图像目标语义层级结构及其非监督学习算法  

Image object′s semantic hierarchy and its unsupervised learning algrithm

作　　者：刘玮[1] 田金文[2] 王勇涛[1] 陈新武[1] 

机构地区：[1]华中科技大学图像识别与人工智能研究所,湖北武汉430074 [2]华中科技大学多谱信息处理技术国防重点实验室,湖北武汉430074

出　　处：《华中科技大学学报（自然科学版）》2009年第5期26-29,共4页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家高技术研究发展计划资助项目(2007AA12Z153)

摘　　要：根据场景图像中某些目标的特征间具有空间关联性和几何连贯性的空间语义属性,对特定场景的图像序列进行学习,首先提取各图像的尺度不变特征变换(SIFT)特征点,分析每对图像中各匹配特征点对间的几何连贯性,然后根据图像序列特征点矢量间的空间紧密程度进行聚类获得目标对象,从而得到基于空间关联性的一种新的图像语义层级结构(SBOSH),此层级结构图是通过非监督自动学习的过程获得的.由SBOSH结构图不仅可得到各节点上的目标对象,而且能够获得各节点间的空间交互性.所得的SBOSH结构图符合人对场景的逻辑分析,可以应用于场景分析、目标及目标局部精确定位和类目标的识别中.According to are the spatial interaction and geometric consistence between objects in scenes of a particular category, a set of scene image was learned, the method could select scale invariant feature transformation （SIFT） feature points and analyze the geometric consistency between the matched points＇ pair. Semantic hierarchy based on spatial relationship based object＇s semantic hierarchy （SBOSH） could be obtained by clustering feature vectors based on their spatial-relationship. This erarchy graph not only obtained the objects at every node, but also obtained the spatial relationship between each node. The semantic hierarchy complies with human＇s philosophical analysis and this hierarchy could be applied to scene analysis, object and its parts＇ detection and category recognition.

关 键 词：目标语义 语义层级结构 空间关联性 几何连贯性 非监督学习 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]