Combining long-term learning and active learning with semi-supervised method for content-based image retrieval

To improve the efficiency of relevance feedback in image retrieval, an integrated method of long-term learning and active learning is proposed. In early stage, more positive samples are obtained through long-term learning. The problem of biased training samples is effectively solved through a semi-supervised method that uses not only labeled training samples but also unlabeled ones; therefore an accurate initial SVM classifier is obtained. In later stage, through active learning algorithm that selects the most useful samples in database to solicit the user for labeling, samples required for labeling by users decreased largely and convergence rate increased greatly. Experimental results on 5000 Corel images library have shown that the proposed method can greatly improve both the efficiency and the performance, and it can accelerate the convergence to user´s query concept as well