Knowledge propagation in content-based image retrieval

Content-based image retrieval (CBIR) systems experience the challenge of semantic gap between the low-level visual features and the high-level semantic concepts. It would be advantageous to build CBIR systems which support high-level semantic query. The main idea is to integrate the strengths of content- and keyword-based image indexing and retrieval algorithms while alleviating their respective difficulties. However, full manual annotation of complete database is often tedious and expensive. To address this difficulty, knowledge propagation (automatic image annotation) has been proposed to automatically assign textual descriptors in the form of keywords or tags to unannotated images. In this paper, a new knowledge propagation scheme is presented which is based on image content analysis and training of keyword classifiers. In particular, genetic algorithm (GA) is utilized to find the salient regions in the labeled images of the same semantic concepts. The importance of the regions is then estimated by a one-class support vector machine (OCSVM). Next, radial basis function (RBF)-based classifiers are trained based on the contents of the labeled images. Finally, the trained classifiers are used for keywords propagation. Experimental results show that the proposed method is effective for image annotation.