Scholarlink

<p>Protein and moisture contents in brewing sorghum are critical indicators affecting Baijiu fermentation efficiency and base liquor quality consistency. However, conventional physicochemical assays are time-consuming and unsuitable for real-time on-site monitoring. This study proposes a hyperspectral reconstruction approach based on a multistage Fourier volumetric attention transformer (MFVAT), combined with machine learning models for low-cost, rapid, and accurate determination of sorghum physicochemical constituents. MFVAT incorporates a Fourier–volumetric attention block (FVAB), where volumetric self-attention (VolSA) models long-range dependencies across spectral bands, while Fourier modulation attention (FMA) recovers edge sharpness and fine textures in the frequency domain, achieving high-fidelity hyperspectral reconstruction from RGB images. Experimental results demonstrate that MFVAT outperforms the compared baseline networks (MST +  + and HRNet) under the present experimental settings (PSNR = 41.7390&#xa0;dB, MRAE = 0.0175). The protein prediction model (SNV + XGBoost, Rp<sup>2</sup> = 0.9174, RMSEP = 0.1368&#xa0;g/100&#xa0;g) and moisture prediction model (MSC + RF, Rp<sup>2</sup> = 0.9700, RMSEP = 0.1274&#xa0;g/100&#xa0;g) achieve prediction accuracy comparable to that of models developed using original full-wavelength spectra (protein: Rp<sup>2</sup> = 0.9229; moisture: Rp<sup>2</sup> = 0.9797). These findings suggest that visible-range hyperspectral reconstruction from RGB images can provide a practical and cost-efficient supplementary solution for nondestructive quality evaluation of liquor-making raw materials.</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

Multistage Fourier Volumetric Attention Transformer for Hyperspectral Reconstruction and Physicochemical Quality Prediction of Brewing Sorghum

  • 期刊论文

  • 发表日期：

    2026年7月13日

查看全文
  • Anying Cai,
  • Liangliang Xie,
  • Juan Wang,
  • Jianping Yang,
  • Haili Yang,
  • Xinjun Hu,
  • Manjiao Chen,
  • Hao Zhang,
  • Kaiyang Yuan,
  • Haonan Yi,
  • Xiang Wan,
  • Rongzhi Wang,
  • Xin Wang,
  • Jianping Tian,
  • Dan Huang,
  • Huibo Luo,
  • Shunbo Zhang,
  • Yuansong Peng

摘要

Protein and moisture contents in brewing sorghum are critical indicators affecting Baijiu fermentation efficiency and base liquor quality consistency. However, conventional physicochemical assays are time-consuming and unsuitable for real-time on-site monitoring. This study proposes a hyperspectral reconstruction approach based on a multistage Fourier volumetric attention transformer (MFVAT), combined with machine learning models for low-cost, rapid, and accurate determination of sorghum physicochemical constituents. MFVAT incorporates a Fourier–volumetric attention block (FVAB), where volumetric self-attention (VolSA) models long-range dependencies across spectral bands, while Fourier modulation attention (FMA) recovers edge sharpness and fine textures in the frequency domain, achieving high-fidelity hyperspectral reconstruction from RGB images. Experimental results demonstrate that MFVAT outperforms the compared baseline networks (MST +  + and HRNet) under the present experimental settings (PSNR = 41.7390 dB, MRAE = 0.0175). The protein prediction model (SNV + XGBoost, Rp2 = 0.9174, RMSEP = 0.1368 g/100 g) and moisture prediction model (MSC + RF, Rp2 = 0.9700, RMSEP = 0.1274 g/100 g) achieve prediction accuracy comparable to that of models developed using original full-wavelength spectra (protein: Rp2 = 0.9229; moisture: Rp2 = 0.9797). These findings suggest that visible-range hyperspectral reconstruction from RGB images can provide a practical and cost-efficient supplementary solution for nondestructive quality evaluation of liquor-making raw materials.