Machine learning study demonstrates stroke and structure priors improve few-shot recognition accuracy in oracle bone inscriptions, highlighting the utility of morphology-aware representations.
Oracle Bone Inscriptions (OBIs) recognition is challenging because glyph classes are numerous, long-tailed, and morphologically variable. We compare repository-based Teacher supervision with stroke and structure priors for class-disjoint few-shot OBI recognition. On the OBC306 dataset, using 42 classes, no-replacement sampling, and 10,000 episodes per shot, original-Teacher, no-Teacher FEAT, and glyph-prior FEAT achieve 74.31/75.44/77.56% at 1 shot and 88.99/89.88/91.67% at 50 shots across three seeds. Glyph priors improve no-Teacher FEAT by 1.79–2.15 points; paired 95% intervals exclude zero. Teacher targets are nearly uniform. Objective-weighting and label-agreement controls recover most of the gap, whereas a balanced subset reverses the effect. On near-saturated HWOBC, seed-0 gains are only 0.015–0.044 points. Baseline, query, module, complexity, and qualitative controls bound conclusions. Morphology-aware features improve OBI few-shot recognition, while Teacher utility depends on repository, objective, and data regime.
No takes yet. Share an insight, caveat, or question.
Xiong et al. (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: