A Deep Network based on Dual Attention for Oracle Bone Inscription Character Recognition

Oracle bone inscriptions, as the earliest systematic written records in China, are invaluable for understanding ancient history and the evolution of writing. However, their diverse forms, irregular shapes, and ambiguous carvings pose significant challenges for recognition. To address these issues, we propose a Dual Attention Network (DANet) for oracle bone character recognition. DANet introduces a Dual Attention Module (DAM) consisting of Position Attention and Channel Attention, which effectively capture position and channel features to improve recognition accuracy. Experiments on the two public datasets HWOBC and OBC306 demonstrate that our proposed method achieves superior performance compared to classical deep learning models. The results show improvements in both Top-1 Accuracy and Top-5 Accuracy, confirming the effectiveness of the proposed framework.