Improving image annotation via representative feature vector selection

How to bridge the semantic gap is currently a major research problem in Content-Based Image Retrieval (CBIR). Most applications are based on supervised machine-learning classifiers to match images with their related categories. Noisy training information has resulted in current systems having low accuracy, especially when using large numbers of vocabulary categories. In this paper, we describe the use of the Information Gain (IG) and AdaBoost learning algorithms for noise and outlier information filtering in the system training stage, thus improving the performance of image classification. Our experiments look at different numbers of target categories and image segmentation schemes.