Scholarlink

<p>In a rapidly shifting agricultural landscape, farmers face the critical challenge of managing plant leaf diseases that endanger their crops and, ultimately, food security. Traditional methods of disease detection often struggle to cope with data imbalance, variability in leaf appearance due to environmental factors, and the complexity of multiple co-existing diseases. This situation becomes even more critical with the emergence of rare diseases that can devastate crops if not identified early. To address these challenges, an integrated framework AgriVisionNet is introduced to revolutionize plant disease detection and classification. At its core, AgriVisionNet employs a Multi Generative Adversarial Network (GAN) to generate diverse synthetic images of plant leaves. This innovative approach expands the training dataset significantly, allowing the model to learn from a wide array of disease manifestations, including those of newly emerging diseases. By effectively augmenting the dataset, we enhance the model’s ability to adapt to various conditions and ensure robust performance in real-time scenarios. For disease detection and severity classification, AgriVisionNet utilizes the Class-Aware MultiScale Vision Transformer (CMViT) Network. The encoder incorporates Multiscale Contextual Feature Attention (MCFA), which captures detailed disease patterns at various scales, enabling the model to detect both fine and coarse features indicative of early-stage diseases. On the decoder side, Class-Aware Grouped Query Attention (CGQA) organizes attention into specialized groups that focus on different regions of the leaf. This targeted approach enhances detection and classification accuracy, allowing reliable categorization into Early (30%), Moderate (30%-70%), and Severe stages (above 70%) with percentages. This timely assessment empowers farmers to take appropriate action to protect their crops. Overall, AgriVisionNet demonstrates impressive performance metrics attaining an F1 score of 97%, accuracy of 97.2%, precision of 96.5%, and recall of 96.8% on benchmark datasets, underscoring its effectiveness in supporting sustainable agricultural practices and bolstering food security.</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

AgriVisionNet: multi-generative adversarial network and Class-Aware MultiScale Vision Transformer for plant disease detection and severity classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月22日

查看全文
  • Preethi Sambandam Raju,
  • A R Revathi,
  • Murugan Mahalingam

摘要

In a rapidly shifting agricultural landscape, farmers face the critical challenge of managing plant leaf diseases that endanger their crops and, ultimately, food security. Traditional methods of disease detection often struggle to cope with data imbalance, variability in leaf appearance due to environmental factors, and the complexity of multiple co-existing diseases. This situation becomes even more critical with the emergence of rare diseases that can devastate crops if not identified early. To address these challenges, an integrated framework AgriVisionNet is introduced to revolutionize plant disease detection and classification. At its core, AgriVisionNet employs a Multi Generative Adversarial Network (GAN) to generate diverse synthetic images of plant leaves. This innovative approach expands the training dataset significantly, allowing the model to learn from a wide array of disease manifestations, including those of newly emerging diseases. By effectively augmenting the dataset, we enhance the model’s ability to adapt to various conditions and ensure robust performance in real-time scenarios. For disease detection and severity classification, AgriVisionNet utilizes the Class-Aware MultiScale Vision Transformer (CMViT) Network. The encoder incorporates Multiscale Contextual Feature Attention (MCFA), which captures detailed disease patterns at various scales, enabling the model to detect both fine and coarse features indicative of early-stage diseases. On the decoder side, Class-Aware Grouped Query Attention (CGQA) organizes attention into specialized groups that focus on different regions of the leaf. This targeted approach enhances detection and classification accuracy, allowing reliable categorization into Early (30%), Moderate (30%-70%), and Severe stages (above 70%) with percentages. This timely assessment empowers farmers to take appropriate action to protect their crops. Overall, AgriVisionNet demonstrates impressive performance metrics attaining an F1 score of 97%, accuracy of 97.2%, precision of 96.5%, and recall of 96.8% on benchmark datasets, underscoring its effectiveness in supporting sustainable agricultural practices and bolstering food security.