Scholarlink

Plant diseases have significant impacts on crop yields and therefore need accurate and timely distinguishing information. This study compares eight Convolutional Neural Network (CNN) models. They are EfficientNetB0, EfficientNetB4, EfficientNetB7, ResNet50, VGG16, MobileNet, MobileNetV3Large, and InceptionV3, to differentiate between six classes of cucumber leaf diseases (Anthracnose, Bacterial wilt, Downy mildew, Fresh (healthy), Gummy stem blight, and Powdery mildew). An in-house dataset of cucumber leaf images was augmented with rotation, flipping, and scaling. The models were trained and tested based on accuracy, precision, recall, F1-score, cost types of parameters, and floating-point operations per second. As a result, ResNet50 achieves the highest accuracy of 93.06% and an F1-score of 92.98%, while MobileNetV3Large offers a trade-off between performance of 79.86% accuracy and cost, 3.12M parameters. The findings determine tradeoffs between computational cost and performance, providing advice for the design of scalable diagnostic tools for precision agriculture.

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

Comparative Analysis of Convolutional Neural Networks for Automated Classification of Cucumber Leaf Diseases

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • S. T. K. Peiris,
  • H. K. I. S. Lakmal

摘要

Plant diseases have significant impacts on crop yields and therefore need accurate and timely distinguishing information. This study compares eight Convolutional Neural Network (CNN) models. They are EfficientNetB0, EfficientNetB4, EfficientNetB7, ResNet50, VGG16, MobileNet, MobileNetV3Large, and InceptionV3, to differentiate between six classes of cucumber leaf diseases (Anthracnose, Bacterial wilt, Downy mildew, Fresh (healthy), Gummy stem blight, and Powdery mildew). An in-house dataset of cucumber leaf images was augmented with rotation, flipping, and scaling. The models were trained and tested based on accuracy, precision, recall, F1-score, cost types of parameters, and floating-point operations per second. As a result, ResNet50 achieves the highest accuracy of 93.06% and an F1-score of 92.98%, while MobileNetV3Large offers a trade-off between performance of 79.86% accuracy and cost, 3.12M parameters. The findings determine tradeoffs between computational cost and performance, providing advice for the design of scalable diagnostic tools for precision agriculture.