Negative Examples Supervised Feature Filtering Strategy in Relevance Feedback

Relevance Feedback is a particular method of feature selection in Content-based image retrieval, and also is a very effective strategy to improve retrieval accuracy and bridge the ´semantic gap´. In this paper, a new feature filter whose parameters are computed by negative examples supervised approach is proposed in order to select the unique features to characterize the positive example clusters. A new query refinement method based on the nearest neighbor is used to rank the candidates after getting rid of irrelevant feature components. Experiment results show that the proposed approach is more efficient and robust than the traditional method.