稀疏特性分析在遥感图像处理中的应用研究

发布时间：2018-04-21 00:05

本文选题：稀疏特性 + 压缩传感　；参考：《复旦大学》2014年硕士论文

【摘要】：近年来,作为最流行的概念之一,稀疏特性被广泛应用于信号处理领域,包括信号压缩与编码、信号加密与传输、信号源分离、信号去噪与重构、特征提取等。同时,稀疏特性作为理论和实用相结合的信号特性,一直以来在应用数学的诸多领域广受关注,例如：理论信号处理、统计估计以及计算谐波分析等。学界对于稀疏特性的热情源于一项有别于著名的香农采样定理的新型采样理论,这个理论被称为“压缩传感”。压缩传感理论基于信号本身具有稀疏性的先验知识,而香农采样理论是针对普适的频率带限信号进行设计的。如今的压缩传感理论就像二十多年前的小波变换理论一样,将各个不同的研究领域紧密联系在一起。通过在采样和稀疏特性之间建立起一种直接的联系,压缩传感理论已经对医学成像、编码与信息理论、信号采集与图像处理以及地球物理与天文数据分析等诸多科研领域产生了深远的影响。压缩传感理论体现了稀疏特性的重要价值,同时也阐明了一系列如小波变换、曲线波变换等有关稀疏数据表示方法的原理。事实上,一个信号一般说来在直接空间(即像素空间)中是不具有稀疏特性的,但通过在某一特定函数集合上的分解之后,该信号可能变得非常稀疏。本论文重点研究稀疏特性在遥感领域中的一系列应用场景,从传统的信号处理方法(图像去噪以及修复重构)扩展到遥感图像中目标场景的自动分类等。第二章介绍了稀疏特性分析与形态多样性概念之间的内在联系,在此基础上,第三至第五章分别讨论了稀疏特性分析的具体应用。第三章重点讨论了稀疏特性分析在遥感图像去噪中的应用。本章介绍了一种基于广义形态成分分析的新型遥感图像去噪方法。这个新算法将经典的形态成分分析算法进一步扩展到了盲源信号分离的架构中。广义形态成分分析算法采用迭代限定阈值的策略来调整优化各参数。算法首先处理图像中最显著的那些特征,然后逐步合并那些没有那么显著的特征来细细调整整个模型所有的参数。文中对广义形态成分分析算法的计算复杂度进行了数学分析,并与当前最先进的遥感图像去噪算法进行了一些对比实验。为了对实验中遥感图像去噪算法的效果进行衡量和定量化分析,本文分别采用峰值信噪比指数和结构相似度指数,从灰度保真度和结构保真度两个方面来评估去噪效果。对实验结果的定量分析已经证明了所提出的广义形态成分分析算法出色的去噪性能,这一定量分析结果也与去噪后的图像所展示出的视觉效果相吻合。观察者甚至很难通过视觉效果来区分出原始的无噪声图像和采用广义形态成分分析算法去噪后恢复出的图像。第四章重点讨论了稀疏特性分析在遥感图像修复中的应用。本章介绍了一种基于增速广义形态成分分析的新型遥感图像修复方法。由于广义形态成分分析算法具有很好的表示和分离形态多样性的能力,所以是一种具有当前最先进的图像修复效果的算法。本章提出的增速广义形态成分分析算法能够提高广义形态成分分析算法迭代过程的速度。通过对广义形态成分分析算法添加更多的一些假设条件,新提出的增速广义形态成分分析算法被证明是一种更加快速的算法,这样就能够处理非常大规模的一些应用问题。一系列实验结果说明增速广义形态成分分析算法能够很好地修复带有不同模式丢失像素的遥感图像。观察者甚至很难通过视觉效果来区分辨别原始遥感图像和经过修复之后的图像。峰值信噪比指数和结构相似度指数也进一步证实了增速广义形态成分分析算法惊人出众的图像修复性能,并且解释了为什么采用增速广义形态成分分析算法能够获得这么好的图像修复视觉效果。对于时间消耗对比实验结果的量化分析证明了,增速广义形态成分分析算法确实能够极大地提高广义形态成分分析算法的迭代速度。第五章重点讨论了稀疏特性分析在卫星图像场景分类中的应用。本章提出了一种基于对纹理层和底图层进行形态成分分析的高分辨率卫星图像分类方法。在该分类算法中所使用的字典矩阵是通过独立元分析的方法构造而成的。经过形态成分分析分解之后,就得到了纹理层和底图层的形态系数向量。这些向量被称为输入的高分辨率卫星图像的稀疏表示。结合纹理层和底图层这两层各自的特征,就可以根据最大似然估计机制,计算出对目标图像进行分类的总体概率大小。对实验结果进行的量化分析以及与一些经典的图像分类算法的对比都证明了,基于对纹理层和底图层进行形态成分分析的卫星场景分类算法比大多数最先进的分类算法有着更高的效率、更好的分类准确度和分类效果。
[Abstract]:In recent years, as one of the most popular concepts, sparsity has been widely used in the field of signal processing, including signal compression and coding, signal encryption and transmission, signal source separation, signal denoising and reconstruction, feature extraction and so on. At the same time, sparsity is used as a signal characteristic of combination of theory and practice. Domains are widely concerned, such as theoretical signal processing, statistical estimation, and calculation of harmonic analysis. The enthusiasm of the academic community for sparse characteristics is derived from a new sampling theory different from the famous Shannon sampling theorem. This theory is called "compression sensing". The theory of compressed sensing is based on the prior knowledge of the sparse characteristic of the signal itself. Shannon sampling theory is designed for universal frequency band limited signals. Today's compression sensing theory, like the theory of wavelet transform more than 20 years ago, closely links different research fields. By establishing a direct connection between sampling and sparsity, the theory of compression sensing has already been applied to medicine Imaging, coding and information theory, signal acquisition and image processing, and geophysical and astronomical data analysis have had a profound impact. The theory of compressed sensing embodies the important value of sparse characteristics, and also expounds a series of principles on sparse data representation, such as wavelet transform, curve wave transformation, etc. In fact, a signal generally does not have a sparse characteristic in direct space (i.e. pixel space), but the signal may become very sparse after the decomposition of a particular set of functions. This paper focuses on a series of application scenarios in the field of remote sensing, from the traditional signal processing method (image). The second chapter introduces the internal relations between the sparse characteristic analysis and the concept of morphological diversity, and on this basis, the third to fifth chapters discuss the specific application of the sparse characteristic analysis respectively. In the third chapter, the sparse characteristic analysis is discussed in the remote sensing map. In this chapter, a new remote sensing image denoising method based on generalized morphologic component analysis is introduced in this chapter. This new algorithm further extends the classical morphological component analysis algorithm to the architecture of blind source separation. The generalized morphological component analysis algorithm uses the strategy of iterative threshold threshold to adjust the parameters. The algorithm first deals with the most significant features in the image, and then gradually combines those less significant features to finely adjust the parameters of the whole model. In this paper, the computational complexity of the generalized morphological component analysis algorithm is mathematically analyzed, and some comparison is made with the most advanced remote sensing image denoising algorithm. In order to measure and quantify the effect of the remote sensing image denoising algorithm in the experiment, this paper uses the peak signal-to-noise ratio index and the structure similarity index respectively to evaluate the denoising effect from two aspects of the gray fidelity and the structure fidelity. The excellent denoising performance of the algorithm is analyzed, which is consistent with the visual effect shown in the image de-noising. The observer is even difficult to distinguish the original noise free images and the generalized morphologic component analysis algorithm to restore the image. The fourth chapter focuses on the sparse characteristics. In this chapter, a new method of remote sensing image restoration based on generalized morphological component analysis is introduced in this chapter. Because the generalized morphological component analysis algorithm has the ability to express and separate the morphological diversity, it is an algorithm which has the first image restoration effect. The speed of the generalized morphologic component analysis algorithm can improve the speed of the iterative process of the generalized morphologic component analysis algorithm. By adding more assumptions to the generalized morphologic component analysis algorithm, the newly proposed growth generalized morphologic component analysis algorithm is proved to be a quicker algorithm, so that it can handle very big rules. A series of experimental results. A series of experimental results show that the growth generalized morphological component analysis algorithm can well repair remote sensing images with different patterns of lost pixels. The observer is even difficult to distinguish between the original remote sensing image and the restored image by visual effect. The peak signal to noise ratio index and the structure similarity degree are very difficult to distinguish. The index also further confirms the remarkable image restoration performance of the generalized morphological component analysis algorithm, and explains why the growth generalized morphologic component analysis algorithm can obtain such good image restoration visual effects. The analysis algorithm really can greatly improve the iterative speed of the generalized morphological component analysis algorithm. The fifth chapter focuses on the application of the sparse characteristic analysis in the satellite image scene classification. In this chapter, a high resolution satellite image classification method based on the morphological composition analysis of the texture layer and the bottom layer is proposed. The dictionary matrix used in the method is constructed by the method of independent meta analysis. After the decomposition of the morphological component analysis, the morphological coefficient vectors of the texture layer and the bottom layer are obtained. These vectors are called the sparse representation of the high resolution satellite images of the input. They are combined with the two layers of the texture layer and the bottom layer. According to the maximum likelihood estimation mechanism, the overall probability of the target image classification is calculated. The quantitative analysis of the experimental results and the comparison with some classic image classification algorithms prove that the satellite scene classification algorithm based on the morphological composition analysis of the texture layer and the bottom layer is more advanced than most. The classification algorithm has higher efficiency, better classification accuracy and classification effect.

【学位授予单位】：复旦大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【共引文献】