Leveraging progressive domain adaptation for unsupervised cross-domain oracle bone inscription recognition

Abstract Oracle Bone Inscriptions (OBIs) are the earliest mature Chinese writing system, irreplaceable for decoding ancient Chinese civilization. However, automated OBI recognition is severely hampered by a lack of annotated data, as surviving rubbings suffer from critical abrasion, fragmentation, and extreme scarcity. To address this challenge, we propose an unsupervised cross-domain OBI recognition framework that transfers glyph knowledge from labeled handwritten oracle characters to unlabeled scanned rubbings, reducing dependence on large-scale annotated rubbing data. Our framework is anchored by a Progressive Domain Adaptation Network (PDAN), built on feature-space linear interpolation and a dynamically-weighted gradient inversion mechanism, paired with a structure-aware target-domain augmentation strategy. We also propose UDA-HS-1K, a large-scale benchmark dataset for the field. Comprehensive experiments show that our method achieves state-of-the-art performance, outperforming previous work on Oracle-241 (74.5% vs. 61.8%) and UDA-HS-1K (43.5% vs. 38.4%), validating the effectiveness of our method and the practical utility of our dataset.