Graph regularized low-rank feature learning for robust multi-label image annotation

Abstract

Automatic image annotation has drawn much attention in the last two decades due to its application in social images organization. Most studies treat image annotation as a typical multi-label classification problem, which requires sufficient number of training samples with complete and clean tags to train reliable prediction models. Being aware of this drawback, we develop a novel graph regularized low-rank feature mapping for image annotation under semi-supervised learning framework. Specifically, our approach refers to the idea of matrix recovery and uses the matrix trace norm to capture the correlations among different tags. It also helps to control the model complexity. In addition, by using graph Laplacian regularization as a smooth operator, the proposed approach can explicitly take into account the local geometric structure on both labeled and unlabeled images. Moreover, considering the tags of labeled images tend to be missing or noisy, we introduce a supplementary ideal label matrix to solve the problems of missing tags and noisy tags for given training samples. A variety of experiments conducted on image datasets ESPGame, IAPRTC-12, and NUS-WIDE prove the effectiveness of the proposed approach.