Image Annotation with Weak Labels

Abstract

In this paper, we address the problem of image annotation when the given labels of training image are incomplete, inaccurate, and unevenly distributed, in the form of weak labels, which is frequently encountered when dealing with large scale web image training set. We introduce a progressive semantic neighborhood learning approach that explicitly addresses the challenge of learning from weakly labeled image by searching image’s semantic consistent neighborhood. Neighbors in image’s semantic consistent neighborhood have global similarity, partial correlation, conceptual similarity along with semantic balance. We also present an efficient label inference algorithm to handle noise by minimizing the neighborhood reconstruction error. Experiments with different data sets show that the proposed framework is more effective than the state-of-the-art algorithms in dealing with weakly labeled datasets.