遥感图像建筑物检测的关键技术研究

发布时间：2018-08-27 07:59

【摘要】：从遥感图像中提取地面目标信息是当今重要的研究课题之一,已经在军事、民用等领域得到了广泛应用。地面目标中有很大一部分都是建筑物,尤其是在城市区域内,这个比例能够达到80%左右。同时,研究建筑物检测对于其他目标的检测也有非常重要的参照价值。因此,遥感图像建筑物检测非常具有研究意义和应用价值。在基础理论研究方面,论文介绍了遥感图像目标识别的主要步骤、目标识别层次划分和分辨率、目标识别中的主要特征,特征的变换不变性以及分类器。重点介绍了特征提取和特征分类,给出了常用的特征提取方法和特征分类方法,同时介绍了距离测度学习及相关算法。在建筑物检测方法的研究方面,论文指出了遥感图像建筑物检测算法现存的问题:其一是提取特征时忽略建筑物形状,这不利于准确描述建筑物特征;其二是进行建筑物分类时忽略建筑物间的相似性和建筑物与背景之间的差异性,这将导致分类准确率降低。针对上述问题,论文将基于形状的特征和区分性学习结合(SSF+DFC)进行建筑物检测。其中,基于形状的特征(SSF)只在同质区域内收集定向梯度数据,并通过不断迭代进行特征更新,最终可以很好地描述建筑物的形状和结构。区分性的特征分类(DFC)通过学习到的距离测度矩阵M得到样本间的相似性或差异性,进而能获得更准确的分类结果。通过在Quick Bird 2数据集上进行实验,并和其他方法进行比较,我们得出将基于形状的特征和区分性学习结合进行建筑物检测能够检测出外部轮廓更准确、数量更多的建筑物。原因是基于形状的特征只在同质区域内收集光谱特征和定向梯度信息,能够很好地捕获建筑物的形状和结构;基于区分性的特征分类,可以获得建筑物之间的相似性和建筑物与背景之间的差异性来提高建筑物检测的精度。
[Abstract]:The extraction of ground target information from remote sensing images is one of the most important research topics and has been widely used in military, civil and other fields. A large part of the ground targets are buildings, especially in urban areas, which can reach about 80%. At the same time, the study of building detection also has very important reference value for the detection of other targets. Therefore, the remote sensing image building detection has the research significance and the application value. In the aspect of basic theory research, the paper introduces the main steps of target recognition in remote sensing image, the classification and resolution of target recognition, the main features of target recognition, the invariance of feature transformation and classifier. This paper mainly introduces feature extraction and feature classification, gives common feature extraction methods and feature classification methods, and introduces distance measure learning and related algorithms. In the research of building detection methods, the paper points out the existing problems of building detection algorithm in remote sensing image: first, ignoring the building shape when extracting features, which is not conducive to accurate description of building features; The other is to ignore the similarity between buildings and the differences between buildings and backgrounds, which will lead to the reduction of classification accuracy. In order to solve the above problems, building detection based on shape feature and distinguishing learning is combined with (SSF DFC). Among them, the shape based feature (SSF) only collects the directional gradient data in the homogeneous region, and updates the feature through iteration, which can describe the shape and structure of the building well. The discriminative feature classification (DFC) can obtain the similarity or difference between samples by learning the distance measure matrix M, and then obtain more accurate classification results. Through experiments on Quick Bird 2 data set and compared with other methods, we conclude that building detection based on shape based feature and distinguishing learning can detect more accurate and more buildings with external contour. The reason is that shape-based features collect spectral features and directional gradient information only in homogeneous regions, which can capture the shapes and structures of buildings well; and classification based on distinguishing features, The similarity between buildings and the differences between buildings and backgrounds can be obtained to improve the accuracy of building detection.
【学位授予单位】：中国地质大学(北京)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP751

【相似文献】