Scholarlink

In order to improve the accuracy and expression of video semantic understanding and description generation, a cross-modal dynamic attention-driven multi-dimensional feature enhancement model is constructed to study the synergistic mechanism between vision and text. MSR-VTT and ActivityNet-Captions are taken as examples to analyze the role of dynamic attention mechanism in feature alignment and semantic construction. The results show that the BLEU-4 scores of the full model are 43.1 and 41.8, the CIDEr scores are improved to 75.2 and 72.6, the validation loss is reduced to 0.184, and the feature mutual information alignment value reaches 0.72, which reflects a good semantic extraction capability and generation performance.

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

Cross-Modal Dynamic Attention-Driven Multidimensional Feature Enhancement with Video Semantic Understanding and Hierarchical Description Generation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Qiaojuan Hui

摘要

In order to improve the accuracy and expression of video semantic understanding and description generation, a cross-modal dynamic attention-driven multi-dimensional feature enhancement model is constructed to study the synergistic mechanism between vision and text. MSR-VTT and ActivityNet-Captions are taken as examples to analyze the role of dynamic attention mechanism in feature alignment and semantic construction. The results show that the BLEU-4 scores of the full model are 43.1 and 41.8, the CIDEr scores are improved to 75.2 and 72.6, the validation loss is reduced to 0.184, and the feature mutual information alignment value reaches 0.72, which reflects a good semantic extraction capability and generation performance.