Mean version space: a new active learning method for content-based image retrieval

In content-based image retrieval, relevance feedback has been introduced to narrow the gap between low-level image feature and high-level semantic concept. Furthermore, to speed up the convergence to the query concept, several active learning methods have been proposed instead of random sampling to select images for labeling by the user. In this paper, we propose a novel active learning method named mean version space, aiming to select the optimal image in each round of relevance feedback. Firstly, by diving into the lemma that motivates support vector machine active learning method (SVM<i><inf>active</inf></i>), we come up with a new criterion which is tailored for each specific learning task and will lead to the fastest shrinkage of the version space in all cases. The criterion takes both the size of the version space and the posterior probabilities into consideration, while existing methods are only based on one of them. Moreover, although our criterion is designed for SVM, it can be justified in a general framework. Secondly, to reduce processing time, we design two schemes to construct a small candidate set and evaluate the criterion for images in the set instead of all the unlabeled images. Systematic experimental results demonstrate the superiority of our method over existing active learning methods