Image retrieval using relevance feedback can be considered a classification process. In practice, the generalization of classifier is often constrained by the insufficiency of training samples. In this paper, we propose a novel relevance feedback approach capable of collecting more representative samples. Image labeling and classifier training are conducted in two complementary image feature spaces. The complementarities between feature spaces are also studied. Our experimental result based on 10,000 images indicates that the proposed approach significantly improves image retrieval performance.
