Scholarlink

<p>A practical and explainable deep learning (DL) framework for the multi-classification of fruit quality conditions is developed in this study. The proposed system provides a non-intrusive and scalable automated quality inspection solution for fruits, addressing concerns related to food adulteration and spoilage. This approach achieves high accuracy and interpretability by combining DL models with optimization and visual explanations based on Grad-CAMs, and it has potential applications in agriculture, retail, and food safety. The dataset comprises 10,160 images of five common fruits (apple, banana, grape, mango, and orange), with two quality classes (formalin-mixed, fresh, rotten) for each fruit and three quality classes for the rotten classification (formalin-mixed, fresh, rotten). The data is separated into 70% for training, 15% for validation, and 15% for testing to achieve good learning performance and generalization. To overcome the problem of overfitting, the sixfold cross-validation method is employed to validate its stability. This study analyzes four DL models, including Vision Transformer (ViT), InceptionResNetV2, RepVGG, and CoAtNet. The latter is compared with the method presented herein, SA-CoViT-XAI, which combines CoAtNet with simulated annealing (SA) for hyperparameter tuning and Grad-CAM for interpretability. The developed model is intended to possess both reliable predictive performance and interpretability features. All models performed well in classifying multiple classes in the multi-class dataset. The proposed SA-CoViT-XAI model achieved the best accuracy, precision, recall, F1-score, and specificity of 99.61%, 99.71%, 99.60%, 99.65%, and 99.81%, respectively. It was also superior to the other models in terms of specificity, precision, recall, and F1 Score. The Grad-CAM visualizations demonstrated that the model successfully identifies and highlights regions of the fruit that are indicative of decay. The outcomes indicate that the SA-CoViT-XAI model is robust and explainable, and potentially applicable to real-world cases in fruit quality supervision, food detection, and grading. It uses simulated annealing for end-to-end training. Simultaneously, Grad-CAM offers rich information on decision-making; hence, the approach can be positioned as a valuable development in smart agriculture as well as AI-based food inspection.</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-class Fruit Freshness and Adulteration Detection Using Deep Learning Models Optimized by Simulated Annealing and Grad-CAM

  • 期刊论文

  • 发表日期：

    2025年9月4日

查看全文
  • Saranya S.,
  • Dhanya D.,
  • Saravanan Srinivasan,
  • Rose Bindu Joseph P.,
  • Suresh kulandaivelu,
  • Sandeep kumar Mathivanan

摘要

A practical and explainable deep learning (DL) framework for the multi-classification of fruit quality conditions is developed in this study. The proposed system provides a non-intrusive and scalable automated quality inspection solution for fruits, addressing concerns related to food adulteration and spoilage. This approach achieves high accuracy and interpretability by combining DL models with optimization and visual explanations based on Grad-CAMs, and it has potential applications in agriculture, retail, and food safety. The dataset comprises 10,160 images of five common fruits (apple, banana, grape, mango, and orange), with two quality classes (formalin-mixed, fresh, rotten) for each fruit and three quality classes for the rotten classification (formalin-mixed, fresh, rotten). The data is separated into 70% for training, 15% for validation, and 15% for testing to achieve good learning performance and generalization. To overcome the problem of overfitting, the sixfold cross-validation method is employed to validate its stability. This study analyzes four DL models, including Vision Transformer (ViT), InceptionResNetV2, RepVGG, and CoAtNet. The latter is compared with the method presented herein, SA-CoViT-XAI, which combines CoAtNet with simulated annealing (SA) for hyperparameter tuning and Grad-CAM for interpretability. The developed model is intended to possess both reliable predictive performance and interpretability features. All models performed well in classifying multiple classes in the multi-class dataset. The proposed SA-CoViT-XAI model achieved the best accuracy, precision, recall, F1-score, and specificity of 99.61%, 99.71%, 99.60%, 99.65%, and 99.81%, respectively. It was also superior to the other models in terms of specificity, precision, recall, and F1 Score. The Grad-CAM visualizations demonstrated that the model successfully identifies and highlights regions of the fruit that are indicative of decay. The outcomes indicate that the SA-CoViT-XAI model is robust and explainable, and potentially applicable to real-world cases in fruit quality supervision, food detection, and grading. It uses simulated annealing for end-to-end training. Simultaneously, Grad-CAM offers rich information on decision-making; hence, the approach can be positioned as a valuable development in smart agriculture as well as AI-based food inspection.