Boosting the weights of positive words in image retrieval

In this paper, an image retrieval system based on the bag-of-words model is developed, which contains a novel query expansion technique. SIFT image features are computed using the Hessian-Affine keypoint detector. All feature descriptors are taken into account for the bag-of-words representation by dividing the full set of descriptors into a number of subsets. For each subset, a partial vocabulary is created and the final vocabulary is obtained by the union of the partial vocabularies. Here, a new discriminative query expansion technique is proposed in which an SVM classifier is trained in order to obtain a decision boundary between the top ranked and the bottom ranked images. Treating this boundary as a new query, words appearing exclusively in top-ranked images are further boosted by rewarding them with larger weights. The images are re-ranked with respect to the their distance from the new boosted query. It is proved that this strategy improves image retrieval performance.