A new CBIR approach based on relevance feedback and optimum-path forest classification

Abstract

Recently some CBIR approaches have shown the use of relevance feedback to train a pattern classifier to select relevant images for retrieval. This paper revisits this strategy by using an optimum-path forest (OPF) classifier. During relevance feedback iterations, the proposed method uses the OPF classifier to decide which database images are relevant or not. Images classified as relevant are sorted and presented to the user for a new iteration. Such images are ordered according to the normalized distance using relevant and irrelevant representative images, computed previously by the OPF classifier. Our experiments show that the proposed approach requires fewer iterations, being faster and more effective than methods based on SVM.