Scholarlink

<p>To overcome the limitations of conventional quality assessment methods for Schisandra chinensis, we propose a Dual-Modal Cross-Attention Network (DMCA-Net) that integrates cross-modal attention, adaptive weighting, and CBAM modules. The network employs a dual-branch architecture to extract image and spectral features, a bidirectional cross-attention module to enable deep modality interaction, and an adaptive multi-level fusion strategy to dynamically balance the contributions of both modalities. On a self-built dataset of 639 samples, DMCA-Net achieved an average classification accuracy of 98.59% in stratified five-fold cross-validation, significantly outperforming the baseline concatenation fusion model (95.30%). These results demonstrate that the proposed fusion strategy effectively exploits the complementarity of hyperspectral and RGB image modalities, providing an efficient and reliable deep learning solution for non-destructive, automated quality grading of <i>Schisandra chinensis</i>.</p> Graphical Abstract <p></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

Non-destructive quality assessment of schisandra chinensis using a dual-modal cross-attention fusion network

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Pengtao Zhang,
  • Ao Mei,
  • Juan Xi,
  • Jiawei Zhang,
  • Dayang Liu,
  • Hao Lan

摘要

To overcome the limitations of conventional quality assessment methods for Schisandra chinensis, we propose a Dual-Modal Cross-Attention Network (DMCA-Net) that integrates cross-modal attention, adaptive weighting, and CBAM modules. The network employs a dual-branch architecture to extract image and spectral features, a bidirectional cross-attention module to enable deep modality interaction, and an adaptive multi-level fusion strategy to dynamically balance the contributions of both modalities. On a self-built dataset of 639 samples, DMCA-Net achieved an average classification accuracy of 98.59% in stratified five-fold cross-validation, significantly outperforming the baseline concatenation fusion model (95.30%). These results demonstrate that the proposed fusion strategy effectively exploits the complementarity of hyperspectral and RGB image modalities, providing an efficient and reliable deep learning solution for non-destructive, automated quality grading of Schisandra chinensis.

Graphical Abstract