Multi-kernel Co-SVM with a new strategy using unlabeled data for image retrieval

Relevance feedback based on SVM classifier shows a good performance recently but the finite feedback counts limited by user´s patience and the small sample size problem are not solved well, Co-SVM does a good job in solving these problems but still has some flaws. We propose three strategies to try to improve this algorithm: (1) different kernel functions are used to characterize the color and texture visual similarities; (2) a new method is proposed to caculate the confident scores of the contention samples; (3) a bunch of the most irrelevant images with the highest confident score are added into the labeled images to extend the size of labeled data while choosing a bunch of images for user labeling. Experimental results verify the superiority of our method over Co-SVM.