【作者】 杨粤涛；

【导师】 朱明；

【作者基本信息】 中国科学院研究生院（长春光学精密机械与物理研究所） ， 机械电子工程， 2012， 博士

【摘要】 多源图像融合技术是以图像为研究对象的信息融合，隶属于多源信息融合范畴内的一个重要分支--可视信息融合，是一种综合了传感器理论、模数信号转换、数字图像处理、计算机视觉以及人工智能等多种学科的现代高新技术，在军事应用和民用等众多领域都有着广泛的应用。近年来，基于变换域的多分辨率分解的像素级图像融合算法被广泛应用在多源图像融合领域中，有效地克服了空间域中的频谱失真问题，并取得了较好的效果。本论文在前人对像素级多源图像融合的研究工作的基础上，在多尺度变换的图像融合中，针对应用广泛的小波变换对二维图像中视觉效果最突出的大量曲线和曲面的奇异性不能最优的表示，只能用点奇异去逼近线面奇异，造成图像的轮廓、纹理等特征的模糊的缺点，研究了具有多尺度、多方向、各向异性和平移不变性的一种超完备的多尺度变换方法--非采样Contourlet变换(NonsubsampledContourlet Transform,NSCT)。而后以NSCT为基础，主要对红外和可见光图像融合与多聚焦图像融合的融合算法进行研究，并且在Matlab7.5和VC++6.0工具上对提出的所有算法进行了验证。本论文的主要工作可以总结为以下五点：1.深入研究具有具有多尺度、多方向、各向异性和平移不变性的非采样Contourlet变换理论，并且通过仿真实验呈现了NSCT变换对图像进行分解和重构的效果，构建了NSCT用于图像融合的算法框架和具体步骤。2.针对多源图像融合领域中的红外和可见光图像融合，提出了一种有效的基于改进OTSU区域分割和非采样Contourlet变换相结合的红外和可见光图像融合算法。根据红外和可见光成像传感器的成像特性，对源图像进行区域分割和图像的联合区域表示，再针对各区域的NSCT分解系数设计相应的融合规则的方法，实现了系数的最优化融合，有效的提高了红外和可见光融合图像的质量。3.通过对非负矩阵分解(Non-negative Matrix Factorization, NMF)的理论分析，深入研究了投影梯度非负矩阵分解理论(Projected Gradient Non-negative MatrixFactorization,PGNMF)并将其引入图像融合中，仿真实验的融合结果表明PGNMF不管是直接用于源图像融合还是用于NSCT分解后的低频系数融合规则中，都能够得到较好效果的融合图像，并且减小了算法的计算复杂度和耗时，使得其用于对图像质量要求不是很高而对实时性有要求的快速图像融合算法中成为可能。而且分别对红外和可见光图像、多聚焦图像两类不同传感器组合的图像进行融合实验，也在一定程度上体现了该算法的鲁棒性。4.通过对粒子群优化算法的理论研究，针对其存在的早熟收敛和粒子缺乏多样性而导致陷入局部最优的问题，借鉴了人工免疫中的克隆选择学说的思想，提出了一种改进克隆选择的免疫粒子群算法(Improve Clonal Selection ParticalSwarm Optimization,ICSPSO)，进而将其成功的应用在多聚焦图像融合领域，将图像融合问题转化为融合质量最优化的问题，很大程度上提高了融合图像的可靠性和融合效果，较好的融合效果和低时间消耗使得ICSPSO融合算法成为一种高效的快速图像融合算法。5.将压缩感知理论引入高分辨率图像融合中，提出了压缩感知的融合算法与压缩感知和NSCT相结合的融合算法，实验结果表明将这两种基于压缩感知的图像融合算法用于高分辨率图像融合，在牺牲微小的图像质量的前提下可以大幅度的降低融合算法的耗时，所以压缩感知理论的引入，为高分辨率图像融合给予了一种具有可行性的低耗时方法。更多还原

【Abstract】 The multi-source image fusion technique take the image as the object of study inthe information fusion field, one of the most important branches of the Multi-sourceinformation fusion-Visual Information Fusion. It s a mordern high-tech combinedsensor theory, analog to digital signal conversion, digital image processing, computervision, artificial intelligence and many other disciplines, has been widely used inmany range of areas such as military and civilian applications. In recent years, themulti-resolution decomposition based on transform domain pixel-level image fusionalgorithms have widely used in the field of multi-source image fusion, effectivelyovercomed the spatial spectrum distortion and have obtained the better effect.In this dissertation, on the basis of the previous pixel-level multi-source imagefusion algorithms research, in the field of multi-scale image fusion, we found that thewidely used method which called wavelet transform has some limitations. It cannoteffectively represent the most prominent visual effects such as the line discontinuities and the curve discontinuities in the two-dimensional image, and onlyuse the singular points to approximate the singular lines or curved surfaces. Thislimitation resulted in the mistiness of the profile and texture in the image. Therefore,aim at the wavelet s limitations, the research works in this dissertation focus onan overcomplete multi-scale transformation method-the nonsubsampled contourlettransform (NSCT) which has property such as multi-scale, multi-direction, anisotropyand translational invariance.Then we primarily research with the infrared and visiblelight image fusion algorithm and the multi-focus image fusion algorithm, moreover all the proposed algorithms are verified by Matlab7.5and VC++6.0tools.The main contributions of this dissertation can be summarized in the followingfive points:1. In-depth and comprehensive research with the nonsubsampled contourlettransform theory which has property such as multi-scale, multi-direction, anisotropyand translational invariance. After that we present the effects of the simulationexperiment in image decomposition and reconstruction with NSCT. At last weconstructed the framework and specific steps of the image fusion algorithms usedNSCT.2. Aiming at the infrared and visible light image fusion in the field ofmulti-source image fusion, we propose one kind of effective infrared and visible lightimage fusion algorithm which combined the improved OTSU regional segmentationand NSCT. On the basis of the characteristic of infrared and visible light imagesensors, regional segmentation and regional association are used in the source imagesat first, and then project the corresponding fusion rules to the NSCT decompositioncoefficients in different regions. This fusion algorithm achieved a most optimizedfusion method of the coefficients and effectively improved the quality of the infraredand visible light fused image.3. By theoretical analyzed the non-negative matrix factorization (NMF) theory,we detailed study on the projected gradient non-negative matrix factorization(PGNMF) and introduced it to the image fusion field. Experimental results indicatethat PGNMF either directly used in the source image fusion or used in the fusion rulesof NSCT decomposition low-frequency coefficient, it s able to reduce thecomputational complexity and time-consuming, when acquire manifest better fusedimage at the same time. Therefore, the proposed fusion algorithm can be used forreal-time image fusion system which called for less quality requirement of fusedimage. Moreover, we take both of the infrared and visible light image and themulti-focus image to experiment with the proposed fusion algorithm, to some extent,reflect the robustness of the algorithms. 4. By theoretical analyzed the particle swarm optimization (PSO) theory, wefound that PSO method is likely to converge prematurely and the lack of particlediversity lead the swarm to converge to the local optimum. Aimed at the disadvantageof PSO, combined with the artificial immune clonal selection theory, we proposed animproved clonal selection particle swarm optimization (ICSPSO), and the improvedalgorithm had application in multi-focus image fusion field successfully. The fusionalgorithm not only transforms the image fusion question as the optimization problemsbut also enhanced the fused image reliability and the fusion effect to a great extent.The better fusion effect and the lower time consumption causes ICSPSO fusionalgorithm to become one kind of effective fast image fusion algorithm.5. Introduced the compressed sensing (CS) theory to the image fusion field andproposed a new high-resolution image fusion algorithm based on CS. We alsocombined the CS and NSCT to resolve the high-resolution image fusion problem.Experimental results indicate that under the premise of slightly reducing the quality offused image, the proposed algorithm can greatly reduce the time-consuming. Therefor,the high-resolution image fusion algorithm based on compressed sensing is feasibleand effective, also can greatly reduce the time-consuming of the algorithm.更多还原