Scholarlink

<p>Fruit diseases represent a&#xa0;significant challenge to global food security and agricultural sustainability, leading to substantial economic losses. Conventional manual diagnosis remains time-consuming, subjective, and prone to inaccuracies, particularly under complex environmental conditions. Recent advances in deep learning and computer vision have improved disease detection; however, existing approaches often struggle with multi-class classification, threshold sensitivity, and varying illumination conditions. This study introduces a&#xa0;robust framework, termed “multi-model hybrid deep learning” (MMHDL), designed to enhance classification accuracy and generalization. The framework integrates three architecturally complementary convolutional neural network backbones—ResNet-50, VGG-19, and EfficientNet-B4—within a&#xa0;dual fusion strategy combining feature-level and decision-level fusion. A&#xa0;Grey Wolf Optimizer-based feature selection module reduces the high-dimensional feature space (7936 features) by approximately 75%, retaining only the most discriminative representations. The hybrid pipeline incorporates multi-backbone feature extraction, adaptive feature selection, and learnable weighted voting through softmax-based decision fusion. Evaluation conducted on the PlantVillage and Fruits-360 datasets, comprising 87,138 images across 10&#xa0;fruit types and 38&#xa0;disease classes, includes extensive data augmentation techniques such as flipping, rotation, brightness adjustment, Gaussian noise, and cutout regularization. Experimental results using fivefold cross-validation demonstrate superior performance over single-backbone, attention-based convolutional neural network, and Vision Transformer baselines, achieving 98.74% accuracy with strong precision, recall, and F1-score metrics. Real-time inference capability is validated on high-performance and edge devices, supporting deployment in precision agriculture and intelligent farming systems.</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

Multi-Model Hybrid Deep Learning Approach for Accurate and Real-Time Fruit Disease Identification Using Feature-Level and Decision-Level Fusion

  • 期刊论文

  • 发表日期：

    2026年5月20日

查看全文
  • Samruddhi Patil,
  • Balaji Bodkhe,
  • Tejas Sanjay Londhe,
  • S. H. Gawande

摘要

Fruit diseases represent a significant challenge to global food security and agricultural sustainability, leading to substantial economic losses. Conventional manual diagnosis remains time-consuming, subjective, and prone to inaccuracies, particularly under complex environmental conditions. Recent advances in deep learning and computer vision have improved disease detection; however, existing approaches often struggle with multi-class classification, threshold sensitivity, and varying illumination conditions. This study introduces a robust framework, termed “multi-model hybrid deep learning” (MMHDL), designed to enhance classification accuracy and generalization. The framework integrates three architecturally complementary convolutional neural network backbones—ResNet-50, VGG-19, and EfficientNet-B4—within a dual fusion strategy combining feature-level and decision-level fusion. A Grey Wolf Optimizer-based feature selection module reduces the high-dimensional feature space (7936 features) by approximately 75%, retaining only the most discriminative representations. The hybrid pipeline incorporates multi-backbone feature extraction, adaptive feature selection, and learnable weighted voting through softmax-based decision fusion. Evaluation conducted on the PlantVillage and Fruits-360 datasets, comprising 87,138 images across 10 fruit types and 38 disease classes, includes extensive data augmentation techniques such as flipping, rotation, brightness adjustment, Gaussian noise, and cutout regularization. Experimental results using fivefold cross-validation demonstrate superior performance over single-backbone, attention-based convolutional neural network, and Vision Transformer baselines, achieving 98.74% accuracy with strong precision, recall, and F1-score metrics. Real-time inference capability is validated on high-performance and edge devices, supporting deployment in precision agriculture and intelligent farming systems.