VTFCGNet: Dual Domain Attention Mechanism Multimodal Fusion Network for Visual Text Question Answering

Although traditional Transformer architectures are capable of generating fine-grained representations, they often struggle to suppress redundant features during cross-modal attention computation, impairing semantic focus and reasoning performance. Visual Text Question Answering (VTQA) further raises the bar by requiring precise alignment and deep reasoning across images, scene text, and natural language questions. To address these challenges, we propose VTFCGNet, a dual-domain attention network that integrates Fourier-spatial self-attention (FsSA), CBAM-guided fine-grained enhancement, and a structured gather-reason-scatter pipeline for multimodal inference. On the VTQA dataset, VTFCGNet achieves 75.83% accuracy on Y/N questions, 70.47% E-F1, and 45.36% G-F1. On VQA-v2, it achieves an overall accuracy of 71.93%, surpassing all previous baseline models. Ablation studies reveal that VTFCNet and CRGNet contribute complementary gains: removing either causes a drop of over 1.3%, and removing both reduces accuracy to 69.87%. These results confirm the effectiveness and robustness of our proposed architecture. Attention visualization further confirms the model’s superiority in semantic focus and redundant information suppression.