图像中复杂形变物体的外轮廓搜索方法研究

项目名称： 图像中复杂形变物体的外轮廓搜索方法研究

项目编号： No.61305051

项目类型： 青年科学基金项目

立项/批准年度： 2014

项目学科： 自动化技术、计算机技术

项目作者： 苏远歧

作者单位： 西安交通大学

项目金额： 25万元

中文摘要： 物体形状的复杂变化是精确自动分割面临的主要困难之一，本项目将利用物体的外形特征，通过求解复杂形变物体的外轮廓，解决物体的精确分割问题。当把物体外形的所有变化定义为一个形状空间，物体外轮廓的解析过程就转化为一个形状搜索问题，在形状空间中定位与给定图像最匹配的形状。形状空间结构的复杂性、物体外观的变化及背景噪声的影响都极易使搜索过程陷入局部最小。因此，本项目从形状空间的表示与搜索过程的设计两方面入手，提出融合物体形状先验与图像底层特征的物体外轮廓解析方法。主要特色与创新点在于：（1）提出聚类与部件模型结合的方式，解决了复杂形状变化和部分缺失下的形状表示；（2）通过物体的局部匹配估计外形的局部近似，渐进缩减物体外形子空间，实现复杂形变物体外轮廓的精确搜索。

中文关键词： 形状匹配；轮廓片段；物体检测；物体分割；层次模型

英文摘要： Object segmentation is important for computer vision and image understanding. It is challenging and yet unsolved mainly for the complex deformation in object's geometric structure. The proposal studies the deformation for the outline resolving. It defines the outlines coming from the same class as a shape space; and the resolving procedure is then reduced to search for a point in the space such that it explains the test image. However, because of the complex structure of the shape space, plus the noise led by background and the variaty in appearance, the searching process is easily trapped into local minimum. To get rid of it, the proposal conducts the research on shape representation and its matching, and proposes to search for the outline through a procedure integrating both the prior and bottom-up information. Main contributions of the proposal include: (1) proposes a way combining the shape clustering and part-based model for outline representation, (2) designs an algorithm that gradually shrinks the subspace containing the outline through the matching of local features.

英文关键词： Contour fragments；Shape matching；Object detection；Object segmentation；Hierarchical model