Visual question answer(VQA) is a task that combines computer vision and natural language and plays an important role in visual-text interaction tasks. In the last few years, Transformer has begun to appear in a multimodal set of tasks. The emergence of Transformer has advanced the field of artificial intelligence. However, there are certainly drawbacks. For example, fine-grained relationships within modalities are often ignored, poor representation of features and interference of irrelevant information between different modalities. In the inter-modals fusion module, the useful information between different modalities can not be fused.

In this paper, in order to solve the above problems, we improve the visual attention mechanism based on the MCAN model and enhance the attention features and design a fusion module to improve the fusion of two different modalities. Firstly , we propose a module to enhance the ability of visual attention, which is used to learn the relationship of fine-grained features within the image, discard irrelevant information and enhance the effective information, so as to obtain a more interesting region of the image. Secondly a cross-modal information fusion module is proposed to enhance the interaction of different modal information. The fusion module is the core of the whole network model, through which different modal information is effectively combined to predict the correct answer. The experiments are evaluated on the VQA2.0 dataset, and compared with the existing methods. The method has significant advantages.