Learning user intention in relevance feedback using optimization

We present an optimization based approach to simultaneously extracting user interested objects from multiple relevance feedback images. We introduce a novel three-term cost unction; the first term measures the smoothness of local image regions within each individual image; the second term measures the homogeneity of user interested objects across different images; the third term favours the assumption that user interested objects will appear most frequently in the positive feedback examples. To model user interested regions in the query image and all multiple positive feedback images simultaneously, we employ a set of local image patch appearance prototypes to link image pixels across multiple images in order to reduce the complexity. Optimizing the cost function segments out the user interested objects from the query and all positive user feedback images simultaneously, which in turn enables the selection of relevant image features for refining image retrieval. We also present an optimization based manifold learning method which uses feedback samples as constraints to perform image retrieval. We present experimental results to demonstrate the effectiveness of our new methods.