Generating oracle bone inscriptions based on the structure aware diffusion model

The unearthed oracle bone fragments containing inscriptions often suffer from vagueness, absence, damage, etc., which significantly obstructs the interpretation of oracle bone inscriptions.While GANs have been applied to the oracle bone inscriptions restoration (OBIR), their performance often degrades in intricate regions due to instability during training (e.g. model collapse).Thus, we propose a new Structure Aware Diffusion Model (SADM) for the OBIR task, which aims to generate more accurate inscriptions. Specifically, we first introduce a progressive Gaussian mask to simulate damage on character rubbings. Subsequently, we propose an Oracle Bone Sampling algorithm (OBS) to progressively recover the oracle bone inscriptions. Then, we develop an adaptive dynamic adjustment mechanism to perceive the hierarchical structure of the reconstructed image. Finally, experiments on several public datasets prove the method’s effectiveness. Our method achieves 41.3% lower FID on OBC306 (64 × 64) compared to the Cold Diffusion model.