Scholarlink

<p>Automated interpretation of cultural heritage symbols requires shifting from texture matching to semantic reasoning. Chinese auspicious symbols, such as the “Five Blessings,” suffer from high intra-class variation, challenging data-driven approaches. We introduce FiveBlessNet, a knowledge-infused framework mimicking experts’ multi-scale observation. By integrating an Atrous Spatial Pyramid Pooling (ASPP) module for context capture and a Center-guided joint loss for feature compactness, our model enforces domain-grounded discriminability. Validated on FiveBless-1K—a curated benchmark of 1243 images annotated by art historians—the model achieves 86.74% balanced accuracy, outperforming baselines. Grad-CAM visualizations confirm the decision basis shifts from background clutter to culturally diagnostic parts, such as the symmetry of “囍” (Xǐ) and the morphology of the God of Longevity (shóu), establishing a reproducible paradigm for interpretable heritage digitization. Notably, the model shifts decision-making from background clutter to culturally diagnostic parts, and its lightweight ConvNeXt-Tiny backbone enables real-time inference for on-site heritage digitization.</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

Bridging perception and semantics for Chinese five blessings interpretation in heritage digitization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月30日

查看全文
  • Dan Han,
  • Xin Hu

摘要

Automated interpretation of cultural heritage symbols requires shifting from texture matching to semantic reasoning. Chinese auspicious symbols, such as the “Five Blessings,” suffer from high intra-class variation, challenging data-driven approaches. We introduce FiveBlessNet, a knowledge-infused framework mimicking experts’ multi-scale observation. By integrating an Atrous Spatial Pyramid Pooling (ASPP) module for context capture and a Center-guided joint loss for feature compactness, our model enforces domain-grounded discriminability. Validated on FiveBless-1K—a curated benchmark of 1243 images annotated by art historians—the model achieves 86.74% balanced accuracy, outperforming baselines. Grad-CAM visualizations confirm the decision basis shifts from background clutter to culturally diagnostic parts, such as the symmetry of “囍” (Xǐ) and the morphology of the God of Longevity (shóu), establishing a reproducible paradigm for interpretable heritage digitization. Notably, the model shifts decision-making from background clutter to culturally diagnostic parts, and its lightweight ConvNeXt-Tiny backbone enables real-time inference for on-site heritage digitization.