导　　师： 徐智勇

学科专业： 0810

授予学位： 硕士

作　　者： ;

机构地区： 中国科学院研究生院

摘　　要： 高光谱成像技术能够同时探测目标的二维几何空间信息和一维光谱信息。利用高光谱图像所携带的精细光谱信息进行目标探测与识别是当前目标探测与识别和遥感信息处理领域的一个热点研究问题。高光谱图像目标探测与识别系统通常包括图像预处理、数据降维、混合像元分解以及目标探测与识别等环节。高光谱目标探测与识别主要利用高光谱图像含有的精细光谱信息，通过光谱分析实现目标探测与识别。本篇论文从高光谱图像目标探测与识别中面临的难点以及实际应用出发，从光谱分析的角度入手，对基于高光谱图像目标探测与识别技术的关键环节作分析与研究，为高光谱图像的目标探测与识别技术的实际应用提供了理论前提和基础。论文的主要研究工作和成果包括以下四个方面： 1.针对实验室获取的未配准的原始高光谱图像，结合多分辨率思想采用基于频域的傅里叶梅林变换（Fourier-Mellin transform, FMT）方法实现了原始高光谱图像的配准，取得了良好效果。 2.针对实验室原有的端元数目估计方法容易丢失掉高光谱图像中的弱小目标这一问题。论文采用了基于噪声调整的主成分分析（noise-adjustedprincipal components, NAPC）方法对端元数目估计方法作了改进，并取得一定效果。 3.对基于顶点成分分析（vertex component analysis, VCA）的端元提取和混合像元分解结果作了定量评价。验证了端元提取和混合像元分解算法的有效性。 4.对高光谱图像目标探测与识别技术典型情形和常用数学模型下的经典算法作了分析研究，并结合信号检测理论对探测结果作了分析和评价。 Hyperspectral imaging technology can detect both two-dimensional geometricspatial information and one-dimensional spectral information of the target. Ashyperspectral images carried by fine spectral information, target detection andrecognition based on hyperspectral is currently a hot research problem of targetdetection and recognition and remote sensing information processing.Hyperspectral image target detection and recognition system typically includesimage preprocessing, data reduction, unmixing, target detection and recognition et al.Hyperspectral target detection and recognition can be achieved by fine spectral analysisof hyperspectral images.This paper began with the difficult and application of hyperspectral image targetdetection and recognition, discussed and analyzed some key steps of hyperspectralimage target detection and recongnition by spectral analysis and provided a theoreticalbasis and premise for hyperspectral image target detection and recognition applications. The main research work and achievements of the paper include the following four parts: 1. We registrated the original laboratory hyperspectral images by combined themultiresolution analysis /(MRA/) with Fourier Mellin transform /(FMT/) method,and achieved good results. 2. As the original method of endmember estimation may lose small target, weuses the method based on noise-adjusted principal components analysis /(NAPC/)to estimate the number of endmember and improved the results. 3. We evaluated the result of endmember extraction based on vertex componentanalysis /(VCA/) and unmixing, so that verified the validity of endmemberextraction and unmixing algorithm. 4. We discussed and analyzed some classical algorithms of hyperspectral imagetarget detection and recognition based on typical application and mathematicalmodel. We analyzed and evaluated the results of these algorithms by signaldetection theory.

关 键 词： 高光谱 图像配准 端元估计 混合像元分解 目标探测 目标识别

分 类 号： [TP391.41]

领　　域： [自动化与计算机技术] [自动化与计算机技术]