Joint shape and texture analysis of objects boundaries in images using a Riemannian approach
Joint shape and texture analysis of objects boundaries in images using a Riemannian approach
复制标题
使用黎曼方法对图像中的对象边界进行联合形状和纹理分析
DOI：
--
复制
发表时间：
2008
期刊：
影响因子：
--
通讯作者：
E. Klassen
中科院分区：
文献类型：
--
作者：
Wei Liu;Anuj Srivastava;E. Klassen
为了检测、跟踪和识别图像中的目标，必须为代表这些目标的特征建立统计模型。形状和纹理是用于此目的的两个主要特征。虽然过去的研究已经开发了单独的技术来分析这两个特征，但我们的目标是一起研究它们。我们将边界的2D坐标和沿着该边界的纹理函数表示为高维欧几里得空间中的复合参数化曲线。然后，我们定义这些表示的等价关系，以获得所需的不变性。形状分量对刚性运动、全局缩放和重新参数化是不变的，而纹理分量仅对后两者是不变的。使用黎曼结构的商空间，我们计算形状纹理函数之间的测地线路径比较不同的对象。除了一个对象到另一个对象的最佳变形之外，该方法还使用形状和纹理信息提供跨对象的点的内聚配准。我们将使用人工和真实的图像中的对象的例子来演示这个框架。
In order to detect, track and recognize objects in images, one has to develop statistical models for features representing those objects. Shapes and textures are two main features that are used for this purpose. While past research has developed separate techniques for analyzing these for these two features, our goal is to study them together. We represent the 2D coordinates of the boundary and the texture function along that boundary as a composite parameterized curve in a high-dimensional Euclidean space. Then, we define equivalence relations on these representations to obtain the desired invariances. The shape component is made invariant to rigid motions, global scalings and re-parameterizations, while the texture component is made invariant only to the last two. Using a Riemannian structure on the resulting quotient space, we compute geodesic paths between shape-texture functions to compare different objects. In addition to optimal deformations of one object into another, this method provides a cohesive registration of points across objects using both shape and texture information. We will demonstrate this framework using examples of objects in artificial and real images.