Scholarlink

<p>Oracle bone inscriptions, the earliest mature writing system in ancient China, require reliable artefact authentication for archaeological scholarship and museum curation. Traditional expert-based methods are subjective and difficult to reproduce at scale. We propose OracleNet, a few-shot multi-scale deep learning framework integrating ResNet-18, a Feature Pyramid Network, a Convolutional Block Attention Module, data augmentation, Focal Loss and Temperature Scaling calibration. Fivefold artefact-level cross-validation on 272 images yields a mean accuracy of 91.50% ± 4.20% and an ROC–AUC of 99.12% ± 0.80%. A positivity-constrained Temperature Scaling step reduces Expected Calibration Error from 0.269 to 0.054 (79.8% reduction) and Brier score by 52.6%. A symmetric calibrated comparison establishes post-hoc Temperature Scaling as a generalisable calibration tool under few-shot conditions. Sensitivity reaches 98.53%. The backbone-only variant outperforms the full model in raw accuracy on single-fold evaluation (McNemar <i>p</i> = 0.041), confirming that calibrated confidence is the primary contribution.</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

OracleNet: a few-shot multi-scale deep learning framework for calibrated authentication of oracle bone artefacts

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Xiaokui Liu,
  • Shengwei Han,
  • Niannian Liu,
  • Yushuang Liu,
  • An Guo,
  • Yongge Liu

摘要

Oracle bone inscriptions, the earliest mature writing system in ancient China, require reliable artefact authentication for archaeological scholarship and museum curation. Traditional expert-based methods are subjective and difficult to reproduce at scale. We propose OracleNet, a few-shot multi-scale deep learning framework integrating ResNet-18, a Feature Pyramid Network, a Convolutional Block Attention Module, data augmentation, Focal Loss and Temperature Scaling calibration. Fivefold artefact-level cross-validation on 272 images yields a mean accuracy of 91.50% ± 4.20% and an ROC–AUC of 99.12% ± 0.80%. A positivity-constrained Temperature Scaling step reduces Expected Calibration Error from 0.269 to 0.054 (79.8% reduction) and Brier score by 52.6%. A symmetric calibrated comparison establishes post-hoc Temperature Scaling as a generalisable calibration tool under few-shot conditions. Sensitivity reaches 98.53%. The backbone-only variant outperforms the full model in raw accuracy on single-fold evaluation (McNemar p = 0.041), confirming that calibrated confidence is the primary contribution.