Incorporating Multiple SVMs for Active Feedback in Image Retrieval Using Unlabeled Data

Active learning with support vector machine(SVM) selects most informative unlabeled images for user labeling, however small training samples affect its performance. To improve active learning and use more unlabeled data, we propose a new algorithm training three SVMs separately on the color, texture and shape features of labeled images with three different kernel functions according to the features' distinct statistical properties. Different algorithms are used in the selection of disagreement and agreement samples from unlabeled data and also in the calculation of their confidence degrees. The lowest confident disagreement samples are returned to user to label and added to the training data set with the highest confident agreement samples. Experimental results verify the high effectiveness of our method in image retrieval.