Abstract: 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.

Key words: spatial relationship feature, image retrieval, RAID feature

摘要： 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。

关键词: 空间关系特征, 图像检索, RAID特征