Smart Image Search by Boosted Shape Features

An one-line image database search method, which utilizes the boosted-shape feature similarities, is proposed. Salient common feature informations provided by the relevance feedback or multi-instance query are boosted for improving retrieval results. Weak classifiers are successively refined to yield a final strong classifier. The similarity between two shape samples was measured in statistic space of features, through which relative instead of absolute similarity was targeted for visual information retrieval. Experiments of query by the boosted features on thirty thousand trademark images showed that the retrieved results meet visual similarity of shape very well. Objective evaluations, precision-recall hit curve and averaged normalized modified retrieval rank, ANMRR, demonstrate improved retrieval performances of the proposed method. It shows that only 5 7 boosted features out of 100 or more were enough to represent subjective recognition on shape similarity. Key–Words: MPEG-7, boosting, retrieval, image database, Zernik moments, rotation and scale invariant