【作者】 廉蔺；

【导师】 李国辉；

【作者基本信息】 国防科学技术大学 ， 控制科学与工程， 2013， 博士

【摘要】 图像配准是将在不同时间、从不同视角或用不同传感器拍摄的同一场景的两幅或多幅图像在空间上进行对齐的过程。红外与可见光遥感图像的配准是多传感器图像配准的一类重要组合形式，二者有机结合可以增强场景信息的互补性，减少对场景理解的不确定性，因此在军事情报获取、自主导航、末制导、目标跟踪、图像融合、变化检测和环境监测等应用领域中受到了广泛关注。由于成像机制不同，同一场景的红外与可见光遥感图像存在比较明显的视觉差异，此外图像间还可能存在几何形变、局部遮挡和噪声干扰等因素，使得红外与可见光遥感图像的自动配准富有挑战性。论文从红外与可见光的光谱特性出发，分析了红外与可见光遥感图像的成像特征，探讨了二者的差异性和相似性，并总结了两种常用的图像配准框架。以此为基础，本文提出了多种红外与可见光遥感图像自动配准方法，并通过对比实验对所提方法的有效性进行验证和分析。论文的主要创新点及取得的研究成果包括：(1)提出了一种新的局部不变特征描述子——加窗灰度差直方图(WindowedIntensity Difference Histogram, WIDH)，该描述子能够充分利用特征局部区域内的灰度差信息构建描述矢量，实验结果表明该描述子对图像中可能出现的照度变化、结构模糊、几何形变和噪声干扰等具有较好的不变性，描述矢量维度低但辨识力强。在此基础上，使用斑块特征及WIDH描述子提出了一种近红外与可见光遥感图像自动配准算法。该算法使用SURF检测子提取图像中的斑块特征，并在每个斑块特征周围的局部邻域内构造WIDH描述矢量，在Euclidean距离准则下使用最近邻搜索算法确定矢量间的匹配关系，最后使用随机采样一致性算法剔除误匹配特征对并确定变换参数。实验结果表明，该算法能够比基于SURF的配准算法更有效地实现近红外与可见光遥感图像间的自动配准。(2)提出了一种基于分支点匹配的红外与可见光遥感图像自动配准算法。在特征检测阶段，基于方位一致性提出了一种新的分支点检测和特征化算法，检测到的分支点不仅具有位置信息，而且包含分支边缘数和分支边缘倾角等局部结构信息。实验结果表明，本文提出的分支点检测算法不仅具有较高的定位精度，而且对噪声干扰及对比度变化等具有较好的鲁棒性；在分支点匹配阶段，提出了局部结构相容度的概念，用于检验两个分支点间分支边缘数目和分支边缘倾角的一致性程度，并将其作为匹配约束项嵌入到高斯混合模型分量的后验概率计算中，有效增强了算法对噪声和外点等干扰因素的抵抗能力，并提高了算法的收敛速度。(3)提出了一种基于参数分步估计的红外与可见光遥感图像自动配准算法。使用矩阵正交分解将仿射变换模型的六个参数分离为易于估计的切变、尺度比例、旋转、尺度以及平移（包括x和y方向上的平移量）等因子，并将分解后的参数分成两组，一组由切变、尺度比例和旋转因子构成，另一组由尺度和平移因子构成，然后利用图像中的线段特征分步估计这两组参数的取值。在参数估计的第一步中，基于图像中的线段方向一致性模型构建以切变、尺度比例和旋转因子为变量的目标函数，在给定的参数空间中使用最优化方法搜索使目标函数取最优的参数取值，将仿射变换简化为相似性变换。在参数估计的第二步中，基于线段对齐度准则构建以尺度和平移因子为变量的目标函数，并采用和第一步相同的最优化方法计算参数值。参数寻优过程使用模拟退火算法和改进Powell算法相结合的混合优化策略，有效保证了寻优结果的全局性和精度。该算法通过参数分步估计，将仿射变换的六维参数空间简化为两个三维空间，不仅显著缩减了搜索空间的大小，而且有效降低了参数寻优过程陷入局部最优的概率。(4)提出了一种基于结构最优映射的红外与可见光遥感图像自动配准算法。算法以遥感图像中空间连贯的边缘结构为配准基元，抑制噪声及不一致信息的影响，并基于边缘结构构造对齐测度函数，将红外与可见光遥感图像的配准问题转化为该目标函数在参数空间中的最优化问题。在参数寻优过程中同样采用了混合优化框架，先使用并行遗传算法搜索参数的全局近似最优解，再使用改进的Powell算法对参数的全局寻优结果进行局部求精。为了提高配准效率，提出了一种在两级尺度空间中使用逐级逼近模型的灵活框架，并在粗细两级尺度中分别使用仿射变换模型和投影变换模型。由于使用图像中更具普遍性的边缘结构特征，并融合了形变描述能力更强的投影变换模型，因而该算法在红外与可见光遥感图像配准具有较好的鲁棒性和适应性，降低了由于特定配准基元缺失导致配准失败的可能性。更多还原

【Abstract】 Image registration is the process of overlaying two or more images of the same scenetaken at different times, from different viewpoints, and/or by different sensors.Infrared-visible remote sensing image registration is an important combination form ofmulti-sensor image registration. Organic integration of the two types of images canenhance the complementarity of the scenes information, and reduce the uncertainty forscene understanding. Therefore, Infrared-visible remote sensing image registration hasreceived extensive attention in many application fields, such as military intelligenceacquisition, autonomous navigation, terminal guidance, target tracking, image fusion,change detection, environment monitoring and so on.There exist obvious visual differences between the infrared and visible images of thesame scene due to their different imaging mechanisms. Other influence factors, moreover,may also occur between the images, for example, geometric deformation, partial occlusionand noise perturbation, etc. These make the automatic registration of the infrared andvisible images challenging. This dissertation starts from the spectral characteristics ofinfrared ray and visible light, analyses their imaging characteristics, explores thedifferences and similarities between infrared and visible images, and summarizes twocommonly used image registration frameworks. Based on this knowledge, a variety ofinfrared-visible remote sensing image registration methods are proposed, and theeffectiveness of these methods are validated and analysed throuth comparative experiments.The main achievements and contributions of the dissertation are listed as follows:(1) A novel local invariant feature description opeartor, Windowed IntensityDifference Histogram (WIDH), is proposed. The opeartor can effectively utilize intensitydifference information of the local area around the feature to construct descriptive vectors.The experimental results show that the descriptor has good invariance to luminance change,structure blurring, geometrical distortion and noise disturbance, and possesses lowdimensionality but strong discrimination. Subsequently an automatic registration algorithmfor near-infrared and visible remote sensing images is proposed based on blob features andWIDH operator. The algorithm detects blob features using SURF detector, costructsWIDH description vectors for the blobs around their local neighborhoods, and identifiescorresponding relationships between the vectors based on the nearest neighbor searchalgorithm in terms of the Euclidean distance criterion. Finally the incorrectcorrespondences are removed by a Random Sample Consensus step and then thetransformation parameters are determined. The experimental result shows that theproposed algorithm can realize the registration of near-infrared and visible images moreeffectively than the SURF-based algorithm. (2) An automatic registration algorithm for infrared-visible remote sensing images isproposed based on junction point matching. At the feature detection step, a novel junctionpoint detection and characterization algorithm is prosposed based on azimuth consensus.The extracted junction points possess not only position information, but also local structureinformation, such as the number and slope angles of branch edges. The experimentalresults show that the algorithm has high accuracy in junction localization, and haspreferable robustness to noise disturbance and contrast change. At the junction pointmatching step, the conception of local structure compatibility is proposed to verify theconsistency degree of the number and slope angles of branch edges between two junctionpoints. As constraint terms, Local structure compatibilities are embedded in the posteriorprobability computation of GMM (Gaussian Mixture Models) components. This canreinforce the algorithm’s resistibility to the disturbance caused by noises and outliers, andcan improve the convergence speed of the algorithm.(3) An automatic registration algorithm for infrared-visible remote sensing images isproposed based on parameter step estimation. The parameters of affine transformationmodel is separated into some more easily estimated factors using matrix orthogonaldecomposition method, which are skew, scale ratio, rotation, scaling and translations in xand y directions. These six factors are categorized into to two groups. The first groupincludes skew, scale ratio and rotation, and the second group includes scaling andtranslations in x and y directions. Subsequently the values of the parameters in two groupsare estimated step by step using segment features extracted in the images. At the first step,an objective function with the skew, scale ratio and rotation factors as variablesconstructed based on the consensus model of the segments’ orientations. The optimalsolutions are obtained by an optimization method at the given parameter space. After thefirst step, the affine transformation is simplified to the similarity transformation. At thesecond step, another objective function, with respect to scaling and translations in x and ydirections, is constructed by the alignment degree between segments. The optimalparameter values of the second group are computed by the same optimization method asthe fisrt step. A hybrid strategy is adopted in the optimization process, which is combinedby the simulated annealing algorithm and the improved Powell algorithm. Theoptimization strategy guarantees the globality and accuracy of the results effectively. Theproposed algorithm simplifies the6D parameter space of the affine model to two3Dspaces through the parameter step estimation method, which can reduce the size of thesearching space significantly, and can effectively lower the properbability of theoptimization process trapped into local optimum.(4) An automatic registration algorithm for infrared-visible remote sensing images isproposed based on optimal mapping of edge structures. The algorithm selects thecontinuous edge structures in the image as registration elements, which can suppress the influence of noises and inconsistent information. The registration problem is transformedinto an optimization problem in the parameter space by constructing an alignment measurefunction with respect to the edge structures. A similar hybrid framework is employed in theoptimization process. The approximate global optimum is obtained by the parallel geneticalgorithm, which is followed by the improved Powell algorithm to refine the solutionlocally. To improve the efficiency of the registration algorithm, a flexible framework isproposed, which combines hierarchy approximation models in the two-scale space. Affineand projective transformation models are used in the coarse and fine scales respectively.The proposed algorithm has good adaptability and robustness due to the use of morecommon structure features in images and the integration of more flexible projective model,and reduces the probability of misregistration caused by lack of the particular registrationelements.更多还原