| 摘 要: 甲骨文作为中国最早的成熟文字系统,其科学分期断代对历史学与语言文字学研究至关重要。传统专家分类方法存在效率低、主观性强及难以应对大规模数据等问题。因此,提出一种深度学习方法,通过融合“组别-时期-贞人”甲骨学元数据实现甲骨文组别的智能分类。构建了含24个组别、24889个单字图像的数据集,并设计了基于 YOLOv11的深度学习模型。该模型集成 CoordAtt注意力机制以增强对甲骨文字形细微空间特征的捕捉能力;采用元数据嵌入策略,将甲骨学先验知识融入模型训练。实验结果显示,该方法在甲骨文组别分类任务上取得了78.06%的 Top-1准确率和92.35%的 Top-5准确率,性能优于对比基准。 |
| 关键词: 甲骨文组别分类 深度学习 YOLOv11 元数据嵌入 |
|
中图分类号: TP391.4
文献标识码: A
|
| 基金项目: 国家自然科学基金项目资助(62306172);陕西科技大学国际化教育教学改革研究项目资助(GJ22YB09);陕西科技大学教学改革项目资助(23Y080) |
|
| Research on Intelligent Oracle Bone Script Group Classification and Metadata Embedding Based on Deep Learning |
|
WANG Weiling1, LI Jian1, WANG Shuai2
|
(1.School of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, Xi’an 710021, China;
2.School of History and Civilization, Shaanxi Normal University, Xi’an 710062, China)
20733838@qq.com; lijianjsj@sust.edu.cn; 447561195@qq.com
|
| Abstract: As the earliest mature writing system in China, the scientific periodization and dating of Oracle Bone Script (OBS) are crucial for historical and philological research. Traditional exper-t based classification methods suffer from low efficiency, high subjectivity, and difficulty in handling large-scale data. This study proposes a deep learning framework that integrates Oracle Bone Studies metadata—specifically “group-period-diviner”information—to achieve intelligent classification of OBS groups. A dataset comprising 24 889 individual character images across 24 groups was constructed, and a deep learning model based on YOLOv11 was designed. This model integrates the CoordAtt attention mechanism to enhance the capture capability of subtle spatial features in oracle bone character forms and employs a metadata embedding strategy to incorporate prior knowledge from Oracle Bone Studies into the model training process.Experimental results demonstrate that this method achieves a Top-1 accuracy of 78.06% and a Top-5 accuracy of 92.35% on the OBS group classification task, outperforming existing benchmarks. |
| Keywords: oracle bone script group classification deep learning YOLOv11 metadata embedding |
