Scholarlink

<p>Ancient murals are cultural treasures with high research value, yet few are well preserved. The intricate textures within murals pose substantial challenges for super-resolution reconstruction. To address issues such as detail loss, color shifts, and inadequate noise control in mural super-resolution reconstruction, this study introduces a mural conditional diffusion model (MCDM) for enhanced image reconstruction. The model integrates three core modules: a residual feature distillation network for feature encoding and detail extraction, a residual self-attention module to enhance global consistency, and a Kolmogorov–Arnold-based implicit representation module for high-frequency detail reconstruction. Furthermore, this study establishes three training schemes across two datasets. Experiments show that MCDM performs best on the mixed dataset, with a PSNR of 23.3 dB and an SSIM of 0.8399. Tests across scenarios show that MCDM ensures smooth images while preserving details. Transfer learning further enhances performance by reducing noise and restoring fine features, providing guidance for future work.</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

Ancient mural super-resolution reconstruction based on conditional diffusion model for enhanced visual information

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年5月22日

查看全文
  • Yunsheng Chen,
  • Aiwu Zhang,
  • Feng Gao,
  • Juwen Guo

摘要

Ancient murals are cultural treasures with high research value, yet few are well preserved. The intricate textures within murals pose substantial challenges for super-resolution reconstruction. To address issues such as detail loss, color shifts, and inadequate noise control in mural super-resolution reconstruction, this study introduces a mural conditional diffusion model (MCDM) for enhanced image reconstruction. The model integrates three core modules: a residual feature distillation network for feature encoding and detail extraction, a residual self-attention module to enhance global consistency, and a Kolmogorov–Arnold-based implicit representation module for high-frequency detail reconstruction. Furthermore, this study establishes three training schemes across two datasets. Experiments show that MCDM performs best on the mixed dataset, with a PSNR of 23.3 dB and an SSIM of 0.8399. Tests across scenarios show that MCDM ensures smooth images while preserving details. Transfer learning further enhances performance by reducing noise and restoring fine features, providing guidance for future work.