Scholarlink

<p>Computer vision-based automated and precise fruit and vegetable disease detection is essential in precision agriculture. However, many of the current deep learning methods are computationally inefficient and only classify the disease but do not estimate its location or severity. To address these issues, we suggest a&#xa0;lightweight two-stage deep learning architecture, which also incorporates a&#xa0;version of ResNet50 (ResNet50-Lite) to classify the disease, and a&#xa0;specialized segmentation network, FruitSegNet, to extract diseased lesions and assess their severity at pixel scale. Two datasets are used to assess the framework: (1)&#xa0;a&#xa0;self-made fruit and vegetable disease dataset (containing over 800 images covering 9&#xa0;crop types and 21&#xa0;categories) that includes various disease types with annotated lesion areas and (2)&#xa0;an open-access (Kaggle) fruit disease dataset (containing approximately 3000 images representing 14&#xa0;common fruits and vegetables across 28&#xa0;classes) to test the generalization performance. The novelty of the self-made dataset lies in its rich, multi-level annotations hand-constructed by experts specifically to support supervised lesion segmentation and severity tracking. Large-scale experiments prove that the ResNet50-Lite can achieve a&#xa0;total classification accuracy of 92% on the self-generated dataset and 87% on the publicly available dataset, surpassing other lightweight models, including EfficientNetB0, MobileNetV2, DenseNet121, InceptionV3, and VGG16. The FruitSegNet segmentation model has a&#xa0;mean intersection over union (IoU) of 87.4% and a&#xa0;Dice coefficient of 93.1%, showing accuracy in localizing the areas affected by disease. Furthermore, the level of disease severity is estimated quantitatively by the area of the segmented lesion and is classified as mild, moderate, or severe. The proposed framework provides an efficient, interpretable, and deployable solution for real-time detection, validation, and measurement of the severity of fruit and vegetable diseases in real-world agricultural settings.</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

Fruit and Vegetable Disease Detection Using a Lite Transfer Learning Model and FruitSegNet

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Anuradha Shivajirao Varal,
  • Shabnam Farook Sayyad

摘要

Computer vision-based automated and precise fruit and vegetable disease detection is essential in precision agriculture. However, many of the current deep learning methods are computationally inefficient and only classify the disease but do not estimate its location or severity. To address these issues, we suggest a lightweight two-stage deep learning architecture, which also incorporates a version of ResNet50 (ResNet50-Lite) to classify the disease, and a specialized segmentation network, FruitSegNet, to extract diseased lesions and assess their severity at pixel scale. Two datasets are used to assess the framework: (1) a self-made fruit and vegetable disease dataset (containing over 800 images covering 9 crop types and 21 categories) that includes various disease types with annotated lesion areas and (2) an open-access (Kaggle) fruit disease dataset (containing approximately 3000 images representing 14 common fruits and vegetables across 28 classes) to test the generalization performance. The novelty of the self-made dataset lies in its rich, multi-level annotations hand-constructed by experts specifically to support supervised lesion segmentation and severity tracking. Large-scale experiments prove that the ResNet50-Lite can achieve a total classification accuracy of 92% on the self-generated dataset and 87% on the publicly available dataset, surpassing other lightweight models, including EfficientNetB0, MobileNetV2, DenseNet121, InceptionV3, and VGG16. The FruitSegNet segmentation model has a mean intersection over union (IoU) of 87.4% and a Dice coefficient of 93.1%, showing accuracy in localizing the areas affected by disease. Furthermore, the level of disease severity is estimated quantitatively by the area of the segmented lesion and is classified as mild, moderate, or severe. The proposed framework provides an efficient, interpretable, and deployable solution for real-time detection, validation, and measurement of the severity of fruit and vegetable diseases in real-world agricultural settings.