SIA OpenIR  > 光电信息技术研究室
Alternative TitleImage Matching Based on Affine Invariant Features
Thesis Advisor李德强
Keyword最大稳定极值区域(Mser) 尺度不变特征变换(Sift) 尺度不变 仿射不变
Call NumberTP391.41/H75/2008
Degree Discipline模式识别与智能系统
Degree Name硕士
Degree Grantor中国科学院沈阳自动化研究所
Place of Conferral沈阳
Abstract图像匹配是计算机视觉中的一个重要研究领域,无论在民用还是军用上都有着重要的应用价值。本文以研究室国防重点预研究项目自动目标识别为背景,采用图像匹配方法,实现飞行器定位导航。具体工作流程是:事先利用侦察手段获取飞行器途经下方的地物景象(基准图)并存于飞行器载计算机中,然后当携带相应传感器的飞行器飞过预定的位置范围时,拍摄当地的地物景象(实时图),将实时图和基准图在飞行器载计算机中进行匹配比较,可确定当前飞行器的准确位置,完成定位导航功能。由于对同一场景使用相同或不同的传感器(成像设备),以及在不同条件下(天候、照度、摄像位置和角度等)成像的复杂性和多样性等困难的存在,传统的相关匹配方法对上述困难的克服在方法原理上存在先天不足,所以无法胜任。故本文采用的方法是基于局部不变量特征的图像匹配。局部不变量特征因为能更灵活地描述图像,有效地处理图像复杂和遮挡问题,所以基于局部不变量特征的图像匹配方法对于视点的大变化,图像背景变化,以及目标场景识别等都有较好的效果。基于局部不变量特征的图像匹配方法的步骤通常分为三部分:(1)用图像区域检测算子提取图像相关区域,(2)构造合适的特征描述区域,(3)选择特征相似度度量准则实现图像区域特征的匹配。本文详细研究了最大稳定极值区域 (MSER)方法,在此基础上进行了改进,具体工作如下:(1)利用高斯核函数对图像平滑采样,建立图像的高斯尺度空间,(2)在图像的高斯尺度空间中,利用MSER检测算子检测出图像在不同尺度下的所有仿射相关区域,(3)由于区域不规则,再用仿射不变的椭圆拟合并归一化,这时所有的区域仅存在旋转的不同,(4)用SIFT特征描述图像区域,得到所有区域的128维特征向量集。(5)采用欧式距离度量特征间的相似度,以最近邻和次近邻的比值作为特征匹配准则进行匹配。本论文的主要研究工作在于把图像的高斯尺度空间引入到MSER算法中,进而大大改善了MSER算法对于图像的尺度变换、仿射变换以及图像模糊的性能。由于建立了高斯尺度空间,增加了MSER检测算子检测的范围,所以使得改进算法的性能得到了改善。论文第四章给出四组实验,分别为尺度变换,仿射变换,图像模糊和大视点变换。最后通过对匹配结果正确数量和错误数量的统计,论证了改进方法的性能要好于MSER算法。通过对算法复杂度的分析,得出虽然在改进算法引入了图像的高斯尺度空间,但是算法复杂度却并未增加,与MSER算法相同,为O(nloglogn)。
Other AbstractImage matching is a very important research issue in the field of computer vision, and it is of great important value for both civilian and military area. This thesis studies image matching supported by the laboratory national defense project, with purpose of realizing navigation for aerocraft using automatic target recognition technique. The basic process of image matching can be described as follows. Firstly, acquire the reference image and store it to the computer ROM mounted on the aerocraft. As the aerocraft flies over the target area, the camera on the aerocraft acquires the real time image below it. Finally, the algorithm matches the real time image with the reference one. Mathcing result enables the aerocraft to fulfill the location and navigation. Specifically, image matching techniques compare one region in reference image to another region from real-time image which is possibly acquired from various viewpoint or sensors. Because of the adverse and unstable environment in image acquisition, it is very difficult to fulfill the matching for some traditional methods, such as template matching and relative matching. The method studied in this thesis is based on the local invariant features which are robust and flexible to describe image characteristics. Local invariant features based method can be applied under many adverse and complex conditions even in the case of broad viewpoint or conclusion. The local invariant feature based method consists of three steps, (1) detection of affine covariant regions, (2) description of the affine covariant regions feature, and (3) similarity evaluation on features obtained from different images. The major research work of this thesis is that we propose an image matching method (Scale-MSER) based on maximally stable extremal regions (MSER). The Scale-MSER method includes five procedures. (1) Build the Gaussian Scale Space with the Gaussian kernel. (2) Detect affine covariant regions with the MSER detector in the Gaussian Scale Space. (3) fit all detected regions with ellipse. (4) describe all the regions with the SIFT feature. (5) measure the similarity of all the features based on Euclidean distance, and the ratio of the closest neighbor to that of the second-closest neighbor is regarded as the matching criterion. To summarize, the contribution of Scale-MSER algorithm is that we introduce the Gaussian Scale Space into the MSER algorithm that can improve the performance in the case of scale transform, affine transform and image blur. Because of building gauss scale space of images, expanding the set of detected maximally stable extremal regions, make the MSER algorithm possible to perform well. In the chapter four we demonstrate the conclusion with four experimental results. They are respectively scale transform, affine transform, image blur and big scale and affine transform. The algorithm complexity of the Scale-MSER method is the same as the MSER proposed by Matas, they are both O(nloglogn).
Contribution Rank1
Document Type学位论文
Recommended Citation
GB/T 7714
黄杰华. 基于仿射不变特征的图像匹配[D]. 沈阳. 中国科学院沈阳自动化研究所,2008.
Files in This Item:
File Name/Size DocType Version Access License
10001_20052801472802(3188KB) 开放获取--Application Full Text
Related Services
Recommend this item
Usage statistics
Export to Endnote
Google Scholar
Similar articles in Google Scholar
[黄杰华]'s Articles
Baidu academic
Similar articles in Baidu academic
[黄杰华]'s Articles
Bing Scholar
Similar articles in Bing Scholar
[黄杰华]'s Articles
Terms of Use
No data!
Social Bookmark/Share
All comments (0)
No comment.

Items in the repository are protected by copyright, with all rights reserved, unless otherwise indicated.