Ora-NSC: A Novel Semi-supervised Approach for Oracle Bone Fragment Classification with Imbalanced Classes

Yunyue Hu, Weijie Gao, Yijing Si, Yue Zhang, Jie Liu2025 · DOI

Abstract

Oracle bone inscriptions (OBI), as precious historical records of The Shang Dynasty, approximately 3,000 years ago, marked the origins of Chinese characters and the development of the Shang civilization. Classifying oracle bone fragments is essential for enhancing fragment assembly efficiency and facilitating the interpretation of divination. To address data imbalance and label scarcity in oracle bone fragment datasets, we propose a novel semi-supervised approach for oracle bone fragment classification, which is named Ora-NSC. The approach integrates Mean Teacher with FixMatch, utilizing the Exponential Moving Average (EMA) strategy to enhance the accuracy and stability of pseudo-label prediction. Additionally, an improved loss function is developed to address class imbalances in fragment segments, thereby further enhancing classification performance. Experimental results demonstrate that our method achieves superior performance on both the Oracle Fragments and CIFAR-100 datasets compared to several state-of-the-art approaches. It improves precision by 4.14%, recall by 4.03%, and F1-score by 4.46%, achieving an overall accuracy of 84.44%. This study represents the first application of semi-supervised learning for oracle bone fragment recognition, offering new possibilities for the preservation of cultural heritage.
BibTeX
@misc{hu2025ora,
  author = {Yunyue Hu and Weijie Gao and Yijing Si and Yue Zhang and Jie Liu},
  title = {{Ora-NSC: A Novel Semi-supervised Approach for Oracle Bone Fragment Classification with Imbalanced Classes}},
  year = {2025},
  doi = {10.1145/3743093.3770992},
}

Loading related content...

Figure