Hypergraph Transformer for Multi-Label Image Classification

Multi-label image classification is a crucial task in computer vision. To improve the classification accuracy of multiple objects within an image, many researchers attempt to construct graph or hypergraph structures to capture the high-order relationships between various objects. However, existing graph models tend to have weak resistance to noise. To address this issue, we introduce an attention mechanism and propose a transformer-based hypergraph neural network model(IC-HGT). We design a hypergraph transformer layer that propagates the relationships between label features within the hypergraph through an attention mechanism. Additionally, we utilize multi-head attention on the hypergraph, allowing the model to learn a broader range of vertex representations. Our model achieve a mean Average Precision (mAP) of 84.6% on the MS-COCO 2014 dataset, outperforming the state-of-the-art methods.