Scholarlink

<p>The widespread use of packaging in commercial food products introduces optical occlusion and spectral mixing that hinder non-destructive internal quality assessment. Therefore, we designed an integrated spatially offset Raman spectroscopy (SORS)-deep learning pipeline and validated its effectiveness using butter adulteration quantification as a representative through-packaging application. First, adulterated mixtures (10–90% w/w margarine, 10% increments), measured both unpackaged and under representative packaging, were acquired with a line-scan SORS system and the resulting scattering regions were subjected to signal-to-noise ratio optimization. Second, an Attention-Gated U-Net (AGUNet) was constructed to generate Raman spectra, aiming to strengthen sparse and low-intensity Raman data. Finally, the processed spectra were fed into a multiscale dilated transformer network (MDTNet), which leverages the receptive-field expansion of dilated convolutions and incorporates a spectral-angle loss to enable accurate quantitative prediction of margarine content. Experiments across four representative packaging types demonstrate that the proposed AGUNet generator followed by the MDTNet regressor outperforms alternative generative strategies and a suite of machine and deep baselines in both spectral-reconstruction fidelity and concentration-estimation accuracy and robustness. Experimental results demonstrated that the AGUNet-MDTNet pipeline achieved exceptional prediction performance, yielding <i>R</i><sup><i>2</i></sup> values of 0.998 and 0.973 for unpackaged and packaged systems, respectively, and outperforming alternative generative strategies with average RPD gains of up to 270%. Moreover, systematic evaluation of the model’s performance under reduced experimental sample sizes demonstrates that synthetically generated spectra can effectively compensate for the loss of information caused by packaging-induced signal attenuation, enabling high predictive accuracy even with substantially fewer experimental samples. Overall, the integrated SORS-deep learning framework provides a practical and scalable route for rapid, non-destructive assessment of product integrity in diverse packaged food systems.</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

A Generative-Transformer Framework for SORS-Based Butter Adulteration Quantification

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Zhenfang Liu,
  • Jinlei Li,
  • Zhen Bi,
  • Jungang Lou,
  • Qing Shen,
  • Xiongtao Zhang,
  • Xin Wang

摘要

The widespread use of packaging in commercial food products introduces optical occlusion and spectral mixing that hinder non-destructive internal quality assessment. Therefore, we designed an integrated spatially offset Raman spectroscopy (SORS)-deep learning pipeline and validated its effectiveness using butter adulteration quantification as a representative through-packaging application. First, adulterated mixtures (10–90% w/w margarine, 10% increments), measured both unpackaged and under representative packaging, were acquired with a line-scan SORS system and the resulting scattering regions were subjected to signal-to-noise ratio optimization. Second, an Attention-Gated U-Net (AGUNet) was constructed to generate Raman spectra, aiming to strengthen sparse and low-intensity Raman data. Finally, the processed spectra were fed into a multiscale dilated transformer network (MDTNet), which leverages the receptive-field expansion of dilated convolutions and incorporates a spectral-angle loss to enable accurate quantitative prediction of margarine content. Experiments across four representative packaging types demonstrate that the proposed AGUNet generator followed by the MDTNet regressor outperforms alternative generative strategies and a suite of machine and deep baselines in both spectral-reconstruction fidelity and concentration-estimation accuracy and robustness. Experimental results demonstrated that the AGUNet-MDTNet pipeline achieved exceptional prediction performance, yielding R2 values of 0.998 and 0.973 for unpackaged and packaged systems, respectively, and outperforming alternative generative strategies with average RPD gains of up to 270%. Moreover, systematic evaluation of the model’s performance under reduced experimental sample sizes demonstrates that synthetically generated spectra can effectively compensate for the loss of information caused by packaging-induced signal attenuation, enabling high predictive accuracy even with substantially fewer experimental samples. Overall, the integrated SORS-deep learning framework provides a practical and scalable route for rapid, non-destructive assessment of product integrity in diverse packaged food systems.