一种用于图像检索的新型空间关系特征

New Spatial Relational Feature for Image Retrieval

为了提高对二维图像中空间关系的描述能力,提出了一种新的空间关系量化特征IBS-SC(Interaction Bisector Surface-Shape Context)。以IBS为基础,采用形状上下文描述子SC对IBS与空间关系中的主体区域之间的分布进行描述,将区域与区域之间的交互分布转换为一个点直方图的集合,再使用SC对这些点直方图进行二次统计,得到最终归一化的空间关系特征IBS-SC。为了验证IBS-SC特征的性能,实现了一个图像检索的原型系统,通过对检索结果的对比分析与评价,证明了IBS-SC特征对于空间关系具有良好的表征能力。

This paper proposes a novel spatial relationship descriptor named Interaction Bisector Surface-Shape Context(IBS-SC)to describe the complex relationships among different objects in images. IBS-SC is an image descriptor to model the spatial distribution of subject regions and IBS. The method firstly generates IBS results, and then converts inter-region relations into a set of point histograms. Based on the histograms, this paper uses SC to conduct a second statistical treatment with normalization and obtains the final feature IBS-SC. In experiments, this paper designs an archetype system for image retrieval to demonstrate the effectiveness of IBS-SC. The successful retrievals of spatial relationships in different types and the exhaustive comparison results show that the IBS-SC feature outperforms the existing approaches.