随着航空航天技术的不断发展，航拍图像在军事侦查、地形绘制等方面取得了广泛的应用，由于航拍图像拍摄高度很高，图像上的一点误差对应到地面上可以造成几十米甚至上百米的误差，因此，图像配准技术应运而生。一般的图像配准方法在计算时间和配准精度上都存在一些不足，经过调研和实验，本文从这两个方面对图像配准算法进行了改进。图像配准算法一般采用基于特征的配准方法，其中Scale Invariant Feature Transform（SIFT）算法提取出来的点特征具有仿射不变性，非常适合对不同角度、不同方位拍摄的航拍图像处理，经过实验证明效果也最佳，因此本文采用SIFT算法来进行特征提取和匹配。本文的主要研究内容有以下几个方面：1.由于航拍图像分辨率很高，有些图像甚至达到上百兆，用一般的图像配准算法来处理会因为内存的问题变得非常缓慢，甚至出现无法处理的情况。本文采用了将航拍图像进行分块的方式，并用多线程来分别对这些图像块进行SIFT特征提取，大大节省了特征提取的时间。2.为了避免由于图像分块的原因造成特征提取的丢失，本文在用SIFT特征提取的时候采用了跨图像的极值检测方法，保证了特征提取的完整性。3.提取特征并完成特征匹配之后，将得到一系列匹配点对，本文用最小二乘法对仿射变换参数进行求解，并提出了一种参数迭代更新法，通过剔除大误差的匹配点，循环修改参数，提高了原SIFT算法的配准精度。

With the continuous development of the aviation and aerospace technology, aerial images have been widely used in the military investigation and terrain rendering fields. Because the aerial images are shot from a very high distance, it can cause dozens of meters or hundreds of meters of error on the ground correspond to a little error on the images. Image registration was brought in to solve this kind of problem. However, there are some inadequacies in computing time and registration accuracy of the general image registration methods. After some research and experimentations, this paper improved the algorithm from both aspects.Image registration algorithm commonly uses the feature-based registration method. The point features extracted by Scale Invariant Feature Transform (SIFT) algorithm have affine invariance, is very suitable for the aerial images which are shot by different perspectives and different directions. And it was proved to be the best by our experiments, so we used the SIFT algorithm to do the feature extraction and points matching.This paper contains the following aspects:1. For high resolution of aerial images, some even up to hundreds of megabytes, the general image registration algorithm makes the processing become very slow or even lead to failure because of memory problems. Consequently we divided the aerial image into blocks, and extract point features using multiple threads, which saves a lot of time.2. In order to avoid losing features caused by image segmentation, we detected the extremum value between the image blocks, so that we can ensure the integrity of the feature extraction.3. After the feature extraction and feature matching, we get a series of matched points. We calculated the affine transformation parameters using the least squares method and iterative updated the parameters by removing large error point, by which we finally achieved high-precision registration.