【作者】 张倩；

【导师】 朱策；

【作者基本信息】 电子科技大学 ， 信号与信息处理， 2017， 硕士

【摘要】 角点作为计算机视觉、图像处理技术领域中最为常用的局部特征之一,被广泛地应用于各种视觉任务之中,如视觉追踪、目标识别、图像匹配、三维场景重建以及图像注册等。经历了几十年的发展,到目前为止,角点检测算法大致可以被分为两大类,即基于图像灰度信息的角点检测算法以及基于图像边缘轮廓信息的角点检测算法。相对于基于图像灰度信息的角点检测算法,基于图像边缘轮廓信息的角点检测算法因其具有较高定位精度以及较低的错检率而受到广泛的关注。近十多年间,研究者们提出了大量的基于图像轮廓信息的角点检测算法。其中,以CSS(曲率尺度空间)角点检测算法最为著名。并且在此基础上,衍生出很多经典的基于轮廓的角点检测算法。但这类算法通常采用轮廓曲率计算角点响应函数,导致角点响应函数对噪声和局部琐碎细节比较敏感。因此,一些研究者相继采用其他的测度方式来计算图像轮廓上点的角点响应函数值,如CPDA、CTAR、DoG角点检测算法等。虽然这些算法在一定程度上解决了基于曲率尺度空间的角点检测算法的不足,但也存在一些其他方面的缺陷,如计算效率、算法性能等。鉴于此,本文首先系统地探讨了基于图像边缘轮廓信息的角点检测算法,分析各种算法的优缺点。并针对现有算法存在的不足,主要做出了以下两个方面的学术贡献:(1)改进CPDA角点检测算法并由此提出ACRA角点检测算法,使其算法的性能和计算效率相比CPDA算法而言有极大的提升。(2)通过分析图像轮廓高阶差分图的基本规律,提出两种利用轮廓二阶差分进行角点检测的算法SODC-E和SODC-M。其中SODC-M角点检测算法采用曼哈顿距离来计算角点响应函数,在性能与其他经典的基于轮廓的角点检测算法如CSS、MSCP、CPDA、CTAR等相当的情况下,其计算效率有极大地提升。与算法CPDA和CTAR一样,算法SODC-E采用欧式距离来计算角点响应函数,在计算效率与当下最优的算法CTAR以及CCR相当的情况下,其检测性能有极大地提升。更多还原

【Abstract】 After decades of development,until now corner detection algorithms can be devided into two kinds: intensity based methods and contour based methods.Compared with intensity based ones,contour based methods reveal their superiority in corner localization accuracy and earn lots of attention.In recent decades,researchers proposed a lot of contour based corner detectors,in which CSS(Curvature Scale Space)algorithm is the most popular one,and there are many algorithms derived from CSS method.However,most of them utilize curvature of contour to measure the corner response value,resulting in the sensitivity to the noise.Hence,many researchers try to explore other measures to calculate the corner response function,such as CPDA,CTAR and DoG methods.Although these methods solve the deficiency of CSS based algorithms in some extent,there are other kinds of defects,such as the low computational efficiency.So in this thesis we systematically introduce the characteristics of the state-of-theart contour-based corner detectors and analyze the advantages and disadvantages of them.We make academic contributions in two aspects:(1)Propose the ACRA corner detector based on CPDA algorithm,successfully increasing the corner detection performance and computational efficiency.(2)Propose two original corner detector SODC-M and SODC-E based on second order difference of contour in the image.We analyze the regularity of second order difference map of contour and utilize Manhattan distance and Euclidean distance to measure the corner response value respectively in the two methods.Compared with other nine corner detectors,our SODC-M achieves higher computational efficiency and our SODC-E obtains higher detection performance.更多还原