ViT-based classification of oracle bones for periodization of the Shang Dynasty

Oracle bone inscriptions are the oldest systematically written characters in China, and their periodization and chronology are essential tasks in historical research. However, current deep-learning models face significant challenges when they are used to recognize oracle bone images directly. These models can only achieve an accuracy of no more than 45%. As the research on the periodization of oracle bones mainly relies on the characters inscribed on the bones, a domain knowledge-based oracle bone classification model is proposed that improves the recognition accuracy of Vision Transformer (ViT) by 5.28%. The proposed approach incorporates results obtained from oracle bone character detection to provide a more effective recognition object. Moreover, a novel model with a style network and category embedding is proposed in this paper that focuses on characters, resulting in a notable improvement in recognition accuracy of 16.01%. Our study reveals the importance of domain knowledge in the classification of historical documents, and it presents new insights into using artificial intelligence for assisting in historical document organization tasks.