Enhancing Content-Based Image Retrieval by Exploiting Relevance Feedback Logs
发掘相关反馈日志中关联信息的图像检索方法

Relevance feedback (RF) has been successfully used in content-based image retrieval (CBIR). However, most CBIR systems seldom reuse the latent semantic correlation among images revealed by RF log to guide retrieving across sessions. In this paper, concurrence of images in a RF record is regarded as a kind of semantic homogeneity in certain context and the image-retrieving problem is cast as an authority-image-finding task. Records in RF logs first extend the result from traditional CBIR systems. This produces a relevant graph of images related to the query with multiplex contexts. Then, a modified HITS algorithm is applied to it to distill consensus about semantic relevance. As a result, both visual content and semantic relevance can be maintained in image retrieval and the efficiency is much improved compared with traditional CBIR methods. Experimental results demonstrate its superiority in both objective criteria and semantic clustering capability against the Corel database with 60 000 images.