【作者】 刘德连；

【导师】 张建奇；

【作者基本信息】 西安电子科技大学 ， 光学工程， 2008， 博士

【摘要】 近年来,随着遥感技术的发展,遥感图像的目标检测引起了广泛的关注。研究遥感图像的目标检测对于资源勘探,自然灾害评估,军事目标检测和识别等都具有重要的意义。由于受到地表温度以及大气衰减等因素的影响,遥感系统所受到的干扰比普通的地面监控系统更强,干扰有时会大大超过所要检测目标的强度,因此,实现遥感图像的目标检测更具挑战性。本文的工作主要是研究遥感图像的目标检测方法,包括无目标先验信息的异常检测和已知目标先验信息的目标检测,开展了如下工作:1.分析了遥感图像的背景分布特性,提出了一种改进的背景描述方法。首先分析了单波段遥感图像和高光谱图像的统计分布特性,并比较了高斯模型与实际遥感图像背景特性的偏差,在此基础上,提出采用背景高斯化的方法来描述背景,得到了一种基于背景高斯化的单波段遥感图像异常检测方法和一种基于背景高斯化的高光谱图像异常检测方法。2.分析了检测算法中常用的目标模型,并将目标光谱模型应用于序列图像的目标检测。首先分析了单波段遥感图像目标检测算法中常用的目标模型,然后讨论了高光谱图像目标检测算法中的多元高斯模型、奇异值分解模型和自回归模型,在此基础上,将光谱模型应用于序列图像的时域目标检测,得到了一种新的基于时域剖面分析的运动弱小目标检测算法。3.分析了遥感图像的异常检测方法。首先介绍了基于亮度聚类的异常检测算法的优缺点,以此为基础,提出利用背景分解的思想来描述复杂背景的影响,并采用二维纹理分割来实现复杂背景分解,得到了一种基于二维纹理分割的单波段遥感图像的异常检测算法。接着分析了高光谱图像的特点和高光谱图像异常检测的经典算法—RX算法,将背景分解的思想应用于高光谱图像的异常检测,并对二维纹理分割方法进行扩展,获得了一种基于三维纹理分割的高光谱图像异常检测算法。4.分析了遥感图像的目标检测方法,根据目标大小与成像系统分辨率的关系,分别讨论了高光谱图像的全像素目标检测方法和亚像素目标检测方法,重点分析了Kelly检测法,自适应余弦估计法和自适应子空间匹配法三种高光谱图像亚像素目标检测方法。将背景分解的思想应用于高光谱图像的目标检测,给出了一种基于背景分解的全像素目标检测法和三种基于背景分解的亚像素目标检测方法。更多还原

【Abstract】 Along with the development of remote sensing, target detection in remote sensing images attracts more and more attentions in recent years. Target detection in remote sensing images is of great importance in a variety of applications such as resources survey, disaster evaluation, and military target detection etc. For the affection of land temperature, atmosphere attenuation and remote sensing systems, the interferences in remote sensing are more severe than common surveillance systems on land, which are the key challenge of target detection in remote sensing images.In this article we will investigate target detection in remote sensing images including anomaly detection and target detection. The following works are carried out.1. Background models in target detection algorithms and their disadavantages are analyzed, based on which a new model is proposed. First, the distribution of gray scale images and hyperspectral images is discussed. And the derivation of Gaussian model and real background distribution is also evaluated. Through analyzing the limitation of Gussian model, we present a new anomaly detection algorithm by transforming background into Gaussian space, and the Gaussian transformation based model is also extended to the hyperspectal images.2. Target models in target detection algorithms are reviewed and an application of a target spectral model is addressed. The target models in both gray scale images and hyperspectral images are discussed especially hyperspectal target models such as multidimensional Gaussian model, singular value decomposition model, and autoregression model. By comparing the similarity of target models in hyperspectral images and temporal profiles of sequence images, we apply a target spectral model to temporal based target detection in infrared image sequence and present a new temporal profile based small moving target detection algorithm in infrared image sequence. Experiment with real infrared image sequence has proved the validity of the new approach.3. Anomaly detection algorithms are discussed. First, a detection algorithm based on gray scale cluster is analyzed, and a background segmentation model is proposed to avoid the influence of complex background, with which a new algorithm is developed by introducing 2D Markov model to segment complex background into homogenous regions. Then the characteristic of hyperspectral image is analyzed as well as the limitation of the RX algorithm. The background segmentation model is also employed. We have extended the 2D texture segmentation model to 3D to model hyperspectral images, and obtained a new texture segmentation based anomaly detection algorithm in hyperspectral images.4. Target detection algorithms are examined. According to the relationship between the resolution of imaging systems and targets size, several full-pixel detection algorithms and sub-pixel algorithms are discussed especially the Kelly detection algorithm, the adaptive cosine estimator algorithm, and the adaptive background subspace algorithm. Employing the idea of the background segmentation model, a new background-segmentation-based full-pixel target detection algorithm is presented. And three background-segmentation-based sub-pixel target detection algorithms are also presented including background segmentation based Kelly algorithm, background segmentation based adaptive cosine estimator algorithm, and background segmentation based adaptive background subspace algorithm.更多还原