首页|期刊导航|深圳大学学报(理工版)|Topo-Diff:基于拓扑引导扩散模型的甲骨文转换方法

Topo-Diff:基于拓扑引导扩散模型的甲骨文转换方法OA

Topo-Diff:a topology-guided diffusion model for oracle bone script transformation

中文摘要英文摘要

针对甲骨文因载体残损、严重噪声干扰及字形变异导致难以准确考释与数字化复原的问题,提出一种基于拓扑引导的扩散模型(topology-guided diffusion model,Topo-Diff).构建一个从甲骨文到现代汉字的结构保持型复原框架,利用持续同调技术从待考释的甲骨文图像中提取对噪声具有高度鲁棒性的拓扑不变量,将其量化为多尺度的结构特征向量.通过加性投影机制将该拓扑特征注入以U-Net为噪声预测网络扩散模型的逆向去噪过程,显式引导模型在重构像素纹理的同时,严格遵循原始字形的连通分支与孔洞结构约束.模型采用联合噪声预测、拓扑一致性与边缘几何损失的多目标优化策略,有效克服了传统方法在处理断笔与粘连时的"结构幻觉"问 题.Topo-Diff在1 762个字类的测试设置中,结构相似性指数(structural similarity index measure,SSIM)、峰值信噪比(peak signal to noise ratio,PSNR)、拓扑相似度和识别准确率分别为0.702 4、25.35 dB、0.933 2和0.660,优于所比较的图像翻译和生成方法.Topo-Diff能够在保持甲骨文字形结构完整性的同时,有效提升数字化复原质量与识别可靠性,为甲骨文智能释读和数字化保护提供了一种可行的新方法.

Aiming at the difficulties in accurate decipherment and digital restoration of oracle bone scripts(OBSs)caused by carrier damage,severe noise interference and glyph variations,we propose a topology-guided diffusion model named Topo-Diff.In this model,we construct a structure-preserving restoration framework to convert OBSs into modern Chinese characters.Persistent homology is adopted to extract noise-robust topological invariants from the OBS images,which are further quantified into multi-scale structural feature vectors.By using an additive projection mechanism,these topological features are then injected into the reverse denoising process of the diffusion model with U-Net as the noise prediction network.This explicitly guides the model to reconstruct pixel textures while strictly complying with the structural constraints of connected components and holes of original glyphs.The model adopts a multi-objective optimization strategy combining noise prediction loss,topological consistency loss,and edge geometric loss.It effectively alleviates the"structural hallucination"issue of conventional methods when handling broken strokes and stroke adhesion.Tested under the setting of 1 762 character classes,Topo-Diff achieves a structural similarity index measure(SSIM)of 0.702 4,a peak signal to noise ratio(PSNR)of 25.35 dB,a topological similarity of 0.933 2,and a recognition accuracy of 0.660.Experimental results demonstrate that the proposed method outperforms the compared image-translation and generative approaches.Topo-Diff effectively improves the quality of digital restoration and recognition reliability while preserving the structural integrity of OBS characters,offering a viable new approach for their intelligent interpretation and digital preservation.

刘光南;秦广军;杨子旋;吴梦洁;梁小云

北京联合大学智慧城市学院 北京 100101北京联合大学智慧城市学院 北京 100101北京联合大学智慧城市学院 北京 100101北京联合大学智慧城市学院 北京 100101北京联合大学智慧城市学院 北京 100101

信息技术与安全科学

计算机应用技术甲骨文扩散模型拓扑数据分析结构保持生成持续同调智能考古

computer application technologyoracle bone script(OBS)diffusion modeltopological data analysis(TDA)structure-preserving generationpersistent homologyintelligent archaeology

《深圳大学学报(理工版)》 2026 (5)

541-548,8

National Natural Science Foundation of China(62172045) 国家自然科学基金资助项目(62172045)

10.3724/SP.J.1249.2026.05541

评论