【作者】 李忠新；

【导师】 王执铨；

【作者基本信息】 南京理工大学 ， 控制科学与工程， 2004， 博士

【摘要】 图像镶嵌是一种将从真实世界获取的图像序列合成一个宽视场的场景图像技术。目前，图像镶嵌技术在摄影测量学、虚拟现实、视频图像处理等领域中得到越来越广泛的应用，具有重大的应用价值。本文对图像镶嵌技术进行了较深入的研究，主要研究成果如下： 回顾了图像镶嵌的由来、研究现状以及在各个领域中的应用，并系统地介绍了图像镶嵌的基础理论及其关键技术。 提出了一种基于块匹配的视频图像镶嵌算法，该算法首先利用基于相位相关的块匹配算法计算出图像的运动矢量场，然后根据运动矢量场确定图像子块之间的对应关系，最后通过迭代算法估计出图像间的变换模型参数来进行图像镶嵌。该算法无需从图像中提取特征，能够充分利用图像所包含的信息，在图像不存在较大的畸变以及镜头的缩放情况下可以获得良好的镶嵌效果。 提出了一种基于角点匹配的图像镶嵌算法，该算法首先采用Harris算子从图像中提取角点，然后通过归一化相关法获取匹配的角点对，再利用鲁棒的随机采样估计算法计算出图像之间的变换模型参数进行图像镶嵌。该算法能够较好的解决图像间对应点难以确定的问题，即使存在较高比率的错配点也能获得正确的变换模型参数完成镶嵌，具有较强的鲁棒性。 研究了一种基于光流场的图像镶嵌算法，根据光流约束方程建立了不同的目标函数来估计图像间的变换模型参数，并在原有算法的基础上，进一步优化了变换模型参数，以提高变换模型参数的估计精度。该算法无需提取图像的特征，能够自动的实现图像的镶嵌。经过实验表明在相机存在抖动或拍摄的图像不满足平面场景假设的条件下以及图像间存在光照变化的情况下，该算法也可以获得较好的镶嵌效果。 研究了柱面全景图像镶嵌技术，并提出了一种基于频域信息的图像镶嵌算法，利用图像的互功率谱来估计相机的像素焦距以及图像间的平移运动参数和旋转运动参数进行柱面全景图的镶嵌。该算法在图像的初始运动参数的估计和迭代求精过程中，均增加了对相机绕光轴小角度旋转的情况的考虑，有效地避免了误拼接现象的出现。摘要I.士论文最后论文分析了当前在图像镶嵌技术研究中存在的问题，并指出了进一步的研究方向。关键词:图像镶嵌，块匹配，角点匹配，光流场，柱面全景图更多还原

【Abstract】 Image mosaicing is the art of constructing scene representation with wide view field by use of image sequences shot from real world. With great practical value, now it exerts profound influence on the research and application field of photogrammetry, virtual reality and video processing. This dissertation systematically investigated the topic and its related techniques, which are further concluded as follows:The origin, development and applications of image mosaicing are reviewed first. The fundamental theory and key techniques of image mosiacing are systemically introduced.Later, an algorithm of image mosaicing using method of block matching based on phase correlation is described, in which the motion vector field is first calculated to establish the correspondences of adjacent images, then iterative estimation on parameters of transformation model is performed to create the panoramic image. The proposed method does not need to extract features of images and takes full advantage of the image information. When the geometric deformation and the changes of the scale factor among images are not serious, the method can obtain good panoramic image.Then, an algorithm based on corner matching is put forward, which constructs the panoramic images by a sequence of operations including: the extraction of the feature points by using Harris corner detector, corner matches between image pairs with normalized cross correlation method, and the estimation of the parameters of the transformation model through Robust Random Sample Consensus (RANSAC) approach by which the outliers could be effectively eliminated.An algorithm based on optical flow field is also presented in the thesis. The algorithm builds objective function upon the parameters of transformation model according to the optical flow constraints and estimates these parameters through iterative solution. It is shown to be robust to illumination variety between images and deviations from the assumption of planar scene, or the dithering of camera.The technique of cylindrical panoramic image mosaicing is investigated. The frequency domain information of images by Fourier Transform is employed to estimate the focus of camera and the motion parameters to complete the cylinderimage mosaicing. The rotation of camera around optical axis, which resulting in mis-registering between images, is further considered during the initial estimation and refining process of motion parameters.Finally, some open issues on image mosaicing as well as the further works are pointed out.更多还原