OracleGCD: Generalized Category Discovery for Oracle Bone Scripts

Hetao Wu, Kunchi Li, Xu-Yao Zhang, Qiufeng Wang, Da-Han Wang

Abstract: Oracle bone inscriptions, the oldest known Chinese writing system, pose formidable challenges in character recognition and interpretation due to the large number of undeciphered characters and the scarcity of labeled data. In this paper, we introduce OracleGCD, an enhanced framework that adapts Generalized Category Discovery (GCD) to oracle bone inscriptions, which simultaneously handles the recognition of known characters and the discovery and clustering of novel categories. Specifically, OracleGCD uses a stroke-aware asymmetric view augmentation mechanism, which utilizes dual-grained (coarse- and fine-grained) augmentations to extract more discriminative features. Coarse-grained augmentation applies data transformations to the entire image while fine-grained one first uses adaptive threshold segmentation to extract salient stroke regions and then applies targeted transformations to these key areas to optimize local feature representation. Furthermore, we propose the confidence-guided collaborative feature learning, which predicts the pseudo-labels of unlabeled samples based on the logit output of the model, enabling the use of unlabeled data in supervised contrastive learning. We also implement a logit adjustment strategy to address the long-tailed class distribution in the dataset. Extensive experiments demonstrate that our framework sets new performance benchmarks for the discovery of oracle bone character. It significantly outperforms strong baselines in both known character recognition and novel category identification and cluster, advancing the automated analysis of oracle bone inscriptions and potentially aiding in deciphering ancient Chinese scripts.

External IDs:dblp:conf/icdar/WuLZWW25