未知空间目标纹理缺失表面三维重建算法

针对传统三维重建方法难以对纹理缺失表面进行完整重建的问题，提出一种基于深度学习与截断符号距离函数(TSDF)融合的未知目标三维表面完整重建算法。首先设计一种基于深度学习的图像逐像素深度估计框架，通过在训练过程中引入多个复杂结构模型，提高该深度估计框架的泛化能力；其次，利用TSDF对各帧图像所估计的深度信息进行融合，实现对纹理缺失区域的空间目标完整三维重建。根据仿真校验，对于300 mm尺寸的卫星模型图像，像素深度估计平均误差约为13 mm，通过TSDF融合后尺寸精度误差小于5.10%。实验结果表明该算法可以对未知空间目标光学图像进行逐像素深度估计，并获得目标完整的三维结构与纹理信息，有效解决无纹理区域的重建结构缺失问题。

Three dimensional Reconstruction for Texture lacking Surfaces in Unknown Targets

Aiming at the problem that traditional three dimensional (3D) reconstruction methods are difficult for the reconstruction of the texture lacking surfaces, a 3D surface complete reconstruction algorithm of unknown target based on the fusion of deep learning and truncated signed distance function (TSDF) is proposed. Firstly, a pixel by pixel depth estimation framework based on deep learning is designed. By introducing multiple complex structural models in the training process, the generalization ability of the framework is improved. Secondly, TSDF is used to fuse the depth information estimated by each frame image to achieve the complete three dimensional reconstruction of the space target of the texture lacking area. It is validated that pixel wise depth estimation for an unknown object is achieved by the proposed learning based framework, and the average depth estimating error for each pixel is about 13 mm, and the reconstruction error is less than 5.10% by the TSDF fusion. Compared with the traditional methods, it is addressed in our method for texture lacking regions, and the complete three dimensional structure is obtained.