基于高光谱成像的目标检测算法研究

发布时间：2018-02-04 11:16

本文关键词： 高光谱图像目标检测高阶统计量稀疏算法　出处：《西安电子科技大学》2014年硕士论文　论文类型：学位论文

【摘要】：近几年，由于科学技术的迅猛发展，人们对“感知”提出了更高要求并得到了有效的延伸，同时，对事物的认识能力也得到了不断的提高。过去几十年正是成像光谱发展突飞猛进的阶段，高光谱图像的分析和处理成为当前国内外遥感图像处理领域的研究热点之一。高光谱图像的突出特点是光谱分辨率高，可获得观测对象的几十个或几百个光谱波段的图像信息,而成像光谱系统获得的连续波段宽度一般都小于10nm。高光谱图像是一种三维数据，，成像光谱仪为每个像素点提供一条近似连续的光谱曲线，而所有像素的相同波段对应一个二维图像。高光谱遥感图像目标检测技术是高光谱遥感理论以及实践应用的核心环节。所谓高光谱图像目标检测，即利用已知的目标光谱信息在高光谱图像中对感兴趣的目标进行检测、确认的技术。高光谱图像目标检测技术在军事和民用领域中都有重要的应用价值。在军事领域可用于对飞机、坦克等军事目标进行检测、定位，也可对伪装的军事目标进行检测。在民用领域可应用于公共安全、环境监控等领域。本文在深入研究经典高光谱图像目标检测方法的基础上，提出了两个新的高光谱目标检测框架。由于目前存在的高光谱图像目标检测算法，大多是基于统计模型的检测方法，利用了二阶统计量进行目标检测。然而，现实中的目标往往服从的是非高斯分布。根据ICA的理论基础，针对非高斯分布目标的检测问题应使用高阶统计量进行检测。本文中提出两种采用高阶统计量的检测方法，多种目标材料检测器（MultipleMaterials Detector，MMD）和基于拟牛顿法多种目标材料检测器（Quasi-Newtonbased Multiple Materials Detector，QNMMD）。文章中从理论和实验结果均说明，相对于现有的基于二阶统计量的检测方法，基于高阶统计量的检测方法有更好的检测效果。在本文中，利用高光谱图像的稀疏模型，提出了两种检测方法。第一种是基于凸松弛法高光谱图像目标探测器（Convex Relaxation Based Target Detector,CRBTD）。这个算法中的创新点在于提出了一个连续的凸函数近似l0范数。利用这个方法，可以将很难求解的NP-hard优化问题转化为容易求解的凸优化问题，并且可以找到更准确的稀疏解。在实验中，相比于目前存在的基于稀疏模型的高光谱目标检测算法，CRBTD具有更好的检测结果。第二种提出的算法是，基于k-mean聚类重建光谱库的高光谱图像目标检测算法。在此算法中，通过对高光谱图像进行k-mean聚类、目标光谱剔除并整合的处理，实现了光谱库的自动构造。在真实高光谱数据的实验结果中可以看到，基于自动光谱库构造的稀疏检测算法，在检测效果上优于传统的统计算法。
[Abstract]:In recent years, due to the rapid development of science and technology, people have put forward higher requirements for "perception" and have been effectively extended at the same time. The ability to understand things has also been continuously improved. The past few decades is the stage of the rapid development of imaging spectrum. The analysis and processing of hyperspectral image has become one of the research hotspots in the field of remote sensing image processing at home and abroad. The outstanding feature of hyperspectral image is high spectral resolution. The image information of dozens or hundreds of spectral bands can be obtained, and the width of continuous band obtained by imaging spectral system is generally less than 10nm.Hyperspectral image is a kind of three-dimensional data. The imaging spectrometer provides an approximate continuous spectral curve for each pixel, and the same band of all pixels corresponds to a two-dimensional image. The object detection technology of hyperspectral remote sensing image is the core link of hyperspectral remote sensing theory and practical application, so called hyperspectral image target detection. That is, using the known target spectral information to detect the interested targets in hyperspectral images. Hyperspectral image target detection technology has important application value in both military and civil fields. It can be used to detect and locate military targets such as aircraft tanks and so on in the military field. It can also detect camouflaged military targets. It can be used in public safety, environmental monitoring and other fields. Based on the study of classical hyperspectral image target detection methods, two new hyperspectral target detection frameworks are proposed in this paper. Because of the existing hyperspectral image target detection algorithms, mostly based on the statistical model of detection methods, using second-order statistics for target detection. In reality, goals tend to conform to the distribution of non-#china_person0#. According to the theoretical basis of ICA. High order statistics should be used to detect non-#china_person0# distributed targets. In this paper, two detection methods using high order statistics are proposed. Multiple Materials Detector with multiple target material detectors. MMD) and quasi-Newton-based Multiple Materials Detector based on quasi Newton method. In this paper, the theoretical and experimental results show that the detection method based on higher order statistics is more effective than the existing detection methods based on second-order statistics. In this paper, the sparse model of hyperspectral images is used. Two detection methods are proposed. The first is based on convex relaxation hyperspectral image target detector (. Convex Relaxation Based Target Detector. The innovation of this algorithm is to propose a continuous convex function approximating l _ 0 norm. The NP-hard optimization problem which is difficult to solve can be transformed into a convex optimization problem which is easy to solve, and a more accurate sparse solution can be found. Compared with the existing hyperspectral target detection algorithm based on sparse model, CRBTD has better detection results. The target detection algorithm of hyperspectral image based on k-mean clustering reconstructing spectral database. In this algorithm, the target spectrum is eliminated and integrated by k-mean clustering of hyperspectral image. The experimental results of the real hyperspectral data show that the sparse detection algorithm based on the automatic spectral database is better than the traditional statistical algorithm.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】