Year: 2026
Type: article
Abstract: Oracle bone inscriptions represent the earliest known mature Chinese writing system and constitute invaluable primary sources for studying Shang-dynasty history; yet automated recognition remains hindered by an intrinsically long-tailed class distribution, in which common characters vastly outnumber rare ones, as well as by high inter-class visual similarity. Existing approaches address this imbalance primarily through data augmentation or re-weighting, leaving the underlying label-space skewness largely intact. Here we propose ROOTS (Recognizing Oracle Bone Inscriptions via an Organized Tree Structure), a hierarchical framework that restructures the original 1,771 fine-grained classes into 80 size-balanced superclasses via DINOv2-based prototype extraction and size-aware hierarchical clustering, reducing the mean-to-median class-size ratio from 6.6× to 1.2×. Recognition then proceeds coarse-to-fine: a shared root branches into semantic superclasses before resolving individual characters, with a winner-take-all mask confining predictions to the relevant branch. We further integrate a vision-language pathway in which Qwen3-VL, guided by structured glyph descriptions, performs fine-grained discrimination under stochastic prompt regularization. On a mixed benchmark of 395,845 images spanning five public corpora, ROOTS achieves 95.28% Top-1 accuracy and 99.38% Macro Top-1 accuracy, an absolute gain of 14.2 percentage points over the strongest visual baseline, while lifting tail-class accuracy to 99.83%. These results establish ROOTS as a robust, interpretable solution for long-tailed ancient-script recognition with potential extensibility to other logographic writing systems.
Source: npj Heritage Science
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Award: G9000381079
OpenAlex ID: W7214067223
FWCI:
Cites:
Cited by:
Related to: