OBI-CMF: Self-supervised learning with contrastive masked frequency modeling for oracle bone inscription recognition

Oracle bone inscriptions (OBI) are precious historical records of the Shang Dynasty around 3000 years ago and have profound cultural significance. OBI recognition reveals the origin of Chinese characters and enriches linguistic theory. However, existing OBI images generally have problems with noise interference and intra-class variability, which pose significant challenges to the accuracy and efficiency of traditional manual recognition methods and supervised learning networks. To address this issue, we propose a novel self-supervised learning approach with contrastive masked frequency modeling for OBI recognition, named OBI-CMF. Firstly, the contrastive learning and masking model are integrated within the self-supervised learning framework so that our OBI-CMF can efficiently extract and learn both global and local features of OBI images. Secondly, inter-domain supervision is achieved by learning OBI features in the spatial and frequency domains, which significantly improves the robustness and generalization ability of our OBI-CMF. Finally, a linear classifier is trained using the representations obtained from the pre-trained model. Furthermore, based on the OBC306 dataset, we construct three subsets, OBIR-10, OBIR-100, and OBIR-ID, to validate the advantages of our OBI-CMF from the perspective of balanced and imbalanced data. The performance of our model is evaluated on three subsets of the OBC306 public dataset. Experimental results show that our OBI-CMF achieves 99.57% on OBIR-10 and 95.08% on OBIR-100, which outperforms some supervised and self-supervised learning methods in the OBI recognition tasks. Moreover, it is the first attempt to apply self-supervised learning based on the frequency domain for OBI recognition. We hope this success can provide a new avenue for the virtual protection of cultural relics.