced manifold clustering method capable of automatic model selection using powerful Dirichlet Processing Mixture Models. The results of deep learning are to be applied in a prototyping system of 3D deformable object analysis.