A Transformer-Based Approach for Effective Visual Question Answering

Visual Question Answering (VQA) presents a multifaceted challenge, requiring models to process and integrate both textual and visual data. The Transformer architecture, with its self-attention mechanism, has proven highly effective in managing global dependencies, making it a popular choice for many VQA models. However, a key challenge lies in achieving an optimal balance between global and local dependency modeling within conventional Transformer frameworks. Addressing this issue, our research introduces an innovative Transformer-based approach that employs region-specific self-attention for VQA. In contrast to models that concentrate exclusively on global dependencies, our method concurrently executes intra and inter-window attention operations to effectively capture contextual information from images. By utilizing regional windows for visual feature processing, our model eliminates redundancy in global self-attention while still preserving rich contextual details. The efficacy of our proposed technique, which relies on the utilization of visual features, has been extensively evaluated through experiments. These assessments were conducted using widely recognized VQA benchmark datasets, specifically VQA 2.0 and CLEVR. Our findings demonstrate that the model outperforms existing benchmark alternatives across various performance metrics. Notably, it achieved superior test results of $\mathbf{7 1 . 9 3 \%}$ and $\mathbf{9 8 . 6 2 \%}$ on the VQA 2.0 and CLEVR datasets, respectively, underscoring its effectiveness in tackling complex VQA tasks.