基于成像机理和人类认知的异源图像匹配新方法研究

There is an urgent requirement for high reliable multi-sensor images matching algorithm which is used for all-weather scene navigation and guidance. Starting from the imaging principle and the human's cognition mechanism for multi-sensor image, this project researched a novel multi-sensor images matching algorithm based on salient edges formed between objects of different material. According to the forming mechanism of salient edges, a novel salient edges detection method was proposed based on possibility fuzzy C-means clustering and segmentation. According to the idea of data reduction and the principle that kernel-space is linear separable, its rapid calculation was researched. On this basis, in order to fundamentally solve impact on the matching caused by noises such as local mutation and edge burr, a salient edge description method was put forward on the basis of data compression-Douglas-Peucker algorithm. In view of edge deformation caused by different angle of view and affine transformation, this project described the feature based on local scale, this describing method has invariance to deformation, rotation and scale. Then non-uniform sampling method was used to fuse local information and global information, which is not affected by initial point. In order to further improve the matching reliability, a contour matching algorithm was raised based on improved DTW by scratchable latex path constraint conditions. This project will provide a general practical theory and core method for high reliable multi-sensor images matching.

国家对面向全天候景象导航制导的红外、可见光和雷达图像间高可靠匹配方法有着迫切需求。本申请突破常规借鉴可见光图像匹配思想研究异源图像匹配的局限性，首次从成像原理和人类对其认知机理根本点入手，探索研究基于不同材质对象间粗大边缘的异源图像匹配新理论与新方法。根据粗大边缘形成机理，提出一种基于模糊C均值聚类鲁棒分割对象再提取其粗大边缘的新方法，同时借用数理分析中的数据约减方法大幅降低聚类分割特征空间数据量以提高实时性。另提出将数据压缩领域中具有去噪特点的道格拉斯-普克抽稀算法引入到边缘概貌描述中，以从根本上解决边缘局部突变和毛刺等噪声对匹配稳健性的影响。针对视角形变、几何变换和残缺对匹配的影响，提出基于局部距离相似稳定的抗形变和旋转的特征表征新方法，并采用非均匀采样法融合局部与全局信息以达到抗残缺和初始点无关性，又提出基于九宫格约束的DTW粗大边缘最优化匹配方法，以期得到一种可行的高可靠匹配方法。

针对同一地物的异源图像成像机理和成像条件不同，难以获得图像在灰度、亮度、颜色等特征上共性特征的问题，从人类对异源图像的认知规律入手，提出了一种稳定性好、可靠性高的粗大轮廓异源图像匹配新算法。.针对现有边缘检测算法很难快速准确地提取出实际含噪声的异源图像粗大边缘问题，研究粗大边缘形成机理，提出了一种基于核空间PFCM聚类的异源图像粗大边缘检测新方法。计算灰度和纹理特征值，并以此构建二维特征空间，根据数据约减思想，压缩参与聚类的数据集，根据核空间的线性可分思想，将约减后的样本空间映射到高斯核空间，进而采用FKPFCM算法对图像分割，以此对边缘轮廓记性跟踪。.针对现有轮廓匹配算法难以稳健地处理局部轮廓匹配及不同拍摄视角或非相似变换引起的一定形变问题，根据局部轮廓结构在产生形变时具有相对稳定性的规律，以及一个好的轮廓描述方法需要融合轮廓局部和全局信息的主流思想，研究了基于局部尺度特征的轮廓描述算法。.针对线性轮廓匹配算法匹配效果不佳以及存在几何变换或者一定程度形变的两轮廓之间并不是一对一线性关系的问题，从非线性动态规划理论出发，将轮廓匹配问题转化为最优化问题。.根据优化路径的线性度需满足采样特性需要的问题，研究了基于九宫格路径约束的DTW匹配算法。实验表明，该算法对存在几何变换及一定形变的异源图像整体和局部匹配具有较好的稳健性，匹配准确率平均约为92%，较HD算法提高了30%，较传统DTW算法提高了26%。.以上基础研究工作，为地形匹配、自主导航系统、导弹等武器的精确制导、光学和雷达的图像模板跟踪、天气预报、病理诊断研究等需求提供了理论支撑和方法指导。

内容获取失败，请点击重试