Scholarlink

<p>This study presents a&#xa0;hybrid Vision Transformer-Residual Neural Network (ViT-ResNet) model for automated strawberry ripeness classification, which achieves a&#xa0;98.4% classification accuracy, significantly surpassing EfficientNet (93.2%) and Residual Network-50(Resnet-50; 95.7%). The model was trained on 512 high-resolution images, which were subsequently augmented to 1500 images, leading to an accuracy increase from 92.1% to 98.4%, with precision increasing from 90.5% to 98.1% and recall increasing from 91.2% to 98.6%. Despite these enhancements, training time remained stable at 5 h. To validate the model’s stability and generalizability, a&#xa0;fivefold cross-validation approach was applied, yielding an average accuracy of 97.2% with a&#xa0;variance of ± 0.5%, indicating consistent performance across different data splits. The model was also tested for real-world robustness, outperforming traditional convolutional neural networks (CNN) in low-light conditions (96.5% vs. 89.3%), sunlight conditions (97.8% vs. 91.5%), occlusion handling (91.2% vs. 83.5%), and perspective variations (98.3% vs. 90.1%). Additionally, the lighting variation resilience ranged from 96.5% to 97.8%, whereas the occlusion robustness reached 91.2%. Class activation map (CAM) visualization confirmed that the model focuses on color intensity, texture, and seed distribution for classification. A&#xa0;spatiotemporal analysis using a&#xa0;long short-term memory (LSTM) based framework predicted ripeness progression with 95.9% accuracy on the day before harvest, improving from 82.3% accuracy 7&#xa0;days prior. Furthermore, uncertainty estimation flagged 8.4% of the classifications for manual review, ensuring high confidence in the predictions. These results establish hybrid ViT-ResNet as a&#xa0;state-of-the-art, artificial intelligence (AI) driven solution for real-time, energy-efficient, and automated strawberry classification, making it a&#xa0;scalable and reliable approach for precision agriculture and autonomous harvesting 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

Hybrid ViT-ResNet: A High-Accuracy AI Model for Automated Strawberry Ripeness Classification in Precision Agriculture

  • 期刊论文

  • 发表日期：

    2025年9月17日

查看全文
  • Eshika Jain,
  • Vinay Kukreja,
  • Pratham Kaushik,
  • Vandana Ahuja,
  • Ankit Bansal

摘要

This study presents a hybrid Vision Transformer-Residual Neural Network (ViT-ResNet) model for automated strawberry ripeness classification, which achieves a 98.4% classification accuracy, significantly surpassing EfficientNet (93.2%) and Residual Network-50(Resnet-50; 95.7%). The model was trained on 512 high-resolution images, which were subsequently augmented to 1500 images, leading to an accuracy increase from 92.1% to 98.4%, with precision increasing from 90.5% to 98.1% and recall increasing from 91.2% to 98.6%. Despite these enhancements, training time remained stable at 5 h. To validate the model’s stability and generalizability, a fivefold cross-validation approach was applied, yielding an average accuracy of 97.2% with a variance of ± 0.5%, indicating consistent performance across different data splits. The model was also tested for real-world robustness, outperforming traditional convolutional neural networks (CNN) in low-light conditions (96.5% vs. 89.3%), sunlight conditions (97.8% vs. 91.5%), occlusion handling (91.2% vs. 83.5%), and perspective variations (98.3% vs. 90.1%). Additionally, the lighting variation resilience ranged from 96.5% to 97.8%, whereas the occlusion robustness reached 91.2%. Class activation map (CAM) visualization confirmed that the model focuses on color intensity, texture, and seed distribution for classification. A spatiotemporal analysis using a long short-term memory (LSTM) based framework predicted ripeness progression with 95.9% accuracy on the day before harvest, improving from 82.3% accuracy 7 days prior. Furthermore, uncertainty estimation flagged 8.4% of the classifications for manual review, ensuring high confidence in the predictions. These results establish hybrid ViT-ResNet as a state-of-the-art, artificial intelligence (AI) driven solution for real-time, energy-efficient, and automated strawberry classification, making it a scalable and reliable approach for precision agriculture and autonomous harvesting systems.