Scholarlink

<p>Generating structured reports from event images is a crucial task across domains such as journalism, security analysis, and event documentation. Manual report writing is labor-intensive and time-consuming, motivating the need for automation. This paper introduces “Vision Narrate,” a novel AI-driven architecture that uniquely integrates state-of-the-art techniques in image captioning, context segmentation, and report generation. Unlike existing approaches, our system innovates by combining a fine-tuned VGG16-based encoder–decoder model for detailed image captioning with a new context identification strategy that employs MPNet sentence embeddings and a reverse sigmoid-weighted cosine similarity method to dynamically group captions into event-specific segments. Finally, a generative AI model is used to compose coherent, structured reports from these segmented captions. Experimental results on a dataset of 900 event images from 58 events demonstrate that our method achieves competitive performance with BLEU-4 (0.35), CIDEr (0.97), Pk (0.23), and WindowDiff (0.19) scores compared to traditional text segmentation baselines, while human evaluation yields an average coherence score of 4.74/5 and a fluency score of 4.88/5. Our findings indicate that “Vision Narrate” offers a scalable and efficient solution for automated report generation by uniquely fusing advanced techniques, paving the way for future enhancements in dataset diversity, factual accuracy, and multimodal context segmentation.</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

Vision Narrate: an AI architecture to write reports based on event images

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • Shashank Parmar,
  • Rahul,
  • Rakhee Sharma

摘要

Generating structured reports from event images is a crucial task across domains such as journalism, security analysis, and event documentation. Manual report writing is labor-intensive and time-consuming, motivating the need for automation. This paper introduces “Vision Narrate,” a novel AI-driven architecture that uniquely integrates state-of-the-art techniques in image captioning, context segmentation, and report generation. Unlike existing approaches, our system innovates by combining a fine-tuned VGG16-based encoder–decoder model for detailed image captioning with a new context identification strategy that employs MPNet sentence embeddings and a reverse sigmoid-weighted cosine similarity method to dynamically group captions into event-specific segments. Finally, a generative AI model is used to compose coherent, structured reports from these segmented captions. Experimental results on a dataset of 900 event images from 58 events demonstrate that our method achieves competitive performance with BLEU-4 (0.35), CIDEr (0.97), Pk (0.23), and WindowDiff (0.19) scores compared to traditional text segmentation baselines, while human evaluation yields an average coherence score of 4.74/5 and a fluency score of 4.88/5. Our findings indicate that “Vision Narrate” offers a scalable and efficient solution for automated report generation by uniquely fusing advanced techniques, paving the way for future enhancements in dataset diversity, factual accuracy, and multimodal context segmentation.