Separability-based relevance feedback for content-based image retrieval

Relevance feedback is an interactive querying process of CBIR, which can help the retrieval system adapt to the dynamic user demand and achieve more accuracy for the representation of image similarity according to the user´s view. In this paper, we bring up the concept of the separability of image data as the theoretical criterion to guide generating the best feature vector set to better classify images. We build an energy function for the training of the eigenvectors on the basis of separability, and then apply simulated annealing optimization algorithm to minimize energy function in order to accomplish the generation of the optimal feature vectors set in a particular database. The experiment result shows that this new algorithm for relevance feedback can effectively improve the retrieval accuracy.