Scholarlink

<p>Guava (<i>Psidium guajava</i>&#xa0;L<i>.</i>) is a&#xa0;vital tropical fruit in terms of nutrition and economic value. Correctly identifying its maturity stage is essential for extending shelf life, maintaining nutritional quality, and meeting export standards. However, this is challenging due to the changes in color, texture, and aroma during ripening. Three guava varieties (‘Local Sindhi’, ‘Thadhrami’, and ‘Riyali’) were classified into three maturity levels using a&#xa0;dataset of 2309 images: 232 green, 829 mature green, and 1248 ripe guavas. Three convolutional neural network (CNN) models (MobileNetV2, ResNet50, and VGG&#xa0;19) were trained on 80% of the dataset and tested on the remaining 20%. Classification accuracy was 85.7% for MobileNetV2, 86.6% for ResNet50, and 82.9% for VGG19. Evaluation metrics included sensitivity, specificity, precision, and F1&#xa0;score. ResNet50 achieved the highest performance. The results confirm the potential of deep learning techniques for accurately determining guava maturity levels. Such methods can support producers in making informed decisions, minimizing waste, and improving overall fruit quality for markets.</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

Benchmarking Analysis of CNN Models for Maturity Levels of Guavas

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Ayşegül Toprak

摘要

Guava (Psidium guajava L.) is a vital tropical fruit in terms of nutrition and economic value. Correctly identifying its maturity stage is essential for extending shelf life, maintaining nutritional quality, and meeting export standards. However, this is challenging due to the changes in color, texture, and aroma during ripening. Three guava varieties (‘Local Sindhi’, ‘Thadhrami’, and ‘Riyali’) were classified into three maturity levels using a dataset of 2309 images: 232 green, 829 mature green, and 1248 ripe guavas. Three convolutional neural network (CNN) models (MobileNetV2, ResNet50, and VGG 19) were trained on 80% of the dataset and tested on the remaining 20%. Classification accuracy was 85.7% for MobileNetV2, 86.6% for ResNet50, and 82.9% for VGG19. Evaluation metrics included sensitivity, specificity, precision, and F1 score. ResNet50 achieved the highest performance. The results confirm the potential of deep learning techniques for accurately determining guava maturity levels. Such methods can support producers in making informed decisions, minimizing waste, and improving overall fruit quality for markets.