SUBSPACE LEARNING BASED ACTIVE LEARNING FOR IMAGE RETRIEVAL

The goal of relevance feedback is to improve the performance of image retrieval by leveraging the labeling of human. It is helpful to introduce active learning method in relevance feedback to alleviate the human burden. In the traditional active learning the samples which can improve the classifier the most if they were labeled are selected for the user's labeling. However, the change of the geometrical structure of the data distribution caused by such expensive labeled samples is not fully exploited. By mining user's labeling information, we can reduce the original feature space dimension to ease the classifier's instability brought by the small sample size. In this paper, we propose a novel batch mode active learning method for informative data selection. The labeled samples are not only used to retrain the classifier, but to learn a subspace which efficiently encodes user's intention as well. Especially, a scheme of certainty propagation on the subspace effectively integrates uncertainty sampling and subspace learning into the proposed Subspace learning based batch mode Active Learning method (SubAL) in relevance feedback. Extensive experiments on publicly available dataset shows that the proposed method is promising.