基于投影寻踪的高光谱图像降维算法研究

【作者】 王秀朋；

【导师】 张洪才； 赵永强；

【作者基本信息】 西北工业大学 ， 控制理论与控制工程， 2006， 硕士

【摘要】 高光谱图像数据具有波段多、光谱分辨率高、波段宽度窄、数据量庞大等特点,因而对地物的描述与测度更加细致,但由此也带来了光谱维数高,处理难度大,计算量大,实时性差的问题。显然,如何有效地“降维”,将数据投影到低维空间进行分析是解决这一问题的有效途径。 本文对高光谱图像的降维方法进行了研究,并根据实验需要搭建了基于液晶可调谐滤波器的光谱图像采集系统。主要工作如下: 1.为满足实验需要,搭建了一个基于液晶可调谐滤波器(LCTF)的光谱图像采集系统,在不同的场景中进行了多次实验,结果表明该系统所得光谱图像可以满足研究需要,为具体算法研究提供了必备条件。 2.对常用的高光谱图像降维方法进行了研究,给出了常用的波段选择方法、自适应子空间划分方法、主成分分析方法的仿真分析。分析表明利用波段选择和划分数据源的降维方法,依靠单一指标进行波段选择和数据源划分容易损失有效信息,而主成分分析方法虽然能在保留大部分信息前提下实现有效降维,但其通常只能反映数据的全局或主要特征,而不能反映局部或细节特征。 3.对投影寻踪方法在高光谱图像降维中的应用进行了研究,分别给出了以偏度、峰度、Jones矩指标等为投影指标,以遗传算法为优化算法的仿真分析,仿真结果表明投影寻踪算法能突出数据的局部特征,但实时性较差。 4.提出了一种基于自适应子空间划分的序贯投影寻踪方法,利用自适应子空间划分优化高光谱图像的分组方式,再应用序贯投影寻踪方法对其实现降维。仿真表明此方法不但能降低计算时间,进一步突出数据的局部特征,而且能有效地利用高光谱图像相关性。更多还原

【Abstract】 Compared with multispectral images, the hyperspectral images have more bands, higher spectral resolution, narrower bandwidth, thus can describe and measure the objects in more detail. But for practical purpose, the dimensionality is too high and the computation burden is too heavy. In this thesis, the dimensionality reduction methods of hyperspectral images are studied in detail. The main contributions are as follows:1. A spectral imaging system, which is based on Liquid-Crystal Tunable Filters (LCTF), is constructed. Lots of experiments with different scene are carried out. The experiment results show that the constructed spectral imaging system can capture effective spectral images, which provides the basis for concrete algorithm research.2. Several dimensionality reduction methods of hyperspectral images are analyzed, and the experiment analyses are given included conventional methods of band selection, adaptive subspace decomposition and Principle Component Analysis (PCA). The experiment results show that the band selection methods may lose lots of interesting information. Using PCA to reduce the dimension of the hyperspectral images can remain most of information. It can in general reflect the global features but may ignore the local feature.3. The application of Projection Pursuit (PP) to dimension reduction of the hyperspectral images is studied. The experiments which use skewness, kurtosis, Chiang’s product and Jones moment as the projection index and Genetic Algorithm (GA) as the optimization algorithm are given, and the results show that PP can make the local feature of the spectral data prominent, but it can not realize the real-time processing.4. A new method to reduce the dimension of the hyperspectral images is presented, which combines the adaptive subspace decomposition method and Sequential Projection .Pursuit (SPP). The experiment results show that this method can increase the computation speed, and show the local feature further than the conventional projection pursuit, as well as make use of the relativity of the hyperspectral images.更多还原