【作者】 李壮；

【导师】 于起峰；

【作者基本信息】 国防科学技术大学 ， 航空宇航科学与技术， 2011， 博士

【摘要】 在基于视觉的飞行器导航与制导、遥感卫星灾害监控与环境监测、医学图像分析等应用中,常常需要对不同类型传感器获取的异源图像进行匹配。图像传感器本身结构、成像原理等方面的不同,导致异源图像上对应区域的灰度、对比度都存在较大的差异。因此,异源图像匹配是一项非常有难度的工作。本文以飞行器视觉导航和景象匹配制导为研究背景,以开发可靠的和实时的的异源图像匹配算法为研究内容,对异源图像匹配中的关键技术进行了重点研究。论文主要包含以下五部分内容:第一部分研究自适应NL-means滤波器。对常用的图像去噪方法进行了综述,介绍NL-means滤波器原理,提出自适应NL-means滤波算法。自适应NL-means滤波器不仅能够对未知噪声分布的输入图像进行滤波,而且其滤波效果常常优于传统NL-means滤波器。另外,自适应NL-means滤波器还能有效滤除SAR图像中的斑点噪声。第二部分研究基于引力模型的边缘匹配方法。对常用的边缘匹配方法进行介绍,针对传统方法易受图像噪声和图像畸变影响的问题,提出基于引力模型的边缘匹配方法。作者分别定义了一点到一点、一点到点集、点集到点集的引力矢量。引力的强度描述了待匹配边缘的相似性,而引力的方向为匹配算法在参数空间搜索提供了有效的引导信息。作者还通过在点集引力模型中加入边缘法向一致性权值,进一步提高了边缘匹配的可靠性。第三部分研究基于空间子区一致性的图像匹配方法。空间子区一致性是对图像结构特征的描述,它由图像内部相邻子区之间的相似性构成。与传统的图像梯度特征相比,空间子区一致性特征受图像噪声影响较小,适合对成像质量较差的异源图像进行匹配。第四部分研究图像中的旋转不变特征,提出一种新的旋转不变的图像描述方法——多尺度梯度径向夹角直方图(Multi-scale Histogram of Angle of Radius and Gradient, MHARG)。MHARG特征不仅具有旋转不变性,还对图像灰度级非线性变换和图像噪声有很强的适应性,因此适用于异源图像匹配。利用MHARG特征进行图像匹配,可以避免在旋转空间内的搜索问题,大幅缩短图像匹配算法的运算时间。第五部分研究基于控制区约束的异源图像精匹配方法。主要工作包括:1)提出基于方向Moran信息的控制区可匹配性评估方法。2)针对异源图像的局部区域匹配可靠性和精度较低的问题,提出一种鲁棒求解变换参数的方法。3)提出对基准图进行几何校正,并在校正后的图像上进行精匹配的迭代匹配方法。更多还原

【Abstract】 Matching multi-sensor images is an important work for vision based navigation and guidance, remote sensing based environmental monitoring and medical image analysis. In multi-sensor imagery, the relationship between the intensity values of the corresponding pixels is complex and unknown. Contrast reversal may occur in some regions, and the contrasts of the images may differ from one another. The multiple-intensity values in one image may map to a single intensity value in another. Further, features present in one image may not appear in another, and vice versa. The matching of multi-sensor images thus constitutes a challenging problem.In this dissertation, the vision navigation and guidance based on scene matching are taken as the research background, and the robust and real-time multi-sensor images matching system is taken as the research content. The emphatically work of this dissertation is on the key technologies of multi-sensor images matching.The kernel content of this dissertation is composed of the following five parts.In the first part, the adaptive NL-means filter is studied. The traditional image denoising methods are reviewed. The NL-means filter is introduced and a novel adaptive NL-means filter is proposed. The adaptive NL-means filter can denoise image with unknown noise distribution. And, in most cases, the adaptive NL-means filter can get better denoising result than original NL-means filter. In addition, the adaptive NL-means filter can also be applied on SAR images with speckle noise effectively.In the second part, the gravity model based edge matching method is studied. Traditional edge matching methods are introduced. As these methods are susceptible image noise and deformation, the author proposes a more robust method based on gravity model. The gravity vectors between point and point, point and point set, point set and point set are defined. The magnitude of gravity vector describes the similarity of the edge sets, while the direction of gravity vector can alleviate the search complexity. The author powers the gravity model by the consistency of normal direction of edge to further improve the reliability of edge matching.In the third part, the spatial structure feature is studied and a novel spatial sub-region congruency feature is proposed. The spatial sub-region congruency feature is constructed by the similarities among adjacent sub-regions. It can describe the structure of image while not have to detect edges or local maximums. Compared to traditional gradient based features, the spatial sub-region congruency feature is more robust to image noise, and hence fit to match multi-sensor images with inferior imagery quality.In the fourth part, rotate-invariant features are studied. A novel rotate-invariant feature, the Multi-scale Histogram of Angle of Radius and Gradient (MHARG) is proposed. The MHARG is not only invariant to image rotation, but is very robust to the nonlinear transform of the gray level and image noise. So, it is naturally suit for multi-sensor images matching. When matching by MHARG feature, the search process in the rotation space can be saved, thereby reducing the computational time substantially.In the fifth part, the refined matching on the basis of control region is studied. The main contributions are as follows. (1) A novel method to evaluate the ability of control region to be used for matching is proposed based on oriented Moran information. (2) As the reliability and precision of multi-sensor images matching are always lower than mono-sensor images matching, the author proposes a novel method for robustly solve the transformation parameters. (3) The author geometrically corrects the reference image and carries the refined matching on the corrected image iteratively. The iterative refined processes can further improve the matching precision.更多还原