Image Annotation by Learning Label-Specific Distance Metrics

Abstract

Recently, weighted k nearest neighbor based label prediction model combined with distance metric learning (KNN+ML) [10,14,17], has become more attractive and showed exciting results on image annotation task. Usually, in KNN+ML framework, a uniform distance metric is learned given a collection of similar/dissimilar image pairs from training data. Thus, for a couple of images, their distance is globally unique. However, this might not be sufficient for label prediction on annotation task because it is impossible to distinguish the multiple labels attached to each image. In this paper, we are motivated to learn multiple label-specific distance metrics, and measure the distance of an image pair under different labels’ distance metrics. We also propose a novel label specific prediction model, in which the weight of each label is determined by its specific distance value rather than previous global distance value. Compared with previous KNN+ML methods, our proposed method is able to exactly discriminate each label in each neighbor, and efficiently reduce the prediction of false positive and false negative labels. Extensive experimental results on three benchmark datasets demonstrate that proposed method achieves more accurate annotation results and competitive overall performance.