Learning a discriminative model for image annotation

This paper introduces a new discriminative model for image annotation. To learn the discriminative model, our method divides each training image into patches, and embeds the patches into a hypergraph, so as to find the representative instances (also called exemplars) for every single class by solving the graph. Then, the feature differences between the training samples and the exemplars are used to form new feature vectors for the training process. We aim to prune the specific features for each single label and formalize the annotation task as a discriminative classification problem. The kernel methods are also employed to solve the problem. Experiments are performed using the Corel5K dataset, and provide a quite promising result when comparing with other existing methods.