多策略融合相关反馈技术的研究.pdfVIP

多策略融合的相关反馈技术研究 谢晓霞，朱振峰，赵耀 （北京交通大学信息科学研究所，北京 100044 ） 摘要：在基于内容的图像检索系统中，相关反馈技术的引入在一定程度上缩小了图像的视觉底层特征 与图像的高层语义之间的鸿沟。本文提出一种综合利用反馈信息,融合多种策略的相关反馈算法：根据 核密度估计理论并引入均值漂移优化技术实现查询向量的有效更新；为实时地把握图像的语义信息以 及用户的查询意图，在更新特征各维之间权重分配的同时，还对各个特征之间的权重进行动态更新； 最后，为削弱那些在底层特征上相似但在高层语义上不相关的非相关图像给检索带来的负面影响，使 库中未被检索到的相关图像通过反馈有更大的机会被检索到，本文还提出了基于对反馈图像进行奖惩 制度的特征库自动更新方法。在 Corel 通用图像测试库上的实验表明，本文算法可以明显提高检索的准 确率和收敛的速度。 关键词：基于内容的图像检索 相关反馈 核密度估计 均值漂移 Research on Multi-Strategy Fusion Based Relevance Feedback Abstract: In the content-based image retrieval systems, relevant feedback is popularly considered as an effective way to narrow the gap between high-level concepts and low-level vision features. In this paper, the research on multi-strategy fusion based relevance feedback is conducted. To make the query instance be more representative for user’s semantic concept，the non-parametric kernel density estimation isfirst applied to character its distribution in feature space and furthermore the mean shift based optimization technique is implemented to refine the query. In order to catch the users’ retrieval intention synchronously, both the inter relations among the feature vector and the intra relations among the different kinds of vision low-level features are explored. In addition, we also introduce a rewards and punishment scheme to all the feed backed images to weaken the negative effect from those images, which are relevant in the low-level feature space but irrelevant in the high-level semantics, and endow the relevant images in the database more chance to be retrieved in next round of feedback. The final experimental results on the Corel database show that the proposed approach greatly improves the retrieval performance. Keywords: Contend-Based Image Retrieval （CBIR ）, relevance feedback，Ker