Automatic generation method of oracle bone facsimiles based on improved CycleGAN network

摘要：

甲骨摹本作为记录甲骨文字形态与细节的重要载体，在汉字起源与演变研究中具有重要学术价值.针对传统获取方法耗时费力、受限于实物稀缺及人工绘制主观性强等问题，提出一种基于改进CycleGAN的甲骨摹本自动生成方法.通过引人多尺度特征提取模块(split conv concat,SCC)在不同尺度上同时捕获局部细节特征与全局语义信息.为进一步优化甲骨特征提取过程,利用风格强化注意力残差模块(enhanced style attention residualmodule，ESARM)实现甲骨关键特征的选择性增强并有效抑制甲骨图像背景噪声干扰.实验结果表明，本文方法显著提升了SSIM、PSNR等指标，在保持字形结构完整、实现细节精准重建的同时，使边缘清晰度与连续性接近标准摹本，适用于高标准甲骨摹本生成，为古文字研究提供支撑。

Oracle bone facsimiles. as important carriers for recording the forms and details of oracle bone characters.hold signilicant academic value in the study of the origin and evolution of Chinese characters. Addressing the issues of traditional acquisition methods being time consuming. labor-intensive, limited by the scarcity of physical specimens, and highly subjective due to manual drawing. this paper proposes an automatic oracle bone facsimiles generation method based on improved CycleGAN. By introducing a multi-scale feature extraction module(split cony concat, SCC), local detail features and global semantic information are simultaneously captured at different scales. To further optimize the oracle bone feature extraction process, a style-enhanced attention residual module(enhanced style attention residual module, ESARM) is utilized to selectively enhance key features of the oracle bones and effectively suppress background noise interference in the images. Experimental results show that the proposed method significantly improves SSIM and PSNR, preserving character topology and accurately reconstructing details with edge clarity and continuity approaching the standard reference. Our method is suitable for highstandard generation of oracle bone facsimiles, providing strong support for the study of ancient Chinese characters.

作者：

刘杰，王晨祥，盛文悦，张鑫，刘国奇

Liu Jie,Wang Chenxiang,Sheng Wenyue,Zhang Xin,Liu Guoqi

机构地区：

河南师范大学a.计算机与信息工程学院(人工智能学院)；b.甲骨智能计算实验室；c.软件学院；d.历史文化学院

引用本文：

刘杰，王晨祥，盛文悦等.基于改进CycleGAN网络的甲骨摹本自动生成方法[J].河南师范大学学报(自然科学版),2026,54(3) :44-51. (Liu Jie, Wang Chenxiang, Sheng Wenyue, et al. Automatic generation method of oracle bone facsimiles based on improved CycleGAN network[J].Joumal of Henan Normal University(Natural Science Edition),2026,54(3) :44-51.DOI:10.16366/j.cnki.1000-2367.2024.12.12.0003.)

基金：

国家自然科学基金青年科学基金；国家社会科学基金重大项目；河南省科技攻关项目；河南省高等学校智库研究项目

关键词：

甲骨摹本;CycleGAN;图像生成;多尺度特征融合;注意力模块

oracle bone facsimiles; CycleGAN; image generation; multi-scale feature fusion; attention mechanism

分类号：

TP391

基于改进 CycleGAN 网络的甲骨摹本自动生成方法.pdf