Scholarlink

<p>Agriculture is the foundation of global food security and economic growth. Fruits are a&#xa0;vital source of essential nutrients for a&#xa0;healthy diet. Guava, with its distinct flavor, is not just a&#xa0;delicious tropical fruit but also holds immense economic significance. In developing nations, guavas stand out for their exceptional nutritional value, loaded with vitamin&#xa0;C, dietary fiber, and powerful antioxidants like lycopene. However, the viability of guava cultivation is under severe threat from destructive diseases. One of the most alarming is anthracnose, caused by <i>Colletotrichum gloeosporioides</i>, which creates deep, dark lesions on the fruit. Fruit fly infestations, particularly those caused by the genera <i>Anastrepha</i> and <i>Bactrocera</i>, pose a&#xa0;serious threat by causing internal decay through larval feeding. The importance of accurately classifying healthy fruit for quality assurance cannot be overstated. Relying on traditional manual detection methods, which depend on visual inspections, is not only time-consuming but also prone to errors and difficult to scale. We need efficient solutions to ensure timely and reliable detection for better outcomes. Artificial intelligence effectively bridges the gaps in disease identification through automated and objective methods powered by deep learning and computer vision. Convolutional neural networks (CNNs) excel at capturing intricate local patterns but often fall short in understanding the broader context. In contrast, Vision Transformers and Swin Transformers are adept at modeling long-range dependencies, yet their high computational demands can hinder performance on domain-specific datasets where precise local features are critical. To overcome these limitations, we introduce the Triple Attention Fusion Network (TAF-Net), a&#xa0;powerful tri-branch hybrid framework designed to seamlessly integrate EfficientNetB3 for comprehensive global semantic features, DenseNet121 for optimal dense feature reuse, and a&#xa0;lightweight CNN for precise fine-grained texture extraction. An advanced attention mechanism—combining channel attention, multi-head self-attention, spatial attention, and squeeze-and-excitation convolutional block attention module (SE-CBAM) refinement—effectively integrates a&#xa0;variety of features. Meanwhile, gradient-weighted class activation mapping++ (GradCAM++) produces explainable heatmaps that enhance transparency in decision-making, ensuring trust and clarity in the results. TAF-Net sets a&#xa0;new standard in disease detection, achieving an impressive 98.69% across accuracy, precision, recall, and F1-score. The robust receiver operating characteristic (ROC) and precision-recall curves highlight its nearly flawless classification of all disease categories. GradCAM++ visualizations provide strong evidence that the model effectively targets pathologically relevant areas. Comprehensive class-wise and comparative analyses reveal that TAF-Net consistently surpasses existing state-of-the-art techniques. This paper improves fruit cultivation.</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

An Explainable Artificial Intelligence-Driven Hybrid Model with a Modified Attention Mechanism for Efficient Guava Disease Recognition

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Rajeev Kumar,
  • Anil Sandhi,
  • Reeta Bhardwaj,
  • Shrey Gandhi,
  • Manjot Singh,
  • Devander Kumra

摘要

Agriculture is the foundation of global food security and economic growth. Fruits are a vital source of essential nutrients for a healthy diet. Guava, with its distinct flavor, is not just a delicious tropical fruit but also holds immense economic significance. In developing nations, guavas stand out for their exceptional nutritional value, loaded with vitamin C, dietary fiber, and powerful antioxidants like lycopene. However, the viability of guava cultivation is under severe threat from destructive diseases. One of the most alarming is anthracnose, caused by Colletotrichum gloeosporioides, which creates deep, dark lesions on the fruit. Fruit fly infestations, particularly those caused by the genera Anastrepha and Bactrocera, pose a serious threat by causing internal decay through larval feeding. The importance of accurately classifying healthy fruit for quality assurance cannot be overstated. Relying on traditional manual detection methods, which depend on visual inspections, is not only time-consuming but also prone to errors and difficult to scale. We need efficient solutions to ensure timely and reliable detection for better outcomes. Artificial intelligence effectively bridges the gaps in disease identification through automated and objective methods powered by deep learning and computer vision. Convolutional neural networks (CNNs) excel at capturing intricate local patterns but often fall short in understanding the broader context. In contrast, Vision Transformers and Swin Transformers are adept at modeling long-range dependencies, yet their high computational demands can hinder performance on domain-specific datasets where precise local features are critical. To overcome these limitations, we introduce the Triple Attention Fusion Network (TAF-Net), a powerful tri-branch hybrid framework designed to seamlessly integrate EfficientNetB3 for comprehensive global semantic features, DenseNet121 for optimal dense feature reuse, and a lightweight CNN for precise fine-grained texture extraction. An advanced attention mechanism—combining channel attention, multi-head self-attention, spatial attention, and squeeze-and-excitation convolutional block attention module (SE-CBAM) refinement—effectively integrates a variety of features. Meanwhile, gradient-weighted class activation mapping++ (GradCAM++) produces explainable heatmaps that enhance transparency in decision-making, ensuring trust and clarity in the results. TAF-Net sets a new standard in disease detection, achieving an impressive 98.69% across accuracy, precision, recall, and F1-score. The robust receiver operating characteristic (ROC) and precision-recall curves highlight its nearly flawless classification of all disease categories. GradCAM++ visualizations provide strong evidence that the model effectively targets pathologically relevant areas. Comprehensive class-wise and comparative analyses reveal that TAF-Net consistently surpasses existing state-of-the-art techniques. This paper improves fruit cultivation.