Scholarlink

Timely and accurate detection of plant diseases is essential for safeguarding crop yields and ensuring sustainable agricultural practices. This study presents a deep learning-based framework for guava leaf disease classification that leverages transfer learning and attention mechanisms to enhance model performance. Five state-of-the-art convolutional neural networks (CNNs)—DenseNet201, EfficientNetV2B0, InceptionV3, MobileNetV2, and Xception—were evaluated on the Guava Foliar Lesion Dataset (GFLD) to benchmark their feature extraction capabilities. While DenseNet201 achieved the highest baseline accuracy (85.99%), its recall and F1-score remained relatively low. To address this, we propose DAGuavaNet, an enhanced architecture that integrates a squeeze-and-excitation (SE) attention module into DenseNet201. DAGuavaNet maintained the same accuracy (85.99%) but significantly improved recall (from 79.89% to 85.99%) and F1-score (from 80.22% to 85.81%), indicating more balanced and consistent performance across disease classes. These results demonstrate that attention-enhanced CNNs can provide a robust and effective solution for automated guava leaf disease detection.

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 Deep Learning Framework with Squeeze-and-Excitation Attention for Guava Disease Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Tuong Le,
  • Duc-Tan Nguyen,
  • Hoang-Duy Nguyen,
  • Thang Cap

摘要

Timely and accurate detection of plant diseases is essential for safeguarding crop yields and ensuring sustainable agricultural practices. This study presents a deep learning-based framework for guava leaf disease classification that leverages transfer learning and attention mechanisms to enhance model performance. Five state-of-the-art convolutional neural networks (CNNs)—DenseNet201, EfficientNetV2B0, InceptionV3, MobileNetV2, and Xception—were evaluated on the Guava Foliar Lesion Dataset (GFLD) to benchmark their feature extraction capabilities. While DenseNet201 achieved the highest baseline accuracy (85.99%), its recall and F1-score remained relatively low. To address this, we propose DAGuavaNet, an enhanced architecture that integrates a squeeze-and-excitation (SE) attention module into DenseNet201. DAGuavaNet maintained the same accuracy (85.99%) but significantly improved recall (from 79.89% to 85.99%) and F1-score (from 80.22% to 85.81%), indicating more balanced and consistent performance across disease classes. These results demonstrate that attention-enhanced CNNs can provide a robust and effective solution for automated guava leaf disease detection.