Scholarlink

<p>Grape leaf diseases pose a&#xa0;significant challenge to viticulture, as delayed or inaccurate diagnosis can lead to substantial yield losses and deterioration in crop quality. Early and reliable disease detection is therefore essential for ensuring sustainable grape production and effective crop management. This study presents a&#xa0;comprehensive comparative analysis of deep learning-based approaches for grape leaf disease detection, focusing on a&#xa0;conventional convolutional neural network (CNN), a&#xa0;transfer learning model based on ResNet50, and a&#xa0;proposed hybrid architecture that integrates both methodologies. The hybrid model is designed to exploit the complementary strengths of shallow feature learning offered by a&#xa0;custom CNN and the deep hierarchical feature representations extracted by the pretrained ResNet50 network. Experiments were conducted on the publicly available PlantVillage grape leaf dataset to evaluate the effectiveness of each approach in a&#xa0;real-time disease recognition context. The experimental results demonstrate that the proposed hybrid model achieves a&#xa0;classification accuracy of 98.9%, outperforming the standalone CNN model (94.2%) and the ResNet50-based transfer learning model (97.1%). The hybrid model achieved an F1-score of 98.7%. Furthermore, the hybrid architecture exhibits improved generalization capability while maintaining computational efficiency, making it suitable for deployment in resource-constrained agricultural environments. By combining deep residual feature learning from ResNet50 with shallow CNN-based texture extraction, the hybrid CNN–ResNet50 model outperforms individual architectures. By improving the ability to distinguish minor illness patterns, this hierarchical multi-level feature fusion improves generalization, accelerates convergence, and increases validation accuracy. ResNet50 is justified as the foundation of the hybrid architecture since the performance benefit results from the combination of complementing features rather than just more depth. These findings highlight the potential of hybrid deep learning frameworks in balancing feature richness and efficiency, thereby offering a&#xa0;robust and scalable solution for smart agriculture and precision viticulture applications.</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

Performance Evaluation of CNN, ResNet50, and Hybrid Architectures for Automated Grape Leaf Disease Detection

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Jayashri D. Palkar,
  • Anuradha S. Deshpande

摘要

Grape leaf diseases pose a significant challenge to viticulture, as delayed or inaccurate diagnosis can lead to substantial yield losses and deterioration in crop quality. Early and reliable disease detection is therefore essential for ensuring sustainable grape production and effective crop management. This study presents a comprehensive comparative analysis of deep learning-based approaches for grape leaf disease detection, focusing on a conventional convolutional neural network (CNN), a transfer learning model based on ResNet50, and a proposed hybrid architecture that integrates both methodologies. The hybrid model is designed to exploit the complementary strengths of shallow feature learning offered by a custom CNN and the deep hierarchical feature representations extracted by the pretrained ResNet50 network. Experiments were conducted on the publicly available PlantVillage grape leaf dataset to evaluate the effectiveness of each approach in a real-time disease recognition context. The experimental results demonstrate that the proposed hybrid model achieves a classification accuracy of 98.9%, outperforming the standalone CNN model (94.2%) and the ResNet50-based transfer learning model (97.1%). The hybrid model achieved an F1-score of 98.7%. Furthermore, the hybrid architecture exhibits improved generalization capability while maintaining computational efficiency, making it suitable for deployment in resource-constrained agricultural environments. By combining deep residual feature learning from ResNet50 with shallow CNN-based texture extraction, the hybrid CNN–ResNet50 model outperforms individual architectures. By improving the ability to distinguish minor illness patterns, this hierarchical multi-level feature fusion improves generalization, accelerates convergence, and increases validation accuracy. ResNet50 is justified as the foundation of the hybrid architecture since the performance benefit results from the combination of complementing features rather than just more depth. These findings highlight the potential of hybrid deep learning frameworks in balancing feature richness and efficiency, thereby offering a robust and scalable solution for smart agriculture and precision viticulture applications.