Boosted Inductive Matrix Completion for Image Tagging

Abstract

Search engines have traditionally used manual image tagging for indexing and retrieving image collections. Manual tagging is expensive and labor intensive, motivating the research on automatic tag completion. However, existing tag completion approaches suffer from deficient or inaccurate tags. In this study, we formulate the task in the boosted inductive matrix completion (BIMC) framework, which combines the power of the inductive matrix completion (IMC) model together with a standard matrix completion (MC) model. We incorporates visual-tag correlation and semantic-tag correlation properties into the model for better exploration of the latent connection between image features and tags. We exploit CNN features and word vectors to narrow the semantic gap. The proposed method achieves good performance on several benchmark datasets with missing and noisy tags.