基于深度学习的三维模型重构研究

Research on 3D model reconstruction based on deep learning

由单个图像建立其三维模型是计算机视觉领域的一个热门且具有挑战性的问题。现有的传统单视图三维重构算法在处理低分辨率图像时效果不好,在训练中由于三维图形的高维性,使网络也变得高度不稳定,导致模型重构效果差。针对传统三维重构算法存在的缺点,提出一种基于深度学习网络的改进模型,在模型中加入超分辨率、投影、对抗生成网络(generative adversarial network, GAN)等模块,采用模块化设计强制生成的三维形状与深度图像对齐,使得映射更加规则。在损失函数上运用Wasserstein GAN思想,引入惩罚项,使网络训练难度降低,减小网络模型对训练数据集的依赖,克服了传统算法存在的问题。实验证明,提出的模型较传统方法重构的三维模型更加逼真,符合客观事实。

It is a hot research direction in the field of computer vision to build its three-dimensional model from a single image,and this is a challenging problem.The traditional single view 3D reconstruction’s algorithm is not good at dealing with low-resolution images.Due to the high dimension of 3D images,the network will become highly unstable in training,resulting in poor model reconstruction effect.In view of the shortcomings of traditional 3D reconstruction algorithm,this paper proposes an improved model based on deep learning network,in which modules,such as super-resolution,projection,generative adversarial network,are added,and the 3D shape forced by modular design is aligned with the depth image,which makes the mapping more regular.In addition,Wasserstein Gan idea is used in loss function,and penalty term is introduced to reduce the difficulty of network training.This idea reduces the dependence of network model on training data set and solves the problems of traditional algorithm.The experimental results show that compared with the traditional three-dimensional model,the model in this paper is more realistic and conforms to the objective facts.