ROOTS: recognizing oracle bone inscriptions via an organized tree structure

Abstract

Oracle bone inscription recognition is hindered by long-tailed class distributions and strong inter-class visual similarity. Existing augmentation and re-weighting methods modify samples or optimization while retaining competition within a flat label space. We propose ROOTS (Recognizing Oracle Bone Inscriptions via an Organized Tree Structure), which reorganizes 1771 fine-grained classes into 80 size-balanced superclasses using DINOv2 prototypes and size-aware Ward clustering. A coarse-to-fine visual pathway uses hierarchical prediction and winner-take-all masking; a vision-language pathway combines Qwen3-VL-2B with structured glyph descriptions and paired corpus-conditioned and visual-only instruction training. On a benchmark of 395,845 images from five public datasets, ROOTS achieves 95.28% Top-1 and 99.38% Macro Top-1 accuracy, a 14.22-point Macro Top-1 gain over the strongest non-ROOTS visual baseline. By restructuring the label space, ROOTS reduces head-tail competition and provides interpretable superclass and semantic cues for tail-class recognition.