Researchers have introduced a new deep learning architecture designed to improve the segmentation of oracle bone drill chisels, a critical component in analyzing artifacts from China’s Shang Dynasty. The model, termed the Multi-Feature Global and Local Extraction Network (MFGL-Net), employs a dual-branch structure to address the significant challenges posed by physical degradation and variations in ancient drilling techniques. By achieving high-precision segmentation, the study provides essential spatial priors that enhance the chronological classification of these historically significant bones.
Oracle bones serve as vital primary sources for studying the social, cultural, and political landscape of the Shang Dynasty. However, the drill chisel marks—the physical remnants of the tools used to create these artifacts—are often obscured or damaged after millennia of burial. Accurate identification of these features is essential for researchers to assign oracle bones to specific time periods, yet existing imaging and analysis techniques frequently struggle with the complexities of these degraded surfaces.
To overcome these limitations, the design of MFGL-Net centers on a global-local collaborative encoder. This architecture is specifically engineered to model long-range dependencies across the bone surface while simultaneously capturing fine-grained local features of the chisel marks. By integrating these two perspectives, the network compensates for the loss of information caused by surface wear and damage, allowing for more reliable identification of target areas.
The researchers further enhanced the model’s capabilities by introducing multi-scale context modeling. This approach utilizes multi-scale receptive fields to strengthen target localization, ensuring the network can distinguish between genuine chisel marks and other surface artifacts. Additionally, the network incorporates semantic reweighting and cross-level feature collaboration to supplement edge features, which are often the first to disappear during artifact decay.
Extensive testing on a manually annotated dataset revealed that MFGL-Net achieves state-of-the-art performance in segmentation tasks. The researchers focused particularly on boundary precision and overall accuracy, two metrics that have historically been difficult to maintain in the presence of archaeological degradation. The data suggests that this refined segmentation is robust enough to handle the complex, noisy images characteristic of excavated oracle bone fragments.
The impact of this segmentation improvement extends directly into the classification process. When the team integrated the output of the MFGL-Net into their classification workflow, they observed a measurable increase in accuracy, with the success rate rising from 82.72% to 85.19%. This improvement suggests that precise segmentation provides high-quality spatial information that is critical for the rigorous classification of oracle bone chronology.
By automating the identification of these subtle physical markers, the proposed network offers a scalable tool for processing the vast collections of oracle bones found in archaeological sites. The findings demonstrate that deep learning-based spatial analysis can provide a reliable foundation for cultural heritage research, helping to standardize the study of artifact distribution and population history across different phases of the Shang Dynasty.
Subject of Research: Artificial intelligence applications in archaeological artifact analysis.
Article Title: Dual-branch segmentation of oracle bone drill chisel and its impact on group classification
Article References: Liu, G., Kang, X., Han, Y., Li, W., Yang, Y., Li, X., Ru, L., Song, L., & Liu, D. (2026). Dual-branch segmentation of oracle bone drill chisel and its impact on group classification. npj Heritage Science. https://doi.org/10.1038/s40494-026-03006-4
Image Credits: AI Generated
DOI: 10.1038/s40494-026-03006-4
Keywords: Oracle bones, Shang Dynasty, Archaeological informatics, Image segmentation, Deep learning, MFGL-Net, Drill chisel analysis, Cultural heritage, Pattern recognition, Chronological classification, Spatial priors, Artifact preservation, Semantic segmentation, Multi-feature extraction, Computational archaeology
Cite Scienmag News
Courtney Benton. (November 16, 2026). Dual-Branch AI Network Improves Oracle Bone Drill Chisel Segmentation Accuracy. Scienmag. https://scienmag.com/dual-branch-ai-network-improves-oracle-bone-drill-chisel-segmentation-accuracy/
Courtney Benton. "Dual-Branch AI Network Improves Oracle Bone Drill Chisel Segmentation Accuracy." Scienmag, 16 November 2026, https://scienmag.com/dual-branch-ai-network-improves-oracle-bone-drill-chisel-segmentation-accuracy/. Accessed 16 November 2026.
Courtney Benton. "Dual-Branch AI Network Improves Oracle Bone Drill Chisel Segmentation Accuracy." Scienmag. November 16, 2026. https://scienmag.com/dual-branch-ai-network-improves-oracle-bone-drill-chisel-segmentation-accuracy/








