【作者】 王鹏；

【导师】 巫兆聪；

【作者基本信息】 武汉大学 ， 摄影测量与遥感， 2019， 硕士

【摘要】 高光谱图像可以提供数十甚至数百个覆盖范围从可见光到红外区域的高光谱分辨率图像,它包含了与物体材料的物理性质、化学性质和几何特性相关的空间与光谱分布的丰富信息,已经广泛应用于环境监测、大气与海洋遥感、行星探测等领域。近年来随着高光谱数据处理技术和三维重建技术的发展,许多研究者将物体三维几何信息与光谱信息结合,生成高光谱三维模型,高光谱三维模型对于需要精细光谱分析的三维应用很有意义,可以用于矿物分类和定量分析、文物研究、植物定量分析等,具有广阔的应用前景。目前以高光谱图像为基础进行三维重建的主流方法是逐波段三维重建,然后将所有波段的三维点云融合生成高光谱三维模型。但是高光谱图像的数据量很大,造成计算负担,而相邻光谱的高相关性造成数据的冗余,逐波段计算三维模型会造成计算浪费,也没有必要。那么数据降维就变得非常重要,现有高光谱数据降维技术主要面向遥感图像地物分类等应用,面向三维重建的降维研究很少,现有方法对高光谱三维重建的适用性也没有相关研究。本文重点研究面向三维重建的高光谱图像降维方法,主要的工作如下:(1)研究现有的高光谱图像波段选择方法的算法原理,现有的高光谱图像波段选择方法是从原始波段集中选择最优波段子集,波段子集代表原始图像的主要信息,同时冗余度较低。研究三种类别的波段选择方法,然后通过实验验证三种具有代表性的方法:基于子空间划分的方法、基于图表示的方法、基于密度聚类的方法,分析三种方法对于面向高光谱图像三维重建的适用性。(2)视觉词袋模型是计算机视觉图像处理中的常用方法,本文首次将视觉词袋模型应用于高光谱图像降维问题中,根据高光谱图像所有波段图像的特征点描述子集构造视觉词袋模型,根据视觉词袋模型计算波段间相似度,以此为基础进行波段选择,根据实验证明基于视觉词袋模型的波段选择方法稳定有效。(3)高光谱图像三维重建的特殊性是高光谱图像包含多个波段图像,不同波段能够反映目标的不同特征,特征点描述子是图像特征的描述,面向三维重建的波段选择要尽可能保留图像不同波段的特征。与传统降维方法相比,面向三维重建的图像降维要求波段子集在特征点描述子层面的高代表性和低冗余度。据此,本文提出两种面向三维重建的波段选择方法,基于SIFT特征的波段选择方法和基于ORB特征的波段选择方法,通过实验验证,本文两种方法的特征代表性均优于常用的方法,也能够反映更加完整精细的光谱信息,并且在相同条件下基于ORB特征的方法优于基于SFIT特征的方法。(4)根据面向三维重建的目的,本文提出基于特征代表性的评价方法,用波段子集特征点描述性平均差异化测度衡量波段子集的特征代表性;提出子区间代表性评价方法,将原始光谱范围划分为若干个区间,通过在每个子区间讨论波段冗余度和代表性,评价波段集表达原始光谱信息的完整性和精细程度。更多还原

【Abstract】 Hyperspectral images can provide tens or even hundreds of high spectral resolution band images,which cover the visible to infrared region,and contain rich information about the spatial spectral distribution associated with the physical,chemical,and geometric properties of the material of the object.It has been widely used in environmental monitoring,atmospheric and ocean remote sensing,military field and planetary exploration.In recent years,with the development of hyperspectral data processing technology and 3D reconstruction technology,many researchers combine object 3D geometric information with spectral information to generate hyperspectral 3D models.Hyperspectral 3D models are meaningful for 3D applications requiring fine spectral analysis.It can be used for mineral classification and quantitative analysis,cultural relics research,plant quantitative analysis,etc.,and has broad application prospects.At present,the mainstream method for 3D reconstruction based on hyperspectral imagery is band-by-band 3D reconstruction,and then the 3D point cloud of all bands is fused to generate a hyperspectral 3D model.However,the amount of data in hyperspectral images is large,which causes computational burden,and the high correlation of adjacent spectra causes redundancy of data.Computation of 3D models by band will result in computational waste and is not necessary.Then data dimensionality reduction becomes very important.The existing hyperspectral data dimensionality reduction technology is mainly applied to remote sensing image classification and other applications.The dimensionality reduction research for 3D reconstruction is rare,and the applicability of existing methods to hyperspectral 3D reconstruction is also no related studies.This paper focuses on the method of dimensionality reduction of hyperspectral imagery for 3D reconstruction.The main work is as follows:(1)Study the algorithm principle of the existing hyperspectral image band selection method.The existing hyperspectral image band selection method is to select the band subset from the original band,and the band subset represents the main information of the original image,and the redundancy.Lower.Three kinds of band selection methods are studied,and then three representative methods are verified by experiments: subspace division method,graph representation method,density clustering based method,and three methods for hyperspectral imagery.The applicability of 3D reconstruction.(2)Visual word bag model is a common method in computer vision image processing.For the first time,the visual word bag model is applied to the hyperspectral image dimensionality reduction problem.The visual word is constructed according to the feature point description subset of all band images of hyperspectral image.The bag model calculates the similarity between the bands according to the visual word bag model,and based on this,the band selection is performed.According to the experiment,the band selection method based on the visual tape model is stable and effective.(3)According to the characteristics of three-dimensional reconstruction of hyperspectral images,hyperspectral images contain multiple band images.Different bands can reflect different features of the target.The feature point descriptor is the description of the image features.The feature point descriptor is the basis of image matching and 3D reconstruction.Compared with the traditional dimensionality reduction method,the image dimensionality reduction for the 3D reconstruction requires the band.Set high representation and low redundancy at the feature point description level.Based on this,two band selection methods for 3D reconstruction are proposed in this paper.The band selection method based on SIFT feature and the band selection method based on ORB feature are verified by experiments.The feature representation of the two methods is better than the commonly used method.It is also able to reflect more complete and detailed spectral information,and the method based on ORB features under the same conditions is superior to the method based on SFIT features.(4)According to the purpose of 3D reconstruction,this paper proposes an evaluation method based on feature representation,and uses the descriptive mean difference measure of the feature subset of the band subset to measure the representativeness of the band subset.The spectral representative evaluation method is proposed.The spectral range is divided into several intervals.By discussing the band redundancy and representativeness in each sub-interval,the integrity and fineness of the original spectral information expressed by the band set are evaluated.更多还原