Multi-Output Regression with Tag Correlation Analysis for Effective Image Tagging

HongYun Cai, Zi Huang, Xiaofeng Zhu, Qing Zhang, Xuefei Li

Abstract: Automatic image tagging is one of the most important research topics in multimedia. How to achieve accurate image tagging to bridge the semantic gap between images’ content and users’ semantic understanding has been widely studied in the last decade. One common approach is to convert image tagging to a multi-task learning problem. However, most existing methods ignore tag correlations in the learning process. In this paper, we show the importance of tag correlations in conducting multi-task learning. We formulate image tagging as a multi-output regression problem accounting for tag correlations, which are captured by the covariance matrix of the regression coefficients and the noise across all tags respectively. The combination of multi-output regression with tag correlation analysis takes advantage of the latent dependencies among tags to overcome limitations of existing work. Extensive experiments have been conducted on two benchmark datasets, and the results confirm that our approach outperforms the state-of-the-art methods.