MRAN-VQA: Multimodal Recursive Attention Network for Visual Question Answering

Visual Question Answering (VQA) is a fundamental challenge in multimodal AI, requiring models to integrate and reason over both visual and textual information. Despite advancements in deep learning, existing VQA models struggle with multi-step reasoning, hierarchical feature fusion, and multilingual generalization, limiting their effectiveness in real-world applications. This paper introduces MRAN-VQA, a Multimodal Recursive Attention Network for VQA, designed to address these limitations through a three-stage reasoning pipeline. The proposed approach first employs Recursive Attention Encoding, where a Vision Transformer (ViT) extracts visual features, and BERT-based embeddings encode textual information. A recursive self-attention mechanism iteratively refines these representations, improving contextual alignment. Hierarchical Feature Fusion integrates multi-level visual–text interactions through bilinear attention pooling and gated cross-modal operations. Finally, Answer Prediction with Attention Grounding applies a self-attentive reasoning module to responses while optimizing an Attention Grounding Score (AGS) for improved interpretability. Experiments on VQA v2.0, CLEVR, and our custom BanglaVQA datasets demonstrate that MRAN-VQA outperforms state-of-the-art models, achieving 75.6% accuracy on VQA v2.0, 96.1% on CLEVR, and 72% on BanglaVQA—notably surpassing transformer-based baselines. The model exhibits superior multi-step reasoning capabilities in compositional queries and significantly enhances performance in low-resource multilingual settings. The model extracts textual features via BERT and visual features via a Vision Transformer (ViT). These are refined through a Multimodal Recursive Attention Network (MRAN) using iterative Text-to-Image and Image-to-Text Attention. Attentive fusion combines the modalities, and a softmax layer produces the final prediction. Recursive attention enables deeper contextual understanding, enhancing multimodal reasoning in complex VQA tasks. • MRAN-VQA introduces recursive attention for enhanced multi-step reasoning in VQA. • Proposes hierarchical fusion of low, mid, and high-level multimodal features. • Defines an Attention Grounding Score (AGS) with a trainable auxiliary loss. • Demonstrates strong multilingual and low-resource performance on BanglaVQA. • Outperforms recent SOTA on VQA v2.0, CLEVR, and BanglaVQA (best trade-off at R = 4).