基于图理论的图像特征匹配算法研究

发布时间：2018-09-11 07:00

【摘要】：图像匹配是指将不同时间、不同成像条件下获取的的两幅或多幅图像进行空间上的对准,确定图像之间的几何映射关系,进而使得图像能够匹配的过程。作为计算机视觉的核心技术之一,图像匹配是图像分析与处理中的基础问题。图像匹配在目标对象识别、纹理发现与分析、图像信息融合、图像检索等领域的应用越来越广泛,具有十分重要的研究意义。基于特征的图像匹配算法由于对图像的尺度变化、仿射形变等具有良好的稳定性和鲁棒性,受到了国内外学者的广泛关注。图模型作为一种描述数据的工具,可以有效的表示图像的结构特征,同时保留区域之间的相互联系,利用图模型来实现图像特征点匹配的研究受到了学术界的青睐。基于图理论的图像特征点匹配方法,由于具有较好的适应性和较高的匹配精度,是近年来研究的热点和难点问题。本文围绕基于图理论的图像特征匹配方法进行了相关研究,主要研究内容和研究成果如下:(1)研究分析了图像匹配的理论意义和实用价值,对国内外关于图像匹配的研究现状进行了概括和总结。重点对图像特征匹配进行了理论方面的概述,首先重点介绍了图的基本概念和矩阵表示,然后介绍了图像特征匹配中的两个关键技术:特征提取和特征描述,最后对经典的SIFT图像特征匹配算法进行了妔}0的分析。图的相关理论和对SIFT算法的研究,为本文图像匹配算法的提出奠定了重要的理论基础。(2)针对图像特征点匹配,结合层次聚类的思想,本文给出了一种基于自顶向下分裂聚类的图像匹配算法。该算法的主要思想是采用互k近邻图模型来表示图像之间的对应关系,在互κ近邻图表示模型中,顶点代表特征点之间的对应关系,顶点之间的边代表对应关系的几何相容性。定义的团密度函数可以衡量是否属于同一个团,一般情况下,团密度的值越大,越有可能是正确的团。该算法不仅可以获得图像之间的对应关系,还可以指示出哪些对应关系属于同一个目标。同一个团内的对应关系之间几何相容性较高,不同团之间的对应关系相容性则较低,因此不同的目标会呈现出不同的团。在互k近邻图表示模型的基础上,通过团检测方法获得图中的团,利用的是分裂聚类的思想。最终,根据团内包含的顶点恢复出团内的对应关系,从而达到图像匹配的目的。在真实图像上的对比实验表明,自顶向下分裂聚类的图像匹配算法在匹配性能上要优于ACC算法,提高了图像匹配的查全率和查准率,实验的效果图和定量分析结果都表明该算法具有较好的匹配结果。(3)为了进一步提高图像特征匹配算法的准确度,本文提出了一种基于局部近邻图的特征描述与特征匹配算法,通过为每个特征点构建局部近邻图来深层次挖掘图像上的结构信息。该算法首先通过FAST和SURT算法检测初始的特征点,然后为所有的特征点构造局部近邻图,每个局部图由该特征点及其近邻特征点组成,至此形成一种新颖的特征描述方法。在这个新颖的特征描述符的基础上,给出了一个相似性度量函数和一个能量函数,鉴于此,提出了一种基于局部近邻图模型的特征匹配算法。为了验证该算法的有效性,进行了两个方面的实验:高斯噪声模拟实验和真实图像匹配实验。高斯噪声模拟实验的目的是为了分析离群点和变形噪声对算法性能的影响,而在真实图像库上进行实验,是为了验证该算法在图像特征匹配中的准确度。实验的实例图和定量分析结果表明,基于局部近邻图的特征匹配算法较SM算法具有一定的优越性。
[Abstract]:Image matching refers to the process of spatial alignment of two or more images acquired under different imaging conditions at different times, to determine the geometric mapping relationship between images, and then to make the image matching. As one of the core technologies of computer vision, image matching is a basic problem in image analysis and processing. Matching is more and more widely used in object recognition, texture discovery and analysis, image information fusion, image retrieval and other fields. Feature-based image matching algorithm has been widely used by scholars at home and abroad because of its good stability and robustness to image scale change, affine deformation and so on. Graph model, as a tool for describing data, can effectively represent the structural features of an image while preserving the relationship between regions. The research on feature point matching based on graph model is favored by academia. High matching accuracy is a hot and difficult problem in recent years. This paper focuses on the image feature matching method based on graph theory. The main research contents and achievements are as follows: (1) The theoretical significance and practical value of image matching are analyzed, and the research status of image matching at home and abroad is summarized. In the end, the classical SIFT image feature matching algorithm is analyzed by_} 0. The related theory of graph is analyzed. The research of SIFT algorithm has laid an important theoretical foundation for the proposed image matching algorithm. (2) Aiming at image feature point matching and combining the idea of hierarchical clustering, this paper presents an image matching algorithm based on top-down splitting clustering. In the mutual kappa nearest neighbor graph representation model, vertices represent correspondence between feature points and edges between vertices represent geometric compatibility of correspondence. The defined clique density function can be used to measure whether a clique belongs to the same clique. In general, the larger the clique density, the more likely the clique is to be correct. The correspondence between images can also indicate which correspondence belongs to the same target. The geometric compatibility of correspondence in the same clique is higher, but the correspondence between different cliques is lower. Therefore, different targets will present different cliques. Methods The clique in the graph was obtained by using the idea of split clustering. Finally, the corresponding relationship in the clique was recovered according to the vertices contained in the clique, so as to achieve the purpose of image matching. (3) To further improve the accuracy of image feature matching algorithm, a feature description and feature matching algorithm based on local nearest neighbor graph is proposed, which constructs local nearest neighbor graph for each feature point. The algorithm first detects the initial feature points by FAST and SURT algorithm, and then constructs a local neighborhood graph for all the feature points. Each local graph consists of the feature points and their neighborhood feature points, thus forming a novel feature description method. Based on this, a similarity measure function and an energy function are given. In view of this, a feature matching algorithm based on local nearest neighbor graph model is proposed. In order to verify the effectiveness of the algorithm, two experiments are carried out: Gaussian noise simulation experiment and real image matching experiment. The effect of outliers and distortion noise on the performance of the algorithm is analyzed, and experiments on real image database are carried out to verify the accuracy of the algorithm in image feature matching.The experimental results show that the feature matching algorithm based on local nearest neighbor graph is superior to SM algorithm.
【学位授予单位】：安徽大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】