Abstract
A new relevance feedback based on a tree structure with visual keys is proposed to reflect the user's sense in the region based image retrieval (RBIR) system employing query with visual keys. For relevance feedback, the region image which is obtained from a feedback image is linked to the correspondent leaf node in a tree structure, but it is not effective to add the leaf node to the query like in the usual relevance feedback because of the small increase in the number of the retrieved images. In the proposed relevance feedback, the device in which an upper node from the leaf node is added to the query is implemented. Image retrieval experiments using 1,000 test images in COREL gallery show that an average retrieval precision is improved from one of conventional relevance feedback using leaf nodes by the proposed method using upper nodes, and the effect of the relevance feedback is confirmed from an obtained recall-precision curve.
