基于小波核滤波器和稀疏表示的遥感图像融合

发布时间：2019-04-13 07:43

【摘要】：随着遥感技术的发展，其应用也越来越广泛，在地学科学、农业、气象、林业、城市规划、环境监测等等领域均有不同程度的应用。然而，由于遥感传感器技术本身的限制，所获得的遥感数据往往不能反映出区域的全部信息。为了更好的理解该地域的内容，将不同遥感器获得的图像信息进行融合便成了一项十分经济且有效的方案。近年来，为了提高对遥感图像的解译能力，信息融合的技术被引入到融合多遥感器图像及遥感卫星图像中。本论文以遥感图像融合为研究背景，结合国家自然科学基金、国家“863”计划、“973”计划以及“111”创新引智计划等项目的任务与需求，利用多尺度几何分析、机器学习方法和优化算法等工具，完成了遥感图像融合方法的研究工作。论文主要工作概括如下： 1.借助支撑矢量机逼近原理对图像进行逼近建模，由此实现以核函数来描述图像，并提出一种多尺度变换工具——小波核滤波器。将小波核滤波器应用在桥梁分类、移动和静止目标获取和识别数据库（Moving and Stationary TargetAcquisition and Recognition，MSTAR）的数据识别及合成孔径雷达图像去斑中。桥梁分类和MSTAR数据识别应用结果表明，图像数据经过小波核滤波器之后得到系数能更好的表达原图像的信息，而图像去斑的应用也显示出由于该滤波器具有平移不变性，针对去斑中出现的振铃效应基本清除。 2.将提出的小波核滤波器应用到遥感图像融合中。小波核滤波器具有多尺度性、平移不变性、完全重构性等，使得该滤波器在图像融合的应用中具有优势。针对多传感器图像的特点，以区域能量最大值作为融合策略，实现基于小波核滤波器的多传感器图像融合，并与其他多尺度变换工具如小波变换、非下采样的小波变换、Contourlet变换、非下采样的Contourlet变换进行比较。四组针对曼彻斯特大学图像融合库的多源图像融合结果表明，小波核滤波器应用于多传感器图像融合是有效的，克服了图像融合中常出现的振铃效应，细节保持较好，取得更为清晰的融合结果。针对多光谱与全色图像的融合问题，在小波核滤波器的基础上，提出两种融合策略：其一是与传统的亮度-色调-饱和度变换相结合，对亮度I分量进行处理，，将全色图像的细节加入到I分量中；其二是采用改进的空间分辨率增加框架法（Amélioration de laRésolution Spatiale par Injection de Structures，ARSIS）作为融合框架，利用多尺度分析手段为多光谱图像补充上缺失的细节成分。随后给出的来自于光学卫星的多光谱图像结果表明，小波核滤波器能够应用在多光谱图像与全色图像的融合中，获取融合结果，两种融合框架均能获得所需的具有高分辨率的多光谱图像，为后续多光谱图像的处理及应用奠定了基础。 3.针对遥感图像融合问题，提出了小波核滤波器结合优化算法的遥感图像融合方法。首先，结合小波核滤波器，将粒子群算法应用到多传感器图像融合中。针对细节子带仍采用区域能量最大值的融合策略，而近似子带则选择粒子群算法去搜索得到一个最优的近似子带。实验结果表明，结合小波核滤波器和粒子群算法的方法是有效的，可以得到相对最优的融合结果。针对多光谱与全色图像的融合问题，结合小波核滤波器和克隆选择算法给出两种融合策略：其一是通过小波核滤波器中参数的变化，给出多组小波核滤波器，结合亮度-色调-饱和度变换获得多组融合结果，克隆选择算法用来寻找到最优权值组合给出最优融合结果；其二是利用克隆选择算法寻找最优的亮度I分量，得到一个最逼近全色图像的I分量进行随后的融合处理。结果表明，结合优化算法的融合策略能够找到最优值，得到相对最优的融合结果。 4.随着稀疏表示理论的发展，该理论已被成功的应用于图像处理领域中。由于图像能够采取稀疏表示的方式来得到系数，稀疏的系数用来表达源图包含的信息，因此利用该系数便可完成图像融合的要求。根据多源图像的特点及稀疏表示获得的稀疏系数的特点，给出五种融合策略下得到的融合结果，并进行了比较，选择出适合于稀疏系数的融合规则，并与传统的多尺度变换方法进行比较，结果表明稀疏表示理论应用到图像融合领域亦能获得较优的结果。针对多光谱图像融合的问题，首先参照多光谱图像的特点，将基于稀疏表示的超分辨方法应用到多光谱图像与全色图像的融合中，通过超分辨方法先获得对应于低分辨率多光谱图像的高分辨率图像，结合ARSIS框架与全色图像融合，获得了较好的结果。 5.考虑到多光谱图像融合的目的在于增加多光谱图像的细节信息含量，将二维经验模式分解引入到前文提到的基于广义的亮度色度饱和度变换和小波核滤波器结合的多光谱图像融合方法中，并且为了找到既能提高光谱性又增加细节信息的结果，将折中参数引入了融合方法中，由此获取高分辨率多光谱图像，使得新获取的图像能够在保持光谱特性的基础上增加尽可能多的细节信息。
[Abstract]:With the development of remote sensing technology, its application is becoming more and more extensive, and there are different applications in the fields of geoscience, agriculture, meteorology, forestry, urban planning, environmental monitoring and so on. However, due to the limitations of the remote sensing sensor technology itself, the acquired remote sensing data often cannot reflect all of the information of the region. In order to better understand the content of the region, the fusion of the image information obtained from the different remote sensors is a very economical and effective solution. In recent years, in order to improve the interpretation capability of the remote sensing image, the information fusion technology is introduced into the integrated multi-sensor image and the remote sensing satellite image. Based on the research background of remote sensing image fusion, this paper makes use of multi-scale geometric analysis, machine learning method and optimization algorithm in combination with the task and demand of national natural science fund, national "863" plan, "973" plan and "111" innovation and intelligence program. The study of the method of remote sensing image fusion is completed. The main work of the paper is as follows: In that follow:1. the image is approximated by means of the approximation principle of the support vector machine, and the kernel function is used to describe the image, and a multi-scale transform tool _ wavelet kernel is proposed. Filter. Application of wavelet kernel filter in data recognition and synthetic aperture radar image of bridge classification, mobile and stationary target acquisition and recognition database (MSTAR) The results of the application of the bridge classification and the MSTAR data recognition show that the image data can express the information of the original image better after passing through the wavelet kernel filter, and the application of the image despot also shows that the filter has the translation Invariance, the ringing effect base that appears in the spot-to-spot this cleanup.2. Applying the proposed wavelet kernel filter to remote sensing In image fusion, wavelet kernel filter has multi-scale, translational invariance, complete reconstruction and so on, so that the filter is used in image fusion The multi-sensor image fusion based on the wavelet kernel filter is realized based on the characteristic of the multi-sensor image and the maximum value of the regional energy is taken as a fusion strategy, let transform, non-downsampled Contourlet change The results show that the wavelet kernel filter is effective in multi-sensor image fusion, and the ringing effect in the image fusion is overcome. In order to solve the fusion problem of multi-spectrum and full-color image, two fusion strategies are proposed on the basis of wavelet kernel filter: one is combined with the traditional brightness-tone-saturation transformation, and the brightness I component is processed, and the detail of the full-color image is added. The second is to use the improved spatial resolution adding frame method (ARSIS) as the fusion frame, and the multi-scale analysis is used to supplement the multi-spectral image. The results of the multi-spectral images from the optical satellite show that the wavelet kernel filter can be used in the fusion of the multi-spectral image and the full-color image to obtain the fusion result, and the two fusion frames can obtain the required high resolution. Multi-spectral image, processing and application of subsequent multi-spectral image In order to solve the problem of remote sensing image fusion, a method of combining the wavelet kernel filter with the optimization algorithm is proposed. The invention relates to a sensing image fusion method, which comprises the following steps of: firstly, applying a particle swarm algorithm to a multi-pass filter in combination with a wavelet kernel filter, In the image fusion, the fusion strategy of the maximum value of the region energy is still used for the detail sub-bands, while the approximate sub-bands select the particle swarm algorithm to search for one. The experimental results show that the method of combining wavelet kernel filter and particle swarm optimization is effective and can be obtained. In order to solve the fusion problem of multi-spectrum and full-color image, two fusion strategies are given in combination with the wavelet kernel filter and the clonal selection algorithm. a group wavelet kernel filter is combined with the brightness-tone-saturation transformation to obtain a plurality of sets of fusion results, the cloning selection algorithm is used for finding the optimal weight combination to give the optimal fusion result, The optimal brightness I component is obtained by obtaining an I component of the most approximate full-color image. The result shows that the optimal value can be found with the fusion strategy of the optimization algorithm to get the relative value. 4. With the development of sparse representation theory, the theory has been successfully in that field of image proces, since the image is able to obtain a coefficient in a sparse representation, the sparse coefficient is used to express the information contained in the source map, according to the characteristics of the multi-source image and the characteristic of the sparse coefficient obtained by the sparse representation, the fusion result obtained under the five fusion strategies is given, the fusion rule suitable for the sparse coefficient is selected, The results show that the sparse representation theory is applied in the field of image fusion. in order to solve the problem of multi-spectral image fusion, firstly, the super-resolution method based on the sparse representation is applied to the fusion of the multi-spectral image and the full-color image by referring to the characteristic of the multi-spectral image, High-resolution image of image, combined with ARSIS frame and full-color image fusion 5. The aim of the multi-spectral image fusion is to increase the detail information content of the multi-spectral image, and the two-dimensional empirical mode decomposition is introduced into the above-mentioned generalized luminance-chroma saturation transform and the wavelet kernel filter combination. In the multi-spectral image fusion method, and in order to find the result that both the spectral property and the detail information can be improved, the compromise parameter is introduced into the fusion method, thereby obtaining a high-resolution multi-spectral image, so that the newly acquired image can be on the basis of keeping the spectral characteristic,
【学位授予单位】：西安电子科技大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP751

【相似文献】