Scholarlink

<p>Cassava leaf disease detection plays a critical role in safeguarding crop yield and supporting sustainable farming practices in regions where cassava is a primary food source. The task remains challenging due to visually similar disease symptoms, variability in lesion shapes and colors, inconsistent lighting conditions in field images, and overlapping infections. Existing deep learning and hybrid vision models, although effective in controlled environments, often suffer from high computational demands and limited capability to jointly capture fine lesion textures and long-range spatial relationships. To address these limitations, this research introduces Lite-RGA-GTNet, a lightweight residual graph–augmented graph-transformer network with spectral–directional preprocessing and progressive token pruning. The approach integrates RGB data with directional gradient and vegetation index maps, employs residual graph reasoning before attention layers, and fuses local–global features through hierarchical graph–transformer modules to produce compact yet context-rich representations. Experimental evaluation on a benchmark cassava leaf image dataset, consisting of five classes including healthy and diseased samples, demonstrates that Lite-RGA-GTNet achieves 96.84% accuracy, 96.25% precision, 96.72% recall, and 96.48% F1-score, surpassing existing models such as CassNet and LeafXFormer by up to 2.65% in accuracy, while maintaining an average inference time of 14 ms—indicating its suitability for real-time agricultural deployment.</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

Lightweight residual graph augmented transformer for cassava leaf disease recognition using spectral directional features

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月30日

查看全文
  • T. Satheesh,
  • M. S. Geetha Devasena

摘要

Cassava leaf disease detection plays a critical role in safeguarding crop yield and supporting sustainable farming practices in regions where cassava is a primary food source. The task remains challenging due to visually similar disease symptoms, variability in lesion shapes and colors, inconsistent lighting conditions in field images, and overlapping infections. Existing deep learning and hybrid vision models, although effective in controlled environments, often suffer from high computational demands and limited capability to jointly capture fine lesion textures and long-range spatial relationships. To address these limitations, this research introduces Lite-RGA-GTNet, a lightweight residual graph–augmented graph-transformer network with spectral–directional preprocessing and progressive token pruning. The approach integrates RGB data with directional gradient and vegetation index maps, employs residual graph reasoning before attention layers, and fuses local–global features through hierarchical graph–transformer modules to produce compact yet context-rich representations. Experimental evaluation on a benchmark cassava leaf image dataset, consisting of five classes including healthy and diseased samples, demonstrates that Lite-RGA-GTNet achieves 96.84% accuracy, 96.25% precision, 96.72% recall, and 96.48% F1-score, surpassing existing models such as CassNet and LeafXFormer by up to 2.65% in accuracy, while maintaining an average inference time of 14 ms—indicating its suitability for real-time agricultural deployment.