【作者】 洪日昌；

【导师】 吴秀清；

【作者基本信息】 中国科学技术大学 ， 信号与信息处理， 2008， 博士

【摘要】 随着成像传感器技术的发展,越来越多的传感器被应用于各个领域中。传感器数量的增多导致获取的数据量急剧增加并呈现多样性,传统的信息处理方法已经不能满足新的需求。因此多源信息融合技术在这样的背景下应运而生。多源信息融合是指对来自多个传感器获得的多源信息进行多级别,多方面,多层次的处理和综合,从而获得更为丰富,精确和可靠的有用信息。而多源图像融合作为多源信息融合中以图像为对象的研究领域,综合了传感器、图像处理、计算机和人工智能等多个学科,其主要思想是通过采用一定的算法,将两个或多个具有互补或冗余特性的源图像融合成为新图像,使得融合后的图像能最大限度地利用互补信息,减少冗余,从而获得更高的清晰度和可理解性。本论文集中探讨了多源图像融合这一课题产生的背景,研究的意义,融合的基本原理和体系结构,以及国内外已有的研究成果,并指出已有的研究工作中存在的不足之处。论文的研究内容即是围绕多源图像融合研究中一些存在的问题和不足之处展开的,具体内容包括基于多分辨率分解融合方法的工作总结和新思路,利用变分偏微分方程实现融合的探索,基于区域和多分辨率分解的特征级融合方向上的研究,如何实现图像融合性能的客观评价,以及多源图像特征级融合在实现道路目标识别上的具体应用。主要工作和创新成果如下:1.总结已有的基于多分辨率分解的多源图像融合方法,提出利用双正交多小波变换的多光谱与全色图像融合。其中多小波变换作为单小波的扩展,增加了小波基的个数,从而能很好地解决对称性与正交性,支集长度与消失矩之间的矛盾。在融合策略上,为保留源图像多小波分解后尺度系数的显著性信息,提出利用平均与选择相结合实现融合。2.提出基于特征保持的一阶对比度多聚焦图像融合方法。融合过程中根据源图像反映的不同特征的重要性,由特征图强度设计不同的权重,然后根据源图像的对比度信息的相应加权求取其主分量作为融合的目标梯度场,从而实现特征保持的图像融合。最后把融合算法拓展到彩色域,利用特征强度得到的权重实现彩色域中的高保真度融合。3.由于在图像融合中,人们往往关注于图像中的实际目标或区域,而不是单个像素。因此我们提出结合双正交多小波和图分割的特征级融合方法,通过匹配度测试确定分割得到区域中的小波系数的融合规则。并且总结了基于多分辨率和区域的特征级融合的统一框架,分析单源分割和联合分割对于融合性能的影响,以及具体的融合策略的选取方法。4.基于已有的客观评价方法指出具体的融合中评价指标的选取,提出基于图像结构相似性的客观评价方法。评价中主要考察源图像和融合图像之间的均值、方差相似性和相关系数,并且从灰度和梯度两个方向分别来做评价。5.提出一种在低分辨率遥感图像中道路目标的全自动识别算法。算法利用感知编组确定由边缘检测和无效线段去除后得到的道路边缘线段组,然后确定道路种子点,基于动态规划的道路跟踪算法跟踪道路种子点生成整个道路网,最后利用知识推理去除道路虚警。6.由于高分辨率遥感图像中的道路目标不仅包括低分辨率遥感图像中道路的拓扑特性,还包括高分辨率图像中的目标特性、光谱特性等。因此我们提出结合线性检测算子和分类结果的特征层融合道路识别算法。其中利用道路在高分辨率图象中可以被建模为具有某些统计行为的图像块,而这些图像块可以通过分类获得的特性,然后基于道路的拓扑关系辅助目标检测。更多还原

【Abstract】 Various image sensors appear and their acquired data become explosive and multifarious with the fast development of micro-electronics. In this situation, traditional information processing cannot meet requirements and information fusion rises. The concept of information fusion indicates the information processing and synthesization of multilevel, multiaspect and multilayer. Multi-source image fusion is derived from information fusion and it includes digital image processing, computer science and artificial intelligence and so on. Image fusion is implemented by integrating multiple source images with redundant and complementary information into one result with better intelligibility and definition. In this way, redundancy can be reduced, while complementary information can be utilized more effectively.This thesis studies the research backgrounds of multi-source image fusion and its significances, basic theories and systematic constructure of image fusion. Then we review the achievements of image fusion research at home and abroad. Following that, we set to our works on the deficiencies of existed works. Detailed contents consist of the reviewal and novel idea about multiresolution based image fusion, explorations in image fusion using PDEs, investigations into combination of multiresolution analysis and region segmentation, objective evaluations of various fusion methods and the applications of feature level fusion, such as road object recognition in high resolution multispectral images. The main work and innovations are listed as follows:1. Based on a reviewal and summarization of existed image fusion methods which utilize multiresolution analysis, we propose a novel biorthogonal multiwavelet based multispectral and PAN image fusion algorithm. Increased number of basis in multiwavelet can handle the tradeoff between symmetry and orthogonality, filter length and vanishing moment. Meanwhile, combination of average and selection is used in the fusion scheme for preserving the salient information of scale coefficients from sources.2. We propose a salience preserving based first-order contrast multi-focus image fusion method. According to the different importance of each source, we set different weights to each pixel in different source. Then weighted contrasts of sources are integrated into the target gradient of fused result. Moreover, we extend it into color domain by performing importance weight based color channel combination.3. In image fusion, people pay more attentions to real object or region within sources, but not single pixels. Here we fuse optical and IR images by combing multiresolution analysis and region segmentation. Hereinto, match measure is utilized as the decision rules of wavelet coefficients. Meanwhile, we give a detailed analysis of the framework of multiresolution and region based feature level fusion, impacts of uni-modal and joint segmentation for fusion, selection of fusion rules.4. We set forth most existed objective evaluation methods and how to adopt the objective evaluation method. An objective evaluation method based on structure similarity between sources and fused result is presented. The evaluation method takes the similarity of mean, variance and correlation into accounts and evaluates the fusion method from the viewpoint of gray values and gradients.5. An automatic road recognition method in low resolution remote sensing image is proposed here. Hereinto, perceptual grouping is introduced to decide the road edge line group after edge detection and redundant line segments exclusion. Road seeds are generated accorading to candidate road edge line groups. Then road network is delineated by dynamic programming based road tracking. Finally false road segments are eliminated by knowledge ruling.6. Road object has more traits in high resolution remote sensing images. It includes not only the topology, but also the spectral and shape traits and so on. Thus we proposed a road network recognition method, which combines both edge detection and spectral classification. Bacause roads in high resolution images can be modeled as image blocks with same statistic characteristics and these blocks can be acquired by classification.更多还原