【作者】 周波；

【导师】 祝忠明；

【作者基本信息】 成都理工大学 ， 信号与信息处理， 2012， 硕士

【摘要】 多源遥感图像的融合主要是将不同传感器所获取的同一地区或同一场景的图像进行融合，融合后的图像可信度更高，模糊较少，可理解性更好，更适合人的视觉及计算机检测、分类、识别、理解等处理。多源遥感图像的融合目前更多的还是处于像素级层面的研究，但是，随着人们对遥感图像中感兴趣的目标进行识别和跟踪的需求越来越多，像素级的融合方法便显得越来越不能满足这种需求。基于特征级的图像融合是依赖于提取图像区域特征的一种融合层次，由于是基于图像特征的前提，因此这种融合方式能有效的满足图像目标识别和跟踪的目的。遥感图像的融合的一般步骤是，对图像进行预处理、配准，然后再对图像进行融合。本文在图像配准方面，首先分别详细介绍了基于空间关系和基于特征相似性的配准方法；然后，在这两种配准方法的基础上，充分利用这两种方法的优点提出了一种新的组合基于空间关系和特征相似性的配准方法。通过配准实验表明，组合空间关系和特征相似性的配准方法所得到的配准图像效果要比单独用这两种方法得到的配准效果更佳，配准所需要的时间也更短。在图像配准的基础上，本文在第三部分研究了两种基于特征级的融合方法：Kalman滤波与多特征相结合的含噪图像融合、多通道Gabor滤波用于多特征图像融合，并在此基础上提出了一种基于二次融合多特征的图像融合算法。Kalman滤波与多特征相结合的含噪图像融合方法是通过将源图像经过Kalman滤波后，再提取出图像的特征以及对图像进行窗体分割，然后通过窗体相似度对图像融合。Gabor滤波用于多特征图像融合主要是针对图像的纹理特征进行处理以达到最终对图像进行有效融合的目的。基于二次融合多特征的图像融合方法主要是对通过FCM方法获得的分割区域的特征矩阵，经过主成分分析方法处理降维后再进行融合处理。在对三种融合方法进行研究的同时，本文紧接着分别用实验对每一种融合方法的性能进行了分析，实验表明，基于二次融合多特征的图像融合方法对图像的融合都是有效的。最后本文利用文中提出的新的配准方法和融合方法相结合进行了实验，并且获得了比较理想的效果。然而，任何一种方法都不是十全十美的，这些方法也都有其各自的缺点。但是，随着人们对基于特征级的多源遥感图像融合的不断研究，相信不断会有更多更好的融合方法出现，以弥补这方面的不足。更多还原

【Abstract】 Multi-source remote sensing image fusion is the same area or the image of thesame scene obtained by the different sensors are fused, the fused image higherreliability, fuzzy less understandable and better, more suitable for human visualComputer detection, classification, recognition, understanding and treatment.Multi-source remote sensing image fusion is more or at the level of the pixel level,However, as more and more demand for people interested in remote sensing imagetarget identification and tracking, pixel-level fusion method appears to more to theless able to meet this demand. Images based on feature level fusion is dependent onthe extraction of image area features a fusion level, because it is a premise based onimage features, so this combination method can effectively meet the image targetrecognition and tracking purposes.The general steps of the integration of remote sensing images, the imagepre-processing, registration, and then the image fusion. Registration in the imageaspects, the first were described in detail the similarity based on spatial relationshipsand feature-based registration method; Then, on the basis of these two registrationmethods make full use of the advantages of these two methods presents a The newcombination based on spatial relationships and characteristics similar to theregistration method. Registration experiments showed that the combination of spatialrelationships and characteristics similar registration method with quasi-image effectthan separately by the two methods better alignment, the shorter the time required forregistration.Image registration based on research in the third part of the two methods basedon feature level fusion methods: Kalman filtering and multi-feature combination ofthe noisy image fusion, multi-channel Gabor filtering for multi-feature fusion, and Based on this proposed image fusion algorithm based on quadratic integration ofmulti-feature. Kalman filtering and multi-feature combination of the noisy imagefusion method is through the source image after the Kalman filter, to mention toremove the image characteristics and image segmentation form, and then by formsimilarity fusion. Gabor filter for image fusion of multi-feature texture features forimage processing in order to achieve the final image for the effective integration of.Secondary fusion of multi-feature-based image fusion method is obtained by FCMmethod to split the regional characteristics of the matrix after principal componentanalysis approach to dimensionality reduction and then fused. Three fusion methods,followed respectively by experiments on the performance of each fusion methodanalysis, the experiments show that the image fusion method based on the secondaryfusion of multiple features of image fusion are valid.Finally, the use of the proposed new registration method and fusion methodcombining experiments were carried out and obtained the desired results. However,either method is not perfect, these methods also have their respective shortcomings.However, with the ongoing research of the integration of multi-source remote sensingimage based on the characteristics of class, I believe, continue to have more and betterintegration method appears, in order to compensate for this deficiency.更多还原