Scholarlink

<p>Thangka image captioning is crucial for cultural heritage preservation but challenging due to the visual and semantic complexity of Thangka paintings. Existing deep learning methods often fail to capture detailed features and semantic accuracy, leading to incomplete or incorrect captions. To address this, this paper proposes Geo-TCAM, a Thangka captioning model integrating topic modeling and geometry-guided spatial attention. The model adopts a multi-level feature integration strategy to enhance feature extraction of gestures and objects. By combining LDA topic weights and visual features (TIF), it incorporates external domain knowledge for improved semantic understanding. The GFSA module further enhances spatial layout recognition. Experiments show notable improvements—BLEU-1, BLEU-4, METEOR, and CIDEr increase by 11.9%, 17.1%, 9.7%, and 119.5%, respectively, over baselines. On the COCO dataset, Geo-TCAM also achieves competitive results, supporting accurate Thangka captioning and promoting digital cultural heritage preservation.</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

Geo-TCAM: a Thangka captioning method integrating topic modeling with geometry-guided spatial attention

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月7日

查看全文
  • Ping Zhong,
  • Wenjin Hu,
  • Yinqiu Zhao,
  • Fujun Zhang

摘要

Thangka image captioning is crucial for cultural heritage preservation but challenging due to the visual and semantic complexity of Thangka paintings. Existing deep learning methods often fail to capture detailed features and semantic accuracy, leading to incomplete or incorrect captions. To address this, this paper proposes Geo-TCAM, a Thangka captioning model integrating topic modeling and geometry-guided spatial attention. The model adopts a multi-level feature integration strategy to enhance feature extraction of gestures and objects. By combining LDA topic weights and visual features (TIF), it incorporates external domain knowledge for improved semantic understanding. The GFSA module further enhances spatial layout recognition. Experiments show notable improvements—BLEU-1, BLEU-4, METEOR, and CIDEr increase by 11.9%, 17.1%, 9.7%, and 119.5%, respectively, over baselines. On the COCO dataset, Geo-TCAM also achieves competitive results, supporting accurate Thangka captioning and promoting digital cultural heritage preservation.