融合跨模态transformer的外部知识型VQA
DOI:
作者:
作者单位:

中国人民公安大学

作者简介:

通讯作者:

中图分类号:

TP391

基金项目:

公安部技术研究计划项目(2020JSYJC22);中央高校基本科研业务费专项资金资助( 2020JKF501)


External knowledge-based VQA integrating cross modal transformers
Author:
Affiliation:

1.People'2.'3.s Public Security University of China

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    针对外部知识型的VQA任务性能效果不佳的问题,本文构建了一种融合跨模态Transformer的外部知识型VQA模型框架,通过在VQA模型外引入外接知识库来提高VQA模型在外部知识型任务上的推理能力。进一步地,模型借助双向交叉注意力机制提升文本问题、图像、外接知识的语义交互融合能力,用于优化VQA模型在面对外部知识时普遍存在的推理能力不足的问题。实验可得,与基线模型LXMERT相比,在OK VQA数据集上,本文模型整体性能指标overall提升了15.01%。同时,与已有最新模型相比,在OK VQA数据集上,本文模型整体性能指标overall提升了4.46%。可见本文模型在改进外部知识型VQA任务性能方面有所提升。

    Abstract:

    In response to the issue of poor performance of external knowledge-based VQA tasks, this paper constructs a framework for external knowledge-based VQA models that integrates cross-modal Transformers. By introducing an external knowledge base outside the VQA model, the inference ability of the VQA model on external knowledge-based tasks is improved. Further, the model utilizes a bidirectional cross attention mechanism to enhance the semantic interactive and fusion ability of text problems, images, and in order to optimize the problem of insufficient reasoning ability commonly found in VQA models in the face of external knowledge. According to the experiment, compared with the baseline model LXMERT, the overall performance index of our model OVERALL improves by 15.01% on the OK VQA dataset. Meanwhile, compared with the existing latest model, the overall performance index of our model overall improves by 4.46% on the OK VQA dataset. It can be seen that this paper's model improves the performance of external knowledge-based VQA tasks.

    参考文献
    相似文献
    引证文献
引用本文

王虞,孙海春. 融合跨模态transformer的外部知识型VQA[J]. 科学技术与工程, , ():

复制
分享
文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2023-07-19
  • 最后修改日期:2023-11-02
  • 录用日期:2023-11-14
  • 在线发布日期:
  • 出版日期:
×
亟待确认版面费归属稿件,敬请作者关注