Linking unknown characters via oracle bone inscriptions retrieval

Retrieving useful information from existing collections of oracle bone rubbing images plays a pivotal role in the study of oracle bone inscription decipherment. However, current systems for processing oracle bone information rely on expert-curated databases, which entail a time-consuming and labor-intensive process. Moreover, solely depending on oracle bone databases fails to yield any relevant information about undeciphered characters. Therefore, to address these challenges, in this paper, we present a deep learning retrieval framework named LUC, specifically designed for searching arbitrary oracle bone characters (both deciphered and undeciphered). Specifically, LUC takes clear glyph images as input, which can be handwritten by users or downloaded from websites, and extracts similar characters from raw oracle bone rubbing images through feature extraction and metric learning. Furthermore, unlike conventional image retrieval frameworks, we introduce an additional domain-aware embedding module to bridge the significant domain gap between clear glyphs and image patches. This module utilizes domain-specific information to generate a set of oracle bone radical prototypes, enhancing the structural features of oracle bone characters. Lastly, to mitigate the impact of increased feature output dimensions on retrieval performance, we construct a novel loss function. This loss function, based on the principle of maximum coding rate in metric learning, alleviates the performance degradation caused by dimensionality increase. Importantly, we establish a customized oracle bone retrieval benchmark comprising known characters for training and unknown characters for testing. Extensive comparative experiments demonstrate that LUC achieves superior performance compared to other classical retrieval methods. Furthermore, experiments on three publicperson ReID benchmarks also verify the effectiveness and generalization of our method.