Structured max-margin learning for multi-label image annotation

Abstract

In this paper, a structured max-margin learning scheme is developed to achieve more effective training of a large number of inter-related classifiers for multi-label image annotation. First, a visual concept network is constructed to characterize the inter-concept visual similarity contexts more precisely and determine the inter-related learning tasks automatically. Second, multiple base kernels are combined to achieve more precise characterization of the diverse visual similarity contexts between the images and address the issue of huge intra-concept visual diversity more effectively. Third, a structured max-margin learning algorithm is developed by incorporating the visual concept network, max-margin Markov networks and multi-task learning to address the issue of huge inter-concept visual similarity more effectively. Our structured max-margin learning algorithm can leverage the inter-concept visual similarity contexts to learn a large number of inter-related classifiers simultaneously and improve their discrimination power significantly. Our experiments have also obtained very positive results.