Co-training for search-based automatic image annotation

Search-based automatic image annotation is an effective technology to enhance the performance of annotation. By integrating the co-training technique, this paper addresses on a novel scheme for search-based image annotation, in which two classifiers can contribute to each other during the training phase. Since each classifier can select some most confident images to enhance the generalization ability of the other one, the co-training learning algorithm is triggered out for automatically mining more and more relevant images, which improve the annotation performance greatly. To characterize the various contribution of each relevant image, the probability output of the classification is taken as the corresponding weight. Moreover, the histogram of retrieved keywords is proposed to re-rank the final reliability of the keywords to be annotated, which can also guarantee the scalability property of annotation to some extent. With the promoted search precision on the basis of co-training strategy, the experimental results demonstrate the improvement of annotation performance.