Improved D-FINE model for element detection in Han portrait stone rubbings

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Han portrait stone rubbings are vital for understanding Han dynasty society. Element detection in these images is challenging due to weathering-induced blurring, low and uneven contrast, complex background noise, and significant element overlap. To address this, we constructed a specialized dataset of 2268 high-quality images, categorized into seven classes (beast, bird, human, fish, deity, mythical beast, and monster), and proposed an enhanced D-FINE model integrating three targeted modules. First, the Interleaved Large Kernel Global Attention (ILKGA) captures fine line-engraved textures while suppressing background noise. Second, the Contrast-Adaptive Position Enhancement (CAPE) module improves feature representation in low-contrast regions. Third, a Matchability-Aware Loss (MAL) enhances the detection of dense and small overlapping targets. The proposed model achieves an mAP@50–95 of 38.3% (+1.6% over the baseline), an mAP@50 of 56.5%, and an F1-Score of 69.8%, utilizing only 27.0 M parameters and 60.9 GFLOPs. It outperforms similar-scale state-of-the-art detectors like YOLOv11-M, YOLOv12-M, and LW-DETR-M.