Stroke-Based Perception: Discover Novel Oracle Characters

Oracle bone script is recognized as the oldest ancient Chinese character, which dates back to the Shang Dynasty, providing invaluable information in understanding ancient civilizations. Current automatic recognition methods of oracle characters are often designed in a closed-world setting, where labeled and unlabeled classes are identical. However, novel oracle character categories are often acquired during staged excavation, demanding a generalized ability to decipher the new ones. In this paper, we tackle a novelOracleCharacterCategoryDiscovery (OCCD) task, which aims to identify known oracle categories and cluster novel ones by using the knowledge learned from known classes. We propose a novel semi-supervised learning method for OCCD, which leverages the shared knowledge learned from known and unseen categories as guidance to distill informative features for classification. Specifically, we introduce a keypoint matrix to encode the shared knowledge, which is iteratively learned in a coarse-to-fine manner by gradually removing redundant components. To obtain generalized features, we decouple the character features into parts and aggregate relevant information with guidance from the shared keypoint matrix. In addition, we introduce a character memory bank to support the information exchange between the previous and current updates, enhancing the features' robustness. Moreover, we develop multi-level constraints at character, stroke and feature levels to improve the discriminativeness, sharpness and independence of the features while effectively avoiding the collapse of representation. Extensive experimental results on oracle datasets demonstrate that the proposed model outperforms the related state-of-the-art. Code is available at:https://github.com/Clarence-CV/Oracle_GCD.