【作者】 曹红杏；

【导师】 阮萍；

【作者基本信息】 中国科学院研究生院（西安光学精密机械研究所） ， 光学工程， 2008， 硕士

【摘要】 图像拼接是数字图像处理的一个重要研究领域,它是将具有一定重叠的两幅或者多幅图像进行匹配对准,融合形成一幅大视场图像的技术,在医学、工业、航天等诸多领域有着广泛的应用。由于成像系统特性、拍摄角度和时间的不同,以及噪声干扰和遮挡,使图像拼接变得十分困难。按照图像配准的依据不同,图像拼接分为基于特征、基于灰度信息和基于变换域的图像拼接。基于特征的图像拼接不仅不易受光照、旋转等因素影响,而且特征相对像素数量较少,有利于提高速度,因此成为最受关注的一类拼接方法。在本论文中提出了一种稳健的基于特征点的图像拼接算法。本文首先介绍了图像拼接的几何基础,并根据平面场景的两图像之间的变换关系,确定了射影变换为本文的变换模型。它将作为本文变换矩阵和融合等计算的基础。基于特征点的图像拼接主要包括特征点提取、特征点的匹配、变换矩阵的计算和图像融合四步。特征点提取是整个图像拼接过程的第一步,对于最终拼接结果至关重要。本文介绍和实现了Harrris、SUSAN、SIFT三种最为流行的特征点提取算法。由于SIFT算法对平移、尺度缩放、旋转等保持不变性,对光照变换、仿射变换、甚至射影变换也保持一定程度的稳定性,而且SIFT特征描述符具有很强的匹配能力,因此选择了稳健的SIFT算法作为特征点提取算法。在特征点匹配阶段,本文首先对两图的特征点特征描述符建立k-d树,并利用它寻找特征点在另一幅图上的最近邻、次近邻匹配特征点,进而筛选出初始匹配点对。在计算变换矩阵阶段,本文使用PROSAC方法对初始匹配点对进行过滤,得到变换矩阵的初值和内点集合,然后采用L-M算法对变换矩阵参数进行求精。最后,在图像融合时,为了防止灰度拼接缝,本文依据匹配点对的灰度差异,对拼接图像进行匀光处理,使其灰度一致,再根据得到的变换矩阵,进行加权融合。实验结果显示,最终得到的拼接效果良好。更多还原

【Abstract】 Image stitching is an important research area of image processing. It is a technology that matches a series of images which are overlapped with each other and mixes these images into one image which has big field of view and has been widely used in the medical, industrial, aerospace and other fields.Because of difference of imaging system, point of view, the time as well as noise interference and obstructions, image stitching becames very difficult. According to the different basis of image registration, image stitching is divided into feature-based, gray-information-based and transformation-domain based. The feature-based image stitching is not only vulnerable to the light, rotation, but also helpful to increasing efficiency because the number of features is quite fewer than image piexel. Therefore, it become one type of most attended methods.This paper presents a robust image stitching algorithm.based on feature points.At first, the geometric basis of image stitching is introduced, and in accordance with the transformation between the two images of planar scene, projection transformation is choosed as transformation model. So, projection transformation is the basis of computation of transformation matrix and mixing images.The image stitching based on feature points is mainly composed of four steps: feature points extraction, matching feature points, computing transformation matrix and image mixing. Feature extraction is the first step in the process of image stitching, and is also very crucial. This article introduces and implements Harrris, SUSAN, SIFT three kinds of most popular feature point extraction algorithms. Because the SIFT algorithm is invariant to image translation, scaling, and rotation, and partially invariant to illumination changes and affine or 3D projection, SIFT descriptor has strong ability of matching. So the robust SIFT algorithm. is choosed as feature extraction algorithm. In stage of feature point matching, the k-d tree of SIFT features descriptor is build for two stitching images respectively, then use k-d tree to look for the nearest and the second nearest neighbor feature point in another image for each feature point in one image, and select initial matching point pairs. In the stage of computation of transformation matrix, PROSAC algorithm is used to filter the initial matching point pairs, and get the initial transformation matrix as well as inliers, then the L-M algorithm is applied to obtain higher precision transformation matrix. In the process of image mixing,in order to prevent the gray joints, gray of image is adjusted before mixing image based on the gray differences of inliers., then, according to transformation matrix, weighted mix images. Experiment results show that the quality of final stitching images.is well.更多还原