Scholarlink

<p>This paper presents an innovative approach for determining fruit freshness (fresh or rotten) using advanced deep learning techniques. By employing pre-trained Convolutional Neural Networks (CNNs)–such as InceptionV3, VGG16, ResNet50, DenseNet121, and EfficientNetB0–alongside data augmentation and a hybrid deep learning model that integrates CNNs with Long Short-Term Memory (LSTM) networks to capture pseudo-temporal relationships (i.e., simulated ripening progressions generated from augmented static images), we aim to enhance the accuracy and efficiency of fruit freshness classification. The suggested hybrid framework derives spatial features through CNNs while capturing pseudo-temporal relationships (i.e., simulated progressions of ripening) with LSTM layers. To increase resilience, a variety of augmentation methods, such as rotation, flipping, brightness modifications, and noise introduction, were applied. We performed experiments on the Fruits fresh and rotten dataset utilizing a stratified 70/10/20 train-validation-test distribution. The evaluation of performance relied on metrics including accuracy, precision, recall, F1-score, and ROC-AUC. Among various models evaluated, the hybrid CNN-LSTM secured the top classification accuracy of 98.9%, precision 97.5%, recall 98.1%, F1-score 97.8% surpassing the performance of isolated pre-trained models. An ablation study further showcased the unique contributions of transfer learning, LSTM incorporation, and data augmentation. Our results underscore the efficiency and scalability of the proposed hybrid architecture for practical fruit freshness classification scenarios in smart agricultural applications.</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

A hybrid CNN-LSTM model for accurate fruit freshness classification using deep learning

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月21日

查看全文
  • Debolina Ghosh,
  • Jay Prakash Singh

摘要

This paper presents an innovative approach for determining fruit freshness (fresh or rotten) using advanced deep learning techniques. By employing pre-trained Convolutional Neural Networks (CNNs)–such as InceptionV3, VGG16, ResNet50, DenseNet121, and EfficientNetB0–alongside data augmentation and a hybrid deep learning model that integrates CNNs with Long Short-Term Memory (LSTM) networks to capture pseudo-temporal relationships (i.e., simulated ripening progressions generated from augmented static images), we aim to enhance the accuracy and efficiency of fruit freshness classification. The suggested hybrid framework derives spatial features through CNNs while capturing pseudo-temporal relationships (i.e., simulated progressions of ripening) with LSTM layers. To increase resilience, a variety of augmentation methods, such as rotation, flipping, brightness modifications, and noise introduction, were applied. We performed experiments on the Fruits fresh and rotten dataset utilizing a stratified 70/10/20 train-validation-test distribution. The evaluation of performance relied on metrics including accuracy, precision, recall, F1-score, and ROC-AUC. Among various models evaluated, the hybrid CNN-LSTM secured the top classification accuracy of 98.9%, precision 97.5%, recall 98.1%, F1-score 97.8% surpassing the performance of isolated pre-trained models. An ablation study further showcased the unique contributions of transfer learning, LSTM incorporation, and data augmentation. Our results underscore the efficiency and scalability of the proposed hybrid architecture for practical fruit freshness classification scenarios in smart agricultural applications.