Scholarlink

<p>Spectral rendering approaches are nowadays the most accurate method for predictively simulating the light transport equation with advanced wavelength-dependent phenomena such as light dispersion or metamerism. However, these approaches require a much larger number of samples and longer computational times than standard trichromatic ones—e.g., expressed in RGB color spaces—to properly sample the visible spectrum domain and achieve noise-free images. To address these issues, denoising solutions can be applied to a spectral renderer in order to remove potential color noise, while ensuring that the reconstruction can properly conserve wavelength-dependent phenomena. We introduce a deep learning-based denoising framework tailored to spectral Monte Carlo rendering. Instead of operating on trichromatic RGB data, our approach performs denoising directly in the spectral domain, where radiance is discretized into an arbitrary number of wavelength bins, allowing wavelength-dependent effects to be preserved. To remain agnostic to the spectral resolution, each spectral bin is denoised independently by a light spectrum features denoiser (LSFD), while a compact parametric spectral representation, obtained through a PCA-like whitening transformation, provides global spectral context as auxiliary input to ensure spectral coherence. A second network, referred to as the details sharpener (DS), further refines the output by reintroducing high-frequency details guided by a consistently discretized spectral albedo. By combining spectral radiance, spectral albedo, and geometric features, our method effectively reduces chromatic noise at low sample counts while preserving radiometric accuracy and consistently outperforms state-of-the-art RGB denoising techniques. These results thus open the door for the industrial use of spectral predictive rendering in a quasi-interactive fashion. In accordance with FAIR principles, related code is available at <a href="https://github.com/laurent-lu/SMCImage-Denoising">https://github.com/laurent-lu/SMCImage-Denoising</a>.</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

Spectral monte carlo denoising

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Noizet Mathieu,
  • Rouphael Robin,
  • Deleau Hervé,
  • Prévost Stéphanie,
  • Steffenel Luiz-Angelo,
  • Lucas Laurent

摘要

Spectral rendering approaches are nowadays the most accurate method for predictively simulating the light transport equation with advanced wavelength-dependent phenomena such as light dispersion or metamerism. However, these approaches require a much larger number of samples and longer computational times than standard trichromatic ones—e.g., expressed in RGB color spaces—to properly sample the visible spectrum domain and achieve noise-free images. To address these issues, denoising solutions can be applied to a spectral renderer in order to remove potential color noise, while ensuring that the reconstruction can properly conserve wavelength-dependent phenomena. We introduce a deep learning-based denoising framework tailored to spectral Monte Carlo rendering. Instead of operating on trichromatic RGB data, our approach performs denoising directly in the spectral domain, where radiance is discretized into an arbitrary number of wavelength bins, allowing wavelength-dependent effects to be preserved. To remain agnostic to the spectral resolution, each spectral bin is denoised independently by a light spectrum features denoiser (LSFD), while a compact parametric spectral representation, obtained through a PCA-like whitening transformation, provides global spectral context as auxiliary input to ensure spectral coherence. A second network, referred to as the details sharpener (DS), further refines the output by reintroducing high-frequency details guided by a consistently discretized spectral albedo. By combining spectral radiance, spectral albedo, and geometric features, our method effectively reduces chromatic noise at low sample counts while preserving radiometric accuracy and consistently outperforms state-of-the-art RGB denoising techniques. These results thus open the door for the industrial use of spectral predictive rendering in a quasi-interactive fashion. In accordance with FAIR principles, related code is available at https://github.com/laurent-lu/SMCImage-Denoising.