高分辨率遥感图像道路检测与中心线提取算法研究

CASIA OpenIR > 毕业生 > 博士学位论文

	高分辨率遥感图像道路检测与中心线提取算法研究
	程光亮
	2017-05
学位类型	工学博士
英文摘要	道路作为高分辨率遥感图像中的重要目标，在人们的生产和生活中发挥着十分重要的作用。由于在路网更新、车辆导航和智能交通等应用市场存在巨大潜力，高分辨率遥感图像道路提取得到了国内外研究人员的广泛关注。近年来，有关高分辨率遥感图像道路提取的研究取得了较大进展，但其性能仍无法达到应用要求。高分辨率遥感图像道路提取主要面临如下几方面的难点：(1）树木、车辆、高楼阴影等多种形式的遮挡并存，导致现有道路提取方法的鲁棒性不高；(2) "同物异谱"现象和"同谱异物" 现象导致现有方法漏检率和虚警率较高。另外，现有的道路中心线提取方法难以获得准确、平滑、完整且单像素宽度的道路中心线。本文围绕高分辨率遥感图像道路区域及其中心线提取开展相关研究工作。针对现有方法的不足，本文提出了相应的解决方案。本文的主要研究内容和贡献如下: 1. 提出了两种基于图割的高分辨率遥感图像道路区域提取方法。a）针对现有方法无法获得光滑连贯的道路区域提取结果这一缺陷，引入了一种基于概率传播图割方法的道路提取方法。该方法可以有效地将光谱信息和空间信息进行结合。同时，构建了一种基于几何先验的非道路目标剔除方法，可以有效地剔除所得结果中的非道路目标。 b）为了克服传统图割方法在图像分割任务中存在的过分割和欠分割问题，引入了一种基于多特征自适应图割方法的道路提取方法。该方法根据图像中的边缘信息自适应地确定图割方法中区域项和边界项的相对权重，可以获得比传统方法更优的道路提取结果。对比实验验证了所提两种方法的有效性。 2. 提出了一种基于多尺度分割与张量投票的道路中心线提取方法。该方法的核心思想是引入基于融合的多尺度协同表示与图割技术来获得精确且连续的道路区域。在此基础上，通过联合应用张量投票方法、非最大值抑制方法和道路中心线补全方法，可获得完整、无毛刺的道路中心线。在视觉效果和数值结果两方面验证了所提方法的有效性。 3. 提出了一种基于半监督学习框架的道路中心线提取方法。针对高分辨率遥感图像道路提取任务中带标签样本不易获得这一问题，本文引入了一种基于半监督学习框架的道路提取方法。该方法通过发掘带标签样本与无标签样本之间潜在的结构关系，可以有效地将标签向无标签样本传播。在此基础上，引入了一种基于多尺度滤波和多方向非最大值抑制的方法，可获得准确且完整的道路中心线。实验结果表明，本文所提方法优于目前主流道路提取方法。 4. 提出了两种基于深度卷积神经网络的自动道路提取方法。1）针对现有的道路检测和中心线提取方法中存在的问题，提出了一种基于端到端级联卷积神经网络的道路区域提取和中心线提取方法。该方法将道路检测网络和中心线提取网络统一到一个学习框架中，并可以实现端到端地训练。在测试阶段，该级联网络可以同时获得道路区域提取结果和平滑的道路中心线。 2）根据道路所特有的方向特性和结构特性，提出了一种基于残差网络和方向信息的道路提取方法。所提方法包括两部分：回归网络和分割网络。前者对图像中每条道路方向进行回归，并将所学特征传递到分割网络中；后者通过融合回归网络所学特征及其自身网络所学特征对图像中的区域进行判断，从而可得到平滑连贯的道路区域提取结果。同时，本文将回归网络和分割网络统一到同一个学习框架之中，并可实现端到端地训练。对比实验验证了上述所提两种方法的有效性。 ; The road, as an important object in high-resolution remote sensing images, plays an important role in people's daily life. Due to its huge potentials in road network update, vehicle navigation and intelligent transportation, etc., road extraction from high-resolution remote sensing images has been paid more and more attention by researchers. In recent years, the research on road extraction from high-resolution remote sensing images has made great progress, while its performance fails to meet the application requirements. The main difficulties of road extraction from high-resolution remote sensing images lie in the following aspects: (1) Some road areas are under occlusions of trees and vehicles as well as shadows of tall buildings, which makes the existing road extraction methods show little robustness against the occlusions. (2) There exists phenomena of ``the same spectrum in different objects" and ``the same object with different spectra", which results in the existing methods with high miss rate and false alarm rate. In addition, the existing road centerline extraction methods can not get accurate, smooth, complete and single-pixel width road centerline. This thesis focuses on the research of road extraction and centerline extraction from high-resolution remote sensing images. According to the shortcomings of the existing approaches, some corresponding solutions are put forward in this thesis. The main contents and contributions are listed as follows: 1. Two graph cuts based methods on road area extraction from high-resolution remote sensing images are proposed: a) To obtain smooth and consistent road area extraction result, an urban road extraction approach via graph cuts based probability propagation is introduced, which can incorporate spectral information and spatial information simultaneously. Meanwhile, to effectively eliminate non-road objects from the obtained road area result, a non-road elimination algorithm based on road-geometrical prior is considered. b) To overcome the problem of over-segmentation and under-segmentation in the image segmentation task, a road area extraction approach via adaptive graph cuts with multiple features is introduced. This approach adaptively determines the relative weight of the region term and the boundary term in the graph cuts based method according to the edge information in the image, which can achieve better road area extraction result than traditional graph cuts based method. Contrast experiments validate the effectiveness of the two proposed methods. 2. A road centerline extraction method based on multiscale segmentation and tensor voting is proposed. The core idea of this approach is to introduce fused multiscale collaborative representation (FMCR) and graph cuts based technique to achieve accurate and continuous road regions. Based on the above stage, tensor voting (TV) algorithm, non-maximum suppression (NMS) based algorithm and centerline connection algorithm are jointly applied to achieve complete and spur-free road centerline. It is verified to achieve better performance than the state-of-the-art methods in terms of visual and quantitative aspects. 3. A new road centerline extraction approach based on semi-supervised learning is proposed. To solve the problem of limited labeled samples in road extraction task, a semi-supervised learning framework is introduced to tackle road area extraction task. By exploring the intrinsic structures between labeled samples and unlabeled samples, this framework can propagate the labels of labeled samples to those unlabeled ones. Based on the above stage, an algorithm based on multiscale filtering and multidirection non-maximum suppression is constructed to obtain smooth and complete road centerline. Experimental results show that the proposed method is superior to the state-of-the-art road centerline extraction approaches. 4. Two automatic road extraction approaches based on deep convolutional neural network are proposed: a) To overcome the defects in the existing road detection and centerline extraction methods, a new automatic road detection and centerline extraction algorithm is introduced, which is based on cascaded end-to-end convolutional neural network. This approach incorporates road detection network and centerline extraction network into a unified learning framework, which can be trained in an end-to-end manner. During the test stage, the cascaded network can simultaneously obtain road area extraction result and smooth road centerline extraction result. b) According to the directional characteristics and structural characteristics of road, a new road extraction approach based on residual network and directional information is constructed. The proposed network consists of two parts: regression network and segmentation network. The former regresses the direction of each road in image, and transmits its characteristic to the segmentation network. The latter can obtain smooth and coherent road area extraction result by incorporating characteristics of regression network and segmentation network. It should be noted that the proposed approach unifies the regression network and the segmentation network into a learning framework and can be trained via end-to-end learning strategy. Contrast experiments verify the effectiveness of the above two proposed methods.
关键词	高分辨率遥感图像道路检测中心线提取多尺度分割深度卷积神经网络
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/14826
专题	毕业生_博士学位论文
作者单位	中国科学院自动化研究所空天信息研究中心
推荐引用方式 GB/T 7714	程光亮. 高分辨率遥感图像道路检测与中心线提取算法研究[D]. 北京. 中国科学院研究生院,2017.

条目包含的文件
文件名称/大小	文献类型	版本类型	开放类型	使用许可
学位论文最终版20170527_subm（23132KB）	学位论文		限制开放	CC BY-NC-SA