Dual-branch segmentation of oracle bone drill chisel and its impact on group classification

Oracle bones are important artifacts from the Shang Dynasty. Segmenting the oracle bone drill chisel (OBDC) facilitates the identification of oracle bone periods, which is essential for studying cultural and population changes. However, due to variations in drilling techniques across different periods and millennia of burial, oracle bones have suffered varying degrees of damage, posing significant challenges for accurate OBDC segmentation. To address these issues, this paper proposes a Multi-Feature Global and Local Extraction Network (MFGL-Net) based on a dual-branch architecture, achieving precise segmentation of complex degraded OBDC. Specifically, a global-local collaborative encoder is designed to model long-range dependencies and capture fine-grained local features. Meanwhile, multi-scale context modeling is introduced to enhance target localization through multi-scale receptive fields. Furthermore, through cross-level feature collaboration and semantic reweighting, semantic information of the target area is mined and edge features are supplemented. We manually annotated images to create an OBDC dataset. Extensive experiments on this dataset demonstrate that MFGL-Net achieves state-of-the-art segmentation performance, particularly in terms of boundary precision and overall accuracy. Additionally, incorporating the segmentation results into the OBDC classification task improved classification accuracy from 82.72% to 85.19%, providing high-quality spatial priors for oracle bone chronological classification research.