图像检索中的动态相似性度量方法

为提高图像检索的效率，近年来相关反馈机制被引入到了基于内容的图像检索领域。该文提出了一种新的相关反馈方法－－动态相似性度量方法。该方法建立在目前被广泛采用的图像相拟性度量方法的基础上，结合了相关反馈图像检索系统的时序特性，通过捕获用户的交互信息，动态地修正图像的相似性度量公式，从而把用户模型嵌入到了图像检索系统，在某种程度上使图像检索结果与人的主观感知更加接近。实验结果表明该方法的性能明显优于其它图像检索系统所采用的方法。

A Dynamic Similarity Metric Approach for Image Retrieval

Due to the rapid growing amount of digital images on World Wide Web, there is a need for large multimedia database management. In the past few years, content-based image retrieval has been becoming an active research area. At the early stage of CBIR, research primarily focused on exploring various feature representations, and ignored the subjectivity of human perception. There exists a gap between high-level concepts and low-level features. Relevance feedback is a promising approach to finding a mapping between semantic objects and low-level features.Dynamic Similarity Metric,a novel relevance feedback approach for image retrieval, is proposed in this paper. It is a new technique that explores similarity information in relevance feedback. Differing from the general similarity metric, dynamic similarity metric takes into account the characteristics of all examples in the retrieval process. In general image retrieval, the similarity metric gives the similar degree between two images. Every feature is associated with the weights for reflecting the importance of corresponding feature. Although some relevance feedback image retrieval systems can change weights according to the user's feedback, previous feedback information is no longer used in these systems. Dynamic similarity metric can memorize the previous feedback information. At the same time, time sequence characteristic of relevance feedback is taken into account by the approach. To optimize the retrieval result, the similarity metric formula can be improved dynamically according to the user's interactive information. Through embedding all positive and negative examples' characteristics, it can find good results. The more promising images are emphasized more and the less promising ones are emphasized less. To evaluate the efficiency and contribution of dynamic similarity metric, we have implemented an image retrieval system, ImageSeek, based on the proposed method and performed some experiments. Experiments show that the dynamic similarity metric is a robust algorithm, which adapts to image retrieval based on different visual features and gets high recall by much fewer interactions than other methods.