Detecting unknown and classifying known oracle bone characters via novel data augmentation

This study tackles oracle bone character (OBC) recognition, a challenging task due to the script’s antiquity and variability. Despite recent progress, real-world deployment remains limited, particularly in handling out-of-distribution (OOD) characters. We frame this as two sub-tasks: detecting unknown OBCs and classifying known ones, noting that prior work largely neglects the former despite its archaeological significance. To address these issues, we propose two data augmentation strategies: Dynamic GridMask, which simulates occlusion and structural randomness, and intra-class image fusion, which reduces the adverse effects of conventional image fusion on OOD detection. Experiments on Oracle-MNIST, OBC306, OBI125, and Oracle-241 demonstrate consistent improvements, with classification accuracy gains up to 2.41% and AUROC increases up to 4.50%. These results highlight a practical and robust framework for OBC recognition in open-world scenarios.