Scholarlink

This study discusses a multimodal fusion system that uses cross-domain learning and hierarchical attention to make intelligent imaging apps more accurate, scalable, and fast in real time. Textual, visual, and sensor data are all part of the suggested design. Dynamic feature weighting, cross-modal gathering, and domain-specific standardization all make it easier for links to form within and between modes. The suggested method did better than large-scale baselines like In-ternVL-2.5 and Qwen2-VL in tests on several datasets, with an AUROC of 0.931, an F1-score of 0.861, and a Top-1 accuracy of 88.2%. It has the lowest ECE score (0.036) and Brier score (0.109), which means it is well calibrated. It can be used in real time because it has a 54 ms delay and can handle 168 samples per second. These results demonstrate the method's effectiveness, adaptability, and comprehension in complex mixed environments. It's a cutting-edge way to make smart imaging tools work in the future.

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

AI-Enabled Multimodal Fusion with Cross-Domain Learning Techniques for Intelligent Real-World Imaging Applications

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Teba Mageed Hameed,
  • Huthaifa Ayad Al-Ani,
  • Hameed H. Khalaf,
  • Hamdan Raheem AlKubaisi,
  • Mahmood Hameed Majeed,
  • Hayder Musaad Al-Tmimi,
  • Tarun Goma

摘要

This study discusses a multimodal fusion system that uses cross-domain learning and hierarchical attention to make intelligent imaging apps more accurate, scalable, and fast in real time. Textual, visual, and sensor data are all part of the suggested design. Dynamic feature weighting, cross-modal gathering, and domain-specific standardization all make it easier for links to form within and between modes. The suggested method did better than large-scale baselines like In-ternVL-2.5 and Qwen2-VL in tests on several datasets, with an AUROC of 0.931, an F1-score of 0.861, and a Top-1 accuracy of 88.2%. It has the lowest ECE score (0.036) and Brier score (0.109), which means it is well calibrated. It can be used in real time because it has a 54 ms delay and can handle 168 samples per second. These results demonstrate the method's effectiveness, adaptability, and comprehension in complex mixed environments. It's a cutting-edge way to make smart imaging tools work in the future.