An improved EfficientNetV2 with category-aware attention for ancient ceramic dating

Accurate dating of ancient ceramics is a long-standing archaeological challenge constrained by two bottlenecks: extreme long-tail distribution (50.10% Qing vs 0.3% pre-Qin) and complex fine-grained classification. Conventional models often underperform in underrepresented classes and overlook nuanced features. We propose a framework integrating a category-aware attention (CAA) module and focal loss into an improved EfficientNetV2 backbone. A comprehensive dataset of 50 391 high-resolution images across eight historical periods (pre-Qin to Qing) was curated, using stratified sampling (7:1:2 split) to ensure tail-class representation (e.g., 158 pre-Qin samples). CAA directs the model toward period-specific attributes (e.g., Song celadon glaze) via category prototypes, whereas focal loss mitigates majority-class gradient bias. Our framework achieves 88.59% overall accuracy and 70.35% Macro-F1 score, outperforming the baselines (ResNet50 and ConvNeXtV2) by 5.5%–8.3% in accuracy and 12.1%–15.7% in Macro-F1. It significantly enhances tail-class robustness (e.g., F1 = 0.77 for Wei–Jin/Northern and Southern Dynasties). Ablation studies confirmed that CAA and focal loss contributed 11.2% and 9.8% to Macro-F1, respectively. This study resolves long-tail challenges in ceramic dating, supporting preservation and chronology refinement. Future research will explore multimodal fusion and few-shot learning for data-scarce cultural-heritage tasks.