Scholarlink

<p>Chinese mural faces embody significant cultural, religious, and historical values, but their preservation is challenged by natural degradation and data scarcity. Existing generation methods struggle with poor detail retention, style inconsistency, and overfitting due to limited data. To address these limitations, we propose an enhanced StyleGAN2 architecture integrating FreqSplitAttention and Dual Mask Discriminator (DMD). FreqSplitAttention employs Fourier transforms to decouple and process high- and low-frequency components independently enhancing both global structural coherence and fine-grained texture reconstruction. The DMD operates in spatial and spectral domains to mitigate overfitting and shortcut learning. Furthermore, We construct the Dunhuang Mural Face (DMF) dataset, comprising 9552 high-quality images at 256 × 256 resolution. Experimental results demonstrate that our method outperforms state-of-the-art models on DMF, achieving superior performance with an FID of 27.74, KID of 10.22, and IS of 2.19. Our work offers a robust solution for mural face generation and contributes to digital cultural heritage preservation.</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

Enhanced Chinese mural face generation via FreqSplitAttention and dual mask discriminator

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年5月18日

查看全文
  • Qiangying Huang,
  • Liangfei Cheng,
  • Xiayong Li,
  • Weikang Zhong,
  • Hao Wu,
  • Guowu Yuan

摘要

Chinese mural faces embody significant cultural, religious, and historical values, but their preservation is challenged by natural degradation and data scarcity. Existing generation methods struggle with poor detail retention, style inconsistency, and overfitting due to limited data. To address these limitations, we propose an enhanced StyleGAN2 architecture integrating FreqSplitAttention and Dual Mask Discriminator (DMD). FreqSplitAttention employs Fourier transforms to decouple and process high- and low-frequency components independently enhancing both global structural coherence and fine-grained texture reconstruction. The DMD operates in spatial and spectral domains to mitigate overfitting and shortcut learning. Furthermore, We construct the Dunhuang Mural Face (DMF) dataset, comprising 9552 high-quality images at 256 × 256 resolution. Experimental results demonstrate that our method outperforms state-of-the-art models on DMF, achieving superior performance with an FID of 27.74, KID of 10.22, and IS of 2.19. Our work offers a robust solution for mural face generation and contributes to digital cultural heritage preservation.