Beyond aggregate metrics: diagnosing the long-tailed failure of oracle bone character detection and recognition

Abstract

Oracle bone inscriptions, the earliest mature Chinese writing system, are a primary source for studying ancient Chinese civilization, yet their interpretation remains labor-intensive. Automatically detecting and recognizing visible characters on rubbing images is therefore an entrance task for AI-assisted interpretation. The Oracle Bone Inscription Multi-modal Dataset exposes the problem of extreme long-tailed class distribution: aggregate evaluation conceals severe performance defects on rare classes—the detector can reach an overall AP50 of 0.63, while the few-shot class AP50 is only 0.06. We conduct a systematic diagnosis with a frequency-aware stratified evaluation protocol. Controlled comparisons show that localization is less limiting than rare-glyph recognition, and that in-head and crop-level remedies trade aggregate accuracy against tail accuracy. We thus propose FAR-YOLO, a minimal decoupled baseline that separates localization and recognition training, optimizing the recognition branch only with frequency-aware sampling and morphology-preserving augmentation. While maintaining nearly unchanged overall AP50, FAR-YOLO improves the few-shot AP50 by 10.5 percentage points. It ranks second on all four frequency-stratified metrics among representative long-tailed methods, with consistent tail gains across seven recognition backbones.