伴随无人机技术的快速兴起，城市三维重建发展越来越快。在利用无人机影像进行大场景三维重建过程中，为了获得与实际地物更加契合、精度更高的模型，模型优化成为了重建过程中不可或缺的一环。但利用无人机影像三维重建的过程中，存在影像数量多、重叠度高等特点，因此在优化过程中往往存在由于数据量大导致计算机内存占用过大而造成优化失败的可能性。 本文提出一种基于影像信息的三维模型网格自适应快速优化方法，从而解决大场景倾斜影像三维重建过程中网格优化阶段出现计算机内存占用过大、优化效率低的问题。 网格优化的实质是在网格顶点梯度的不断计算过程中实现同名像点间相关系数达到最大。在网格顶点梯度迭代计算过程中，基于初次梯度计算结果，自适应将网格三角形划分为两种不同的标记：特征丰富的活跃三角形及位于平面区域特征不丰富的怠惰三角形。对活跃三角形继续迭代计算顶点梯度从而驱动顶点移动，对怠惰区域三角形进行简化减少平面区域三角形数量从而实现网格优化过程中对优化效率的快速提升以及减少对计算机内存的占用。 本文主要内容和成果如下： （1）在网格优化过程中，为了获得更好的模型，通过构建影像金字塔，降低影像分辨率实现在不同分辨率等级上的优化，获得由粗到细的优化效果。在网格优化过程中进行分块优化，减小优化过程中的内存消耗，间接提升优化效率。在顶点梯度的解算过程中构建稀疏矩阵，通过稀疏矩阵求解可直接求出网格中所有顶点的梯度值。 （2）本文采用影像包含的灰度信息对三维模型进行优化。在优化过程中，经几何投影将影像经过物方网格三角形诱导投影到与其有公共可视区域的影像上组成相关影像对，利用投影后的同名像点间的灰度信息计算影像对间的零均值归一化相关系数并对网格中物方三维点坐标进行求偏导函数，得到顶点梯度值，通过顶点梯度驱动网格的优化。 （3）针对使用大场景倾斜影像三维重建过程中网格优化效率过低的问题，本文构建了网格优化精度与优化效率函数，求该函数的最优解，得到网格优化中精度与效率的平衡点，对顶点梯度值进行评价，并依据评价结果将网格中的三角形自适应标记为活跃三角形或怠惰三角形，通过对活跃三角形继续进行优化，放弃怠惰三角形的优化并将其简化，减少对计算机内存的占用，实现了网格优化效率的大幅度提升。

With the rapid rise of drone technology, urban 3D reconstruction is developing faster and faster. In the process of 3D reconstruction of large scenes using UAV images, model refinement has become an indispensable part of the reconstruction process in order to obtain a model that is more compatible and more accurate with actual objects. However, in the process of 3D reconstruction of UAV images, there are many features such as large number of images and high degree of overlap. Therefore, in the optimization process, there is a possibility that the optimization of the computer memory is too large due to the large amount of data. In this paper, 3D model mesh adaptive fast refine method based on image information is proposed to solve the problem of excessive memory usage and low efficiency in the refinement stage during the 3D reconstruction process using large scene images. The essence of mesh refinement is to achieve the max ZNCC between the same-named pixels in the continuous calculation process of the mesh vertex gradient. In the mesh vertex gradient iterative calculation process, based on the initial gradient calculation results, the mesh triangle is adaptively divided into two different markers: an active triangle with rich features and a inactive triangle with less features in the planar region. Continue to iteratively calculate the vertex gradient for the active triangle to drive the vertex movement, simplify the inactive area triangle, reduce the number of plane area triangles, and achieve rapid improvement of refine efficiency and reduce computer memory usage during grid refinement. The main contents and results of this paper as follows: (1) In the mesh refinement process, in order to obtain a better model, the image resolution is constructed, and the image resolution is reduced to achieve refinement at different resolution levels, the coarse-to-fine refinement effect is obtained. In the mesh refinement process, the block refinement is solved, the memory consumption in the refinement process is reduced, and the refinement efficiency is indirectly improved. and the sparse matrix is constructed in the process of solving the vertex gradient. The sparse matrix solution can directly find the gradient values of all the vertices in the mesh . (2) This paper uses image gray information to refine the 3D model. In the refinement process, the image is projected onto the image with its common visible area by geometric projection to form a related image pair, and the image pair is calculated by using the gray level information between the image pairs. The zero-mean normalization of the cross-correlation coefficient and the three-dimensional point coordinates of the object in the grid are derived, and the gradient change value of the vertex is obtained, and the mesh optimization is driven by the gradient of the vertex. (3) For the low efficiency in the refinement process of 3D reconstruction using large scene UAV images, this paper constructs the mesh refinement precision and mesh refinement efficiency function, finds the optimal solution of the function, and obtains the mesh refinement. Balance between accuracy and efficiency, evaluate the gradient values of the vertices, and adaptively mark the triangles in the grid as active triangles or slotted triangles according to the evaluation results. By refining the active triangles, the refine of the slotted triangles is abandoned and Simplify, reduce the occupation of computer memory, and achieve a significant increase in mesh refinement efficiency.