A Study of Visual Question Answering Techniques Based on Collaborative Multi-Head Attention

In visual question answering task, the dominant approach recently has been to use a unified model for pre-training and fine tuning it. This unified model typically uses a transformer to fuse image and text information. In order to optimize the performance of the model on visual question answering task, this paper proposes a transformer architecture based on a collaborative multi-head attention mechanism to address the key/value projection redundancy problem in the multi-head attention mechanism of the transformer. In addition, this paper uses the Swin transformer model as the image feature extractor to extract multi-scale image information. Validation experiments are conducted on the VQA v2 dataset in this paper, and the experimental results show that applying the collaborative multi-head attention approach and the Swin transformer backbone to the visual question answering model can effectively improve the correct rate of the visual question answering task.