Scholarlink

<p>This study addresses the decipherment of oracle bone script by formalizing the mapping between oracle bone glyphs and modern Chinese characters as an image-to-image translation problem. We design a dual-path generative framework combining Pix2PixGAN and CycleGAN to automatically produce candidate modern glyphs. The models are trained and tested on 1160 pairs of deciphered “oracle bone–simplified character” images, and then applied to 150 undeciphered inscriptions, forming a standardized pipeline for machine prediction and human–AI collaborative decoding. For an additional set of 160 deciphered samples, we compare AI-only, Human-only, and AI+Human outputs against standard glyphs using PSNR, SSIM and LPIPS, and further quantify human–AI synergy via multi-expert ratings and statistical analysis. Results show that GAN-assisted collaboration markedly outperforms both purely AI-based and purely manual outputs in structural fidelity, perceptual quality and expert acceptability, establishing a quantifiable and scalable paradigm for intelligent oracle bone decipherment and heritage-oriented digital humanities.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Human–computer collaborative approach to the decipherment of oracle bone inscriptions with generative adversarial networks

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月10日

查看全文
  • Shaoting Zeng,
  • Jiayue Bai,
  • Junyi Shi,
  • Yining Li,
  • Yujing Zhao,
  • Yifeng Shi

摘要

This study addresses the decipherment of oracle bone script by formalizing the mapping between oracle bone glyphs and modern Chinese characters as an image-to-image translation problem. We design a dual-path generative framework combining Pix2PixGAN and CycleGAN to automatically produce candidate modern glyphs. The models are trained and tested on 1160 pairs of deciphered “oracle bone–simplified character” images, and then applied to 150 undeciphered inscriptions, forming a standardized pipeline for machine prediction and human–AI collaborative decoding. For an additional set of 160 deciphered samples, we compare AI-only, Human-only, and AI+Human outputs against standard glyphs using PSNR, SSIM and LPIPS, and further quantify human–AI synergy via multi-expert ratings and statistical analysis. Results show that GAN-assisted collaboration markedly outperforms both purely AI-based and purely manual outputs in structural fidelity, perceptual quality and expert acceptability, establishing a quantifiable and scalable paradigm for intelligent oracle bone decipherment and heritage-oriented digital humanities.