Parallel Hypergraph Convolutional Neural Networks for Image Annotation

In recent years, the application of graph neural networks in automatic image annotation has become more mature. However, there still exist several problems. First of all, the regular graph structure network modeling is not accurate enough, and it is not flexible enough to deal with multi-label and heterogeneous data. Second, some algorithms only use graph convolutional networks to construct sample or label graphs, limiting the fusion and extension of multi-modes. In this paper, we propose a semi-supervised automatic image annotation method with parallel hypergraph convolutional neural networks(PHCN). The algorithm combines hypergraph convolutional network (HCN) with graph convolutional network (GCN) to improve the annotation performance under semi-supervised learning. In addition, We connect the label graph with the sample hypergraph, further consider the distribution of labels and features, and aggregate the features. Experiments on three benchmark image annotation datasets show that our method is superior to other existing state-of-the-art methods.