基于深度张量学习的三维重建关键技术及并行研究

图像三维重建通过二维几何信息和拓扑信息进行三维描述，因此，三维重建技术是机器视觉、机器人识别检测、自动驾驶和3D电影等领域的共性科学问题和核心技术。本项目围绕深度三维重建问题实现高效和精准三维重建估计，研究如何提高深度三维重建方法计算高效性，研究对于自然无约束条件下物体人脸和人体姿态三维重建和估计鲁棒性。本项目深入结合深度网络图像深度语义信息表达能力和张量精简高维表达能力及充分反映事物属性特性，以及注意力机制和非监督学习充分利用图像自身结构和特性信息学习能力，提出基于深度张量三维重建并行算法、基于形状结构学习的物体三维重建、基于多属性结构约束的人脸三维重建、基于多映射拟合的人体姿态三维估计。本项目所研究技术能够有效解决深度三维重建网络计算存储耗费大和低效率问题，有效对抗物体人脸和人体姿态在自然无约束环境下的变化，有效提升视觉领域问题处理效率，通过并行计算技术有效提升视觉领域解决问题能力。

Image 3D reconstruction is described by 2D geometric information and topological information. 3D reconstruction technology is a common scientific problem and core technology in many fields such as machine vision, robot recognition and detection, automatic driving and 3D movie. For the sake of improving both the efficiency and accuracy of 3D reconstruction in the natural unconstrained environment. Based on the deep network's deep semantic information expression ability, Tensor's simplified high-dimensional expression ability and fully reflecting the characteristics of things' attributes, as well as the attention mechanism and unsupervised learning which make full use of the image's structure and feature information learning ability, this project proposes the 3D reconstruction parallel algorithm based on deep tensor learning, the shape structure learning network for object 3D reconstruction, the multi-attribute structure constraints network for face 3D reconstruction，and the multi mapping fitting network for human pose 3D estimation. These methods can effectively solve the problem of large storage costs and low efficiency of deep 3D reconstruction networks computing, effectively against the changes of the face and human posture in wild conditions, effectively promote the processing efficiency of visual field problems, and effectively promote the ability to solving visual field problems through parallel computing technology.