基于轮廓信息的形状匹配方法研究
文献类型：学位论文
| 作者 | 蔡慧英 |
| 学位类别 | 博士 |
| 答辩日期 | 2017-05-25 |
| 授予单位 | 中国科学院沈阳自动化研究所 |
| 授予地点 | 沈阳 |
| 导师 | 朱枫 |
| 关键词 | 形状匹配 启发式混合遗传 CHP特征 多尺度积分特征 分层射影不变特征 |
| 其他题名 | Research on Contour Information Based Shape Matching Methods |
| 学位专业 | 模式识别与智能系统 |
| 中文摘要 | 识别和定位是计算机视觉的两项重要任务，识别是在图像中找到期望物体的过程；定位是确定物体位置与姿态的过程。识别与定位的依据是物体的特征，包括颜色、纹理、结构、形状等，其中形状是常用的特征之一。形状匹配是实现物体识别和定位的重要方法，它作为计算机研究领域的热点问题之一，已有很长的研究历史，取得了很多重要的研究成果，并且已有部分成果取得了应用。尽管如此，由于形状匹配的复杂性，还存在尚未完全解决的理论问题，它目前仍然是计算机视觉领域的一个研究重点。本文对基于轮廓的形状匹配方法开展研究，具有重要的理论意义与实际应用价值。 形状匹配要考虑的因素很多，如物体是否为刚性物体、匹配的变换群、轮廓点集是否有序等。以实际的工程应用需求为背景，本文利用形状的轮廓信息，在不同条件下对匹配方法开展了研究： (1) 欧氏变换下基于启发式混合遗传的形状匹配方法 基于轮廓点梯度信息的形状匹配方法适用于各种形状的物体，但是需要全局遍历。由于梯度方向和物体的结构有关系，对于特定形状的物体，就可以结合合适的优化算法加快搜索速度，以避免耗时的全局遍历。图像的宽边缘特性使得利用梯度信息的匹配方法在最优解附近是局部上升的，因此可以通过爬山法等局部优化算法快速找到最优解。所以，对于杂乱场景下的直线型物体，研究如何利用这些特点设计避免遍历以加快搜索速度的方法。 (2) 相似变换下基于点对上下文直方图特征的形状匹配方法 由于轮廓段在遮挡和混乱的情况下是断开的，而之前的匹配方法通常要求模型轮廓段和检测图像中轮廓段对应或者用局部的容易受噪声影响的特征进行匹配，所以设计一种可以表达部分轮廓段的描述子是必要的。又柔性的物体通常是由独立的部分构成的，例如人体的头部、躯干和四肢，这些独立的部分构成一个类似于弹簧的模型。每个独立的部分是带形变的具有独立旋转平移缩放等变换的模型。利用这些独立部分的相互约束关系，可以构造一个轮廓段整合方案。针对上述问题，研究能够表达轮廓段特征的描述子，轮廓段整合方法，和基于两者的形状匹配方法。 (3) 仿射变换下基于多尺度积分特征的形状匹配方法 在有序点集闭合的情况下，形状可以表示为轮廓点构成的曲线，也可以表示为内部的区域。将多尺度积分不变量应用到仿射变换下，需要已知形状之间的仿射变换关系。由于利用特征点是计算仿射变换最直接的方法，而现有的方法存在着提取的特征点个数少或提取方法复杂的问题。针对上述问题，研究仿射变换下基于多尺度积分特征的形状匹配方法，包括 “基于凸包的特征点提取方法”，“基于各向异性高斯核的多尺度积分特征”和基于这两者的匹配方法。同时研究如何将形状之间的仿射变换降级为正交变换，然后直接利用多尺度积分特征进行匹配的方法。 (4) 射影变换下基于分层射影不变特征的形状匹配方法 由于共面二次曲线对存在着射影不变量，对轮廓上的点拟合二次曲线可以避免其它方法中特征点不稳定的问题，所以射影变换下的形状匹配可以利用共面二次曲线对的方法；对此，研究如何利用共面二次曲线对构成能够描述形状局部到全局信息的丰富特征。 本文针对不同条件下的形状匹配方法进行了研究，根据不同的应用需求，对速度和识别率分别进行了提升，丰富和补充了现有的形状匹配方法，具有一定的应用价值。 |
| 英文摘要 | Recognition and localization are the two important tasks of computer vision, Recognition is the process of finding the expect object in the image; Localization is the process of insuring the position and pose of the object. The basis of recognition and localization is the features of the object, including color, texture, structure, shape and so on, where, shape is one of the features used usually. Shape matching is the important method to realize recognition and localization. As a hot problem of the computer research field, it has a long research history. It has achieved a lot of important research results and some results have been used in application. However, because of the complexity of the shape matching, there still exists theoretical issues unresolved completely, and it is still a research focus in the field of computer vision. This paper researches on shape matching methods based on contour information, and this has important theoretical significance and actual application value. There are many factors to consider in shape matching, such as the rigidity of the object, the transformation group, the ordered information of the point set and so on. With the actual engineering demands as background, this paper uses the contour information of the shape, and makes researches for shape matching methods under different conditions: (1) Shape matching method based on heuristic hybrid genetic algorithm under European transformation The shape matching method based on contour gradient information is suitable for all kinds of objects with a global ergodicity. For objects with special shape, it can be combined with suitable optimal methods to speed up to avoid the exhaustive global ergodicity according to the relationship of gradient orientation and the structure of objects. The wide edge character of the image make an up trend around the optimal solution of the shape matching method based on the gradient information, so the optimal solution can be found by a partial optimal method such as climbing method. Therefore, for the objects with linear structure, this paper studied how to make use of these features to design a searching method to avoid the ergodicity to speed up the method. (2) Shape matching method based on point pair context histogram feature under similarity transformation The contour segments are usually broken under occlusion and clutter, and the former matching method is usually required that the contours of the model and the contours in detecting image are corresponding with each other or matching with the local features affected by noises easily, so it is necessary to design a feature to express the partial contour segment. Deformable objects are usually composed of separate parts, such as head, torso and limbs of the human body, and these separate parts form a spring like model. Each part is a deformable model with separate rotation, translation, and scaling transformation. With the constrains of these separate parts, an integrating scheme of the contour segments can be constructed. To solve the above problems, this paper studies the descriptor which can represent the contour features, the integrating method of contour segments, and the shape matching method based on them. (3) Shape matching method based on multiscale integrating feature under affine transformation In the case of ordered point set closure, the shape can be represented as a curve of the contour points, or an internal region. It is necessary to know the affine transformation of shapes to apply the multiscale integral invariant to the affine transformation. Although using the feature points is the direct method to compute the affine transformation, the existing feature extracting method have the problems of the less number of feature points or it is too complicate. At the same time, this paper studies how to degrade the affine transformation to orthogonal transformation, then, the multiscale integral feature can be used directly. (4) Shape matching method based on projective invariant features under projective transformation As there exists projective invariants of two coplanar conics and fitting conics by contour points can avoid the unstable feature points of other methods, the projective shape matching method can use the coplanar conics. This paper studies how to use the feature constructed by coplanar conics to obtain a rich feature which can get the local to global information. This paper makes researches for shape matching method under different conditions. For different application requirements, the matching speed and recognition rate are improved. It enrich and supplement the existing shape matching method, and has a certain application value. |
| 语种 | 中文 |
| 产权排序 | 1 |
| 源URL | [http://ir.sia.cn/handle/173321/20541]  |
| 专题 | 沈阳自动化研究所_光电信息技术研究室 |
| 推荐引用方式 GB/T 7714 | 蔡慧英. 基于轮廓信息的形状匹配方法研究[D]. 沈阳. 中国科学院沈阳自动化研究所. 2017. |
入库方式： OAI收割
来源：沈阳自动化研究所
浏览0
下载0
收藏0
其他版本
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。