低空大倾角立体影像自动匹配方法研究

发布时间：2019-05-19 06:36

【摘要】：近年来,随着“数字地球”和“智慧城市”的建设,三维建模技术迅速发展。低空倾斜摄影测量采用多角度拍摄,可同时获取地物顶端以及侧面纹理等多方位的信息,满足三维建模对影像数据的要求。将倾斜摄影测量获取的数据应用于三维建模中,可节约三维建模成本,同时能提高三维建模的速度,在三维建模中比传统低空摄影测量更具优势。然而,低空倾斜摄影测量由于在多个角度对地物进行拍摄,影像包含复杂的三维场景信息以及更多的信息量、数据量和数据冗余,具有更大的辐射畸变和几何变形,影像中地物互相遮挡、地物阴影、视差断裂等现象也较为普遍,这些使倾斜影像匹配难度大大增加。本文针对低空大倾角立体影像自动匹配方法进行研究,结合当前计算机视觉研究热点之一的深度学习方法,提出深度学习辅助的影像匹配方法,采用深度学习方法对影像进行分类,基于分类结果进行影像匹配。研究内容主要包括:(1)提出深度学习辅助的影像匹配方法。提出一种深度卷积神经网络递归识别模型,选取影像样本对卷积神经网络进行训练,然后输入待分类影像进行分类。模型首先对输入影像进行初始识别,然后自动判别单位格网内的分类情况,并根据分类情况进行精细识别与分类,最终精确识别和定位场景目标;(2)在影像精细场景分类结果的基础上,对影像进行特征提取和匹配。以VS2008为编程平台,采用C++语言,将基于SIFT和Harris-Laplace的影像匹配算法进行实现。首先,建立影像多尺度空间,检测其中的Harris兴趣点,并采用LoG算子对其进行筛选得到特征点,采用SIFT方法进行特征描述。在搜索匹配点阶段,根据特征点所属场景类型在待匹配影像相同场景中进行搜索。从而减少匹配搜索量,提高匹配精度和效率;(3)采用本文场景分类方法以及深度学习辅助下的影像匹配方法进行实验,对实验结果进行分析。实验结果表明,基于深度学习的场景分类方法对高分辨率遥感影像场景的分类可取得较高精度的结果,配合本文的影像匹配算法,可有效增加匹配点数量、提高匹配精度和效率。
[Abstract]:In recent years, with the construction of Digital Earth and Smart City, 3D modeling technology has developed rapidly. Low altitude tilt photogrammetry adopts multi-angle photography, which can obtain multi-directional information such as top and side texture of ground object at the same time, which can meet the requirements of 3D modeling for image data. The application of the data obtained by tilt photogrammetry to 3D modeling can save the cost of 3D modeling and improve the speed of 3D modeling. It has more advantages than traditional low altitude photogrammetry in 3D modeling. However, low altitude tilt photogrammetry contains complex 3D scene information and more information, data amount and data redundancy, and has greater radiation distortion and geometric deformation because of taking pictures of ground objects from multiple angles. The phenomena such as mutual occlusion of ground objects, shadow of ground objects, parallax fracture and so on are also common in images, which greatly increase the difficulty of oblique image matching. In this paper, the automatic matching method of low altitude and large inclination stereo image is studied, and combined with the depth learning method, which is one of the hotspots of computer vision research, a deep learning aided image matching method is proposed. The deep learning method is used to classify the images, and the image matching is carried out based on the classification results. The main contents of this paper are as follows: (1) A deep learning aided image matching method is proposed. In this paper, a recurrent recognition model of deep convolution neural network is proposed. The image samples are selected to train the convolution neural network, and then the images to be classified are input for classification. The model first carries on the initial recognition to the input image, then automatically distinguishes the classification situation in the unit grid, and carries on the fine recognition and the classification according to the classification situation, and finally accurately identifies and locates the scene target. (2) based on the classification results of fine scene, the features of the image are extracted and matched. Using VS2008 as programming platform and C language, the image matching algorithm based on SIFT and Harris-Laplace is implemented. Firstly, the image multi-scale space is established, the Harris interest points are detected, and the feature points are screened by LoG operator, and the feature points are described by SIFT method. In the stage of searching for matching points, according to the scene type to which the feature points belong, the same scene of the image to be matched is searched. Thus, the matching search quantity is reduced and the matching accuracy and efficiency are improved. (3) the experimental results are analyzed by using the scene classification method and the image matching method assisted by depth learning. The experimental results show that the scene classification method based on depth learning can obtain high precision results in the classification of high resolution remote sensing image scenes, and the number of matching points can be effectively increased by combining with the image matching algorithm in this paper. Improve matching accuracy and efficiency.
【学位授予单位】：东华理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P23

【参考文献】