Scholarlink

Background <p>The Qingming Festival scroll stands as a fragile piece of cultural heritage, yet conventional digitisation methods tend to wash out finer details, such as brushstrokes and pigment textures. The study introduces Structure-Informed Learned-Kernel Super-Resolution (SILK-SR), and fuses it with hyperspectral spectral restoration to jointly recover brushwork structure and pigment fidelity in a single, scroll-aware workflow.</p> Methods <p>The Palace Museum in Beijing supplied 96 high-resolution scans of the Qingming Festival scroll, covering four distinct regions and focusing on four primary pigment types. For super-resolution, we used SILK-SR, a structure-informed learned-kernel model, trained on 12,000 image patches measuring 128 by 128 pixels. Performance was assessed using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). For spectral restoration, the team turned to hyperspectral imaging data spanning 400 to 1000 nanometres, integrating pigment identification through a hybrid method involving Support Vector Machine-Markov Random Field (SVM-MRF) models and Convolutional Neural Networks (CNNs).</p> Results <p>Mean resolution differed across regions (6.2 ± 0.8 to 7.1 ± 1.0&#xa0;μm/pixel; F(3,92) = 4.11, <i>p</i> = 0.037). Super-resolution improved PSNR from 28.6 ± 2.3 to 34.1 ± 2.5 decibels (t(46) = 7.42, <i>p</i> = 0.004) and SSIM from 0.71 ± 0.05 to 0.84 ± 0.06 (t(46) = 6.95, <i>p</i> = 0.007). Brushstroke recovery rose from 0.63 ± 0.07 to 0.78 ± 0.08 (t(46) = 5.84, <i>p</i> = 0.012). Spectral restoration achieved pigment classification above 93% and increased the tonal gradient range from 0.42 ± 0.06 to 0.57 ± 0.07 (t(40) = 6.02, <i>p</i> = 0.009). Integration yielded seam continuity of 93.2 ± 3.1% (F(2,84) = 9.11, <i>p</i> = 0.004), visibility indices rising from 2.9 ± 0.4 to 4.2 ± 0.5 (t(44) = 8.01, <i>p</i> = 0.001), and narrative detail nearly doubling in restored outputs.</p> Conclusion <p>The SILK-SR-based dual-pipeline links structure-informed super-resolution with hyperspectral pigment reconstruction to produce culturally legible restorations suitable for scholarly analysis and public access.</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

Super resolution and spectral restoration enable digital restoration and visual analysis of along the river during the Qingming Festival

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月20日

查看全文
  • Wang Xufeng

摘要

Background

The Qingming Festival scroll stands as a fragile piece of cultural heritage, yet conventional digitisation methods tend to wash out finer details, such as brushstrokes and pigment textures. The study introduces Structure-Informed Learned-Kernel Super-Resolution (SILK-SR), and fuses it with hyperspectral spectral restoration to jointly recover brushwork structure and pigment fidelity in a single, scroll-aware workflow.

Methods

The Palace Museum in Beijing supplied 96 high-resolution scans of the Qingming Festival scroll, covering four distinct regions and focusing on four primary pigment types. For super-resolution, we used SILK-SR, a structure-informed learned-kernel model, trained on 12,000 image patches measuring 128 by 128 pixels. Performance was assessed using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). For spectral restoration, the team turned to hyperspectral imaging data spanning 400 to 1000 nanometres, integrating pigment identification through a hybrid method involving Support Vector Machine-Markov Random Field (SVM-MRF) models and Convolutional Neural Networks (CNNs).

Results

Mean resolution differed across regions (6.2 ± 0.8 to 7.1 ± 1.0 μm/pixel; F(3,92) = 4.11, p = 0.037). Super-resolution improved PSNR from 28.6 ± 2.3 to 34.1 ± 2.5 decibels (t(46) = 7.42, p = 0.004) and SSIM from 0.71 ± 0.05 to 0.84 ± 0.06 (t(46) = 6.95, p = 0.007). Brushstroke recovery rose from 0.63 ± 0.07 to 0.78 ± 0.08 (t(46) = 5.84, p = 0.012). Spectral restoration achieved pigment classification above 93% and increased the tonal gradient range from 0.42 ± 0.06 to 0.57 ± 0.07 (t(40) = 6.02, p = 0.009). Integration yielded seam continuity of 93.2 ± 3.1% (F(2,84) = 9.11, p = 0.004), visibility indices rising from 2.9 ± 0.4 to 4.2 ± 0.5 (t(44) = 8.01, p = 0.001), and narrative detail nearly doubling in restored outputs.

Conclusion

The SILK-SR-based dual-pipeline links structure-informed super-resolution with hyperspectral pigment reconstruction to produce culturally legible restorations suitable for scholarly analysis and public access.