Multi-label image annotation based on neighbor pair correlation chain

Image annotation plays an important role in content-based image understanding, various machine learning methods have been proposed to solve this problem. In this paper, label correlation is considered as an undirected bipartite graph, in which each label are correlated by some common hidden topics. As a result, given a label, random walk with restart on the graph supplies a most related label, repeating this precedure leads to a label chain, which keep each adjacent labels pair correlated as maximally as possible. We coordinate the labels chain with its respective classifier training on bottom feature, and guide a classifier chain to annotate an image. The experiment illustrates that our method outperform both the baseline and another popular method.