Scholarlink

<p>This paper studies sequential image captioning with stronger object grounding and nonlinear visual interaction modeling. We propose <b>RacGK</b>, which combines Region-Specific Feature Extraction (RSFE), a Region-Aware Attention Mechanism (RAAM), and Gaussian RBF-based Kernelized Self-Interaction Attention (KSIA) in an attention-guided LSTM decoder. On MS COCO 2014 with the Karpathy split, RacGK achieves competitive results, including 40.3 BLEU-4 and 59.4 ROUGE-L after CIDEr optimization. Ablations support the contribution of region-aware fusion and Gaussian kernel attention. Code is available at <a href="https://github.com/alamgirustc/RacGK">https://github.com/alamgirustc/RacGK</a>.</p>

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

RacGK: Region-aware image captioning with LSTM and Gaussian kernel attention

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • Mohammad Alamgir Hossain,
  • ZhongFu Ye,
  • Md. Bipul Hossen,
  • Md. Atiqur Rahman,
  • Md Shohidul Islam,
  • Md. Ibrahim Abdullah

摘要

This paper studies sequential image captioning with stronger object grounding and nonlinear visual interaction modeling. We propose RacGK, which combines Region-Specific Feature Extraction (RSFE), a Region-Aware Attention Mechanism (RAAM), and Gaussian RBF-based Kernelized Self-Interaction Attention (KSIA) in an attention-guided LSTM decoder. On MS COCO 2014 with the Karpathy split, RacGK achieves competitive results, including 40.3 BLEU-4 and 59.4 ROUGE-L after CIDEr optimization. Ablations support the contribution of region-aware fusion and Gaussian kernel attention. Code is available at https://github.com/alamgirustc/RacGK.