Scholarlink

Visual Question Answering (VQA) has emerged as a complex and challenging interdisciplinary task that requires the fusion of computer vision and natural language processing. In this research, we present a detailed investigation into the VQA domain, focusing on our implementation of the cutting-edge Vilt-B32-MLM model on the popular MS COCO dataset. With a compelling accuracy of 79%, our study highlights the effectiveness of Vilt-B32-MLM in comprehending and responding to intricate questions related to visual content. This paper provides an in-depth analysis of our methodology, highlighting the crucial role of data preprocessing, model architecture, and fine-tuning strategies in achieving such a significant accuracy milestone. The findings presented here offer critical insights into the potential of advanced deep learning models in addressing complex multimodal tasks, paving the way for further advancements in the field of VQA and its applications in various real-world scenarios. Furthermore, we compare our approach with other CNN-based TSDR methods, demonstrating its superiority.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing Visual Question Answering with Beam Search in Transformer Models

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Pratiksh Kumar,
  • Rishik Gupta,
  • Vanshika Mishra,
  • Prakhar Shukla,
  • Bagesh Kumar,
  • Pratham Bhatia,
  • Abhinav Upadhyay

摘要

Visual Question Answering (VQA) has emerged as a complex and challenging interdisciplinary task that requires the fusion of computer vision and natural language processing. In this research, we present a detailed investigation into the VQA domain, focusing on our implementation of the cutting-edge Vilt-B32-MLM model on the popular MS COCO dataset. With a compelling accuracy of 79%, our study highlights the effectiveness of Vilt-B32-MLM in comprehending and responding to intricate questions related to visual content. This paper provides an in-depth analysis of our methodology, highlighting the crucial role of data preprocessing, model architecture, and fine-tuning strategies in achieving such a significant accuracy milestone. The findings presented here offer critical insights into the potential of advanced deep learning models in addressing complex multimodal tasks, paving the way for further advancements in the field of VQA and its applications in various real-world scenarios. Furthermore, we compare our approach with other CNN-based TSDR methods, demonstrating its superiority.