Scholarlink

<p>Ancient Tai Lue palm-leaf manuscripts commonly exhibit complex deteriorations, including irregular perforations, seal impressions, missing strokes, and blurring. This paper introduces MSPR-TL, a progressive restoration framework designed for such compound degradations. The network employs a three-level architecture operating at local, structural, and global scales, where multi-scale branches collaboratively integrate cross-level features to recover both fine details and overall layout consistency. A bidirectional feature fusion module facilitates inter-scale information exchange, while Swin Transformer blocks model long-range dependencies to enhance stroke reconstruction and suppress background interference. The training objective combines pixel fidelity, structural preservation, and perceptual quality through a dynamically weighted hybrid loss that incorporates a nonlinear diffusion regularizer. To support this work, we present TLADIRD2026, a dedicated dataset comprising palm-leaf and cotton paper manuscript images. Experimental evaluations on this dataset and public benchmarks demonstrate that MSPR-TL achieves competitive performance on multiple metrics, exhibiting robust generalization across diverse degradation patterns.</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

Multi-scale progressive Swin Transformer for ancient Tai Lue palm leaf manuscript restoration

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月12日

查看全文
  • Jingying Zhao,
  • Aoqi Jiang,
  • Hai Guo,
  • Shuyao Liu,
  • Zhenwei Guo

摘要

Ancient Tai Lue palm-leaf manuscripts commonly exhibit complex deteriorations, including irregular perforations, seal impressions, missing strokes, and blurring. This paper introduces MSPR-TL, a progressive restoration framework designed for such compound degradations. The network employs a three-level architecture operating at local, structural, and global scales, where multi-scale branches collaboratively integrate cross-level features to recover both fine details and overall layout consistency. A bidirectional feature fusion module facilitates inter-scale information exchange, while Swin Transformer blocks model long-range dependencies to enhance stroke reconstruction and suppress background interference. The training objective combines pixel fidelity, structural preservation, and perceptual quality through a dynamically weighted hybrid loss that incorporates a nonlinear diffusion regularizer. To support this work, we present TLADIRD2026, a dedicated dataset comprising palm-leaf and cotton paper manuscript images. Experimental evaluations on this dataset and public benchmarks demonstrate that MSPR-TL achieves competitive performance on multiple metrics, exhibiting robust generalization across diverse degradation patterns.