DHHG-TAC: Fusion of Dynamic Heterogeneous Hypergraphs and Transformer Attention Mechanism for Visual Question Answering Tasks

Amidst the burgeoning advancements in deep learning, traditional neural networks have demonstrated significant achievements in unimodal tasks such as image recognition. However, the handling of multimodal data, especially in visual question answering (VQA) tasks, presents challenges in processing the complex structural relationships among modalities. To address this issue, this article introduces a dynamic heterogeneous hypergraph neural network (HGNN) model that utilizes a Transformer-based combined attention mechanism and designs a hypergraph representation imaging network to enhance model inference without increasing parameter count. Initially, image scenes and textual questions are converted into pairs of hypergraphs with preliminary weights, which facilitate the capture of complex structural relationships through the HGNN. The hypergraph representation imaging network further aids the HGNN in learning and understanding the scene image modalities. Subsequently, a transformer-based combined attention mechanism is employed to adapt to the distinct characteristics of each modality and their intermodal interactions. This integration of multiple attention mechanisms helps identify critical structural information within the answer regions. Dynamic updates to the hyperedge weights of the hypergraph pairs, guided by the attention weights, enable the model to assimilate more relevant information progressively. Experiments on two public VQA datasets attest to the model's superior performance. Furthermore, this article envisions future advancements in model optimization and feature information extraction, extending the potential of HGNNs in multimodal fusion technology.