【作者】 王军伟；

【导师】 刘文予；

【作者基本信息】 华中科技大学 ， 通信与信息系统， 2012， 博士

【摘要】 形状是一种高层次的视觉特征,在计算机视觉、图像分析与理解等领域中常用于表达目标物体。形状的关键性和重要性,使得形状表示和形状匹配成为计算机视觉领域中的基础性问题。虽然由于日常生活中同类物体形状的复杂多变性,使得形状匹配面临着许多困难,但长期以来在学者们的不懈研究下,已产生出一大批经典的形状表示技术和形状匹配算法。到目前为止,形状匹配仍然是计算机视觉和图像分析领域的前沿和热点问题。本文考察了形状匹配技术的发展历程,并发现：目前已有的形状描述子均未能简单有效的将局部与全局形状信息结合起来。针对这一问题,本文分别从局部特征和全局特征的定义出发丌展研究,并提出了一种能够将这两者有机结合起来的框架,定义了性能上更为优越的轮廓描述子。本文的研究成果主要有：(1)提出了有效的局部特征局部特征是定义形状描述子的基础,它不仅需要准确的描述形状的特性,而且要在线性变换(平移、旋转、尺度缩放)下保持不变。本文从轮廓点的相对几何关系出发,从不同的角度分别定义了数种有效的局部特征,如特征三角形、高度函数等。这些局部特征都能够较为准确的表示形状,并且能够在线性变换下保持不变性,从而为定义有效的形状描述子打下了坚实的基础。(2)提出了有效的全局特征全局特征是形状描述子所蕴含的重要信息,它直接关系到描述子对噪声或局部形变等因素的抵御能力。本文通过借鉴形状匹配算法的经验,创造性的将轮廓的顺序关系定义为全局形状特征,并将它与局部特征有机结合起来。本文是轮廓的顺序关系首次作为形状特征在形状描述子中得到使用和体现。事实证明,轮廓的顺序关系不仅能够作为约束条件应用于形状匹配算法中,也完全可以作为形状特征融合在形状描述子中,并能显著提高描述子的描述性能。(3)提出了局部与全局形状特征的有机结合框架只有将局部与全局形状特征相结合,才能使所定义的描述子兼具有刻画形状的准确性和对抗噪声的稳定性。本文提出一种新的描述子定义框架,该框架能够将局部与全局特征有机结合在一起。我们发现,利用轮廓的顺序关系对局部特征进行处理,将局部特征根据轮廓顺序关系排列成一个序列,能够自然的将这两种性质不同的特征有机结合起来。与多尺度描述子和多方面描述子相比,本文所定义的描述子能够有效克服已有描述子不易使用或信息不全等问题,本文方法同时含有局部与全局信息,且简单易用。(4)提出了三种新型的有效的形状描述子在特征三角形、形状上下文、高度函数等有效的局部特征基础上,结合轮廓的顺序关系,构造了三种新的轮廓描述子。实验结果表明,这些方法均能取得较高的形状检索准确率,特别是基于高度函数的轮廓描述子,该方法能够取得到目前为止所有描述子中国际上最高的形状检索准确率(在标准测试集MPEG-7数据库上,仅靠描述子的检索精度可达90.35%,结合图转导算法后检索精度可达96.45%)。同时,这些描述子几何意义明确,易于计算、维度低,其综合性能超越了常用的轮廓描述算法。本文对基于轮廓的形状表示技术展丌了深入研究,敏锐的抓住了轮廓的顺序关系这一核心而本质的全局形状特征,通过将不同的局部特征与轮廓顺序信息相结合的方式,提出了数种几何意义明确、描述性强、表达准确、计算复杂度适中的轮廓描述子,在形状匹配这一计算机视觉领域的关键问题上取得了进展,相关成果已经得到了大量的实验证明。本文所提出的算法对于其它视觉理论及应用也有一定的启发性。更多还原

【Abstract】 Shape, as a high-level visual characteristic, has been frequently used to represent the properties of objects in Computer Vision, Image Analysis and Understanding. Shape is a kind of critical and fundamental feature, and shape representation/matching has become one basic problem in the field of Computer Vision. Because of serious intra-class variances among object shapes in our life, there are a lot of difficulties when performing shape matching. However, a huge number of classical shape representation technologies and shape matching algorithms have appeared by hard works of researchers. Shape matching is still one of the leading and hot topics in Computer Vision and Image Analysis so far.In this thesis, the development of shape matching technologies is reviewed, and it is found that none of current descriptors is able to combine local and global features effectively and efficiently. Accordingly, some researches are performed on how to define local and global features, and a framework to naturally combine these two issues is presented, resulting in some novel descriptors with better performances. The main contributions of this thesis are summarized as follows:First, some effective local features are proposed. Local features are fundamental for defining shape descriptors. Local features should give a precise description for shape contours, and remain unchanged under linear transformations (translation, rotation, and scaling). Based on the geometric relationship between contour points, several local features, such as feature triangle and height functions, are defined to represent different aspects of shape contours. These local features are able to represent shapes exactly, and achieve linear transformation invariance. They provide a solid basis for presenting effective shape descriptors.Second, some effective global feature is proposed. Global features are important issues contained in shape descriptors. They directly determine the robustness of shape descriptors to noise and local deformations. Based on the experience of shape matching algorithms, the order sequential information of contours is creatively defined as the global feature. This is the first time for contour order information to be treated as shape feature used in shape descriptors. It is proved that the contour order information can be used not only as constraints in shape matching algorithms but also as shape features in shape descriptors. It is able to help significantly improve the performance of descriptors.Third, a new framework to combine local and global shape features is presented. Shape descriptors will never become not only description precisely but also robust to noise until both local and global features are combined together. It is discovered that after local features are rearranged according to the contour order information, the local and global features are naturally combined. Compared with current multi-scale or multi-aspect descriptors, the proposed methods are able to overcome the problems of usage difficulties and lack of shape information. The proposed descriptors include both local and global shape information, and they are easy to use and compute as well.Fourth, three novel shape descriptors are presented. Based on the local features of feature triangles, shape contexts and height functions, combined with contour order information, three novel shape descriptors are constructed correspondingly. The experimental results show that the proposed methods are all able to achieve excellent shape retrieval rates. Especially for height functions, it is able to get the highest accuracy rate of shape retrieval among all descriptors all over the world. It achieves on the well-known MPEG-7 shape benchmark the best ever high rate of 90.35% only by the descriptor and 96.45% when combing with the graph transduction algorithm. What is more, each of these descriptors has a definite geometric meaning, is efficient to compute, and has a low feature dimension. The general performances of them outperform other widely used shape description algorithms.In this thesis, contour based shape representation is carefully studied and the contour sequential information (a kind of global feature) is regarded as the core and fundamental shape characteristic. Several contour descriptors with clear geometric meaning, high discriminative power, precise representation and moderate computational complexity are presented by combining the sequential information with different local shape features. Some successes are obtained for shape matching, which is one of the key problems in the field of Computer Vision, and these successes are already confirmed by extensive experimentations. Some vision theories and applications can also benefit from the studies and algorithms proposed in this thesis.更多还原