【作者】 陶超；

【导师】 田金文； 谭毅华；

【作者基本信息】 华中科技大学 ， 控制科学与工程， 2012， 博士

【摘要】 在高分辨率遥感影像中,城区与建筑物作为两类重要的地物目标,其自动提取和解译在城市建设,GIS系统更新,数字化城市以及军事侦察等多个领域都有着重要应用。为此,本文在从高分辨率遥感影像上进行城区和建筑物提取方面进行了有意义的探索和尝试。概括而言,论文主要进行了如下四个方面的研究工作：首先,对高分辨率遥感影像中常用的特征分析方法进行了较为全面的综述研究。根据特征类型不同将其抽象为三大类：光谱特征、纹理特征和局部关键点特征,然后阐述了每一类特征分析方法的基本思想,并对其中具有代表性的特征提取方法进行了较为深入的分析和讨论,在目标特征提取方面为后续的研究工作奠定了良好的基础。其次,提出一种有监督的多特征融合的城区检测算法。该方法将最优的特征融合问题转化为多个特征核函数的组合问题,并采用多核学习的方法在进行多特征融合的同时,完成城区SVM分类器的学习。与传统的基于单种图像特征的城区检测方法相比,该算法通过有效地融合多种图像特征,大大提高了城区检测算法性能。然后,传统的城区检测算法通常一次只能处理一幅输入影像,且为提高识别精度,需要提供大量人工标注的数据样本,过程十分繁琐。面对海量级的遥感影像数据,难以满足当前各种应用的自动化和实时化需求。针对这一问题,提出一种无监督的基于多幅高分辨率遥感影像的城区协同检测算法。首先对输入影像集中的各个影像进行独立处理：利用改进的Harris算子检测影像中的角点,然后根据影像中的角点分布提取其中的候选城区区域。在得到各个影像中的候选城区区域后,把它们合并起来得到一个候选城区区域集合,然后对它们进行纹理特征建模,最后利用谱聚类和graph cut算法从中筛选出真正的城区区域。与现有的无监督城区检测方法相比,由于该算法充分利用了多幅影像提供的信息,因此检测精度得到提高。与已有的有监督城区检测方法相比,该算法检测精度与它们相当,但由于无需任何训练样本,因此在效率上得到提高。最后,在人类视觉认知理论的启发下,通过仿效人类认知复杂对象时通常采取的由易到难的策略,提出一种基于视觉认知启发的城区建筑物分级提取算法。该方法将面向对象的思想融入到基于邻域总变分的建筑物分割方法中,并通过分析分割后不同类型建筑物提取的难易程度,提出一种多特征融合的建筑物对象分级提取策略：首先通过形状分析检测一部分分割完整的矩形建筑物目标,然后采用新提出的多方向形态学道路滤波算法将建筑物与邻近光谱相似的道路目标分离,确保每一个候选建筑物目标都是独立的对象,最后利用初提取的建筑物对象和已剔除的非建筑物对象作为样本建立概率模型,根据贝叶斯准则进行复杂建筑物后提取。与传统的建筑物检测算法相比,该算法的优势在于能够直接从测试影像中提取训练样本用于建筑物目标模型学习,由于训练样本和测试对象在同一尺度和光照条件下获得(例如,它们均来自于同一幅影像),这充分保证了模型的置信度和稳健性；而且样本集的建立过程不需要人工参与,这也满足了算法自动化的需求。实验表明：该方法可以检测同一幅影像中具有不同形状结构和光谱特性的建筑物目标,准确率高、鲁棒性好,具有较高的实际应用价值。更多还原

【Abstract】 In the past few years, urban area and building detection from high-resolution remotely sensed image have become crucial for several applications. The main one is to update the geographic information databases, which are critical sources of information in diverse fields such as cartography, city planning and change detection. To this end, this dissertation is trying to propose several algorithms for urban and building detection from high resolution remotely sensed images. Concretely, main contents of this dissertation include the following four parts:Firstly, we briefly review previous works on feature extraction approaches, and divide them into three categories:spectral feature extraction, texture feature extraction and local feature extraction. Afterwards, we descible in detail the basic concept and principle of feature extraction approaches, which are commonly used in high-resolution image interpretation.Secondly, a supervised urban detection approach based on mutil-feature fusion model is proposed. In this method, we treat the problem of mutil-feature fusion as estimating a weighed linear combination of mutilple feature kernel functions. And the weight for each feature kernel function is automatically estimated in a multi-kernel support vector machine (SVM) learning framework during the training stage. In the classification stage, we first divide the test image into several non-overlapping image blocks, and then apply the SVM classifier to determine whether each image block belongs to urban or not. Compared to traditional approaches using only texture information for urban detection, experimental results demonstrate that fusing multiple features can help improving urban detection accuracy rate.Thirdly, given a set of high resolution satellite images covering different scenes, an unsupervised approach to simultaneously detect possible urban regions from them is proporsed. The motivation behind is that:the frequently recurring appearance patterns or repeated textures corresponding to common objects of interest (e.g. urban area) in the input image dataset can help us discriminate urban area from others. With this inspiration, our method consists of two steps. First, we extract a large set of local feature point by Harris corner detector. In order to achieve a reliable extraction of corners from urban areas, we further propose two criterions to validate and filter them. Afterwards, we incorporate the extracted corners into a likelihood function to locate candidate regions in each input images. Given a set of candidate urban regions, in the second stage, we formulize the urban detection process as an unsupervised classification problem. The candidate regions are modeled through their histogram representation of Gabor texture features, and the classification problem is solved by spectrum clustering and graph cuts. The experimental results show that the proposed approach is capable of and efficient at simultaneously detecting urban regions from multiple high-resolution satellite images, and performs comparable or even better in comparison with the state-of-the-art supervised method.In high-resolution satellite image, buildings can be considered as clustered objects belonging to the same category. Human perception of such objects consists of an initial identification of simple instances followed by a recognition of more complicated ones by deduction. Inspired by this theory, a novel hierarchical building extraction framework is proposed to simulate the process, which includes three major components. Firstly, a total variation based segmentation algorithm is presented to decompose the given image into object-level elements. Then, shape analysis is applied to extract some common and easily identified rectangular buildings. To ensure each candidate of building target is isolated, a multidirectional morphological road-filtering algorithm is designed to separate the buildings from their neighboring roads with similar spectrum. Finally, the detection of buildings with complex structures is formulated as a deduction problem based on preceding extracted information in terms of maximum a posteriori (MAP) estimation, and a Bayesian based approach is put forward to deal with it. Comparing to the conventional way of detecting objects through the information learned from previously collected training samples, our method has two advantages. First, our approach can learn building models directly from the original images. Therefore, it is highly automatic, for no manual aid is required in the collection of training data. More importantly, since the training data are collected from the identical scale and illumination conditions (e.g., in the same image), our model is more discriminating. This enables that the proposed framework has the ability to detect building with complex structures and varying spectral response, independent of pre-defined and limited building models.更多还原