Scholarlink

<p>Tomato cultivation is widespread worldwide, and the crop serves as a&#xa0;nutrient-tionally significant vegetable globally. Yet a&#xa0;variety of plant diseases severely threatens their productivity. To address the challenge of timely diagnosis of these diseases, deep learning models have been proposed, which have shown strong per- Performance on laboratory-based tomato leaf datasets. However, their accuracy often deteriorates in real-world field environments. This degradation in performance is due to complex backgrounds and the high variability among field images belonging to the same disease class. To overcome these challenges, this paper introduces a&#xa0;two-stage framework that improves tomato disease classification under field conditions. The proposed approach minimizes background interference during training and also mitigates inter-class variability. In the first stage, a&#xa0;Faster Region-based Convolutional Neural Network (FRCNN) Integrated with Feature Pyramid Networks (FPN) is employed. This stage is used to detect and isolate tomato leaves. This ensures the effective removal of non-essential background content. In the second stage, a&#xa0;variability factor is estimated using Fuzzy C‑Means (FCM) clustering. In this stage, disease-similarity weights are assigned to each field image. These weights represent the variability among classes-specific field samples. These weights are then incorporated into a&#xa0;custom loss function of deep learning models. It enhances training robustness and identifies diseased regions effectively under highly variable field conditions. An evaluation was conducted using two datasets: the FieldPlant dataset (real-world images) and the PlantVillage dataset (images collected from laboratories). The model, FRCNN-FCMWeight, achieved 87% accuracy, while baseline CNNs achieved about 82% accuracy. The data collected from laboratory images were found to significantly improve tomato disease classification under diverse real field conditions and to control the influence of environmental variability. As such, the model provides strong support for agricultural use cases and improves overall model generalization.</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 Robust Method for Real-Field Tomato Disease Classification Using FRCNN-Based Leaf Isolation and FCM-Guided Variability Estimation

  • 期刊论文

  • 发表日期：

    2026年7月16日

查看全文
  • Shashank Yadav,
  • Anand Shanker Tewari

摘要

Tomato cultivation is widespread worldwide, and the crop serves as a nutrient-tionally significant vegetable globally. Yet a variety of plant diseases severely threatens their productivity. To address the challenge of timely diagnosis of these diseases, deep learning models have been proposed, which have shown strong per- Performance on laboratory-based tomato leaf datasets. However, their accuracy often deteriorates in real-world field environments. This degradation in performance is due to complex backgrounds and the high variability among field images belonging to the same disease class. To overcome these challenges, this paper introduces a two-stage framework that improves tomato disease classification under field conditions. The proposed approach minimizes background interference during training and also mitigates inter-class variability. In the first stage, a Faster Region-based Convolutional Neural Network (FRCNN) Integrated with Feature Pyramid Networks (FPN) is employed. This stage is used to detect and isolate tomato leaves. This ensures the effective removal of non-essential background content. In the second stage, a variability factor is estimated using Fuzzy C‑Means (FCM) clustering. In this stage, disease-similarity weights are assigned to each field image. These weights represent the variability among classes-specific field samples. These weights are then incorporated into a custom loss function of deep learning models. It enhances training robustness and identifies diseased regions effectively under highly variable field conditions. An evaluation was conducted using two datasets: the FieldPlant dataset (real-world images) and the PlantVillage dataset (images collected from laboratories). The model, FRCNN-FCMWeight, achieved 87% accuracy, while baseline CNNs achieved about 82% accuracy. The data collected from laboratory images were found to significantly improve tomato disease classification under diverse real field conditions and to control the influence of environmental variability. As such, the model provides strong support for agricultural use cases and improves overall model generalization.