【作者】 王宏；

【导师】 谷延锋；

【作者基本信息】 哈尔滨工业大学 ， 信息与通信工程， 2014， 硕士

【摘要】 高光谱成像实现了光谱维度上的细致观测，光谱分辨率得到了很大提高，这就为地物辨别提供了更丰富的分辨信息。然而，对地物分类而言，更高的光谱分辨率并不意味着更有效的分类结果，分类器的观测尺度影响着类内和类间样本的相似性度量，而传统分类器的观测尺度不能自动适应数据的变化。因此，本文着眼于高光谱图像特征的多分辨率挖掘与利用，通过多核学习方法将空间信息和光谱信息进行整合，以提高高光谱图像地物分类应用的能力。本文的主要研究内容是将信号的多分辨率分析理论引入到多核学习方法中，以实现高光谱图像的地物分类目的，分以下三个部分进行展开：首先，从高光谱图像的成像原理和数据特点出发，本文分析了多核学习方法在分类问题上的优势和不足，在多核学习理论框架之上着重介绍了核函数对样本相似性的度量性质，并以此为基础进行多核函数的构造，将抽象的映射特征空间形象化、具体化。接着，研究了典型的多核学习问题求解方法，引入多核函数的线性和非线性组合方式，为后续多分辨率多核学习方法研究提供支撑。然后，为将核函数的映射关系进行推广，将映射特征具体化，给出典型的二维高光谱图像特征提取方法，针对高光谱图像数据在空间和光谱维度上的相关性，将二维特征提取方法又推广到了三维特征提取。更重要的是，3D-Gabor实现了特征空间的多分辨率分析，通过对特征空间的伸缩变换，得到具有多分辨率特性的特征，并进一步通过实验验证了具有多分辨率特性的特征提取方法用于地物分类时的优越性能。最后，以多分辨率分析理论为指引，在核特征空间中建立了具有多分辨率学习能力的多核学习模型，实现了多分辨率多核分类器的构造。先以特征映射关系为基础构建基核函数，通过对映射空间进行多尺度化，得到具有多分辨率特性的多核函数。再针对多分辨率特性的特点，给出多核函数的非线性组合方法及多核学习问题的优化求解过程，以完成分类器的设计。最终验证了多分辨率多核学习在高光谱图像分类应用上的优越性。更多还原

【Abstract】 Hyperspectral imaging achieves detail observation in the spectral dimension,and the spectral resolution has been greatly improved, which provides richdiscriminative information for classification. However, in terms of land-coversclassification, the great high spatial and spectral resolution does not mean theeffective classification. The observation scale of classifier affects the similaritymeasurement of the inter class or the within class, but the observation scale of thetraditional classifier cannot adapt to the data changes automatically. Therefore, thispaper focus on the property of multi-resolution on hyperspectral image, andintegrates spatial information and spectral information by multiple kernellearning(MKL), which can improve the ability of land-covers classification onhyperspectral image.The main research of this paper is to introduce the theory of multi-resolutionanalysis to MKL, which is used to achieve effective classification, three aspects asfollowing in details:First, starting from imaging principle of hyperspectral image and datacharacteristic, the advantages and shortages of MKL in classification are analyzed inthis paper. The measurement property of kernel function is introduced significantlybased on the framework of MKL, providing an opportunity to construct multiplekernels, which visualizes the abstractive mapping feature space. Then, to introducethe linear and non-linear combination of kernels, several typical MKL algorithmsare given in details, providing the theory foundation to expand MKL tomulti-resolution MKL.Second, in order to expand the mapping relationship of kernel and visualize themapping feature, two typical two-dimensional hyperspectral image featureextraction methods are shown. In connection with the correlation between spatialand spectral on hyperspectral image data, the2D feature extraction method isextended to3D feature extraction. More importantly, multi-resolution analysis infeature space is realized by3D-Gabor filter, which can obtain multi-resolutionfeature through scaling feature space. Corresponding experiments are given to provethe superior performance of multi-resolution feature in land-covers classification.At last, taking the multi-resolution analysis theory as the guide, the MKLmodel with the ability of multi-resolution learning is created in kernel space toconstruct MKL classifier based on multi-resolution. Firstly, construct basic kernelbased on feature mapping and scale the mapping space to get the multi-resolution multiple kernels. Then, considering the characteristic of multi-resolution, give thenon-linear combination of multiple kernels and the solving process of MKL tocomplete the classifier design. Finally, validate the multi-resolution MKL onhyperspectral classification to emphasis the advantages of multi-resolution theory onhyperspectral image processing.更多还原