CASIA OpenIR  > 毕业生  > 硕士学位论文
图像对应点确定及其应用
其他题名Image Point Correspondence and its Applications
万崇玮
学位类型工学硕士
导师胡占义
2007-05-26
学位授予单位中国科学院研究生院
学位授予地点中国科学院自动化研究所
学位专业模式识别与智能系统
关键词对应点 图像匹配 局部特征 Sift 匹配扩散 视频镜头检测 遥感影像配准 Point Correspondence Image Matching Local Feature Sift Match Propagation Video Shot Detection Remote Sensed Image Registration
摘要图像之间对应点的确定问题是计算机视觉中的一个关键问题,在目标检测、物体识别、三维重建、图像配准、视频理解等具体应用中发挥着重要作用。由于图像类型和几何形变复杂多样,而且具体应用需求各有不同,使得对应点匹配问题成为一个极具挑战性的热点研究课题。 本文的工作主要是围绕如何确定图像之间的对应点以及如何将这些匹配算法应用于计算机视觉的一些具体问题展开的,主要工作有: 针对图像在复杂形变下的匹配问题,实现了一个基于SIFT(Scale Invariant Feature Transform)特征的匹配软件。整个匹配过程主要分为三个步骤:首先采用DoG(Difference-of-Gaussian)算子在图像的尺度空间中检测关键点并进行精确定位,然后计算关键点的主方向并构造特征描述子,最后根据描述子之间的欧氏距离来度量图像特征点之间的匹配程度。基于SIFT特征的匹配方法能够很好地适应图像之间的复杂形变,并且具有很强的辨别能力。我们的软件经过大量实验测试,具有很好的稳定性。 针对三维重建等需要大量高精度对应点的问题,实现了一个对应点扩散的软件。该软件能够得到具有子像素级精度且分布均匀的“准稠密”匹配点。本文在“准稠密”匹配的实现中,采用了一种鲁棒的初始点匹配方法来选择种子匹配,将SIFT特征与几何约束相结合,由粗到精地获取可靠的特征点匹配,并用灰度相关算法进一步提高匹配点的精度。该软件已应用于大场景的三维重建中,获得了逼真的重建效果。 将基于SIFT特征匹配的方法应用于视频镜头检测,提出了两种帧间匹配机制:两两匹配机制和连续匹配机制,通过分析帧间SIFT特征匹配点数目随时间变化的特性来检测镜头的边界。这种镜头检测方法将突变和各种类型的渐变统一于一种检测模式,避免了模型的选择和参数的调整。 将SIFT特征应用于遥感图像匹配。我们采用了基于SIFT描述子的匹配方法来确定初始控制点,并通过对仿射模型参数进行优化以得到精确的控制点位置,然后依据得到的控制点计算控制区域的配准模型。另外,我们还将对应点扩散算法初步应用到航空影像的匹配。
其他摘要Image point correspondence is a key problem in many applications of computer vision, such as object detection, object recognition, 3D reconstruction, image registration and video understanding. Due to a large family of image types, various shape distortions, and diversity of applications, the correspondence problem has been a challenging problem for a long time. In this dissertation, the main work is focused on the implementations of some state-of-art algorithms, as well as their applications. It can be summarized as: The SIFT (Scale Invariant Feature Transform) is implemented for image point correspondences. In this algorithm, firstly keypoints are detected by DoG (Difference-of-Gaussian) operator and refined in the image scale space. Then a dominant orientation is assigned and a local image descriptor is computed for each keypoint. Finally keypoints are matched using the Euclidean distance between descriptors. SIFT features are invariant to complicated image transformations and distinctive enough for matching. Experimental tests demonstrate the stability of our SIFT implementation. In some applications like 3D reconstruction, a large quantity of high accuracy corresponding points are needed. To this end, a match propagation software is implemented to get Quasi-Dense point correspondences with even distribution and sub-pixel accuracy. In this thesis, a coarse to fine matching strategy is adopted to obtain seed matches, which combines SIFT features with some global geometric constraints. In addition, a correlation method is adopted to refine the point into sub-pixel accuracy. Obtained seed matches can be used for propagation of reliable point correspondences. The software is applied to 3D reconstruction of wide-ranging scenes, and satisfactory results are obtained. A SIFT feature based matching method is proposed for video shot detection, and two different measure schemes are introduced, namely pair-wise based matching and chain matching. Shot boundaries are detected by analyzing temporal evolution of the number of matched SIFT features across frames. In this method, all kinds of shot transitions are detected within the same scheme, with the advantages of avoiding model selection and parameter adjustment. SIFT features are applied to remote sensed image matching. Control points are computed using SIFT descriptors and refined through optimization under an affine model. Then the refined control points are used for control region registration. Besides, match propagation is applied to aerial image matching, and preliminarily encouraging results are obtained.
馆藏号XWLW1061
其他标识符200428014628007
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/7397
专题毕业生_硕士学位论文
推荐引用方式
GB/T 7714
万崇玮. 图像对应点确定及其应用[D]. 中国科学院自动化研究所. 中国科学院研究生院,2007.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
CASIA_20042801462800(10608KB) 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[万崇玮]的文章
百度学术
百度学术中相似的文章
[万崇玮]的文章
必应学术
必应学术中相似的文章
[万崇玮]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。