OBDC-Unet: a U-shaped network leveraging wavelet dynamics and multi-scale receptive fields for oracle bone drill chisel segmentation

Precise segmentation of oracle bone drill chisel (OBDC) regions is crucial for Shang Dynasty archeological dating and historical analysis, yet long-term burial often results in fractures, blurred boundaries, and irregular morphologies that hinder accurate extraction. To address these challenges, we propose OBDC-Unet, a U-shaped segmentation network integrating wavelet dynamics and multi-scale receptive fields for OBDC segmentation. The network incorporates Multi-directional Parallel Convolution (MPC), a Feature Enhancement Fusion (FEF) module combining WTConv and MLPs for cross-scale interaction, and a Global Context Channel Attention (GCCA) module to strengthen key feature modeling. Experiments on the OBDC dataset show that OBDC-Unet achieves an IoU of 90.91%, outperforming state-of-the-art methods and demonstrating robust performance in complex preservation conditions. To our knowledge, this study is the first to integrate wavelet dynamics with deep learning for oracle bone analysis, providing an effective tool for automated archeological documentation.